S

La

b

c

a

ARRAA

JCC

KTESC

1

opttm

taeeststp

vT

p

2h

Journal of Behavioral and Experimental Economics 53 (2014) 97–110

Contents lists available at ScienceDirect

Journal of Behavioral and Experimental Economics

j ourna l h o mepage: www.elsev ier .com/ locate / jbee

ocial influence in trustors’ neighbourhoods

uigi Luinia, Annamaria Neseb, Patrizia Sbrigliac,∗

University of Siena, ItalyUniversity of Salerno, ItalyUniversity of Naples II, Italy

r t i c l e i n f o

rticle history:eceived 18 June 2013eceived in revised form 18 August 2014ccepted 23 August 2014vailable online 1 September 2014

EL classification:72

a b s t r a c t

We offer new and clean evidence that social interactions impact on individuals’ choices. In an experi-mental trust game we study whether and how trustor’s behaviour is affected by social influence of othertrustors’ choices over time. We account for three important factors of trustors’ preferences: risk atti-tude, generosity and expected trustworthiness. Our results confirm that trustor’s behaviour is affectedby peers. We find a general convergence in trusting behaviour: the effect of social influence is (for mostof subjects) significantly reducing the amount sent by trustors in each period. Furthermore, analyzingcontagion within the neighbourhoods, we find that agents tend to imitate similar types ((un)-generous or

91

eywords:rust gamexperimentsocial influenceontagion

(un)trusting) when placed in the same neighbourhood. Indeed – in the few neighbourhoods with a preva-lence of generous and risk-loving subjects – trust substantially increases over time. Nearness, withoutany strategic component, is a clear element of contagion in trustors’ behaviour.

© 2014 Published by Elsevier Inc.

ewit

tmtTfppbss

. Introduction

The aim of this paper is to study the impact of social influencen individuals’ trusting behaviour. As noted in Fehr (2009), “Trustlays a role in almost all human relationships. . . Trust also seems par-icularly important in economic exchanges because it seems obvioushat the absence of trust among trading partners severely hampersarket transactions. . .”

The basic objective is to clearly identify the determinants ofrusting behaviour in market transactions, as social motivationsre mixed with standard profit motivations that are generallyxamined in all economic exchanges. In the vast experimental andconomic literature where trusting behaviour has been analysed,everal contributions stress the relationship between sociality andrust. For example, Berg, Dickhaut, and McCabe (1995) find that

ocial history is important in that under particular conditionsrust and reciprocity are stronger when individuals can observeeers’ behaviour. Indeed, in the absence of rewards and sanctions,

∗ Corresponding author at: Department of Economics, Faculty of Economics, Uni-ersity of Naples II, Gran Priorato di Malta, 83100 Capua, Caserta, Italy.el.: +39 823274022; fax: +39 823274042.

E-mail addresses: luini@unisi.it (L. Luini), anese@unisa.it (A. Nese),atrizia.sbriglia@unina2.it (P. Sbriglia).

eiNrpGtdtm

214-8043/$ – see front matter © 2014 Published by Elsevier Inc.ttp://dx.doi.org/10.1016/j.socec.2014.08.007

ndogenous social norms can emerge if individuals clearly identifyith a group. Accordingly, social history, by providing common

nformation on the use of trust in groups, may increase social iden-ity.

Recently, a number of experimental papers have focused onhe effects of peer influence on behaviour in economic environ-

ents, an area that had not previously received attention. Similaro our research work, an example of analysis of peer effects in therust Game is presented by Mittone and Ploner (2011). Their paperocuses on the behaviour of receivers and studies the effects of peerressure (when the receivers’ choices are being observed by otherlayers) and the effect of social spillovers (when the interactionetween trustees’ choices is observed). They find that peer pres-ure has a positive effect on reciprocity and – to a lesser extent –o do social spillovers. A direct observation of peer actions is alsoxamined by Falk and Ichino (2006), who find clear signs of socialnfluence among workers engaged in the same task. Gaechter,osenzo, and Sefton (2012) find that information regarding the

eciprocal behaviour of peers affects the individual’s level of reci-rocity in a gift exchange game. Finally, Falk, Fishbacher, andaechter (2013) find that individuals adapt their behaviour to

he neighbourhood to which they are randomly allocated in coor-ination and public good games. The interesting point made byhese studies is not only that price or consumption strategies

ay be affected by peer pressure, but also that reciprocity, trust,

9 Expe

coo

g(iitiNaitUasntr

as(eata

gegivsSuos

mma

ftgbafcplgrf

tctaeoba

ntgb

VsHtbogas

Mt(ouobtigc

iSFsieianahwah

aadtaLastb

benbanet

8 L. Luini et al. / Journal of Behavioral and

ooperation and work efforts are influenced by either “convergencer dispersion of behaviour” in social networks and among groupsf individuals.

Studies on social influence in strategic settings (the dictatorame) have been conducted by Cason and Mui (1998) and Servatka2009). In the latter research, the author focuses on the correctdentification of social influence (i.e., how the proposer’s behaviours determined by the respondent’s observed level of generosity)hus separating this aspect from reputational effects that existn game theoretical models. Finally, in a recent paper, Deck andikiforakis (2012) analyse whether observation of peers’ actionsccelerates the convergence to the payoff dominant equilibriumn a minimum-effort game. The authors devise two differentypes of observation of peers: perfect and imperfect monitoring.nder perfect monitoring (players are able to observe all peers’ctions in real time), there is a clear effect on the equilibriumelection. Under imperfect monitoring (players observe only theireighbours), the effect is minimal. The authors argue that uncer-ainty may play an important role in determining the oppositeesults.

Our experiment is also connected to two rather close researchreas in which the individuals’ choices are affected by otherubjects’ decisions. These research areas address the following:i) the discrepancy between individual and group decisions forxactly the same game (including all the relevant parameters)nd (ii) the role played by new incoming information in shapinghe individuals’ choices (e.g., observations previously unavailable,dvice).

With reference to the discrepancy between individual androup behaviour in games, Kugler, Kausel, and Kocher (2012)xperimentally observe that groups behave closer to theame-theoretical assumption of rationality and selfishness thanndividuals: the authors compare the behaviour of groups and indi-iduals in a two-person trust game and find that groups of sendersend smaller amounts of tokens than individuals. Charness andutter (2012), upon reviewing the literature regarding individ-al decision making in situations with salient group membership,bserve that, in most of the experiments, groups are generally moreelfish than individuals.

The relevance of new information received by the decisionaker is the focus of another research area that focuses on theodification of choices after observing peers and/or after receiving

dvice from peers.Schotter and Sopher (2006) investigate the development of the

ollowing conventions of trust: (i) people receive advice from thosehat made the same decision before them in an intergenerationalame and (ii) advice facilitates the creation of a convention ofehaviour and decreases the amount of trust. Chaudhuri, Graziano,nd Maitra (2006) study a linear public good game with advice (viaree-form messages) being passed from one generation to the suc-eeding generation. Such advice may consist of private knowledge,ublic knowledge or common knowledge. Common knowledge

eads to subjects leaving more “exhortative” advice and therebyenerates a process of social learning. Sbriglia (2008) examines theole of advice in p-beauty contest games and finds that messagesrom winning players accelerate individual learning.

Though the importance of peer influence has been widely inves-igated, there is no previous research which tries to ascertain theausal relationship between social influence and trust. Specifically,here are three research questions motivating our study. Our firstim consists of determining whether trust is affected by peer influ-

nce in neighbourhoods of trustors engaged in the repeated playingf different trust games. Second, in the case of convergence ofehaviour over time, we assess whether social influence producesn increase (or a decrease) in the overall level of trust. Finally, a

btc

rimental Economics 53 (2014) 97–110

ovel feature of the research consists in studying whether and howhe individual’s inner preference for trust (i.e., risk aversion andenerosity) affects the individual’s propensity to imitate others’ehaviour.

Our approach is closer to that adopted by Fortin, Lacroix, andilleval (2007). Their paper focuses on the relationship betweenocial interactions and tax evasion in a repeated public good game.ere, groups of subjects play a repeated public good game where

here is a probability of auditing and where they can observe theehaviour of the other components of the group over time. As inur experiments, the sessions are divided into a “NO-INFO” cate-ory and an “INFO” category, based on whether or not the subjectsre informed about the others’ behaviour, to measure the effect ofociality on the individuals’ choices.

Evaluating social influence is not an easy task. As detailed inanski (1993), the identification of peer effects involves con-

rolling several confounding factors. These include the following:i) the self-selection of individuals into homogeneous groups asbserved correlation in individuals’ actions may reflect individ-als’ similar preferences rather than a causal effect of one’s actionn another; (ii) the exogenous (contextual) effects as individualehaviour may vary according to the socio-economic characteris-ics of different groups; (iii) the correlated unobservable that mightnfluence all group members in a similar way, as individuals in aiven group may behave similarly because either they have similarharacteristics or they face a similar institutional environment.

Even if several studies based on observational data have mademportant steps towards the solution of such problems (i.e.,acerdote, 2001), many authors (i.e., Falk and Ichino, 2006; Falk,ishbacher, and Gaechter, 2013; Hartmann et al., 2007) empha-ise the possibility to better determine the existence of peer effectsn a fully controlled context with laboratory experiments. Thexperiment in this paper circumvents the problems related to thedentification of peer effects as follows: (i) subjects are randomlyssigned to different neighbourhoods; (ii) contextual effects doot occur as interactions are anonymous, and correlated effectsre overcome because subjects face the same context (they allave equal economic incentives and share equal information); (iii)e explicitly check for correlated effects either due to experience

nd strategic learning variations during the trust game or due toomogenous trustors’ characteristics.

As in Ashraf, Bohnet, and Piankov (2006) and Chaudhurind Gangadharan (2007), information on subjects’ characteristicsnd trust attitude (i.e., risk attitude, social preferences, socio-emographic characteristics, and beliefs about the behaviour of therustees) are drawn through the following: (i) a dictator game, (ii)

questionnaire, and (iii) the lottery method suggested by Holt andaury (2002). Because of the random group formation, these char-cteristics are exogenous, which enables us to investigate whetherpecific types of agents (i.e., generous or untrusting) are more likelyo trust and imitate similar types (i.e., because they observe aehaviour consistent with their own preferences).

Our findings show that there is a convergence in trustingehaviour in most neighbourhoods, and the effect of social influ-nce is (in the majority of the cases) to significantly reduce theumber of tokens sent by trustors in each period. Furthermore,y analysing contagion within the neighbourhoods, we find thatgents tend to imitate similar types when placed in the sameeighbourhood. Indeed, in those few groups with a majority of gen-rous and risk-loving subjects, trust substantially increases overime.

This paper is organised as follows. Section 2 presents theehavioural hypotheses and our experimental design, while Sec-

ions 3 and 4 describe our empirical findings. Section 5 providesonclusions and possible extensions.

Expe

2

tt

x

wr

�

warna

uhid

n

x

t

c

x

ti

x

wepczivn

Cb

e(

∣∣

e

6

wbi(oisno

Cott

2

wmaktht

nicoebItdpgtpi

msl(b

An important aspect of the design was that the sender wasaware that his/her neighbours were playing with different recip-ients. In fact, trustors were re-matched with different trustees in

4 As suggested by a referee, an interesting extension to our study, as well as to sim-ilar experimental analyses of contagion, is to examine how the behaviour adopted byan individual trustor changes when the same subject migrates to a different neigh-bourhood. This would allow to test whether contagion produces the emergence ofa single behavioural rule, rather than a series of group-specific rules.

L. Luini et al. / Journal of Behavioral and

. Evaluating social influence: the behavioural hypotheses

Assume that in the trust game the number of tokens sent byrustor i, randomly allocated to neighbourhood L, is represented byhe following function1:

i = �i + ˇ

3∑j ∈ L,j /= i

xj (1)

here x is the number of tokens sent by the neighbours and �iepresents the individual’s preference for trust, specifically2:

i = ˛(zi) + ϑE(y)

here ˛(zi) is an aggregate measure of the individual’s risk attitudend generosity, ϑE(y) measures his/her belief about the trustee’eciprocity, and both measures are exogenous to the choice of theeighbours. is often defined as the “social interaction coefficient”,s it measures the impact of sociality on the agents’ choice.

Eq. (1) shows that if > 0 (positive peer effects) then an individ-al’s trust increases (or decreases) with an increase (or decrease) inis or her neighbours’ trust; if < 0 (negative peer effects), then an

ndividual’s trust increases (decreases) when the neighbours’ trustecreases (increases).

We are interested in the value of ˇ; specifically, if sociality haso effect on the choice of i (i.e., = 0), then:

i = �i; (2)

Trustors’ choices only depend on the individual’s preference forrust.

In the alternative scenario, where the sender’s behaviour fullyonverges within each neighbourhood, then3

i =�i + ˇ

∑3j ∈ L,j /= i�j

(1 − 4ˇ2)(3)

Going back to Eq. (1), our main interest is to determine whetherrusting (or untrusting) behaviour is contagious. Therefore, follow-ng (1), we model the individual decision to trust as follows:

i = xi(ri, gi, ti, zi, x−j)

here xi is the number of tokens sent to an anonymous recipi-nt by trustor i in each period, ri and gi are measures of socialreferences, respectively (individual risk and generosity), ti indi-ates the trustor’s beliefs (namely, expected trustworthiness), andi indicates individual socio-demographic characteristics. Our mainnterest, however, concerns the impact on the trustor’s decision ofariable x−j , the observed average number of tokens sent by his/hereighbours in the previous period.

Consequently, we consider the following hypothesis:

laim 1. ∂xi/∂x−j /= 0 Peer influence exists, senders modify theirehaviour as a result of their observations of other trustors’ behaviour.

However, an analysis of individual’s characteristics allows us toxtend our inspection to an additional aspect. In fact, starting from1), it can be shown that:

xLi − xI

i

∣∣ > 0 if

(∑3

j�j − 2ˇ�i

)< 1 (4)

1 See Falk and Ichino (2006) for a similar theoretical framework.2 In Section 2.2, we discuss the determinants of trusting behaviour and their

xperimental measures.3 In our experimental design (see next section) the convergence value of xi ∈ [0,

00], therefore 0 ≤ |�| < 0.5.

etiot

ae

ep

rimental Economics 53 (2014) 97–110 99

here xLi

is the number of tokens sent by individual i as a mem-er of neighbourhood L and xI

iis the number of tokens the same

ndividual sends in the absence of social interaction. Notice that4) indicates that the specified difference in behaviour dependsn the sociality factor, (ˇ), and also on the distance between thendividuals’ preferences for trust within the neighbourhoods. Themaller the differences in the preferences, i.e., the more homoge-eous the neighbourhood, and the stronger the effect of socialityn the individual. We therefore consider a second hypothesis4:

laim 2. Individuals’ preference for trust (�i) influences the effectf sociality within neighbourhoods. Specifically, the smaller the dis-ance between trusting attitudes in the neighbourhoods, the higherhe impact of sociality.

.1. The experimental design

The experimental design was based on a standard trust gamehere two agents acted sequentially. Player A was given an endow-ent of 600 experimental tokens at the beginning of each period

nd was required to decide how much of this endowment would beept and how much he/she would transfer to player B, who receivedhe amount sent by A multiplied by a factor, � = 3.5 Player B thenad to decide how many tokens he would send back to Player A, andhe game ended. The stage game was repeated for twenty periods.

The relevant feature of the experimental design was that begin-ing with the sixth period and onward, each sender was placed

n a neighbourhood of three senders and was able to observe thehoices (i.e., the number of tokens sent in the previous period)f the other members of the neighbourhood. The assignment ofach subject to a specific neighbourhood was randomly decidedy the computer. The trust game was divided into two parts: a NO-

NFO part (periods 1–5) and an INFO part (periods 6–20).6 Once therust game is repeated, correlated effects can arise. If each subjectecides to reduce his/her own trust (and contribution) from oneeriod to the next, a correlation in the observed actions may beenerated. As in Fortin, Lacroix, and Villeval (2007), we determinehat an initial 5-period phase completes the individual learningrocess and stabilises the behaviour of the subjects participating

n the experiment.7

Both in the instructions (see Appendix A) and in the infor-al introduction to the sessions presented by the experimenters,

enders were aware that (1) the trustors’ payoffs were uncorre-ated, (2) the neighbours had the same information about the game,3) the initial endowments were equal for all subjects in all neigh-ourhoods.

5 The experimental exchange rate was 1 token = 1 Euro cent. Player B had nondowment and Player A was conscious of the different allocations (see Instruc-ions). The amount sent by A and returned by B in the Trust Game is sensitive to thenitial endowment distributions (Cox, 2004). Here, we do not discuss the issue sinceur interest is in examining the effect of social interaction on trust. In our design ofhe Trust Game, the recipient has no initial endowment.

6 As in Fortin, Lacroix, and Villeval (2007), we introduced a NO-INFO phase tollow subjects to learn the rules of the game and avoid confusion between learningffects and social influence effects.7 Moreover we regress – as a further check to exclude time effects – the choices of

ach sender against players who were not in his/her neighbourhood. Our estimationrocedures are explained in Sections 4.1 and 4.2.

1 Expe

emenisoiotd(fdaloaohiteieArttrgtiif“t

isteota

to

ie

asoamVs

wB

pt

eitcb

ttwed

dewp

rfoi9tmotS

2

tgtt“te

iBbbbo

00 L. Luini et al. / Journal of Behavioral and

ach period, and there was only a small probability that they wouldeet the same partner again. Specifically, we ran 4 sessions of the

xperiment, 3 with 24 players, and the last with 18 players. Theumber of times each sender was matched with the same recip-

ent was given by the fraction N/(T/2), where N is the number ofubjects, T is the number of periods and 2 represents the numberf roles within the experiment.8 Furthermore, as illustrated in thenstructions, senders were aware that they would be paid for onlyne period, randomly extracted at the end of the experiment. Bothhe matching protocol and the random payment procedure wereesigned to separate social influence from “observational learning”Manski, 2000). Once players were placed in a neighbourhood, inact, the convergence of behaviour could be explained by both airect effect of the social interaction on the individual’s behaviournd on the indirect effect that sociality had on the individual’searning process. As an example, if senders could observe the levelf reciprocity of player B, then the imitation of the neighbour’sctions might be motivated by strategic reasons such as reputationr imitation of the best observed strategy. Even when A playersave no information about the level of reciprocity of B players, this

ndirect effect may be present if the actions of the neighbours affecthe belief of the individuals regarding the reciprocity of the B play-rs. Our experimental design addresses this point according to fourmportant aspects: (1) we provide no information about the B play-rs; (2) we implement the random matching procedure; (3) we pay

players for only one period (i.e., for only one B player), which isandomly selected at the end of the session; (4) trustors learn abouthe actual number of tokens returned in each period only at the endhe session (i.e., at the payment stage).9,10 Our experimental designeproduces a one-shot trust game, where the repetition of the sin-le stage decision has the only purpose to enhance convergence. Tohe best of our knowledge, our design is the first among the papersn this field (see Section 1) that explicitly considers the direct andndirect effects of sociality and the first to isolate social influencerom observational learning. Accordingly, our senders were likecard players” facing one opponent, but they were able to observehe other players’ moves at different tables.

Overall, the experiments were composed of four parts (includ-ng the trust game), which were randomly presented to theubjects.11 The three sections of the experiment, in addition to therust game, were devoted to measuring �i, the individual’s pref-rence for trust, and they consisted of: (1) a questionnaire, (2) ane shot dictator game and (3) a section where participants hado indicate their choices in ten lotteries (Holt and Laury, 2002; seelso Eckel and Wilson, 2004).

In the dictator game, players A had the role of dictators. Dic-ators and respondents did not receive any initial endowment,nly the dictators – at the beginning of this specific section of the

8 The exact number of times they would meet the same B player was not reportedn the instructions, but how the matching procedure had been organised wasxplained by the experimenters at the beginning of the sessions.9 We thank a referee for her useful suggestions on this aspect of the design. There

re experimental designs (Servatka, 2009) that separate observational learning andocial influence by measuring the latter across groups (i.e., subjects in group Abserve the behaviour of subjects in group B) rather than within groups, therebyvoiding the correlation of beliefs. We preferred the alternative form of measure-ent (used also in Falk, Fishbacher, and Gaechter, (2013), and in Fortin, Lacroix, andilleval, 2007) because we believe that, in the former types of designs, the effect ofocial influence is diminished and could be confused with a “social history” effect.10 We thank A. Chaudhuri for his useful comments on this specific point. In otherords, trustors were unable to assess how many tokens had been returned by the

players in each period.11 Each part was introduced with specific instructions that appeared on the com-uter screen (see Appendix A), and the order in which the subjects participated inhe different stages was randomly set by the computer programme.

sCa

a

btdaC

(s

spg

rimental Economics 53 (2014) 97–110

xperiment – were endowed with 200 tokens. For the dictators,n the payment stage, the total profit was calculated subtractinghe amount of tokens they had sent to the anonymous respondent;onversely the profit of the respondents (B players) was obtainedy adding the tokens sent by the dictators.

The questionnaire12 contained three types of questions. Oneype of questions was related to the subject’s demographic charac-eristics; in another set of questions (as in Fehr, 2009), the subjectas asked to define his/her attitude towards risk and his/her own

xpectation of trustworthiness; such questions are reported andiscussed, also in light of previous studies, in Section 2.2.

At the beginning of the session, participants were randomlyivided into two groups (players A and B) and the recipients (play-rs B) participated in all parts of the experiment; however, theyere not placed in neighbourhoods from the sixth to the twentietheriods.

All subjects were paid at the end of the experiment. Theyeceived a show up fee and were paid on the basis of their per-ormance in the dictator games, while they were paid only forne period (chosen randomly at the end of the session) for thenvestment game and the lotteries. The sample was composed of0 subjects (45 senders and 45 recipients), the average payoff ofhe participants was nearly D 19 and each session lasted approxi-

ately an hour and a half. The experiments included four sessions,ne of which was conducted in 2009 at the University of Siena andhree of which were conducted in May 2010 at the University ofalerno.

.2. Measuring individual preferences for trust (�i)

In the experimental trust literature, the decision of the sendero send tokens to an anonymous recipient in a one-shot trustame has been explained on the basis of two main motivations:he expectation of monetary returns (or expected reciprocity) andhe unconditional desire to be kind to another human being (e.g.,unconditional kindness”, “warm glow”, “altruism”). Both motiva-ions are mixed every time a subject decides to send part of thendowment to an anonymous recipient.13

Many authors also claim that the individual’s attitude to trusts highly correlated with his/her risk preferences (see Ashraf,ohnet, and Piankov, 2006; Cox, 2004; Fehr, 2009). The relationshipetween trust and risk attitude, however, is often questioned on theasis of the assumption that trust is a primitive of the individual’sehaviour and the propensity to send tokens to recipients dependsn the willingness of the sender to establish a profitable relation-hip with the recipient, especially when � = 3 (Chaudhuri, 2009;haudhuri and Gangadharan, 2007), and it cannot be compared to

risky choice.There are several papers that aim at measuring each component

nd separating the two basic motivations.Here, the main research question is to assess whether trust can

e affected by contagion while considering the individual’s atti-ude towards trust. Accordingly, in each section of the experimental

esign, we focus on measuring the individual’s preferences anddopt a within-subject design (Ashraf, Bohnet, and Piankov, 2006;haudhuri and Gangadharan, 2007).14

12 A copy of the questionnaire is available upon request.13 On the determinants of trusting behaviour, see Ashraf, Bohnet, and Piankov2006), Chaudhuri and Gangadharan (2007), Cox (2004), and Fehr (2009), for exten-ive analyses.14 Ashraf, Bohnet, and Piankov’s (2006) experimental work focuses on the mea-urement of the different components of trust; Chaudhuri and Gangadharan, 2007’saper focuses on the relationship between trusting/reciprocating behaviour andenerosity, analysing also the role played by gender. Both paper, however, use a

Experimental Economics 53 (2014) 97–110 101

t

temteBStwtrii

archnte

sdtiitotga

teet

si

WcWott

o

aeolqritc

0

1

2

3

4

5

6

0 1 2 3 4 5amou

nt s

ent

by t

rust

or in

the

5th

per

iod

(in

100

toke

ns)

average amoun t sent by the neighb ours in the 4th per iod(in 100 tokens)

sta(u(

ofih

3

3

tsiTtois

L. Luini et al. / Journal of Behavioral and

First, expected trustworthiness is measured using the followingwo questions:

“Generally speaking, would you say that you can trust most peo-ple?” “Generally speaking, would you say that one can never be toocareful?”.15 Trustors’ responses were collected on a scale 0–10scale (higher values indicate greater trustworthiness).

The experimental literature contains several methodologieshat have been used to elicit the trustor’s belief regarding thexpected behaviour of the recipient. More specifically, experi-entalists have often adopted a direct measure of the expected

rustworthiness by asking the subject to express his/her ownxpectation about the recipient’s trustworthiness (see Ashraf,ohnet, and Piankov, 2006; Costa-Gomes and Weizsäcker, 2008).16

apienza, Toldra, and Zingales (2007) run a modified trust gameo extrapolate the “belief component” from the trustor’s actions,hich are also affected by his/her own generosity and risk atti-

ude. Their main finding is that “the sender’s expectation of theeceiver’s trustworthiness is a good predictor of the quantity sentn the trust game and it is highly correlated with the trust questionn WVS” (p. 3).17

In our design, we adopted the answers to the survey questionsbove as a proxy of expected reciprocity for two reasons. First, weun the experiments in two Universities located in two economi-ally and socially different Italian regions; thus our sample has aigh degree of heterogeneity with respect to the social and eco-omic backgrounds (see footnote 17). Second, we wanted to avoidhe confusion that a direct elicitation may induce in a complexxperimental design.

More conventionally, as stated in Section 2.1, we measuredome preference characteristics of the senders’ behaviour by usingictator games, lotteries and self-reported measures of the risk atti-ude. With regards to the individual level of altruism and generosityncorporated in the utility function of the individual, several stud-es (Cox, 2004; Ashraf, Bohnet, and Piankov, 2006) have consideredhe behaviour of dictators as a proxy of the generosity componentf the trustors’ decision in the trust game. Therefore, in modellinghe sender’s choices, we used the data from the one shot dictatorame described above as a measure of individual generosity (Eckelnd Grossman, 1998).

Finally, we selected two measures of risk based on both ques-ionnaire and laboratory data. In the former case, we used an

xperimentally validated measure of risk preference (Dohment al., 2005; Fehr, 2009) that is based on a question derived fromhe German socio-economic panel (GSOEP): “Are you, generally

imilar experimental methodology (the within-subject protocol) that we replicaten the present paper.15 Subjects were not paid to answer the questionnaire. The similar question in the

VS is: “Generally speaking, would you say that you can trust most people or that onean never be too careful?”. WVS responses are usually collected on a 2-point scale.

e adopted two questions on a 0–10 scale in order to provide higher variability inur data; the indicator measuring trustors’ expected trustworthiness, used later inhe empirical analysis (i.e. see Sections 3.2 and 4), has been obtained by averaginghe scores reported on both the questions.16 In some cases, to improve the accuracy of the prediction, subjects are rewardedn the basis of the success of the expectation.17 Specifically, the authors find that the subjects’ answers to the questionnaire,nd their reported beliefs, are correlated when the individuals are calculating thexpected amount returned, especially if they are sending a large amount of money;therwise, for a small amount of money sent, the belief reflects the “anticipatedevel of retaliation rather than the general level of trust. This suggests that the WVSuestion is a good measure of the expectation component of trust in economically-elevant situations” (p. 3). In their concluding notes, the authors also stress that,n order to assess the exact nature of this correlation, it is important to evaluatehe level of homogeneity of the population that is the object of the study, as theorrelation tends to be significant in heterogeneous samples.

h

cdstiss

es

(4oftmatLH

Fig. 1. Scatter plot – 5th period.

peaking, a person who is fully prepared to take risks, or do you tryo avoid taking risks?”. The respondents answered this question onn 11-point Likert Scale ranging from 0 (very risk averse) to 10very risk-seeking). In the latter case, we measured risk preferencessing the well-known lottery method suggested by Holt and Laury2002).

It must be specified that, in the empirical analysis in Section 4,ur measures of individual risk preference will rely only on the datarom the questionnaire in that (i) lottery choices are inconsistentn 10 out of 45 cases18 and (ii) our questionnaire data provide aigher variability in risk aversion across the different observations.

. The trust game: the results

.1. A descriptive analysis

In the trust game, each player had 600 tokens in each round;he average number of tokens sent by the trustors varies con-iderably across the 15 neighbourhoods, from a minimum of 15n the 1st neighbourhood to 360 in the 9th neighbourhood (seeable 1). Furthermore, some neighbourhoods exhibit decreasingrust across periods (i.e., neighbourhoods 1, 2, and 13), while inther cases, trust increases (i.e., neighbourhood 12 and very slightlyn neighbourhoods 6 and 9). In the other cases, trust does not varyignificantly or it follows an up and down pattern (i.e., neighbour-oods 3, 10 and 15).

Our main interest in the trust game concerns the presence ofontagion and, in this respect, we find some first descriptive evi-ence in Figs. 1–5. Figs. 1 and 2 plot the average number of tokensent by each trustor in the 5th period and in the 20th period, respec-ively, as a function of the average amount sent by their neighboursn the previous period. In the absence of contagion, these graphshould fluctuate around 0 (imitation is not expected in rounds 1–5ince the trustors do not observe their neighbours).

In Fig. 1, the data points are spread out, as in a circle, and, asxpected, the “trend line” has a negative slope equal to −0.1, nottatistically significant (t-stat = −0.507). In Fig. 2, the data points are

18 In the lottery game, our subjects make 10 decisions between a safe optionoption A) and a risky one (option B). The risk neutral choice pattern consists of

safe choices (when the probability of a high payoff for both the safe and the riskyption is low) followed by 6 risky choices (when the probability of a high payoffor both the safe and the risky option increases to 10/10). When individuals switcho the risky option before the fourth choice, they are considered risk seeking (sym-

etrically, they are considered as risk averse when they switch to the risky optionfter the fourth choice). We observe 10 individuals making inconsistent choices inhat they switch back from the risky to the safe option later in the game. Holt andaury (2002, p. 8) found similar inconsistencies in their data. Main disadvantages ofolt and Laury tables are reviewed in Maier and Rüger (2010).

102 L. Luini et al. / Journal of Behavioral and Experimental Economics 53 (2014) 97–110

Table 1Amounts of tokens sent by the trustors in each neighbourhood (std. dev. in parentheses).

(a)

Period Neighbourhoods

1 2 3 4 5 6 7 8

1 70 (121) 133 (58) 83 (70) 167 (152) 200 (115) 160 (79) 68 (114) 93 (75)2 53 (84) 235 (229) 50 (80) 133 (115) 150 (115) 203 (172) 108 (183) 110 (36)3 50 (86) 267 (208) 63 (109) 67 (115) 200 (173) 272 (199) 64 (109) 122 (81)4 50 (87) 216 (247) 87 (81) 67 (115) 90 (79) 288 (191) 38 (46) 103 (110)5 53 (84) 60 (53) 42 (72) 67 (115) 245 (236) 249 (217) 143 (231) 96 (81)6 7 (11) 50 (50) 53 (92) 167 (58) 150 (86) 183 (146) 50 (87) 70 (26)7 2 (3) 43 (51) 40 (69) 117 (104) 100 (50) 270 (199) 83 (104) 70 (29)8 3 (6) 60 (53) 53 (92) 135 (112) 98 (92) 210 (115) 40 (69) 73 (21)9 0 67 (58) 47 (81) 133 (58) 123 (155) 103 (61) 133 (231) 67 (29)10 0.3 (0.6) 50 (50) 60 (109) 133 (153) 97 (55) 165 (119) 70 (121) 90 (53)11 0.6 (0.1) 73 (23) 0 133 (115) 82 (55) 178 (68) 72 (107) 83 (29)12 1.2 (2.1) 10 (17) 0 33 (58) 100 (100) 196 (089) 93 (162) 70 (35)13 1 (1.7) 0 0 167 (153) 133 (153) 219 (70) 33 (58) 90 (36)14 3.3 (5.7) 33 (58) 40 (69) 83 (29) 117 (29) 280 (108) 167 (289) 70 (26)15 0.3 (0.6) 50 (50) 44 (75) 50 (87) 133 (116) 200 (199) 27 (46) 117 (65)16 0.3 (0.6) 27 (46) 0.3 (0.6) 133 (153) 67 (58) 233 (58) 33 (57) 97 (15)17 1.3 (1.5) 37 (55) 37 (55) 133 (115) 117 (126) 283 (161) 70 (121) 78 (68)18 2 (2) 20 (35) 43 (75) 33 (58) 33 (29) 243 (172) 66 (114) 68 (58)19 1.7 (1.5) 33 (58) 53 (92) 199 (105) 84 (76) 317 (202) 12 (20) 67 (47)20 0.7 (1.1) 50 (50) 50 (87) 0.7 (0.6) 121 (156) 254 (215) 6 (10) 63 (23)

Tot. 15 76 42 108 118 225 68 85

(b)

Period Neighbourhoods

9 10 11 12 13 14 15 Tot.

1 263 (182) 85 (99) 200 (264) 233 (266) 173 (283) 83 (104) 67 (115) 139 (143)2 283 (182) 117 (159) 266 (252) 250 (229) 170 (286) 100 (173) 50 (50) 152 (162)3 260 (208) 240 (242) 233 (225) 257 (240) 168 (287) 133 (115) 17 (28) 161 (172)4 308 (166) 160 (208) 83 (333) 200 (265) 200 (264) 67 (115) 33 (58) 131 (161)5 346 (147) 67 (41) 267 (252) 320 (270) 167 (289) 100 (173) 136 (229) 164 (188)6 383 (065) 48 (45) 167 (289) 350 (132) 233 (231) 67 (115) 101 (172) 141 (157)7 386 (11) 90 (98) 133 (231) 383 (125) 233 (231) 33 (58) 69 (113) 139 (161)8 403 (0.15) 90 (0.95) 100 (1.73) 383 (1.04) 200 (2.65) 33 (0.58) 100 (1.00) 136 (151)9 393 (40) 37 (23) 233 (252) 333 (126) 200 (265) 33 (58) 100 (173) 125 (150)10 356 (4.9) 123 (154) 150 (260) 283 (225) 167 (289) 33 (58) 69 (114) 131 (152)11 417 (72) 110 (85) 267 (252) 367 (189) 200 (265) 50 (87) 33 (58) 128 (158)12 427 (47) 197 (184) 250 (250) 383 (161) 200 (265) 33 (58) 67 (115) 138 (172)13 370 (118) 123 (139) 200 (264) 450 (50) 200 (265) 50 (87) 33 (58) 142 (176)14 380 (46) 177 (143) 250 (250) 383 (202) 33 (58) 33 (58) 35 (57) 140 (173)15 333 (136) 137 (120) 133 (126) 236 (176) 0 50 (87) 67 (115) 115 (140)16 358 (52) 68 (76) 183 (275) 433 (115) 33 (58) 33 (58) 34 (57) 110 (143)17 435 (64) 175 (239) 250 (250) 386 (79) 33 (58) 50 (87) 35 (57) 136 (167)18 419 (84) 50 (87) 233 (250) 299 (171) 33 (58) 33 (58) 67 (115) 111 (154)19 335 (128) 123 (196) 233 (250) 380 (208) 0 0 33 (58) 136 (174)20 330 (72) 220 (252) 167 (288) 458 (370) 7 (11) 7 (11) 67 (115) 124 (172)

Tot. 360 128 205 338 139 62 61 135

0

1

2

3

4

5

6

0 1 2 3 4 5 6

.am

ount

sen

t by

tru

stor

in t

he 2

oth

peri

od(i

n 10

0 to

kens

)

average amount sent by the neighbours in the 19th period(in 100 tokens)

Fig. 2. scatter plot – 20th period.

0

0.5

1

1.5

2

2.5

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

amou

nt se

nt b

y tr

usto

r (in

100

toke

ns)

periods

trustor 1

trustor 2

trustor 3

Fig. 3. Neighbourhood 1.

L. Luini et al. / Journal of Behavioral and Expe

00.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

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

amou

nt se

nt b

y tr

usto

r (in

100

toke

ns)

periods

trustor 9

trustor 8

trustor 7

Fig. 4. Neighbourhood 3.

0

1

2

3

4

5

6

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

amou

nt se

nt b

y tr

usto

r (in

100

toke

ns)

trustor 25

trustor 26

trustor 27

tsh

iountp1tsain

t(G

Rbcti

c(a

3

gStmstTw

sttedftbas

phieta

weEapte

x

ttf

tngiat(toprely on our survey measures of trustworthiness (see Section 2.2).When we compare trustors’ expectation of trustworthiness and

Periods

Fig. 5. Neighbourhood 9.

ighter (particularly in the lower left corner of the diagram) and thelope of the trend line is 0.6 (t-stat = 3.362), consistently with theypothesis of a positive peer effect.

Figs. 3–5 display the number of tokens sent by each trustor dur-ng the 20 rounds of the game in some neighbourhoods19; theyffer interesting insight in that they document different individ-als’ reactions to observing other trustors. Fig. 4, for example, doesot report evidence of imitation in neighbourhood 3 since similarrends in individual trust are observed before and after the fiftheriod (see in particular, trustors 8 and 9). On the contrary, trustor

in neighbourhood 1 (see Fig. 3) sends large amounts to his/herrustee in the first five rounds of the game, but his trust collap-es to “0” after observing his/her neighbours’ behaviour. Finally,s can be seen from Fig. 5, someone else exhibits increasing trustmitating his/her neighbours’ trusting behaviour (i.e., trustor 27 ineighbourhood 9).

Overall, the data20 indicate heterogeneity in individuals’ atti-ude to imitate peers, which is consistent with previous studiesGlaeser, Sacerdote, and Sheinkman, 1996; Falk, Fishbacher, andaechter, 2013).

esult 1. We observe that the majority of subjects are affectedy peers’ actions in periods 6–20. However, we can state that twolasses of subjects exist: those whose behaviour is influenced by

he behaviour of their neighbours and those whose behaviour isndependent of the behaviour of others.

19 The data for the other neighbourhoods are available on request.20 Overall, a small number of individuals (only 7 out of 45 individuals) display noontagion effects at all (i.e., they send “0” tokens each time). Most neighbourhoodsat least 8 out of 15) display contagion effects, with an increasing trust in some and

decreasing trust in others.

t

ot

o

rimental Economics 53 (2014) 97–110 103

.2. Looking at social preferences, beliefs and behaviour

Table 2 summarises information on subjects’ risk attitude,enerosity, and beliefs about trustees’ behaviour. As reported inections 2.1 and 2.2, the measure of risk preference derived fromhe answers to the GSOEP question (Dohmen et al., 2005), while, as

easures of subjects’ generosity, we indicate the amounts of tokensent by each player to his/her anonymous receiver during the dic-ator game (Eckel and Grossman, 1998). According to the data inable 2, sampled individuals are mainly selfish and risk averse forhatever index we consider.

The last column in Table 2 also reports the amounts of tokensent by the trustors during the first five rounds of the game: morerusting subjects are expected to send higher amounts of tokenso their recipients. Such measure represents an additional way toxamine the exogenous propensity to trust in that the individuals’ecisions in the NO-INFO part of the trust game are independentrom the others’ choices.21 To this regard, it is interesting to notehat most individuals classified as risk loving or generous on theasis of the questionnaire and of the dictator game, respectively,re also “more trusting” in the NO-INFO part of the trust game (i.e.,ee neighbourhoods 9, 12).

When comparing individual profiles in Table 2 and the observedatterns during the game (in Table 1), we find that the neighbour-oods characterised by a majority of generous and less risk averse

ndividuals – i.e., neighbourhoods 6 and 9 – exhibit higher lev-ls of trust, as well as increasing trust; the opposite evidence ishen reported for the neighbourhoods composed of selfish and riskverse individuals – i.e., neighbourhoods 1, 2, and 13.

In order to verify whether the observed patterns are significant,e apply a model of convergence first suggested by Ashenfelter

t al. (1992) and used by Noussair, Plott, and Riezman (1995) andckel and Grossman (2005). This model is particularly appropri-te to investigate questions about the initial levels of trust of thelayers, the variation in the levels of trust during the game, andhe asymptotic level of trust for each neighbourhood. The basicstimating equation is given by:

it = ˇ1iDi

(1t

)+ ˇ2

(t − 1

t

)+ u (5)

The dependent variable xit is the amount of tokens sent byrustor i at time t; the dummy variable Di takes a value of 1 forrustor i, and 0 otherwise. The origin of the convergence processor trustor i is given by ˇ1i; ˇ2 is the asymptotic level of trust.

Table 3 (Model 1) indicates that the convergence points (sta-istically significant in 14 cases out of 15) estimated for theeighbourhoods 6, 9 and 12 are consistently above the conver-ence points reported for the other neighbourhoods.22 The datan the lower section of Table 3 (Model 2) differ for the inclusionmong the regressors of a dummy variable, “Rounds 6–20”, con-rolling for the periods from 6 to 20. In some neighbourhoodsi.e. neighbourhoods 1, 2), “Rounds 6–20” significantly decreaseshe amounts of tokens sent by trustors to their recipients; thepposite is true in neighbourhoods 9 and 12, where an upwardattern is observed.23 Trustors’ belief about trustee’s reciprocity

rustors’ behaviour during the game, in some cases we find a good

21 We thank Ananish Chaudhuri for his suggestion on this point.22 In order to control for serial correlation and heteroscedasticity, the methodf correction employed is Kmenta (1986) cross-sectionally heteroskedasticity andime-wise autoregressive model.23 In some neighbourhoods, however, (i.e. neighbourhood 13) convergence isbserved in the last rounds of the game (more details are available on request).

104 L. Luini et al. / Journal of Behavioral and Experimental Economics 53 (2014) 97–110

Table 2Exogenous measure of individual preferences: Individual generosity, risk preferences, expected trustworthiness and trust in the NO-INFO rounds.

Neighbourhood Trustor Generosity Expectedtrust. Risk preference Tokens in NO-INFO PART

Question Lottery methoda

11 50 2 5 n.c. 1622 0 5 4 Extremely risk averse 43 5 1 3 Very risky averse 0

24 10 3 5 Risk averse 1075 0 4.5 4 Highly risk averse 1006 0 7 4 Risk averse 340

37 0 3.5 3 Very risk averse 508 0 3 5 Very risk averse 09 120 5.5 7 Highly risk loving 145

410 0 5.5 5 Risk neutral 20011 0 3 3 Risk averse 6012 50 6 5 Risk neutral 40

513 50 1.5 2 n.c. 20014 0 3 4 Risk neutral 6115 100 4 7 n.c. 280

616 1 3 6 Risk averse 43017 120 6.5 6 Risk loving 14418 60 6 3 Risk averse 130

719 0 4 2 Very risk averse 11220 0 4 8 Risk averse 021 30 5 10 Slightly risk averse 242

822 50 3 5 n.c. 6723 30 7 7 Slightly risk averse 5624 70 3.5 2 Very risk averse 192

925 110 5 8 Very risk averse 27026 150 5.5 3 Very risk averse 46827 80 3 6 Risk averse 141

1028 50 5 8 n.c. 8129 50 7.5 10 n.c. 30030 50 5 5 n.c. 20

1131 0 5 2 Risk averse 032 200 2.5 7 Slightly risk averse 27033 100 2.5 5 Risk averse 400

1234 200 6 10 n.c. 48035 50 3.5 7 n.c. 26036 20 4 4 n.c. 16

1337 50 2 4 Very risk averse 2738 0 5.5 3 Risk averse 039 200 3 5 Very risk averse 500

1440 0 3 7 Risk neutral 041 50 1.5 4 Very risk averse 5042 100 4 5 Highly risk averse 240

1543 10 5.5 4 Risk averse 11644 50 3 6 Highly risk averse 17045 0 5 2 Risk averse 0

Mean 50.35 4.14 5.11 147Median 50 4 5 116

Generosity: number of tokens sent by the players to their anonymous receivers during the dictator game. Expected trust.: higher values (on a scale 0–10) indicate moret e mort

c(wg

asi

R

rustworthiness. Risk preference (question): higher values (on a scale 0–10) indicatrustor to his/her recipient during the first five rounds of the trust game.

a n.c.: no consistent choices.

orrespondence between attitudinal and behavioural measures

for example, see trustor number 9 in Fig. 4), while in other cases,e observe such a correspondence only in the first rounds of the

ame (see trustor 27 in Fig. 5, for example).24

24 In 11 out of 45 cases, we do not find consistency between questionnaire datand trustors’ behaviour (for example, trustors 12, 18, 23, 38 and 43, reported highcores on the survey questions, but they sent 0 or few tokens to their recipientsn many rounds; opposite findings are reported for trustors 1, 24, 32, 33, 39 and

gora

4p

e risk loving. Tokens in the NO INFO PART: average number of tokens sent by each

esult 2. Some neighbourhoods, characterised by a majority ofenerous and more risk loving individuals, exhibit higher levelsf trust, as well as increasing trust; the opposite evidence is then

eported for some neighbourhoods composed of selfish and riskverse individuals.

4). In most of these cases, individual behaviour is consistent with individual socialreferences, however.

L. Luini

et al.

/ Journal

of Behavioral

and Experim

ental Econom

ics 53

(2014) 97–110

105

Table 3Starting and convergence points (std errors in parentheses).

(a)

Variables Neighbourhoods

1 2 3 4 5 6 7 8

Model 1Starting point trustor i, i = 1 (std. errors) 206.698** (36.07) 121.561** (33.973) 97.474** (23.355) 329.262** (82.199) 274.738** (48.079) 517.589** (128.202) 9.517 (10.579) 76.129** (30.144)Starting point trustor i, i = 2 (std. errors) 7.062** (1.76) 136.258** (34.237) −8.349 (17.873) 114.407◦ (83.853) 57.673 (49.196) 67.714 (100.458) −7.270 (6.685) 33.831◦ (21.486)Starting point trustor i, i = 3 (std. errors) 0.004 (0.64) 394.271** (121.749) 264.675** (75.798) 39.344 (65.121) 479.625 ** (108.44) 80.645* (50.384) 473.262** (147.465) 261.797** (36.962)Convergence point (std errors) 0.018 (0.056) 57.304** (8.208) 9.958** (4.102) 92.990** (14.814) 66.082 ** (8.667) 195.873** (8.552) 6.012** (1.531) 68.537** (3.95)Log-lik. function −55.089 −331.573 −301.733 −354.974 −343.707 −358.049 −262.978 −294.863

Model 2ˆ

Convergence point 16.854** (0.817) 111.529** (22.113) 13.890 (11.614) 2.220 (37.338) 132.331** (20.983) 240.059** (24.111) 10.247** (4.593) 117.955** (7.191)Rounds 6–20 −17.366** (0.72) −58.559** (20.439) −4.029 (10.736) 77.731** (34.527) −61.875** (19.39) −34.617 (22.286) −3.571 (4.244) −43.98** (6.647)Log-lik. −166.480 −330.51 −301.529 −351.114 −340.371 −358.273 −264.921 −283.981

(b)

Variables Neighbourhoods

9 10 11 12 13 14 15

Model 1Starting point trustor i, i = 1 (std. errors) 108.099◦ (80.632) 80.109◦ (51.329) −175.667* (101.654) 538.500** (39.829) 12.814 (37.384) −7.372 (6.163) 12.979 (19.079)Starting point trustor i, i = 2 (std. errors) 501.134 ** (32.741) 504.619** (166.298) 378.221** (176.943) 97.432 (96.637) −35.022 (39.254) 65.764** (33.289) 375.237** (117.518)Starting point trustor i, i = 3 (std. errors) −0.1034 (61.532) −40.941 (46.275) 500.910 ** (155.06) 278.255* (150.333) 93.37** (21.12) 435.853** (118.146) −13.596 (9.610)Convergence point (std errors) 410.411** (7.192) 66.988** (9.604) 137.937** (13.462) 425.817** (9.197)) 46.61** (9.23) 6.008** (1.339) 10.849** (1.842)Log-lik. function −335.068 −341.989 −385.310 −358.536 −347.42 −281.207 −279.544

Model 2ˆ

Convergence point 308.97** (18.025) 91.422 ** (27.458) 171.575** (38.655) 380.551** (25.015) 29.925 (26.449) 18.943** (6.096) 10.535 ** (5.319)Rounds 6–20 92.007** (16.664) −21.431 (25.355) −28.382 (35.648) 40.133* (23.119) 17.926 (24.447) −9.513* (5.634) 0.587 (4.916)Log-lik. −332.227 −342.727 −385.039 −357.482 −348.954 −289.852 −280.169

* Statistically significant at 10% level;** Statistically significant at 5% level;◦ Statistically significant at 20% level.ˆ Starting points included in model 2, but not reported.

106 L. Luini et al. / Journal of Behavioral and Expe

Table 4Tobit estimates of peer effects: fixed effects (lower and upper limits:0.6).

6th–20th periods 6th–20th periods

Variablesa I IINeighbours’ action 0.322 (0.087)*** −0.0009 (0.0008)Log-lik. −731.085 −738.082N. of observations 675 675

Std. errors in (). Time effects included.a The dependent variable is the number of tokens sent by each trustor in the

round t. Neighbours’ action: average number of tokens sent by the neighbours inthe period t − 1.

***

4

4i

tober

ldfds2etciht

oraab

4

e

tobnaoeto

IaT

bn

rtievWbppaoea

ohdntgaebaAir

vtgaisas a mean between the two indexes – described in Sections 2.1 and2.2 – measuring trustor’s generosity and risk aversion respectively.

25 The variable “neighbours’ action” is the average number of tokens sent by thetwo neighbours in period t − 1. In preliminary estimates, we included the amountsof tokens sent by each neighbour in the previous round as regressors (instead of themean “neighbours’ action”) and we obtained similar results.

26 The results in Table 4 (as well the results in Table 6) are robust to the alter-native specifications of the model controlling for the amount sent by each trustorin the previous period (t − 1): the coefficient on the lagged variable is −0.039 (stderr. = 0.116) in the rounds 6–10, and it is equal to −0.094 (std err. = 0.069) in therounds 11–20.

27 We also test for spurious correlation in our data by regressing own action onneighbours’ actions during the 2nd–5th rounds and, as expected, we do not find sta-tistically significant peer effects (results not reported here but available on request).

28 Sacerdote (2001) measures peer effects among college roommates by regress-ing their own outcome (grade point average, or GPA) on peer outcomes. Randomassignment of the students to dorms and to roommates implies that all the room-mate’s background variables (i.e., ability affecting individual GPA) are uncorrelatedwith his/her own background characteristics. This makes it possible to measure areduced-form effect of student outcome on his/her roommate j’s background, thussolving the problem of simultaneity in the observed outcomes. SAT scores and highschool class rank are used as noisy measures for unobserved student ability.

29 The expected trustworthiness index is not included in the variable “own type”since the relative coefficient in Table 5 is statistically significant (at 10 % level)only in the first rounds of the game (and, in the rounds 6–10, the coefficient of

Statistically significant at 1% level.

. Estimating social influence

.1. Experimental and econometric issues in the analysis of socialnfluence

Our analysis of peer effect relies on a simple framework wherehe number of tokens sent by each subject to the trustee dependsn both the subject’s own propensity to trust and the observedehaviour of the other trustors. The null hypothesis of no peerffects predicts a statistically insignificant coefficient from theegression of the trustor’s own behaviour on peers’ behaviour.

The correct identification of peer influence poses several prob-ems, extensively addressed in Manski (1993). The experimentalesign, described in Section 2, has overcome these problems asollows: (i) the self-selection problem has been addressed by ran-omly allocating our subjects to groups in our experiments (for aimilar experimental procedure, see Falk, Fishbacher, and Gaechter,013; for observational data, see Sacerdote, 2001); (ii) contextualffects do not arise in the game because the interactions amongrustors are completely anonymous; (iii) the experiment avoidsorrelated effects because the subjects make their decisions indentical contexts; that is, the three subjects in each neighbour-ood are provided with the same budget, the same incentives andhe same information (Falk, Fishbacher, and Gaechter, 2013).

Estimating peer effects by regressing one’s own behaviour onthers’ actions could be still problematic, however, because of theeflection (or simultaneity problem). Our estimation procedure,s we will explain in more detail in the next section, explicitlyddresses both the problem of correlated effects and simultaneityias.

.2. The econometric model and the results

In this section we provide evidence of the existence of peerffects (Claim 1).

In our econometric model, the trustor’s action (the number ofokens sent to an anonymous recipient) in period t depends bothn the trustor’s own propensity to trust and on the observed neigh-ours’ actions (e.g. the average number of tokens sent by eacheighbour to his/her trustee in the previous period, t − 1). Over-ll, we consider the fifteen periods in which individuals actuallybserve their neighbours’ actions. The availability of panel datanables us to mitigate problems of unobserved (individual andime) correlated effects via the inclusion in the model specificationf both fixed and random effects (Hartmann et al., 2007).

Our estimates in Table 4 are based on a Tobit fixed-effect model.n column I, the coefficient estimated on the variable “neighbours’ctions” is 0.322, and it is statistically significant at the 1% level.he null hypothesis of no peer effects would predict no relationship

“wgi

rimental Economics 53 (2014) 97–110

etween his/her own action and others’ actions. The data reject thatull hypothesis.25,26

Correlated effects could arise, however, particularly withespect to time. For example, if all subjects decided to reduce theirrust from one period to the next, then we would find correlationn observed actions and we would misattribute it as a causal peerffect (i.e., correlated effects in trust games are likely to occur givenariations in experience and degree of learning during the game).e explicitly take this issue into account in column II of Table 4

y regressing the behaviour of each trustor on the behaviour of aair of trustors that is randomly drawn from the sample (not theair actually observed). If we observed a significant correlation inn agent’s actions then we would admit the existence of a spuri-us correlation in our data (see also Sacerdote, 2001). However, asxpected, we do not observe a statistically significant correlationmong individuals’ choices.27

Interpreting the coefficient obtained by regressing an agent’swn behaviour on trustors’ behaviour as a causal peer effect may,owever, still be misleading because of the simultaneity of theecisions (i.e., either a reflection or endogeneity problem). Theeighbours’ behaviour affects the trustor’s behaviour, which inurn, affects the neighbours’ behaviour (these problems are evenreater in small neighbourhoods; Moffitt, 2001; Krauth, 2002). Toddress this problem, following Sacerdote (2001), we consider asvidence of peer effects the finding of a significant correlationetween trustors’ decisions (number of tokens sent in each period)nd the (exogenous) propensity of his/her own neighbours to trust.n agent’s own preference for trust affects his/her decision, but

t can be a priori excluded from the decision of the others in theeference group.28

Therefore, we need a valid measure of the (exogenous) indi-idual propensity to trust. Taking into account previous studies onhe correlates of trust (see Section 2.2), we suggest that individualenerosity, risk aversion and expected trustworthiness are likely toffect the trustors’ behaviour. Having confirmed that this is the casen Table 5, we then construct exogenous measures of the propen-ity to trust for each player: the variable “own type”29 is obtained

expected trustworthiness”, statistically significant at 10% level, is not estimatedith the expected positive sign); one could argue that beliefs change during the

ame and that when trustors observe neighbours’ actions, they no longer explainndividual behaviour.

L. Luini et al. / Journal of Behavioral and Experimental Economics 53 (2014) 97–110 107

Table 5Tobit estimates of individual trust: random effects.

Variablesa 2nd–5th periods 6th–10th periods 11th–15th periods 16th–20thperiodsCoefficients Coefficients Coefficients CoefficientsI II III IV

Age −0.636 (0.096) −0.356*** (0.061) −0.391*** (0.100) −0.207*** (0.076)Sex (male = 0) 0.146 (0.396) −1.114 (0.214)*** −1.686*** −0.912*** (0.312)Generosity 1.721*** (0.303) 2.487*** (0.210) 2.727*** (0.505) 2.991*** (0.429)Risk preference 0.310*** (0.091) 0.451*** (0.059) 0.529*** (0.098) 0.586*** (0.092)Expected trustworthiness 0.169* (0.104) −0.110*(0.058) 0.030 (0.139) −0.027 (0.137)Log-lik −254.179 −266.817 −289.012 −282.317Restr. L. L. −297.775 −337.526 −333.525 −338.580n. observ.s 180 225 225 225

Std. errors in parentheses. Constant and period dummies included.a Generosity: index for trustor’s generosity derived from the dictator game (dummy = 1 if generous, 0 otherwise). The risk preference index is derived from the GSOEP

question (higher values indicate higher risk loving). The expected trustworthiness index is derived from the similar WVS questions (higher values indicate higher trustwor-t

tav

fit

vscs

scTwws“1amee

todst“

vwnrsd

c

ititp

ath

Rtd

ib

atvvr“etul“v

cto“

hiness).* Statistically significant at 10% level.

*** Statistically significant at 1% level.

A measure of each player’s neighbours’ exogenous propensityo trust – the variable “neighbours’ type”30 – is then included as

regressor in some of the results presented in Table 6 in order toerify whether the estimates are robust to simultaneity.

In fact, in Table 6, we address as evidence of peer effects thending of a significant correlation between trustors’ decisions andhe propensity to trust of his/her own neighbours.

We report separate estimates for different periods to elicit anyariations in peer effects during the twenty rounds. As age andex dummies were the only socio-demographic variables signifi-ant at least at the 10% level, they are included in the final modelpecification.

The coefficient on “neighbours’ actions” is large and statisticallyignificant only after the 6th round. As argued above, this coeffi-ient could be biased by the reflection (or endogeneity) problem.herefore, we consider the exogenous variable “neighbours type”,hich has a small and insignificant effect in columns III and VIII;hen the “neighbours’ action” variable is dropped from the model

pecification in columns IV and IX, the estimated coefficients forneighbours’ type” become statistically significant at least at the% level.31 This result is not surprising given that generosity andttitude to risk significantly predict individual decisions to trust;oreover the coefficient for this variable is not affected by the

ndogeneity problem. We interpret such findings as supporting thevidence regarding peer effects (see also Sacerdote, 2001).

Moreover, the estimates in columns V and X are based on addi-ional exogenous measures of the propensity to trust, respectively,f each trustor and of his/her neighbours32: “TR-NOINFO” is aummy variable equal to 1 when the mean amount of tokens

ent by the trustor to his/her trustee in the “NOINFO” part ofhe trust game is above the average in the sample, 0 otherwise;Num.TR-NOINFO” is the number of his/her neighbours classified

30 Firstly, for each neighbour, we construct a further variable – that is similar to theariable “own type” – measuring his/her propensity to trust; then, for each trustor,e construct the variable “neighbours’ type”, that is the mean between his/her twoeighbours’ propensity to trust. In preliminary estimation, the inclusion among theegressors of the indexes measuring each neighbour’s generosity and risk preferenceeparately (overall, 4 different variables instead of the “neighbours’ type” variable)oes not affect the main results in Table 6 (particularly after the 10th round).31 In preliminary estimates, we regressed trustors’ characteristics on neighbours’haracteristics and we did not report any significant correlation.32 These exogenous measures of the individuals’ propensity to trust have beenntroduced in Section 3.2 (see also Table 2 and note 21). It is interesting to note thathe variable “own type” relies on the evidence reported in Table 5 and, accordingly,t explicitly incorporates risk aversion; the classification of our trustors accordingo the variable “TR-NOINFO”, however, does not incorporate any measure of riskreferences (on this point, see also Section 2.2).

tpot

Rbe

a

uUdpf

s “TR-NOINFO”. Once again, we find a positive correlation betweenhe trustors’ decisions and the exogenous propensity to trust ofis/her own neighbours.

esult 3. We address as evidence of peer effects, robust to simul-aneity, the finding of a significant correlation between trustors’ecisions and the propensity to trust of his/her own neighbours.

A further aim is to understand whether and to what extentndividual preferences affect individual decisions to imitate others’ehaviour (Claim 2).

In this respect, in columns VI and XI, we focus on trustors’nd neighbours’ preferences: in the set of regressors, we includehe dummy variable “trusting”, that is equal to 1 when both theariables “own type” and “neighbours’ type” are above the meanalue in the sample (so that “more trusting” individuals – i.e. moreisk loving and/or more generous subjects – are grouped withmore trusting” neighbours); the variable “untrusting” is a dummyqual to 1 when both the variables “own type” and “neighbours’ype” are below the mean value in the sample. (so that “morentrusting” individuals are grouped with neighbours with simi-

ar characteristics). The excluded dummy variable indicates thatmore untrusting” players meet “more trusting” neighbours (andice versa).33

The coefficient reported for the “trusting” dummy variable (inolumn VI) indicates that more trusting people are more likely torust when they are grouped with more trusting neighbours. Thepposite is true when we consider the coefficient estimated for theuntrusting” dummy variable: more untrusting individuals reduceheir trust when they meet similar types. In the last rounds, therevailing trend is a reduction of trust, but once again, more gener-us and less risk averse individuals are more willing to trust whenhey observe trusting behaviour.34

esult 4. We observe that individuals are more likely to imitateehavioural rules that they recognise as homogenous with their

xpressed or hidden preference structure.

33 To be precise, the variable “own type” is below the mean value in the samplend the variable “neighbours’ type” is above (and vice versa).34 The coefficient on the variable “untrusting” is not statistically significant in col-mn XI, thus there are not significant differences with the “excluded” variable.nfortunately, the low variability of the variable TR-NOINFO across the observationsid not permit us to further investigate whether the peer effect were stronger wheneople meet similar types (e.g. as reported in Table 2, there are no neighbourhoodsormed by three trustors classified as “TR-NOINFO”, or otherwise).

108 L. Luini et al. / Journal of Behavioral and Expe

Tab

le

6To

bit

esti

mat

es

of

pee

r

effe

cts:

ran

dom

effe

cts.

Var

iabl

esb

2nd

–

5th

per

iod

sa6t

h–1

0th

per

iod

sa11

th–2

0th

per

iod

sa

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

Coe

ffici

ents

I

II

III

IV

V

VI

VII

VII

I

IX

X

XI

Age

−0.0

61

(0.1

01)

−0.2

21**

*(0

.036

)

−0.2

41**

*(0

.047

)

−0.2

54**

*(0

.047

)

−0.0

65

(0.0

64)

−0.0

312

(0.0

51)

−0.0

72*

(0.0

42)

−0.0

71

(0.0

49)

0.01

7

(0.0

77)

0.02

8

(0.0

45)

−0.1

77

(0.0

55)**

*

Sex

(mal

e

=

0)

−0.8

71**

(0.4

48)

−1.0

65**

*(0

.159

)

−1.1

39**

*(0

.165

)

−0.5

27**

*(0

.164

)

−0.3

33

(0.2

26)◦

−0.9

39**

*(0

.233

)

−0.7

42**

*(0

.159

)

−0.7

32**

*(0

.191

)

−1.0

61

(0.3

15)

−0.2

29

(0.1

89)

−1.0

54

(0.2

12)**

*

Ow

n

typ

e

2.74

7***

(0.6

71)

1.81

5***

(0.2

34)

1.83

7***

(0.2

94)

2.34

2***

(0.2

40)

0.80

6◦(0

.537

)

2.21

1***

(0.2

19)

2.28

5***

(0.2

40)

2.57

2***

(0.4

08)

2.16

9***

(0.7

13)

Nei

ghbo

urs

’ act

ion

0.65

3

(0.6

46)

0.57

6***

(0.0

56)

0.69

8***

(0.0

85)

0.39

9***

(0.0

52)

0.29

6***

(0.0

69)

Nei

ghbo

urs

’ typ

e

−0.6

90

(0.4

89)

0.80

4***

(0.2

79)

−0.5

07

(0.7

46)

0.68

9

(0.4

24)

1.71

9***

(0.5

71)

−0.0

34

(0.9

99)

Un

tru

stin

g

−1.1

28**

*(0

.485

)

0.25

4

(0.5

16)

Tru

stin

g

2.12

7***

(0.6

05)

2.08

2***

(0.7

09)

TR-N

OIN

FO

3.25

6**(0

.281

)

2.87

9**(0

.351

)N

um

. TR

-NO

INFO

1.49

4**(0

.237

)

1.24

9**(0

.324

)Lo

g-li

k

−258

.107

−268

.752

−266

.133

−275

.369

−267

.981

−271

.374

−520

.530

−517

.959

−523

.135

−521

.918

−518

.361

Res

tr. L

. L.

−310

.494

−345

.832

−345

.780

−352

.845

−346

.464

−349

.997

−664

.117

−662

.552

−673

.769

−682

.612

−666

.654

N. o

bser

v.

180

225

225

225

225

225

450

450

450

450

450

Std

. err

ors

in

par

enth

eses

. Con

stan

t

and

per

iod

du

mm

ies

incl

ud

ed.

aEs

tim

ates

in

colu

mn

s

I are

rela

ted

to

the

2nd

–5th

per

iod

s;

esti

mat

es

in

colu

mn

s

II–V

I are

rela

ted

to

the

6th

–10t

h

per

iod

s;

esti

mat

es

in

colu

mn

s

VII

–XI a

re

rela

ted

to

the

11th

–20t

h

per

iod

s.b

The

dep

end

ent

vari

able

is

the

nu

mbe

r

of

toke

ns

sen

t

by

each

tru

stor

in

the

rou

nd

t.

Nei

ghbo

urs

’ act

ion

: ave

rage

nu

mbe

r

of

toke

ns

sen

t

by

the

nei

ghbo

urs

in

per

iod

t −

1.

Ow

n

typ

e:

exog

enou

s

tru

stor

’s

pro

pen

sity

to

tru

st,

obta

ined

as

the

mea

n

valu

e

betw

een

his

/her

gen

eros

ity

and

the

risk

pre

fere

nce

ind

exes

(th

ese

are

du

mm

ies

equ

al

to

1

wh

en

the

tru

stor

is, r

esp

ecti

vely

, mor

e

gen

erou

s

and

mor

e ri

sk

lovi

ng,

0

oth

erw

ise)

. Nei

ghbo

urs

’ typ

e:m

ean

valu

e

betw

een

the

two

nei

ghbo

urs

’ pro

pen

siti

es

to

tru

st

(an

alog

ousl

y

to

the

“ow

n

typ

e”

vari

able

, eac

h

nei

ghbo

ur’

pro

pen

sity

to

tru

st

is

calc

ula

ted

as

the

mea

n

valu

e

betw

een

his

/her

gen

eros

ity

and

risk

pre

fere

nce

ind

exes

).

Tru

stin

g:

du

mm

y

equ

al

to

1

wh

en

both

the

vari

able

s

“ow

n

typ

e”

and

“nei

ghbo

urs

’ typ

e”

are

abov

e

the

mea

n

valu

e

in

the

sam

ple

(so

that

mor

e

risk

lovi

ng

and

mor

e ge

ner

ous

ind

ivid

ual

s

are

grou

ped

wit

h

agen

tsw

ith

sim

ilar

char

acte

rist

ics)

. Un

tru

stin

g:

du

mm

y

equ

al

to

1

wh

en

both

the

vari

able

s

“ow

n

typ

e”

and

“nei

ghbo

urs

’ typ

e”

are

belo

w

the

mea

n

valu

e

in

the

sam

ple

. Th

e

excl

ud

ed

du

mm

y

vari

able

ind

icat

es

that

“mor

e

tru

stin

g”in

div

idu

als

mee

t

“mor

e

un

tru

stin

g

nei

ghbo

urs

”

(an

d

vice

vers

a). T

R-N

OIN

FO:

du

mm

y

equ

al

to

1

wh

en, i

n

the

firs

t

5

rou

nd

s

of

the

tru

st

gam

e,

the

tru

stor

sen

t

to

his

/her

reci

pie

nt

an

amou

nt

of

toke

ns

hig

her

than

the

mea

nam

oun

t;

Nu

m. T

R-N

OIN

FO:

nu

mbe

r

of

tru

stor

s

clas

sifi

ed

as

“TR

-NO

INFO

”.*

Stat

isti

call

y

sign

ifica

nt

at

10%

leve

l.**

Stat

isti

call

y

sign

ifica

nt

at

5%

leve

l.**

*St

atis

tica

lly

sign

ifica

nt

at

1%

leve

l.◦

Stat

isti

call

y

sign

ifica

nt

at

20%

leve

l.

5

traottia

(bip

pedioi

rK(ktaewHIc

ss

dAscirot

caotvp(vcps

nvwt

rimental Economics 53 (2014) 97–110

. Concluding remarks

Might trust be influenced by the observed behaviour of otherrustors rather than by the expectation of monetary benefits? Ouresearch provides clean experimental evidence of contagion. Onverage, subjects send a lower (higher) number of tokens afterbserving their neighbours sending few (many) tokens. While inhe presence of social influence, we observe a neat change in con-ribution before and after the fifth period, in the absence of socialnfluence, we observe similar trends in individual trust before andfter the fifth period.

Overall, contagion can affect trust in two different directions.Firstly, within neighbourhoods, we observe positive peer effects

> 0) in the sense that the majority of subjects adapts theirehaviour to the observed behaviour of their neighbours, increas-

ng (decreasing) the number of tokens sent to the recipients, in theeriods 6–20, in line with what his\her neighbours do.

Secondly, comparing the average level of trust over the entireopulation (the fifteen neighbourhoods), we find that social influ-nce has a negative impact on trust, since the overall averageecreases in periods 6–20. However, the negative effect of social-

ty on the population’s average level of trust can be dependentn the fact that we have a prevalence of selfish and untrustingndividuals.

It is a well-known result that trustors differ in their giving witheference to their risk attitude. For example, in field experiments,arlan (2005) and Schechter (2007) observe that less risk averse

risk taking) trustors pass more to the trustees. It is also a well-nown result that trustors differ in their giving with reference toheir generosity (Cox, 2004; Ashraf, Bohnet, and Piankov, 2006)s the most generous trustors pass more to the trustees. Consid-ring social preferences, individual risk aversion and generosity,e provide new evidence regarding future research questions:ow do an individual’s characteristics relate to social influence?

s social influence greater in groups that have homogeneousharacteristics?

Though experimental methods are useful tools for measuringocial influence, there are several concerns that must be empha-ised.

First, our neighbourhoods are artificial; that is, subjects are ran-omly assigned to neighbourhoods at the beginning of the sessions.ccordingly, artificial social groups may be a poor proxy for realocial groups, and one may argue that social influence is, in thisase, a biased indicator of the level of social influence one may findn the real world. However, in our opinion, the identification of theeal social groups is a complicated task regardless of the methodol-gy adopted because too many unobserved factors may influencehe allocation of individuals to specific networks.

Secondly, using experimental behavioural measures helps over-ome some of the possible problems which can arise withttitudinal data on trust (e.g., lack of incentives), conversely, itpens the doors to some possible identification problems betweenhe measurement of peer effects and the measurement of obser-ational learning (Manski, 2000). We explain how we tackled theroblem when designing the experiment; however, one may argueas for the artificial neighbourhoods) that abstracting from obser-ational learning is an artefact that may introduce a bias as strategiconsiderations are crucial factors that determine behaviour whenlaying for real monetary benefits. In other words, in a real context,ocial influence is mixed with strategic considerations.

Measuring social influence is a growing area of research in eco-omics, and we believe that the experimental methods can beery useful instruments in finding answers while the problems,

hich specific methodological tools may produce, must spur fur-

her research on this topic.

Expe

A

iaeCgwfa

A

cttmfap

I

eysysas

G

ecstttcgeuisp

dtmff

T

T

d

mr

T

pew

hawmw3ihots

t

TWOd

tmyafyiiass

A

T

cP

T

a

(

(

L. Luini et al. / Journal of Behavioral and

cknowledgements

We would like to thank participants in the SEET 2010 meet-ng in Marrakesh and participants in the workshop on Cognitivend Experimental Economics (Capua, May 2010), the ESA confer-nce (Tucson, 2012) for their useful comments. We thank Ananishhauduri, John H. Kagel and two anonymous referees who sug-ested valuable changes. A special thank-you to Niall O’Higginsho participated in the experimental design. Financial support

rom Miur (prin 2007) and the Universities of Siena and Salernore acknowledged. The usual disclaimers apply.

ppendix A. The Instructions

Please note that subjects read the Instructions directly on theomputer screen and each part was presented separately. Beforehe experiment started, all they knew was that the session wouldake between one hour and one hour and a half and the experi-

ents comprised different parts. We here report the Instructionsor Player A. The Instructions for player B differed for the Dictatornd the Trust Game. While playing as recipients in the Trust Game,layer B was not allocated to groups.

nstructions for Player A

Welcome to our experiment! Today you are participating in anconomic experiment that will help our research and will enableou to earn a fair amount of money. The experiment compriseseveral parts, and the Instructions for each section will appear onour screen when a part of the experiment is completed by all theubjects. Read these Instructions carefully and do not hesitate tosk if you find them unclear. Please do not communicate with otherubjects in the room.

eneral Information

Earnings: You will receive a 2 Euro participation fee. During thexperiment, you will receive tokens (Exchange rate 1 token = 1 Euroent). The experiment is divided into four parts, one of which is aimple questionnaire. You will not be paid for filling in the ques-ionnaire. As for the remaining parts, you will be paid accordingo your decision in the part where a single decision is required. Inhe two remaining parts where a series of choices are required, theomputer (at the end of the session) will randomly select one sin-le period and you will be paid accordingly. Please note that at thend of the experiment, the payment scheme, i.e. the table summingp your earnings for each decision in the three parts of the exper-

ment will be presented on your screen; then, the computer willelect the period to which the payment is referred and you will beaid immediately after.

The roles: At the start of the experiment, the computer willivide all the participants into two groups: A and B. You will beold in the next screen which role you are playing: please bear in

ind that, once the role is selected, it will be kept throughout theour parts of the experiment: if you are A, you will play that roleor the entire session.

he Questionnaire

Please fill in the questionnaire that will appear on your screen.

he Dictator Game

The organisers are allocating 200 tokens to you; you have toecide how many tokens you are keeping for yourself and how

(

rimental Economics 53 (2014) 97–110 109

any tokens you wish to send to an anonymous player B whoeceived nothing. Your will keep the remaining tokens. for yourself

he Trust Game

This part is composed of 20 periods (decisions). From the 5theriod onwards, you will be placed in a group with two other play-rs A. The composition of the groups is fixed: for fifteen periods youill therefore be in the same group.

In each period, the organisers allocate 600 tokens to you; youave to decide how many tokens you want to keep for yourselfnd how many tokens you want to send to an anonymous player Bho has no endowment. Player B receives the tokens sent by youultiplied by three. For example, if you send 200 tokens, he\sheill receive 600 tokens; if you send 100 tokens, he\she will receive

00 tokens, and so forth. Player B with whom you are paired isnformed on the number of tokens he received; then, he is askedow many tokens he wants to return to you. You will be informedn player B’s decisions only at the end of the experiment, when aable appears on the screen reporting – for each stage – the tokensent and the tokens returned.

You will be asked to take 20 decisions on the number of tokenso send to the different player B.

PLEASE NOTICE THAT, FROM THE FIFTH PERIOD TO THE TWEN-IENTH (THE PERIODS IN WHICH YOU ARE PART OF A GROUP) YOUILL BE INFORMED OF THE NUMBER OF TOKENS SENT BY THE

THER TWO A PLAYERS WHO ARE IN YOUR GROUP: the screenisplay will make this information clear.

Summing up: (a) in each of the 20 periods you receive 600okens; the decision you have to take is always the same: how

any tokens you want to keep for yourself and how many tokensou send to a player B who receives the amount multiplied by threend is then asked if he wishes to return tokens to you. Player B variesrom period to period; (b) from the fifth to the twentieth period,ou will be in a group with two more players A and you will benformed of their choices for that period; at the end of the exper-ment, when the payment is computed for each part, a table willppear on your screen summing up your 20 decisions and the deci-ions of the player B with whom you were paired for that specifictage.

READ THESE INSTRUCTIONS CAREFULLY: DO NOT HESITATE TOSK FOR EXPLANATIONS IF YOU FIND THEM UNCLEAR.

he lotteries

The table that will appear shortly on your screen asks you tohoose between a “X” and a “Y” option for ten different choices.lease indicate your decision for each of the ten options.

he payment stage

Welcome to the payment stage! We remind you that you earn 2 Euro participation fee.

1) Payment for the Dictator Game: You earn 200 tokens less theamount of tokens sent to the B player.

2) Payment for the lotteries: your screen will display the tablesumming up your ten choices. Now the computer randomlyselects one of the ten choices and then the card will be turned

up and you will be paid accordingly.

3) Payment for the Trust Game: your screen will display the tablesumming up your twenty decisions and the twenty decisionstaken by the specific B player to whom you were paired for that

1 Expe

A

i

R

A

A

B

C

C

C

C

C

C

C

D

D

E

E

E

F

F

F

F

G

G

H

H

K

KK

K

M

M

M

M

M

N

S

SS

S

10 L. Luini et al. / Journal of Behavioral and

specific period. Now the computer randomly selects one of thetwenty decisions and you will be paid accordingly.

ppendix B. Supplementary data

Supplementary data associated with this article can be found,n the online version, at doi:10.1016/j.socec.2014.08.007.

eferences

shenfelter, O., J. Currie, H.S. Farber and M. Spiegel, 1992. “An experimental com-parison of dispute rates in alternative arbitration systems”. Econometrica 60,1407–1434.

shraf, N., I. Bohnet and N. Piankov, 2006. “Decomposing trust and trustworthiness”.Experimental Economics 9, 193–208.

erg, J., J. Dickhaut and K. McCabe, 1995. “Trust, reciprocity and social history”.Games and Economic Behaviour 10, 122–142.

ason, T. and V.L. Mui, 1998. “Social influence in the sequential dictator game”.Journal of Economic Psychology 42, 246–265.

harness, G. and M. Sutter, 2012. “Groups make better self-interested decisions”.Journal of Economic Perspectives 26(3), 157–176.

haudhuri, A., 2009. Experiments in Economics. Playing Fair with Money. Lon-don/New York: Routledge, pp. 101–105.

haudhuri, A. and L. Gangadharan, 2007. “An experimental analysis of trust andtrustworthiness”. Southern Economic Journal 73(4), 959–985.

haudhuri, A., S. Graziano and P. Maitra, 2006. “Social learning and norms in a pub-lic goods experiment with inter-generational advice”. The Review of EconomicStudies 73(2), 357–380.

osta-Gomes, M.A. and G. Weizsäcker, 2008. “Stated belief and play in normal formgames”. The Review of Economic Studies 75(3), 729–762.

ox, J.C., 2004. “How to identify trust and reciprocity”. Games and Economic Behav-ior 46, 260–281.

eck, C. and N. Nikiforakis, 2012. “Perfect and imperfect real-time monitoring in aminimum-effort game”. Experimental Economics 15(1), 71–88.

ohmen, T., A. Falk, D. Huffman, U. Sunde, J. Shupp and G.G. Wagner, 2005. “Indi-vidual risk attitudes–Evidence from a large, representative, experimentallyvalidated survey”. IZA W P. 2380.

ckel, C.C. and P.J. Grossman, 1998. “Are women less selfish than men? Evidencefrom dictator experiments”. The Economic Journal 108(May), 726–735.

ckel, C.C. and P.J. Grossman, 2005. “Managing diversity by creating team diversity”.Journal of Economic Behavior and Organization 58, 371–392.

ckel, C.C. and R.K. Wilson, 2004. “Is trust a risky decision?” Journal of Economic

Behavior & Organization 55, 447–465.

alk, A. and A. Ichino, 2006. “Clean evidence on peer effects”. Journal of LabourEconomics 24, 39–58.

alk, A., U. Fishbacher and S. Gaechter, 2013. “Living in two neighborhoods – socialinteraction effects in the lab”. Economic Inquiry 51(1), 563–578.

S

S

rimental Economics 53 (2014) 97–110

ehr, E., 2009. “On the economics and biology of trust”. Journal of the EuropeanEconomic Association 7(2–3), 235–266.

ortin, B., G. Lacroix and M.C. Villeval, 2007. “Tax evasion and social interactions”.Journal of Public Economics 91, 2089–2112.

aechter, S., D. Nosenzo and M. Sefton, 2012. “Peer effects in pro-social behavior:social norms or social preferences?” IZA DP 6345.

laeser, E.L., B.I. Sacerdote and J.A. Sheinkman, 1996. “Crime and social interaction”.Quarterly Journal of Economics 111, 507–548.

artmann, W.R., P. Manchanda, H. Nair, M. Bothner, P. Dodds, D. Godes, K. Hosana-gar and C. Tucker, 2007. “Modelling social interactions: identification, empiricalmethods and policy implications”. In: Seventh Triennial Invitational ChoiceSymposium,. U. of Pennsylvania’s Wharton School.

olt, A.C. and S.K. Laury, 2002. “Risk aversion and incentive effects”. The AmericanEconomic Review 92(5), 1644–1655.

arlan, D.S., 2005. “Using experimental economics to measure social cap-ital and predict financial decisions”. American Economic Review 95(5),1688–1699.

menta, J., 1986. Elements of Econometrics. Macmillan: NY.rauth, B., 2002. Simulation-based Estimation of Peer Effects. W.P., Simon Fraser

University.ugler, T., E. Kausel and M. Kocher, 2012. “Are groups more rational than individ-

uals? A review of interactive decision making in groups”. Wiley InterdisciplinaryReviews: Cognitive Science 3(4), 471–482.

aier, J. and M. Rüger, 2010. “Measuring risk aversion model-independently”. Dis-cussion Papers in Economics 33, 2010.

anski, C.F., 1993. “Identification of endogenous social effect: the reflection prob-lem”. Review of Economic Studies 60(3), 531–542.

anski, C.F., 2000. “Economic analysis of social interactions”. Journal of EconomicPerspectives 14(3), 115–136.

ittone, L. and M. Ploner, 2011. “Peer pressure, social spillovers, and reciprocity: anexperimental analysis”. Experimental Economics 14(2), 222.

offitt, R.A., 2001. “Policy interventions, low-level equilibria and social interac-tions”. In Durlauf, S.N. and H. Peyton Young (Eds.), Social Dynamics. BrookingsInstitution.

oussair, C.N., C.R. Plott and R.G. Riezman, 1995. “An experimental investiga-tion of the patterns of international trade”. American Economic Review 85,462–491.

acerdote, B.I., 2001. “Peer effects with random assignment: results for DartmouthRommates”. The Quarterly Journal of Economics 116(2), 681–704.

apienza, P., A. Toldra and L. Zingales, 2007. “Understanding Trust”. NBER WP 13387.briglia, P., 2008. “Revealing the depth of reasoning in p-beauty contest game?”

Experimental Economics 11(2), 107–121.chechter, L., 2007. “Traditional trust measurement and the risk confound: an exper-

iment in rural Paraguay”. Journal of Economic Behavior & Organization 62(2),272–292.

chotter, A. and B. Sopher, 2006. “Advice, trust and trustworthiness inan experimental intergenerational game”. Experimental Economics 9,123–145.

ervatka, M., 2009. “Separating reputation, social influence, and identificationeffects in a dictator game”. European Economic Review 53, 197–209.