The economic implications of reciprocity in teams and markets [Elektronische Ressource] / Elke Renner
99 Pages

The economic implications of reciprocity in teams and markets [Elektronische Ressource] / Elke Renner


Downloading requires you to have access to the YouScribe library
Learn all about the services we offer


The economic implications of reciprocity in teams and markets Elke Renner 2005 DISSERTATION zur Erlangung eines Grads einer Doktorin der Wirtschaftswissenschaft (Dr. rer. pol.) der Universität Erfurt Staatswissenschaftliche Fakultät Gutachter: Prof. Dr. Bettina Rockenbach Prof. Dr. Manfred Königstein Datum der Disputation: 19.12.2005 urn:nbn:de:gbv:547-200600171 [] 2 Acknowledgements Many thanks to Klaus Abbink, Christoph Albrecht, Andrea Bäcker, Simon Gächter, Mario Gruppe, Özgür Gürerk, Sebastian Händschke, Bernd Irlenbusch, Manfred Königstein, Jürgen Kumbartzki, Thomas Lauer, Rosemarie and Peter Meyer, Mark Peacock, Christiane Pilz, Bettina Rockenbach, Karim Sadrieh, Paul and Dorothea Schindegger, Martin Sefton, Martina Stamm, Jean-Robert Tyran, Anne Ulmke, Katrin Uschmann, Gerhard Wegner, Tim Wenniges and Egon Zimpel. 3 Contents 1 Introduction and overview 6 1.1 References 8 2 Leading by example in the presence of free rider incentive 9 2.1 Introduction 9 2.2 Related literature 10 2.3 Design and procedures 10 2.3.1 Design 10 2.3.2 Procedures 12 2.4 Leading by example in a one-shot experiment – a benchmark 13 2.4.1 Results 13 2.5 The dynamics of leadership and voluntary cooperation 15 2.5.1 Does the leader matter? 15 2.5.2 Are there good and bad leaders? 17 2.5.3 The impact of increased gains from cooperation 19 2.5.



Published by
Published 01 January 2006
Reads 127
Language English

The economic implications of reciprocity in teams and markets

nner eke RlE



aft (Dr. rer. hder Wirtschaftswissensczur Erlangung eines Grads einer Doktorin pol.) der

Universität Erfurt

Staatswisshe Fakultät enschaftlic


eProf. Dr. Bttina Rockenbach Prof. Dr. Manfred Königstein

ion: Datum der Disputat




ledgements Acknow

Many thanks to Klaus Abbink, Christoph Albrecht, Andrea Bäcker, Simon Gächter, Mario Gruppe,
ÖzgThomaür Gsü Lareuer, Rork, Sebastian Händsemarie and Peter Meschke, Bernd Irleyer, Mark Peaconbucsch, Manfred Kk, Christiönigstein, Jüane Pilz, Bettina Rorgen Kumbckenartbazki, ch,
Anne UlmKarim Sadrike, eKatrin Uschmh, Paul and Dorotheann, Gearha Schirdndeg Wegnger, Maer, Timrt Wennigein Sefton, Martina Stams and Egon Zimpm,el. Jean-Robert Tyran,


nts Conte 1 1.1 2 2.1 2.2 2.3 2.3.1 2.3.2 2.4 2.4.1 2.5 2.5.1 2.5.2 2.5.3 2.5.4 2.5.5 2.6 2.7 3 3.1 3.2 3.3 3.3.1 3.3.2 3.4 3.5 3.6 4 4.1 4.2 4.3 4.3.1 4.3.2 4.3.3 4.4

wnd overviection aIntroduReferences
Leading by example in the presence of free rider incentive
ion tIntroduc uretd literaRelateDesign and procedures
signDes reProceduLeading by example in a one-shot experiment  a benchmark
sultsReThe dynamics of leadership and voluntary cooperation
Does the leader matter?
Are there good and bad leaders?
The impact of increased gains from cooperation
ample?xng by eLeadi ader?Shall I be a bold or timid leding remarks lucConReferences
When in doubt...  Cooperation and punishment under incomplete
information ion tIntroducdures d procegn andesiExperimental sultsReur bution behavioContri rehavioument bPunish sationcWelfare impliConclusions
Group size and social ties in microfinance institutions
ion tIntroduces studirimental d expeRelate signental ded experimModel an elThe modents Treatmures edcExperimental pro sultsRe

6 8 9 9 10 10 10 12 13 13 15 15 17 19 22 23 24 25 27 27 29 30 30 33 37 38 40 41 41 43 44 45 46 47 48


4.4.1 The effect of the group size
play cs of Dynami 4.4.2cial ties The effect of so 4.4.3s er effectGend Summary and discussion
4.6 References
5 Price rigidity in customer markets
ion tIntroduc 5.15.2 An experimental approach to price rigidity in customer markets
5.3 Experimental design
5.3.1 General description of the experimental design
5.3.2 Procedures, parameters and information conditions
5.3.3 Predictions
sultsRe 5.45.5 Concluding remarks
5.6 References
ns Instructio:Appendix 1.A nd 1.b Tables 1.a a:Appendix 1.Bns Instructio:Appendix 2.A 2.d Tables 2.a :Appendix 2.Bns Instructio:Appendix 3.Ans Instructio:Appendix 4.AAppendix 4.B: Screens announcing the cost shock in period 3
Appendix 4.C: Figure 4.a

50 50 51 52 52 54 56 56 57 58 58 59 60 62 68 69 71 77 79 85 89 91 97 98


Introduction and overview 1 TraditioThe question nally economiof how to dests hasign efficient eve built this analysiconsomic o on thrganeoretiizatiocns ial mosdels a a key issssuming ue in economic afully rational and nalysis.
tractablpurely selfish e matagenthematical mos. This assumptiodels. Howeven has a clr, it has oftear adven been criticiseantage in its simplicity whid as unrealistcih alloc. In the last few ws building
decades a growing number of researchers in economics have investigated actual human behaviour in
controlled laboratory experiments. These experiments explore the boundaries of human rationality in
search for alternative behavioural assumptions that can feed back into the development of new
models with more explanatory power.
One central finding in this field is the prominent impact of reciprocity. Many experiments have
kind ademoncts strated(positive reci that people arprocity) and e not unbouthey punish unndedly selfishkind but acts (neg care aboative reciput others wellrocity). The existen being. They rewace of rd
positive reciprocity has been demonstrated in gift exchange and trust games (e.g. Berg, Dickhaut and
rejeMcCabe, 19ction beh9a5 and Fehviour in ultir, Kirchsmatum gameteiger asn (Gdüth, Riedl, 19Schm93ittberge). Negative rer and Schwarze, ciprocity ca1982n be ob) and in servepunid in shment
behaviour in public good games (Fehr and Gächter, 2000). In the moonlighting game (Abbink,
Irlenbuthese asch annd mad Renny other expener, 2000) poriments resitive aciprnd negocal bativehavioe recipur carocn be conity ares obseiderred aved sis a multaneostylized fauslyct (Came. From rer
). 2003environThis thesmentis investigates studied in ths the ecoe next fournomic ch implicatapters shions of recipare several corocmmon fity in teams and meatures.arkets. The First, in all of them,
and group traditional thesettory sugings invegesstigts oppoated efficienrtunistic behaviocy requires ur byind individualividuals to cos will undermine eoperate, but they have privfficiency. In the team ate
incentives to free ride; in the market setting we investigate efficiency requires consumers to trust
selleto distrurs nst sellot to supers. Seply subcon-dstanda, in the four envrd producironmts, but moralents we hazard pr study, the inefficienoblems give concy prosbumerlem may be s good reason
considmitigated throerations caugh recipn overcorocity. These chame team inefficienpters tocy and gethemar showrket failure d that whether in faepends on sect reciveral factoprocity rs. First,
and moinformatiost obviously, individn that allows themuals mu to act on their rest be sufficiecipnroctly real ciprocal in nature. Secomotives in a way that improves group ond, they must hautcomeve the s.
act oppThird, they mortuniustically. st have instruments at their disposal to effectively monitor and discipline individuals who
to leading by A more detailexample in teed outline of the spams anecifid do lc reseaeaderrs mch questionsake a differen is asce? followsSecon: Firsd, can t, doespeer p reciuniprocshmity give risent e
enhance cooperation in teams when it is difficult to identify free riders? Third, can reciprocity sustain
group lending schemes in microfinance institutions? Fourth, can reciprocal loyalty help overcome
market failures in markets with incomplete information about quality? Each of these four questions will
be addressed in separate chapters. The next section gives a brief overview.
The second chapter examines the impact of positive reciprocity in the form of leading by example in
the context of a simple team cooperation problem with free rider incentives. In a public goods game
linear pone playeublirc was a good. Hissignes contribution was d to the role of the leadannouner anced tod had to cho three followeors se how muwho then dch toecid contried bute to a
recisimultaneouprocity, their cosly on their conntributiotns will ribution deincreacisionse with the les. If followers aader's rcontributie sufficiently stronglon. By allowing the ley motivated by ader to set
wholan example, e growe provide thup. We find that leadee leader and follr with oweran instrume contributnt to harneions are poss thsitively correlese motivations to benated even in a efit the one-
in a repeshot gamated leadee, testifying to the relevanrship game did not ence of recihanprocity.ce the ov Despite thierall level of coopesration , we find that the presenrelative to a situation ce of a leader
peoplwith an abe follow sethe leadent leader. Even thougrs exh on ample, the leadeaveragrs e it paywho do so eas to be a bolrn led leadess than follor becaweuse rs ansufficied leaders are ntly many
reduce theinot willing to tolerate thir contributiosns. : they resist being taken for a sucker and the only way they can resist is to
The third game. Previochapus reseter examinarch hesa negs showative ren that recipciprocity inroc the form of peeally motivated playr puniers arshme willingent in a public to punish free goods


rideimprovinrs, ang grod soup outcome the opportunis. Here ty to punish pwe studrovidey the effectivs an instrument for eness of this instrumdisciplining free rideent when it is difficult to rs and
identify free rito evaluate others' contribders. The exputerimeions to a pubntal setlic gooup discrid. minaIn both treatmenttes between ts contributiohe ability to obsens are orbve and thservable but e ability
taken. When endoonly in one treatment are the individual wments are publicly known puniendowmenshts publicment exerts aly know strong n before pdiscipluniishmening effect ant decisind ons are
cooperation rates are high. In contrast, when endowments are private information, cooperation rates
reluare subctanstantice to punially lower ash; rathendr punishment those who are punihas virtuallyshed n doo not cha effect. Surpnge risingtheir behavioly this is not due to a ur in response to
punishment. We conclude that transparency about others endowments is crucial for successful
peration. cooThe fourth chapter demonstrates that reciprocity can sustain group lending schemes in microfinance
seleprogramcted gmreous. These props. In a typical microfinangrammes provice schde pooer peome bplorrowers withe with small loan individual ris given tosky projec jointly liable self-ts form groups
which apply for loans together. The whole group is liable if any group members default. Thus, joint
that recipliability providrocites iny based gsuranrcoue againp lendinsg outt individual riperforms insks butdi relies ovidual lendinn recig.proc This real solidasult is robrust ity alone. We showith respewc t
al ties. size and socito group long-teThe last rm relchapationshipter analyses s bereciptween brocuyers and sellity in customer maers thatrkets. Cu evolve if stomerbuyers tru marketss aret sellers to provi characterizede high d by
antagoquality and if sellenize customers arrs ande trus disrutworpthy. Hot the rewever, lationship. changeWe experims in the termentally shos of this impliwcit cont that mutually benract may eficial
long-tethat price rm relrigiationships fredity after a temporary coquently prevast shocil in marketk is mus for expech more prrienonceo goodunced if prices. Howe increases ver, we also show cannot be
of the lemonjustified by cost increas type, but arses. Hene pronce, lone to prig-tce stierm relckiatineonss. ships in customer markets mitigate market failure
Taken together the four chapters of this thesis provide evidence for the power and limitations of
reciPositive reprocity anciprocity is d illustrate thstronge signifi enoucant egh to sustain miffect altecrrationofinans in ince institutional rulestitutions and hels can psexert on outcomes. to mitigate market
failure in customer markets. Substantial positive reciprocity is also observed in leadership games.
However, the effect is not strong enough to attain efficient team outcomes. Substantial negative
recitranspaprocity herenclps to suy about othestain rs capcooperation iabilities to n peercoop perate punishunishmment gameent is dse. Yet, if trimental. there is a lack of


1.1 References

Abbink, Klaus, Bernd Irlenbusch, and Elke Renner (2000): The Moonlighting Game  An Experimental
Study on Reciprocity and Retribution. Journal of Economic Behavior and Organization, 42, 265-277.

and EcoBerg, Joycnomie, Johcn Behavio Dickhaur, 10, 122-142. t, and Kevin McCabe (1995): Trust, Reciprocity and Social History. Games

Camerer, Colin (2003): Behavioral Game Theory: Experiments in Strategic Interaction, Princeton
University Press

Fehr, ErnAmerican Economist and Simon Gäc Reviewchter , 90 (4(2000): ), 980Coo-94. peration and Punishment in Public Goods Experiments,

Fehr, ErnsExperimental Investigation.t, Georg Kirchste Quarteiger, anrly Jod Arno Riedurnal of Econl (1993): Doeomics, 108, 4s Fair37-4ness Preve59. nt Market Clearing? An

BargaiGüth, Wernening. Journra, Rolf Schmittbergl of Economic Beer, anhaviod Bernd Sr and chOrgwaranize (198zation, 3 , 367-88. 2): An Experimental Analysis Ultimatium


Leading by example in the presence o2 free rider f1 incentives okay to go over the line people think itOnce you as a CEO go over the line, then selves. themLawrence Weinstein, Head of Unisys st, July 27, 2002, p.58) i (quoted after The Econom

2.1 Introduction A central question in leadership research concerns how a leader can effectively influence his
followers. Of the many types of such influence one of the most common and elementary is leading by
out unpleexample. Pubasant but socilic figures often demoally desired ancstrativities: actste exemplary of blbehavioood dour in nation ordeand giving to r to motivate peoplcharity are often e to carry
phenpublicomenized in ordeon frequently used in ordr to motivate other peer toop enhance follle to follow suit. Role modowers moetivation, espelling is a ciallwidye if spreathe desid red
other habehaviour cand, is held to detrannot be enforced by othect from the wir meanllingnesss like rewa of others to coords or penpaerate: a boss lties. Setting a bad exampcannot leave the le, on the
manager asoffice before the end of the k employees toworkin defer expeg day while dcted paye increamandses whiling that his e awarding him a rise. staff works overtime; neither can a
In this paper we approach the behavioural impact of leading by example with a series of
goodexperims gamental e) in whistudiecs. Ouh comr basicplete free ri designdi involves a ng is a domisimnant strateple team cooperatiogy but the total surplun dilemma (os wor lineauld be r public
seqmaximizeuential decisid if all players contrion process: In treatmenbute theitr whols with a le endoeadwment. We imer, one memplement leaber of a groudip is selng by examplectede as a to be the
person who decides first on his contribution to the team project. The decision of the leader is
conveyeof leadershidp experim to the followers, who theents without a nlea deder, whcide privately ere allon their o members of wn a contrigroup debution. cidTo sepe indepaendentrate the effect ly serve
struas control trectures. Thaus, assutments. It is straigmingh selfish, money maximitforward that the mere existenzing individualce ofs, the game the a leader dooeretic prs not affect iedictinon of centive
non-cooperation is the same for both settings and leadership should be irrelevant.
We believe that our simple leadership game allows us to isolate one important, and hitherto rather
inceneglecntives. Oted asur simpect of leadple game isolateership, nams exely the role of actly this aspect withexample settingout confoun in the presending it with other feace of free ridetr ures of
leadership that surely are important in reality, like managerial skills, charisma, coordination, signalling
etc. We ran two series of experiments, a strictly one-shot leader game and a finitely repeated leader game.
The formexample. Theer gives us a b second expenecrhmaimentrk on ths will allow ue importas ncto assess the dynae of a non-strategic follomics of leawingdership. of the leaders
observed iiThe main obje) whethective of our r it positively experiments iaffects the behs to examavioine i) whethur of the leader thiers and leadsubjer-effect can reects in the leadally be er role
raise their level of cooperation and iii) whether the presence of a leader enhances efficiency and
groups with leader achieve higher cooperation rates.
Our main results are as follows. We find that many followers indeed follow the leaders example and
stricontribute thectly one-shot experime more to the team projent where strategic inct the more cethe leadentives for follor puts owingn the example are ab the line. This happesens alrent. This ady in a
over to the reobservation tepeated lestifies the padeor game, whitential relevchance of leadi we set up to ng by examplstudy the dynamics of leade. We also find that this reership. Givensult carrie the s
effectiveness of leading by example, we expected to find an improved efficiency in the repeated
1This chapter is based on the working paper Leading by example in the presence of free rider
incentives, joint work with Simon Gächter.


leadership game. Yet, it turned out that leaders were not able to improve the efficiency of cooperation
sucker, sincrelative to a situation whe followers folloere leadew their lears weders examre absent. Theple only half-hea reason is thartt being a leadedly, i.e. they contriber meaunte s being a
systematically less than the leader. As a consequence leaders reduce their cooperation over time,
leadewith the consequer. Empirically, it turns nce that followers out, that the correlatireduce as well.on of l The queeader and follstion ariseowesr beh whetheaviour is stronr it pays to be a bold g enough
to make it worthwhile on average to be bold and contribute high amounts. But a large variance
ns. remaiWe proceed as follows. In the next section we put our work into perspective of the previous literature.
provideSection 2.3 ds the one-shescribes the ot leadbasicership expe decisiorins situment thatations will provide the b and the expeenchmarimental procerk for duresthe relevan. Sectiocen 2.4 of
preleadinsengce of free ride by example. Sectior incentn 2.5 will inveives. Sectionstigat 2.6 conce the dyludes. namics and difficulties of leading in the

literature 2.2 Related Although there is a huge literature on various aspects of leadership, there is to our knowledge only
one economic approach to the phenomenon of leading by example. Benjamin Hermalin (1998)
provides a signalling model in which a leader can induce following by moving first. If the leader is
privately informed on the productivity parameter of his teams project, he can increase welfare through
expanding his effort and thus, leads by example. Meidinger, Sinclair-Desgagné and Villeval (2002)
reciprovide expprocity than throuerimental evidegh singnalling. ce on HeWebrmer, Camalins moeredel ar, Rottenstreind find that leadershich and Knez p wo(2001) rks mostudy lere through adership
in a coordination game. Although they primarily focus on the attribution of leadership quality and
whether subjects exhibit the fundamental attribution error, they also find that a leaders speech only
slightly increases efficiency in large groups, whereas small groups coordinate well, independent of the
leader. ce of aenspreSince we focus on a type of leadership that induces following not by solving coordination games or
incentive problems assuming perfectly rational agents, we mainly consider the experimental literature
on related issues. The studies most akin to our set-up are the experiments by Moxnes and van der
Heijdcontext usien (2n00g0) and van d - similar to oeur Heijdr study - a en andseque Moxnes (200ntial decision pr0). They introotocolduce lead. They find that followers inership in a public bad vest less
in a public bad when a leader is introduced but this effect is attenuated if the leader faces no costs for
rship. leadeThere is al(Andreoni 2so 0a literature that inve01; Potters, Sefton and Vestigatessterlu the impond 2001rtan), ce of leain politics dership givin(Arce M. 20g01 ch; Rhodearitable funs andd raisin Wilsog n
1999) and strategic management (Miller 1990; Foss 2001).

2.3 Design and procedures

2.3.1 Design Our basic design involves a simple team cooperation dilemma (or linear public goods) game. Each of
the four team members has to decide on how many out of 20 tokens to keep and how many tokens to
contribute to a team project. For simplicity, the size of the team project is just the sum of all
contributions to it. The payoff for each subject is the following:

4πi=20−gi+α∑gj0<α<1 (1)
j=1In our experiments, the marginal per capita return of a contribution to the team project, α, was either α
= 0.4 or individual haα = 0.s an ince6 (for details see bntive to free ride oelonw). Compa the otherr ging the rmaoup memrgibernal pays contributionsoffs it is obvious that a ratio. Regardless of what nal
the other group members contribute, every individual is better off by keeping all tokens for him- or
herself, since the marginal payoff of a token contributed to the team project is only α < 1 compared to


a margimonetarnal pay income, the domyoff of 1 from keepinant strateing the token. gy in this game Thus, asis to contsuming that ributsubjectse nothing. A social dil maximize theiemmar own arises,
sincteam membere it is in the joint intere has an incest of the grountive to free ride. p to contribute the whole amount to the team project but each
henThe leadceforth er gwiame is a veryll call the leader de simple varicideation of s first in his this staor hendard publir team. The c goods gamother three group meme. One subjecbers (callet that we d
simultanthe followers) leaeouslrn aby about their contriout the leaderbutionss contrito the team projebution toct.2 the team project an The payoff functiond then decide s of all team
membestrategy of fulrs, inl free ridincluding the lega is indepeder, are identindent cal anof the decd equisioal n of others. The only Nato (1). It is straightforwarsh equilibd that the equilirium is brium full free
ts. criding by all subjeAll our experiprovide the bmaselients ne to cowere condumpare the cted with two baeffect on coopsic treatment variaeration in the presetnions: Experice of a leadements r. Table without a leader 2.1 gives
an overview of all treatments that we employed in this study.
ew. tment overviaTab. 2.1: Tret: hmark experimenBenc One-shot public goods game with or without leader
= 0.4 αNo. of Treatment Criterseleion for lction eader With leader No leader subjects
s) (group0 Randomly 20 groups 8 groups 112 (28)
Repeated puDynamic leablic goods game dership exwipth aned rimentswithou: t a leader
α = 0.4 No. of
Treatment Criterseleion for lction eader Period 1-10 Period 11-20 subjects
s) (group1 Randomly With leader 48 (12)
butor in the est contriHigh2a group in period 1-10 (ConNo leatrol for treatment 1) der With leader 36 (9)
cial) (pro-sobutor in the Lowest contri2b group in period 1-10 (ConNo leatrol for treatment 1) der With leader 36 (9)
r) de(free riDynamic leadership experiments:
α = 0.6  TheRepe impact of inated pucreasblic goods ged gainame s from coopwith and wieration on lethout a leader aders and followers
No. of Treatment Criterseleion for lction eader Period 1-10 Period 11-20 subjects
s) (group3 Randomly With leader 36 (9)
Highest contributor in the No leader
4a grou(pro-sop in periocial) d 1-10 (Control for treatment 3) With leader 20 (5)
butor in the Lowest contrider No lea 4b group in period 1-10 (Control for treatment 3) With leader 20 (5)
r) de(free ri 2 In the experiment we never talk of leaders and followers, but instead of the subject that decides
first in his or her group. See the instructions in appendix B for further details.

Treatmand compare ent 0 is a benit to a simultaneouchmark study. Hes one-shreot public g we investigoods gamate the leadee withr gaout a leademe in a purer. Since this exly one-shot siperimtuation ent
is just played once, there is no strategic reason whatsoever for the followers not to free ride, if they are
interecontristed in bute to the team projemaximizingct. Thus, ou their payoffs. A leader one-shr whot gamo anticipatee measures ts this, has as he extent andwell no reaso effectiveness of n to
efficleadinigency. by example in a very simple team leader game that requires voluntary cooperation to yield
TreatmThe saents 1me group of team  4 take the one-shmembers plot leadayer gams the same ine teamto a strate gamgic coe for ten times. Thntext of a repeated team gis study allows uasme. to
assess the dynamics of leadership in our context. We can investigate whether leaders are able to lead
their groups toward efficiency over time. To check for the robustness of our findings, we also vary the
gains from cooperation by changing α: In treatments 1 and 2, α = 0.4; in treatments 3 and 4, α = 0.6.
The main design of our experiments in treatments 1  4 is again very simple. Our subjects played a
seqDuriuenng the whole expece of ten team gamriment groues with a leps readmained uer and a nchsequanged. The subjeence of ten team gamect who hs wiad to decithout a leadede first in a r.
group (i.e., the leader) was the same person during all leader periods.
Treatment 1 is an experiment with a leader.3 The subjects who had to decide first in their group were
chosen randomly before the first period started.
Our treatment 2 will change the perspective of equally qualified leaders and investigate whether
seqsome leuenadece. This hars are better ths the added advaan others. Tontage that end, that the grsubojups ects first playwithout leed a no-laders from theadere first se and then a lqueneadece can r
serve as control groups for the groups with leaders in treatment 1. To address the impact of different
types of leadaccording to their contriers on the leabutidership pon in the first roblem, sequenthe leadece. Bers in the fore peseriod 1co1, subjend treatmects wint were chth either the hosen ighest
or the lowest average contribution in their group in the previous periods were assigned to the leader
role. We will discuss the exact rationale for these design features below.
on leadThe main pers aurpose of treatnd followers contriments 3 anbution bd 4 is to cheehaviour. ck Thefor the impa experiments ct of increwere coandused ctgains from coed in exactly the operation
gainsame s from coway as operation (4in treatment×αs 1 and 2. The ) from 1.6 to 2.4. only difference was that α = 0.6, which increased the social

2.3.2 Procedures All experiments were conducted in the computerized experimental laboratory at the University of
independeErfurt. In total 308 studnt groups of four subjents from variouects. Ths e last coldisciplineus mnparticipate of table 2.1 containd in our expesriment information hos. In total we had 77 w many
participants (and groups of four subjects) took part in the various treatments.
The suexperimbjects ental softwawere enrcoue z-trerae ged to partidevelopcied by pate thUrs Firouscgh hbachposters oer (199n9). the campus. At the beginning of eaWe used the ch session
the participants were randomly assigned to one of the booths in the laboratory. The booths visually
sepand indarateepend thed participently. The wriants and entten instrusured tctions werehat every handindied out in whividual made chis or her dh the above social dilecision aenonymmma ously
assignsituation aed, sund ch that subjethe experimecntstal pro did not knocedwures were which of explthained in de other paetairticilpants . The grouwere inps were ra their grondomly up. To make
questionssure that every participa. We did not startnt unde the experimerstood thne decit before sioall subjects han problem, they had to answed correctly answered all r a set of contcontrol rol
. sstionqueIn all treatments 1  4 subjects received the instruction only for the condition in the first ten periods
3repo We also playrt these resulted ten peris here, since ods with no lewe do noadt need ter after the lehe obseader exprvationserim from these expents were finierimeshednts for o. We do ur not
. sanalysi 12


(with othe experimr withoenut a leader t will continreuespec for anothetively). Before tr ten periohe ds asecond seqund the coenrrespondice begang instn, they were infoructions were rmedhand that ed
out. In our benchmark experiment (treatment 0) subjects received only one set of instructions, since
they made only one decision in either the game with or without a leader.
Immediately after each session subjects were paid anonymously in cash. On average, our subjects
earnlasted aed  10.pproximately 40 minutes a3 in the one-shot experimnd theen repeatts and  9.5 ed game experimin treatmentents took 7s 1  4. The on5 minutes.e-shot expe4 riments

example in a one-shot experiment  a benchmark yLeading b2.4 The set of experiments that we describe in this section will provide the benchmark for our analysis. In
these experiments the leader game is played exactly once. One of the four team members (the
leader) is randomly selected to make the first contribution decision. The rationale for the random
leader selection is straightforward. In our experiment, all team members have equal incentives (see
(1)) and no further qualifications to be a leader are necessary. Therefore, ex ante, each team member
is equally qualified to be the leader. By the one-shot design we rule out any reputation considerations,
which might be decisive in repeated interactions. Thus we have a very strong test for a leadership
effect since followers can free ride on the leaders contribution without fearing any influence on future
contribution behaviour of either the leader or any other member in their group.

2.4.1 Results Figure 1a shows the average contribution of followers as a function of their leaders contribution (i.e.,
are the folloeach dot reprwers esecontnts a grouributip). The moons. An interesting ore the leadebsers rvination is this one-sh that the followersot experimen, in sot contribute, the me groups, higher
contribute on average more than the leader (i.e., their mean contribution lies above the diagonal). Yet,
this holds only for contributions ≤ 10. For contributions above ten, followers in all groups contribute
less than the leader. On average, followers contribute more for leader contributions above ten than
regbelorewssi ten. Overall, the coon of mean follower contrrelation ributioof leaders' ann on the resped followecrs tive leader cocontributionsntribution yiel is pods a higsitive. A simple linear hly
significantly positive coefficient (t = 3.76). The more conservative Spearman rank correlation
coefficient also yields ρ = 0.56, which is significantly positive at p = 0.01.

4one-sh We delibot decieratsely had relatiion very seriously. vely high stakes in the one-shot game to ensure that our subjects take their


The obmany subjseects are rvation from Figuwilling to coore 1a is consperate if others istent wicoth reopceerate as nt findings in publiwell (see, e.g., Croc goods son 1experim998; Keents tser ahat nd
van Winden 2000; Clark and Sefton 2001; Fischbacher, Gächter, and Fehr 2001; Fischbacher and
). 03Gächter 20Does this leader effect induce leaders to raise their level of cooperation and thus, can leading by
example lead to higher overall cooperation rates? Figure 1b gives the answer by showing the
cumuland leaaders itive con the teams ntribution functionwith a leads of subjer.ects in t On averaghe e, cosubjectntrol treatmes in the teams nt without a leadewithout a leader and of r contributefollowers d
8.5 tokens; the followers in the teams with a leader invested 9.2 tokens and the leaders put 12.5
tokens into the project. The cumulative distribution function of leaders contributions lies below the
cumulative distributions of both the control experiment and the follower contributions, which indicates
that leaders contribute more than followers and control subjects. A nonparametric Wilcoxon rank-sum
contritest shobutionws of subjes that the contribcts in uthe conttions of subjrol groupects in ts (p = he le0.0ader 28, two-siderole are d)signifi. As can bcantly heigher seen from Figcomupared to the re 2,
the modal choice of the leaders was to contribute the maximum possible amount of 20 tokens, which
wachois chces in theosen by control treat 30 percent of the leader ment. Only 10 percent ofsubjects, com the leadpareders co to only 12.5 percentributed zeront of full contrib to the team projeution ct;
whereas the followers and the control subjects did so in 19 and 23 percent of the cases, respectively.
On average, followers contributed only slightly more to the team project than the benchmark subjects.
Morepercent. Thisover, the followe differencrs ae is vsignifierage contricant (Wilcoxon mabution fell shtcorhed-pt of the leaders airs test, p = 0.01, twocontributi-sided). The aon by an average of 24 verage
contribution only slightly but insignificantly (p = 0.401, Wilcoxon rank-sum test, two sided) increased
tendenfrom 8.5 tokecy to higher ns in the contcontriburol grouptions but no s to 10.0 tokesignificant effens in the teams ct of leadershwith a leadip on the over. Henerall cooce we find a peration rates.
We summarize these results from our benchmark experiment as
followeObservation rs and1: Followe leaders are significars follow theintly por leadersitivelsy exam correlated. Leadple in the one-shot teamers contribute mo game: Contribre than usubjtionects in s of
control groups and also more than their followers. Followers contribute insignificantly more than
control subjects. Overall cooperation rates in one-shot interactions are slightly but insignificantly higher
in the presence than in the absence of a leader.
The fact that leaders contributed more than control subjects suggests that leaders even in the one-
Hoshot experimw profitable waent correctly anticipats it for leaders to chooed that their contse their rerispecbution will infltive contribuuention? Did it pace their followers conty to be a bold leader ribution.
who contributes the whole amount or would timidness and low contributions have paid off? As Figure
1 shofollowewr cos there is indentribution. Yet, is the obed a positive correserveldation relation betweestronng eno the leadeugh to rsse contribution cure profits above 20 (which iand the average s
the minimum payoff from full free riding)? From the payoff function we can calculate the necessary
strencontribution tgth of followeors contri the team project. Assumbution behavie that tour such that a leadehe estimated linears mer relation betwean payoff increases in hien a leaderss and his
or her average followers contribution is gf=c+βgl. Plugging this relation into the payoff function
(1) gives us the following relation between the leaders payoff, πl, and his or her own contribution gl:
πl = 20 + 3αc + (3αβ + α  1)gl (2)
From (2) we can easily calculate the critical c* and β*, such that (i) the leaders mean payoff is at least
20 (which he would get from full free riding) and that (ii) the leaders payoff increases in gl. Setting
dπl/dgl = 0 and solving for β gives

β*=13−αα (3)
With averagαe to contribute hi = 0.4 is straightforward to see thas or her endowmt c*ent; if β = 0 and < β*β* he or she ea = 0.5; i.e., if βrns mo > β*re if by contri it pays for the leadebuting nor on thing.5

5 Of course, if α < 0, the critical β* has to increase to secure a mean payoff > 20 for all leader

As Figure 1 shows, c = 4.81 and β = 0.35. From this it follows that the average leader would have
ro tokens to the team he or she had contributed zeoff, if s material payfared better, in terms of hit. cprojedireFrom the viectly benefwpoiits from the lent of a followeadersr, there contribuare two tion glsou. Secorces of sucndly, if another follower folcessful leading lows the leadby example. First, he ers
followeexample anrs bened increasefit from an incres his contriabution sed contribution of (on average aothcecording to gr followers. Both argumf = c + βgl witenh βts are > 0), thenstraig all htforward
from the followers payoff function (adapted from (1) and setting α = 0.4):
πf = (20  0.6gf) + 0.4gl + 2×0.4(c + βgl). (4)
if β The follower> 0; the mean mas mean parginal followeyoff net of the ler payoadersff in an increa contribuse in gtion (which is l. is 0.8πβf  0.4 (i.e., in our cgl) isa simplyse (s increasinee Figure 1) g in β
this is 0.8×0.35 = 0.28). Of course, (4) says just that the mean follower payoff will be increasing in gl,
and nothithe followeng ars averagbout signe payoificance. Thff net of the lee Speaaderrmans ran contk ribcorrelation bution is 0.56, etweewhicn the leah is highlders coy significantributnt (p = ion and
). 0.0111contriObservation bution increa2: Timised sigd leaders enificantlayrned mo in the leadere than brold s contribution. leaders. Followers payoff net of the leader
In summabetter than bry, leadinold and ggenerous leade by example works even in a oners in terms of the-shot context. payoff they can se Yet, timid ancure for d greethemselvedy leads. ers fare
How does leading by example work in case of repeated interactions, which characterize many
leadeincentives? rship problemWe will invests in reaility? What is the dynamgate these issues in the foics llowing of leadershisection. Given that p in the presence of free ri(i) there is a positive der
leadecorrelrs ation becontributed motween the core than control ntributionssubje of leadecrs ts, whiach suggnd followers evenests that they were in a one-shotawar game, and (iie of their impact )
on followers contribution behaviour, there should be the possibility that the leaders, by setting the right
example, caIn the repeatned game lead their gs thirso shups toould reallward myo be feasible,re efficien sit contributionce we hans tve seenhan in simultaneo that already in the ous choice setupne-shot s.
game contributions are (albeit insignificantly) higher in the presence than in the absence of leaders.
Yet, we have also seen that this will be no easy task, since in the one-shot experiment timid leaders,
who diEven if the leaders ad not set a good exare willimple and ng to set a bocontld examplributed little or nothine, the benchmga, fared betterk experimr thaent nhas geneshorouwsn th leaders. at they
are in jeopardy of being exploited by their followers, if the followers only half-heartedly cooperate.
cooThus, the realperators is an empi dynamics of leaderical questiorsn, whhip in the preich we try seto enlightence of free riden in the remr inceantives aninded conr of this paper.ditional

2.5 The dynamics of leadership and voluntary cooperation
We will first look at hoIn the discussion of our results from the w cooperation evolexperimved over ents intime in the pre treatmentsence as 1  4, we will nd absence of a leaproceed as follows. der in
treatments 1 and 2. Second, we will investigate whether there are different types of leaders and
whethtrade by looekir the type of leader ng at treatmemakents 3 and 4. Fs a differencourth, we e. Third,will che we anck whalyze thethee impact of inr we find evidcreaence for leadised gains from ng by
example anfinal step we will take ud how the resulp atgain the qus from the onestion whethe-shot exeper iritment ca pays more trry ovo be a bold oer to the repreated game a timid leadesr. . As a

ader matter? Does the le2.5.1 We start with the evolution of cooperation in the presence and absence of a leader in treatment 1,
whehighere the lear than inder is their absenrandomce. Figure ly selecte2 dsh. We expect ows the average evolthat contribuutiotionns in the of contributiopresence of a leans in groups der awherre e in
the leader was randomly determined. For the sake of comparison, we also include the cooperation
levels of the control groups (from treatments 2a and 2b) where there was not yet a leader in the first

contributions gl > 0.


table 2.1). ece (seuenseqLeaders contributed more than subjects in the control groups and more than their followers. On
tokenaverags. A Wile a leacder contrioxon rankbute-sum test and ad 10.8 toke Wilcoxons. Contronl masubjtchedects -paicontributrs test rejeed only 6.68 cts the and follonull hypothesiwers 6.54 s of
equal contribution of leaders compared to subjects in the control groups and to followers respectively
(significant at p < 0.01, one-sided).

201816141210oitubirtne coagverAsn4



control groups without leader
Figure 2. Average contribution over time of leaders, followers, and control groups without a leader
This is veOverall, and ry remarkacontrary to ouble, gir expectatioven the correlnsati, the preon of leadesenr and folloce of a leadweer seemr contribus not totion that we have have mattered much.
observed in the one-shot experiment. The average cooperation rate in the experiments with a leader
was not significantly higher than in the control experiment (Mann-Whitney test, two-sided, p = 0.80). In
a leadethe contr corol experimntributionsents, su amobjectsunted to 7.61 contributokensted 6.67 tok. We summariens to the teaze our resultm projes acs t; in the experiments with
subjectObservation s from3: Rand the control teamomly determined les without a leadeaders contrir. Yebute mt, the oveorall coopere than followeration rs arand also mte is not signifoire than the cantly
higher in the presence than in the absence of a leader.
presenObservation ce, efficien3 seems to hcy of team prodave a sobuctioer inmplicatio is not incrn:ea leaders do nosed relative to a team situatit matter in the senson whe that, inere they their are
seleabsent. Hocted acwecording to somever, one migh pre-definedt object that in reality lead criteria. Yet, iners a contrare usuallst to many real life siy not randomly selectetuations, leadd, but pre-ers in
our setup neequally skilleed no specid ex ante. This justifieal manageris our al leaderandom rship skilleadelsr sele. In our expection. Neverthelerimesnt, every subjes, the equcal skill t seems to be
argusuited thament mignh others to be thet not be true beca leader. Thuse some leadee reason for thrs mais conjecy in fact even in this simplture is that evidence sugeg situation bests theat better
people differ in their cooperative inclinations. Some people are free riders and others are conditional
cooquarteperators. For inr to a third of the sustanbjectce, Fischbas cachn beer claset al. (2001sified as free rid) and Fischebacrs, and rouher and Gäghly 50chter (20 percent are 03) find that a
conditional cooperators.6 These types of subjects may differ in their leadership qualities.

6 See Ockenfels (1999); Gächter, Herrmann and Thöni (2003); Falk and Fischbacher (2002), and

followers. ThHere is why. As the above e followers contributioresults shown in all perio, randomds fellly determin short of the led leadeadeers rs contcontributerid more thabution. In other won rds,
the leaders got suckered in each period. Different types of players may be differently suited to stand
simultaneouthis situation. Intuitively, ons team gamese (the pro-socials), migmight assume that those subjects ht be the better leaders, becauwho are willing to cooperate in the se they are maybe
more willing than others to sacrifice payoffs for the sake of the group, or because they feel less
suckerewhen others d than others. Anocooperate and who arther reason could be te therefore willing hat they are condito set the extional coopeample of being a cooperator. rators, who cooperate
Apparently, setting an example works on average, as our results so far have shown. By contrast, free
ridethat more thar types min 20 perceght not be willnt of our followeing to set an rs anexample and control subjed to contribcuts in the one-shte to the team project. Remot experiments ember
contributed exactly zero to the team project. Maybe, if they were the leaders, low contributions would
emerge. One may argue that making a free-rider type the leader of the group is like setting the fox to
keep the geese. When the leader is randomly selected, it is inevitable that we sometimes select a
aggregatiofox, and somn eetimes a cooffect that hides the true diffperative leaderener. Our ceresus in relelt about ineffevant leadective leaders mrship skills. Our following ight then be due to an
experiments test these arguments.

Are there good and bad leaders? 2.5.2 To test the argument about leader quality, we look at the experiments of treatments 2a and 2b in
which, in the first ten periods, subjects played a simultaneous team game with no leader (see table
2.1). We then ranked the subjects in each group according to their average contribution to the team
project of their group over all ten rounds. We call a subject with the highest contribution over all ten
periods in his or her group the pro-social type. Subjects with the lowest contribution are called free
rider types.7 In treatment 2a we made the pro-social the leader in the second sequence; in treatment
2b the freof the leader. e rider beAll subjeccame the leadets were only infor. Subjecrmed that tts were hey would lenot awararn e of this rawhethenking ar they are the leand the determination der or not.
This avoids any influence of the leader determination rule on the followers behaviour, and allows us to
)famoudetermis pane the mere least. For a similader r reason, subjeeffect that is not confcts did nooundt know abed with the inout the leadeformation ar expebouriments in the t the leadersse (incon-d
sequence, which avoids confounding different contribution types and their anticipation of possible
consequences of the leader or follower role.
we shoFigure 3 depiw the averagcts the evolutie coon of contribution pathoperas tof our two ion in the first and sedifferent future leadecond sequenr types. We alce. In the first seso showqu the ence
average contribution path of our future followers. The differences in average contribution levels
between pro-socials and free riders are remarkable. In the initial period, average contributions differ by
almost ten points. Pro-socials always contribute more than free riders, and even strongly increase their
contributions in the final period. The difference in contribution patterns suggests that our classification
. ensiblof types is se

Kurzban and Houser (2001) for related results.
7 This procedure implies that the classification of types is relative to the average contribution of a
particular group. This is justified because group composition stays the same over all 20 periods, i.e.,
r. out a leadewithwith and 17

20 periods without leader periods with leader
1816141210snoitubirtage conrevA4
free rider (leaders)pro-social (leaders)followers
socials) anFigure 3. Average d followers in pecontributriodions in tres 1-10 (batmefore they ents 2a and 2are leaders ab of different leadend followers, rer types (fsp.) anree rid peders anriod 11-2d pro0 -
(when they are leaders and followers, resp.).
The sein periocods 11nd seque-20. We shonce is thw the average one we are e contrimainly interebution of the prosted in he-socire baleca and free riuse there deis nor leadew a tear, m leader
8leaderespecrs conttively, riand the averabute more than followers, whige follower contributioch is conn.s A first main obistent with our previoservatious findings. n is that both types of
But who is thto keep the geese? Mue better leadch erto our surp? Is it true that making a rise, free riders formebehaverd free rider the lead more like wolves in sheeer is like sps etting clothing. the fox
Once they became leaders, they immediately and significantly changed their contribution behaviour:
perioon average thd 11. When we comey contributepd 3.7 tokenare this to thes in first perio the firsd,t sequen then the immece and jumped updiate increawardse in contri to 9.9 tokens in butions is
test). Thueven more res, this chmarkablange is e. The chamore thange ins hi just a restghly significaart effect. On averant (p = 0.0079, two-ge, free ridsided er leadeWilcoxon sirs contgrined rabuted nks
seq7.87 tokeuence! Thns in the see null hypothcond seqesiuens that tce, which ihe averagse co an increase of almontribution of the free-rist 114 peder learcent compaders is the red tosa the first me in the
first and the second sequence can be rejected at p = 0.01 (Wilcoxon matched-pairs test).
chaAre pronge thei-socialr behavios better leaur suders? Figure 3bstantially. They only sligh suggests nottly. When pro-social but insignificantly increas become leadese their contrs, theriybution do not s
from an average of 8.7 in the first to 9.3 tokens in the second sequence (Wilcoxon signed ranks test, p
Whe= 0.441n we ). For copro-mpare perisocials odwe can 1 and perinot od excl11 coude that their ntributioincreans, the chseang in contributioe is pons issitive, but insig just a restanificarnt t effect.
(Wilcoxon signed rank test, two-sided, p = 0.127). We keep these findings as
Observation 4: Free riders strongly increase their contribution rate when they become leaders,
whereas the pro-socials only slightly increase their contributions, once they are leaders.
A compamore (on averisonrage 1.4 of pro-so3cial leade tokens), but this rs and free-ridedifferencr e islead ers not signifishowcants that the fo (Mann-Whitnrmer contriey test, two-sibute slightly ded, p =
0.453). In other words, once a leader, the (in-)famous past does not matter for the leaders behaviour.
The average contribution behaviour of pro-social and free rider leaders is also not significantly
8 We pool the follower contributions, because they are basically identical.


different from the contribution of randomly selected leaders (Kruskal-Wallis test, p = 0.236). Thus, we
have no evidence for the existence of good and bad leaders, when we define good and bad by the
aviour. tion behupast contribAs with randomly determined leaders, cooperation rates are not significantly higher than in the control
experiments without a leader. This holds for both leader types.9 All types of leaders are apparently
rather ineffective in stimulating higher levels of voluntary cooperation than is achieved in the absence
of a leader. We summarize our results in the following
Observation 5: There are no evident differences between leader types: former free riders show similar
contribution rates as former pro-socials, or randomly selected leaders. Neither type of leader is able to
induce significantly higher cooperation levels than in simultaneous team games.

gains from cooperationThe impact of increased2.5.3 How robinvestigate thiust ars quee the restion sultwes of Obse look atrvatio the rens 3  5 with sults from trerespecatments 3 ant to gains from d 4,coo which were exacperation? To tly identical to
treatmenttreatments. The social gs 1 and 2 except that the margiains from coopnal per eration are thecapitar return efore inαcrea is 0.6 insteased from 1.6 to 2.4. Yet, to d of 0.4 in the former free ride
behaviois still a domiur between treatmenant strategy, which implients 1 and 2 as that nd 3 and 4, respthere should necotively. Hot be any differenwever, it is knoce in coown from peration previous
experiments that the higher α is the higher are contributions to the public good (see, e.g., Ledyard
que1995). Thusstion therefore is h the coopeow irationncrea problsed gem is ains from eased wcoohen tperation he gainscha fromnge the re cooperatiosults frn increaom Obsese. The rvations 3 
5. Three arguments suggest that there will be a significant leader effect on cooperation rates. First, if we
can replicate the positive influence of α on contributions, it should be easier for followers to contribute
sinarguce the loment holds of coss they incuurse r from contribalso for the leuader ting, 1  who, αin , is lower if addition, shoαu = 0.6 than if αld feel less suckered if followe = 0.4. Second, the same rs
have a higher willingness to contribute. Third, if α = 0.6 the critical β (see (3)) is β* = 0.22 compared to
β* = 0.5 if α = 0.4. Put differently, since β measures the followers marginal increase to a contribution
contriby the leader,bute his or her whol already a rather small e endowmeβ nt. Since we suffices to mahavke it woe observed thrthwhile on averaat in the one-shot bege for the leadenchmr to ark
experiments β = 0.35 (see Figure 1), the threshold of β* = 0.22 should be easily met.
We report the results of the experiments of treatments 3 and 4 in Figures 4 and 5, respectively.
from the αSeveral interesting o = 0.4 experimbservenations ts cacan brry oveer to the made from Fαi = 0.6 experimgures 4 anents. As in treatd 5. Qualitatively, the main findingment 1 the randomly s
determined leaders of treatment 3 (Figure 4) contribute more on average than followers and subjects
ents. in the control experim

99 We look at the group averages in sequence 1 and compare the group averages of groups with a
pro-social leader and groups with a free-rider leader, respectively. The Wilcoxon-matched pairs test
returns p = 0.549 for groups with a pro-social leader and p = 0.515 for groups with a free-rider leader.

201816141210iront cgearevAonsibut2


control groups without leader


Figure 4. Average contribution over time of leaders, followers, and control groups without a leader

20 periods without leader periods with leader
18161412 cgearevAonsibutiront4
free rider (leaders)pro-social (leaders)followers

socials) anFigure 5. Average d followers in pecontributriodions in tres 1-10 (batmefore they ents 4a and 4are leaders ab of different leadend followers, rer types (fsp.) anree rid peders anriod 11-2d pro0 -
(when they are leaders and followers, resp.) when α = 0.6.
Hoexperimwever, theents difference in contriof treatment 3. Statistibutionscal tests of leaderreveal ths and followeat leaders in trrs is cleaeatmerly diminint 3 coshentribute sid in the gnificantly
standmore thas in conn followetrs rast to the re(p = 0.008) bsults from treut only insigatment nificantly1 where lead more than ers cocontribute sintrol subjegnifcits (p = 0.40cantly more than 0). This

followers and control subjects.
riders In treatments show4a and 4 strongly differeb (Figure 5nt coop)eratio, we fin patternnd that, as in treatments in sequence 1. s 2aAs in treatme and 2b, pro-sonts 2a anciald 2bs and free we find
whereathat, once thes the freey-rid are leadeers significantrs, the pro-socially increase their cos do not chantributionnge their s cont(p = 0.042ributions). Lea signiders coficantly (p = 0ntribute .225),
insignificantly more than followers, but significantly more than control groups.
contriA compabutionrisons. of Figure The increassed gain 4 and 5 with Fis from gurecoos 3 anperation if d 4 αsuggests tha = 0.6 exert a strontg there is a strong level effe influence on the ovct in erall
level of cooperation. Mann-Whitney U-tests that compare cooperation levels find that in groups with
and withexperimout aents (all p-value leader, avesrage co < 0.05). operation is significantly higher in α = 0.6 experiments than in α = 0.4
gainsIn summa from cory, consisoperation intent withcrease. First, a our conjectusre we hs, we avefind that the leade just seen, the overall cooprship problemerat is reduion rate is ced if the
significantly higher when α = 0.6 than when α = 0.4, irrespective of whether there is a leader or not.
Second, the leader is indeed less of a sucker under high gains from cooperation. In the α = 0.4
experiment the followers contribution falls short of the leaders contribution by 3.74 and 3.21 tokens,
respecfollowetively. rs averBy cage ontrascontributt, in the ion is reαduce = 0.6 experimd to ents t1.67 in sequehe differennce 1 ance betwed 1.97 in seen the lqueneadece 2. rs and the
Figure 6 looks at the contribution of followers relative to the leaders as a function of the leaders
contribution. In this figure we distinguish between the α = 0.4 and the α = 0.6 experiments. Figure 6
reveals an int0.4) followers contribereute absting difference betweeout 30 percennt less t treatmenthan les. Wheaders. Thin the gais rens from lation holdcoosp for all coeration arntribe loutw (ion α =
0.6), followelevels of leaders, exrs often contricept if the leadebute more than leadr contributesers, onlyif the leader ma 1 token. If gains from kes a low (,eration a smaller thare hin 11) gh (α =
contribution to the team project. For leader contributions above 11, followers contribute less than
leadeare the suckerrs, but still more than for all contriin the bution levelαs, whe = 0.4 experimereas with nts.high gai Thus, with lons fromw gain coopseratio from coon they are onlperation ly eaders
suckered if they make a high contribution. = .6
o teivtlae rsrewlloo ff oniotubirtnoC
0.8MPCR = .4
noitubirtnos c'ereadl0
Leader's contribution
Figure 6. Followers relative contribution as a function of leaders contribution
(for leader contributions > 0 and α = 0.4 and α = 0.6, respectively).
However, despite the increase in the cooperation levels in the experiments with higher gains from
cooperation and despite the reduced sucker problem, we still get no effect of leaders on contribution
levels. This holds for both 21

= .6RCPMPM = .4RC

gainsObservation from co6: Qualitativeloperation. Quy, all reantitativelsults fry, the oveom Obserrvall cooatperatioions 3 to 5 can level is strorry over to tengly inamcre gamased anes with high d the
leader is less systematically suckered than in team games with low gains from cooperation.
Maybe the reason for Observation 6 is that the positive correlation of leader and follower contributions
an impaobserved in tct of the prehe onesen-shot gace of a leader omes dind not survive t overall coophe repeateeration leveld gams? We investige, which is why we do not ate this question in observe
tion. ecbsuthe following s

Leading by example? 2.5.4 Figure 7 shows that leading by example  in the sense of positively correlated leader and follower
contricontributionbution as s is well a a functiond n ofalive in all rep the mean leader eatconted team gributames. Thion. Each date figure plota point repress the mean folloents one groweur p and
we do not distinguish between different leader types (i.e., we pool the leader sequences from
1, period 1treatment 1 a0, and pond 2). We sholed ovow the relatioer all ten periods, ann betwed seen avpaeragrately for e leader α = 0.4 (Figand follower ure contri7a) and butionα = 0s .6 for period
). e 7br(Figu

Figure 7. Mean follower contribution as a function of mean leader contribution: (a) α = 0.4 and (b) α =
0.6. perioThe positive correld. The Spearman ranation bek cotween the lerrelation acoefficiders and the ent befollowetween the lers acontders coribution holntribudts alreions anady in the first d the average
0.59 (p = 0.0followers contribution 007). Correlatis ρ ions a= 0.54, whire pocsithi is signifive and sigcant nif(p=0.002icant at lea1). Lsiket at p = 0.002 iwise, in penriod 10, we g all other peet riods aρ = s


well. Pooled over all ten periods we get ρ = 0.79, which is significant at any level.10
Moreover, we find positive correlation coefficients not only in the aggregate data but also for the single
independent groups. Tables 1.a and 1.b in Appendix 2.B report these coefficients. Since there is no
negative coefficient, the binominal test highly significantly rejects the null hypothesis that positive and
negative correlation coefficients are equally likely. Thus, we conclude:
Observation 7: The ineffectiveness of leaders to affect overall contribution levels is clearly not due to a
genecontriral labutionck s. Obseof following thrvation 1, me leaderade in ths exame one-shple, in the ot experimsense of haents, cavirrieng sigs ovnificaer to tntly pohe repsitiveleated y correlated
interactions. How cahints. As Figun we squres 1 to 7 are Oshowbservation 6 and from different Observatiangleson 7? The follo, it is a tough life being a leawers behaviourder. Although, on might give some
half-heaveragartedly, in particulae, the followers will follow the ler in team gameadsers exa with lomw ple gaiand inns from creacoopese their coration. Giventribun this situatiotion, they do so only n, is it
better to be bold or timid?

bold or timid leader? Shall I be a 2.5.5 In Figure 8 we look at this issue. In Figure 8a we show the leader and the follower payoffs as a
function of the leaders mean contribution over all periods to the team project in team games with α =
0.4. Figure 8b does the same for games with α = 0.6. The black circles indicate the payoffs of each
construindividual leaction, der, and the tthe triangle thriangles at represents represa particulaent the mear ngroup lie payoff of a s eithergrou exactly abop of three followeve or belors. Bwy the circle
that represents the leader of this particular group.

with low and Figure 8. Mean follower ahigh gains from coopnd mean leeratioadn. er payoff as a function of mean leader contribution in games
the followeSeveral interesting ors bmean payoff stronservations can bgly increease made from Fs in the leiadergure 8. s contribuWe start by discussing Figtion. This is iunevitable, for two re 8a. First,
reasonpayoff directls, as ywe have increaseshos inwn in Sectio the leaderns contri III (see ebutiquaton toion (4 the team project. Moreov)). By virtue of a public gooer, as longd, a followe as the rs
senfollowesrs will also initivity of followers readirecctly benefit from a triggetion to an increase inred in the lecreaderasse in oth contriebur followertion is non-ns contributioegative, the n to the

10 For these calculations we have pooled the data from the two sequences. When we do the
calculations separately for the two sequences, we get ρ = 0.79 if the leader sequence is played first
and ρ = 0.72 if the leader sequence comes second. Both are highly significant (p<0.0001).


team project.11 Second, the followers payoff is in all 30 groups, except one, strictly higher than the
leaders payoff. This just reflects again that the leader is the sucker in this game. Third, and most
interesting for our main question at hand, on average it pays to be bold and contribute high amounts.
his average payoff. Remember that this can is erage the higher The more a leader contributes on avmeanonly happs that en iβf > 0.5. Over al the followers sufficil periodsently stron, the actual gly foβ = 0.7llow the lead882 (see Fiersgu example, which in ore 7), i.e., on averagur coe this ntext
sugaveraggese cots thatntribution an it pays to be a bold lead his realized The Spearmpayoff is indeean rand sigk ornder correlatioificantly positive (n beρtwee = 0.36n the le8, p = aders
0.0452, two-sided). A linear regression (see Figure 7) confirms this result. Fourth, the variance in
average payopayoffs is very large for leaffs that were loweders and ver than 20, which is try small for fohe llseowers. For a third of the leacure payoff from full free riding. Loders this resuok at the lted in
unluprojeccky guy it, but stn the loweill ended up r riwith an averagght part of Figure 8ae payoff of 18.. He contribute3, whered on as hiaverags followers ee 17.3 atokenrnesd a mea to the team n payoff
of 28.9. Goodslightly more than his u and bad lunlucky colleack can be very gueclos (18.8 on ave. The toeragep earni) anng lead tookder in this ex home a meperiment coan payoff of almost 29. ntributed
Figure 8b depicts the results from the experiments with high gains from cooperation. We conclude this
cording ction by resepayoff increObservation ases8: On ave in his rconagtribution bee, it pays to bcae a high cuse the followers inontributing leadcrease ter in this expheir contriebutirimons sufficieent. The leaderntly s
Leadstrongly to mers therefore faake leading bce a trade oy example woff between the rirthwhile on ask avnd retuerage. Yet, the varn of leadingrian by exce iamnple. payoffs is very large.
Can economiwe draw cs, it relates tosome advice for leade ones risk and retrs from Oburn servatipreferences. Given that on 8? If there is one, then, as so often inyou will be suckered any way,
ld you really want to be. o your mind how bpmake u

Concluding remarks 2.6 have shIt is a tough life being a leaown, is not in fact thatder in the p everyone resenis a free rice of free der.ride That would r incentives. Ironimake the procally, the reablem eason, asy. The s we
difficulty arises exactly because, on average, people are following the leaders example.
Interesting questiontemplate. First, what is the role of beliefs for further researchs about emergother foe for which ollowerur sis compntribution ale experimsent a function can serve a of the s a
leadea contrirs butocontrsribution? belief aboutWe know fro the average m prevcontriious expebution riofm others aents that thnd hiere is a ps or heor ositiwn contrive correlatiobution tn betweeo the n
public good. An interesting issue therefore is, how a leader shapes these beliefs about other followers
contrileaders bution bcannehaviouot stand this rdisa. Second, an Thereforent problemre, is what would th that the leader gete optimal cos sumpenckered asation fnd moor the st
leader be to solve the problem, if it can be solved at all? Paying the leader better than the followers
may not solve the problem because it may give the followers a further excuse to rip off the leader.
Fourth, hoThird, would tw dhe prooes the pbrlem bobleem change solved if the leadewith team sizer would have a sa? And so on. nctioningAll these prob device at his dispolems arse interesting al?
because many people dont free ride, even if they easily could.

11we find that the indirect be If we just look at the indirect effect, i.e., nefit from an increaif we sed subtconractt the direct paribution of the other folloyoff from the leadewers, asrs a funcontcrition of bution,
the leaders contribution, is significantly positive (ρ = 0.792, p < 0.0001).

2.7 Referencesstudy on reAbbink, Klauciprocs, Bernd Irlenbity and retributiusch anon, Jod Elke Renurnal of Econoner (2000): The mmic Behavioor on& Organilighting gazatiomen  An experi, Vol. 42, 26527mental 7.
Andreoni, James (1988): Why free ride? Strategies and learning in public goods experiments, Journal
04. 913nomics, 37, 2of Public EcoAndreWisconsioni, Jan-Madison. mes (2001): Leadership Giving in Charitable Fund-Raising, Mimeo, University of
RaiAndresing, Mimoni, James (200eo, University3): Public Go of Wisconodsin-Madis Experiments son. Without Confidentiality: A Glimpse into Fund-
EconoArce, M. Damic Paniel, G. (2001pers, Vol. 53, 114-1): Lead37. ership and the aggregation of international collective action, Oxford
and EcoBerg, Joycnomice, John Behavio Dickhaur, 10, 122-42. t and Kevin McCabe (1995): Trust, Reciprocity and Social History,. Games
ReClark, Ken, aciprocationnd Martin Sef,. Economic tJournon (200a1): Tl Vol. 111, 51-6he Seque8. ntial Prisoners Dilemma: Evidence on
3(1Coo)k, 55-79. son, R. (2000): Framing Effects in Public Goods Experiments. Experimental Economics, Vol.
Croson, Rachel T.A. (1996): Partners and strangers revisited. Economics Letters 53, 25-32.
Crogamesosn, Wo, Rachrking Papeel T.A. (1998): Theorier 98-11-04, The Wharton Ss of altruism chan reoociprocity: evidence from linl of the University of Pennsylvania. ear public goods
Falk, Armin and Urs Fischbacher (2002): Crime in the lab  Detecting social interaction, European
-869. view, 46, 859Remic EconoZüriFischbach, Institucher, tUrs (1999e for Empirical Re): ztree. Züsearrich Tch in Econooolbox fomics,r Readymade E Working Pacpoer No. 21. nomic Experiments, University of
FischbaEvidence frocher, m a public gUrs, Simon Goäods expchteerimenr and Ernst Feht, Economir (2c001): Are peos Letters, Vol. 71, 397-4ple conditionall04 y cooperative?
Fischbacher, Urs, and Simon Gächter (2003): Heterogeneous motivations in voluntary contributions to
Mimeo. s, goodcpubliGächter, SimSocial Capital:on, Benedi Trust, Coopekt Herrmaration, annn, and Chrid Informal Sastian Tnchöni (2tions in a Cro003): Undss-Soerstandincietal Perspeg Determinanctive, Mimeo. ts of
Foss, Nicholas (2001): Leadership, Beliefs and Coordination: An Explorative Discussion, Industrial
7-388. l. 10(2), 35nge, Voorate Chaand CorpHeijdexperiment, SNF Reen, Eline van der anport 1d3/00, Berge Erling Moxnens (20, Norway00): Fou: Leadership andation for Resend followingarch in Economi in a publiccs a in a pund blic bad
ministration. ss AdBusineHempin: R.M. Stogdill and A.E. hill, J.K., and Coons,Coon A.E. (1957s (Eds.):Lead): Develoer bepmhaventior: Its descri of the leader pbehaviotion and mer deasuscriptioremn questent, pp.6-38, ionnaire,
Columbus: Bureau of Business Research, Ohio State University.
Hermalin, BeAmerican Economicnjamin E. (19 Review98): To, Vol. 88 (5), 1188-1ward an Economi206.c Theory of Leadership: Leading by Example,
publiKesecr, Clau goods,di Scandina, and Fraavians n Jouvan Windrnal of en (2000Econo): Condmics Vol. 102 (1itional coope), 23-39. ration and voluntary contributions to
GoodKurzban, Robs Gamee, Europeart, Housern Journ, Daniael (20l of Person01): Individual Diality, Vol. 15, 37-5fferences in 2. Cooperation in a Circular Public
Ledyard, JohAlvin E. (Eds.), Handn (199boo5): Pubk lic Gooof Experimends: A Survey oftal Economic Experims. Princeton ental ReseUniversity Press, Princetoarch. In: Kagel, Johnn. & Roth,
Miller, Gary J. (1990): Managerial Dilemmas: Political Leadership in Hierarchies, in: Karen S. Cook


and Margret Levi (Eds): The Limits of Rationality, Chicago: The University of Chicago Press.
Moxnes, Erling and Eline van der Heijden (2000): The effect of leadership in a public bad experiment,
SNF Report 15/00, Bergen, Norway: Foundation for Research in Economics and Business
ration. tAdminisOckenfels, Axel (1999): Fairness, Reziprozität und Eigennutz  ökonomische Theorie und
experimentelle Evidenz. Die Einheit der Gesellschaftswissenschaften, Bd. 108, Mohr Siebeck,
en. TübingPotters, Jan, Martin Sefton, and Lise Vesterlund (2001): Why Announce Leadership Contributions? An
Experimental Study of the Signaling and Reciprocity Hypotheses,. CentER DP 2001-100, University of
Tilburg. Rhodes, Carl M. and Rick K. Wilson (1999): Leaders, Followers, and the Institutional Problem of Trust,
Mimeo. Weber, Roberto, Colin Camerer, Yuval Rottenstreich and Marc Knez (2001): The Illusion of
Leadership: Misattribution of Cause in Coordination Games, Organization Science Vol. 12, No. 5,
8. 582.59Yukl, Gary, and W. Nemeroff (1997): Identification and measurement of specific categories of
leadership behavior: A progress report, in: J.G. Hunt and L.L. Larson (Eds.): Crosscurrents in
leadership, Carbondale: Southern Illinois University Press, 164-200.


When in doubt...  3 Cooperation and punishment under incomplete information I say that individuals stheir levels of effort. The consequence ofhould be held account an action for a person should not depend able for their degrees of effort but not
on his circumstances, but it should depend on how hard he tries. er, 1998, p.18) (John E. Roem

3.1 Introduction Recent experimental studies have shown that peer monitoring is a particularly effective means of
Walkattenuating iner, 1992; Fehr and Gcentive problächteems anr, 2000 d alleand 20viating free riding be02; Sefton, Shupp ahaviond Walur (e.g., Ostkerrom, Gard, 2002; Masclet,ner and Noussair,
Villeval and Tucker 2003; Page and Putterman, 2000; Carpenter 2003). While in standard
experimental social dilemma games cooperation is substantial in early rounds but significantly erodes
at remaover time, in gamerkabs with aly high levels. In these dditional pustudinishmees memnt beroppos of peer grtunities cooproups caeration increan identify free ridses and iesrs an sustained their d
willingnnorm. Peeess tor punishment th punish them suus seemccessfully disciplins to be a majoesr ke those y to persiwho destent coopeviate from a groupration. s cooperative
Almost all previous studies investigate voluntary cooperation and punishment in environments in
which group members are homogeneous both with respect to their endowments and their preferences.
This homogsymmetry is a natural eneity is not a feature of way to begin many everyday sostudying coocial peratiosettings:n and mepunimbers of woshment behaviourk teams or r. Yet,
communities most likely differ with respect to wealth, ability, experience, strength, health, or a host of
other resource endowments that limit the extent to which they are capable of contributing to their
p. grouIn this paper we are interested in how cooperation and punishment are affected if team members differ
problwith respecem in which gt to their caproup mabemberilities (wes halth,ave different end abilitieowments, etc). We model ts. Spechis situifically, the bation aasisc de a publicisicon goproblods em
is as follows (for details see the next section): in each of the ten periods each subject receives an
endowment from a uniformly distributed set of potential endowments. Subjects simultaneously decide
how much of this endowment they to contribute to the public good.
crucIf group memial. For instanbers are heterce, it is often easy to oogeneous wibseth resprve ect whto their enat others are dodowming, butents informati not what they coulon may becomd be e
sickness or mdoing. It is, for example, obalingering. Pevious that a colleagople who live on soue cial is awelfbseare mighnt from work bt do so beucat this muse they need it or ight be due to a
because they want to exploit welfare. Low contributions to a work teams project can be undeserved or
due to low effort. In such situations it difficult to judge others behaviour: the mere observation of
others levels of cooperation allows no inference of their cooperativeness.
To capture these effects, we vary the information condition in two experimental treatments. In one
groutreatment, whp membich we call Pers. In PRIVATE, subjeUBLIC, everycts only kno group mwe mbtheierr own e is infondormedwment in a part about the enidowmcular group. ent of all other This
endosetup allwmenows ts influenus to ance cooswer peratioour first ren ratesearch qs, i.e., whatuestion: Ho is the differenw does thece betwe information aen PUBLbout the IC and PRIVATE?
Our seprevious expcond reseerimearch nts havinterest e shocowncernn that spuni the shmepowernt of puniin homogshmeeneount in this enviros groups cann stronment. Regcall that ly mitigate the
punifree ridshmeer problem inhent is used when rentgroups a in the public re heteroggoodenes gamous with respee. Therefore, we are intect to their endorewmentsted in (i) hs and ow (ii) how is
mplete information? oed by inctit affecor relHeterogeative coneous enntributiondoswme (i.ents ma., absolute ke the punicontributioshmens ant decsis a propion moortionre comple of the endowment) might tx: First, in PUBLIC absolute rigger


punishment. Second, in PRIVATE a potential punisher lacks the information about the actual
endowment of other group members; he or she sees only the absolute contributions of the other group
membetype I errorr: s. If he or she dethe punishercides n might believeot to punish a that the certsubjainect ha (low) cos a low ntriendobution he owmernt but actually ha she might commit a s a high
endoled to puniwmeshnt and neverthment. Of couelerss decided tose, a type II error is al contribute liso possiblttle  a behavie: if a subjour that undect decideres to puni PUBLIC sh woa certaiuld have n
contribution he or she might actually hurt someone who contributed a large fraction of his or her
t. nwmeendoOur maitake advan rentasults are the foge of this situation and llowing: Ficontrst, as soon ribute coansis thde endoerably lowmewenr proportiot is not publins ofcly their endo known, subjwments. ects
Seconis quite smd, the preall. Surprisence of psingly, low coopunishment operation rateportunities promotes s are not due to a reluctacoopnce to punieration. Yet, in PRIVATE this effect sh. Punishment
expenses are not significantly different in PRIVATE and PUBLIC. Third, punishment has negative
welfare implications in both treatments: even with a strong increase in cooperation rates in PUBLIC
teams are better off if they cannot punish each other.
with reWe beliespecve that our ret to their casultpabis are of practilities to cooperatcal and theorete. It is therefore impical releovanrtacen.t to understa In reality groups nd how certainly differ
heterogeneity and the information about it affect cooperation. Our results show that in PUBLIC people
largely base their punishment of other group members by how hard they try. Thus, in PUBLIC
peoples behaviour is consistent with John Roemers normative request in the opening quote. By
who actually contrast, in PRIVATE punicontributed thshmeeir whole ennt can ddo conowmesniderablet (which may happe harm because pun to be smallnishme). Therefont may hurt people re,
informatioit provides a n arobustnebout others ess chndoeck ofwment the pos is cruwer of pucial. Thinishs finding is alment to mitigate the free riso of theoreticalder pr relevanoblcem with e because
respect to heterogeneous endowments and incomplete information.
The rethe experimmaindeental pror of this chacedpter is oures. Section 3.rganized a3 pressent follows: in the next ses the results. Section 3.4 discuction we desses scribe the dwelfare esign and
s. udeSection 3.5 conclns. implicatio


Experimental design and procedures 3.2 Our experimental set up is closely related to the design of Fehr and Gächter (2002). The basic
decidecissioion situan in the setconion involves d staa simple linge. In the first stage eaear public gooch odfs game in the four team member first stage as receinvd a punies an endoshment wment of
tokens and can allocate these either to a team project or to oneself. The first- stage earnings of group
member i are

4πi=ei−gi+α∑gj (1)
j1=where ei denotes individual is endowment and gi is her contribution to the team project. The marginal
per capita return of a contribution to the team project is set at α = 0.4, which implies that it is socially
invested yieldoptimal if all group ms a net retuembernrs co of 4*0.4 ntribute t−heir entire endowm 1 = 0.6 to the group. In contrast to this soent to the team project: each tocially optimal solution ken
however, it is individually not rational to contribute since every token contributed yields a private net
return of 0.4 the first stage if subject− 1 = − 0.6. Tsherefo are rational and re the unique equilibrium consisonly care about their own monetary payoff. When ts of all group members free-riding in
punishment opportunities are available, the first stage is followed by a punishment stage in which each
subjon puniect can ashmessint reducegn up to tes the punin punishing shmemembent points to ears paych of the other group moff by one point and the pembunished meers. Each poimbers nt spent
equilibrium predictpayoff by three points. Since punis no punishmeshment in the sent is coconstly d stait is ge and not a crediblezero contributio threat and thns ine sub the first stagame pge. erfect
with puniAs in Fehrshm anent and then d Gächter (20anothe02), sor sequme of our subjeence of cts ften public girst played a ood gamseqes withuenout punice of ten publishmecnt good games
opportunities. For other subjects we reversed the order of the sequences to obtain control groups
.ncenishment in the first sequewithout pucontriIn the presbute by randent study we impomly determinilement unng endceowmertainty nts. about the capAt the beginning of stage onability of individual group meme each individubers to als
endoendowmewmennt wat is 20 tokens indepens for eadently drach indiviwn from a uniform didual. By this feature stribuwe motiondel gro on {0, 1, , up me40}. Thumbers with ths the expected e same
expected capabilities; however the actual capability of a given subject can vary across games.12
we condTo investigatuctee the impact d two treatmeof informationnts with different inform about endation conowments on ditionscont: the publiribution ancd puni informatioshmen treant behavioutment r
(PUBLIC) antheir individual endowmed the private informatnt before staiogen treatment (PRI one. After they have madVATE). In the PUBLIC coe their contribndution deition subjecisiconts learn s and
before they have to decide whether and how many punishment points they wish to assign to each of
many tokenthe other teas m membeout of this enrs tdowmhey also leent a pearn the rson contriendobuwments of every single grouted to the team projp membeect. An overview ofr and how our
design is presented in table 3.1.

12 We drew two sets of 40 numbers (set A and set B) which determined the endowments for each of
in the sethe four groucondp membe treatment wers for ten peri used the odsran. In 9 out of dom set A 18 grouin the first sequeps in the nce afirst and 10 ond B in the seucot of 19 of the groupnd sequence s
and revein the endorsed wment struthis ordctuer in tre betwehe remen thaining groue two treatmeps. By ntthis pros, whiccedh enure ableswe control fo us to dir randorectly compam differenre the ces
lts. ures. 29


Tabelle 3.1:. The treatments of the experiments
Information on individuals Sequence I Sequence II No. of independent
endowment (10 periods) (10 periods) groups
PRIVATE without puwith punishmnishment ent without puwith punishmnishment ent 10 9
PUBLIC with punishment without punishment 10
without punishment with punishment 8

The PRIVATE treatment differs only with respect to the information available in the punishment stage.
At the beginning of the punishment stage each group member is informed of the other subjects
endocontriwmebutionnts,s, but is not informe nor the information on thed of their endo endowmwmentents,s. Neithe affects the equilibrir the random deteum solrminatioution describn of the ed
requireabove: if it is s that in every pericommon knoowledd no subjege that subject punicsh tsan maxiothemise owr, and that aln earningl subjesc then suts free ride. bgame perfection
The experiments were conducted in the experimental laboratory at the University of Erfurt using the
particiexperimental pated, thus softwarwe collee z-tree (Urs Ficted a total 37 indepschbacher, 199ende9).nt observation A total of 14s.8 stude The subjectnts frosm were e variouncours disciplinaged to es
participate through posters on the campus and we made sure that each subject participated only once.
At the beginning of each session the participants were randomly assigned to one of the booths in the
decision anolaboratory. The boothnymously ans visually sepd independearatnte the particily. The groupps weants anre randd enomsure that everyly assigned, individual ma so that subjekecs his ts did
whinot knocwh the de whiccision h of the othesituationr parti of the first secipants quenwere in tce ofheir tengrou periodp. Writtes ann instrud the experictionmsen were htal proceduandedre out in s were
explained in detail (a translation of the original instructions is provided in Appendix 2.A). The
instru(and ectionndoswm were written in ents in the caa neutral lase of PUBLIC) wenguage. The re displaoryed after stagder in which the other ge one was raroundomly shp members duffled in ecisions
behavioeach periur in od, suprevious perich that it was not poods. We usedssible this to attri procedbute the current deure to exclude reputation effecision of a persocts. n to his or her
received the After period tcorreen suspbjecondingts were info instructionrmed s in whicthat another h the modified desequecince of 10 periosion sids tuation was will follow. Subjeexplainecd. ts then
Immediately after each session subjects were paid privately in cash. Average earnings were  13.20
(about US $13 at that time). An experimental session lasted approximately 75 minutes.

3.3 Results We start with an overview of contribution behaviour in our two information conditions before we
scrutinize punishment behaviour and the reactions to punishment.

behaviour 3.3.1 Contribution Figures 1a and 1b plot the evolution of contribution rates over time.13 Figure 1a shows the results from
sesessissionons withs with punishment punishment oppoopportunitiertunitiess in the fi in the serst secondquen sequence; figure 1b shce. Figures 1a aows the rend 1b clesults froarly indim cate
displathat the averayed to the other gge contributioroupn rate membes rare s. suThis differenbstantially locwer e is sigwhenifican thent regardle individual enss ofdowm whether pents aunrie not shment
oppo0.000 for rtunitieseqsuence are availas with pble unishment, uor not (signifiscant ating two- p < 0.sided Mann01 for sequen-Whitney U-teces withosts aut punishment anpplied to group d p =
13the cont Most of our dributiaon rate ata analysis gi/esi is ba. Figure 1 dised onsplays contributio the averagn rates. For each subje of these across all subjeect and pecriod ts withiwe comn the pute
ent. and treatmdriorelevant pe 30


es). averag

differenFigure 2, whice is ch deobserved apictcs the averagross the entiree contribution range of enratedowmen for different endots. For all endowmenwment intervals, t intervals thshowe s that this
contribution rates in PUBLIC are well above those in PRIVATE.

itubirtne coagverA ere pat rno0.1PRIVATE with punishment
0.3PUBLIC with punishment
0.2PUBLIC without punishment
PRIVATE without punishment
Endowment interval
Figure 2. Average contribution rates subject to the endowments
r first result: servations in oubrize these oWe summaResult 1: Contribution rates are lower when endowments are private information.
Figure 2 also suggests that in both information conditions subjects with higher endowments contribute
endoa lowewmer propont interval and thrtion of their endoe average cowment. Spearmntribution an ranrateks in co the grorrelation upcoeffici provide statients bsetweetical evidn the ence for this
decreaAppendix 2.Bsin)g trend at the ind. In PUBLIC we obtaiependent gron 17 negative anup level (the resultd 3 positive coes are repfficients anorted in detaild 16 neg in table 2.a in ative and 4
negative anpositive for the seqd 1 positive coeuences without anfficients for both seqd with punishmeuences. Usint respng theectivelsye coeffici. In PRIVATents as a summaE we obtain 1ry 8
equally likely statistic two-sided binfor all seqomialuence tests s (p reje< 0.03). We rect the null hypotcord theses that pohese findings as our sitive and nesecond result: gative coefficients are
Result 2: Subjects with higher endowments contribute at a lower rate.
evident from iWe next turn nspectioto the impact n of figuresof punishmen 1a andt 1b that t opportunities on coophe availability of sanctioeration behns has a stronaviour. It is immedig positive ately
impact on contribution rates in the PUBLIC condition: the availability of punishment opportunities
increaand by 17.95 se the averagpercentae coge pontribution ints (wheraten ps by 16.66 (whunishment is en puniavailable in thshment is e seconavailabled sequen in the first sece). A two-quensidceed )
puniWilcoxon matshment sequeched-pance is plirs teayed first andst, applied to indep at p = 0.028 endent grouif the punishmp averages, is sient seqgnificuenceant at p = 0.036 if the is played second. In
contribution PRIVATE however, the imrates increasepact of puni by only 7.84 percentashment opporge ptounities isints if the pu considenishmrably weaent seqkuener. Thce is ple average ayed first
sided Wiland by only 6cox.on match93 percentaed-pgeairs te points if thests, again applie punishment d toseq indepeuence indent sgrou played seconp averages, are ind. The assosigcinificant at ated two-
ectively. respd p = 0.169 np = 0.093 aFigures 1a and 1b also show that contribution rates in the PUBLIC treatment tend to increase over
time when pstatistical repurnishmesentation foent is por the ssibleevolution of contrib. In PRIVATE however, there iutions for thse indepe no apparendent gnroupt time trend. To s we calculate gain a
Spearmare repoan rarted innk correlatio detail in tabn coefficiele 2.b, Appennts betweedix 2.B). n averaIn PUBLIC 14 gge contributioron rates ups yield pand posieriodtive and 4 gros (the reups sults
and nenegative coefgative coefficificients in the ents arpunie equshmeally likentl sequey (binomial nce and test, signifiwe can rejcanectt the null hypo at p = 0.031). In PRIVATE the thesis that positive
coefficient is positive in 11 groups and negative in 8 group and we cannot reject the null hypothesis

that positive and negative coefficients are equally likely (binomial test, insignificant at p = 0.648).

behaviour 3.3.2 Punishment Looking again at figure 2, we see that the effect of punishment opportunities on cooperation rates is
stronger in PUBLIC at all endowment intervals. In PUBLIC it is further remarkable that the presence of
puniendoshmewments:nt opportu average nities econtxribution erts a paraterticularly sts increase byrong more thadisciplininn 20g effect on su.9 percentagbjeects with hi points in the gh
endowment interval (31;35) and by 25.4 percentage points in the (36;40) interval. In PRIVATE
however, for rates by only all endoabout 5 pewmercennts above 1tage poi0 tokennts. In PRIVATE s, the poonly subjssibility of being puects with endnisheowmd increaents belseow 1s coope0 tokenration s
exhibit a slightly higher increase. Our third result summarizes these findings:
Rerates and highsult 3: In PUBLIC the p cooperationrese rates anrce of punishme sustained. In ent opportunitiePRIVATE the pres signsificantenlyce of pu increases nishmcooent oppoperation rtunities
leads to a small, but insignificant, increase in contribution rates.
subjWhy is puects in PRIVATE simpnishment less effective whely refrain fromn endo punishment. wments arHoe prwever, theivate informatiore is little evidence in the dn? A first conjecature is that ta to
support this conjecture. Punishment is slightly less frequent in PRIVATE than in PUBLIC: in PRIVATE
36.3% and in PUBLIC 39.7% of all decisions are punished. Altogether, subjects spend slightly more
althougon punishmeh the difference is nt in PUBLIC (1.14 tonot signifikencants per pe (two-sidrioded ) thMann-Whitnean in PRIVATE (0.86 tokey U-test, p = 0.126). Thins per pesri is od), our fourth
result: RePUBLIC treatsult 4: Subjects mspeent, but the nd sldifferenightly lecss oe is n puninot signifishmcant. ent in the PRIVATE treatment compared to the
modeIf readinerate impss to puniact of punish is sishmilar ament on cocroopss the two treration rateeatments, then whs in PRIVATE? The reat else might ason is that subjeaccount for the cts who
a are puperson we calnished cdo not chulate angthe relative e their behchange in the coaviour in responsentribution to the saratenctions from pes. To meriod t to asut+r1e the rea for each ction of
subject.14 In PUBLIC we find clear statistical support for a disciplining effect of punishment. Spearman
rankthe averag correlation coe relative chaefficientnge in a grous between thps contribe sum of puniushmetion rate in periont pointd t+s sp1 yield poent by a grousitip in peve coefficieriond t and ts in 15
groups and negative coefficients in only 2 groups (in one group we did not calculate the coefficient
since the obin the vast majority of the grouserved punishps, moment levels were lere points spent ss thanon puni 3; see table 2shme.nt lead to higher inc, Appendix 2.B for details). Thus, creases in
contribution rates in the next period. In PRIVATE however, the opposite is the case for 10 groups,
whecoefficire higent poher sancsitive. (We tions leacods to deuld not calcreculasing coate coefficintributienon ratets in 4 grous. Onlyps in which the observed p in 5 of the PRIVATE grouunishmeps is the nt
are eqlevels were less thaually likely can be rejne 3). Hence, the nullcted for PUBLIC (two hypothesis-sided binomi that positive aand negl test, p = 0.002ative correlation ) but not for PRIVATE coefficients
(p = 0.302). This is our fifth result:
Rehigher increasult 5: In PUBLIC puses in coopenishmration rateent has a disciplinins in the following g rouneffect: subjects d. In PRIVATE cooperespond to higher ration ratesanctions do not s with
ent. nishmd to puresponWhy is punishment a rather blunt disciplining device in PRIVATE? A closer look at the punishment
patterns in our two treatments provides an explanation. We first look at punishment with regard to
of others and absolute contreaributioncst to this in the puni. In both informatioshment stage.n treatment Fis, subjgure 3 ectshos caws the relation observe the anship bbsoletweute coenntri receivebutiond s
14 We use the following measure that expresses the reaction of a person as the change in her
contridenote the abution bsolute rate in the followinchangge in the contri period bution rate forelative to the maximum por subject i as Δssii = (ble ingit+1cre/eiat+1  gse or deit/eitcrea). A decrese. We ase
in contribution rate from period t to t+1 implies that Δi < 0 and we can denote the relative decrease by
δ = Δi/( git/eit). For Δi > 0 the relative increase in the contribution rate is denoted by δ = Δi/(1 - git/eit).



punifundameshmentallnt points any across our td abwsolute coo treatmntribuent condtions for botitions. h Subjectreatmentts who contribute s. Punishment panothing retterns doceive si not differ milar
punishment in both treatments: on average 3.5 and 3.3 points in PRIVATE and PUBLIC respectively.
points, anFor any positid is slightly higheve contribution level punir in PUBLIC (with the shexcement lies in the ratherption of the 21-25 narrow racontnge bributetween 0 and 1ion level). Across all .5
positive contributions punishment averages 0.7 points in PRIVATE and 1 point in PUBLIC.
This punishment pattern implies that in both treatments it clearly pays off to contribute at least one
token (if popositive contribution ressible) to avoid ceives a slithe high peght punanishmlty associent anyated with contway, no matter hoributing nothing. w high itHo is (with the ewever, sinxce any ception
inceof 0 punishmntive to respent for coond to puntribnishmutions hient withghe increr than 35 ased contributionin PRIVATE) there appes.15 ars to be no monetary
eceiRstnio ptmenshinu pved1
Absolute contribution
. Figure 3. Absolute contributions and received punishment points

15 To show this more formally we ran linear regressions of contributions as the explanatory variable
responsible for losses from being punished (ignoring observation2s where subjects contributed nothing
in a roun4,3477, R2d). T = he reg0,1635 for PressionUBLIC. The s yield y = -0.margi0664x + 2.8nal ben522, Refits throug = 0,17h reduced penal23 for PRIVATE and y = -0ties are 0.0664 and .0693x +
0.0693 respectively, which do not compensate for the marginal cost of 0.6 incurred from contributing
re token. one mo 34



54io ptmenshinu pvedeceiRstn1


0[-100,-70)D[-ev7i0,-at50i)on[- f50r,o-3m0) av[-3er0,-a10)ge c[-1ont0,10)ribut[1i0,30on r)at[e30, of50) othe[50r, 70g)roup [70,100]
members in percentage points
Figure 4. Percentage deviation from other group members average contribution rate and received
nt points shmepunicontribution This result is of particulrates are not oar ibsemportarved. In PUBLIC, honce for PRIVATE, wewhevere abr, this is onlsolute contributioy part of the story sinns are obcsee surved, but bjects
whican alch shoso awss sess the punicooperatioshmen rates. nt receivedPuni conditionshment accoal on rda subjing to contriects debution rateviation from the other gs is displrouayed in figure 4, p
16membepattern: subjers avceragts are coe punintribshed moution he In PUBLIC, we avily the more theisee a r cocleantribr and proution rate fallnounces shd punishmeort of the rest of the nt
grouaveragps averae of the other gge corntributiooup members ann rate. Punishment is lod it is particulwest whearly high if son the contributiomeone unden ratr-coe is clontribsutes te to the o a high
degree. Spearman rank correlation coefficients between deviations from the groups average
indepcontriendebution nt grourate and the avp level: in PUBLIC 15 erage punicoefficishment pointents as rere negceived provative and 1 coide furtheefficier stantistical evidet is positive (two-nce at the
sided binomial test significant at p = 0.001; for details see table 2.d in Appendix 2.B).17 This
punishment pattern closely resembles the punishment behaviour observed in Fehr and Gächter (2000
preand 20sent stu02) dwhere aby, where asolbsoluute and relte and relaative contributiotive contribuns tioncoincis are not pede due trfectly coo uniform errelandtowmeed, punishmnts. In the ent
contriaccording bution. This cleato contrirly bution rates suggestappeas that coors more pronouperation rates are the nced than pudonishment accordinminant criteria for punig to absolute shment
ns. iosdeciThis hano incentive to reas important implict to puncatishmeions for the int sincence contrintives buto resption rates are unoond to pubnishmserved. In PUBLIC, howevent. In PRIVATE there is er, still
coounder-contribperation rate in the groutors can escape siup. Whethgnificaner itt pays o payoff to doff reductions by clo so depenseds on the size oly matching the averaf the endowmge ents.
For higher endowments it is more costly to cooperate at a specific rate. For example, to match a

16 It is clear that in PRIVATE, where cooperation rates are unobserved, figure 4 merely serves as an
over- and unillustration of how punider- contributshmors in ent pointrelative terms. s, applied on the basis of absolute contributions, are distributed on
17PUBLIC an In PRIVATE d 3 grou 8 groupps in Ps havRe positive anIVATE, because less thd 8 negative an 3 punicoefficishmeents. We hant levels wed to excludre obsere 2 groupved in these s in
ps. grou 35


whecontrireabution s someone with an rate of 50% a person endowmewith an endont of 20 only has wmento spent of 40 tokend 10 tokes needns. Thus to cos, the lowentribute 20 tor the kens,
endocontriwmebutionns.t the more likel In order to demony it is that strate ththe benefit ofat the obser a loweved punir penshmealty outweint is sufficieghsn the cotly strong to msts of higher ake it
deciworthsiowhilen we es to increatimated the se coocoperatiosts ann rated bes we conefit of a nductedhypothetical in a simple cocrease in st-benefcoitntri anabutionlysis.s. Usi For evnerg the y
average contribution rate of the other group members r as a reference point, we calculate for every
undecontrir-cobution ntribution derate (i.e. 0.6*(endocisionwme the loss of ent*r - actual arnings incont stage one that would be nribution)). Assuming a contribution rate ofecessary to match this r would
induce losses from punishment of 1.41 points (as observed in the data), we can calculate the
associated wihypothetical net impact of th the actual punidecisshiomen and nt consequthe reduencection s as the differeof 1.41. The rence bsults shetweow ten the reduhat 55% of all under-ction
contributors would have fared better if they had matched the contribution rate of the other group
averagmembere rate is. Fosr all unde 1.83. We r-cosuntributommarirze thes the saverage findinge net bs in our next reenefit of matching the result: st of the groups
Recooperativesult 6: In PUBLIC ness. Subjectsubjesc are punits punishsh othede mr ogroup mre severelemybers a the lower thccordeir coing to their dentributiongree rate rels of ative to the
average contribution rate of the other group members. Punishment is strong enough on average to
create a mPRIVATE, where puonetary incenishmntive to increent cannot be conase cooditiperationed on ion orden contrir to mbutionatch the group rates, there are no ms averaonetge rate. In ary
incentives to increase cooperation.
puniIn the psychshmeont. Modellogy literatures of proced there ural juis evidstice, for examplence that not only monetae, emphasizery concerns tri that the perception of whgger reactionether s to
punishmePRIVATE punishmnt is viewedent ca as fnair or u be reganjurded stified matteas violating thrs for subse fundameequentant behaviol principles inur (e.g. Tyler, 1988 dubio pro re). In o and
thus might be generally perceived as an unfair treatment. Similarly, it might matter whether
punicontrishmebution nt hits an innrate. To see ocent persowhether the jun withstification a high contribution rate ofor punishment mattersr an actual fre, we consetruc rideted figuresr with a low 5a
and b shoand 5b that show w how relawhethtive (per over- aossible)nd un increase oder-contribr deucreators respose in contribnd diution rates frofferently to punishment. Fim period t to guret+1s are 5a
related to received punishment points in period t. We again use the average contribution rate of the 18
grouPUBLIC subjp as a reects gferencenee poirallnt and diy tend to increase coscriminate betwentribuen tiunder- aon levels. This rend over-contriactionbutor is particulas relative to this.rly strong for In
contrithose buted awho unnd reder-contribceived 0uted. Co or more thanoperation rate 3 punishmens only decreatse (sli points. In PRIVATE the reactioghtly) for those who ovens of thr-e two
prontypes cleouncearly diverge: und than in PUBLIdeCr-contri). Over-contbutorributos tend to incrrs decreeaase theise their cor contribution ntributionsrate ans (althoud this reaction gh this is less appears
ed. r the more they get punishngeto be stro

18 We chose this reference point for com figure 4. Different reference points such as the arability withpaverage contribution rate over all rounds, or 50% contribution, give a very similar picture. Of course, in
PRIVATE subjects cannot infer whether their contribution rate was higher or lower than other group
members. 36


Overall 62% of the over-contributors in PRIVATE contribute at a lower rate after having received
punishmenegatively. In PUBLIC the percentant points in the previous pgeeriosd, wher of over- aneas od nlundeyr-co 28% of the ntribundeutors who redr-contribuuce contritors reacbutit on rates
puniin responshmese to being pnt apparently matters for unisubshed are only 23% and 12sequent behavi% respour, and uectivelynjustified puni. Thus, the justificatioshment seen for ms to crowd
peration. out coo

implications 3.4 Welfare increaPunishment ise in coopes coration rastly and the overates that is sufll welfficiently hiare implicatigh to outweions are pogh thsitie loss throuve only if punishmegh sanctionint triggengrs an . The fact
punithat punishmshment does moent only has re haa moderm than good irate imnpact on this treatmecooperation ratent. Figure 6 compas in PRIVATE resalread the welfare effecty suggs in ests that
terms of realized efficiency gains from cooperation of all treatments. In the PRIVATE punishment
condactivities in stition groups realiage two dzee 46stroy a conside% of the gains of full corable part oof these gperatioains, whn after the first stagich finally ree. However, punisults in only 17% shment
realized efficithe net effect of punishmenency. Compart is cleaed to the condrly negative. In ition without puniPUBLIC the reshmesunt wherlt is similar: ale 40% efficienthough the pcy is reaocsheitive d,
impact on cooperation is much higher, it is not sufficient to compensate for the losses from punishing
and beistage. After png punishunishment haed. The realized gains taken placs eare 78% of realized gaithns are e full coopredueracedtion benchma to 40%, which is clerk after the first arly lower
than the realized gains of 58% in the no punishment condition.


loss through punishment






100loss through punishmentPUBLIC
78%75PRIVATE 58%
46%50tcener pns inaicy genciif efedziealR0
40%40%2517%no punishmentpunishmentno punishmentpunishment
cy gains enlized efficiFigure 6. Rea


3.5 Conclusions cooIn this papeperation ar we have nd punishmstudient decised the role of ions iheteron the contexgeneous ent of a public gdowmeonts and inod. In each pcoeriod, gmplete informroup membeation for rs
had differefully informednt endo about all endwments.owme We also varied the informnts; in PRIVATE they only knew tation conditioheir own: in PUBLIC groun endowment. p members were
We belirelevanece comes frove that this enviromn the fact that in rement is inteality grresting for poup mracembetical ars often nd theoretical rediffer in their asoncaps: The pabilities, aracnd tical
therefore in their potential to contribute to their group. The theoretical relevance stems from the fact
anonymothat punishmus oent is cone-shot gansidmeers (for theed to be a moechretical argumanism thaentts se can mitigatee e.g., Fehr and the free rideGächter 200r problem eve2; Gintis, n in
Bowles, Boyd and Fehr 2003; Boyd, Bowles, Gintis and Richerson 2003). Yet, all evidence that
suprespeportcs this t to their endomechaniwmentsm comes from exs. Our experiperimments theents in reforwhich group e provide a robumembestners ss wechre homogeck of the preeneovus with ious
lts. uresOur experiment conveys two main insights: First, heterogeneity per se does not compromise voluntary
cooon their contriperation whenbution rel puniashment is potive to their endssiowmeble. In PUnt. As a conBLIC peoseqple puuennish othece, contribur groutions inp membecrears dse to high ependent
levels, as they do in experiments in which group members are homogeneous with respect to their
endowithout puwment. Our nishment oppsecond ortunities main finding icontrisbution that informs ination about oth PRIVATE are lower thaers enndowm in PUBLIC. Thuent is crucs,ial. Even the
possibility to hide when there is incomplete information leads to more free riding behaviour.19
Morememberover, pus are obsenishmrent doevable but endos not enhanwmence coots are noperatiot (i.e., nin PRIVATE). Surpri if only the cooperation levsingels of other gly, this is not due to a roup
reluctance to punish. Punishment is still widespread and therefore bound to involve many type II
errors implicatiowhen is re people that establishiwith low enng informdowmeationn about othets but high rs cacontributiopabilitien rates get pus is crucial to anishved. Thusoid type II erro, a practirs. cal
19Güth, Hu This finding ick as cond Ockensisnfels tent 1996with eviden) or fromce gift exchang from ultimatue experimm gamesent ( (e.g., He., Mitzkewitnnizg-Schmidt, and Nagel 1993;
Romore ckenbselfish behavioach and Sadrur. ieh 2003; Irlenbusch and Sliwka 2005), where incomplete information also led to


This is consistent with what Knez and Simester (2001, p.767) report about the behaviour of employees
who get partly paid according to a team compensation system: Employees [] began contacting
colleemployeeagues ths who had ese callcalles enad in sick, to ask whbled them to monitor etherwhet any assiher the abstansece ncwaess nee were duded. eAccordi to valid illnesseng to s.
Thus, establishing transparency is crucial to avoid detrimental effects of punishment.


3.6 ReferencesBoyd, Robert, Herbert Gintis, Samuel Bowles, and Peter J. Richerson (2003): The evolution of
altruistic punishment, Proceedings of the National Academy of Science 100(6), 35313535.
Carpenter, Jeffrey (2003): The demand for punishment, Mimeo.
Fehr, Ernst and Simon Gächter (2000): Cooperation and Punishment in Public Goods Experiments,
American Economic Review, 90(4), 980-94.
Fehr, Ernst and Simon Gächter (2002): Altruistic punishment in humans, Nature, Vol 415, pp. 137-40.
Fischbacher, Urs (1999): ztree. Zürich Toolbox for Readymade Economic Experiments, University of
Zürich, Institute for Empirical Research in Economics, Working Paper No. 21.
Gintis, Herbert, Samuel Bowles, Robert Boyd, and Fehr Ernst (2003): Explaining altruistic behavior in
humans, Evolution and Human Behavior 24, 153-172.
Güth,Werner, Huck, Steffen, Ockenfels, Peter (1996): Two-level ultimatum bargaining with incomplete
information. Economic Journal, 106, 593604.
Hennig-Schmidt, Heike, Bettina Rockenbach, and Abdolkarim Sadrieh (2003): Incomplete and
Asymmetric Surplus Information in Labor Relations. Discussion Paper No.: 2003-012E, University of
Erfurt. Bernd Irlenbusch and Dirk Sliwka (2005): Transparency and reciprocal behaviour in employment
relations, Journal of Economic Behavior and Organization 56, 383-403.
Knez, Marc and Duncan Simester (2001): Firm-wide incentives and mutual monitoring at Continental
Airlines, Journal of Labor Economics, 19 (4), 743-772.
Masclet, David, Charles N. Noussair, Steven Tucker, and Marie-Claire Villeval (2003): Monetary and
Nonmonetary Punishment in the Voluntary Contributions Mechanism, American Economic Review,
), 366-380 (12003, 93Mitzkewitz, Michael and Rosemarie Nagel (1993): Experimental results on ultimatum games with
incomplete information, International Journal of Game Theory, 22, 171198.
Ostrom, Elinor, James Walker and Roy Gardener (1992): Covenants with and without a sword: Self-
Governance is possible, American Political Science Review, 86, 404-417.
Page, T. and L. Putterman (2000): Cheap talk and punishment in voluntary contribution experiments,
mimeo. Romer, John E. (1998): Equality of opportunity, Harvard University Press, Cambridge Massachusetts
Sefton, Martin; Shupp, Robert and Walker, James (2002): The Effect of Rewards and Sanctions in
Provision of Public Goods, Working Paper Nottingham University
Tyler, Tom (1988): What is procedural justice? Criteria used by citizens to assess the fairness and
legal procedures, Law Society Review, 22, 301-355


4 Group sinstitutions20 ize and social ties in microfinance

4.1 Introduction In recent years microfinance institutions (MFIs) have become one of the most important instruments in
developmpilot scheme lendinent policy. The idg small amountea of microfins of money to ance arose invillagers in Ba the mid-70nglas whedesh whn Mohao, dummad Yunue to a lack of s started a
Gramcollateeen Baral, hank to rud no accen suss to convech schemenstion onal loan a larges.r sca Encouralged by hie. Today the Gramegh repaymenten Bank len rates, he fouds to about 2.4 nded the
the worlmillion peod andple. Since Gra a plethora of meenorganisations early success providing ses, thesm conall loancept of mis to the poor hcrocreditas has spve come into beinread throgugho.21 ut
meanMicrofinans confince institutioned to them. Microls are most widendieng prograspread inmme less s hadevelopve receed contly been intrountriesd, althoucedugh they are in transition by no
ecoStates.nomie22 Theres like Bosnia a are mond re thaRunssi 5 million hoa and evuseen in holdwes servstern ecoed by micronomies crelike Canadit schemda and the Unitees in the world d today.
Prior to the mlimited. First, with few subicrofinance restantial possevolution poossionr peopls poes opor phoousertunitieholdss can to take upnot offer colla loans had beteral to back en severely up their
loanrural villages s. Secondbeyond the re, the potential addresseach of the traesd of small loaitional bankins in leng systess dmevelope. Third, althoud cogh loans nuntries often live in remote eeded for
individual projects are small, their myriad nature makes monitoring and enforcement costs prohibitively
programhigh. Poor villmes whicagers only ah providecd subcess to crsidiseedid credt had been thit. However, sinrougch none the-cose schememmercial devs faelopmeced the nt same
comonitorists and weng difficulties are typically dos traditional banomed to failure for that reaks they son. often suffered from poor repayment rates and high
schemeMicrofinansce institution differ vastly in their cos use ncrete iinnovative meanmplementatios to overcons mostme th ofe tse hem shaproblreme somes. Though the main characterisingle stics,
the most prominent of which is that of group lending.23 In a typical microfinance scheme borrowers
one or mwith individuaore groul risky prp membojectsers de form groupfault. Thus, joint liabis which apply for loanlity provides as togetn insuher. The rance awholgainse group is liat individual risks. ble if
Even if an individual projmight still be able to do so. In this sense joint liabilitect fails and some of the bory rowers are unserves as a substitute able to repfor ay, the group acollateral. Unless the s a whole
subsindividual ritantially losks are pweer than witrfectlh individual by correlated, the ovorrowierall ng. risk of involuntary non-repayment can be
Compproved to be ared wia great th traditional succecredit pross. Repaymegrammes in lent rates leapedss developed co to levels previously ununtries, microsecreen in less devdit schemelopees have d
regions. Grameen reports repayment rates of more than 90%; other programmes replicated such
progfiguresram. Homeswe have failed tover, the story is not witho live up to their promiut blemisse.h: whil24 Furthee many wermorre sue, the ultimate gccessful, numerouoal of establis MFI shing

20joint work wit This chaptehr is ba Klaus Abbised onk ann the wod Bernrkidng p Irlenbuaper Gsch. roup size and sozial ties in Microfinance institutions,
21BRAC in Ban Examples ingladecludsh, Pride BancoSe Africa in Keol and PRODnya, Tanzania and UganEM in Bolivia, Banka Rakyat and Badada, Malawi and Ugan Kredit in Indonenda, FINADEV in sia,
Benin. 22 See CONLIN (1999), or ARMENDÁRIZ DE AGHION and MORDUCH (2000) for an overview.
23 Therefore, microfinance institutions are often referred to as joint liability lending institutions (JLLIs),
though Aghion ansome d Mordinstitutionuch (20s al00)). so give small loans to individuals with good reputation (Armendáriz de
249). lin (199 See e.g. Con 41


sustainable credit schemes for the poor has not been reached, and most programmes still rely on
subsidies and donations.
To improve the performance of microlending it is vital to improve the design of these schemes. Among
practitioners as well as academic scholars there is a heated debate on the appropriate design of their
key features. Lending to groups involves a fundamental dilemma: It may insure the credit against
involuntary defaults, but individual borrowers reliance on fellow borrowers to repay the loan gives the
former an incentive to free-ride. Indeed, if the success of an individual project is not sufficiently
verifiable by other group members the dominant strategy for each individual is to shirk and hold others
liable for own default. Being aware of this peril, MFI schemes have usually incorporated a number of
safeguards, the most prominent of which is that borrower groups be self-selected. This is the case in
many programmes, the expectation being that close social ties enhance peer pressure and group
solidarity. In a theoretical study, Besley and Coate (1995) show that the possibility of inflicting social
sanctions on peers helps improve repayment.25 However, the effectiveness of self-selection is not
undisputed; Wydick (1999) empirically investigates MFIs in Guatemala and finds no evidence that
groups made up of acquaintances have higher repayment rates than those consisting of strangers.
Social ties m26ay even hamper repayment discipline if they lead to more forgivingness towards
defaulters.Free-riding incentives may depend crucially on the size of the borrowing groups. In practice, it is
unclear how far group size affects repayment rates. FINCA, the organisation which pioneered the
village banking concep27t, lends to large borrower groups of between 10 and 50 members, and boasts
repayment rates of 96%. On the other hand, Grameen prefers smaller groups with typically only five
members, in order to keep free-riding and in-group co-ordination problems under control. In the
academic literature, both positions have their advocates. Ghatak and Guinnane (1999) argue that
despite the insurance effect of larger groups, smalle28r groups are to be preferred for their better in-
group co-ordination and reduced level of free-riding. On the other hand, Buckley (1996) empirically
rk effectively. bers still can wo memwith ten or morefinds that groups inceWhile muntives is ch ofacknow the literatureledged to a much l focuses on gresser exoup sitent. ze and sociIn general MFal ties, the importaIs aim at forming long tence of dynamic rm
relationships with their client groups. Follow-up loans are frequently made subject to whether precious
loans have been repaid. These two features are intended to encourage compliance with repayment
obligations. It can be argued that this aspect of microcredit schemes is at least as important for
generating repayment discipline as peer pressure between the group members. Creating dynamic
incentives may become vital if microcredit schemes are to be applied to other economies. In the urban
contexts of transition economies, for instance, it may be more difficult to form self-selected borrowing
groups than in closer-knit rural communities. For this reason Armendáriz de Aghion and Murdoch
(2000) argue that in such economies the focus on group-lending should be abandoned and suitable
dynamic incentive schemes should be sought.
The empirical evidence on how the various design features of microcredit schemes affect their
success is still limited and controversial. Murdoch (1999) promotes the need for well-designed
experiments to identify the impact of MFI design features on their performance. However, controlled
experiments in which single properties of institutions are systematically varied are difficult to carry out

25less on soci Conlin (199al9) argues ties. See also Wen that replicner ating th(1995). e MFI concept in western cities requires that schemes rely
26 Guinnane (1994) conjectures that such an effect may have contributed to the failure of Irish credit
27co-operatives in the nineteenth century.
Villages in FI FINCAs scheNCA projecme differs from others witts form self-governeh resped groupct to the interns to which the lal organioan is givezation of n. The distborrower gributiroon of ups.
credit is left largely to the group members themselves.
28 This argument is supported in empirical investigations by Mosley and Dahal (1987) and Devereux
free-ridinand Fishe (19g effects a93). There porretiovical inveded by Impavido (19stigations into the t98) anensid Armendon betweárizen po de Aghion (1sitive insuran999). ce and negative

29variables, e.gin the field, due to probl. the individual projeemct ris of data accessibility and sk, are unobsecomparability.rvable. Therefore, we int Furtherrmoodre,uce an many releva alternative nt
approawe can control for spech to the empicific rical analysiparamseters a of microfinnd obanservce e binehaviourstitutions. In an intera in simulated MFI ctivescena laboratoriorsy experim directly. ent,
Furthermore, we can identify which factors influence behaviour by changing particular variables of the
experimental environment, holding all other aspects unchanged.
behavioAs a startinur cognne point for oucted to group lenr resedarch aging enwe modda, we coel a sinstuation in whitruct a stylisecdh re MFI scepaymenario.nt depends o To study freen grou-ridip ng
solidpast. In our earity alonxperimee. To implemnt each mement dynamic iber ofnce a group of ntives, n follow-up loaplayers investns are subjes in an individct to full repaymeual risky project. nt in the
Whether the project succeeds is known only to the individual investor. Subjects decide individually
whethable to coer orntri nobute. The expt to contributeeriment to the group endrepayms if too few conent. tHoweveribute, that isr, only those , if the grouwith sup as a whole ccessful projecacnnot ts are
success: (1fulfil its repayment obligati) The group size, which we on. We focus set to n = 2, on three inn strumental vari= 4 and n = 8ables id in three coentifiednd as cruitions, (2cial for MFI ) the dynamic
incetreatment" suntive strubjectcture, as alrend (3ad) the inty had to enrolensity of soci as a group toal ties between grou capture the influenp members. Ice of social tien a "grous. p recruitment
We obrepaymseent rarve a high antes are gened robust perally higher thrformance of grouan those acphi lending inevable by indistitutions in all ovidual lendinug.r treatm While indivients. In fact dual
contrigreater dibution spersiorates dencrea of risks. We clese slightly witharly id largeentify tr grouphe importans, the impace of dynamic inct of free-riding icesntives. Toward alleviated by the s the end
of the experiment repayment rates decrease substantially. Furthermore, we find that social ties only
have a relatively moderate effect on repayment rates. Closer-knit groups have higher, but less stable
repayment rates than those composed of strangers. We also observe gender effects: In our
experiment women tend to contribute more than men.

Related experimental studies 4.2 To our knowledge laboratory experiments on microcredit institutions have not yet emerged in the
literature. However, since microcredit institutions allow group members to free-ride at the cost of the
group, valuable insights may be gained from the literature on public good games. In such games, each
subjgood. Everybect of a group of n peody in the grorsons can deup of n individualcis rede to invest aceivesn a retu amount x (urn of cx, where p to a certain c < limit y) in a public 1, but nc > 1. Thus, it
is a dominant strategy for rational players not to invest, but the Pareto efficient solution is realised if
everybody co-operates by investing the maximum amount. In experimental public good games
subjects typically contribute considerable amounts but fail to reach the social optimum.
Inspired by the microfinance theme, Barr and Kinsey (2002) conduct an experiment on such a public
good game in Zimbabwean villages. Though they do not aim at modelling a microfinance scenario,
their research question is closely linked to common MFI practices. Many microfinance institutions
target women as their clients, partly because they consider womens empowerment as a goal as such,
but also because women are often seen as more reliable borrowers.30 The authors test this conjecture
by analysing womens and mens behaviour in the standard public goods game. The differences they
find are small, but qualitatively supportive of the MFI practice: Women tend to contribute more to the
han men. good tcpubliThe study most akin to our experimental set-up is the public good experiment by Suleiman, Weiss,
and Bornstein (2002). In their study, the players endowments y are stochastic and private information.
Thus, as in our design, players cannot identify whether other players failure to contribute is due to bad
luck or shirking. The authors observe less free-riding in this environment than in a standard public

29 These difficulties are discussed in Bolnick (1988) and Hulme (2000).
30respon There is sosibility (Ledmgee evidenrwood ce th1999:3at women8). Morducs repayment dih (1999:1583sciplin) reporte is highs studieser due to a b findingetter se that male lending nse of
focugroused pps at Gramolicy see Goeteen have higher dz and Sen Guepta (1996fault rates than fem), anad Kabeele grour (20ps01). . For critical accounts of the women-

good setting with fixed endowments. Thus, players are less tempted to free-ride themselves when
ding. ot identify others free-rithey cannSeveral authors have examined the role of social factors in experimental public good games. All these
studieto all players.s, howe Gächtever, deal withr and Fehr symmetri(199c9) i situationnvestigate s in whiwhch enether socidowmeal anpproval ints are thcee santives redume for all andce free kno-wn
riding btowards theiehavir grouour in a repep members afteated public gr the experimood game. Subjecent. It turned out thts have the oppoat social apprtunity for sociroval alone coal approval uld not
enhabefore the exnce co-operimperatioent con. Ho-operation increwever, if in addition subjeases significactsntly. coul31 This suggestd familiarise themsselv that ex-ante familiarites with each other y may
(200be impo2), and Brandtrtant in establiss, Riedl,hing co-ope and van Winration. Inden (2 another 002) instudy Van Dijk,vestigate the developm Sonnemans,ent of soci and van Winal den
conattachmenectednent becomss ases stron a result of repeated iger after sunteraccessful ctcoion in -opera publiation. c good setting and find that social
experimThe effect of grouent with 4, 10, 40, p size has beeand 10n studi0 partiecipants. Td first by Isaaheyc, Walker, an find that contrary to the cod Williams (1994) in a pmmon conjeublcture ic good
contributions even increase with very large groups. A similar result is obtained by Carpenter (2002),
who increacompse huaregs grouely as the grops of 5 and 10 up size insubjcreecasets. Hos, whicweveh r,may account in both studies mafor this effect.rgin Unleal social bss therenefits e are
strong synergies between the individual projects within an MFI borrowing group this is typically not a
characteristic of microfinance institutions.
Unlisolely deke in mopenst public god on the willinod gnegamess bs mentiut also on the aboned above ilitrepayment in y to repay. Througa microfih no faultnance scena of their own, rio does not
individualfellow group s whose projmembers. Indiect fail canviduals not cosensntribute to e of solidarerity is assepayment and ssed inhen the solidce they rearitly on the solidy game experimented arity of
by Selten and Ockenfels (1998). Three players each roll a die to determine whether they win a prize.
Winnmake subers castan tranntial transfesfer moners where femay to losers. Contrales showry to more solid the game theorearity than maletic predis. ction the great majority
Nonthere is neo st of these studieudy examinins hag grous been carried p solidarity out with a in a dynamic envimicrofinanrce bonackgrounment with follow-up bed in mind. Consnefits equently,
conditional on compliant behaviour. The incentive structure in the existing studies is quite different
from a typicalimmediately tran microfinsferred to ance environmthe microfinanent. Therefce instoritutione, ths in quee findingstion. Hens of these studice the nes caneed for a nenot be w
ent. experim

Model and experimental design 4.3 We consider a very simple experimental set-up capturing essential features of group-lending. Since
microfinance institutions come in many different forms, we were forced to make design choices. In the
present study, we focus on the conflict between free-riding and group solidarity in the borrower group,
thus, we assume that individual repayment cannot be enforced. To focus on the group solidarity
aspect, we assume that indi32viduals have no means of verifying the success or failure of their fellow
borrowers individual projects. Furthermore, by assuming that the success of the individual projects
is uncorrelated, we abstract from complications that arise if risks are connected, as are, e.g., if their
success depends on seasonal conditions. Finally, we consider a symmetric situation in which there
are no differences in the individuals strategic situation.

31Plott (1985), I This is in linsaac ane with earlid Waler ker finding(1988, 19s by Dawe91), Bohns, McTeta and Frey (199vish, and Shaklee (199). Although t77), Isaac, Mhey examine cCue, and different
questions, they all find that pre-play communication increases co-operation.
32 This is in line with the assumptions made by Besley and Coate (1995) in a theoretical study. The
(198issue of pee1), Stiglitz (19r-scree90n)ing a, Varindan (19 peer-m90)o, Ghatak anitoring in grnd Guoupinann lending ie (19s99 ad)dres, Armendárizsed by Stiglitz and de Aghion aWeiss nd Gollier
(2000), and Ghatak (2000).


model 4.3.1 The liable. The loA group of n ian enandividualbless re each grceives a loaoup men, for the repmber to inveastyment of which all grou in an individual risky projp membeect. All projers are jocintly ts are of
the same type, and the probability of success of any given project is 5/6. In case of success, the
investor receives a project payoff of 420 talers (the fictitious experimental currency). If the project fails,
however (the probability of which being 1/6), the subject receives a project payoff of zero.
After all projects have been carried out, the group loan has to be repaid. For repayment to ensue, we
assume that each individual is supposed to repay 210 talers, and hence the group is liable to repay a
total amount of 210n talers (for example, if we assume a loan of 175 talers per individual and an
intereensure that thst rate ofis condition p 20%). Those individualertains we ass whossume thate proje no cinvestor ts failed carennceivot contribues income from anothete to the repaymr source ent; to
and that none possesses savings which could be used to repay the loan in the eventuality that the
. t failsprojecIndividuals whose project succeeds decide whether to contribute to the group repayment. As
can amentionescertaid abn whetheove, informatir an ion on tndividualshe proj deects fault is stratesugiccess or failure ic or ds prue to the failure of the projivate; no other group meect. Hence, we mber
model an idealised scenario in which the repayment of loans must ensue in the absence of means to
ment. enforce repayTo model joint liability in a simple and straightforward way, the debt of 210n talers is split evenly
the higheamong those r is the burden foindividuals who are ar the single ble acontributond willing to r. cSince contribontribute. Thutionus, the less ins can only bedividual financed fros contmribute, the
current rounds project payoffs, full repayment is only possible if at least half of the group members
bute. contriproceeds in tOnly if the group fulfils its repayhe same way with the samment oble grouigation, doep mems bers (A maxithe game contmum of ten roundinue into a further s can beround, whic played). If h
projemore thact failuren half the group m), then the group cannembers dot reefaulpay the t (regardlefull amount, in whiss of whethecr the default is sth case no further rategirounc or dds ueare to
played and the subjects of the group obtain no further payoffs during the experiment. With this feature,
we model the practice of many MFIs which make follow-up loans conditional on the full repayment of
previous loans.
After each round, players are informed about the number of contributors in the respective round (but
payoff minus not their identthe players sharities), their own project payoff, and e of the repayment burden). their round payoff (comprised of ones own project
contriThe gambutions e theoreticat all are mad predie thus bction, assuminringingg play that players maximise th to an end after the first roeir own incound. Thme, is that noe intuition is
straigpenultimate round. table 4.htforward. In the last round, it is ob1 shows the payoff distrivious that no buplayer would tion a player would gever coentribut given the numbte. Now conseir of der the
other players willing and able to play, for eight-player groups (n = 8).33 The entries in italics mark the
casepayoffs are s in whiccomposh the numbeed of the (sure) payr of players contributing offs in the pesufficenultimate ros for the gund plusame to rea the loch ttery that the the final rounplayer d. The
gets in the final round. It can easily be seen that the payoff distribution a player gets from not
regacontrirdlebuting stochass of the actual nustically dominatember of othes thr players e one he orwillin she wog and abluld obte to repay. Theain by contribrefuting. This is ore, it can netrue ver be
round 10. By induprofitable to contribute in thction, it follows that in a sube penultimate round, game pand herfect eence one quilibwould prerium, no contdict thributionat no game es are mnters ade in
any round, and play will end after the first round, in which no contributions are made.

33 The argument is completely analogous for n = 2 and n = 4. It can easily be generalised for a wide
range of other parameter constellations. Notice that the argument does not require assumptions about
s attitude to risk. ectthe subj 45


Table 4.1. Payoff with and without own contribution in the penultimate round (n = 8)
Decision 0 1 number of other player2 3 s willing and able to repay 4 5 6 7
55555cont. 0 0 0 420; 0; 1/6 /6 84; 504; 1/6 /6 140; 560; 1//66 180; 600; 1//66 210; 630; 1//66
5555not cont. 420 420 420 420 840; 420; 1//66 420; 840; 1//66 420; 840; 1//66 420; 840; 1//66
Notes: The entries in italics are lotteries, where each line represents one outcome. Within each line
he payoff and the probability of winning this payoff. colon separates tthe semidraThrows augh the n idealiseunambid pictuguoreus game theo of a microfinretic propance situaterty ion in whiof the subgch reamepayment ca perfect enquilinot be estabrium the gblisaheme d by
liability in a microfinplayers own payoff maximisation alance relationship, sinone.ce re Thereforpaymee, the game allont out of group solidws us to stuarity stdany the impact of ds in sharp contrajoint st
. r behaviouumto equilibriIt is worththe loan if the projewhilect is succe to look at a hypothetical benssful. This occurs withchmark a prob of individual lendability of 5/6. Contraing. A single individual cary to the case of group n repay
lending, an individual would prefer to repay in all rounds bar the last, since the benefits of future credit
canoutweinot excegh the edshort-te 5/6. Thus, while grm profits of shirkinroup-lgending cr. However,eates free since p-ridirojeng icts ncentives (amay fail, expeccording the gcted repaymame-ent rates
genetheoreticrate hig predher repaymentiction there would be n rates aond mo rere profitable payment at all), the dispeloans for the lendersion of risks mr. This, hakeoweves it possibr, requilres e to
group solidarity, and one of our research questions is precisely whether group-lending mechanisms
are able to outperform the benchmark of individual expected repayment rates of 5/6.

4.3.2 Treatments We designed our experiment to examine two major issues in the design of microfinance schemes. The
first issue concerns the effect of different group sizes on repayment performance. In absence of
strategic default, larger group sizes provide some insurance against uncorrelated individual risks.
However, it is unclear how group size behaviourally affects the tendency to free-ride. If free-riding is
more pronounced in larger groups, this might counteract the insurance effect of larger groups. To test 34
for the effect of group sizes, we conducted experiments with group sizes of n = 2, n = 4, and n = 8.
The second issue we address is the effect of social ties on behaviour in a microcredit group. Typically,
MFI borrowers are self-selected groups whose members have known each other for some time. Thus,
we induce different levels of social ties by applying two distinct recruitment techniques: In the
individual recruitment (IR) treatments, subjects register individually for the experiment, thus minimising
social ties between participants interacting with each other; In the group recruitment (GR) treatment,
potential participants are required to register for the experiment in groups of four. The latter method,
ensures that groups are self-selected, since subjects need to form groups in order to register
themselves for the experiment. This method resembles the self-selection process required by real-life
microlen35ders. As in their procedures borrower groups are formed before they enter the microfinance
scheme. To control for the level of acquaintance between the group members, we requested the
subjects to indicate the intensity of their contact to the other group members on a scale from 1 (no
contact) to 7 (frequent contact), separately for each of the other group members. We distinguish
between professional and private contacts. It seems plausible that social ties are more pronounced if

34 This roughly covers the range of most typical MFIs. Only few lend to groups with more than eight
wers. borro35 When they register for our experiment, however, subjects did not know the task that they would
perfoinformation arm. This was explbout the task waaines the samd to them only in the ace for all subjetuacl sessiots at the outn. This enset of a sesured thassiton. the state of


the contacts are intensive and private. All statements were made anonymously.36
We conducted our GR treatment with a group size of n = 4, given our concern that a group size of two
would allow effectively act as one thecouplesn (with a comm the social on hoconfuselict is remhold budgoveet) to regid and the inster as a gcentive struroup.ctur If the two plae is reversedyers . In fact
hand contriwe did buting inot conn all rounducds ext experimecept the lastnts wi one is socith n = 8 for pracally as well atical reasson individually ops: it would have proved totimal. On the other o difficult
experimto find self-seent, and, in bralected grouckets, the numbeps of eight subjr of ects. Tasubjecble 4.ts in the corre2 indicates thsponding tree factorial daetment. sign of our
ment ris of our expee treatmentTable 4.2. ThGrou Group size Rep cruitment method Individual
n = 4 n = 2 GR4 (32) IR4 (28IR2 (16))
)IR8 (64 n = 8 Notes: In brackets are the numbers of subjects.

sprocedureal 4.3.3 Experiment(eLaThe expeb) at the riment was condUniversity of Erfurt, Germucted in the Erfurter Laany. Most subjects boratorium für expwere stuerimedents from variountelle Wirtschaftsfos disciplinrschues, ng
where students of economics, law, and sociology constituted the largest fractions. To minimise
presentation effects, we designed our experiment in a completely context free fashion. We presented
the microfinance situation to the experimental subjects without connecting it to a microfinance story.
We opted for a neutral presentation to avoid the uncontrolled effects of possible connotations raised
by hypothetical stories, and to ensure best possible comparability with other experimental results.
Each session began with an introductory talk after which the written instructions were handed out to
the subjects (Translations are provided in Appendix 3.A). The instructions were read aloud and
explained in detail. After the introduction, the subjects were seated in cubicles, visually separated from
one another by curtains. In the IR treatments, the cubicle numbers were randomly attributed to the
subjects, and no subject was informed about the identity of the other players in his group. In the GR
treatment, the subjects who had registered together formed a group. This was known to the subjects
and further emphasised by decision sheets in different colours for the different groups in a session.
In each round, the success or failure of a project was determined through independent random draws
for each subject, with a probability of one sixth for the failure of the project. Each subject rolled a die to
determine his or her projects success or failure for each round. In order to overcome the difficulty of
monitoring the veracity of subjects reports about their draw, we asked subjects at the very beginning
of the session (before they knew the rules of the game) to roll a die ten times and to enter the
outcomes in the first column of their decision sheet. Later in the experiment in each round the losing
number which was the same for all subjects was determined by a randomly selected participant
advanrolling a dice for this roe. This nuund mmber waatches publid the locsily annong numbunced. er All subjecmet with the projets whose individuct failureal numbe and thus a prr drawn in oject
. payoff of zeroEach round proceeded as follows. The subjects were each handed one decision sheet, on which the
complete history of play was presented on one page. The subjects indicated their repayment decision
by ticking Yes or No boxes for the current round. After all subjects had completed their decisions,
the experimenter collected the sheets. The losing number for the current round was then drawn. By
36 For different research questions other techniques have been applied to study the impact of social
ties. Gächter and Fehr (1999) and Bohnet and Frey (1999) invite strangers, but allow subjects in one
condition to get acquainted with each other in a pre-play communication stage. Van Dijk, Sonnemans,
and van Winden (2002) and Brandts, Riedl, and van Winden (2002) study the evolvement of social
ties through interaction and assess social ties using psychological tests before and after play.

letting the subjects make their repayment decisions before the losing number was drawn, we
compgathered deuted thecision results als of theso in the ca roundse ts anhat d distributean individualds the deci projsioect han shs eefailed.ts contai37ning The experimthe results. enters
All groups in the session38 completed this procedure ten times, even when their groups play had
actually enwere not reveded as a result aled by diverging duof repaymeration of play nt default. This ensure(this would have didstorted both a, first, that the constitution of the nonymity and the groups
not coustatistical indnteraependence ofct the monetary incentives. the group s), and second, that a preference for a short playing time could
Immediately after the session, the subjects were paid anonymously in cash. The exchange rate was
averagset to DM 0.02e of DM 25.50 for 1 per tale½r. The total earni hours, considerngs in the ably moseression ra than a studngeents nod from DM 7.8rmal 0hourly wage i to DM 65.20 nwith an Erfurt.
approximatelOne DM is eqy DM/$ 2.10. uivalent to  0.51. At the time of the experiment, the exchange rate to the US dollar was
We aimed at gathering eight independent observations per treatment. Since a few subjects in one
se16 sussibjecon did not shots, such that in thiw up, osne se treatmession in nt, only sethe IR4 treatven indepenment couldent obd only be conservationdsucted with 1 are available. Sin2c rather than e
subjtreated aectss in different group a statistically indepens within a sedent obssion did servation. not interact with one another, each group can be

4.4 Results The deplay. Most besign of our expehavioural effects in ouriment focuses or datna expre the impact ss theofm group selvessize, social tie in the overall nus, andmber of contrib the dynamics of utions
we observe in the treatment. However, before we conducted the experiment, we decided to also look
at the contribution rates for the rounds 1-9 only. Since the game is certain to end after the tenth round,
we may expect that behaviour in that round to be substantially different. To test for treatment
contridifferencbution es, we chorates in the inse to apply Fishdependent ersus two-bject groupsamples. permutation test to all pairwise comparisons of the
Furtheadvantagr, we aleous sinso ance at thealyse the first rou very beginnind behavioung of play earch. Statistical test individual decis onsio first roun is a nd behavioustatistically indeper can be ndent
observation. Repayment performance
averagTable 4.3 showe in the four treatms how many participents of ouants r experimagreed to ent. Acontribute (chodditionally it reports avese yes on the drageeci contribsion ushetion rates in et) on
the first round as well as averages over the first nine rounds. In the first round, between 82.3%
su(treatmccessful. Tent IR8) and 9he contributio8.9%n rates (GR4) of the subjover all roundects decis of the experimde to contribuent are lote if their projwer but still conect turnsids out to be erably
high. Notice that according to the selfish equilibrium prediction we should observe no contributions at
all. Table 4.3. Contribution decisions and average duration of play
IR8 GR4 IR2 IR4 ent Treatmyes-choiyes-choiceces is an rounll roundd 1 s 79.8% 93.4% 76.0% 85.7% 72.3% 82.3% 74.1% 98.9%
averagyes-choicee nums roundber of ros 1-9 unds played 7.5 81.3% 7.0 79.4% 5.1 72.3% 7.0 75.9%

37spon Brandts ataneound Chars play and comness (2p0lete strate00) examine gy elicitatibehaviouron, bu in t find resultdifferent interactive games to be unaffes compacted. They sugring gest
that both procedures be equivalent for low-complexity tasks.
38coul We cod not idenduncted each setify which of the otherssio n with at leastsubjects be two groulongps to to their group. ensure that subjects (in the IR treatment)


From the lenthey affect the repaymders point of view the mostent rate. To address this qu interestinestg qiuestionon in our exp regaerimerding differental framewont MFI desigrk one ns icouls hod loow k
at the actual repayment rates realised in the sessions. However, these are highly influenced by the
realisations of the random draws. Thus, we rather look at expected repayment rates, which we define
as the repayment rate that the lender could expect if we take observed default rates as a proxy for the
the probability that a borrower is willing ϕegically. Denote by ratprobability that a borrower defaults stto repay. Then, actual repayment probability is 5/6ϕ, since he can only repay when his project is
ually contributing is binomially the number of group members actsuccessful. Given this probability,distributed with a single event probability of 5/6ϕ. Thus, the expected repayment rate ERR can be
suted acompn/2−1n
210⋅n , where B denotes the non-cumulated binomial distribution. The repayment (in talers) in the case that
less than n/2 group members actually repay is given by the first sum of the numerator. In this case, the
sum loan is ocompnutesly partially rep the expeacteidd repaym, but those who pay mustent in ca se that the wholgive their entie loan irse proje repcaid.39 t payoff of 420. The second
ment rates repayExpected1



GR460.123456789round s ent rateFigure 1. Expected repaym Figure 1 depithe observed rates of cts the expecontricted repaymbution deciensiot rates for thns in the correspeondi four treatmeng round annts of the experimed treatment. nt, using for ϕ
As mentioned earlier, individual lending can at maximum generate a repayment rate of 5/6 = 83.3%.
Rerepay. In our call that according to theexperiment, howeve theorr, lendetical predictioers would prens, indifer grovidual borroup lendwers ing to individwould always, grouual lendingps n. In all ever
substreatmenttantially highes, expected repar in earlier roundyment rates as. Thrus, groe above thiup lens dtheoretiing outpceal berfornchmms inark modividualst of the time, lending. Even without and are
39experim The insuental raquestionnnce effect of larger gaire, we asked roups suwas bjectprisnci to estipally undemate the prorstood bybability to rea the subjeccts. In a post-h the final round for
are willidifferent groung to contrip sizes (bute. On ∈ {n averag2, 4, 8}) and round ne the influence umbers (n of the two dime∈ {5, 10, nsions20}) give was an thassetssed qu all group mealitatively mbers
correprevioucst, thoug findings by Gnh overall preeobzy (199abilities 6) andwere quantitativ Abbink, Irlenely undebusch, anrestimatd Renner (2002ed. The latter result repli) in the sense tcahat tes
dynamic effects are not sufficiently taken into account.

makinany monitorig groupng-lendi oppong advartunitiesntageo, trust anus to the lendd group solidaer. rity can establish substantial repayment rates

the group size The effect of4.4.1 While large groups allow for a greater dispersion of risks one could expect that they are also prone to
more free-riding. Indeed, table 4.3 suggests that contribution levels tend to decrease with the size of
the group but only to a moderate extent. The statistical analysis provides only weak support. The non-
grouparametrips, rejec Jonckhects the null hypere teothesis of eqst, applied to the rates ual rates inof yes-ch favour of the hypotheoices in roundsis s 1-9 in thof decreae singlsing e inderatpendes at a ent
40weahypothek sisis ofgnificance level of equal contribαu = 0.tion rate10 (os onne-sidely for the d).com Of all pairwiparisson of thee compari contribsons, we ution rates in rocan rejeundct the null s 1-9
for IR8 versus IR4. The rates tend to be weakly significantly (α = 0.10, one-sided, Fishers two-sample
permutation test) lower in the IR8 condition. All other differences are not statistically significant.
Table 4.3 also indicates a decreasing average duration of play. However, the Jonckheere test does
not rejeis not only dect the null hypothetermined by the susis of an equbjects decial numbesions,r of play but also by the chaned rounds. ce moveNotice that the contins determining uatiosuccess n of play
ts. cprojeand failure of the individual

mics of play Dyna4.4.2 Since we havare fewer rounds left in whie a finite number of roch profits canunds the dynamic be made. Thinceus, we ntives beshouldcome expect coweakentrirbution rate over time, as there s to
decreatwo thirdsse to of subjwardectss the end contribute in rounof play. Figure 2 shod 7, latew rouns thatd contri this is indebutioned the ca rates fall subse. sThougtantially. Hoh still more thawever, n
considthey do not reerable ach number of suzero, even in the bjects contribute, though nlast round of play. It seems o further rounsurprisingds can be expecte that even in the last d. 41 round a

roundsbution decisions overiContr1cesio0.8 chyes0.62IR of yncquel. frereGR4
round Figure 2. Relative frequency of yes choices over rounds
The decreasing trend of the contribution decisions can be observed not only in the aggregate data, but
correlalso in the ation cosingle groefficients bupsetwee. We compute,n the number of yes- for each session choices aseparatend roundly, non-pas. Table 4ram.4 repoetric Sperts thearmsan rane raknk
correlation coefficients. Using these as summary statistics, the binomial test rejects the null hypothesis
40 The analysis of group size effects is based on the three treatments with individual recruitment.
41contribute, then own co Possibly a solidarity motive accountribution can eants fose the fellor this type of behaviow players burdur. If a subjen of repaymect expent. ects other subjects to

α = 0.05 (onthat positive and nee-sided)gative correl for all four treatmation coents.42 efficients are equally likely at a significance level of at least
(excluTable 4.4. ding round 10 Spearm) an rank correlation coefficients of yes-choices over the played rounds
grouTreatmp ent IR2 IR4 IR8 GR4
2 1  0.51 0 0.52 0.82 0.77 0.88* 0.61*
4 3 0 0.27  0.62 0.89 0.67 0.77 0.63*
0.95  0.42 0.73* 5 7 6 0 0.67 0.75 0.57 0.32 0.98*  0.65
0.75* 0.77 0.87 8 * significant at α = 0.05 (one-sided)
 number of played rounds lower than three

ies social tThe effect of4.4.3 In the GR treatment, the groups who registered together can be expected to be a self-selected group
in which social ties are stronger than in the anonymously matched groups of the IR treatment. The
question arises whether these stronger social ties result in higher repayment rates due to a higher
impact of group solidarity in self-selected groups. Our data provide mixed evidence with this respect.
While the first round contribution rate rises from 85.7% in IR4 to 96.9% in GR4 (insignificant, p = 0.12
according to the Fisher exact test), we obtain an overall contribution rate which is even slightly lower in
GR4 than in IR4 (74.1% versus 76.1%). Thus, the comparison of the contribution rate over the two
treatments does not provide strong evidence for an effect of social ties on repayment rates. These
results are in line with those of the survey study by Wydick (1999), who finds a surprisingly small
degree to which social ties within borrowing groups affect group performance.
However, further exploration of the GR4 data indicates some support for an effect of social ties. As
figure 2 indicates, the rate of yes-choices starts at a higher level in GR4, but we can observe a sharp
decreaof the experise in later roundment. If we compars. In the IR4 treae the Spearman ranktment, the rates of yes-cho correlation coefficiices is moents of re stable towathe yes-choirdces in the s the end
singlgreatere gro (in aupbssolute valu of these two treatmentes) than in IR4. s, as repFisheorsrted in two-sample table 4.4, we f randoimisation tend that the coefficientst rejects the null s in GR4 are
hypothesis of equal rank correlation coefficients at a significance level of α = 0.01 (one-sided).
Stronger social ties seem to induce a high, but less stable willingness to repay. Possibly subjects tend
to be less tol43erant towards their peers suspected shirking when the fellow group members are their
s.friendWithin the GR4 treatment, we find some evidence that the extent to which the group members are
socially tied matters. As mentioned earlier, we asked for statements about the level of acquaintance
between the group members. When we correlate the level of private social contacts to the overall
contribution rate in the groups, we obtain a Spearman rank correlation coefficient of rS = 0.73, which is
significant at α = 0.05 (one-sided). Thus, stronger private contacts between the group members seem
to have a positive effect on repayment decisions. The same analysis with the level of professional
contact does not yield a significant result.
42 Notice that the exclusion of round 10 is conservative to our findings.
43 It is well-known that subjects in public good experiments cease co-operation when they observe
free-riding by other group members (see, e.g. Keser and van Winden (2000) and Falk, Fischbacher,
and Gächter (2002)). Conditional co-operation is much harder to achieve in our game since other
players shirking is not unambiguously observable. However, subjects may abandon co-operation if
they feel that observed total contribution levels are suspiciously low.

effects 4.4.4 Gender Thougexperimh the stent, we kept reudy of gendecord of the gendr effects has ner of eaot been onch pae orticipaf our cont. Table 4.5 shore issues at the outws first roset of the und and overall
is highrates of yeer thans-choices, for women an that of men in all treatmend metn seps with iaratelyndividual re. It can be secruiting.en Fishe that womenrs exact test applies rate of yes-choid to the ces
number of yes-choice44s in round 1 yields a one-sided significance level of α = 0.01, if we pool the data
s.IR treatmentfrom all three Though the frequencies in table 4.5 show sizeable differences in the contribution rates of females and
observationmales, the res doesults shos not allould bew to com interppreted rehwith caensively cution. First, the relontrol for all possiatively small nubly relevant socio-ember of indconomiepc endent
can variablesstatistical, such that we canly detect the effect only for the poolnot detect compound eed daffects with otheta of the first round dr backgrouata ofnd variabl the individual es. Second, we
recruitment conditions, such that the evidence is relatively weak (mens contributions in the group
pracrecruitmentice to target womet treatment are even slign seem difficult, as the htly higher). Finallculturayl backgroun, interpretations d of our subjewith respcect pool regt to the caurding rrent MFI
gender relations might be considerably different from that in the target societies of microfinance
. sschemeTable 4.5. Contribution decisions of women and men
IR2 IR4 IR8 GR4 ent Treatmyes-choiyes-choiceces os off men, first ro women, firstund round 50.0% 100.0% 66.7% 94.7% 70.8% 90.0% 96.2% 100.0%
yes-choiyes-choiceces os off men, overall women, ove rall 55.0% 78.4% 55.0% 80.5% 54.9% 78.8% 81.3% 77.0%

nd discussion aSummary4.5 grouWe introdup lendiceng contra an experimcts on enrepaymtal microfinaent pence grformaance. me to sepSmall loanarate the ims are given to gropact of esups sential chwho aararcte joeristiintly cs of
repoliable for repart an experimeyment. Incentnt to investigate the iives to repay nfluenare prce ovidedof those featu through the presrosp on strateegic dect of follow-up loafault of grouns. p We
members. Treatments involve different group sizes and a condition in which self-selected subject
groups register for the experiment together.
We observe high willingness to repay in all treatments, though game theory would recommend free-
rates ariding. Thnd arouregh this hig able to suhsta diegree of solidn the flow of furtherarity, the experimecredits for several pntal lenderioding groups. The willingnes reach high repss to ayment
periocontrids, whebute derclinese the ince as the experimntives to keep upent proceed the borrowis, butng it remain-relations reshipma diminish. rkably high, even in the later
We also examine the question of group size which is lively discussed in MFI policy. A dilemma arises
wheand mon the advre free-riantage of largding. Our results er groupshows through the in that the perfosurance effect irmance of the experims counteracteental d by less microcredit gsoliroupdarity s is
surpritendensingcy towardly robus shist with rerking,spec their supt to grouerior dispep size. While rsion ofthe large risk makes tr groups inhem pedeerform d manifest a higat least as hweer ll as
smaller groups in our parameter constellation.
borroOur results awer grorups. Ovee also roburall perfost againrmanst variatce is not sigions of nsociificanal ties between tly worse with stranthe megermbes thrs an with goof the experiod mental
acqexperimuaintanceent, but their behavs. Self-selected groiour is leupsss st exhibit a higheable, possibly becar willingnuess tose friends a contribrue less te in the beginwilling to tolerate ning of the
supposed free-riding by others.

44 Notice that in the first round, the individual subjects can be treated as statistically independent
s.observation 52


Our data exhibit a gender effect. Females show a significantly higher willingness to contribute than
(199males whi8)). This ch isis also intere in line with previousting as it sees findims tongs in related suppo experimert the commontaln prac games tice of M(Selten and OFI lenders to give ckenfels
worlloans predomd, this finding should inantly to wonmen. Hoot be geneweveralised pr, as gerionder r to future rerelationses are darchi. fferent in different parts of the

We beliof how aned why group leve that the experinding schemmental methoes dsu is especceed orciall fayil in practi well suited to gaining a dce. The present study provideeeper undserstanding a
framework in which the effects of central MFI features can be disentangled. Of course, the present
study shsimple, we haould be seve developeen asd a starting p a very basic oint rathermodel that na than a tucomprerally lacks hensive exploratiomany of the complexities ofn. To keep things real life
groustudep lendint subjngec contrat pool. Microctfs. Furtherminance instituore, outr experimions are imentpl has bemented all oeen convducteer the wod in our rld laboratory uwithin a greatsi variety ng our
of econwith different omic subjand cultural bect pools aseemckgrouns a pds, romsuising ch that the researreplich agcendation of a for the future. our findings in different societies


4.6 ReferencesLaw, EcoAbbink, Klaunomics, ans, Bernd Irlenbd Organizationusch, and Elke Ren, 18, 428-454. ner (2002): An Experimental Bribery Game, Journal of
JouArmendrnal of Deáriz de Aghion, Bevelopment Economiatriz (199cs, 60, 9): On the de79-104. sign of a credit agreement with per monitoring,
seleArmendction moáriz ddel, The Ecoe Aghion, Benomic atriz anJourd Chnalristian Go, 110, 632-6lli43er (2000. ): Peer group formation in an adverse
Armendáriz de Aghion, Beatriz and Jonathan Morduch (2000): Microfinance beyond group lending,
20. sition, 8, 401-4 Tranmics ofEconoBarr, Abigail, and Bill Kinsey (2002): Do Men Really have no Shame? Working paper, University of
Oxford. collateBesley, Timotral, Johurnal of Devey and Stephlan Coopment Econate (1995): Group omics, 46, 1-18. lending, repayment incentives and social
JourBohnet, Iris anal of Economind Brunc Behao S. Frey (1vior and o999rga): The sonizationund of, 38, 43-5 silen7c. e in prisoner's dilemma and dictator games,
DeveloBolnik, Brupmentce R. (1988, 16, 501-510. ): Evaluating Loan Collection Performance: An Indonesian Example, World
Brandts, JoStability in Experimrdi and Gary Charnental Gaemes, Exss (200perim0): Hot ental Ecovs. Conomics, 2, 22ld: Sequential7-238. Responses and Preference
BrandtExperimental s, Jordi, Arno Riedl Study, mimeo. Institut dand Frans van WindAnàlisi Ecoen (2nòmi0c02): Competita (CSIC), Barcelonive Rivalry ana. d Social Ties. An
Buckland the smey, Graalleholdeme (1r ag996ricult): Rural and aural credit adgministratiricultural credion, in: David Hulmt in Malawi: A study of the Malawi Me and Paul Mosley (Edudszi Fund ): Finance
dge. Routlerty, London: Against PoveprovisiCarpenteron of p, Jeublic gffrey P. (2002oods, Middleb): Punishing ury College Free-RideWorkinrs: how gg Paper Seroup siriesze affect No. 02s mu06. tual monitoring and the
Conlin, Michael (1999): Peer group micro-lending programs in Canada and the United States, Journal
of Development Economics, 60, 249-269.
Daweassumptios, Robyns about othen M., Jeanne r people'McTavis besh anhavior in a cod Harriet Shammkons dillee (197e7): mma Behavior, cosituation, Jourmmuninal of Persocation, and nality
y, 35, 1-11. gal Psycholoand SociProgramDevereux, Jos in Develohn and, Rauyping, Develomond EH. Ficonshe (199omies, 313): An Econo, 102-121. mic Analysis of Group Lending
Falk, Armin, Experimental Urs FischInvestigation,bacher, and Sim mimeo, Univon Gächter ersity of St. Gallen. (2002): Isolating Social Interaction Effects  An
Fehr, ErnAmerican Economist and Simon Gäc Reviewchter 90, 980(2000): -994. Cooperation and Punishment in Public Goods Experiments,
Behavior aGächter, Simnd organion and Ernst zatioFehr n, 39, 341-3(1999): 69. Collective action as a social exchange, Journal of Economic
seleGhatakction effe, Maitreesh (20ct. The Econ00): Screomic Joueninrgnal, by the comp 110, 601-631. any you keep: Joint liability lending and the peer
Ghatatheory and pkr, Maitreesh and Timactice, Journaol of Developthy W. Guinmnane (199ent Econom9): ics, 60The econo, 195mi-228. cs of lending with joint liability:
Gneesystematic bizy, Uri (1996ases): Proba, Acta Psycholobility judgmengica, 93, 59-68. ts in multi-stage problems: Experimental evidence of
over LoaGoetz, Anne Marie and Rin Use in Rural Credit Programna Sen Gupta (199mes in Ban6): Whgladeo Takesh, Wos Thrlde Cre Developmdit? Genender, Powet, 24, 45-63. r and Control


1894Guinn-19ane, Ti14, Explorationmothy W. (19s in Economic Hi94): A failed institutional story, 31, 38-61. transplant: Raiffeisen credit cooperatives in Ireland
Hulme, David (2000): Impact Assessment Methodologies for Microfinance: Theory, Experience and
better Practice, World Development, 28, 79-98.
Impavido, Gregorio (1998): Credit rationing, group lending and optimal group size, Annals of Public
and Cooperative Economics, 69, 243-260.
Isaac, R. Mark and James M. Walker (1988): Communication and free riding behavior: the voluntary
contribution mechanism, Economic Inquiry, 26, 585-608.
Isaagoods problec, R. Mark and m, in: LaboratJames ory ReM. Walker (1search991): Cost in Political Econly communiomy, Thocation: An experimas Palfrey (ed.), Ann Arbment in nesteor d public
86. 9-2ss, 26an PreUniversity of MichigIsaac, R. Mark, Kenneth F. McCue and Charles R. Plott (1985): Public goods provision in an
experimental environment, Journal of Public Economics, 26, 51-74.
Isaac, R. Mark, James M. Walker and Arlington W. Williams (1994): Group Size and the voluntary
provision of public goods, Journal of Public Economics, 54, 1-36.
Kabeer, Naila (2001): Conflicts Over Credit: Re-Evaluationg the Empowerment Potential of Loans to
Women in Rural Bangladesh, World Development, 29, 63-84.
KesePublic Gor, Clauoddia and Frans, Scandinavias van Winn Journal of den (2Econo000): Conditimics 102, 23-2onal Coope9. ration and Voluntary Contributions to
Ledgerwood, Joanna (1999): Microfinance Handbook. World Bank, Washington D.C.
Morduch, Jonathan (1999): The microfinance promise, Journal of Economics Literature, 37, 1569-
1614. ExperimentMosely, Paul s in Nepaand Rudra Prasal and Bangladd Daesh, Manchal (1987hester ): Credit for the RuPapers on Deral Poor: A Comvelopment, 3, 45-59. parison of Policy
Selten, Reinhard and Axel Ockenfels (1998): An experimental solidarity game, Journal of Economic
Behavior and Organization, 34, 517-539.
The AmeriStiglitz, Josecan Econoph E. and Andremic Revieww Weis, 71, 393-4s (198110. ): Credit Rationing in Markets with Imperfect Information,
Stiglitz, Joseph E. (1990): Peer monitoring and credit markets. World Bank Economic Review, 4, 351-
366. Individual EnSuleiman, Radowmmzi , Taly Weents oniss a the Provisionnd Gary Born of Publicstein Go (2002): Thods, mimee Enhano, Hebrew Univcing Effect of Unceersity Jerusalrtainetm y about
Varian, Hal R. (1990): Monitoring Agents with Other Agents, Journal of Institutional and Theoretical
46, 153-174.mics, 1EconoVan Dijk, Frans, Joep Sonnemans and Frans van Winden (2002): Social Ties in a Public good
Experiment. Journal of Public Economics, 85, 275-299.
Wenner, M. (1995): Group credit. A means to improve information transfer and loan repayment
performance. Journal of Development Studies, 32, 263-281.
Wydick, Bruce (1999): Can social Cohesion be harnessed to repair market failures? Evidence from
group lending in Guatemala, The Economic Journal, 109, 463-475.


45sPrice rigidity in customer market5

5.1 Introduction has been eloThe intuition that implicit contraquently develctoped by Arths betweur Oken buyers anun (1981), d sellewho coinrs aree dan impo the term customertant sourr mace of pricrkets for e rigidity
sellermarkets s may be rwith long-teeluctrm relaant to incrtionshieaseps b preicestwee in ren buyspers onse to temporand sellers. In cuary coststome shr mocks becarkets, he aause crghangesued, in
the terms of the implicit contract may antagonize customers and disrupt the customer relationship.
Okun also speculated that customers may be less antagonized if price increases can be justified by
increamarkets asere as in sellers co source of psts. ricHene rigice, acdity, but price ricording to this ingidity is less protuition, long-tnoerm relunced if priceationships in increacuses stocamer n be
justified by cost increases.
Despite the importance of price rigidity in macroeconomics and industrial organization, relatively little
wa1995 for s knsuown arveys). Thibout the empis lack of empirirical validity ocalf kno Okunwleds inge is tuitions usurprinsing sitil recently nce cu(see stomeCarrlton 198 markets ar9 or e Wynne of utmost
economic importance. According to Blinder et al. (1998, p.302), about 85 percent of all the goods and
servian ongcesoing relationship. T in the U.S. nonfarm businhe reasoesn whs sector ay relatively re soldlittle wa to regulas knr cuownstome is that rs most theoriwith whom es of psellerice rs have
stickiness rely on variables that are either unobservable by conventional methods or unobserved in
praccritical issue itice. Hennce, the a the microfobundationsject failure of the standa of macroecordnomics (Blin research methodder et al., p. 3) motivated reseology to make headway onarche this rs in
recent years to use unorthodox approaches to investigate the issue of price rigidity in customer
markets. Empirical evidence on the relevance of Okuns intuitions comes from questionnaire and experimental
studies. In a much-cited questionnaire study, Kahneman et al. (1986) have investigated fairness
perceptiobe more accens of price inptable if they can bcreasees in the ge justifieneral pd by cost inubliccreases. Howe, and find that peoplver,e percei their findivngs ale price inso cresugases to gest that
relationprice rigishipdity doe. In a comps not depellinend on g interview whether buyestudy Blinder rs andet sellers have al. evaluate motives for pa one-timer or a long-teicing decirm sions from
the perspective of the sellers. Systematic interviews with managers who are responsible for pricing
decisionot to antagons shownize thei that long-ter customers rm relationshiby their prips arcing de coecimmsionon and imps. The auortthors find ant, and that managsupport for the ners are careful otion
increathat manageses. Much to their rs believe cusurpstomers to be lrise, they do not fiess antagoniznd that price ed wherigiditn priyce incre is related toases can b long-term e justified by cost
relationships. In a similar interview study with Swedish firms, Apel et al. (2001) find that implicit
contracts are the most important explanation for price rigidity.
interaExperimental ction caevidenn be obce uservesedf in controlleully complemd envirents findionmenntgs from s. Cason anquestiond Friedmnaire studiean (2002s becau) study pse marice rket
rigidity in an opaque market in which buyers develop customer relationships to save on search costs.
The authoin turn increasesrs show that hig the sellhers' maer search corket power and sts tend to indinduces couce cunsistomederablre price rigis to remaidity. However, tn with their sellehe r, which
authors do nbe justified byot cost in investigate tcreases. The relevhe hypothesis that priance of thisce ri hypothesigidity is les is iss pronnvestigateouncedd if price in in a series of tcreahree ses can
closely related papers (Kachelmeier et al. 1991a, 1991b, Franciosi et al. 1995). These papers find
Howeak anwever, thed nosn-pee stursistedies dont price not analyze rigidity when pricustomecre in markets bcreases ecancausnote long-te be justified brm relaytionship cost increas are ses.
excluded by design in these experiments. It is important to note that all experimental studies cited
45Behavior & O This chapterrgani is bazatiosed onn the pap, Vol. 55, 2004, 575-593,er Price rigi jointdity in custome work with r maJean-Robrkets, Journal ert Tyran. of Economic


above analyze markets for inspection goods. In such markets, no quality uncertainty prevails.
Our experimental study is the first to investigate long-term relationships on markets for experience
goods as an explanation of price rigidity. For experience goods, quality uncertainty prevails in the
sense that buyers learn the quality of the good only after the purchase. We analyze markets for
experience goods because our focus is on how trust in customer relationships generates price rigidity.
seaWe belirch coests inve that trust is an i an opaque mportamarket. In fact, Okun (p. 14nt factor in shaping custome1) noted that r markets in athe extent to whidditionc to saving oh firms arn e likely
to enjoy repeat patronage depends both on the satisfaction of customers with previous purchases and
their confidence that the supplier will maintain good performance. Hence, long-term customer
relationships may be upheld if buyers trust sellers to provide high quality and if sellers are trustworthy.
We compare behaviour in two experimental markets that differ with respect to whether buyers and
selleoutcomrs es can dcannot be puevelop longbli-tcerm relly observed ationships. Boand in both mth maarkrketets qus are opaality uncerque in the setainty prevails. Hinse that magh rket quality is
more valuable to buyers, but its provision is more costly to sellers. In the customer market, buyers
can trade repeatedly with the same seller and can thus develop long-term trading relationships. In
contrast, this is not possible on the anonymous market where buyers and sellers do not know with
Cason and Gwhom they trade. In this mangadharan a2002 for exprket, econoermic theoimental evidry predictencs the leme). That is, low quality is providon outcome (Akerlof 197ed at low 0, see
prices because sellers cannot build a relation-specific reputation by providing high quality. In contrast,
in the customer market, upholding customer relationships by providing high quality at a reasonable
price may also be profitable to sellers if it induces customers to return to this seller.
To analyze whether customer relationships are upheld and whether they partially resolve the lemons
problem, we compare the anonymous market and the customer market. To analyze whether the
customer market causes price rigidity, we compare price adjustment in response to a cost shock in the
two markets. To analyze the cost justification hypothesis of price rigidity, we compare price
coadjusst shoctment in two cuk. In one treatmestomer marketnt, buyers ares that exclusivel informed aby doiuffer with ret the cost shspect to the information aock (i.e., information is puboutblic), the
and in the other treatment they are not (i.e., information is private to sellers). Hence, with public
information a price increase is justified by a cost increase, while with private information it is not.
Our main results are as follows. First, we find that long-term customer relationships are frequently
Highupheler qud and thality on the cuat average qstoumaer mality is about three timerket comes at a higs hiher pgher rwhen cuice, but upholstomerding the relationshi relationps arship ie pos ssible.
profitable for both sellers and buyers on average. Hence, we find that customer relationships partially
resolve the lemons problem. Second, we show that customers tend to penalize sellers for price
increases and, in particular, for providing low quality by terminating the customer relationship. Third,
we show that price rigidity is more pronounced if the price increase cannot be justified by a cost
increase. In this case, sellers increase prices much less and bear considerable losses to avoid
antagonizing customers. In all, we show that customer markets mitigate the lemons problem but are
prone to price stickiness. Since long-term customer relationships are upheld in our customer market
because buyers trust sellers to provide high quality, this paper also demonstrates how trust can shape
mes. market outcoWe proceed as follows. Section 2 explains the advantages of an experimental approach to price
rigidity in customer markets and relates this study to the literature. Section 3 explains the experimental
design, section 4 presents the experimental results, and section 5 concludes.

customer markets inAn experimental approach to price rigidity5.2 We test three hypotheses about to how trust shapes customer markets. Our first hypothesis H1 is that
cusuggestomesrted in markets the seminreal arsolve the lemticle by Akeons prlrobleof (see malso A. This rrohypothew 19sis h73). Tas already cahat is, we testsually be whetheen r buyers
trust sellers to deliver higher quality in a customer market than in an anonymous market.46 We expect
46as well as in t Note that the mechahe studies nism that leadcited in this section s to the "lemonis of the moral hs oautcome" of lozard type, while the mechw quality at low priceanissm in ou r paper
described by Aero is of the adverse selection type.

relationthis hypotheships siendos to hold begenouslycause two pre form in markets vious exwith pequalrimental studiity uncertaintesy have foun and that they improve mad that long-term rket
performance. Kollock (1994) obtains these findings by comparing markets with and without quality
uncertainty and private vs. public offers. Brown et al. (2004) investigate a gift-exchange labor market
in which buyers (firms) choose prices and sellers (workers) choose the quality (effort). They compare
two markets that differ with respect to the anonymity of market participants and find that long-term
benefirelationcshiial. Hops are upwever, both held in studies anon-anonymonalyze stable us marketens avironmentnd that these s and relationdo, therefoshirps e,are mutu not investigaally te
whether long-term relationships cause price rigidity.
To test the lemons hypothesis, we compare market outcomes on the customer market and on the
chaanoracnymouterizesd market. Note by Okun. On auction ma that our anorkets homonymous magenous grket does noodot correspons are traded, andd to an au trading is oction mrgaanized rket as
as in exchanges for stocks or raw materials. Such markets are most closely approximated by
experimental double auction markets. In contrast, we compare two posted-offer markets in which
sellers post prices and in which goods are non-homogenous. We do not compare our customer
market to a double auction market because in order to isolate the role of long-term relationships, we
need to compare two settings that are otherwise identical (ceteris paribus variation).
increaThe sesecos on nd hythe cupothesistomesr we test is the in market tend to undevisible harmine trust anndshake hypothd disruepsit the custos H2. We analyze whmer relationshiether pp. rice
Because quality may disruwe inpvestigate mt the custarkets omer relationwith quality unceship. For rtaexaminty, both increasiple, Okun (p. 154) ng onsugges esprice and pted that orovnce the iding low
firm draws a clientele with attractive implicit contracts, any deviation unfavourable to customers is
seen as a violation of these contracts. Hence, our design also serves to investigate the relative
importance of price and quality changes.
The third proposition we test is the cost justification hypothesis H3. We investigate whether
customers are more willing to tolerate price increases if these are justified by cost increases. To this
shocend, we k with a compare psituation ricwhee adjn they are notustment on the cu informed astomerb mouat the shorket whenck. buHenyers acre, we doe informed a not comparbout the coe a cost st
and a demand shock to investigate the validity of the cost justification hypothesis. Okun suggested
that prices are more flexible after a cost shock than after an increase in demand because price
increademansed shs areocks involve perceived as a sloss of experi more fair by mcustomeental cors ntin rol becathe formeurse co case. Host and demanwevedr, comp shocariks may ng cost and differ in
several dimensions. Instead, our approach of implementing a controlled variation of information
conditions tests hypothesis H3 more directly.
Suppothe understartive evidennding that price for H3ces shoul is provided red by Blimain fixed hold whender et al. (p. 157). Thn your costs ine authors acresk managease, or do customrs: Doeers ses e
price increases as justified when costs increase? The responses (n = 111) were as follows: When
costs inmixed (13.5%cre), c) tolease, our customrate ers nprice increormases (7ally a) still wa1.2%).nt  us to hold ouNote that thesr prices e finding(15.3%s reflect manag), b) attitudes ers beliefare s
beliefs traabout customnslers prate into actuaice tolelran pricince. Hog dewevcisions oer, we r cannfirm profits. For exaot infer from thmese responple, are masenages hors w manawillingg to ers
antagonize 15 or maybe 30 percent of their customers to increase profits on the remaining customers?
While questionnaire studies are particularly useful to obtain information about decision-makers beliefs
and attitudes, they are inapt to study market interaction and, ultimately, price rigidity. In contrast, our
experiment analyzes market interaction under controlled conditions, but motives of market participants
must be inferred from observed behaviour.

design 5.3 Experimental and paSection 5.3.1 rameters in dprovides a geetail. Section 5.3.3 neral descriptpresion of tents the predihe design, and sectictions. on 5.3.2 explains the procedures

tal design e experimenGeneral description of th5.3.1 In our expemarket (AM). riment, two maIn both marketrkets opes, an experierate nce gsimultanood ieous sly: a custometraded. That is,r buyers o market (CMn) and an aly learn the qnonyualitmous y of


the good after the purchpossibility to become a repeat customer. In the ase. The essential difference betweCM, buyers and selen the two malers crkets consian trade with the samsts in the e
partnerelationr reshippea with this specific selltedly. Buyers receive an offer from a per. In contrasartt, the AM is anonymicular seller and decide ous. In the AM, buyers fawhether to upholce offers d their
customefrom a numbr relaer of untionshiknop with wn sellea specific rs, ansellder. due to the anonymity of the market, they cannot establish a
In the CM one buyer and one seller are matched at the beginning of a trading phase. In each period
the seller posts a price to his buyer and chooses the quality he provides in case the buyer accepts. If
same the buyer aseller icncepts his sell the next period. Thus, theers price offer, customhe learns the er relationship is uquality of the good apheld as longnd is re-mat as the buyer ched with the
continues to accept his sellers offers. However, if the buyer rejects an offer from his seller, the
customerAM from this perio relationshid on. According to thip is terminated. As as pro concsedequruene, the cuce, both the bustomer relationship yer and the seller pastarts with an rticipate in the
exogenwith an exogouenous match but is endoges match is that the focus of ounously upheld by custr studomy is on trust within the relatiers. The reason why we start the relonship, not oan tionship
consumer search.
offers aIn the AM all nd casellen chors ose frosimultmaneously po these offers in ranst prices dom ordeto all buyers. Buyers r. If a buyer accesee a list opts anf anonymou offer, neither the seller s price
nor the buyer can identify the trading partner. Hence, trading in the AM is completely anonymous. All
phabuyers and se. A conseqsellers whuence of this proceduo are in the AM in a particulre is that markear petriod participaremainnts there focan only switr the rest of the tradch from the ing CM to the
AM, but not from the AM to the CM. This procedure implements a "grim trigger" strategy by customers
and enforces an "unforgiving" behaviour that is empirically plausible (Engle-Warnick and Slonim
(200trust gam3) showes). This g that this straterim trigger stgy is chosrategy penrovide by about s sell80 peers in our expercent of the first moveriment with strong inrs incentive infinitely repeated s not to
antagonize customers since customers cannot return to the seller at a later point.
(seeTo investigat table 5.1). In the treatment labellee the effect of information adbout the PRIVATE, coonst shock, we imly the sellers are informplement two tedreatme about the cont condst shoitions ck,
while in the treatment called PUBLIC both sellers and buyers are informed about it.
Table 5.1: Overview over experimental conditions
Information PRIVATE PUBLIabout cost shock is C
Customer market (CM) CMPrivate CMPublic
Anonymous market (AM) AMPrivate AMPublic

5.3.2 Procedures, parameters and information conditions
Each experimental session has 5 trading phases with 10 periods each. In all sessions, 8 sellers and 4
buyers participate. In each period a seller can at most sell one unit of the good and a buyer can buy at
trading pmost one uairs nit. At starting othe beginnning of ea the CM. All other pach tradingrticip phaantse s star3 sellers and t on the AM. As a con3 buyers are randsequence, thomly matcheere are at d as
customeleast five sellr relaers ationshind one bp in the CM. This uyer on the AMconstella plus all titraon implies that thding pairs where is alo at some poiways an excent have termiss supply on nated the
od. the AM of 4 units of the goParticipexperimants ent at a commocan earn pointnly knos by trwn exading. chaThese pointnge rate. Participas are convents payorted ffs are Price into money at the end of the  Seller Costs for a
seller and Buyer Value  Price for a buyer if a trade is concluded, and zero otherwise.
Tabelle 5.2: Buyer values and seller costs
buyer values seller costs surplus
High quality 200 80 (t = 3: 120) 120 (80)
Low quality 100 40 (t = 3: 60) 60 (40)


Table 5.2 shobuyers, but providing high ws the parameters of bquality is alsou twice ayer values ans cod stly to tseller costhe sellers as. His providing low qugh quality is twice as valuable for ality.
Therefore, thquality differencese total surpl directly tranus of a high-quslate into efficienality trade cy differenis twice the suces, we rpluwill cos of a low-qncentrate on repouality trade. Since rting quality
Hendifferencce, costes bes increalow. In perise fromo period 2 to 3 d t = 3 of each trading pfrom 80 to 12hase, a tempo0 for a highrar-qy cost shouality good, ack of 50 percent nd from 40 tooccurs. 60 for
a low-quality good. Note that costs fall back to the previous level from period 4 on.
The information conditions are as follows: In both treatments the payoff calculation, the number of
trading pexactly how hasesmany selle and periodrs and buyers, and the buyer values there arse on th are comme AM and theon information. The subje CM, but they are informcts do nedo that t know
there are always more sellers than buyers in the AM. The exact seller costs given in table 5.2 are only
knohigh quwn to the selleality and that the cors, but thsts are buyers kne the saow that selleme for all sellr costs aers. re lower for delivering low quality than for
condThe two treitions, seallers atments differ ere infoxrmed aboclusively witut the coh respest shoct cto the informk at the beginniationg of perion about the cost shd 3. Buyers are informeock. In both d
about the cost shock in PUBLIC but are not informed about the shock in PRIVATE, and this condition
iskno knw a) thatown to sellers costs incre. In particaseu by 50 percelar, the information nt, b) that tsellehis cost srs obtain in Phock isR temporarIVATE has three ay (they knospwe tcthat costs. Sellers s
PUBLIC, a) awill fall back to the previound b) hols level in perid for both buyers and sellers, od 4), and c)a nd sellthat buyers aers knre not informeow that buyers are infod about a) armendd about b). In
particia) and bpant). Nots aree inform that our obed aservationbout it at the s frobeginm perioning of d 3 are responperiod 3. ses to the cost shock because
At the beginning of each session the participants are randomly assigned to one of the booths in the
compexplain the euterizedxperime laboratorntal pry (we use the ocedures in software z-tree,detail are han Fischbaded out.47cher All subje1999c). Writtets read in instrunstructionctionss fo which r both
roleRoles are ras. There are no ndomly assigncontrol questioed in the first ns, bperiout subjed cand ts caken ask quept throughstiooutns befo the entire sere the ssiexperiment ston. The written arts.
compinstruuter ctionsscreen contain no infos from permation oriod 1 on. Thn the sellee informatiors con on the cost parameters, whist shock is ch are showonln on a sey displayed oparatne screen sellers'
that appears at the beginning of period 3 (see Appendix 5.B). In PRIVATE, the screen explicitly states
information athat the information abobout the cout the st shocshock. In PUBLIC, tk is shown to sellehe screen is rs only shoanwn to botd that buyers doh buyers and not have any sellers, and the
information about the cost shock is announced aloud by the experimenter.
At the end of each period, subjects in both treatment conditions only get information about the
outcomthe CM noes r in(i.e., price, qualit the AM know about markety, and own outcomeprofit) of their own trans (such assa averaction. ge quality, That is, neitransather paction prticipranices, etc.) ts in
outcomes isin the AM. Hence, both ma provided to market partirkets are opcipaqants. ue in the sense that no information about overall market

5.3.3 Predictions In the AM, the theory of competitive markets predicts the lemons outcome, according to which low
quality is provided at a low price. In particular, because of the anonymity of market interaction in the
AM, a rational and self-interested seller always provides low quality. Because of the excess supply (of
4 units in each period) in the AM, prices should be driven down to the cost of providing low quality, for
AM for three reasonexample, to 40 units (see ts. First,able 5.2). Ho the AM is opaque inwever, we the expect convergsense that market partience to be slcipantow ans only get d imperfect in the
information about their own transaction, but not about market outcomes. This lack of market feedback
information is expected to lead to slow convergence (see Cason and Friedman for an extended
discussion). Second, sellers in our experiment know buyer values, and fairness considerations may
prevent prices from falling to costs. In particular, sellers may be reluctant to undercut each other
knowing that the entire surplus goes to the buyers if prices equal costs. Third, buyers may trust sellers
despite the anonymity of market interaction. It is known from many experimental studies on trust
47 Instructions are available from the authors upon request.


games that people have a tendency to trust others even in perfectly anonymous one-shot interactions
(e.g., Berg et al. 1995). Two features of the CM contribute to the persistence of long-term customer relationships. First,
do not knmarkets arow e oqualitiepaqus and e. Participrices ppants in the rovided on the AMCM only get . Tfeedbhereforack abe, bouuyers ht their own ave insufficiemarket trannsat information ction, but
to determine rationally whether to remain with their seller or to terminate the customer relationship to
expeexperimrimentent may induce cu (i.e., to uncover pricstomers to rees and qualitiemain with their s on the AM). Hoseller: the terminatiowever, a partin of a relationcular rule of ouship r cannot
canbe revonot retuked. Trnhis rul to the CM upe makes expeon finding orimeut that tntation potenhey earn less on the tially more expensive for bAM than on the CM. Houyers becaweusever, they
buyers knmarkets. Henow ce, we that the exwoulperiment has 5 pd expect customer relhases in ationwhich participantships to be terminated at a highes can gather experienr rate in ce early with both
ds. n late perios and isephaselleThe sers to pcorovnd feature ide high qucontaributing to the persility. Suppose a buyer stentruce ofs long-tets a seller to prm relrationshipovide high qsua is that buyers mlity and, upoan y trust
receiving high quality at a fair price, continues to trust the seller. Suppose further that the buyer
the selleterminates thr is tempted in eae relationshipch pe upon receiviriod to reap a onng bad qe-tiualitmey or upon exp gain by proverienciniding low qug a price increaality, but has to trade se. Then,
the selleoff this one-tir beme gain agcause the market is opainst the coaqust ofe, losing hi it may be prs customerofitable for the selle. While this cor to prost is novt knoide high quwn exaactly to lity if the
relationfor a more deshiptailed acco can be maintaiunned for t). A similar reasoninsufficiently many g appperiodlies to pris with csue choices after tfficient probabhe cost shility (see Broock. wn et al.
Accoantagordinizing ngly, a selleand losing his custor may find it profitablmer. e to refrain from increasing his price after the shock to avert
The remaining potential length of the customer relationship may be an important determinant of
early peterminatioriods n for two rebecauasonse the risk of not beis. First, antagonizingng able to tra the cude on the AMstomer on the is higheCM is mor in ere arly periodscostly for the seller in . Second,
if the customthe CM are foregoer terminne (seates e James 2the relation002 foshr a more ip in an eardetaly perioiled analyd, more speriois). These tds of pwroo fafitable interactors lead us to ction on
since termiexpect that customenation incer relantives are otionshipsppo tend site for buyerto be terminats and selled at a higheers, we have no cler rate in late periodar presdictio. Howen for the ver,
rate at which relationships end.


5.4 Results We conducted 12 runs in each treatment at the University of Erfurt. In total, 288 undergraduate
students from various disciplines participated in our experiment. Subjects earned  15.3 (US$ 15,
approx.) on average within about 120 minutes.
Table 5.3 summarizes important differences between the CM and the AM. The table shows the quality
and prices averaged over all subjects, phases and periods. As can be seen in the table, high quality is
providepercent of thed abo transactions ut three times as often in the CM ain the CM, but only s in the AM. In particulain 20 percent in the AM. We henr, high quceforth code hiality is providgh ed in 59
average qquality as 1 aualnd low qity. Higher quality as 0, allowiuality in the CM comeng us to sps atea a higher pk of the percenrictage of high qe. In fact, the average priuality providece isd 45 as
percent higher in the CM (115.1) than in the AM (79.9).
Since tramarket yields conding high qusiderableality fully translate efficiency gains. In s into higheradditio gains fromn, both buyers and trade (sesellee table 5rs .earn m2), the cuore onstomer the CM.
profits aIn particulrae 47.r, averag8 on the CM, e buyer earningbut only 12.2 on the AM. s are 43.9 on the CM and 3Hence, uphol9ding the cu.7 on the AM, and averastomer relationge seller ship is
able on average. tmutually profiTable 5.3: Overview (averages, all periods, all phases, both treatments)
Customer Market Anonymous Market
) (AM) (CMTranTradsaed Quction Price ality 0.59 115.1 0.20 79.9
39.7 43.9 gs nBuyer Earni12.2 47.8 Seller Profits

observations Statistical testing for differefrom the two ncemarkets as betrweee nno the CMt strictly andindependent. For the AM is problematiexample, pric. Thce reason ises in the AM might that
be affected by the experience of subjects who were exogenously allocated to the CM but failed to
upholbetween the d the relaCM and the Ationship. As a conM by consesqueidenrce, seling the stectiorictn lyprobl indepenems predenvt observatioail. Thus, wens of the first p test for differences eriod in
(average 1the first pha03.8) againse. We test trasnt those sawho ction prices of were exogenouslthose who wey allocated tore exogenou the AM (average 85.7sly allocated to the CM ). A Mann-
Comparing qWhitney test reveals that transactiouality choices analogoun pricessly across the are indeeCM (0.44d highe) and r in the CM (p = 0.0the AM (0.22) yields a we10, one-tailaedkly test).
highesignifir (p cant differen= 0.000, Mannce (p < 0.-Whitney test) in100, Chi-square test, the first perioχ2 = 3.24, dfd in the CM (45.9) than in t = 1). Seller profits are sihe AM (9.2). gnificantly Buyer
earnings are also higher in the CM (40.9) than in the AM (32.3) in period 1 (p = 0.000, Mann-Whitney
). ttesWe conbetter macludrkete outcome that while thes. Average qualit AM did not fully convy on the CM iserge much higher to the lemons outcome, th, and tradinge CM produ on the CM is mutually ced much
the lemonprofitable fos problem, as r buyers and suggesellers. Hensted by hce,y long-tepothesisrm rel H1. ationships in the CM at least partially resolved


A) Pricing in customer markets We now discuss how trust in customer markets affects prices, and, in particular, how price responses
of custdiffer acroomer ss trrelatieatmeonships. nts. Subsection B) considers quality, C) profits, and D) analyzes the determinants


011cige praervae90

70Customer market (CMPublic)
Anonymous market (AMPublic)
Figure 1: Average transaction prices with public information (PUBLIC)
Figure 1 shows average transaction prices per period in the two markets with public information
(PUBLIC). To facilitate the exposition, we aggregate prices over all 5 phases (see Appendix 5.C for
higheaveragr in CMe prices in eaPublic than inch phase). Seve AMPublic. Secoral facts and, prire evceids do responent from inspd to the cost shection of figure 1. First, priock in both AMPublic ces are
CMPubliand in CMPcu. Increasing pblic. Third, averagrices in CMPe priceubls ic reare result frlativeom ly a selestable in AMction bias. The buPublic, but tend to increayers with the lose in west
tolerance foquality) tend to drop or high prut of the CM in eaices (i.e., those with the lowerly periods. In addist degree of trution, highst that the seler quality is stler will prongly arsovidsocie high ated
As a cowith highnseer pquricesence, cu, and prostomver reliding low quationships exhibitiality tends tong low pri disrupt the cuces anstomed low qur relationshiality tend to be dip (see sesrupted ction D).
t. firsFigure 2 shAs with publiocws aver information, age transactioaveragne pri pricceess are hi in the twgheor in markets whe CMPrivate tn informatiohan in AMPrivn is private (Pate, and prices have RIVATE).
the coa positive trest shock in perind in CMPrivod 3 inate that is ab AMPrivate, but averagsent in AMPrie pvricesate. We agai do not seem to ren observe a psporonounnd to the coced rest shoaction to ck
becauin price response However, the price res are obscursped onseby the upwars are difficult to assed trend in CMss from mere in that resultspecs from the seletion of thcte figureion biass .




110ci prageerave90


Customer market (CMPrivate)
Anonymous market (AMPrivate)
Figure 2: Average transaction prices with private information (PRIVATE)
The simplest way to correct for this selection bias is to consider price changes. We calculate how
individual sellers' transaction prices responded to the cost shock (i.e., changed from period 2 to 3).
Note that we correct for the bias by considering only transaction prices of those sellers in period 2 who
also transacted in period 3.
Table 5.4: Price adjustment in response to cost shock
Customer market Anonymous market
(4) (5) (6) (1) (2) (3) CMPrivate CMPublic Effect of AMPrivateAMPublic Effect of
information information on AM on CM all phases 4.15 16.40 0.000 20.72 19.61 0.954
phase 1 4.40 23.65 0.013 23.03 29.21 0.507
phase 2 -0.79 12.39 0.028 16.75 31.31 0.225
phase 3 0.71 13.19 0.043 23.78 13.94 0.149
17.00 17.56 0.773 10.63 14.56 0.032 4 se phaphase 5 4.92 17.87 0.002 14.54 14.87 0.954
Notes: Columns (1), (2), (4) and (5) show the average change of transaction prices from period 2 to
period 3. Columns (3) and (6) test for the effect of information on price adjustment. They show p-
values of Mann-Whitney tests (one-tailed for CM, two-tailed for AM). All CM values are changes of
transaction prices for a given customer relationship.
To evaluate the "cost justification" hypothesis H3, we compare price responses in the customer
markets across information conditions. At the individual level, we find that 79.6 percent of sellers
increase their prices in CMPublic, while only 40.3 percent of sellers do in CMPrivate. When
aggregating over all 5 phases, we find that the average price response to the cost shock is almost 4
times more vigorous with public information (16.4, see column 2 of table 5.4) than with private
information (4.2) in the CM. Mann-Whitney tests show that prices respond significantly more in
CMPublic than in CMPrivate. This is true when aggregating over all 5 phases (p = 0.000), and it also
holds at the 5 percent level for each individual phase (see column 3). Note that in contrast to the
customer market, information about the cost shock has no effect on the anonymous market (see
lumn 6). oc




If we analyze the responses of posted prices (rather than accepted prices as above) across
information conditions in the customer market, we find that price offers on average increase by 7.5 in
CMPrivate and by 17.1 in CMPublic. This difference across information conditions is highly significant
(p = 0.000, Mann-Whitney), suggesting that sellers anticipated being turned down by customers if they
increase prices too much without justification (i.e., in CMPrivate).
In all, we conclude that customer markets cause pronounced price rigidity if price increases cannot be
justified by cost increases, but that price rigidity is much less pronounced if the price increase is
justified by a cost increase, as suggested by hypothesis H3.
B) Quality Figure 3 shows the percentage of transactions in which high quality is provided (average quality) for
both markets and information conditions. Average quality is higher in the CM than in the AM from
period 1 on. While quality is relatively stable at low values in the AM, quality further increases in the
CM to its peak in period 9, and then sharply falls in period 10. The increase in quality until period 9
reflects that higher quality tends to be associated with higher prices in the CM. Again, this increasing
trend can be explained by a selection bias. Since providing low quality tends to disrupt the customer
relationship, customer relationships with high quality tend to be upheld at a higher rate (see section D
for details). CMPublicAMPublic
1CMPrivateAMPrivate qugeaeravy
uality Figure 3: Average q








The sharp drop in period 10 is an endgame effect that is frequently observed in finite trust (and
publihigh quc goods) games ality in the CM for strategi(e.g., Selten and Stoeckec reasons. Ser 198llers se6). This sharp drop suem to have understoogged thasts tthat selle providing lors prw ovide
table 5.3). Hoquality tends to disruwever, pprovidt the cuing high qstomeruali relattionshiy ceasesp and t to hat upholbe profitabldinge for sell the relationers in the last peship is profitabriod since it le (see
canremanot indurkable thce customerat finite long-term rel loyalty (i.e., indationuships ace a brue apt yer to upholto induce syd the relationstematically higship byher qu definition). It is ality in the CM
at most last for 10 pethan in the AM up to perioriods. Accod 9. Rememrdinbgly, depeer that all parnding otnicipa players' rationts knew that the custnality and experieomer relationnce, oneship would can
expect quality to fall at some point before the end of the game (see Ch. 5 in Camerer 2003). However,
the refinite trust gasults of me (5 peEngle-Wariodrnic) and an eqk and Slonim suguivalent infigesnite trust gat that players behme. Tavhe authors finde relatively similarly in a si that a lot of mple
experiefinite and the infinite game.nce is needed (more than 20 re We only reppeeat the game titions) to yield cle5 times (par behahases) avioural diffend find norences b signifietwecant en the
differences in buyer trust (breakup) or seller trustworthiness (quality) over phases.
informatioFigure 3 seen. Homs to suwever, thegge declinst that the cost shoe in quality is not sick deprgensses quality more ificantly different in the AM across infowith public than with private rmation
conditions (Mann-Whitney test, p = 0.323). If we control for the selection bias in the CM analogously
as for p3), we find thrices e followin(i.e., by considg. Average qering houalitw quay falls bylity within a cu 12.5 percentagstomer relae potionshiints (from 75p changes fro% to 62.5%) in m period 2 to
CMPrivate and by 17.4 percentage points (from 59.1% to 41.8%) in CMPublic. However, these
changes are not significantly different (p > 0.3, Chi-square test, χ2 = 1.84, df = 2). We conclude that
the information about the cost shock had no significant effect on quality.
s Earnings andC) ProfitUph5.3). Hoolding thwevee customerr, the gains from trustin relationship is pg onr the CM ofitable for bare uneqoth buyers anually distd selleributed. Sellers on avrs beerage nefit mu(see table ch more
from uphthe CM than ioldinng the relation the AM (47.ship tha8 vs. 12.2), bn buyuers. In pat buyers ortnicul averagar, sellers eae only earn 1rn almo0.2 percent most four times as mure in ch in CM than
in AM (43.9 vs. 39.8). This implies that buyers can sanction sellers by terminating the relationship and
low quthat sellers hality. Tahe averave a strong incege sanctionntive not to antago of terminating the renize customelationsrhs byip for a s increeasiller isng prices 35.6 per period (= or by providing 47.8
prices a- 12.2) are lond rewer onsults from two sou the AM than on therces. Fi CM. Seconrst, trading don , the AM is less profitabland more importantly, there is a high rie for sellers because sk of not
being able to trade on the AM, which means that seller earnings are zero.
An interesting question is who bears the burden of the cost shock in the customer market. Of course,
(sethe coe table 5.2st shoc)k depresses . Since the relationoverall eaship is rnings siprofitable fonce the gainr selles from trars, they should bede fall by a third, cete willing to bear thris paribue burds en
of the cost shextent to which it does deock if doing so contripends on the ibutesnformatio to uphold n the relationavailable to custship. Accoomers. Herding to hypothence, we wosis uld exH3, the pect
that sellers bear more of the burden when information about the shock is private than when it is public.
becauThis is inse theydeed the ca are reluctant to increse. With private informatioase pricesn, , and thethe selleyrs b reduce qualear the brunt of the coity only a little. In particulast increase r, seller
losses(46.3 vs. 18.9 in period 3 (a percent) os a penrce the CM when intage of pneriod 2 pformatiorofitn is s) are moprivate. In core than twice as lantrast, with prge aublic infos buyermatr losseion, s
buyers on the CM suffer losses that are about five times as large as seller losses. In particular, with
public information buyers lose 62.3 percent of their earnings, while sellers only lose 12.6 percent of
their profits (again relative to period 2). Hence, sellers are able to shift the burden of the cost shock to
customers with public information but not with private information.
earniInterestinnggly, there as in the CM. If we test sellre no pronouner pcerofd long-teits acrorm effess treacts of the cotments by period st shock on (Manselln-Whitney tester profits or customes), we only r
earnifind a signifingcant differens are only different in perioce in period 3 ds 3 (p(p = 0.010 = 0.002); a) anll other ped 6 (p = 0.0rio1ds are in1) at the 5 pesignificarcennt (p > 0.48t level. 7). Buyer
D) Customer relationships
In total, 360 trading pairs start on the CM (= 12 runs x 3 pairs x 2 treatments x 5 phases). These pairs
can uphold their relationship for at most 10 periods. The average breakup rate is similar in both

treatmenta maximum in the shos (1c6.4% in CMPrivate, k period (34.5% 17.2% in CMPubliin CMPrivate,c). The differen 20.3% in CMce between thPublic). In the last perioe two treatmed, the nts is at
breakup rate is remarkably high (24.3% in CMPrivate, 36.6% in CMPublic). As a consequence, still
47.2 percent of all possible customer relationships are upheld in period 3, while this percentage drops
od 10. to 15.0 in peri(see table 5.5To analyze the determi) so that the enantss of cutimatestomes for dru relationmmyship variabls, we run a maes can be interpreted aximum likelihs the chood probit reange in gression
probability of a breakup. In particular, DF/dx is the marginal effect for a discrete change of the dummy
variable from 0 to 1. Table 5.5: Probit regression
cudepestomendent Vr relaariabltionshie: prop bability of breakup of dF/dx P > l zl
(1) Treatment (dummy, 1 = PUBLIC) .0079 0.731
(3) Period (2) Phase -.0121.0093 * 0.218 0.024
(4) Last (5) Shockpeperriod iod (d(dummyummy) ) .1556*** .3477*** 0.000 0.000
(6) (7) This peShockperioriods posted * public info d price (dummy) -.1177.0030*** ** 0.000 0.001
(8) (9) Price inExtent of price chacrease from last penge from last peririod (dummy) od -.0000 .0702* 0.998 0.020
(10) Large (> 30 units) price increase (dummy) -.1104* 0.010
(11) Large (> 30 units) price cut (dummy) .0624 0.403
(12) Last period bad quality (dummy) .3986*** 0.000
Notes: n = 1346, Prob > Chi2 = 0.0000, Log likelihood = -540.32, Pseudo R2 = 0.1675;
*** = significant at the 0.1 percent level, ** = 1 percent, * = 5 percent.
The insignificant estimate for (1) Treatment shows that the probability to terminate the relationship is
not different with public or with private information, which is what we expect in all non-shock periods.
Phase (2) is insignificant, and Period (3) is weakly negative. This suggests that relationships tend to
break-off early in a phase, but the tendency to uphold relationships is not different across phases.
we would hBoth of theseave expecte findings are d customesurprisinrg relationship. Given our dis to bescussi more pon of serevalenarch in opt in late phaaqusee markets in s when paserticipantction 3.3, s
have gained experience with the market institutions. The (weakly significant) negative estimate for (3)
periois surpds risi(where the cong since firms shst of terminould be moating thre hesitae relationnshipt to provide bad q is lower). Houawever, thislity in early period seemss than in only to hold for late
terminated is the very last period. The estimate 34.77 points higher in the last perifor Last period (4) indicaod than in other petes that riods, suggthe probabilitesting that some y that a relationship is
10 (as mocustomers anticipated that st of them do, see figure 3)sellers are not going to be trust. Hence, many customewrorthy s seemand will deliver l to understaond that sellew quality in period rs can be
more reliably absent in the last peritrusted if theyod, a consid have an inceerable ntive to be trportion of customers ceusasetworthy, but when thiss to trust. incentive is altogether
ShockpeEstimates riod (5) in(5) and (6) cadicates tpture the effehat the probabct of the shoility of sck perioeparation is 15.56d in the two treatm points higents. The eher in period 3, astimate for nd (6)
suggprices inests thatcrease mu the probach less bility of breakuin CMPrivate p is 11.77 than in CMpoPuintsblic in re lower with psponublic infose to the cormatst shock ion. Even though (see table
custome5.4), the brers aarke antagoup rate is hinizegher in d by unjustifieCMPrivdate, i price increndicatinasges. that sellers only partly anticipated that
relationThe estimshiaps.tes (7 Estimate (7) ) to (11)show jointly captures that higher level the effect of s ofprici prices ng on thecontribute to termination o disrupf customtion, and (8er ) shows
that if priceextent of price increas are increasesed,s in a rathe the probabilir surprisity of ding masruptionner. Larger pn is 7.0 pointrice ins highcreases er. Cu(9) astomers reapparentlcyt to the do not
(10)addition reduce the proally contribute to tebability of brearminatiokup. n. SurprisinApparently, gly, very largvery large (by me pricoe increare than 3se0 units) ps seem to be interprice increareted ses
by customers as a signal that the seller is going provide high quality. In contrast, very large price cuts

(11perio)d ha do not res the quantitativeduce the probly strongeability of breast effect on termikup. Finallynation. The hi, the quality provghly signifiided by the sellecant estimate (1r in the pr2) sheviouows s
points in the substhat providing low quality increases equent period. the probability of termination of the customer relationship by 39.9

remarks 5.5 Concluding This paper demonstrates how trust can shape market outcomes. In particular, we provide the first
experieexperimnce gental oodinvestigatios. While we belin of how loneg-tve that our stuerm custodmy prer oviderelations intereships causting inse prisightce s, we also feel the neerigidity in markets for d
to explore pinvestigate risoce rigime additionaldity in custome aspercts of p marice rigirkets furthedity in cur. Fortunately, oustomer markets. For experimr exaentaml design lendple, the design cas itself to n
implemeeasily be andted apin perioted to analyzd 3, and the prie asymmetrice aeds in prijustmece nt corigiuld be compardity. To do so, a negative coed with the price st shorespocnk coseuld be to the
positive shock. It would also be interesting to investigate further the implications of strategic trust for
price rigiand are willindity. We speg to bear the culatecorresp that many seonding collers dosts in pe not resporiod 3 bend tocau tshe cost shoce they hope to reap the gk by increasiaing pricens from s
trading with loyal customers in the remaining 7 periods of the game. To test for the importance of this
trade-off, one could implement a cost shock in, say, period 7 and compare this price response with the
in perioone in ped 7 beriod caus3. If strategic e the gaintrust is imps from tradinortgant, price with loyal customerigidity should brs cane less pro now only be nouncereapd after theed in fewer shock
ds. perioFour remarks concerning the interpretation of our results seem appropriate.
First, we it was knoshown wto all particip that finite long-teants in ourm relar experimtionshipsent that caus the cue price rigistomer relationdity in customeshirp can last for 10 markets. In particulperioads r,
cuat mosstomert. In c relaontrastionshit, sps twithandard theory must as perfectly rational ansud me self-intereinfinite horiszted agons ents ( demongstrat., Klein and Le possibleffler 1981e effects of ).
Fairness motives are likely to provide an explanation for this discrepancy (see Brown et al. 2004 for a
theoretical discussion).
Seconour findind, whilgs ae the intuitionlso draw atte that price inntion to the importacreases nce ofmay antagoni quality reduze ctcustomeions. In partirs mocular, outivated our expr resulterimes shownt,
Sellers anticithat providingpating this re low quality is even more diaction have ansru inceptive to custntive to leave prices omer relationand quships talities han prunchanice incrged afeaseter s.
temporary cost shocks. That customer relationships may not only induce price rigidity but also quality
rigidity is illustrated in the extreme example of the Nickel Coke. Levy and Young (2002) argue that
becaubottle conse of astannt over a peri implicit contract od of 70 years (frowith its customerm 1886 s, Cotoca-Col 1959), but a not onalso hely held the nld quaominal plity constant orice of a ver this
period despite important cost and demand shocks. However, firms may be forced to choose between
price and quality adjustments at some point. Whether price or quality rigidity is likely to be more
pronounexample, the journced may importanal Cotnsumer ly depend onReports (1994, 59 the relative (10), p. 618) repvisibility of price anorts that the d quality changtypical caes. For n of dog food
once held 14 ounces. Now it holds 13.2 ounces. The can itself hasnt changed - and neither has the
price. Rotemberg (2002) provides further examples of how almost imperceptible quality reductions
can be used to mask price increases.
Third, oureputatior stuns fodr providiy demonng histragh quality coted that long-term reluld only prevationships cauail within the relationse price rigiship. That is, selledity in a setting in which r
reputatioother potens ntiawere relationshl buyers. However, if sellers caip-specific in our expen devrimeelop genent, and reputral repuations tations throucould not be trangh marketsms (eitted to .g., by
advertisiterm relationgn) orships social netcan beworks (see very different than Huck and Tyin the abran 2senc003), ine of thecentivses for ue transmissiopholdin chang personnels. nal long-
Fourth, we find that long-term relationships are profitable for sellers in our experiment. However, while
long-teenvironrm relmentationships. For examples clea, Graysorly have benn anefits, td Ambler (19hey may also have d99) suggrest that long-teawbacks in rmmore customcomplers mex ay hold
higher expethe potential for dissatictations ons sefaction with a givrvice quality than oneen servic-time e quacustomers. Since higlity, firms may have to incur hiher expectationghesr co insts to crease
ship. stomer relationion of the cuavert terminat 68

5.6 ReferencesAkerlQuarteorly Jof, G.A. ur(197nal of Eco0): The nomics, 8market for le4, 488m-500. ons: quality uncertainty and the market mechanism,
Apel, M., Friberg, R., Hallsten, K, (2001): Micro foundations of macroeconomic price adjustment:
Survey evidence from Swedish firms, Working paper Sveriges Riksbank no. 128.
Arrow, K.J. (1973): Social responsibility and economic efficiency, Public Policy , 21, 303-17.
Berg, J., Dickhaut, J., McCabe, K.(1995): Trust, reciprocity, and social history. Games and Economic
, 122-42. 0Behavior 1UndeBlinder, A.S., rstandingCane Price Stickitti, E.D., Leboness, Russellw, D.E., Rudd, J. Sage FoundB. (1ation, Ne998):. Asking w York. about Prices. A New Approach to
Brown, M., Falk, A., Fehr, E. (2004): Rational Contracts and the nature of market interaction.
Econometrica, 72 (3), 747-780.
Camerer, C. (2003). Behavioral Game Theory. Experiments in Strategic Interaction. Princeton
Princeton. ,ty PressiUniversCarltovaluable for un, D.W. (1989nderstan): The tding macroeheory anconod facts about mics?, in: Schmalmarkets enseeclear: Is industri and R.D. Willal orgaig (Eds.)nization , Handbook
of Industrial Organization, Vol. I. North Holland, Amsterdam, 909-46.
Cason, T.N., Friedman, D., (2002): A laboratory study of customer markets. Advances in Economic
Policy 2, Article 1. nd Analysis aCalaboratory mason, T.N., Gangrkets, Joadhurnaraanl, L. (2002): E of Environmentnvironmal Econental omics alabeling and Mannd iancompgement,43, 1lete co13-3nsum4. er information in
Engle-Waof Economirnic cBehavior & Ok, J., Slonim, R.L. (200rganization4): , Vol. 55, 575-5The evolution93. of strategies in a repeated trust game, Journal
Fischbacher, U. (1999): z-tree. Zürich Toolbox for Readymade Economic Experiments. University of
Zürich, Working Paper No. 21.
and equilibriumFranciosi, R., Kujal, P., Michelit prices in posted-offer masch, R., Smith, V.L., rkets, EconoDenmgic Jo, G. (19urnal, 105, 93895): effect on temporary
Grayson, K., Ambler, T. (1999): The dark side of long-term relationships in marketing services, Journal
of Marketing Research, 36, 132-41.
Huck, S., Tyran, J.-R. (2003): Social ties and reciprocity in markets for experience goods. Mimeo.
James, H.S. (2002): On the reliability of trusting, Rationality and Society, 14, 231-58.
conKachelsummeieer price respr, S.J., Limbergon,se to informatio S.T., Schadewaln about prd, M.S.oducers (1991coa): A labst and oratoprofits. Accounry market examinatioting Review, 66, n of the
694-717. Kachelinvestigatiomeien, Jourr, S.J., Limbergnal of Economi,c S.T., Schadewal Psychology, 12, 4d, M.S. (199147-64. b): Fairness in markets: a laboratory
Kahnemin the market. Americaan, D., Knetsch, J.n Economic RevieL., Thaler, R., 1986. Fairnew 76, 728-41. ss as a constraint on profit seeking: entitlements
Klein, B., Leffler, K.B. (1981): The role of market forces in assuring contractual performance, Journal
. 1of Political Economy, 89, 615-4commitmenKollock, P. (1994): Tht, and truse emerget, American Journce of exchanange l of Sociologystructures, 100, 313-4: an expe5. rimental study of uncertainty,
Working papLevy, D., Young, A.T. (200er Emory Univ2): Thersity. e real thing: nominal price rigidity of the nickel Coke, 1886-1959,
Okun, A.M. (1981): Prices and Quantities: A Macroeconomic Analysis, The Brookings Institution,


gton, D.C. shinWa

chaRotembnges anerg, J. (2002d monetar): Cuy policstomer ay, NBER wonger atrkin price ing papcreaer 9se320. s, time variation in the frequency of price

Selten, R., Stoecker, R. (1986): End behavior in sequences of finite prisoner's dilemma supergames,
Journal of Economic Behavior and Organization, 7, 47-70.

DallaWynne, M.A. (199s, 1st quarter, 1-5): Sticky12. prices: What is the evidence? Economic Review, Federal Reserve Bank of



Appendix 1.A: Instructions The instructions were originally written in German. We document the instructions of the treatment with
a leadewere adar in peptedriod 1 accordi-10 anglndy. The ori without a leaginal instruder in pectionriso adr 11-2e avail0. The inable fromstructions the author upon in the other treatrequest. ments
nt rmation on the experimel infoaGenerYou are tacarefully, you can, depking part in an eendixperimeng on your dnet on decicisiosniosn-m, earn akia considerablng. If you read the followine amount of g instrumoney. It is therefoctions re
very important that you understand the following instructions.
Earnings In the experiment you earn points. The points you have earned will be converted in Euros at the rate
. 1 point = 0.40 At the end of the experiment your total income in Euros will be paid to you in cash.
Group membership and anonymity
During the experiment you are member of a group of four participants, i.e. in your group will be three
more grouyou form a group with the p members. The compsame partiositiocipants thrn of the grououghoup is the sat the experime during the ment. It will be randwhole expeomly drimeent. Thus termined
with whom of the other participants you form a group.
membeAll participrs of hantis des grocide aup. To enonnsure anonyymously, i.e. no particimity it is panimperative that all particit will ever learn the identity of the other pants observe the following
rule: otherDuring thwise exe express periment alyourself. l communication is prohibited, i.e. you are not allowed to speak or
Should you have any questions please ask the experimenter.
ure cederimental proThe exp

The experiment consists of several periods.
period 1-10 situation in The decisionThe procedure is the same for each of these periods 1-10.
Every participant receives 20 tokens at the beginning of each period. Your task is to decide how you
projeuse youcrt and h endoowmew many of them to keent. You have to decipd for yourself. e how many of the 20 tokens you want to contribute to a
come our inytion of The calculaYour income consists of two parts:
1. The points you keep for yourself = Your endowment  your contribution to the project
2. The income from the project. The income from the project is calculated as follows:
Your income from the project = 0.4 * (total contribution of all four group members to the
t) cprojeYour income in points is therefore:
(20  your contribution to the project) + 0.4*(total contributions to the project)


each groThe income oup mfembe each gr rerceivoup membees the same ir fromncom the projee from tct ishe proj calculated in ect. the same way, this means that
Examples: Suppothe grouse the p receives an insum of the cocontributionme from the sproje of all groucp memt of: 0.4*60 = 24 pbers is 60oin tokets. ns. In this case each member of
of 0.4*9 = 3.6If the total contribution to the proj points from the project. ect is 9 tokens, then each member of the group receives an income
For each token, which you keep for yourself you earn an income of 1 point. Supposing you contributed
incomthis tokee from tn to the projhe project inect wostead, then theuld rise by 0.4*1=0. total cont4 pointribus. Hotion to the prowever thje incomect woeuld ri of the other gse by one poinroup t. Your
woulmembers wod rise by 1.6 points. Yould also rise by 0.4 points ur contributioneach, so that to the projec the total incot therefore also raime of the groses the incomup from the peroje of the other ct
groumemberp membs to ters. Ohe project. Fon the other hand you r each point contributearn an inedcome for ea by any membech proint contri you earn 0.4but*1=0.4 ped by the other oints.

The decisionOne member in each group decides first on his/her contribution. When this member has made his/her
decision, the other members of the group are informed of the amount of tokens he/she has contributed
to the project. Thereafter the other group members decide.
The person, who decides first in a group, remains the same for all 10 periods.
the following input-screen: s first in a group will see The person who decide

The period number appears in the top left of the screen. In the middle of the screen you will find the
contrifollowingbution will be di information: Your splagroup coyed to the other membnsists of 4 ermembes rof your gros. You aurp.e the person Then they will decidwho decide on their es first. Your
contribution. Your endowment is 20 tokens.
your deIf this screen cisionappe by typing a numbears on your mor betwenitor you are the men 0 and 20 inember of your gro the input field. This field can be reup who decideaches first. You take d by


clickingyou have also deci it with the mouse. ded howAs soo many pointn as ys you keou have deep focir yourded how man self: This is (2y points to co0  your contribute to thentribution proje) tokect, ns.
the Enter-keyAfter entering) your cont. Once you hribuave done thition you musts, your deci press the O.sion can no longK. button (either with the er be revised. mouse, or by pressing
As soon as the decision is made the following input-screen will appear for the other three members of
p: the grou

The coYou take yountribr deciution of the grosion by typing a numup member, who hber betas maweende his/he 0 and 20 in thr contribution de input field aecisind coon first is dinfirm bysplayed. pressing
the O.K.-button.


After all three have taken their decision the following income-screen is displayed for all members of
p: the grou

grouYour own contp are displayed. Furthribution aned the cormorntributioe, you are infon of the mermed mber who haof the total contribution of as made his/her decill group msioemn first in your bers
(including your contribution) and your income in the current period.

? sstionany queDo you have


ure in periodcederimental proThe exp 11-20

of these periods. The experiment will continue for 10 more periods (period 11-20). The procedure is the same for each

-20period 11 situation in The decision In the following periods the composition of your group remains unchanged, i.e. your group consists of
the same pAll values likeersons a the endoswm in the previouent of 20 tokes periodn per pes. The riodedcisi and the calcuon situation is balation of your incomsically the same ae remsa in before.
unchanged. The only difference in the following periods is that there will be no one in your group who
decides first anymore. This means that every participant takes his decision without knowing the
contribution of any other member of the group. The following input screen will appear for all
s: pantpartici


your coOn the incomntribuetion) an-screed your n your contribincomeu displation, the total cyed. ontribution of all members of your group (including


Appendix 1.B: Tables 1.a and 1.b followeTable rs con1.a Stribution for epearman ranach inkdep correlatioendnent grou coefficip witenth αs of a lea = 0.4 ders contribution and his average
Leader in period 1-10 Leader in period 11-20
Group No. Random No. Group rider Free Group No. Pro-social
22 .134 13 .848** 1 .877* 23 .984** 14 .839** 2 .012 24 .692* 15 .706* 3 .745* 25 .471 16 .752* 4 .834** 26 .887** 17 .742* 5 .459 18 .873** 6 .757* 27 .806** 28 .199 19 .794** 7 .218 29 .486 20 .904** 8 .696* 30 .908** 21 .228 9 .194 10 .617
11 .500
12 .425
** significant at α = 0.01 (two-sided) * significant at α = 0.05 (two-sided)

followers Table 1.b. Scontribution for epearman ranach inkdep correlatioendnent grou coefficip witenth α =.s of a lead6 ers contribution and his average
Group No. Leader in peRandriod 1-10 om Group Free Leader in pe Group riod 11-20No. Pro-social
rider No. 1 .56 10 .395 15 .412
2 .711* 3 .799* 11 12 .982** .859** 17 16 .346 .781*
4 .304 13 .826** 18 .390
5 .521 14 .854** 19 .567
6 0
7 .787**
9 8 .475 -.020
** significant at α = 0.01 (two-sided) * significant at α = 0.05 (two-sided)

Appendix 2.A: Instructions treatmThe original inent without punistructionsshm were in Geent oppormrtunityan. We do in the firscumt sequeent the instrunce and ctionswith the opp of the publiorc informtunity to puniation sh in
instruthe sectioncons ard sequee availablnce. Thee from instructions the author upon in the other treatmrequest. ents were adapted accordingly. The original
nt rmation on the experimel infoaGenerYou are taaccording to the deking part in an ecisionxperimes you make. Plnt on decisioease read thesn makie instructng. At the end of the exions carefully. periment you will be paid
EarningsDuring the experiment you earn points. The total income in points your earn during the experiment will
s at the rate in Euronverted be co5 . 1 point = 0,02At the end of the experiment your total income in Euros will be paid to you in cash.
At the beginning of the experiment every participant receives 100 points.
onymity d anership anp membGrouDurimore grng the exoup membeperimrs. Thent you are meme compositiobern of the grou of a group of fp is the saour pamerticipan during the ts, i.e. in yowhole expeur group rimwill be three ent. Thus
you form a group with the same participants throughout the experiment. It will be randomly determined
with whom of the other participants you form a group.
All participants decide anonymously, i.e. no participant will ever learn the identity of the other
members of his group. Therefore it is imperative that all participants observe the following rule:
During the experiment all communication is prohibited, i.e. you are not allowed to speak or express
otherwise yourself. Should you have any questions please ask the experimenter.
ure cederimental proThe exptask iThe expes to deciriment conde, how masistsny of ten periods. T points you contribute to your ghe procedurerou of each of these pps project. eriods is the same. Your
a period situation in The decisionwment doYour encan bAt the beginnie between 0 and 4ng of each pe0. Horiod eaw many poch partints eaicipanch pat has anrticipa endont has wmewinll be randt of points. Thomly determinee amount of pod for ints
every single participant. Every number between 0 and 40 is drawn with the same probability. Every
endomembewmern in a grout is nep can therwly determined at the beginefore have a diffening rent endof owmeeach perint between 0 anod for every singld 40 e participoints. The pant.
on cisiThe deYou have to deciyou want to contribute to thde how e projemany points ct of your gof your endoroup. wment you want to keep for yourself and how many
ation of your income ulcThe calYour income consists of two parts:
1. The points you keep for yourself = Your endowment  your contribution to the project
2. Your income The income from the projfrom the project = 0.4 * (toect. The incomteal contribufrom the project is tioncalc of all four group memulated as follows: bers to the
t) cprojeeach groThe income oup mfembe each gr rerceivoup membees the same ir fromncom the projee from tct ishe proj calculated in ect. the same way, this means that


Your income in a period is therefore:
(Your eto the projecndot). wment  your contribution to the project) + 0.4*(total contribution of all four group members
membeExamples: rS of the grouuppop rese the ceives an insum of the concometributions of all gr from the project of: 0.4*6oup mem0 = 24 poibers is 60 tonts. If the total kens. In this case each
contri3.6 points frobution tom the proje the projecct. t is 9 tokens, then each member of the group receives an income of 0.4*9 =
For each token, which you keep for yourself you earn an income of 1 point. Supposing you contributed
incomthis tokee from tn to the projhe project inect wostead, then theuld rise by 0.4*1=0. total cont4 pointribus. Hotion to the prowever thje incomect woeuld ri of the other gse by one poinroup t. Your
woulmembers wod rise by 1.6 points. Yould also rise by 0.4 points ur contributioneach, so that to the projec the total incot therefore also raime of the groses the incomup from the peroje of the other ct
groumemberp membs to ters. Ohe project. Fon the other hand you r each point contributearn an inedcome for ea by any membech proint contri you earn 0.4but*1=0.4 ped by the other oints.
sion ciw to take your deHoAt the beginning of each period the following input screen appears:

On this screen the amount of your endowment in the current period is displayed. You take the decision
the O.K.-button your contrion. You can bution to the projecat maximum contrit by typing tbuthe coe yorrespondiur whole enng numdowmenber in the inpt. If youu do not wat field and pressing nt to
ease enter 0.bute, plcontriAs soon as all members of your group have made their decision the following income screen appears:


There you see your own contribution to the project and the total contribution of all group members
(inclincomudineg from t your cohe projntribuect antion). Furthermd your total income in thore it is displisayed ho period, whiw manych points y is the sum ofou kept for yourse the points yolf, the u kept
and the income from the project.

Another screen will appear on which you can see for each of the four members of your group
septhe endoarately thwmee endont this is. Youwmenr own t, how many pcontriboution is alints were wacoysntributed to th displayed ine proje the first coluct and how mamn in blue, the ny percent of
contributionnote: the order of the s of the other three gother group memroup membebers will rs in colube new in each period. mn two, three and four in random order. Please


sstionany queDo you have



Information for the following 10 periods
procedure is the same in each 10 more periods. Like before the The experiment will continue for d. perioIn the following period the composition of your groups remains the same, i.e. you form a group with
the same participants as before.
The decision situation in the following 10 periods
The decision task now has two stages.
ge First staThe decision in the first stage is the same as in the periods before: You decide again, how many
incompoints of your endoe are the same awment s beyou contrifore and the butendoe to the projewments cwill be rat. The prondocedure amly determinnd the calced for eaulatioch pn of your articipant
d. oin each periWhen all members of the group have taken their decision the income screen again appears, which will
inform you on your income (in the first stage). Thereafter the second stage follows.
Second stage
In the second stage you decide whether and to what extend you reduce the income of the other
members of your group.
e d stag in the seconThe decisionYou will seand the contrie again the bution in pescreen (see brcent of the endellow) oowmen whicnt are dh ithe endosplayewmed for anll four grots, the contriup memberbution to the projes. Your own ct
contriyour group in bution again is dirow two, thresplae and four in yed in the first row in blrandom ordeue, the contr. Please note:ributions of t This order ihe oths er threrenewee memd in each bers of
d. perio


You now must decide how many points you want to give to each other member of your group.
Each point you give to a specific member of your group reduces this members income by three
points. If you choose 0 points, the income of this group member is not reduced. You can distribute at
maximum 10 points to each of the other group members.
The sum of all points you give to other group members are reduced from your income.
After you have decided how many points you distribute to all group members please press the O.K.-
button. After all participants have taken their decision in second stage the following income screen appears:

Your income from stage 1 and the points you have distributed is displayed. You are also informed of
how many negative you received from other group members and how much your income is reduced
s. tnby these poiated as: calculcome is Your final instage from the first come Your inminus costs for distributing points

Minus reduction through received points (=3 * received points)
period in thisYour Income =Please note: Your income from the first stage is at maximum reduced to zero through the points you
distribureceived fromted, such that your other group incommembere cas. Hon bwever you e negative in always ba period. ear the costs for the points you have
? sstionany queDo you have 84


bles 2.a  2.d Appendix 2.B: TaTable 2.a. Spearman rank correlation coefficients of contribution rates and endowments
Treatment PRIVATE without with without PUBLIC with
group no. opportunitpunishmenyt opportunitpunishmenyt opportunitpunishmenyt opportunitpunishmenyt
1 -.220 -.714*-.833* -.476 2 .060 -.738* ** -.810* -1.0003 -.548 .119 -.881** -.690 4 -.195 -.686 -.095 -.357 6 5 -.667 -881** -.667 -.038
-.405 -.333 -.333 -.119 7 -.611 -.452 -.156 -.143 8 -.643 -.143 -.857** -.905** 9 -.805* -.786* -.167 -.905** 10 .143 -.548 -.857** -.738* 11 -.690 .429 .690 -.786* 12 -.805* -.714* -.857** -.857** 13 -.357 -.810* -.595 -.810* 14 .330 -.180 -.619 -.048 15 .452 .262 -.095 .452 16 -.333 -.881** -.619 -.667 17 -.790* -.714* -.738* -.952** 18 -.467 .286 -.333 -.048 19 -.265 -.190 * significant at α = 0.05 (two sided)
** significant at α = 0.01 (two sided)

Table 2.b. Spearman rank correlation coefficients of contribution rates over periods
group no. wipunishmenthout t wipunishmenth t wipunishmenthout t wipunishmenth t
opportunity opportunity opportunity opportunity
1 .874** -.267 -.236 .770** 2 -0.28 .600 .406 .588 3 .939** .876** -.152 .455 4 .813** .073 .345 -.345 5 .500 .333 -.648* .758* 6 .865** -.612 -.745* .079 7 .865** .176 -.176 -.345 8 .442 .212 .564 .115 9 .957** .304 -.612 .855** 10 .588 .164 .358 .042 11 .937** -.721* -.952** -.467 12 .915** .782** .139 .709* 13 .127 .273 -.661* -.164 14 -.074 .091 -.648* -.358 15 .738* .867* .127 -.445 16 -.091 .224 -.571 -.442 17 .608 .673* -.091 .636* 18 -.612 -.006 -.273 .742* 19 -.321 -.394 * significant at α = 0.05 (two sided)
** significant at α = 0.01 (two sided)

Table2.c. Spearman rank correlation coefficients of received punishment points in period t and
average relative change in contributions from period t to t+1

group no. wipunishmenth t wipunishmenth t
opportunity opportunity
1 .649 - 2 .321 -.200 3 .100 .500 4 .700 -.500 5 .543 -.400 6 .238 -.874** 7 .800 -.857* 8 -.486 ** -1.0009 .900* .200 10 .109 .800 11 1.000** -.486 12 - .143 13 .410 .500 14 .781* - 15 .371 -.248 16 .319 -.044 17 .829* -.800 18 -.500 - 19 ) dedwo sisignificant at p = 0.05 (t* ** significant at p = 0.01 (two sided)


contriTable 2.d. Sbution (in percentapearman range points)k and correlationreceived p coefficiunishents of deviationment points from groups average relative
group no. wipunishmenth t wipunishmenth t
opportunity opportunity
-.893** -1.000 1 -.679 -.200 2 -.900* .624 3 -.700 ** -1.0004 -.829* .391 5 -.286 .333 6 7 .179 -1.000**
.029 .500 8 ** -1.000.100 9 -.427 -.800 10 ** -1.000.486 11 - -.257 12 -.700 -.500 13 -.810* - 14 -.943** -.321 15 ** -929-.122 16 -.357 .400 17 - - 18 - 19 * significant at α = 0.05 (two sided)
** significant at α = 0.01 (two sided)


Appendix 3.A: Instructions The original instructions were written in German. The following instructions are those for the treatment
IR8. The instructions for the treatments IR2 and IR4 differ only in the corresponding numbers. The
instructions for treatment GR4 contain one different sentence which is indicated below. The original
instructions are available from the author upon request.
Experimental Instructions During the experiment you belong to a group consisting of 8 randomly chosen members.
{you have regiGR4 treatmentstered}. : During the experiment you belong to a group consisting of the 4 members with whom
The composition of each group does not change throughout the experiment.
grouThe expeps resultriment starts of the previous rowith the first rouund.nd. Whethe A maximum of 10 rour therndse will be a fol are played. lowing round depends on the
Contribution to the group payment?
In every round, each member of a group decides whether to contribute to a group payment or whether
ontribute. not to cIn each round, the group has to raise a total amount of 1680 talers in order to reach a further
d. rounWho is able to pay and who pays how much?
For each gthis, in every rounroup membd the exper it is ranerimenter adomly dseks a rantermindomed whly choether he senre particeives an amouncipant to roll a die. If the number it of 420 Taler. For s
identical to the number for the respective round, which a participant has written down in the first
columnot match, thn before particie the beginnipant receives ng of the experime420 talersn. Thust, this partici, the decisipant receiveon whether a ps 0 talerasrtic. If ipanthe numberst receives 42 do 0
talers takes place at random and independently ofall other participants.
A group meA group memmber who haber can only s contri0 talers cannbute to the groot contribuup payte. His pmenta if he has reyoff is 0 talers. ceived 420 talers in this round.
The amount of 1680 talers that must be raised by the group will be divided equally between those
groucontrip members bute, i.e. have receivwho decided to coed 420 talentribrs.ute to the gro The more memberup payment as of a grnd whoup are o in adwilling and adition are abblle to e to
contribute, the less is the amount each member has to pay. This amount will be subtracted from the
mber. 420 talers received by each paying meThe contributions and round payoffs of a paying member result as follows:
grouNumbp member of paers ying Contripaying membution per ber Group raises 1680 talers? Roupaying memnd payoffber for a
0 - No Æ no further round
1 420 No Æ no further round 0
3 420 2 420 No No ÆÆ n noo furt further rounher round d 0 0
5 336 4 420 Yes Yes ÆÆ further round further round 0 84
7 240 6 280 Yes Yes ÆÆ further round further round 180 140
210 further round ÆYes 8 210 All members who have received 420 talers and have decided not to contribute to the group payment
round. the payoff for this as 420 talersobtain thisperiment the exfStructure o 89


the grouThe structup payre of all roundment by tickings is identi yes or nocal. Each g in coluroup memn 2 mber dof the deciecidesion shes whetheet. r he wantThe decision is taken s to contribute to
onlybut not those for the currenfor the followit roun Thus, pleads. se tick only the box which corresponds with the current round,
numbeThe exper is primeunters blicly anncolleounced act all decisnd reion cosheets arded in colund mna randomly cho 3 by the experimsen particienter. Tpanhe provit rolls a die. sionalThis round
payoff results as explained above (column 4): if the numbers of column 1 and 3 match, the group
member receives 0 for this round; if the numbers differ he receives a provisional round payoff of 420.
The experimenter determines the number of paying group members (column 5) and writes down the
raisifinal rounng the ad pamount of 168yment for each membe0 talers, anrd the in columexperimn 6. Colent continueumn 7 informs wis whth the next round. ether the group succeeded in
conveFor earch pted intartoicipan DM at an exchant the total payoff is determinge rate of 2.00 DM ped (in er 100 taletalers) at the end of the experimrs. Additionally each particient and panis t
DM. mount of 5.00 anreceives aThe total amount is paid in cash to each participant at the end of the experiment.
Please note: All decisions are made anonymously, i.e. at the end of a round each member is only informed about
his oparticiwpantn decis sision, his ot in a cubiclwn pe. No concluayoff and the numbesions can be dr of rpaawn reying grougarding p members. Durithe identity of the other grong the experimup ent all
members, their decisions or payoffs.
To make sure that anonymity is guaranteed, it is necessary to adhere to the following rules:
During the experiment it is forbidden to speak or to communicate in any way.
All members are asked to stay in their cubicles till the end of the experiment.
further roundEven if in a round a group s are taking pldoeace for this not sus grcceeoup, all pad in raisinrticipang the amount ts have to stay in their cubiof 1680 talers, which mecles untianls that no the end
this groof the experiup. Ofm course, thoent. The experimse grouenterp mems continue to habers do not have to take dnd out and to ecicollesionct the des anymore cision papers also for


Appendix 4.A: Instructions Original instructions were in German. The instructions were the same in both treatments. In the private
information treatment the information on the temporary cost shock was shown on sellers screens (see
Appendiabout the cost shock was x 5B) only at the beginnishown on ang of pll pearticipriod 3. In antsthe screen public infos.). Thrme origination treatmal instructionsent the inform are avation ailable
est. qun re the author upofromnt rmation on the experimel infoaGener the experiment you will be paid according to the end ofperiment. At You are taking part in a market exthe decisions you make. Please read these instructions carefully.
experimEarnings. Duent will be convering the experted in Euroriment you ears at the ran points. The te 1 point = 0.012 total income i and pain points you d to you in caearn durish. ng the
particiPlease note: pant willDuring the e ever learn the identitxperiment all particiy of the persopantns s decidewith whom he i indepenterandently and acts. Thernonymousefore it is imperative ly, i.e., no
that all participrohibited, i.e. ypants observou are not alloe the followinwed tog rule: speak or exDurinpress og the expthererimwise yourself. ent all communication iShould yous have any
questions please ask the experimenter.
Overview of the experimental procedures
determiIn the experined atm the beginnient there arnge buyers and of the experiment and disellers. Whether you asplayed on yore aur co buyer or a mputeser screller will be en. You wirandomly ll keep
your role throughout the whole experiment. The experiment is divided into several trading periods.
In every trading period a seller can sell one unit of a good and a buyer can buy one unit of the good.
The seller can produce the good either in high or in low quality. The quality determines the production
d to the buyer. d the value of the gooer incurs ansts the sellcoa price whiSellers and bch uexceeyers ds hican eas costs of prn points by conclroduction, uding a a buyer earntrade. A selles pointr earnss if he buys a g points ifood who he sells the se value good at
exceeds the purchase price.
In every period a sellers task is to determine the price at which he sells the good. Moreover, he
determines whether he will deliver high or low quality. This determines the production costs he incurs
and the valuethe quality of the good and of the good to the buyer. hence he doesAt the time not know thof purchase a be value of the good. uyer knows only the price but not
Calculation of your income in a period
Income of a bquality: 200. These valueuyer. For a bsu are the sameyer, a unit of a good ha for all buyers.s the followi The incomeng values of a buyer is : low qucalculateality: 100, high d as follows:
income = value of the delivered good  price.
If a buyer decides not buy a good in a period, his income is 0. If he buys at a price that exceeds the
value of the good delivered, he makes a loss. Losses will be subtracted from your income.
prodIncome of a uction coseller. A sellests if he producer has los a higw prodh quction couality good. The exact prodsts if he produces uction a low qucosts aality good are onlynd hig knowh n to
the sellers and will be displayed on the sellers computer screen. The production costs are the
same for all sellers. If a seller does not sell his good in a particular period, his income is 0 in this period. If he sells his
ws: ated as follocalculs income is good, hiIncome = price  production costs subject to quality.
Every participant receives an additional 150 points at the beginning of the experiment.
How to trade
There are two markets: market I and market II. Trading rules differ across markets.


s for markeTrading ruleI t

matcheOn market I, a sped trading pair acific ssell loner tradg as they are es with a spetrading coific bun market I. This myer, i.e. a seller and a beanuyes that you deal with r form a fixed the same
person as long as you stay on market I.

prodA seller mauction cokests ans an offer by d the value of the godetermining a price aod for the buynd theer in ca quality he wilse hel accept supply. This s. On market I a selalso determilner sees his es
en: input scregthe followin

price and The productioclicking on consts of hig the qh and louality he delivers if the w quality will be dibuyer accesplayepts. By clid. The selleckir mang on tkes his offer by ehe o.k.-button he ntering a
d. n not be revisefer to the buyer. After this, the offer cas his ofsubmit

Please note: The buyer will only see the price but not the quality the good. Only upon acceptance
does he learn the quality and hence whether the good is of high or low value to him and which income
d. s earnehe ha


ng screen: A buyer will see the followi

The buyer decides, whether to accept or reject the offer by clicking the corresponding button.

• The buyer accepts the offer

good (lThe buyeow orr wil l see an inhigh); the valcome screue of the gooen wid (100 oth the followinr 200)g and hi informatios income for this pen: price; the delivered quriod (= value ality of the  price)

has pThe sellroduceder receives an in; his incomcoe in this perime screen wiod (= pricth: the price  pre; oductiohis produn coctiosts n cosubjsts, suect to thebjec produt to the qualiced quty he ality).
the same buyOnly if the buyer has accepted the oer in the following period. ffer, will both stay on market I, i.e., the seller will submit an offer to

• The buyer rejects the offer

for the cuAfter a rejerrenction on mat period and the buyer can rket I, both buyer and seller trabuy one of the available ofde on market II. fers on mThe sellearket II. r submits a new offer


II s on market Trading rulei.e., even if allOn market II there a buyers are cseveceral sellers anpt an offer there will bd several e sobuyme selleers. There are alwayrs, who will not sels more sellers thal their good. Contrary n buyers,
has made a pto market I an offer canarticular offer.not be su bmitted to a specific buyer and a buyer cannot identify which seller
an offer tw to submiHoEvery seller has to submit an offer in every period by determining a price and the quality of his good.
en: ng input-screwill see the followiA seller

enterinThe produg a prictiocn coe and clists of a hckinigh and log on the qualiw qty he will produality good will be displuce if a buyer accepts hiayed. The seller s omakes hiffer. By clicking on s offer by
revithe sused. bmit-button with the mouse he submits his offer. Once an offer is submitted it cannot be


s er of the buyThe decisionis not possiblA list of all offers on markete to infer whic II inh seller h decreasias made a specifng order of price will be diic offer. splayed to buyers on that market. It
quality will be delivered. Only on the prices offered but not, which can only see Please note: Buyers acceptalow value for nce does a buhim and which incomyer learn the qe he ualithas ey of arnthe ed ingood he boug this trade. ht and thus whether it has a high or a

Buyers will see the following input-screen:
The buyers on market II are asked, one after another, whether they want to accept one of the
he wants to aavailable offers. A first buyccept an offer, he clicks oer is randomlny it with t chosen. Thishe mouse. With a buyer will see another clill availablck on the ae offers on market II. Iccept offer-f
button the trade is concluded and the offer is removed from the list. If he does not want to buy any
offer he clinext buyer is crandks on the no pomly chosurchase-ben. He will see all outton. As soffers ton ashat are the first buyer hastill available on ms finishead his derket II and decidcision, the e
r. r to accept an offeewhethscreen: be informed via an income ded they will Once all buyers have deciwhethor high qualiter they have concludy; the value of the good (1ed a trade; at00 or 20 which pri0) cand the income they have purchaes in this peried; whetod her the go(= value  prod iicse). of low
in this period Sellers, in turn, are inform(= price  productioed via their inn costs come subjscreect to theen: whethe produced quality). r they have sold their good; their income
Therewith the trading period is over and a new one starts in which new trades can be concluded.
Trading phases and access to market I and market II
There will be several trading phases. Every trading phase consists of 10 trading periods.
When a new phase starts the period counter at the top left of the screen will start again with period 1.
as a fixed paiIn every trading phar, as long ase sos tme sellehe buyer ars and buyers staccepts the sellert on market I. rs offers. As sooA seller an as the bnd auyer rej buyer stay oects ann offer market I
they both join market II and stay there until the trading phase ends after 10 periods. It is not possible


to return to market I. The sellers and buyers who start on market I are randomly chosen and randomly
d. matche

suAt the beginnich that nobody will trade ng of a new trading pwith the samhase eselle person rs and whebunyers a starting are aggaaiin randn on market I. omly chosen and matched

trading pAll sellers anhase. d buyers who do not start on market I are on market II for the whole 10 periods of a


st shock in period 3othe cAppendix 4.B: Screens announcing

At the beginning of period 3 the information on the temporary cost increase was shown on the screens
below. In the private information treatment (PRIVATE) the following screen was shown to sellers:

In the public information treatment the screen below was shown to both sellers and buyer.


Appendix 4.C: Figure 4.a

Average transaction prices with public information (upper panel) and private information (lower panel)
se 1-5 in pha


am University of Nottinghcs Economiof School Parkity Univers2RD NG7 am Nottinghm odKingUnited Tel: ++44 (0)115 951 5399
Fax: ++44 (0)115 951 4159

Curiculum Vitae erElke Rennam University of Nottingh ss Addrecs Economiof School Parkity Univers 2RD NG7 am Nottingh United Kingdom
Tel: ++44 (0)115 951 5399
Fax: ++44 (0)115 951 4159
tion nt PosiCurresince 12/2003 Lecturer at the School of Economics, University of Nottingham
Education 2000-2003 Junior researcher (wissenschaftlicher Mitarbeiter) and Ph.D. candidate at the
ity of Erfurt Univers1997- 2000 Junior researcher at the Wissenschaftliche Hochschule für Unternehmensführung
resear(WHU)  Ottocher at the Depa Beisheim Grtmrent for Orgaduate Schoanization aol of Managend Managment in Valleement ndar and junior
1989-1997 Economics atTitle of diploma thesis: Experime the University of Bonn, Diplom Volkswinte zu einem Reziprorztin itätsspiel mit
oppoBestrafungrtunity), supervised bsmöglichkeit (Experimenty Professor s Reinhaon Reciprrd Selteocn ity Games with punishment

Research interests
Behavioural economics, microeconomics, game theory, organization theory