motives for sharing in social networksare.berkeley.edu/~ligon/papers/network_motives.pdfmotives for...

MOTIVES FOR SHARING IN SOCIAL NETWORKS

ETHAN LIGON AND LAURA SCHECHTER

Abstract. What motivates people in rural villages to share? Us-ing variants of dictator games we measure two preference-relatedmotives for sharing (benevolence and directed altruism), and twoincentive-related motives (sanctions and reciprocity). Most shar-ing in the experiment can be explained by benevolence, but vari-ation in sharing is explained more equally by the two classes ofmotives. Finally, variation in sharing within the experiment is cor-related with observed ‘real-world’ gift-giving, while other motivesmeasured in the experiment don’t seem to be useful in predictingbehavior outside the experiment.

JEL codes: C92, C93, D03, D64, D85, O17.

1. Introduction

Villagers in rural Paraguay frequently share resources, making trans-fers to others in their community without any immediate recompense.These transfers are variously referred to as ‘gifts’ or ‘loans’, and thissort of informal credit or sharing seems to be extremely important forhousehold welfare. But for sharing to improve welfare, a householddoes not need to share with everyone in the village—a connection withsomeone may well do, provided that that someone is themselves well-connected with others in the village (Ambrus et al. 2010).

Our contribution is not to show that sharing is important in thevillages we study—there has long been plentiful evidence that ruralvillagers engage in non-negligible amounts of sharing in risky environ-ments (Townsend 1994, Jalan & Ravallion 1999). However, why people

Date: February 17, 2011.Key words and phrases. Risk-sharing,benevolence,altruism,sanctions,reciprocity,Paraguay.Schechter thanks the Wisconsin Alumni Research Foundation, the Russell Sage

Foundation, and the Delta Foundation for funding. We are grateful to Idelin Moli-nas and everyone at Instituto Desarrollo for their support and advice while inParaguay. Paul Niehaus and seminar participants at Dartmouth, Maryland, OhioState, Princeton, UC Berkeley, Yale, BREAD at Duke, MWIEDC at UW Madison,and SITE at Stanford provided helpful comments.

1

2 ETHAN LIGON AND LAURA SCHECHTER

share and who they choose to share with are critically important ques-tions for understanding the interaction between policies designed toaffect household welfare and intra-village sharing.1

To answer the ‘who’ and ‘why’ of sharing, our basic strategy is tovisit villages and offer a randomly selected ‘treatment’ group (i) somemoney; and (ii) the opportunity to invest some or all of this money witha high expected return, but only on behalf of others in the village. Fromthe decisions these subjects make in the experiment we can measurethe importance of different motives for sharing.

The ‘who’ of sharing is interesting for two distinct reasons. First, itis interesting to know what characteristics cause one person to choosethe other to share with; we explore this in a companion paper (Ligon &Schechter 2010). But second, simply being able to choose a ‘who’ withwhom to share may have an important effect on how much actually isshared. This second reason is the focus of this paper.

We consider several different motives which might lead a subject tomake an investment on another’s behalf. First, the subject might in-vest from a motive of “baseline altruism” or undirected general benev-olence—by making an investment she helps another more than sheherself is harmed, and she derives utility from this. Second, the sub-ject might invest on a particular other’s behalf because she wishes thatparticular person well. We call this directed altruism. In our accountaltruism is distinguished from benevolence by being directed towardimproving the welfare of some particular person.2 We think of thesefirst two motives as being determined by the other-regarding prefer-ences of the sharer (e.g., Becker 1981, Camerer & Fehr 2004), and referto these as ‘preference-related’ motives.

Other motives may be essentially selfish. This includes the thirdmotive for sharing which is to avoid social sanctions. The villagerswe study live in a social environment which encourages some sortsof behaviors with rewards, and discourages others with punishments.In addition, the identity of the beneficiary may matter. Making aninvestment on behalf of a particular person might be a simple way forthe subject to repay past debts, or to curry favor with selected membersof her social network. When the subject cares about the identity of the

1For example, transfers targeted at poor households may be shared with therich (Ligon 2004); low interest microfinance loans may be ‘pipe-lined’ to ineligibleothers (Goetz & Gupta 1995); and the direct provision of public goods such as schoolsupplies may result in off-setting reductions in household contributions toward thosepublic goods (Das et al. 2010).

2Alternatively, one might think of benevolence as some baseline level of altruismtoward an average other living in the village.

MOTIVES FOR SHARING IN SOCIAL NETWORKS 3

recipient of her largesse (beyond what can be explained by altruism)we regard this as evidence of a motive of reciprocity. When a subjectshares because of either sanctions or reciprocity, she is responding to aperceived system of rewards and punishments; accordingly, we refer tothese motives as being ‘incentive-related.’

A somewhat unusual feature of our experiments is that they aredesigned to indirectly measure behavior that we expect to occur outsidethe frame of our experiment. When an experimental subject shares forselfish reasons, we do not expect to observe the consequent sanctionsor reciprocal transfers. Instead, by controlling the dictator’s abilityto choose recipients and manipulating the flow of information betweenthe experimental frame and the outside village we can infer somethingabout what those sanctions or reciprocal transfers must be.

Our experiment is most closely related to an experiment carried outby Leider et al. (2009) with Harvard undergraduate students. In Sec-tion 2 and Section 5 we compare our experimental design and resultswith theirs. One of the chief contributions of our paper is that we arewe are able to replicate all of their major findings with our data set,despite extraordinarily large differences in populations and the envi-ronments in which sharing was observed.

Our approach is also closely related to work on risk-sharing in villagesin developing countries. There has been a long-standing understandingthat levels of risk in such settings often are such that sharing is criti-cal to survival (Scott 1976), and there have been systematic efforts tomeasure the importance of such sharing (e.g. Townsend 1994). But inmodeling risk-sharing, different researchers and disciplines rely (some-times implicitly) on different ideas of what might motivate sharing.For “moral economists” such as Scott or sociologists such as Durkheim(2001), some generalized pro-social benevolence is instilled in villagers,and this leads to sharing. For economists with a focus on familiesor dynasties such as Altonji et al. (1992) or evolutionary biologistssuch as Wilson (1978) altruism may seem to be key. For neoclas-sical economists who imagine that sharing arises from the voluntaryexchange of contingent claims (Townsend 1994), sanctions will be nec-essary to enforce these claims. And finally, those who look to repeatedor dynamic games as the key to understanding sharing tend to focuson reciprocity (Posner 1980, Fafchamps 1992, Ligon et al. 2002).

Our main results show that all four motives for sharing are quan-titatively important. The preference-related motives account for thelargest proportion of observed transfers, in the sense that the meantransfer attributable solely to benevolence and altruism is equal to 91percent of the mean total transfer. Based on this observation, one


might conclude that incentive-based motives for sharing were unim-portant. However, though they do not have a large effect on meantransfers, variation in transfers across subjects depends importantlyon the incentive-based motives—we estimate that they account for be-tween 11 and 56 percent of the variation in transfers. Perhaps moretellingly, it is only the incentive-based motives that seem to help explainsharing outside the context of the experiment. In particular, ‘better-connected’ individuals are more motivated by reciprocity, and moregift-giving (outside of our experiment) is not correlated with benevo-lence or altruism, but is correlated with our measure of reciprocity.

In Section 2 we discuss related literature. In Section 3 we describethe set of simple experiments we designed to measure the importanceof these different motives for sharing. Section 4 describes our samplingscheme and some survey data which complements the experimentaldata. In Section 5 we use the data gathered from the experiment toseparately identify the contributions of benevolence, altruism, sanc-tions, and reciprocity to transfer patterns. Section 6 concludes.

2. Related Literature

Many economists and other social scientists have used variations ofthe ‘dictator game’ as a method of measuring motives for sharing withothers (Camerer 2003). What is typically observed in these experimentsis the size of a voluntary transfer from the ‘dictator’ to some otheranonymous person. The magnitude of the observed transfer is ofteninterpreted to reveal something about the preferences of the dictator—indeed, the standard form of the experiment is designed to eliminateother possible sources of variation by insisting upon anonymity.

Experimental economists find evidence of other-regarding prefer-ences, trust, and reciprocity in such anonymous settings (Carter &Castillo 2010). But most of the real world situations in which peoplechoose to share resources with others are not anonymous, and ofteninvolve the ability to choose one’s partner.

A variety of papers manipulate anonymity in games similar to ours,but with different aims and consequently different research designs. Forexample Hoffman et al. (1994) and Hoffman et al. (1996) manipulateanonymity in dictator games, but the only variation in anonymity in-volves whether the subject can be identified by the experimenter ornot. Similarly, Habyarimana et al. (2007) conduct experiments in ur-ban Uganda in which they vary whether or not the subjects meet therecipient in order to deduce his ethnicity (in this urban setting onlyonly 5 percent of subjects report that they know the recipient).


An experiment with aims closer to our own is reported by Glaeseret al. (2000). This involves non-anonymous trust experiments withHarvard undergraduates, eliciting lists of the friends they have in com-mon before they participate in the games. They find that studentswho have more common friends and who have known each other longerare both more trusting and more trustworthy. However, the authorsdo not distinguish whether this is due to higher levels of altruism be-tween more connected partners, or due to the possibility of repeatedinteractions outside of the experimental setting.

There are antecedents in the literature which explore some of thesame questions we do. For example, Leider et al. (2009) and DellaV-igna et al. (2010) try to measure the importance of different motivesfor sharing (the ‘why’ of sharing) in a lab experiment and a field ex-periment respectively, and Barr & Genicot (2008) try to understandhow people choose others to share with (the ‘who’ of sharing). Butthe ‘why’ and the ‘who’ are closely interlinked questions, and ours isthe first effort of which we are aware to address the two questionssimultaneously.

As in Leider et al. (2009), we attempt to distinguish among differentmotives for sharing. Although there are three critical differences be-tween our experiment and theirs, we can replicate all of their relevantmain results with our data (as we discuss in Section 5). The first differ-ence is simply that our settings are very different—where Leider et al.(2009) consider sharing among undergraduates in a Harvard dormitory,we consider sharing among people in a poor agricultural setting. Theimportance of and reasons for sharing seem likely to be different acrossthese two settings; sharing in a Harvard dormitory may be related tosocial and career motivations, while sharing in Paraguayan villages maybe more closely related to credit and insurance. A second difference isthat we have data both on transfers made in the experiments, but alsoon transfers made in daily life in the real world. Thus, we can comparemotives for sharing from the experiment with motives for sharing inthe real world.

The third difference is that in some of our treatments we allow oursubjects to choose the recipient of their sharing, rather than randomlyselecting a recipient for them. In the poor rural setting in which weconduct our investigation this refinement is important. One of thefeatures of this environment is the importance of informal credit andgift-exchange. But to have access to informal credit one need not havea reciprocal arrangement with any randomly selected person in thevillage; one need only establish such a relationship with a single chosenperson. Even in the U.S. credit may be importantly related to social


networks—Leider et al. (2009) cite evidence from the 1995 GeneralSocial Survey that 55% of Americans report that they first approachclose friends and family members when they need to borrow a large sumof money. This suggests the importance of looking at relationships withpartners of the respondent’s choosing, rather than randomly chosenpartners.

Two recent papers also study the ‘who’ of sharing. Barr & Geni-cot (2008) and Attanasio et al. (2009) consider partner choice in ruralZimbabwe and Colombia respectively. For the experiments reportedin these papers subjects form risk-pooling groups.3 While sharing inthese experiments does not increase payoffs per se, it decreases the riskindividuals bear. In contrast to the work we present here, the exper-imenters impose a particular sharing rule on the groups, and (criti-cally) assume that no side-payments or other compensating sanctionsor transfers occur outside the purview of their experiment.

Most experiments do not rely on subjects’ ongoing relationships witheach other. More commonly, repeated interactions are part of the ex-perimental design. There is a growing literature which allows pun-ishments and rewards within experiments. In contrast, in our non-anonymous experiments we do not assume that we can control the‘super-game’ punishments and rewards that may occur outside theframe of our experiment. Instead we try to exploit these to measurethe incentive-based motives for giving. Evidence from other controlledgames and evidence from previous rounds of data collection leads us tothink that these incentives may be relatively large.

Experiments show a relatively large amount of within-game punish-ment (Fehr & Gachter 2000a, Ostrom et al. 2000, Falk et al. 2008),even in games played with undergraduates for stakes of only a coupleof dollars. In addition, the possibility of punishment by a second moverhas a large impact on behavior of the first mover, encouraging him toact more cooperatively in order to avoid punishment. For example,ultimatum games are quite similar to dictator games in that the firstmover decides how to split a pot of money. But, in this case, the secondmover can either reject the split (so neither player receives anything)or accept it. The median proposal in ultimatum games hovers around40-50 percent and offers below 20% are rejected 40 to 60 percent ofthe time (Camerer 2003, Table 2.2). Andreoni et al. (2003) find thatoffers are larger when second movers can both punish and reward and

3Slonim & Garbarino (2008) allow some players to choose characteristics of theirpartner (age and gender) and find that senders in both the dictator game and thetrust game who chose their partner send more than those who did not.


smallest when they can do neither. Second movers chose to punish orreward 43 percent of the time when it was possible. And maximumsecond mover payoffs (for college undergraduates) are US$ 2.40. Thisshows that within-game punishment is often used and is effective, evenin low stakes games.

As secondary evidence that super-game punishments may be largeand impact behavior within the games, we analyze data from a subsetof the same population collected in 2002 (Schechter 2007). The sur-vey asks respondents who experienced theft (and claim to know whostole from them) how they punished the person who stole from them.Sixteen percent say they yelled at the person, 69% say they told alltheir neighbors, and, of the 84% who say they used to drink terere(a Paraguayan tea) with the person, 42% said they stopped drinkingterere with the person. The median item stolen was worth 200 KGs(thousand Guaranies; 1000 Guaranies were worth about 20 US centsat the time of the experiment) while the 25th percentile was 50 KGs.A day’s wages in agriculture at the time was between 15 and 20 KGs.Punishments are not significantly less likely when the stolen item isworth less than or equal to 50 KGs compared with the rest of the sam-ple. As the stakes in our games are 14 KGs, we think that it is notunlikely that the possibility of super-game punishments would effectbehavior, and that selfish or generous decisions might be punished orrewarded.

3. Experiment

Working with a random sample of households from each of fifteenvillages in rural Paraguay, we first conducted a comprehensive house-hold survey, and then encouraged each of our randomly selected samplehouseholds to send someone to participate in an experiment the follow-ing day. Because our experiment involved having subjects play variantsof the dictator game, we refer to the people who came and participatedas ‘dictators’. The actions of the dictators in our games led to house-holds in the village receiving varying amounts of money; we refer tothese households as ‘recipients’.4

Dictators were asked to play four distinct games, in a random or-der. The games varied principally in whether or not the dictator wasanonymous, and in whether the recipient was randomly selected fromthe set of households in the village, or was chosen by the dictator. By

4Although a majority of dictators in our sample are male, we use female pronounswhen referring to dictators, and male when referring to recipient households.


manipulating the dictator’s anonymity and ability to choose the recip-ient, results from each of the games can be used to distinguish betweenthe four distinct motives for sharing.

In each of the games the dictator is given a sum of money and decideshow to divide it between herself and another household. To give thedictator a reason to share within the confines of our experiment (ratherthan choosing to make a transfer later on) we double any money sharedby the dictator—the catch is that any money so invested is given tosome other household. While only those individuals who showed upfor the experiment could act as dictators, any household in the villagecould be a recipient.

From the point of view of each of the dictators, the four games sharesome common elements. In each game, the dictator is given 14 KGs(about $3). The dictator can simply keep this money if she chooses,but she is also offered an opportunity to invest the money on another’sbehalf. The returns to this investment are stochastic, and given by2(τ + d) KGs, where τ is the amount invested by the dictator, and d isequal to an integer 0–5 determined by the roll of a six-sided die.5 Thedictator never learns the value of d.

Any household in the village (not just those randomly included inour survey sample) can be a recipient, except that the money investedby a dictator cannot go to her own household (however, her householdcan receive money from other dictators). Recipient households are alsonecessarily participants in the experiment. Though their participationis mostly passive, the information they receive and the dictator’s beliefsabout this information are important.

From the recipient’s point of view, the different games also havesome common elements. In particular, for those recipients who werenot also dictators, an envelope filled with the random returns from thevarious dictators’ investments from all the different games is deliveredto each recipient household shortly after the conclusion of the game.6

Before observing the amount of money in the envelope, the chosenrecipients are asked to respond to a short survey, which (along withbasic household information) elicits the recipient’s hypothetical play inthe same set of games.

Other elements differ across the four games. As noted above, whatwe vary is whether the dictator remains anonymous or is revealed, and

5Thus, even if the dictator sent nothing, the recipient usually received payoffsdue to the stochastic addition. In the one case in which both τ and d equalled 0,we visited the household to tell them they had not received any money.

6This occurred at the earliest on the same day and no later than 2 days later.


whether the recipient is drawn randomly or is chosen by the dictator.Thus, in the ‘Anonymous-Random’ game, the dictator decides howmuch to share with some randomly selected household in the village,and neither the dictator nor the recipient ever learns who the other was.In the ‘Revealed-Random’ game, the dictator once more chooses howmuch to send to a randomly selected household, but the identities ofboth the dictator and recipient are subsequently revealed to each other(and the dictator knows that this will happen). In the ‘Anonymous-Chosen’ game, the dictator chooses not only how much to invest, butalso who the recipient household will be. The recipient never learns theidentity of the dictator. In the ‘Revealed-Chosen’ game, the dictatorchooses both a recipient household and how much she wants to share,and her identity is subsequently revealed to the recipient. The gameprotocol can be found in Appendix D.

Chosen RandomAnonymous B + A B

Revealed B + A+ S +R B + S

Table 1. Treatments and Motives for Transfers

Thus, differences across the games amount to whether or not thedictator chooses the recipient or not, and what is revealed to the re-cipient about the identity of the dictator (note that when the dictatorchooses, the recipient’s identity is necessarily known to the dictator).These differences are illustrated in Table 1. In the first column thedictator chooses the recipient, while in the second column the recipientis randomly selected from the set of households in the village. The tworows indicate whether the dictator remains anonymous, or whether heridentity is revealed to the recipient.

The elements which appear in Table 1 indicate the motives thatwe expect play a role in each of the different treatments. Considerthe first row of the table, describing games in which the anonymityof the dictator is preserved. Because the dictator remains anonymousshe can be neither punished nor rewarded for her play in the game.Such external incentives can play no role in determining how much shedecides to share, and we follow the literature (e.g., Camerer & Fehr2004) in supposing that in these two anonymous games it is variationin individual preferences that explains variation in sharing.

When there is both anonymity and the selection of recipients is ran-dom (this is the canonical version of the dictator game), not only can


sharing not be motivated by incentives, but the random selection of re-cipients means that the dictator’s sharing in the Anonymous-Randomgame also can not be motivated by a desire to make a transfer to anyparticular person. Accordingly, we term the sole motive for sharing inthis game benevolence or B.7 In the case in which the dictator remainsanonymous but can choose the recipient, she may be motivated by thatsame benevolence, but her sharing may also depend on a desire to ben-efit a particular person—this motive is directed altruism or A, so wetake the sum total of her sharing in the ‘Anonymous-Chosen’ game tobe B + A.

Next consider the second row of Table 1, or the two games in whichthe dictator’s identity is revealed. Preferences can influence sharinghere. But beyond this the fact that the dictator’s identity is revealedand aspects of her behavior observed means that she can be punished orrewarded—accordingly her decision about how much to share dependsboth upon preferences and also upon incentives.

When the dictator’s identity is revealed but the recipient is selectedrandomly, members of the village learn something about the dicta-tor’s generosity, but since no particular person is the object of thatgenerosity neither directed altruism nor reciprocity can play a role inthe dictator’s decision. However, a reputation for being generous maybe rewarded (or its lack punished), depending on the social norms orsanctions present in the village.8 Thus, we take the sum total of thedictator’s sharing in the Revealed-Random game to be B + S.

Finally, when the dictator’s identity is revealed and she chooses therecipient, all of the motives we have discussed may come into play.The dictator may share because she is benevolent; because of altruismdirected toward the recipient; because of social sanctions, and becausethe dictator believes that the chosen recipient may reciprocate if she

7Benevolence captures a combination of generosity and efficiency preferences (asin Fisman et al. (2007)) since the doubling of the amount sent means that aggregatepayoffs are maximized by transfering all funds.

8We are not entirely happy with this terminology. Though ‘sanctions’ capturesthe idea that the village may collectively and impersonally apply rules to encouragesharing, for many people the term also connotes a punishment rather than a reward.Thus we offer ‘social norms’ as a synonym, but that phrase seems to mean differentthings to different people. A third alternative is ‘social pressure,’ as in DellaVignaet al. (2010), but this too seems to carry a somewhat negative connotation. Wetake comfort from the fact that all three of these begin with the letter ‘S,’ andhenceforth use ‘sanctions’ without apology.


shares more. Accordingly, the sum total of the dictator’s sharing in theRevealed-Chosen game is B + A+ S +R.9

3.1. Preserving Anonymity. For some of our games, it is importantthat the recipient not be able to infer who the dictator was, and impor-tant that the dictator not be able to credibly claim to have sent moneyto the recipient. If the dictator thinks that the recipient may be ableto infer her identity, it may cause her to send more in the Anonymous-Chosen version of the game. This would cause us to underestimate theimportance of the incentive-related motives, and particularly of reci-procity. Thus, our measure of reciprocity may be considered a lowerbound. We take several steps to ensure anonymity.

In the anonymous games (Anonymous-Random and Anonymous-Chosen) we do not reveal who the dictator was. However, the dic-tator may wish to reveal herself, and it is this that we take pains tothwart. The only obvious way for a dictator to credibly claim to havesent money is to announce just how much money she sent. We takethree steps to make sure that although the dictator knows how muchshe chose to send, she can not have very much information about howmuch was received. First, in the ‘Chosen’ games, the transfer fromonly one of the two games is actually implemented (determined foreach village by the flip of a coin). Second, in each game we add anadditional random amount of money to the amount received, and thedictator never learned this amount. Third, all of the money sent by allof the dictators to both chosen and random anonymous recipients wasaggregated into a single envelope for each recipient household.

In practice, there were 179 households who received money in thechosen non-revealed game. Of those, thirty-two percent received morethan one anonymous amount of money, making inference more difficult.Although recipients in the chosen non-revealed game are not told whichor how many dictators sent them money, they are probably aware ofthe set of dictators since the games took place in a public place andwere not secret.

We can get a sense of how easy it would be for a recipient householdto guess which dictator chose him by looking at the number of dictators

9Fehr & Gachter (2000b) discuss evidence that economic incentives may crowdout good will. If this is the case, then giving in the revealed random game will notequal the simple sum of benevolence and sanctions. The fact that sanctions areat work will crowd out benevolence. Likewise, the fact that reciprocity is at workwill crowd out both benevolence and directed altruism. Thus, if we would like tomeasure the size of sanctions or reciprocity when all other motives are ‘turned off,’the estimates stated in this paper from the linearly additive approximation may beunderestimates.


with whom he is directly linked according to the survey (assuming thata dictator is more likely to choose recipients with whom she is linked).10

We find that 7 percent of the recipients are linked with no dictators,20 percent are linked with one dictator, 18 percent are linked with twodictators, 39 percent are linked with three to five dictators, and 17percent are linked with six to twelve dictators. Thus, it would not beobvious to a recipient which of the dictators chose him.

Another way to look at anonymity is to compare the actions of thereal dictators with the actions of the recipients in the hypotheticaldictator games. Out of the 371 dictators, 166 chose a recipient whowas also a dictator, 197 chose a recipient who was not also a dictatorand who we were able to interview and ask them to play a hypotheticaldictator game, and 8 chose a recipient who was not also a dictator andwho we were not able to interview. Out of the 166 dictators who choseanother dictator, 37 were also chosen by the dictator they gave to;while out of the 197 who chose a recipient who participated in thehypothetical games, 34 recipients also chose the person who chose him.Since we carefully monitored the experiments, there is no way thatthe recipients who were themselves participating as dictators couldhave been informed by the dictator that she chose him. But, sincewe distributed money to the non-dictator recipients in the day or twofollowing the experiment, it is possible that a dictator could have goneto inform a recipient that she had chosen him. So, the fact that theshare of non-dictator recipients who chose their dictator 34/197 (17%)is, if anything, lower than the share of dictators who chose each other,36/166 (22%), suggests that a recipient chose the dictator who chosehim because they have an ongoing relationship, rather than becausethe recipient knew which dictator chose him and chose to also give tothat dictator to return the favor.

Altogether, these three steps mean that the fact that a householdreceives a certain amount of money provides little information. Ev-ery household in the village can be randomly chosen to receive moneyin the ‘Anonymous’ and ‘Revealed’ games; whatever was sent had anadditional random component added to it; and each recipient house-hold received all its money from anonymous games bundled togetherin a single envelope, even when that money came from several different

10A link occurs between two households who lent or borrowed money, lent orborrowed land, gave or received gifts, gave or received remittances, and helpedout with or received help with health costs in the last year. It also includes theanswer to who they would go to and who would go to them if they needed 20,000Guaranies, as well as close relatives (sibling, parent, or child only) and compadres(the relationship between the parent and godparent of a child).


games or dictators. Thus, except in cases in which we explicitly re-vealed the information, a recipient household could not know whetherthey were chosen or not, could not know how many dictators mighthave chosen them; and could not know how much money was sent byany particular dictator.

3.2. Additional Details. The games in each village took approxi-mately three hours to complete, and players were given 1 KG extra forarriving on time. We used our vehicle to pick up participants who werenot able to get to the game using their own means of transport. In thiscase they were given 1 KG if they were ready when the vehicle arrivedat their residence.

Dictators were not allowed to choose to send money to their ownhousehold, nor could their own household be randomly chosen to re-ceive money from themselves. The dictators were given 14 KGs (a bitless than $3US) in each version of the dictator game. A day’s wagesfor agricultural labor at the time was 15 to 20 KGs.

The players received no feedback about the outcome in each ver-sion until all four sets of decisions had been made. The order of thefour versions was determined randomly for each participant. Thoughthere are twenty-four possible orderings for the four games, we onlyimplemented the twelve orderings which kept the two Chosen gamestogether (we asked players to choose to which single household theywished to send money before asking them how much they would sendin the anonymous and revealed versions).

Importantly, the dictator was allowed to choose only a single recip-ient household for the Chosen games, with the transfer from only oneof the Anonymous or Revealed variants implemented (according to acoin flip). We would expect this to have the effect of weakly reducingthe amount of sharing we observe in both of the Chosen treatments.If we had permitted households to choose a different recipient whenanonymous than when revealed, it is possible that their choices wouldhave been different, and their sharing decision also different. For exam-ple, if a dictator wished to curry favor with a wealthy neighbor who shedisliked, she might well have sent more in each of the two Chosen treat-ments had she been able to choose the wealthy neighbor in the Revealedvariant, but a friend in the Anonymous variant. However, allowing dif-ferent recipients in the two different Chosen treatments would not allowus to do the decomposition of transfers into different motives that wediscuss above. In particular, to compute reciprocity we subtract trans-fers in the Anonymous-Chosen treatment from the Revealed-Chosen,


but if transfers in the former treatment depended on altruism towardsomeone else this decomposition would be invalid.

After participating in the games we asked the dictators two ques-tions. First we asked her the open-ended question why she chose therecipient she did. This yielded the following six categories of answers:a) “he is a good friend”; b) “he is a good person”; c) “he needs moneynow”; d) “he always needs money”; e) “I trust him”; and f) “I owe hima favor.” Dictators could state multiple motives (and in practice neverchose more than two). There seems to be some basis to the answer tothis question. When the dictator stated that the recipient needs moneynow or always, the wealth of the recipient is less than half of what it iswhen poverty is not stated as a reason and the difference is significantat the 1% level.

We also asked the dictator an open-ended question about how shedecided the quantity to invest in the two versions of the games for whichshe chose the recipient. The answers were categorized into one of twopossibilities: a) “the person needs the money and I do not care whetheror not he knows that it comes from me”; and b) “the person will knowthe money is from me and that was important to my decision making”.We can analyze how the size of the four motives differs depending onthe answers to these questions.

4. Data

In 1991, the Land Tenure Center at the University of Wisconsinin Madison and the Centro Paraguayo de Estudios Sociologicos inAsuncion worked together in the design and implementation of a sur-vey of 300 rural Paraguayan households in sixteen villages in threedepartments (comparable to states) across the country. Fifteen of thevillages were randomly selected, and the households were stratified byland-holdings and chosen randomly. The original survey was followedby subsequent rounds of data collection in 1994, 1999, 2002, and 2007.All rounds include detailed information on production and income. In2002 questions on theft, trust, and gifts were added. Only 223 of theoriginal households were interviewed in 2002.11

In 2007, the one non-random village was dropped, and new house-holds were added to the survey in an effort to interview 30 households

11Comparing the 2002 data set with the national census in that year we find thatthe household heads in this data set were slightly older, as we would expect giventhat the sample was chosen 11 years earlier. The households in the 2002 surveywere also slightly more educated and wealthier than the average rural household,probably due to the oversampling of households with larger land-holdings.


in each of the fifteen randomly selected villages for a total of 450 house-holds. Villages ranged in size from around 30 to 600 households. Inone small village only 29 households were surveyed. The survey con-tained most of the questions from previous rounds and also added manyquestions measuring networks in each village.

The process undertaken in each village was the following. We arrivedin a village and found a few knowledgeable villagers to help us collecta list of the names of all of the household heads in the village. Everyhousehold in the village was given an identifier. At this point we ran-domly chose new households to be sampled to complete 30 interviewsin the village. (This meant choosing anywhere between 6 and 24 newhouseholds in any village in addition to the original households.)

The survey included a series of network questions measuring to whomhouseholds made transfers and from whom households received trans-fers, among other things. A full list of the network questions can befound in Appendix D. Since we had a list of the names of all of thehousehold heads in the village, we could match the answers to thesenetwork questions with the identifiers of each household. The surveysprovide evidence of large amounts of in-kind exchange.

We invited all of the households which participated in the survey tosend a member of the household (preferably the household head) toparticipate in our experiments and 371 (83%) of the households didso.12 The experiments were conducted in a central location such asa church, a school, or a social hall. In the experiments the dictatorchose another household in the village to which she wished to transfermoney. If the chosen household had not been surveyed previously aspart of our main sample, then we carried out a ‘short survey’ withthose additional households. This shorter survey contained all of thenetwork questions which were asked in the long survey, but did notcontain the detailed production questions. The short survey also askedthe respondents how they would have played in the games if they hadparticipated.

These villages are mostly comprised of smallholder farmers, of sim-ilar ethnicity. There are no village chiefs, and government is at themunicipal level, which is higher than the village. There are no majormoneylenders, but informal credit is important—in our sample, 42%of households had lent money in the past year (to anyone inside oroutside the village) but only 4% had lent to three or more households.

12The 17% of households which participated in the survey but not the games werericher and more educated, had younger household heads and fewer adults. Theywere no different in terms of the number of links they had with other householdsor in terms of gifts given or gifts received.


Additionally, of the 30% of households who borrowed money in thepast year, 62% also lent money. There are no large plantation ownersin the sample villages.

5. Results

5.1. Description of Play. Recipient households are required to livein the same village as dictator i, but i cannot make transfers to herown household. Thus, let Vi denote the set of households in i’s vil-lage excepting i’s own household, and let #Vi be the number of suchhouseholds.

In each game g ∈ G = AR,AC,RR,RC dictator i chooses toshare some quantity τ gi which could range from 0 KGs to 14 KGs, andin addition some random quantity dgi will be added to her investment.If j is the ultimate recipient of payoffs from i’s play in game g (whetherchosen or randomly selected), then let ψgij equal one, or zero otherwise.The fact that the same person is chosen for both of AC and RC meansthat ψACij = ψRCij = ψCij . A coin-flip determined whether the payoffsfrom AC or RC would be implemented. Let πi be equal to one ifAC payoffs were implemented, and zero otherwise. Thus, the totalresources shared by i are equal to

Senti = τARi + τRRi + πiτACi + (1− πi)τRCi

while the total amount received from the experiment by a recipient jis equal to

Receivedj = 2∑k∈Vj

[ψARkj (τARk + dARk ) + ψRRkj (τRRk + dRRk )

+ψCkj(πk(τACk + dACk ) + (1− πk)(τRCk + dRCk ))

].

Thus, the earnings for a dictator are equal to 42KGs−Senti+Receivedi,not including the possible 1 KGs bonus for punctuality. The averageof these earnings for the dictators was 40.93 KGs with a standarddeviation of 21.71. The maximum earnings for any dictator was 205KGs; the minimum was 0.

In total, there were 371 households which sent someone to play asa dictator in these games, and 752 households which received moneyfrom at least one dictator.13 Thirty percent of these 752 recipientsreceived money from more than one dictator. In total, 299 households

13The sample we have in each village is no longer perfectly representative (theyare now older than average due to the length of the panel) but using data fromthese households to estimate total village income suggests that the total amountdistributed in a village ranged from 0.01% to 0.4% of annual village income.


were chosen by a dictator (rather than randomly selected) to receivemoney. Some households were chosen by more than one dictator.

Game Anonymous Anonymous Revealed RevealedRandom Chosen Random Chosen

Anonymous 5.08 [63%] 3.95 15.57 39.36Random (2.07) (0.061) (0.000) (0.000)

Anonymous 99,143,129 5.39 [65%] 1.23 24.14Chosen (-2.07,2.48) (2.68) (0.250) (0.000)

Revealed 89,132,150 109,136,126 5.47 [65%] 11.56Random (-2.36,2.34) (-2.28,2.18) (2.70) (0.010)Revealed 73,127,171 79,138,154 87,146,138 5.93 [68%]

Chosen (-2.27,2.80) (-2.10,2.36) (-2.11,2.57) (2.84)

Table 2. Comparison of Transfers in Different Games.Cells on the diagonal report the mean transfer given ineach game in KGs; the proportion of the expected to-tal pot shared by the dictator (in square brackets); andthe standard deviation of the transfer (in parentheses).Friedman statistics for a test of a significant differencebetween transfers between each pair of games are pre-sented above the diagonal, with p-values in parentheses.Cells below the diagonal report the count of observationsin which the transfer in the row game was less than, equalto, or greater than respectively; and in parentheses re-port the mean difference in transfers between the rowgame and the column game conditional on the transferin the row game being respectively less than or greaterthan the transfer in the column game.

5.2. Mean Transfers. Table 2 presents comparisons of sharing acrossthe different games. Analogous results for the hypothetical games areshown and discussed in Appendix C. There are three distinct kinds ofstatistics presented in the table. First, in the cells on the diagonal wereport mean transfers made in each of the four games, with standarddeviations of transfers reported in parentheses. Notice that these meansincrease from left to right, from a mean transfer of 5.08 KGs in thebaseline Anonymous-Random game up to a mean transfer of 5.93 KGsin the Revealed-Chosen game. We also report transfers as a share of


the total expected pot; these also increase as one moves from left toright, from 63 percent to 68 percent of the total.14

Second, its of some interest to know more about the distribution oftransfers. Cells in Table 2 below the diagonal report the count of peoplewho made transfers in the row game which were respectively less than,equal to, or greater than in the column game. So, for example, of the371 subjects who played all four games, 73 gave less in the Revealed-Chosen game than in the Anonymous-Random game, while 171 gavemore. Figures in parentheses are conditional mean differences, so inthis same example subjects who gave more in the Revealed-Chosengame gave on average 2.80 KGs more.

Third, we’d like to have a formal test of whether there are signifi-cant differences in transfers across the three games. We can use thecounts reported below the diagonal to construct a two-sided sign testof the hypothesis that transfers in each pair of games is equal. Thistest is a special case of the Friedman test (Lehmann 1975, p. 267).We report the Friedman test statistics in the cells above the diago-nal, along with p-values associated with the null hypothesis of equality.At the ten-percent level, we can reject equality in transfers betweenall games except between the Anonymous-Chosen and the Revealed-Random games. Since our setup assumes that the transfers in thesetwo games are B + A and B + S respectively, and takes no stand onwhich one of these should be larger than the other, this result is notunexpected. Taken together this overwhelmingly confirms the patternhinted at by the increases in mean transfers from left to right—as addi-tional motives for sharing come into play in the different games, morepeople respond by giving more.

A final observation regarding Table 2 is that the baseline level ofsharing in the Anonymous-Random game is high, but may not be outof line with results from dictator games in other settings.15 Cardenas

14The total pot is calculated as follows. Taking into account the fact that theamount shared is doubled and has a random amount (with a positive expectedvalue) added to it before being given to the recipient, a dictator who gives theaverage of 5.084 KGs in the Anonymous Random game creates an expected totalpot of (14− 5.084) + 2× 5.084 + 5 = 24.084.

15Though the order in which games are played is randomized, one still mightwonder whether the order in which these four games are played is important. InAppendix A we reproduce our main regression results, but controlling for order andexperimenter effects. The magnitudes of the results from this exercise differ verylittle from the main results we present in the text (although we lose significanceafter controlling for so many other variables).


& Carpenter (2008) survey the literature employing anonymous dicta-tor games in developing countries, including Henrich et al. (2006). Inexperiments conducted with non-students in developing countries, theaverage share of the initial endowment sent ranges from 26 to 42 per-cent. In comparison, for the Anonymous-Random version of the gamewe play in Paraguay, the average amount sent was 36 percent of theinitial endowment.16

Categories(Players in Category) B A S R

Everyone (Actual) 5.08∗∗∗ 0.31∗∗ 0.38∗∗∗ 0.15(371) (0.14) (0.12) (0.12) (0.16)

Chosen because needy 5.23∗∗∗ 0.53∗∗∗ 0.55∗∗∗ 0(153) (0.23) (0.20) (0.19) (0.26)

Chosen because of affinity 4.98∗∗∗ 0.13 0.26∗ 0.27(230) (0.17) (0.14) (0.15) (0.19)

Revelation unimportant 5.19∗∗∗ 0.39∗∗∗ 0.37∗∗∗ -0.035(285) (0.16) (0.13) (0.14) (0.17)

Revelation important 4.73∗∗∗ 0.047 0.44∗ 0.77∗∗

(86) (0.32) (0.27) (0.25) (0.37)

Table 3. Average value of motives in the actual games,measured in KGs. Numbers in parentheses are standarderrors. Asterisks indicate different levels of significance,with * indicating that the corresponding mean is signif-icantly different from zero at a 90% level of confidence;** at a 95% level of confidence, and *** at a 99% levelof confidence.

We next attribute transfers in our four games to one of the four dis-tinct motives described above. Adopting the arguments in the previoussection, we have the following relationships between sharing in the fourgames and the four motives:(1)

Ti =

τARiτACiτRRiτRCi

=

Bi

Bi + AiBi + Si

Bi + Ai + Si +Ri

=

1 0 0 01 1 0 01 0 1 01 1 1 1

Bi

AiSiRi

= PMi,

16This is considerably less than the 63 percent (of the expected pot) that oursubjects send, but in the experiments reported by Henrich et al. (2006) the amountsent is not doubled, as it is in ours.


where Ti is the column vector of transfers made by person i, and whereP is an invertible matrix which maps motives into transfers. Then,using data on the observed Ti for each person i we can identify motivesMi = P−1Ti. In Table 3, we measure these quantities for the dictatorsin the actual games.

To calculate the reported means and standard errors in Table 3 weestimate the regression

τ gi = B + A1(g = AC) + S1(g = RR) + (A+ S + R)1(g = RC) + εgi

where τ gi is the amount sent by person i in game g and the notation1(g = h) is a function which takes the value of one if the game g is equalto h ∈ AR,AC,RR,RC and zero otherwise. The means reported inTable 3 are then simply obtained by taking differences of the estimatedregression coefficients, and the variances of these means are obtained aslinear functions of the variances of the regression coefficients, allowingfor clustering at the level of the individual.

In Table 3 the column headings indicate different mean values ofcomputed motives; thus, for example, B is the mean value of the Bi.Means of motives for all 371 dictators in the ‘Actual’ games are reportedin the first row labeled “Everyone (Actual)”. In the subsequent rows,we report results from additional regressions which include dummyvariables related to the category of interest (e.g. choosing one’s partnerbecause of affinity) interacted with each of the right-hand side variables.Thus, each row of the table represents a separate regression, and thereported coefficients have the interpretation of conditional means of thevarious motives.

The mean B is much greater than the mean of other motives. Whythis should be so is a puzzle, but not a new puzzle (a large literature isdevoted to understanding ‘high’ levels of sharing observed in one-shotanonymous dictator and ultimatum games; see, e.g., Forsythe et al.1994). The means A and S are of roughly similar magnitude, andare significantly greater than zero. The unconditional mean R is notsignificantly greater than zero. The incentive-based motives of S andR are seen to account for 9% of the total motives.

The rows with the indication “Chosen because” condition on the re-sponse to an open-ended question asked regarding the dictators’ reasonsfor selecting the particular person they did in the “Chosen” games. Therow labeled “Chosen because needy” gives mean values of computedmotives for dictators who indicated that they chose the person they didbecause he “needs money now” or “he always needs money.” The rowlabeled “Chosen because of affinity” gives mean values of computedmotives for dictators who indicated that their choice of person had to


do with the fact that “he is a good person,” “he is a good friend,” “Itrust him,” “I owe him a favor,” or something similar. Some obser-vations are classified in both categories since people were allowed tochoose more than one motive. The values of B and A both increasewhen the recipient is chosen because he is needy.

Turning to the next two rows of Table 3, “Revelation important”means the dictator cares that her identity will be revealed to the chosenrecipient. “Revelation unimportant” means that the dictator says thatshe does not care whether or not her identity is revealed to the chosenrecipient. For those dictators who report that revelation in the chosengames is important we observe a decrease in B and A relative to thecase in which revelation is reported to be unimportant. This makessense, since the motives of altruism and benevolence are exactly thosein which revelation of the dictator’s identity plays no role. Perhapsmost interestingly, there is a rather large increase in the mean valueof reciprocity and a smaller increase for sanctions in the case in whichrevelation is held to be important. These increases in the mean value ofeconomic motives should not be a surprise, for the same reason that adecrease in the preference-based motives should not be. But the muchlarger value of R than S for dictators who care about their identitiesbeing revealed to a chosen recipient household suggests that the kindof individual-specific rewards that one might associate with sharing ina network are considerably more important than are the more generalrewards or sanctions which might influence the value of S.

Unknown to us, Leider et al. (2009) independently conducted a some-what similar experiment, running both an anonymous and a revealedversion of a dictator game with nameless randomly chosen partners liv-ing in the same dormitory. They also conducted both anonymous andrevealed versions of dictator games with named partners of randomlychosen social distances (e.g., direct friend, friend of a friend, etc.). Fromthis they measure benevolence (which they call baseline altruism), al-truism (which they call directed altruism), and a mix of sanctions andreciprocity (which they call “prospect of future interaction”).17

Although our setting and subject pool are quite different from Leideret al. (2009), and our experiment differs in a number of notable respectswe are nevertheless able to essentially replicate their principal results.First, we find somewhat similarly sized effects. When sharing is efficient

17Leider et al. (2009) distinguish between what they term “preference-basedreciprocity” and “enforced reciprocity.” While “preference-based reciprocity” ismore comparable with our measure of altruism, “enforced reciprocity” correspondsto what we simply call reciprocity, which depends on the incentives created byrepeated play.


in their sample (in our sample) giving in the revealed random game is14% (8%) higher, giving in the anonymous chosen/direct friend game is10% (6%) higher, and giving in the revealed chosen/direct friend gameis 40% (17%) higher than in the anonymous random game. Thoughincreases in their games are uniformly larger than ours, we believe thismay be due to the fact that we doubled the amount sent while theytripled it.

More surprisingly, where it is possible to compar we can also repli-cate all of the main results from Leider et al. (2009). Their statementof their “Result 1” is that “Baseline altruism and directed altruismare correlated.” In our terms, this means that the amounts sent inthe Random-Anonymous game (B) and the Random-Revealed game(B + A) should be positively correlated, and indeed they are, withPearson correlation coefficient of 0.64. Their statement of their “Result3” is that “the observability of decisions by partners increases givingmore for friends than for strangers.” In our setting this is simply aprediction that R > 0 (or that the difference in sharing between theRevealed-Chosen and Anonymous-Chosen games exceeds the differencebetween the Revealed-Random and Anonymous-Random games); thisprediction is borne out in Tables 2 and 3.

Finally, “Result 5” from Leider et al. (2009) is that “the nonanonymityeffect and directed altruism are substitutes.” In our terms this can beframed as a claim that the preference-based motives B + A are nega-tively correlated with incentive-based motives S + R. And this resulttoo carries through with our data; Pearson’s correlation coefficient inthis case is equal to −0.32.

5.3. Variation in Motives.

5.3.1. Accounting for Variance in Sharing. We can move beyond anexamination of mean values of motives and look at their contributionto the variance in the amount sent. One method to determine theshare of variation contributed by each motive is to estimate a sequenceof regressions, and to see how the R2 changes as additional motivesare added—this amounts to computing different shares of variancesfor a Gram-Schmidt orthogonalization of motives which eliminates thecomplication of covariances among motives. We begin by regressingτRCi on Bi alone, then regressing τRCi on Bi and Ai, then adding inRi, and finally adding in Si. The final regression has an R2 of 1 by


construction. The contribution of each motive is how much it adds tothe R2.18

(1) (2) (3) (4) (5)Constant 5.93∗∗∗ 2.80∗∗∗ 1.51∗∗∗ 1.38∗∗∗ −0.00∗

(0.15) (0.26) (0.24) (0.19) (0.00)Bi — 0.62∗∗∗ 0.83∗∗∗ 0.81∗∗∗ 1.00∗∗∗

(0.04) (0.04) (0.03) (0.00)Ai — — 0.59∗∗∗ 1.06∗∗∗ 1.00∗∗∗

(0.05) (0.05) (0.00)Ri — — — 0.54∗∗∗ 1.00∗∗∗

(0.03) (0.00)Si — — — — 1.00∗∗∗

(0.00)R2 0.00 0.34 0.53 0.71 1.00

Table 4. Accounting for Variance in Sharing UsingVariation in Observed Motives (Mi). Result of regressingτRCi on motives. Each column adds an additional vari-able; centered R2 statistics are reported in the bottomrow. Sample size is 371.

Results from this exercise are reported in Table 4. Using data onlyfrom the actual games, we find that variation in Bi accounts for 34%of the variance in transfers; in Ai accounts for 19% of the variance, inSi accounts for 29% of the variance, and in Ri accounts for 18% of thevariance. This would imply that incentive-based motives contribute47% of the variance in transfers in the actual games.

5.3.2. Accounting for Variance Using Sums of Motives. A second methodto determine the share of variation contributed by the different motivestaking into account the covariance between motives is to recall that atransfer τRCi observed in the Revealed-Chosen game is equal to the sumof the preference-based motives (Bi +Ai) and the economic incentives

18A disadvantage of this approach is that the order in which one adds variablesto the regression matters—if the variables are not orthogonal, then an “earlier”variable will be credited with variation that could instead be attributed to a “later”variable if these two are positively correlated. Here we have determined the orderby first finding the single regressor which maximizes the R2 in column (2) of Table4; then finding the pair of regressors which maximizes the R2 in column (3), andso on. This procedure leads to a choice of order which maximizes the sum of theR2 statistics across the five regressions.


τRC B A S RτRC 8.07 0.58 0.14 0.08 0.25B 4.47 7.26 -0.43 -0.43 0.31A 0.88 -2.67 5.25 0.56 -0.67S 0.54 -2.69 2.95 5.35 -0.72R 2.18 2.57 -4.66 -5.07 9.33

Table 5. Correlation/covariance matrix of motives(from the actual games). On and below the diagonalare the covariance matrix, while above the diagonal wereport Pearson correlation coefficients. Sample size is371.

(Si +Ri). Here we explore how much of the variation in τRCi is due toeach of these contrasting pairs of motives.

Note that var(τRCi ) = var(Bi+Ai)+var(Si+Ri)+2 cov(Bi+Ai, Si+Ri). Accordingly, the total variance of τRCi can be attributed to vari-ance in one or another of Bi + Ai or Si +Ri, with

1 =var(Bi + Ai)

var(τRCi )+

var(Si +Ri)

var(τRCi )+2

cov(Bi + Ai, Si +Ri)

var(τRCi ).

[0.89] +[0.56] −[0.45]

The figures in brackets are the estimates of these variance shares com-puted from Table 5. Since the covariance term is negative, it followsthat preference-related motives account for at most 89 percent of thevariance in transfers, while incentives account for at most 56 percent;similarly, the least preference-related motives could account for is 44percent, and the least incentives could account for is 11 percent. Fromthis we conclude that although economic motives have a small effecton mean transfers, variation in these motives across individuals seemsto play a larger role in explaining variation in sharing.

5.3.3. Measurement Error. As the previous equation reveals, the meansof these motives conceal some important variation. Estimated covari-ances and Pearson correlation coefficients between the different motivesand between transfers in the ‘Revealed-Chosen’ game are reported inTable 5. There are very striking negative correlations between reci-procity and each of altruism and sanctions, and a positive correlationbetween reciprocity and benevolence.

These correlations among motives may reflect actual variation indictator ‘types’ as described by Fisman et al. (2007). For example,


it may well be that some individuals engage both in the kind of con-ditional generosity one might associate with reciprocity and the un-conditional generosity associated with benevolence. (This associationbetween benevolence and engagement in reciprocal exchange comple-ments the findings of Gurven (2004) and others who find that individ-uals who have more experience with markets are also the most likelyto display what we call benevolence.)

However, there is another, less cheering interpretation of the cor-relations we observe: that these are a consequence of our method ofmeasuring the different motives without taking into account the pos-sibility of measurement error. To see how this might be a possibleproblem, suppose that each individual has a ‘true’ underlying set ofmotives M∗

i = (B∗i , A∗i , S

∗i , R

∗i )

T. But suppose that when dictators playour game and choose their transfers, they do so in a way which onlyimperfectly reflects their underlying motives—transfers for person i ineach game g depend on these motives, but also on some measurementerror εgi , so that the vector of transfers Ti = PM∗

i + εi.If we restrict the measurement error process by assuming that it is

classical,19 we make some progress. In particular, so long as Eεgi = 0,our calculation of mean motives (as in Table 3) remains valid, sinceEM∗

i = EMi. However, the covariances among ‘true’ motives M∗i can-

not be directly inferred from covariances among the observed motivesMi, even exploiting the classical assumption that the measurement er-rors are all orthogonal to each other and to each of the motives. Ifthe measurement errors are identically distributed across the differ-ent games so that Eεiε

Ti = σ2

ε I, the covariance matrix Ω∗ of the ‘true’motives is related to the covariance matrix Ω of the observed motivesby

Ω∗ = Ω− σ2ε

1 −1 −1 1−1 2 1 −2−1 1 2 −2

1 −2 −2 4

.From this we see that the covariances between our observable (buterror-ridden) (Bi, Ai, Si, Ri) depend on the variance of the measure-ment error term in a way which is consistent with the patterns ofcovariances revealed in Table 5.

5.3.4. Correlates of Motives and Bounding Measurement Error. If thevariance of measurement error is large, then we will tend to overstate

19I.e., assume that Eεi = 0, EεgiXi = 0, and EεiεTi = σ2ε I. Since transfers can

take values only in a bounded interval ε obviously can not be Gaussian.


the importance of motives other than benevolence in determining thevariance in transfers. (Remember that measurement error will not af-fect our measures of the means of the transfers.) Here our aim is to finda lower bound on the importance of economic motives in contributingto variance, even if measurement error is large.

Let the true vector of motives depend on a vector of observables Xi

and a set of unobservables ui via a linear relationship M∗i = γTXi +ui;

E(ui|Xi) = 0 by assumption. Let εi denote a vector of measurementerrors, such that Mi = M∗

i + P−1εi. Maintaining the hypothesis thatsuch measurement error is classical, we can estimate γ and calculatea lower bound on the covariance matrix of the true motives M∗

i . Inparticular, we want to find the largest value of σ2

ε such that EuiuTi

remains positive semidefinite.To measure this bound, we choose a small set of variables to include

in Xi, and estimate the vector regression

(2) Mi = γTXi + ui + P−1εi

using ordinary least squares. Results are reported in Table 6. (Sum-mary statistics for all variables in these tables can be found in Ap-pendix B.) The largest variance σ2

ε consistent with our estimated co-variance matrix of unobservables remaining positive semidefinite is ap-proximately 1.90. The resulting estimated limiting covariance matrixfor the true motives M∗

i is reported in Table 7.From Table 7 we see that while the variances of the motives other

than benevolence are now much smaller, variance in reciprocity is stillrelatively large at 32 percent of the variance of benevolence, and has alarger variance than the other motives. While measurement error mayinflate the apparent role played by incentive-based motives in account-ing for variation in transfers, these motives remain important even aftertaking into account the maximal amount of classical measurement er-ror.

Recall the accounting exercise we did from the covariance matrixin Table 5 where we assumed that there was no measurement error.We now assume that there is a classical measurement process with amaximal variance of 1.90 and repeat that exercise. Note that we nowseek to explain variation in τRCi not only by examining variation inmotives, but also in measurement error. Since this measurement errorprocess is classical by assumption, its covariance with the ‘true’ motivesis zero. Then, the accounting for M∗

i which is analogous to the earlier


Bi Ai Si Ri

Dictator Age −0.01 0.00 0.00 −0.02∗∗

(0.01) (0.01) (0.01) (0.01)Dictator Male? 0.38 −0.07 −0.15 0.40

(0.30) (0.26) (0.26) (0.34)log(Income) 0.36∗ −0.07 −0.10 0.03

(0.20) (0.18) (0.18) (0.23)Family Size 0.07 −0.01 −0.06 0.06

(0.07) (0.06) (0.06) (0.08)No. Adult Males −0.24 0.05 0.21 −0.38∗

(0.17) (0.15) (0.15) (0.20)log(Theft) 0.02 0.08∗ 0.09∗∗ −0.10∗

(0.05) (0.04) (0.04) (0.06)OR degree −0.03 −0.07∗∗ −0.04 0.09∗∗

(0.04) (0.03) (0.03) (0.04)R2 0.14 0.09 0.08 0.12

Table 6. Correlates of Motives, Including Motives forPartner Choice. Village fixed effects included. Samplesize is 371. Numbers in parentheses are standard er-rors. Asterisks indicate different levels of significance,with * indicating that the coefficient is significantly dif-ferent from zero at a 90% level of confidence; ** at a 95%level of confidence, and *** at 99% level of confidence.

B∗i A∗i S∗i R∗iB∗i 5.34 −0.76 −0.79 0.64A∗i −0.76 1.46 1.06 −0.85S∗i −0.79 1.06 1.56 −1.27R∗i 0.64 −0.85 −1.27 1.70

Table 7. Limiting Covariance Matrix for True Unob-served Motives M∗

i with Classical Measurement Error.This matrix results when we use the results relating ob-servables Xi to observed motives Mi as reported in Table6, and assume a classical measurement error process witha maximal variance (of 1.90). Sample size is 371.

accounting we did for Mi is:

1 =var(B∗i + A∗i )

var(τRCi )+

var(S∗i +R∗i )

var(τRCi )+2

cov(B∗i + A∗i , S∗i +R∗i )

var(τRCi )+

σ2ε

var(τRCi ).

[0.65] +[0.09] +[0.02] +[0.24]


Thus, classical measurement error can account for at most 24 percentof the variance we observe in τRCi , and at this upper bound variationattributable to the economic motives (S∗i , R

∗i ) is still nine percent. The

covariance between preference-based and economic motives in this lim-iting matrix is now positive, so we can extend our claim to say thatin the presence of classical measurement error, economic motives canaccount for no less than between 9 and 11 percent of variance in theRevealed-Chosen transfers (with maximal measurement error) or be-tween 11 and 56 percent (with no measurement error).

5.4. Interpretation of Correlates. Beyond being useful for bound-ing measurement error, we can use the results of Table 6 to relatevariation in motives to various observables, and this may have someindependent interest. One might expect that reciprocity would mat-ter less to individuals who are more self-sufficient or less invested intheir social network, while individuals who have more invested in theirsocial network would send more. Consistent with this view, Table 6confirms that older people and households with more adult males aresignificantly less reciprocal.

We find that wealthier people are more benevolent, but they donot share differentially more in any of the other treatments. We alsoexplore whether being the victim of property theft is correlated withthe four motives. Most of this theft is agricultural theft committed byfellow villagers (Schechter 2007). In the case of theft, the correlatesof sanctions and reciprocity, the two incentive-based motives, go inopposing directions. Households which experience more theft sharemore due to fear of sanctions. These households may be vulnerableto theft and may wish to avoid future theft by sharing more in theAnonymous-Revealed game. On the other hand, our estimates suggestthat these same victimized households have low levels of reciprocity,suggesting that they would find autarky relatively appealing.

In addition to these characteristics of the household, we also includea variable “OR-degree” meant to capture the position of the householdin the social network of the village (Leider et al. 2009). In the survey,respondents were asked to whom they lent money, lent land, gave gifts,gave remittances, and helped out with health costs in the last year.They were also asked from whom they borrowed money, borrowed land,received gifts, received remittances, and received help with health costsin the last year. They were also asked who they would go to and whowould go to them if they needed 20,000 Guaranies. The degree is thenumber of households within the village with whom they are linked insuch a way. It is called an OR-link because we assume there is a link


either if i says she is linked with j OR j says he is linked with i, orboth.20

The more connected an individual is, the more reciprocal she is andthe more valuable the social network is for her and the less she givesfor reasons of directed altruism. The magnitude of this is such that a 1standard deviation increase in the OR-degree leads to a 0.12 standarddeviation increase in reciprocity. For comparison, one standard devia-tion increases in age of household head or number of adult males leadto 0.10 and 0.11 standard deviation decreases in reciprocity.

We believe that the contrast between the OR-degree’s impact onreciprocity and its impact on altruism is due to the feature of our ex-periment which requires dictators to choose a single recipient in both ofthe two Chosen treatments. The argument is that well-connected dic-tators are more likely to have high returns to making a large Revealedtransfer to an important recipient. But a well-connected dictator whochooses to curry favor from an important recipient may not particu-larly care for that recipient, so that when the dictator decides on whatanonymous transfer to make she will tend to make a smaller transferthan would a less well-connected household with fewer opportunitiesfor strategic sharing.

In Table 8 we add to the regression variables related to borrowingand lending as well as giving and receiving within the village. In resultsnot shown here we included loans and gifts outside the village, but noneof their coefficients are significant. This is as would be expected, sincethe dictator game recipients were limited to within-village boundaries.This tends to confirm our view that the significance of the loan andgift variables is due to the social network within the village.

As expected, households who have borrowed more from their village-mates have less to lose from failing to share in the Revealed-Chosengame (since if others seek to punish them they can renege on their debt)and so they are less reciprocal. Likewise, households which gave moregifts to their village-mates are more reciprocal. Perhaps they are hopingthat their greater gift-giving will be returned in the future. As wasthe case with theft in the previous regression estimates, the correlatesof sanctions and reciprocity, the two incentive-based motives, go inopposing directions for the gift-giving variable. They are not expectingsanctions from randomly chosen village-mates, but focus their givingon and expect returns from specifically chosen individuals. (In Ligon

20If i and j are both in our sample then there is a link if either or both say theyare connected. If i is in our sample but j is not, then there is only a link between iand j if i claims the link exists. We do not have data on j’s perception if he is notin our sample.


Bi Ai Si Ri

Dictator Age −0.01 0.00 0.00 −0.02∗∗

(0.01) (0.01) (0.01) (0.01)Dictator Male? 0.28 −0.20 −0.23 0.53

(0.29) (0.26) (0.26) (0.34)log(Income) 0.38∗ 0.02 0.01 −0.15

(0.20) (0.18) (0.18) (0.23)Family Size 0.05 −0.05 −0.08 0.10

(0.07) (0.06) (0.06) (0.08)No. Adult Males −0.20 0.07 0.19 −0.38∗

(0.17) (0.15) (0.15) (0.20)log(Lent) −0.00 −0.04 −0.04 0.07

(0.06) (0.05) (0.05) (0.07)log(Borrowed) −0.04 0.07 0.06 −0.13∗

(0.07) (0.06) (0.06) (0.08)log(Gifts Given) −0.02 −0.05 −0.08∗ 0.12∗∗

(0.05) (0.05) (0.05) (0.06)log(Gifts Received) 0.05 0.01 −0.05 0.01

(0.06) (0.05) (0.05) (0.06)R2 0.14 0.06 0.08 0.11

Table 8. Correlates of Motives, Including Measures ofFinancial Activity Within the Village. Village fixed ef-fects included. Sample size is 371. Numbers in paren-theses are standard errors. Asterisks indicate differentlevels of significance, with * indicating that the coeffi-cient is significantly different from zero at a 90% level ofconfidence; ** at a 95% level of confidence, and *** at99% level of confidence.

& Schechter (2010), we find that dictators are more likely to chooserecipients to whom they gave money in the past.)

It is worth noting that gift giving is not correlated with the twopreference-based motives of altruism and benevolence. Our evidencetells us that people in these villages give gifts because they expecttheir network partners to reciprocate. This suggests that the transferswe measure in surveys are motivated by self-serving reciprocity ratherthan by benevolence or altruism. Thus, although economic incentivesonly account for nine percent of mean transfers in the experiment, theyaccount for much more of transfers in the real world.


5.5. Partner Choice. In Ligon & Schechter (2010) we examine part-ner choice by running regressions in which the unit of observation is adyad of the dictator with every other household in the village. The de-pendent variable is an indicator variable for whether the dictator chosethat household as recipient, while the explanatory variables are indi-cator variables for different types of relationships the pair might have(lending money, borrowing money, giving gifts, receiving gifts, etc.).

We find that many of the dyad characteristics are positive and sig-nificant, especially those which signify the dictator had made transfersin the past to the chosen recipient. Players are more likely to choosea recipient they are related to. They are also more likely to choosea recipient who would come to them if the recipient needed money, arecipient to whom they gave agricultural gifts in the past year, a recip-ient to whom they gave money to help with health expenses in the pastyear, and a recipient to whom they lent land or money in the past year.They are also likely to choose recipients who lent them money in thepast year. The fact that dictators are most likely to choose as recipi-ents households to whom they give real-world transfers again suggeststhat reciprocity is an important motive in transfers in both real worldand experimental transfers, and that experiments with partner choicecan inform us regarding relationships in daily life.

6. Conclusion

Using four variants of a dictator game, we distinguish four motivesfor sharing benevolence, altruism, sanctions, and reciprocity. We findevidence of sharing due to all four motives. Although the incentive-related motives of sanctions and reciprocity contribute small amountsto mean transfers in the game, they account for a larger share of vari-ation in transfers.

We find that reciprocity as measured in the experiments is correlatedwith a number of real-world variables. The more connected an indi-vidual is, the more valuable the social network is for her and the morereciprocal she is. People who give more gifts to their village-matesare on average neither more benevolent nor more altruistic—rather,these higher levels of gift-giving are correlated with higher levels ofreciprocity within their social network. Thus intra-village sharing out-side our experiment seems to be less a consequence of other-regardingpreferences than of selfish reciprocity within the social network.


References

Altonji, J., Hayashi, F. & Kotlikoff, L. (1992), ‘Is the extended familyaltruistically linked? Direct tests using micro data’, The AmericanEconomic Review 82, 1177–1198.

Ambrus, A., Mobius, M. & Szeidl, A. (2010), Consumption risk-sharingin social networks. Unpublished Manuscript.

Andreoni, J., Harbaugh, W. & Vesterlund, L. (2003), ‘The carrot or thestick: Rewards, punishment, and cooperation’, American EconomicReview 93(3), 893–902.

Attanasio, O., Barr, A., Cardenas, J. C., Genicot, G. & Meghir, C.(2009), Risk pooling, risk preferences, and social networks. Unpub-lished Manuscript.

Barr, A. & Genicot, G. (2008), ‘Risk-pooling, commitment, and infor-mation: An experimental test’, Journal of the European EconomicAssociation 6(6), 1151–1185.

Becker, G. S. (1981), A Treatise on the Family, Harvard UniversityPress, Cambridge, MA.

Camerer, C. (2003), Behavioral Game Theory, Princeton UniversityPress, Princeton, New Jersey.

Camerer, C. & Fehr, E. (2004), Measuring social norms and preferencesusing experimental games: A guide for social scientists, in ‘Founda-tions of Human Sociality: Experimental and Ethnographic Evidencefrom 15 Small Scale Societies’, Oxford University Press, chapter 3,pp. 55–96.

Cardenas, J. C. & Carpenter, J. P. (2008), ‘Behavioural developmenteconomics: Lessons from field labs in the developing world’, Journalof Development Studies 44(3), 311–338.

Carter, M. & Castillo, M. (2010), ‘Trustworthiness and social capital inSouth Africa: Analysis of actual living standards data and artefac-tual field experiments’, Economic Development and Cultural Change. Forthcoming.

Das, J., Dercon, S., Habyarimana, J., Krishnan, P., Muralidharan, K.& Sundararaman, V. (2010), When can school inputs improve testscores? Unpublished manuscript.

DellaVigna, S., List, J. A. & Malmendier, U. (2010), Testing for al-truism and social pressure in charitable giving. Unpublished Manu-script.

Durkheim, E. (2001), The Elementary Forms of Religious Life, OxfordUniversity Press, Oxford.

Fafchamps, M. (1992), ‘Solidarity networks in preindustrial societies:Rational peasants with a moral economy’, Economic Development


and Cultural Change 41(1), 147–174.Falk, A., Fehr, E. & Fischbacher, U. (2008), ‘Testing theories of fairness

- Intentions matter’, Games and Economic Behavior 62, 287–303.Fehr, E. & Gachter, S. (2000a), ‘Cooperation and punishment in public

goods experiments’, American Economic Review 90(4), 980–994.Fehr, E. & Gachter, S. (2000b), ‘Fairness and retaliation: The econom-

ics of reciprocity’, Journal of Economic Perspectives 14(3), 159–181.Fisman, R., Kariv, S. & Markovits, D. (2007), ‘Individual preferences

for giving’, The American Economic Review 97(5), 1858–1876.Forsythe, R., Horowitz, J., Savin, N. & Sefton, M. (1994), ‘Fairness

in simple bargaining experiments’, Games and Economic Behavior6, 347–369.

Glaeser, E. L., Laibson, D., Scheinkman, J. & Soutter, C. L. (2000),‘Measuring trust’, Quarterly Journal of Economics 115(3), 811–846.

Goetz, A. M. & Gupta, R. S. (1995), ‘Who takes the credit? Gen-der, power, and control over loan use in rural credit programs inBangladesh’, World Development 24(1), 45–63.

Gurven, M. (2004), Does market exposure affect economic game be-havior? The ultimatum game and the public goods game among theTsimane of Bolivia, in J. Henrich, R. Boyd, S. Bowles, C. Camerer,E. Fehr & H. Gintis, eds, ‘Foundations of Human Sociality’, OxfordUniversity Press, Oxford, pp. 196–231.

Habyarimana, J., Humphreys, M., Posner, D. N. & Weinstein, J. M.(2007), ‘Why does ethnic diversity undermine public goods provi-sion?’, American Political Science Review 101(4), 709–725.

Henrich, J., McElreath, R., Barr, A., Ensminger, J., Barrett, C.,Bolyanatz, A., Cardenas, J. C., Gurven, M., Gwako, E., Henrich,N., Lesorogol, C., Marlowe, F., Tracer, D. & Ziker, J. (2006), ‘Costlypunishment across human societies’, Science 312(5781), 1767–1770.

Hoffman, E., McCabe, K., Shachat, K. & Smith, V. (1994), ‘Prefer-ences, property rights, and anonymity in bargaining games’, Gamesand Economic Behavior 7(3), 346–380.

Hoffman, E., McCabe, K. & Smith, V. (1996), ‘Social distance andother-regarding behavior in dictator games’, American Economic Re-view 86(3), 653–660.

Jalan, J. & Ravallion, M. (1999), ‘Are the poor less well insured? Ev-idence on vulnerability to income risk in rural China’, Journal ofDevelopment Economics 58(1), 61–81.

Lehmann, E. L. (1975), Nonparametrics: Statistical methods based onranks, Holden-Day, San Francisco.


Leider, S., Mobius, M. M., Rosenblat, T. & Do, Q.-A. (2009), ‘Directedaltruism and enforced reciprocity in social networks’, Quarterly Jour-nal of Economics 124(4), 1815–1851.

Ligon, E. (2004), Targeting and informal insurance, in S. Dercon, ed.,‘Insurance Against Poverty’, Oxford University Press.

Ligon, E. & Schechter, L. (2010), Structural experimentation to dis-tinguish between models of risk sharing with frictions. UnpublishedManuscript.

Ligon, E., Thomas, J. P. & Worrall, T. (2002), ‘Informal insurancearrangements with limited commitment: Theory and evidence fromvillage economies’, Review of Economic Studies 69, 209–244.

Ostrom, E., Walker, J. & Gardner, R. (2000), ‘Collective action andthe evolution of social norms’, American Political Science Review14(3), 137–158.

Posner, R. (1980), ‘A theory of primitive society, with special referenceto law’, Journal of Law and Economics 23, 1–53.

Schechter, L. (2007), ‘Theft, gift-giving, and trustworthiness: Honestyis its own reward in rural Paraguay’, American Economic Review97(5), 1560–1582.

Scott, J. C. (1976), The Moral Economy of the Peasant, Yale UniversityPress, New Haven.

Slonim, R. & Garbarino, E. (2008), ‘Increases in trust and altruismfrom partner selection: Experimental evidence’, Experimental Eco-nomics 11(2), 134–153.

Townsend, R. M. (1994), ‘Risk and insurance in village India’, Econo-metrica 62(3), 539–591.

Wilson, E. O. (1978), On Human Nature, Harvard University Press.

Appendix A. Controlling for Order Effects

Let us number the Anonymous-Random game 1, the Revealed-Randomgame 2, the Anonymous-Chosen game 3, and the Revealed-Chosengame 4. There are 12 orders possible since Games 1 and 2 may beseparated, but games 3 and 4 were never separated. As explanatoryvariables representing order effects we control for the effect of game 2preceding game 1 on play in both games, the effect of games 2 and 1being separated from one another interacted with which came first onplay in both games, the effect of games 2 and 1 being separated on playin all four treatments, and the effect of the order of treatments 3 and4 on play in treatments 3 and 4. We also controlled for experimentereffects.



Everyone (Actual) 5.75∗∗∗ 0.31 0.24 0.20(371) (0.19) (0.25) (0.27) (0.42)

Chosen because needy 5.87∗∗∗ 0.56∗ 0.41 0.01(153) (0.24) (0.29) (0.30) (0.48)

Chosen because of affinity 5.66∗∗∗ 0.12 0.13 0.32(230) (0.22) (0.27) (0.29) (0.42)

Revelation unimportant 5.81∗∗∗ 0.39 0.24 0.00(285) (0.20) (0.26) (0.29) (0.43)

Revelation important 5.52∗∗∗ 0.00 0.29 0.85∗

(86) (0.33) (0.34) (0.32) (0.52)Everyone (Hypothetical) 6.81∗∗∗ 0.68 1.15∗∗ -0.53

(173) (0.37) (0.44) (0.51) (0.71)Chosen because needy 7.42∗∗∗ 0.43 0.71 0.00

(62) (0.48) (0.50) (0.53) (0.76)Chosen because of affinity 6.37∗∗∗ 0.92∗ 1.54∗∗∗ -0.93

(111) (0.41) (0.49) (0.54) (0.74)

Table A-1. Average value of motives, measured inKGs, and controlling for order and experimenter. Thetop panel reports means in the ‘Actual’ games, whilethe bottom reports values in the ‘Hypothetical’ games.Numbers in parentheses are standard errors. Asterisksindicate different levels of significance, with * indicat-ing that the corresponding mean is significantly differentfrom zero at a 90% level of confidence; ** at a 95% levelof confidence, and *** at a 99% level of confidence.

In Table A-1 one can see that the values of all four motives still tendto be positive. The value of sanctions still seems to be higher thanthe value of directed altruism. On the other hand, benevolence andreciprocity in the social network are both higher after controlling fororder effects, while sanctions and altruism are both smaller. It is stillthe case that the value of reciprocity in the social network is greaterwhen players state that the motive behind choosing the person did nothave to do with poverty and that they care whether the recipient knowsthat the money came from them. Directed altruism and benevolenceare still lower in these cases. So, the main results which held previouslycontinue to hold when controlling for order and experimenter effects.


The results from the hypothetical games controlling for order andexperimenter in Table A-1 also don’t change qualitatively. Those whochose a recipient because they were poor continue to have a higher valueof reciprocity and a lower value of directed altruism. These seeminglystrange results could be due to the fact that these people do not consti-tute a random sample as do the people who participated in the actualgames.

Appendix B. Summary Statistics for Players and theirHouseholds

Variable Mean (Std Dev)Player Age 48.01 (16.46)Player Male 63%Hh Annual Income (in $) 4,019 (5,627)Family Size 4.97 (2.39)# Adult Males 1.31 (0.90)OR-degree 6.62 (3.97)

Mean if non-zero (Std Dev) % non-zeroTheft Experienced (in $) 144.30 (283.96) 42.7%Amt Lent in Village (in $) 54.58 (125.47) 33.4%Amt Borrowed in Village (in $) 39.09 (57.09) 24.5%Gifts Given in Village (in $) 74.19 (129.37) 60.3%Gifts Received in Village (in $) 63.91 (109.30) 37.8%Observations 368

Appendix C. Hypothetical Experiments

In the short survey we ask the head of the recipient household hy-pothetical questions about how he would have played if he had beeninvited to participate in the economic experiments. In this case, we donot worry about whether these hypothetical dictators or their decisionscould be deduced so in order to simplify the explanation of the game wedid not propose to randomize the amounts received. Since the random-ization increases the expected amount received by 5 KGs, this meansthat the hypothetical amount received by the recipient is on average 5KGs less than the expected transfers in the actual games. (Note thatthis does not affect the expected marginal returns to sharing, whichremain equal to two.)

For the hypothetical games Table D-1 is the equivalent of Table 2for the actual games. The average dictator shares 63 percent of thetotal resources, while she shares 64 percent of the total resources in the


Game Anonymous Anonymous Revealed RevealedRandom Chosen Random Chosen

Anonymous 6.60 [64%] 6.26 3.57 27.30Random (3.44) (0.000) (0.068) (0.000)

Anonymous 41,65,67 7.08 [67%] 0.14 32.67Chosen (-3.46,3.34) (3.21) (0.770) (0.000)

Revealed 46,61,66 59,59,55 7.17 [68%] 18.24Random (-2.93,3.55) (-3.02,3.55) (3.35) (0.000)Revealed 31,54,88 20,77,76 33,62,78 8.10 [73%]

Chosen (-3.00,4.00) (-3.40,3.22) (-3.27,3.44) (3.29)

Table D-1. Comparison of Transfers in Different Hypo-thetical Games. Cells on the diagonal report the meantransfer given in each game in KGs; the proportion ofthe expected total pot shared by the dictator (in squarebrackets); and the standard deviation of the transfer (inparentheses). Friedman statistics for a test of a signif-icant difference between transfers between each pair ofgames are presented above the diagonal, with p-values inparentheses. Cells below the diagonal report the count ofobservations in which the transfer in the row game wasless than, equal to, or greater than respectively; and inparentheses report the mean difference in transfers be-tween the row game and the column game conditionalon the transfer in the row game being respectively lessthan or greater than the transfer in the column game.

hypothetical game. While sharing in the ‘hypothetical’ games is con-sistently about 30 percent greater than in the corresponding ‘actual’games when measured in terms of Guaranies, the difference when mea-sured as a share of the expected total resources in each game is muchsmaller, since the expected total resources in the hypothetical game are5 KGs less than in the actual games (recall that the random additionisn’t a feature in the hypothetical games). We also find that, again, wecan reject equality in transfers between all games except between theAnonymous-Chosen and the Revealed-Random games.

For the hypothetical games Table D-2 is the equivalent of Table 3for the actual games. Again, we find similar patterns for the two typesof games. The incentive-based motives of S and R are seen to accountfor 13% of the total motives in the hypothetical games compared to9% in the actual games. When we perform the exercise in Table 4



Everyone (Hypothetical) 6.60∗∗∗ 0.47∗∗ 0.57∗∗ 0.45(173) (0.26) (0.24) (0.24) (0.28)

Chosen because needy 7.26∗∗∗ 0.21 −0.48 1.13∗∗∗

(62) (425) (359) (336) (413)Chosen because of affinity 6.25∗∗∗ 0.62∗∗ 0.93∗∗∗ 0.018

(111) (0.33) (0.31) (0.31) (0.37)

Table D-2. Average value of motives for the hypothet-ical games, measured in KGs. Numbers in parenthesesare standard errors. Asterisks indicate different levelsof significance, with * indicating that the correspondingmean is significantly different from zero at a 90% level ofconfidence; ** at a 95% level of confidence, and *** at a99% level of confidence.

attributing variance to the different motives for the hypothetical gameswe find similar results. Variation in Bi accounts for 23% of the variancein transfers, in Ai accounts for 23% of the variance, in Si accountsfor 37% of the variance, and in Ri accounts for 17% of the variance.This would imply that incentive-based motives contribute 47% of thevariance in transfers in the actual games, and 54% of the variance inthe hypothetical games.

Appendix D. Survey and Game Details and GameProtocol

[The following instructions were read to the participants.]Thank you very much for coming today. Today’s games will last two

to three hours, so if you think that you will not be able to remain thewhole time, let us know now. Before we begin, I want to make somegeneral comments about what we are doing and explain the rules ofthe games that we are going to play. We will play some games withmoney. Any money that you win in the games will be yours. [ThePI’s name] will provide the money. But you must understand thatthis is not [his/her] money, it is money given to [him/her] by [his/her]university to carry out [his/her] research.

All decisions you take here in these games will be confidential, or,in some cases, also known by your playing partner. This will dependon the game and we will inform you in advance whether or not yourpartner will know your identity.


Before we continue, I must mention something that is very impor-tant. We invited you here without your knowing anything about whatwe are planning to do today. If you decide at any time that you do notwant to participate for any reason, you are free to leave, whether ornot we have started the game. If you let me know that you are leaving,I’ll pay you for the part of the game that you played before leaving. Ifyou prefer to go without letting me know, that is fine too.

You can not ask questions or talk while in the group. This is veryimportant. Please be sure that you understand this rule. If a persontalks about the game while in this group, we can not play this gametoday and nobody will earn any money. Do not worry if you do notunderstand the game well while we discuss the examples here. Each ofyou will have the opportunity to ask questions in private to make sureyou understand how to play.

This game is played in pairs. Each pair consists of a Player 1 anda Player 2 household. [The PI’s name] will give 14,000 Guaranies toeach of you who are Player 1s here today. Player 1 decides how muchhe wants to keep and how much he wants to send to Player 2. Player1 can send between 0 and 14,000 Gs to Player 2. Any money sent toPlayer 2 will be doubled. Player 2 will receive any money Player 1 sentmultiplied by two, plus an additional contribution from us. Player 1takes home whatever he doesn’t send to Player 2. Player 1 is the onlyperson who makes a decision. Player 1 decides how to divide the 14,000Gs and then the game ends.

The additional contribution is determined by the roll of a die. Theadditional contribution will be the roll of the die multiplied by 2 if itlands on any number between 1 and 5. If it lands on 6, there will beno additional contribution. Thus, if it lands on 1 there will be 2,000additional for Player 2, if it lands on 2 there will be 4,000 additionalfor Player 2, if it lands on 3 there will be 6,000 additional for Player2, if it lands on 4 there will be 8,000 additional for Player 2, and if itlands on 5 there will be 10,000 additional for Player 2. But if it landson 6 there will not be any additional contribution for Player 2.

Now we will review four examples. [Demonstrate with the Guaranimagnets, pushing Player 1’s offer to Player 2 across the magnetic black-board.]

(1) Here are the 14,000 Gs. Imagine that Player 1 chooses to send10,000 Gs to Player 2. Then, Player 2 will receive 20,000 Gs(10,000 Gs multiplied by 2). Player 1 will take home 4,000 Gs(14,000 Gs minus 10,000 Gs). If the die lands on 5 , Player 2 willreceive the additional contribution of 10,000 Gs, which means


he will receive 30,000 total. If the die lands on 1, Player 2 willreceive the additional contribution of 2,000 Gs, which means hewill receive 22,000 total.

(2) Here is another example. Imagine that Player 1 chooses tosend 4,000 Gs to Player 2. Then, Player 2 will receive 8,000 Gs(4,000 Gs multiplied by 2). Player 1 will take home 10,000 Gs(14,000 Gs minus 10,000 Gs). If the die lands on 3, Player 2 willreceive the additional contribution of 6,000 Gs, which means hewill receive 14,000 total. If the die lands on 6, Player 2 will notreceive any additional contribution, which means he will receive8,000 total.

(3) Here is another example. Imagine that Player 1 chooses toallocate 0 Gs to Player 2. Then, Player 2 will receive 0 Gs.Player 1 will take home 14,000 Gs (14,000 Gs minus 0 Gs). If thedie lands on 2, Player 2 will receive the additional contributionof 4,000 Gs, which means he will receive 4,000 total.

(4) Here is another example. Imagine that Player 1 chooses toallocate 14,000 Gs to Player 2. Then, Player 2 will receive28,000 Gs (14,000 Gs multiplied by 2). Player 1 will take home0 Gs (14,000 Gs minus 14,000 Gs). If the die lands on 4, Player2 will receive the additional contribution of 8,000 Gs, whichmeans he will receive 36,000 total.

That’s how simple the game is. We will play four different versionsof this game. Player 2 will always be a household in this community.

1.) In one version, Player 2’s household will be chosen by a lottery.The same family can be drawn multiple times. It could be someonewho is participating in the games here today, or it could be anotherhousehold in this company. It can not be your own household. Youwill not know with whom who you are playing. Only [the PI’s name]knows who plays with whom, and [he/she] will never tell anyone. Theymay be happy to receive a lot of money but can not thank you, or theymay be sad to receive a little money but they can not get angry withyou, because they are never going to know that this money came fromyou. You will not know the roll of the die in this version of the game.

2.) In another version, Player 2’s household will also be chosen bya lottery. The same household can be drawn multiple times. In thisversion you will discover the identity of Player 2 after all of the gamestoday, and Player 2 will also discover your identity. After the gameswe’ll go to the randomly drawn Player 2’s house and we will explainthe rules of the game to him and we will explain that John Smith gaveso much money and then the die landed in such a way, but that when


John Smith was deciding how much to give he did not know who themoney was going to. They may be happy to receive a lot of money,and will be able to thank you, or they may get angry with you if theyreceive little money, because they will know that the money was sentby you.

3 and 4.) In the next two versions, you can choose the identity ofPlayer 2. You can choose any household in this village and we willgive the money to someone in that household who is over 18. Therewill be two versions of this game, only one of which will count for yourearnings today. You must choose the same household as recipient inthese two games, and you can not choose your own household.

3.) In one version, we will not tell Player 2’s household that youchose them and we will make it difficult for them to figure out youridentity. That person will never know that you were the one who sentthe money. They may be happy to receive a lot of money, or sad toreceive little money, but they have no way of figuring out that themoney came from you. Even if you go to them afterwards and tellthem that you chose them and sent them money, they may not believeyou. You will not know the exact amount they received because weadd the additional contribution to the amount sent and also becausethey will receive all their earnings together at the same time as someamount X. They will not know which part of it comes from whom, orif they were chosen by a Player 1, or chosen by the lottery.

4.) In the other version we will tell Player 2’s household that youchose him to send money to and you will both know the roll of the die.He can be angry with you if you send little or thank you if you send alot.

After all of you play all four versions, I will toss a coin. If the coinlands on heads, the Player 2 household you chose will know who chosethem. I will go to their house and give them the money, and explainthe rules of the game to them, and I will tell them that you chose themand tell them how much money you sent them. If the coin lands ontails, the Player 2 household you chose will not know who sent themthe money. We will not tell them that the money came from you, andthey will not be able to find out. Remember, you decide how muchyou want to send when you choose the household and they know thatthe money comes from you, and how much you want to send when thehousehold won’t find out where the money comes from. But in thisvillage only one of these two versions will count for money, dependingon the toss of a coin. I will toss the coin in front of you after you haveall played.


We now are going to talk personally with each of you one-on-one toplay the game. You will play with either [Investigator 1] or [Investi-gator 2] in private. We will explain the game again and ask you todemonstrate your understanding with a couple of examples. You willplay the game with real money. Please do not speak about the gamewhile you are waiting to play. You can talk about soccer, the weather,medicinal herbs, or anything else other than the games. You also haveto stay here together; you can not go off in small groups to talk quietly.Remember, if anyone speaks of the game, we will have to stop playing.

Dialogue for the GameSuppose that Player 1 chooses to send 7,000 Gs to Player 2. In this

case, how much would Player 1 take home? [7,000 Gs ] How muchwould Player 2 receive? [14,000 Gs ] What if the die falls on 3, whatwould the additional contribution be? [6,000 Gs ] So how much wouldPlayer 2 receive in total? [20,000 Gs ] What if the die falls on 1, whatwould the additional contribution be? [2,000 Gs ] So how much wouldPlayer 2 receive in total? [16,000 Gs ]

[The order of playing these games is randomly chosen for each player.]Here I give you four small stacks of 14,000 Gs each, for a total of

56,000 Gs.

• Now we will play the game in which neither you nor Player 2will know each other’s identity. They may be happy to receivea lot of money but they can not thank you, or they may besad to receive little money but they can not get angry withyou. This is because they are never going to know that thismoney came from you. Take one of the stacks of 14,000 Gs.Please give me the amount you want me to give to Player 2’shousehold, or if you do not want to give anything then don’thand me anything. I will double any money you give me andadd the additional contribution to it and give it to a randomlychosen household in your village.• Now we will play the game in which you and Player 2 will know

each other’s identity after the end of the games today. Theymay be happy to receive a lot of money, and will be able tothank you or they can get sad when receiving little money, andwill be able to get angry with you. This is because they willknow that the money was sent by you. Take one of the stacksof 14,000 Gs. Please give me the amount you want me to giveto Player 2’s household, or if you do not want to give anythingthen don’t hand me anything. I will double any money you give


me and add the additional contribution to it and give it to arandomly chosen household in your village and inform them ofthe rules of the game and explain how much you sent and thatyou sent it without knowing to whom you were sending.• In the next two games you choose the household to which you

want to send money. Now, tell me which household do you wantto send money to?• Now we will play the game in which the recipient household is

not going to know that you chose them. Take one of the stacksof 14,000 Gs. Please give me the amount you want me to give to[name], or if you do not want to give anything then don’t handme anything. I will double any money you give me and add theadditional contribution to it. They are not going to be able tofigure out who chose them. They may be happy to receive a lotof money, or sad to receive little money, but they have no wayof figuring out that the money came from you. Even if you tellthem that you chose them and sent them money, they may notbelieve you. You will not know the exact amount they receivedbecause we add the additional contribution to the amount sentand also because they will receive all their earnings together atthe same time as some amount X. They will not know whichpart of it comes from which person, or if they were chosen by aPlayer 1, or chosen by the lottery.• Now we will play the game in which the recipient household will

know that you chose them. Take one of the stacks of 14,000 Gs.Please give me the amount you want me to give [name], or ifyou do not want to give anything then don’t hand me anything.I will double any money you give me and add the additionalcontribution to it and give it to Player 2’s household and tellthem the rules of the game and explain that you chose themand explain how much you sent. They can be angry with youif you send little or thank you if you send a lot.

Now you must wait while the rest of the players make their decisions.Remember that you can not talk about the game while you are waitingto be paid. Please go outside to chat a bit with the enumerator namedEver before exiting.

The End[After all participants have made their decisions, talk to them as a

group one last time.] Now I will flip a coin. [If heads:] The coin landedheads, which means that the Player 2 household you chose will know


who chose them and how much money they sent. [If tails:] The coinlanded tails, which means that the Player 2 household that you chosewill not discover who sent them money. Now I will speak with you oneat a time one last time to give you your winnings and to tell you whowas drawn in the lottery to receive money from you in the revealedversion of the game.

[Call players in one at a time.] In the anonymous game you kept [XGs ]. In the game in which you will discover who you sent the moneyto, you kept [Y Gs ] and [name] received [M Gs ] since their name waschosen in the lottery. In the game in which you chose your partner and[if the coin landed heads ] he will know who sent him the money [or ifthe coin landed tails ] he will not find out who sent him the money, youkept [Z Gs ], [and if the coin landed heads ] so Player 2 received [M Gs ].

[If received in anonymous game or chosen game:] You also received[G Gs ] from an anonymous Player 1. [If received in revealed game:]You also received [H Gs ] from a Player 1 who did not know he wasplaying with you and his name is [name each] and he sent you thisamount [M ] which was doubled and then the die landed on [D ]. [Ifreceived in chosen revealed game:] You also received [J Gs ] in totalfrom a Player 1 who chose you and their name is [name each] and hesent you this amount [M ] which was doubled and then the die landedon [D ].

That means you have won a total of [X + Y + Z + G + H + J Gs ].Thank you for playing with us here today. Now the game is over. Afterwe finish handing out the money here, we will go to the households ofthe appropriate Player 2s to give them their winnings.

University of California, Berkeley

Michigan State University

motives for sharing in social networksare.berkeley.edu/~ligon/papers/network_motives.pdfmotives for...

Documents