Does informal risk sharing induce lower efforts?
Evidence from lab-in-the-field experiments in rural Mexico
16-34
Ingela Alger, Laura Juarez,
Does informal risk sharing induce lower e¤orts?
Evidence from lab-in-the-…eld experiments in rural Mexico
Ingela Alger
y, Laura Juarez
z,
Miriam Juarez-Torres
x, and Josepa Miquel-Florensa
{February 9, 2016
Abstract
How does informal risk sharing a¤ect incentives to avoid risk? While moral hazard is expected under formal insurance, theory suggests that the incentive e¤ects of informal risk sharing are ambiguous: internalization of the external e¤ects of transfers on others may reduce or enhance incentives to avoid risk. To study this issue, which is particularly relevant for developing economies, we designed a novel real-e¤ort lab experiment and conducted it in 16 small villages in rural Mexico. We …nd that subjects internalize the e¤ects of transfers enough for the presence of transfers to signi…cantly increase e¤ort compared to autarky situations.
Keywords: informal insurance, e¤ort, moral hazard, free-riding e¤ect, empathy e¤ect
JEL codes: D64, C93, O12
We thank the Agence Nationale de la Recherche for funding this research through a grant awarded to Ingela Alger (Chaire d’Excellence ANR-12-CHEX-0012-01). Ingela Alger and Josepa Miquel-Florensa further gratefully acknowledge the generous hosting by Banco de México. We are indebted to Ing. Mario Rivera de Labra (INIFAP Campeche), whose help was invaluable. Finally, this paper would not have seen the light of day without the outstanding research assistance, both in the …eld and in the o¢ ce, provided by Mariana Garcia and Ildrim Valley. The views expressed in this article are solely those of the authors and do not necessarily re‡ect those of Banco de Mexico.
yToulouse School of Economics (CNRS) and Institute for Advanced Study in Toulouse. [email protected]
zBanco de México. [email protected] xBanco de México. [email protected]
1
Introduction
Risk is an unavoidable feature of life. In countries where formal insurance institutions are non-existent or weak, transfers between households tend to be widespread, as is shown for instance by the cross-country analysis of Cox, Galasso and Jimenez (2006).1 A large empirical literature
has established that these transfers— which stem mostly from the extended family,2 but also
from friends, and neighbors— do indeed provide insurance, albeit partially; see Rosenzweig (1988), Udry (1990), Townsend (1994), Ligon, Thomas and Worrall (2002), Fafchamps and Lund (2003), Dercon and Krishnan (2003), Dubois, Jullien, and Magnac (2008). Clearly, if the risk people face is beyond their control, risk sharing can only be bene…cial. However, in most cases peoplecan a¤ect the amount of risk they face, e.g., by working, saving or investing. Then, the overall welfare consequences of risk sharing must also account for how it a¤ects the incentives to avoid risk in the …rst place.
Theory suggests that moral hazard, whereby insurance reduces the incentive to avoid risk, is an issue in formal insurance markets (Arrow, 1963, Arnott and Stiglitz, 1988). By contrast, when risk sharing takes place between individuals who are kin, friends, or neighbors, the underlying motivation for why individuals choose to help each other may also a¤ect their incentives to avoid risk, as shown by Alger and Weibull (2008, 2010).3 In their model two
individuals exert productive e¤orts and each individual may transfer part of his/her output to the other. Depending on whether the transfers are driven by a social norm or by altruism, such informal risk sharing may or may not have a detrimental e¤ect on the incentives to exert e¤ort. In theory, productive e¤ort may even be higher with informal risk sharing than if individuals lived in autarky. This result has important implications for economies where informal risk sharing is common: indeed, an important detrimental incentive e¤ect could mean that informal risk sharing hampers economic development, while a positive incentive e¤ect would imply that it promotes economic development. It is therefore important to understand if and how informal risk sharing a¤ects the incentives to avoid risk. In this paper we report the results from a novel experiment that was designed speci…cally for this purpose.
More precisely, our aim is to provide evidence on how the existence of transfers— both the pressure to make them and the expectation to receive them— a¤ects incentives to make productive e¤orts. To this end we designed a real-e¤ort experiment in which incentive e¤ects are distilled by imposing the size and direction of any transfers on subjects. Each subject was exposed to four e¤ort games, in which e¤ort determined the probability of success, and success meant that the subject generated additional income on top of the initial endowment.
1See also Fafchamps (2011) for a survey of the literature on inter-household transfers. 2See Cox and Fafchamps (2008) for a survey of the literature on extended kinship networks. 3For a review of theoretical models of intra-family risk sharing, see Laferrère and Wol¤ (2006).
In autarky games, the subject simply got to keep any additional income in case of success. In transfer games, subjects allocated to the donor role had to make a transfer (of …xed size) in case of success, and subjects allocated to the recipient role received a transfer (of …xed size) in case of failure. Reciprocity motives were ruled out by ensuring that e¤ort choices as well as payo¤s were private information, and by matching subjects anonymously. Across the transfer games played by any given subject, the e¤ect of e¤ort on own payo¤ was kept constant; in contrast, the identity of the (passive) individuals, also a¤ected by the e¤ort and subsequent transfer, varied from being one individual to be a group of individuals or the local health centre. Furthermore, across subjects we varied the size of the transfer as well as the role (donor vs recipient), and hence there were four di¤erent treatments.
The experiment was conducted in 16 small villages in rural Mexico.4 These villages
rep-resent a sample of the Mexican population for which informal risk sharing is still among the main sources of insurance.
The experimental design allows us to test several hypotheses. First, from within-subject comparison of e¤ort under autarky and in transfer games we can deduce whether transfers have a positive, a negative, or no e¤ect on e¤ort. Second, the payo¤ structure in some treatments allows us to use within-subject analysis to infer whether subjects derive positive, negative, or no utility from transfers. Third, between-subjects comparison of e¤ort in donor and recipient treatments reveal whether transfers a¤ect the incentives of donors and recipients di¤erently. Fourth, within-subject comparison allows us to analyze whether the identity of the passive recipients or donors a¤ected the subjects’willingness to provide e¤ort.
Our …ndings are as follows. First, transfers have a signi…cant, positive, and non-negligible e¤ect on e¤ort. This conclusion is based on a large number of within-subject comparisons of the subject’s e¤ort in a game that involves a transfer and in an autarky game, and the …nding is robust across treatments. Second, since transfers have a signi…cant and positive e¤ect on e¤ort even in treatments where the transfer represents 100% of the potential additional income, we can safely conclude that, on average, subjects derive positive utility from transfers. Third, the increase in e¤ort between autarky and transfer games is consistent across donor and recipient roles. Fourth, the positive e¤ect of transfers on e¤ort is signi…cantly larger when the transfer a¤ects the local health center than when it a¤ects other subjects. Finally, di¤erences across villages in religious homogeneity and indigenous population allow us to analyze how the results relate to these village characteristics. We …nd that the positive e¤ect on e¤ort of a transfer is signi…cantly higher in villages where the population is more homogeneous when
4There is a large literature on behaviours in small villages in Mexico, and access to high quality data about
it comes to religious denomination.5
Although our experimental design is original, there is a large experimental literature that is related to our results. First, it is well known that many undergraduate students in developed countries are willing to give in anonymous dictator games, which suggests that they derive a bene…t from giving in such situations (Blanco, Engelmann, and Normann, 2011). These …ndings are corroborated in experiments in developing countries, with twists on the anonymity of the recipient, and the origin of the income to be shared.6 If given the possibility to pay a
price to avoid giving, however, some subjects use this option (Dana, Weber, and Kuang, 2007). In particular, Jakiela and Ozier (2016) report that many of their Kenyan subjects were willing to pay a price to hide income obtained in an investment game, the interpretation being that these subjects want to avoid a “kin tax” (the experiment allowed to control for relatedness). In sum, the experimental evidence on whether individuals derive utility from making transfers to others or from avoiding that others make transfers to them is mixed. Importantly, none of this evidence relies on real-e¤ort experiments, and none of this evidence was collected in Latin America.
In the non-experimental empirical literature on transfers, it appears that all the three motives for helping on which it has focused— altruism, favor exchange, and reciprocity— are at work.7 Individuals thus appear to derive a bene…t from helping others, consistent with our
…ndings. However, there is also some empirical evidence, based entirely on African data, that suggests that “kin taxes” have a detrimental e¤ect on the incentives to invest and/or save.8 5Even though our subjects do not know the individual identity of the donor or recipient with whom they
are being paired, in villages that are more homogeneous in terms of religious denomination the likelihood of being paired with another individual who shares the subjects’ religious beliefs is higher, which can have an e¤ect on sharing behavior. Relatedly, Barr, Dekker and Fafchamps (2012) …nd that individuals belonging to the same religious group are more likely to share risk when agreements are enforced through social sanctioning than when they are so by the experimenter. Regarding other economic outcomes, Fisman, Paravisini and Vig (2012) show that shared religious beliefs between borrowers and lenders a¤ect outcomes like loan amounts and repayment.
6Thus, Ligon and Schechter (2012) conclude that the largest proportion of observed transfers in anonymous
dictator games among villagers in rural Paraguay can be attributed to preference-related motives. Jakiela (2015) compares the behavior of U.S. students to Kenyans to study whether the willingness to share is di¤erent when the endowment in the dictator game derives from luck and when it is the result of a risk-free e¤ort. In a comparative experimental study in which subjects could give own wealth, or steal or destroy others’wealth, Fafchamps and Hill (2015) found that some gave and some stole, and that Kenyan and Ugandan subjects behaved more pro-socially than British subjects.
7See, e.g., Lucas and Stark (1985), Altonji, Hayashi, and Kotliko¤ (1992), Cox, Eser, and Jimenez (1998),
Foster and Rosenzweig (2001), De la Brière et al. (2002), and De Weerdt and Fafchamps (2011).
8See Azam and Gubert (2005), Baland, Guirkinger, and Mali (2011), Di Falco and Bulte (2011, 2013),
Finally, there is no evidence on whether transfers within rural communities9 tend to have a
net negative or a net positive e¤ect on productive e¤orts.
The results of the novel, real-e¤ort experiment in this paper show that individuals get positive utility from sharing. This implies that informal risk sharing within our participating villages might increase material welfare, not only because transfers within these villages allow to smooth income over states of nature, but also because they induce higher productive e¤orts. In the next section we present a theoretical model that delivers predictions and inferences that are used to design the experiment and to interpret the results of the empirical analysis. In Section 3 we describe the experiment, and in Section 4 we provide information about the …eld setting and the sampled population. The empirical strategy as well as the results are presented in Section 5, and Section 6 concludes.
2
Theory
Consider a variant of the e¤ort-and-transfer game proposed by Alger and Weibull (2008) in which an individual i, who has initial wealthyi, may generate additional income by making e¤ort. The e¤ort level determines the probability that success occurs, in which case some additional income, z, is generated; with the complementary probability, when failure occurs no additional income is generated. Let p(e) be the probability of success associated with e¤ort e 2 [0; e], where p is an increasing, strictly concave, and di¤erentiable function. The individual’s …nal wealth also depends on whether there is a transfer, t. We distinguish three cases. First, under autarky, i simply gets yi in case of failure and yi +z in case of success. Second,imay be in adonor role, in which case she makes a transfert to another individual,j, in case of success, so that she getsyi in case of failure andyi+z t in case of success. Finally, i may be in a recipient role, in which case she receives a transfer t from another (passive) individual, j, in case of failure, so that she gets yi +t in case of failure and yi +z in case of success. In contrast with Alger and Weibull (2008), where both agents make productive e¤orts, here we assume that j is passive; as in the original model, however, we take the size of the transfer to be exogenously given.
Furthermore, assume that (a) i has private information about her e¤ort level and her realized material payo¤,10 (b) if i’s e¤ort has a consequence for another individual’s …nal 9Azam and Gubert (2005) collected data on productive e¤orts of farmers in villages in Mali, and compared
the e¤orts of those who receive remittances from abroad on a regular basis to the e¤orts of those who do not receive such remittances. By contrast, we impose transfers between individuals within the same village, thus seeking to capture a more symmetric situation.
material payo¤ then this is the private information of that individual, and (c) any such payo¤ consequence is anonymous in the sense that i does not know the identity ofj and vice versa. Under these assumptions, the willingness to provide e¤ort cannot be driven by a strategic repeated-interaction e¤ect, and must stem from the individual’s preferences.
Next, we derive predictions and behavior-based inferences about preferences for the three situations at hand.
2.1
Autarky
Under autarky,i cares only about own consumption utility and disutility of e¤ort. Assuming that utility is additively separable in consumption utility and disutility of e¤ort, let u(wi) denote the consumption utility that i derives from wealth wi, and c(e) the cost of making e¤orte, whereuis strictly increasing and concave, andcis increasing and strictly convex with
c0(0) = 0. The individual’s expected utility then writes
p(e)u(yi+z) + [1 p(e)]u(yi) c(e): (1)
The autarky e¤ort eA ise if
c0(e) p0(e) [u(yi+z) u(yi)]; (2) and otherwise it is implicitly de…ned as a function ofz by the …rst-order condition
c0(e) = p0(e) [u(yi+z) u(yi)]: (3) Standard arguments imply:
Proposition 1 The autarky e¤ort level,eA(z), is increasing in the income generated by
suc-cess, z. Moreover, eA(0) = 0.
2.2
Donor role
Wheni is in a donor role she has to transfert >0 to someone else in case of success, so that her material payo¤ is yi+z t in case of success and yi otherwise.11 LeteD(z; t) denote i’s e¤ort when she is in a donor role. We here show that inferences about i’s preferences can be drawn by comparing this e¤ort to the e¤ort she would make if she were sel…sh. Thus, as a
11Here we do not model whyiwould …nd herself in a donor role, but focus instead on the consequences that
benchmark, suppose that i issel…sh in the sense that she cares only about own consumption utility and disutility from e¤ort, she maximizes
p(e)u(yi+z t) + [1 p(e)]u(yi) c(e): (4)
Letting e0
D(z; t) denote i’s e¤ort if she is sel…sh, and focusing on the case e0D(z; t) < e, this e¤ort is implicitly de…ned by the …rst-order condition
c0(e) =p0(e) [u(yi+z t) u(yi)]: (5) Note that if t = 0, the e¤ort is the same as under autarky: e0
D(z;0) = eA(z). Hence, Proposition 1 immediately implies:
Corollary 1 Suppose that iis sel…sh and that she makes a positive e¤ort under autarky, i.e.,
eA(z)>0. Then:
(i) She chooses a lower e¤ort level if she has to transfer some or all of her income in case of
success to another individual: e0
D(z; t)< eA(z) for any t >0.
(ii) She makes no e¤ort if she has to transfer everything she earns in case of success: e0
D(z; t)jt=z =
0.
(iii) She makes the same e¤ort for any(z1; t1) and (z2; t2) such that z1 t1 =z2 t2.
Which theoretical predictions arise if individuali is not sel…sh in the sense de…ned above? Then, depending on her preferences she may choose to make more, less, or the same e¤ort as under autarky. As is well-known from the literature, individuals may be non-sel…sh in many di¤erent ways. Since our data does not allow us to make a horse race between alternative non-sel…sh preferences, we here limit ourselves to establishing inferences that can safely be made about an individual’s preferences based on comparison of her behavior in various roles.
To state the inferences, we distinguish two cases: full transfer and partial transfer.
Under full transfer, i has to transfer all of the additional income to j: t = z. As was established above, a sel…sh individual would then make no e¤ort, e0D(z; t)jt=z = 0. Hence, under our assumptions we can conclude the following.
Inferences 1: Donor, full transfer
(i) if eD(z; t)jt=z >0, the individual derives positive utility from making the transfer;
(ii) if eD(z; t)jt=z = 0, the individual is either sel…sh or she derives negative utility from making the transfer.12
12This conclusion would be slightly di¤erent if we had not assumedc0(0) = 0. Indeed, ifc0(0)>0, even an individual who derives a positive bene…t from the transfer could chooseeD= 0.
Under partial transfer, i has to transfer some of the additional income to j: z > t > 0. If an individual reduces e¤ort due to a transfer, i.e., if eD(z; t) < eA(z), it is impossible to infer whether the individual is sel…shly responding to the lower own material payo¤; deriving positive utility from the transfer that is nonetheless too small to outweigh the utility cost associated with the decrease in own material bene…t from making e¤ort; or, deriving negative utility from the transfer that pushes e¤ort below what it would have been without a transfer but with own material payo¤ dropping from z to z t. However, by comparing the e¤orts under autarky and in a donor role that has the same consequence for own material payo¤, the following conclusions can be drawn:13
Inferences 2: Donor, partial transfer
(i)if z1 t1 =z0, andeD(z1; t1)> eA(z0), the individual derives positive utility from making
the transfert1;
(ii)ifz1 t1 =z0, andeD(z1; t1)< eA(z0), the individual derives negative utility from making
the transfert1;
(iii) if z1 t1 =z0, and eD(z1; t1) =eA(z0), either the individual is sel…sh, or the positive or
negative utility she derives from making the transfert1 does not outweigh the change in the
e¤ort cost that an e¤ort adjustment would entail.
2.3
Recipient role
Wheniis in a recipient role she receives a transfert >0from another individual,j, if she fails to generate additional income herself by obtaining success. Letting eR(z; t) denote i’s e¤ort in a recipient role, we here show that several inferences can be drawn abouti’s preferences by comparing this e¤ort to the e¤ort she would make if sel…sh.
As a benchmark, then, if i is sel…sh she maximizes
p(e)u(yi+z) + [1 p(e)]u(yi+t) c(e): (6)
Letting e0
R(z; t) denote i’s e¤ort if she is sel…sh, and focusing on the case e0R < e, this e¤ort
13Note that these inferences can be generalized to increases in transfers:
Inferences 2’: Donor, partial transfer
(i)ifz1 t1=z2 t2,t1> t2, andeD(z1; t1)> eD(z2; t2), the individual derives positive utility from making
the additional transfert1 t2;
(ii)ifz1 t1=z2 t2,t1> t2, andeD(z1; t1)< eD(z2; t2), the individual derives negative utility from making
the additional transfert1 t2;
(iii)ifz1 t1=z2 t2,t1> t2, andeD(z1; t1) =eD(z2; t2), either the individual is sel…sh, or the positive or
negative utility she derives from making the additional transfert1 t2 does not outweigh the change in the
is implicitly de…ned by the …rst-order condition
c0(e) = p(e) [u(yi +z) u(yi+t)]: (7) Since e0
R(z;0) =eA(z), Proposition 1 immediately implies:
Corollary 2 Suppose that iis sel…sh and that she makes a positive e¤ort under autarky, i.e.,
eA(z)>0. Then:
(i) She chooses a lower e¤ort level if she receives a transfer from someone else in case of failure: e0
R(z; t)< eA(z) for anyt >0.
(ii) She makes no e¤ort if the transfer she receives is as large as the income she obtains when successful: e0
D(z; t)jt=z = 0.
Clearly, the following inferences can be drawn under full transfer (t = z) and partial transfer, z > t >0, respectively.
Inferences 3: Recipient, full transfer
(i) if eR(z; t)jt=z > 0, the individual derives positive utility from avoiding that someone else has to make a transfer to her;
(ii) if eR(z; t)jt=z = 0, the individual is either sel…sh or she derives negative utility from avoiding that someone else has to make a transfer to her.14
Inferences 4: Recipient, partial transfer
(i)ifz > t >0andeR(z; t) eA(z), the individual derives positive utility from avoiding that someone else has to make the transfert to her;
(ii) if z > t >0 and eR(z; t) < eA(z), either the individual is sel…sh, or she derives negative utility from avoiding that someone else has to make the transfertto her, or the positive utility that she derives from this does not outweigh the bene…t she incurs from making a lower e¤ort. To establish whether transfers have a positive, a negative, or no e¤ect on e¤ort, and whether individuals derive any utility or disutility from making or receiving transfers, we designed the “e¤ort-and-transfer experiment.” This experiment exposes subjects to variation in the payo¤ consequences of e¤ort for self and others in such a way that the results and inferences derived in this section can be applied. Next we describe the experiment in detail.
14This conclusion would be slightly di¤erent if we had not assumedc0(0) = 0. Indeed, ifc0(0)>0, even an individual who derives a positive bene…t from avoiding that another individual makes a transfer to her could chooseeR= 0.
3
The e¤ort-and-transfer experiment
In the experiment each subject played a total of four games (one of which was randomly drawn to calculate payments). In each game, the subject was given an initial endowment,y, and was presented with an opportunity to exert a real e¤ort that determined the probability, p(e), of generating additional income, z. Sources of variation across the four games played by a subject were: (1) the size of the additional incomez, (2) whether there was a transfer,t, and (3) who was a¤ected by a transfer if there was one. Importantly, the size and direction of transfers were never chosen by the subjects, but were instead imposed by us. In other words, the only decision that a subject had to make in each game was the e¤ort to provide (a detailed description of the e¤ort and lottery is provided below).
Across the subjects we varied the size of the forced transfer relative to the additional income
(partial vs full) as well as the role (donor vs recipient). There were thus four treatments:
Donor-Partial (DP), Recipient-Partial (RP), Donor-Full (DF), or Recipient-Full (RF). In
each session the subject pool was divided into two groups of equal size. One group played a
donor treatment, and the other group a recipient treatment, sequentially. Whichever group
was not actively playing was used (unbeknownst to them) as passive players in some of the games played by the active group.
The detailed description of the games, to which we turn next, shows which within- and between-subject comparisons the design allowed us to make.
Starting with the two Partial treatments, the initial endowment was y = 100 in each game.15 Furthermore, in all the games active subjects received a higher payo¤ in case of
success than in case of failure. Each subject played two autarky games, in which (s)he got to keep any additional income,z, that (s)he generated in case of success: Autarky Low, with z = 50, andAutarky High, withz = 75. Each subject also played two transfer games, both of which involved a forced transfer and entailed the same material payo¤ consequences for the active subject asAutarky Low:
In the Donor-Partial (DP) treatment, in both transfer games the additional income in case of success was z = 75, of which the active subject had to make a transfer t= 25. What di¤ered between the two transfer games was the identity of the recipient(s), all of whom received an initial endowment of 50:
in theOne-to-One game, the transfer was given to one individual, randomly drawn from
15All the numbers are expressed in points, the experimental unit, worth MXN 0.50 each. The overall
average payo¤ turned out to be MXN 92 (it was MXN 59.55.across the donor-full treatments, MXN 122.02 across the donor-partial treatments, MXN 62.30 across the recipient–full treatments, and MXN 122.43 across the recipient–partial treatments).
the group of passive individuals, and whose identity was never revealed to any subject; in the Pool game, the transfer was split equally among all the passive individuals. In the Recipient-Partial (RP) treatment, in both transfer games success led to an addi-tional income of z = 75, and in case of failure the active subject received a transfer t = 25. What di¤ered between the games was the identity of the donor(s), all of whom received an initial endowment of 200:
in the One-to-One game, the transfer was taken from one individual, randomly drawn by us from the group of passive individuals, and whose identity was never revealed to any subject;
in thePool game, all the passive individuals contributed equally to the transfer received by the active subject.
Turning now to the two Full treatments, these were designed to single out the e¤ect of transfers on e¤ort in a particularly stark manner: in the transfer games the active subject received the same payo¤ in case of success and in case of failure. In all the games, including
theAutarky game, an additional income of z = 75 was generated in case of success. Besides
playingAutarky High, each subject played three transfer games.
In theDonor-Full (DF) treatment, the initial endowment of active subjects wasy = 100
in all the games (including Autarky High). In the three transfer games, the subject had to transfer t= 75 in case of success, where the identity of the recipient(s) varied as follows:
in theOne-to-One game, the transfer was given to one individual, randomly drawn by us from the group of passive individuals, and whose identity was never revealed to any subject; passive individuals received an initial endowment of 25;
in thePool game, the transfer was split equally among all the passive individuals, all of whom received an initial endowment of 25;
in the Public Good game, the transfer was added to an initial sum of money (equal to 25 times the total number of passive subjects), and the resulting sum of money was given to a local public good: the health center of the village.
In theRecipient-Full (DF) treatment, the initial endowment of active subjects wasy= 25
in all the games (includingAutarky High). In the three transfer games, the subject received a transfert = 75 in case of failure, and the identity of the donor(s) varied as follows:
in the One-to-One game, the transfer was taken from one individual, randomly drawn by us from the group of passive individuals, and whose identity was never revealed to any subject; passive individuals received an initial endowment of 175;
in thePool game, all the passive individuals contributed equally to the transfer received by the active subject; passive individuals received an initial endowment of 175;
in thePublic Good game, the transfer was taken from an initial sum of money (equal to 175 times the total number of passive subjects), and the remaining sum of money was given to a local public good: the health center of the village.
The payo¤ consequences of success and failure for the active subject in all the games are summarized in Figure 1. In sum, we had a 2x2 factorial design that allowed us to study:
(a) within subjects: whether e¤ort is di¤erent in an autarky game and a transfer game that have the same payo¤ consequences for self (this comparison could be made in each treatment); (b) within subjects: whether e¤ort is di¤erent in transfer games that have the same payo¤ consequences for self but where the transfer a¤ects either one individual or a pool of individuals (this comparison could be made in each treatment);
(c) between subjects: comparing the behaviour of donors who all got the same initial endowment and the same additional income in case of success, we can analyze whether e¤ort is a¤ected by the size of the transfer (compare DP to DF).
Figure 1. Each cell shows the payo¤ consequences of success (S) and failure (F) for an active subject. The …rst number is the endowment; a number in italics is the additional income earned in case of success; any other number is a transfer (to the subject if preceded by a plus
sign, from the subject if preceded by a minus sign).
The e¤ort consisted in threading nuts onto bolts. For each fully threaded nut (all the nut-bolt pairs were identical), the subject increased the probability of generating additional income by 0.1, a probability that we set to zero for zero threaded nuts. In each game the subject had one minute to thread nuts onto bolts.16 Curtains ensured that each subject’s e¤ort
choice was unobservable by the other subjects. Furthermore, to minimize the experimental pressure to exert e¤ort each subject had on his/her table a fresh newspaper to look at during the imparted time. The games were color-coded, and at the end of each session one coloured card was picked from an opaque bag by a child, and this card determined which game would be used to calculate the subjects’payo¤s.17 Protocol and visual materials used in the sessions
are provided in the Appendix.
It is beyond argument that the task to thread nuts onto bolts in a short period of time is a task that requires understanding and ability, and that some learning may appear as the task is performed repeatedly. To get a measure of individual ability and learning, each subject did an ability test at the beginning and at the end of the experimental session. The ability test consisted in rewarding the participants with one point for each properly threaded nut in one minute. The way these two incentivized tests are used to control for ability and learning is explained in Appendix.
Finally, after the second ability test was performed, subjects were asked a question. The question pertained to the One-to-one game played by the subject. In the Donor-Partial
treatment we asked subjects if they would have liked to give a transfer larger than t= 25, in
the Donor-Full treatment we asked if they would have liked to give a smaller transfer than
t = 75, and how much more/less in case of a positive answer. Furthermore, we asked each subject in the recipient treatments if (s)he would reveal his/her e¤ort level to the individual with whom (s)he was matched in the One-to-one game if given the opportunity to do so.
Last, we asked all the subjects to …ll out a post-experimental questionnaire.
Figure 2 summarizes the session structures. The structures were the same for the twoFull
treatments and for the twoPartial treatments. The order of the games varied across sessions. Details about this can be found in the Appendix on learning.
focus on interior solutions. In the experiment the probability of success is linearly increasing in the number of threaded nuts and it may therefore appear that the cost of e¤ort is linear. However, since subjects were time-constrained e¤ort really consisted in the speed at which nuts were threaded, and we believe that it is therefore sensible to believe that marginal cost of e¤ort were indeed increasing in e¤ort. Furthermore, the fact that out of 2,144 observations (536 participants who played four games each) there are 23 observations where zero nuts were threaded, and 209 where ten nuts were threaded, indicates that our focus on interior solutions is sensible.
17Concretely, for each fully threaded nut, the subject earned one ball. Thus, upon completing the four
games, each subject had earned a certain number (between 0 and 10) of balls in each game. Once the game that would be used to calculate the payo¤s had been picked, a traditional style bingo cage with balls numbered from one to ten was used to draw one ball. Any subject who (in the relevant game) had threaded a number of nuts equal to or exceeding the number on the drawn ball generated additional income, whereas those who had threaded a smaller number did not generate additional income.
Figure 2. Session structures
4
Field Setting
4.1
The state of Campeche
The experiment was conducted in July 2014 in the state of Campeche, in Southeast Mexico. Historically this area was dominated by the Maya. According to the 2010 National Population and Housing Census from the National Institute of Statistics and Geography (INEGI), about 12% of the population older than 5 still speak an indigenous language. With its 57 thou-sand square kilometers, Campeche represents about 3 percent of the total surface of Mexican territory, its 822,441 inhabitants accounted for 0.73% of Mexico’s total population in 2010. According to the 2010 Mexican Population Census about 25% of the population resided in rural localities (with less than 2,500 inhabitants).18
Campeche is a low plateau rising no more than 400 meters above the sea; regions with rolling hills are separated by ‡at land. According to the Series IV of land use and vegetation maps 2010 from the INEGI, agricultural land covers 3.3% of the land, while 25.6% is used for livestock, 65.6% for forestry products, and 5.5% for human settlements. Oil and natural gas extraction activities dominate the economy, by generating 94.7% of the gross value added of the whole state (and 37% of Mexico’s total oil and gas production), while manufacturing generated 0.8%, commerce and service sectors jointly contribute with 4.4%, and agriculture, livestock, and …shing together generated 0.1% (National Economic Census 2014). Although the state of Campeche is thus an important contributor to Mexico’s economy as a whole, 43.6% of the state’s population is in poverty situation and 11% in extreme poverty (CONEVAL 2015).19
Our objective was to get a grip on how transfers a¤ect productive e¤orts in villages with a high proportion of self-subsistence farmers. Campeche is well suited for this, since according
18Censo de Poblacion y Vivienda 2010, INEGI, available at www.inegi.org.mx.
19Extreme poverty refers to persons who cannot a¤ord the cost of a minimum food basket that allows
them to carry out a minimal level of physical activity and who, in addition, show between three to six social de…ciencies.
to the data of the Program for Direct Income Support to Farmers (PROCAMPO), among the 23,885 farmers that received the PROCAMPO support in 2013 in Campeche, 19,797 had less than 5 hectares of land; such farmers are usually subsistence farmers. The land consists mostly of calcareous soils and sedimentary rock, which are poorly suited for agriculture. There are important di¤erences in water availability, with water being abundant in the south and scarce in the north, a fact which is particularly relevant in light of the fact that only 2% of the land used for agriculture is irrigated.
When studying small farmers in Mexico it is important to keep in mind that land own-ership is governed by a special regimen of social land tenure called the ejido. The members of each ejido (the ejidatarios) have collective rights over some agricultural land, which may be cultivated collectively or individually, depending on the the land use decisions taken (by majority rule) in the ejido assembly. Traditionally, ejido plots that were cultivated by indi-viduals could not be sold, rented or put forward as collateral for credit, and no labour outside of the ejidatarios’s family could be hired to work on this land.20 In 1992, the government
implemented a reform to give ejidatarios full property rights over their plots and to loosen some other constraints imposed on ejido land.21 However, according to the National Agrarian
Registry O¢ ce (RAN), in Campeche 52% of the total agricultural land area is still cultivated collectively in 384 di¤erent ejidos and by 60,207 farmers (7.3% of the State population; 19.5% are women and 80.5% are men).
4.2
Locality selection
Our aim was to conduct the experiment on a sample of a population living mostly o¤ of agricul-tural production. To achieve this, we selected rural localities with less than 1,000 inhabitants, with a relatively high proportion of subsistence farmers, and relatively far away from major industry and/or tourist areas. To keep the land ownership institutions constant across the sampled population, we chose to focus on villages organized as ejidos. We excluded seven municipalities because of increasing violence (Candelaria, Carmen, and Escarcega), scarcity of agricultural activity and/or closeness to major developed areas (Calakmul, Campeche, Pal-izada, and Tenabo), and important recent immigration from Central America. We selected sixteen villages from the remaining four municipalities, three of which are located in the
north-20Based on the principle that the land should belong to those who actually work on it, any ejido plot left idle
for two or more years by the ejidatario could be withdrawn by the government and given to another peasant.
21First, ejidatarios had to obtain land certi…cates from the government, which only allowed them to trade
among themselves, and the decision to participate in the certi…cation process had to be taken collectively in the ejido assembly. Then, after an additional legal procedure, they could obtain property titles to trade their plot with those outside the ejido (Sanderson, 1984).
ern part of the state (Calkini, Hecelchakán, and Hopelchén) and one about 150km south from the other ones (Champotón). Instead of drawing a random sample of localities from these mu-nicipalities, we focused on obtaining a set of localities that would share some common traits, but also highlight some of the key di¤erences between the north and the south of the state, e.g., the proportion of people speaking an indigenous language, as described below. Thus, we selected eight villages in the north and eight in the south. Half of each set of villages was assigned to a Partial treatment and half to a Full treatment. Figure 3 shows a map with all the localities.
Figure 3. Map of localities with color coding for session type
As shown in Table 1 and Table 2, the two sets of localities are very similar with respect to most of the basic socio-economic variables (population, gender composition, fertility rates, and education levels), they are all organized in ejidos, and they follow similar patterns for the proportion of the population that are ejidatarios. However, the two sets of villages also di¤er. For instance, there is a di¤erence in the share of the land owned by the ejido that is for common use: while the lowest share is 82% in the northern localities, the share exceeds 32% in only two of the southern localities. Furthermore, they di¤er in some dimensions that may be related to the question at hand, namely, transfers. First, although no external data sources on transfers exist at the locality level, data from the post-experiment questionnaire (see Table 3) shows that 64.5% of the subjects in the southern localities and 45.7% of the subjects in the northern localities claimed to have made some transfer. In contrast, 39.4% of the subjects in the south and 38.6% of the subjects in the north said that they had received some kind of transfer during the year prior to the experimental session. Second, the social backwardness
indicator and the asset index22 show that the northern localities are poorer than the southern,
with a higher variance in northern localities (see Table 2). This may mean that within-village transfers are less common in the northern villages. Third, the north has much higher shares of indigenous population (almost 93% versus 9%), which may matter if there are important cultural di¤erences between indigenous and non-indigenous parts of the Mexican population when it comes to sharing.23 Fourth, there is signi…cantly smaller religious diversity in the
north than in the south (82% versus 50% of the subjects declare themselves Catholics, and 13% versus 32% declare themselves as followers of another religion). This suggests that there may be greater community homogeneity and/or cohesion in the northern than in the southern localities, something which may a¤ect the willingness of people to share with others in the community. Finally, southern localities have, on average, higher travel times to the next big town, a feature which may imply that people need to rely more on the community in the south than in the north, a feature that would tend to make transfers more common in the south than in the north. In Section 5.5 we will analyze whether these locality characteristics can explain part of the behaviours that we observe in the experiment.
Table 4 shows how our sample population compares to the population in the 2010 Mexican Census. Compared to the census data, in our sample there is over-representation of women (especially young women), of individuals with smaller dwellings, and of people born in the state. The …rst two are not very surprising: although we tried hard to achieve gender balance, it was often harder to recruit men due to their work; moreover, poorer people may have a smaller opportunity cost of time. It is also important to keep in mind that the Census was collected in 2010 and our survey was carried out in 2014, so some di¤erences are to be expected. In particular, it is conceivable that due to high fertility rates, since 2010 many young people have built and moved into new, small dwellings, and that this explains some of the di¤erences between the Census data and our sample beyond any selection bias that may have arisen due to availability and opportunity cost of time.
22As measures of wealth, we construct indices of dwelling characteristics and household assets by taking
the …rst component in a principal components analysis (Filmer and Pritchett, 2001). The variables used to construct the dwelling index are dummy variables for whether the participant or her family own their home, whether they have running water, a toilet, electricity and dirt ‡oor, the number of rooms and of light bulbs. For the asset index, we use dummy variables for whether the household owns other real estate properties or land, vehicles, TV radio, cellphone, gas stove, small kitchen appliances, refrigerator, washing machine, books; productive assets like agricultural machinery or animals; and small livestock.
23Mexico still has a large number of distinct indigenous groups, and in light of the results reported by
Henrich et al (2005) (see also Jakiela, 2015) it seems interesting to investigate whether there are signi…cant di¤erences between these groups, and also between these groups and individual who do not view themselves as belonging to any indigenous group.
4.3
Experimental Sessions
Participants were invited by public announcement by the village authorities. From the vol-unteers, we selected 40 participants in each village, aiming to obtain an equal number of men and women, and no more than one representative per household.24 Recall that the localities visited were relatively small (see Table 1), and hence the likelihood that the subjects knew each other was quite high. From data collected in the post-experimental questionnaire, we see that participants indeed tended to know almost all the other participants. However, only a very small share were …rst-degree relatives (see Table 5).
The sessions were held in a public meeting space in the village (usually the primary school, although once it took place in a high school and once in the meeting space of the ejido), usually in the afternoon when the agricultural labour day had ended. In each village two rooms where used: one room was used as a …eld lab, while the other room was used for the group that was not actively playing (subjects who were passive …rst and then active had a snack in this room while waiting for their turn to be active, while the others had a snack and …lled in the post-experimental questionnaire in this room). It was ensured that the …eld lab had twenty tables and chairs, arranged in four columns facing the experiment director (the same in all the sessions), and opaque curtains were installed between the columns of participants in order to provide privacy and to ensure that the subjects could properly see and hear the experiment director. Given the age variance of the participants and the low alphabetization levels, no written instructions were handed out and all materials were made language free.25 In the villages where any subject did not understand Spanish the instructions were translated by a native Mayan speaker. A complete session, including the post-experimental questionnaire, would typically last for about three hours. The average payment was 90 pesos, close to the minimum wage of an agricultural operator.26
Upon arrival, the participants were requested to pick a card from an opaque bag. The card had a number and a symbol on it. The number was used by us to track the subject. The participants that had a half-circle on their card (and an even number) were asked to proceed to the lab, while the others (with a full circle and an odd number on their card) were asked to proceed to the other room. It was ensured that both groups had the same
24We cannot calculate a show-up rate since the announcements were made either through loudspeaker
announcements or by word of mouth in the village. In all the sessions volunteers were assembled and the …rst 40 people to show up were selected, unless latecomers were prioritized to ensure gender balance.
25The experimental protocol, that was communicated in full orally by the experiment director, as well as
the visual materials that were handed out to the subjects can be found at http://ingelaalger.weebly.com.
26In 2014, no legal minimum wage for agricultural workers existed in Mexico. For reference, in Campeche in
2014 the general minimum wage was 63.7 pesos per day and the minimum wage for an operator of agricultural machinery was 93.6 pesos per day (http://www.conasami.gob.mx).
number of participants, so that one-to-one matching could be done in the one-to-one games (the matching was anonymous and random; we simply matched subject number 1 to subject number 2, etc). Upon entering the experimental room, the participants would look for the table with their number. They would …nd on the table two bowls with the nuts and the bolts, and a newspaper. They were told that the newspaper could only be used when instructed, which was in the time dedicated to the task. The refurbishment of the bowls was done every two games, and all bowls, even those that were essentially full, were refurbished to ensure that participants could not infer any information about the others’e¤orts from the movements of the assistants, and to ensure that the movements of the assistants were the same for all the participants.
Each session consisted of an ability test, followed by the four games, a second ability test, and one post-experimental question related to the One-to-One game (see Figure 2). The question was asked at the end to avoid contamination to the games. At the beginning of the session the participants were told that they would play four colour-coded games, and that one of the games would be randomly picked for payment. This was done to avoid wealth e¤ects during the session. The order of the games was altered between sessions, as presented in the Appendix, but the participants always played the autarky games …rst. The purpose of this was to start the session with the easiest game and to ensure understanding by participants, who might be foreign to abstract thinking.
The experimental part would last between 30 and 45 minutes, depending on the need for translation. When done, groups would quietly change rooms without interacting. When done with the lab part, the participants were given the post-experimental questionnaire (those who could not not read got help from a native Mexican assistant). After both groups had performed the experimental part, a child would pick a colored card (from an opaque bag) that determined which of the four games would be used to calculate the payo¤s. Finally, the lottery (see Footnote 17) was played in front of all participants to ensure transparency of the payments calculation.
5
Results
Our design allows to ask the following three questions, that we address in turn in this section: 1. Do transfers a¤ect the e¤ort choices of the donors and recipients of these transfers? 2. If there is an e¤ect, does it depend on the origin/destination of these transfers?
3. If there is an e¤ect, does it depend on the size of the transfer, or on whether the subject is a donor or a recipient of a transfer?
5.1
Do transfers a¤ect e¤ort?
To estimate whether the expectation of receiving or giving a transfer a¤ects e¤ort, we per-form a within-subjects analysis, in which we compare the e¤ort of a given participant in two games— an autarky game and a transfer game. Speci…cally, we estimate OLS regressions of the following form separately for each pair of games that we use to make comparisons:
eig = 1+ 1Tig+Li 1 +Xi 1+Si 1+uig (8) whereeig is the number of threaded nuts of participantiin gameg,Tig is an indicator variable equal to 1 if the e¤ort corresponds to a transfer game and zero if it corresponds to an autarky game, Li is a group of variables aimed at controlling for learning during the experimental session, Xi is a vector of characteristics of the participant and her household (see below), Si includes locality …xed e¤ects, anduig is an error term. The coe¢ cient of interest is 1, which
measures the e¤ect of a transfer from/to another anonymous participant on e¤ort, compared to the autarky situation.
Accounting for learning with the group of variables included in Li is important because even though we randomized the order in which games were played, the autarky games were always played before the transfer games. To control for learning we use the ability tests performed at the beginning and the end of the experimental session, and the variation in the order of games. A detailed description of how we use these variables, and some evidence on learning, can be found in the Appendix.
We include as participant controls their age, number of siblings, number of children; dum-mies for married, for whether the participant was born in the municipality where she currently lives, speaks an indigenous language, for whether she speaks both an indigenous language and Spanish, for whether she has at least secondary education (i.e. 6 or more years of schooling), and a dummy for whether the main income source of her household comes from agriculture. We include the indexes of dwelling characteristics and household assets, constructed from our participant survey and described in the previous section. We also include the number of organization types and community festivities in which the subject participates, a measure of how actively she participates in those festivities (0 if she does not attend, 1 if she attends, 2 is she contributes with money, time or goods), a measure of how cooperative the community is as perceived by the participant, and a measure of her trust and attitude towards e¤ort.27 27We measure the participant’s perception of how cooperative is her community using her response to the
question “If a problem arises in your locality, do people cooperate to solve it?”, which is coded from 0 to 3 (never=0, always=3). The trust variable is a dummy equal to 1 if the participant agrees that “one can trust the majority of people” and zero if instead she agrees that “one ought to be more careful when trusting people”. Finally, her attitude towards e¤ort comes from the question: “In your opinion, is work e¤ort rewarded with higher income?”, coded from 0 to 3 (never=0, always=3).
Locality …xed e¤ects are included. In our reported estimations, we cluster the standard errors at the locality level to account for any correlation in errors within villages, and use a t distri-bution to test whether the main coe¢ cients are statistically signi…cant, as it is recommended when the number of clusters is relatively low (Cameron and Miller, 2010).28
Starting with the Partial treatments, we report a …rst set of results in Table 6. Here we compare e¤ort in each of the transfer games to that in Autarky Low, games that are all equivalent in terms of payo¤ to the subject. In all cases (DP One-to-One, RP One-to-One,
DP Pool, and RP Pool) the existence of a transfer induces subjects to make a higher e¤ort
than under autarky, and the di¤erence is statistically signi…cant at 5% in all cases except for
DP Pool (p= 0:2). These results already suggest that transfers have, on average, a positive
e¤ect on the incentive to exert e¤ort. Note, however, that we cannot exclude that this is driven by a desire to maximize the experimental payments for the whole set of participants in the session, since in Autarky Low the additional income generated (i.e., given by us) in case of success is 33% lower than in the transfer games (z = 50 vs z = 75).
Hence, in a second step, we compare, still for the Partial treatments, e¤ort in each of the transfer games to that in Autarky High. These games are all equivalent in terms of the additional income that is generated in case of success (z = 75), and the e¤ect of a transfer is now to reduce the subject’s own material payo¤ gain from exerting e¤ort by 33% (for both donors and recipients this gain drops from 75to 50). This comparison is closer to a real-life situation than the previous one, since in real life a transfer does reduce the material gain from exerting e¤ort. The results, reported in Part A of Table 7, show that, on average, subjects still make a higher e¤ort in the presence of a transfer compared to autarky, an increase which is statistically signi…cant at conventional levels in all cases except RP Pool (p= 0:13). Since here e¤ort does not increase the total experimental payments to the community, we can safely conclude that it is the transfer that has a positive e¤ect on the incentive to exert e¤ort.
Finally, we turn to the Full treatments, where all the games give the same additional income in case of success, and the e¤ect of a transfer is to reduce the subject’s material payo¤ gain from exerting e¤ort by 100% (for both donors and recipients this gain drops from75 to
0). The results, summarized in Part B of Table 7, show that even in this case the transfer has a positive e¤ect on e¤ort, an e¤ect which is signi…cant at 5% in all cases except forDF Pool. In sum, on average subjects increase their e¤ort in the presence of a transfer, and in most cases the e¤ect is statistically signi…cant. Furthermore, this e¤ect is robust in the following ways. Firstly, the increase in e¤ort occurs both for those who are on the giving end and for
28We also tried clustering the standard errors at the individual level, to account for any serial correlation
in e¤ort between games. These results, which are similar to our reported ones, are not shown, but they are available upon request.
those who are on the receiving end of a transfer. Secondly, it occurs whether the transfer partially or fully removes the material gain to self from making e¤ort. Thirdly, the increase in e¤ort occurs whether the transfer a¤ects one other individual or a pool of other individuals.
Together with Inferences 1 to 4 in the theory section, these …ndings lead us to our main conclusions. Firstly, on average our subjects derive positive utility from making a transfer to other(s) when in a donor treatment and from avoiding imposing a transfer on other(s) when in a recipient treatment. Secondly, this utility e¤ect is large enough to outweigh the utility loss associated with a decrease in own material payo¤, since the net e¤ect on e¤ort is positive. Thirdly, the e¤ect is economically signi…cant:29 when comparing e¤ort in One-to-One games
to that inAutarky High games, on average e¤ort is 9.4% higher in theDonor Full treatments, and 14.4% higher in the Donor Partial treatments; for recipients, e¤ort increases by 7% in theFull treatments, and by 9.4% in thePartial treatments.
Our data does not allow us to estimate the motivation for why a subject would increase her e¤ort in the presence of a transfer. However, it is easy to check that the results at hand
(i) cannot be explained by purely altruistic preferences, (ii) is consistent with aversion to advantageous inequity (since theFull treatment allows to reduce such inequity more than the
Partial treatment does), and (iii) is consistent with subjects deriving utility from the mere
act of giving rather than only from the e¤ect of the transfer on its recipient.
5.2
Does the source/destination of the transfer matter?
In this section, we study whether the source/destination of transfers matters for e¤ort choices, we again conduct within-subjects analysis using the regression described in Equation 8, but now we compare e¤ort in di¤erent transfer games. Indeed, since the consequences of e¤ort on own material payo¤ as well as the size of the transfer are identical across all the transfer games for any given subject, these comparisons allow us to detect whether the source/destination of the transfer has an e¤ect on subjects’ e¤orts. In these regressions we again control for individual characteristics and locality …xed e¤ects, and for learning (a di¤erence, however, is that for these game pairs, the game order was varied across subjects; see Appendix).
The results, reported in Table 8, show that, on average, the di¤erence between e¤ort in the Pool game and that in the One-to-One game is not statistically di¤erent from zero at conventional signi…cance levels, both in thePartial and in theFull treatments. These results indicate that subjects do not attach a higher or a lower utility to a transfer that a¤ects one individual in the other room compared to a transfer that a¤ects all the individuals in the other room.
In Table 9 we exploit the transfer game that is unique to the Full treatments, namely, the
Public Good game, to see whether subjects make more or less e¤ort when the transfer a¤ects
the local health center as compared to when it a¤ects some passive participant(s). In this case there is a signi…cant di¤erence: subjects exert a higher e¤ort when the transfer a¤ects the local health center, an increase which is statistically signi…cant at 5% in all cases except
for DF One-to-One.30
5.3
Does the size or the direction of the transfer matter?
The within-subject comparisons reported above show that in general a transfer has a positive e¤ect on e¤ort. Here we rely on within- and between-subjects comparisons to ask whether the e¤ect is di¤erent depending on the size and the direction of the transfer.31
First, we study whether the size of the transfer matters by exploring the fact that all the subjects allocated to a Donor treatment played the exact same Autarky High game to compare e¤ort choices of subjects allocated to theDonor-Partial treatment to those allocated to the Donor-Full treatment. More precisely, we analyze whether the di¤erence between the e¤ort in a transfer game and that in Autarky High di¤ers between the Partial and the Full
treatments. The estimating equation in this case is given by:
eig = 2+ 21Tig + 22Fi+ 23Ti Fi+Li 2+Xi 2+Si 2+vig (9) where Fi is a dummy equal to 1 if participant i was assigned to the Full treatment and zero if she was assigned to the Partial treatment, and the other variables are de…ned as before. In this equation, the coe¢ cient of interest is 23, which measures the e¤ect of the transfer (relative to autarky) in theFull treatment after di¤erencing out the same e¤ect for thePartial
treatment. Put di¤erently, the estimate of 23 gives the double di¤erence, i.e., the average di¤erence in e¤ort for the same pair of games between theFull and thePartial treatments.
We run the regressions both forOne-to-one and forPool. The results are reported in Table 10. The transfer game dummy alone is positive and signi…cant at 1%, which could imply that e¤ort increases with transfers. Furthermore, the interaction of this variable with the Full
treatment dummy is negative, suggesting that such positive e¤ect is smaller for donors in the
Full than in the Partial treatment (which was expected from the e¤ort increases in the Full 30The fact that there is a signi…cant di¤erence indicates that subjects do respond di¤erently in di¤erent
transfer games, and it lends further support to our claim (see Tables 7A and 7B and the accompanying comments) that the e¤ort increases associated with transfers are not simply driven by a desire to maximize the sum of the experimental payments to the subject pool.
31As for the within-subject estimations, in all the between-subjects estimations reported in this subsection
andPartial treatments reported in Table 7), but the interaction is not statistically signi…cant. This is remarkable, since the transfer in the Full treatment is three times as large as that in
thePartial one.
Second, we analyze whether the direction of the transfer matters by examining the fact that all the subjects allocated to a Partial treatment played the exact same Autarky High
game to compare e¤ort choices of subjects allocated to theDonor-Partial treatment to those allocated to theRecipient-Partial treatment. In a more precise approach, we analyze whether the di¤erence between the e¤ort in a transfer game and that inAutarky High di¤ers between
theDonor and the Recipient treatments, by estimating the following equation:
eig = 3+ 31Tig + 32Ri+ 33Ti Ri+Li 3+Xi 3 +Si 3 +ui: (10) Here,Ri is a dummy equal to 1 if participant i was assigned to the Recipient treatment and zero if she was assigned to the Donor treatment. The other variables are de…ned as before and the coe¢ cient of interest in this case is 33, which captures the average di¤erence between
recipients and donors of the e¤ect of a transfer compared to autarky.
We again run the regressions both for One-to-one and for Pool. The results, reported in Table 11, show that the transfer game dummy alone is positive in all cases and signi…cant in three cases, implying that e¤ort increases with transfers. The interaction of this variable with the Donor treatment dummy is negative in one case and positive in three cases, but it is never statistically signi…cant. Thus, for our subjects the willingness to provide e¤ort was no di¤erent in the case where e¤ort made it more likely that the subject helped someone else, than in the case where it made it more likely to avoid being helped by someone else.
5.4
Robustness
In sum, the experimental data shows that transfers have a positive e¤ect on e¤ort in the four treatments, and that in most cases the e¤ect is statistically signi…cant and economically non-negligible. In this subsection we use the post-experimental questionnaire and locality-level data from the 2010 National Population and Housing Census to perform some robustness checks.
5.4.1 Are some e¤ects driven by the fact that transfer size is …xed?
By contrast to many experiments (e.g., Ligon and Schechter, 2012 and Jakiela, 2015) where subjects are asked to choose the size of a transfer (and typically the size can be zero), in our e¤ort-and-transfer experiment the transfers are forced and of …xed size: the participant’s choice is on whether to perform e¤ort, but in case of success of this e¤ort (or failure for the
case of the recipients), transfers of a given size are automatic. To control for the fact that these transfers may be smaller or larger than the individuals would have chosen themselves, we use data collected from the subjects in theDonor treatments about this. More precisely, at the very end of each experimental session (see Figure 2) in aDonor treatment, we handed out a sheet of paper and a pencil to each subject. The sheet reminded the participants about the transfer size in the One-to-One game and asked a question; to ensure that all subjects understood, we also showed them the corresponding visual material and read out the text on the sheet aloud. The question was di¤erent in the Partial and the Full treatments.32 To
subjects in the Donor-Partial treatments we asked: “Would you have liked to give more?”, and “If so, how much more?”. It turns out that 40.6% of the subjects would have liked to give more. To subjects in theDonor-Full treatments we asked: “Would you have liked to give less?”, and “If so, how much less?”. We found that 52.7% of the subjects would have liked to give less.
It is quite remarkable that such a high proportion of donors would have liked to give more than 33% of their extra earnings in the Partial treatment, and that only 53% would have liked to give less than 100% of their extra earnings in the Full treatment. However, we note that this is consistent with the conclusion that we drew from the e¤ort choices, namely, that on average subjects derive positive utility from making a transfer.
Next, we study whether the e¤ects of transfers on e¤ort may be due to the fact that transfer size was imposed in the games by checking: (1) for subjects, in theDonor-Partial treatments, whether donors who would like to give more respond di¤erently to a transfer compared to autarky than the others; (2) for subjects, in theDonor-Full treatments, whether donors who would like to give less respond di¤erently to a transfer compared to autarky than the others. Part A of Table 12 shows that it is not the case: the interaction between the answer to the question and the treatment variable is insigni…cant in both cases.
For donors, we further worried about the potential impact of the fact that those who were on the receiving end of the donors’transfers were completely passive. To check to what extent this issue mattered we asked donors whether their answer to the questions mentioned above would change if the recipient had been able to do an e¤ort. The answer was positive for 42.5% of the donors in the Full treatment and 25.8% of the donors in the Partial treatment. This suggests that there is some willingness to reward e¤ort, which in real situations could play a role in e¤ort decisions. In our experiment, however, the way subjects answered the question at hand does not have a signi…cant e¤ect on the di¤erential e¤ort between autarky and transfer games (see Part B of Table 12).
32In both cases the subjects answered this question before knowing which game would be used to calculate
5.4.2 Does the fact that e¤ort is unobservable matter?
In the experiment a subject’s e¤orts are his or her private information. This allows us to focus on pure preference e¤ects and to avoid signalling e¤ects. In reality, however, e¤ort is often observable, and hence our results may under-estimate the e¤ects that transfers have on e¤orts compared to real-life situations. Although we did not expose subjects to di¤erential observational treatments, we asked a question related to e¤ort observability to all subjects who were assigned to Recipient treatments. More precisely, at the very end of each experimental session (see Figure 2) in aRecipient treatment, we handed out a sheet of paper and a pencil to each subject. On this sheet we reminded the participant about theOne-to-One game, and about the fact that his (her) e¤ort would not be revealed to the passive subject with whom (s)he was matched, and then we asked: “If we gave you the opportunity to reveal your e¤ort in the One-to-One game to the individual to whom you were matched, would you do so?”. The answer turned out to be positive for 39.3 % of the subjects in theFull treatments and for 29.07% of the subjects in thePartial treatments (numbers which are signi…cantly di¤erent at the 0.07 level). When interacted with the treatment variable of interest, however, the answer to the question has no signi…cant impact on the e¤ect of the transfer in theOne-to-One game compared to autarky (see Part C of Table 12).
Summing up, the data collected through questions lend no support to the hypotheses that the e¤ects of transfers on e¤orts are driven by the fact that transfers are forced and of predetermined size, or the fact that e¤ort is unobservable.
5.5
Do individual and locality characteristics matter?
In this section, we rely on the subjects’ responses on the post-experiment questionnaire as well as the census data to study whether individual and/or locality characteristics matter for the e¤ects of transfers on e¤orts.
Starting with locality characteristics, recall from Section 4 (see Table 2) that the selected localities di¤er on certain dimensions that may matter when it comes to transfer patterns, in particular, poverty (as measured by the social backwardness and asset indices), the share of indigenous population, the degree of religious homogeneity, and the distance to a larger town. Table 13 reports results of regressions, each of which includes as control one of these variables. Estimates suggest that neither the social backwardness index nor the share of indigenous population a¤ects the di¤erence in e¤ort between the autarky and the transfer game at hand. By contrast, the indicators of religious homogeneity do have an impact in all
theRecipient treatments, whether religious homogeneity is measured as the share of Catholics