The World Bank Economic Review, 36(4), 2022, 857–888 https://doi.org10.1093/wber/lhac016 Article Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Keep It Simple: A Field Experiment on Information Sharing among Strangers* Cátia Batista , Marcel Fafchamps, and Pedro C. Vicente Abstract SMS information campaigns are increasingly used for policy. A field experiment is conducted to study informa- tion sharing through mobile phone messages. Subjects are rural households in Mozambique who have access to mobile money. In the baseline intervention, subjects receive an SMS containing simple instructions on how to redeem a voucher for mobile money. They can share this non-rival information with other exogenously as- signed subjects unknown to them. Few participants redeem the voucher. They nonetheless share it with others and many share information about the voucher they do not use themselves. The voucher is shared more when no information is provided on the receiver. When partial information is provided, no evidence is found of more sharing with subjects who have similar characteristics. Treatments are introduced to increase the cost of send- ing a message, shame those who do not send the voucher to others, or allow subjects to appropriate the value of the voucher. All these treatments decrease information sharing. To encourage information diffusion among strangers, the best is to “keep it simple.” JEL classification: D83, D64, O33 Keywords: SMS information campaign, nudging; mobile money, anonymity, shared identity 1. Introduction Many policy interventions rely on messages to convey information to a target audience so as to induce behavioral changes—for example, letters (e.g., Hjort et al. 2021), SMS (e.g., Abebe et al. 2018; J-PAL 2020; Banerjee et al. 2021b; Afzal et al. 2022, mobile phones (e.g., Cole and Fernando 2021; Kelley et al. 2022), and social media (e.g., Alatas et al. 2022). The use of messages has further increased since the onset of the COVID-19 pandemic (e.g., Alsan et al. 2020). In many interventions, the policy maker has individual identifiers (e.g., phone numbers) for only a fraction of those they wish to target. Consequently, reaching other interested individuals through information diffusion is often essential for the policy to Cátia Batista is Associate Professor of Economics at the Nova School of Business and Economics, Campus de Carcavelos, Rua da Holanda, N 1, Carcavelos, Lisboa, Portugal. Her email address is catia.batista@novasbe.pt. Marcel Fafchamps is Senior Fellow in the Freeman Spogli Institute, Encina Hall, 616 Serra St., Stanford, CA 94305, USA. His email address is fafchamp@stanford.edu. Pedro C. Vicente is Professor of Economics at the Nova School of Business and Economics, Campus de Carcavelos, Rua da Holanda, N 1, Carcavelos, Lisboa, Portugal. His email address is pedro.vicente@novasbe.pt. We benefited from comments from Yves Zenou, Sanjeev Goyal, Francis Bloch, Markus Goldstein, participants to the Monash University Conference on Social Networks 2018, and from the editor and two anonymous referees. We wish to thank Stefan Leeffers, Timóteo Simone, and the NOVAFRICA office in Mozambique for excellent research assistance. We are particularly grateful to Carteira Móvel/MKesh for institutional support. We wish to acknowledge financial support from the International Growth Centre. All errors are our responsibility. © The Author(s) 2022. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 858 Batista, Fafchamps, and Vicente succeed. Yet we know little about how to motivate people to share information with others, especially people they do not know. This paper seeks to address this knowledge gap. The sharing of valuable information is at the heart of many important economic processes such as the diffusion of new technology (e.g., Ryan and Gross 1943; Griliches 1957; Foster and Rosenzweig Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 1995; Bandiera and Rasul 2006; Vilela 2019; Beaman et al. 2021; Carter et al. 2021; Cole and Fernando 2021), the adoption of new consumer products (e.g., Fafchamps et al. 2022), credit reference services (e.g., Kandori 1992; Greif 1993), information about market opportunities (e.g., Granovetter 1974; Fafchamps and Minten 2012; Kelley et al. 2022), and the referral of workers and trainees (e.g., Beaman and Magruder 2012; Fafchamps et al. 2020a). Information sharing is also essential to social learning, i.e., the process by which crowds form inference by aggregating dispersed information (e.g., Golub and Jackson 2010; Chandrasekhar et al. 2020). Two key maintained assumptions underlie much of this work. First, it is often implicitly assumed that people are willing to share information when doing so brings no immediate or delayed benefit. In practice, however, even when the information itself is non-rival, sharing it typically imposes a cost on the sender. Second, the recipient is assumed to trust the information provided. This assumption is made even though, in many cases, the quality of the information cannot be verified, or can only be verified at a cost. If these two assumptions are violated, some valuable information may not be shared, and some shared information may not be believed (e.g., Allcott and Gentzkow 2017). Epidemiological models of diffusion on networks (e.g., see the excellent reviews by Vega-Redondo (2007) and Jackson (2010)) have demonstrated that small changes in the probability that a message is successfully transferred between two nodes can have dramatic effects on the spread of information.1 Given this, it is somewhat surprising that little empirical research has sought to ascertain the extent to which individuals successfully share valuable information with each other. We know little about whether recipients actually read or believe the messages they receive and under which conditions they forward these messages to others. This lack of knowledge is particularly acute for information shared among strangers on social media—or, in lower income countries, on mobile phone platforms such as those introduced to share information among farmers (e.g., Cole and Fernando 2021) or between employers and jobseekers (e.g., Kelley et al. 2022).2 The purpose of this paper is to investigate these research questions formally using an original field experiment implemented through text messages on mobile phones in Africa. All the social diffusion pro- cesses mentioned at the onset of this paper share a common difficulty: the value of non-rival information varies across recipients in ways that are difficult if not impossible for senders to predict. Not only does this uncertainty disincentivize the sharing of non-rival information, it also creates variation in the willingness to share that depends on (unobserved) expectations about benefits to others. To sidestep this difficulty and maximize the power of our experiment, we standardize the value of information across all subjects: it is about a voucher for free money, the value of which is fixed, revealed to all senders, and verifiable by them. In our baseline intervention, selected subjects receive an SMS voucher that they can redeem for mobile money. Having received the SMS, subjects can offer the same voucher opportunity to up to four other 1 For instance, in Poisson random networks with n nodes, a giant component emerges when the link probability p rises above 1 /n and it grows in size until p reaches log(n )/n, at which point the network becomes fully connected. This means that if p represents the probability with which information is successfully transferred between two arbitrary nodes in a large network, when p < 1/n only a vanishingly small proportion of nodes will be informed, while if p > log(n )/n, all nodes will be informed. It follows that small frictions in information sharing can have large consequences on information spread and thus on efficiency. 2 For agricultural extension, other examples include the work of BRAC and of NGO Self Help Africa. For jobs, other examples include JobTalash in Pakistan, which is currently being studied by Erica Field, Rob Garlick, Nivedhitha Sub- ramanian, and Kate Vyborny. The World Bank Economic Review 859 subjects who, in turn, can redeem it for cash and get the same voucher to others. This information transfer process goes on for several rounds. We focus our attention on whether people redeem the voucher and whether they pass it on to others. This experimental design mimics the process by which people share information by passing on or reposting messages they have received on social media or on information Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 platforms shared through mobile phones. We use redeeming behavior to measure the extent to which messages are read, understood, and believed. Sending behavior is used to measure the willingness to share valuable but non-rival information with strangers. We study a sample consisting of heads of households and their spouses in rural areas of Mozambique. Subjects are only allowed to give the voucher to strangers selected by us from a different village in the sample. The purpose of the stranger matching is to avoid in-person communication and behavior coordi- nation between subjects. In addition, all communication between subjects is done via text messages that go through the experimenter’s switchboard and the identity (or phone number) of linked individuals is never revealed. Stranger matching also eliminates the correlation in preferences and behavior that characterize self-selected social networks as a result of homophily. We see these features as a strength of our study because they greatly facilitate causal inference in the study of information diffusion.3 While information sharing among strangers is probably a lower bound on information sharing among socially connected individuals, it is nonetheless empirically relevant—first because it occurs frequently in practice (e.g., on social media and in casual conversation with strangers), and second because information from a distant stranger is less likely to repeat information already present in one’s social circle, and is thus often more valuable (e.g., Jackson and Wolinsky 1996; Jackson 2010). Since the vouchers can only be redeemed for mobile money, familiarity with mobile money is essential. For this reason, we recruit all the participants from a pool of individuals who were previously introduced to mobile money services, have used the services, and have an active mobile money account on their mobile phone. We find nevertheless that a surprisingly small proportion of recipients redeem the free- money voucher: 26 percent in the baseline intervention, and even fewer in most other treatments. This is an unexpected result given that redeeming the voucher is a low-cost, high-return action. This suggests that many subjects either ignore the messages they receive, or do not trust them. At the same time, we find that subjects often share the voucher message with others, even when they do not redeem it themselves. In other words, some people are willing to incur a cost to share information with others, implying that they understand our messages well enough to do this. Yet, by not redeeming the voucher, they implicitly reveal that they do not believe it. This type of behavior is consistent with a warm glow motivation (e.g., Andreoni 1990), rather than with pure altruism.4 Information sharing remains limited, however, and many participants never get the opportunity to receive the free money. We then introduce a number of treatments in an effort to increase the diffusion of the valuable non- rival information—i.e., making the voucher available to someone else. These treatments are divided into two batches of three and each group of baseline subjects is assigned to either of these two batches, with equal probability. The first treatment of batch one is to give some information to subjects about the sender or recipient of the SMS. While this information is not sufficient for subjects to identify the other party, it nonetheless should facilitate information sharing if subjects identify more easily with similar people and, as a result, behave in a more altruistic way towards them. Contrary to expectations, disclosing key characteristics of the sender or recipient reduces information sharing: both redeeming and sending vouchers fall. These 3 The direction of causality in the diffusion of information is always difficult to ascertain on social networks: linked individuals often share similar interests and, as a result, may simultaneously get new information from a third source instead of each other; they can also search for information by asking their contacts, possibly triggering the diffusion of the information itself. Our design abstracts from these difficulties. 4 By definition, an altruist cares about the utility of others, not just about the action of giving. An altruist who believes that paying to redeem the voucher is not beneficial would presumably not want to share it with others. 860 Batista, Fafchamps, and Vicente patterns indicate that subjects behave in a more altruistic and trusting manner when uninformed about the specific characteristics of the sender or recipient. This could arise because revealing differences may induce some subjects to reduce their trust. This is not what the data show, however: subjects are not more likely to redeem a voucher received from someone with similar characteristics—or to send it to someone similar. Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 A more likely explanation is that processing the information contained in the SMS becomes cognitively more demanding when characteristics of the other party are added, leading more subjects to dismiss the SMS. Next, we vary the costs of sending vouchers. We find that information sharing falls when the mone- tary cost of sending the SMS increases. This is in accordance with standard theory. We also experiment with a non-monetary cost, namely, (anonymously) shaming subjects who do not send the voucher to others. If subjects are concerned about their self-image, this treatment could increase sharing because it introduces a cue that not sharing violates a social norm. We do not, however, observe any significant effect of the shaming treatment. In the third treatment of batch one, we introduce the ability for sub- jects to circulate misleading information more cheaply than useful information. If subjects value their social image—i.e., they want to pretend to be doing the “right thing”—but they do not care about oth- ers, they may be tempted to pay less and send information, an action that makes them look good since the recipient does not know a priori that the information is not useful. We see very little take-up in this case, indicating that most subjects do not purposefully set out to deceive others by sending misleading information. In the second batch of treatments, we introduce the possibility for subjects to appropriate part or all of the value of the vouchers destined for others. The motivation behind these treatments is that subjects may be more willing to share valuable information—i.e., the voucher—if they get a monetary compensation for doing so. To this effect, we design three treatments along the lines of the dictator, ultimatum, and re- verse dictator games, and adapt them to our design. These are chosen because they resemble mechanisms that have been used in sharing a non-rival good or service. The reverse dictator mimics situations where the provider lets the user “pay what they like”—an approach practiced online (e.g., shareware) and offline (e.g., alms giving when visiting a church). The ultimatum mimics situations where the provider sets a price for the service that is presumably below its value to the user, but the user can refuse. This resembles a simple market transaction: the information is non-rival and the sender bears little or no cost for sharing it, but nonetheless extracts a payment because the user is willing to pay for it. The dictator game corre- sponds to situations where the provider appropriates part of the value of the information to the user, but does not reveal the value of what has been appropriated. Here the comparison is slightly more tenuous, but this treatment bears some resemblance to the business model of Facebook, Google, and others, which is to appropriate non-rival information they collect on users and sell it to third parties. Surprisingly, we find evidence that, if anything, allowing senders to extract or solicit payment reduces information cir- culation. When the vouchers are used, however, the treatments have a large effect on how their value is split between sender and receiver. When we combine these findings to those from the treatments dis- cussed above, the same pattern emerges: when SMS messages become more complex and cognitively challenging—as they do in all our treatments relative to the baseline intervention—they get more readily dismissed. This paper contributes to the literature in several ways. First it complements a theoretical literature on diffusion that takes information transfer in human populations as a given (e.g., Bloch et al. 2008; Jackson et al. 2012). Our results cast some doubt on the implementability of strategic mechanisms that rely on the near perfect sharing of non-rival information. Second, our work generalizes earlier findings by Mobius et al. (2015) who examine how relative strangers share and aggregate information that helps them win movie tickets. Like us, they find that the sharing of information is highly imperfect: signals travel only up to two links. It is, however, unclear how general their findings are, due to the strategic complexity of their The World Bank Economic Review 861 design and the fact that information is partially rival. Our results confirm that information sharing is far from perfect even in the absence of such considerations. Third, our results echo those of Drexler et al. (2014), who find that teaching financial literacy using a simpler curriculum based on rule-of-thumb heuristics works better than relying on standard accounting Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 training. They similarly interpret their results as suggesting that “keeping it simple” makes external inter- ventions more successful at reaching a less literate target population. Similar sentiments can be found in Naeher and Schündeln (2021) and in McKenzie (2021) regarding entrepreneurship training, and they tally well with the many studies showing that cognitive load can impede optimal decision making in humans (e.g., Deck and Jahedi 2015; Drichoutis and Nayga 2020). Our study extends these insights to types of interventions that, a priori, could be conceived as unsophisticated and therefore unlikely to tax human cognitive abilities—namely, short SMS messages delivered on individual phones. Our findings have far-reaching policy implications. Mobile telephony has revolutionized the way many interventions are conducted. This is particularly true in parts of the developing world—such as sub- Saharan Africa—where the penetration of mobile phones has massively increased in recent decades. A growing number of policy interventions employ mobile phone messages to pursue a development objec- tive. Some of these messages nudge recipients into taking a particular action, e.g., reminders regarding savings (Karlan et al. 2016a; Blumenstock et al. 2016; Abebe et al. 2018), debt repayment (Karlan et al. 2016b; Afzal et al. 2018), or preventive health (Obermayer et al. 2004; Patrick et al. 2009; Raifman et al. 2014). Other interventions have taken the form of information and awareness campaigns. Recent exam- ples include information about agricultural prices (Fafchamps and Minten 2012), water quality (Okyere et al. 2019), and the electoral process (Aker et al. 2017).5 Such interventions have the potential of reaching beyond the recipient of the original message. Indeed many policy interventions have long sought to increase their impact by relying on social diffusion. A number of recent studies have tested whether such interventions diffuse through existing social links (e.g., Banerjee et al. 2013, 2019; Fafchamps and Vicente 2013; Fafchamps et al. 2020b; Comola and Prina 2021). But little work exists on information sharing in the more anonymous settings now permitted by social media and information platforms based on mobile phones such as Whatsapp and similar. Infor- mation Technology (IT) can potentially make diffusion much easier because messages (e.g., SMS, emails, tweets, Facebook, or Whatsapp posts) can easily be reposted or forwarded to others. Its potential could be further strengthened by using mobile money to incentivize diffusion. Firms sometimes reward customers for introducing them to new clients, for instance. Similar approaches have been discussed in public policy circles, e.g., whether HIV-positive individuals can be incentivized to identify possible carriers for testing from within their community or sector of activity, or whether slum dwellers can be incentivized to iden- tify COVID-affected people at home for testing and treatment. More generally, most development actors recognize the potential for running inexpensive nudging or information campaigns through IT. Yet we know little about whether recipients actually read, understand, or believe the messages they receive, and whether they forward or post these messages to people they do not know. Our paper fills this knowledge gap. It also shows that incentivizing the spread of non-rival information may backfire: the sharing of non-rival information between strangers is best helped by keeping things simple. 2. Baseline Intervention The purpose of our experimental design is to test two main assertions: whether people believe truthful and valuable information received from a stranger, and whether people are willing to transmit non-rival information that is valuable to strangers. The intervention to which subjects are exposed—i.e., receiving an SMS message that can be shared with others—is similar to many policy interventions in developing 5 Mobile phones have also been used to conduct surveys (e.g., Garlick et al. 2020). 862 Batista, Fafchamps, and Vicente countries. We then introduce a number of treatments that, based on theory and past evidence, can be expected to increase information usage and circulation. The causal chain we have in mind for our interventions can be summarized as a sequence of five neces- sary conditions. The subject (a) receives the message, (b) reads the message, (c) understands the message, Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (d) believes the information contained in the message, and (e) acts upon that information by sending a message. Care is taken in the experiment to ensure that condition (a) is satisfied and that condition (c) is satisfied conditional on reading the message with care. Because reading and understanding a message re- quires effort, subjects may not make this effort if they do not trust unsolicited messages, i.e., if they do not believe such messages to be in their interest. To attenuate this concern, we alert participants beforehand and obtain their voluntary participation in the study. Unlike other studies of information sharing that rely on existing social networks, we randomly assign subjects to a set of strangers with whom they can share the SMS; they cannot share it with anybody else (e.g., Centola 2010). The purpose of this design choice is to eschew endogeneity concerns that affect causal inference about interventions that rely on preexisting social links.6 Exogenous peer assignment has been used in a number of recent randomized controlled trial (RCTs) (e.g., Cai and Szeidl 2018; Fafchamps and Quinn 2018) to eliminate confounds due to the self-selection of social links (e.g., Berg et al. 2019; Bandiera et al. 2020). This design choice has one disadvantage: given that trust and altruism are likely to be lower between strangers than between socially connected individuals, our findings on information sharing should be seen as a lower bound on the propensity to make use of valuable information received by SMS (i.e., redeeming the voucher) and to share that information with others by SMS (i.e., sending the voucher). Our design does, however, offer a number of advantages in terms of external validity that experiments using existing social networks often do not have. First, it obviates some serious endogeneity concerns associated with using existing social networks, as discussed above. Second, information sharing among strangers is not rare. Messages uploaded on open forums or social media can be reposted and, as such, have a vocation to be shared with strangers. It is indeed common on social media for people to disseminate information that originates from an unknown source (e.g., Allcott and Gentzkow 2017; Alatas et al. 2022). Our findings throw new and valuable light on these processes. In the remainder of this section we present the experimental design in detail. We first describe the link structure used throughout the experiment. We then discuss the baseline intervention and the anonymity treatment and present the main empirical results of this intervention. 2.1. Link Assignment Random assignment of links is organized as follows. After having selected 192 experimental participants among rural dwellers with experience of mobile money, we divided them into 12 groups of 16 individuals 6 To illustrate, imagine that the experimenter “seeds” an existing social network by giving a piece of information to one person, and then documents that individuals close to the seed are more likely to have that information at endline (e.g., Banerjee et al. 2013). Can this be interpreted as evidence that information diffuses along the existing social network? Not necessarily. One possibility is that the original social network “rewired” as individual interested in the information sought to access it (e.g., Comola and Prina 2021; Banerjee et al. 2021a), such that information actually diffused among new links. To the extent that distance to the seed in the original network is correlated with distance to the seed along new links, it will “predict” receiving the information without having channeled it. Another possibility is that the seed shares the received information to those who ask, and proximity to the seed in the original network is correlated to individual propensity to seek out information in general, and hence to obtain information from the seed. In this case, information is transferred directly from the seed to the respondent, but this transfer is triggered by the respondent. Again, distance from the seed in the original network predicts getting the information but, in this second example, there is no diffusion along any social network, old or new. Our design circumvents these difficulties since (a) information can only pass from the seed to the target through an SMS transfer that we observe directly and (b) subjects are assigned links exogenously and at random. The World Bank Economic Review 863 Figure 1. A Square. Position Position Position Position 1 2 3 4 Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Round 1 I₁₁ I₁₂ I₁₃ I₁₄ Round 2 I₂₁ I₂₂ I₂₃ I₂₄ Round 3 I₃₁ I₃₂ I₃₃ I₃₄ Round 4 I₄₁ I₄₂ I₄₃ I₄₄ Source: Authors. that we call squares (see fig. 1). When assigning people to a square, we make sure that individuals in the same square are initially unrelated to each other. This minimizes the likelihood of communication outside the control of the experiment. As illustrated in fig. 1, a square is a 4 × 4 grid of 16 subjects Irp , where r denotes the round and p denotes the subject’s position in the round. We build information-sharing links between rows of the same square as follows: each element in row 1, i.e., subjects I11 to I14 , is allowed to transfer the SMS to each and every subject in row 2, I21 to I24 ; each subject in row 2 is similarly to transfer the voucher to each and every subject in row 3, I31 to I34 ; and each subject in row 3 is allowed to transfer to subjects I41 to I44 . Since subjects in row 1 receive SMS’s that originate directly from the experimenter, they may trust and share them more. Rows 3 and 4 are added to test this possibility by comparing the behavior of subjects in row 1 to that of subjects in rows 2 to 4. Since subjects are randomly assigned to a square and to a position in that square, there is no self- selection possible across pairs. The fact that a subject in rows 2 to 4 receives a voucher is therefore random from their point of view. There may be characteristics that induce subjects in row 1 to send the voucher to others. But, by design, these characteristics are uncorrelated with the recipients of the voucher. This eliminates the risk that the characteristics of sender and receivers may be correlated with each other in some way. Put differently, selection into receiving a voucher is random since subjects are randomly assigned to a position that, in that square, receives a voucher. All contacts between participants take place through text messages mediated by the experimenter, i.e., subjects pass information to each other by using text messages relayed by our switchboard from one subject to another. Subjects are never told the identity or phone number of the person with whom they are sharing information. All the messages received by participants come from the switchboard and are written in Portuguese—see supplementary online appendix tables S1.1 to S1.8 for the full list of original messages used in the experiment, together with their English translations. For each message sent, an experimental subject incurs at most a cost of 1–2 meticais charged by the phone operator.7 In compensation for this— and their participation time—each subject receives a participation fee of 70 meticais paid in mobile money at the end of the experiment. When the experiment took place, USD 1 was approximately equivalent to 35 meticais. 7 Virtually all subjects in our experiment use pay-as-you-go. Phone operators run occasional promotions of the form “Earn X free SMS if you top up your account by Y meticais.” 864 Batista, Fafchamps, and Vicente All interventions and treatments are implemented as a game played at the level of the square. Each game takes place over a period of a little less than a week each and is divided into rounds that each take approximately 24 hours, i.e., subjects in a round have 24 hours to redeem the voucher and to share it with up to four others. This basic structure applies to each game, with some differences across treatments, Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 as described below. We now describe in more detail the baseline intervention at the level of a square. 2.2. Intervention Design The baseline intervention (T0) starts with a seeding round, i.e., round 1. In this round, after an introduc- tory message by the experimenter, each individual in the first row of the square—i.e., I11 to I14 —receives an SMS from the experimenter asking whether they want to receive 35 meticais—approximately USD 1—on their mobile money account. To receive the money, the subject has to send a message back with the word “yes” or “sim” (its Portuguese equivalent), upon which mobile money is automatically trans- ferred to the respondent’s mobile money account. No further action is required to receive the 35 meticais. Subjects who fail to respond do not receive the transfer. Each round 1 subject then receives messages asking if they want us to give the same voucher to round 2 participants. Subjects receive four such messages, one for each of the four round 2 participants. To instruct us to send the voucher to this other person, the subject has to reply to each initial SMS with another SMS message containing the word “yes/sim.” Since each of the four senders in round 1 can send the voucher to each of the receivers in round 2, subjects in round 2 can receive up to four vouchers. Those who do not receive any SMS voucher from round 1 participants are dropped from the game. The remaining round 2 participants first receive an introductory message from the experimenter before receiving the SMS voucher itself. In round 2 the SMS voucher is worded slightly differently: it explicitly states that the voucher is sent at the request of another participant in the experiment. Since there are four round 1 subjects who could have sent the voucher, a round 2 subject can receive up to 4 × 35 meticais. To receive the money, the subject has to reply to each of these messages with the word “yes/sim.” After this, round 2 subjects receive messages asking if they want us to give a voucher to round 3 participants. As in round 1, they receive four such messages, one for each round 3 participant. Round 2 participants have to reply “yes/sim” by SMS to each of those messages if they wish to send the voucher to the corresponding round 3 participant. Based on these responses, a list is drawn of those round 3 subjects who are to receive the SMS vouchers. Round 3 follows the same structure as round 2. Round 4 starts in the same way: subjects I41 to I44 receive the SMS voucher for each of the round 3 subjects who has instructed us to do so. But since this is the last round, they are not asked about sending the voucher to other players. Each reply to the experimenter, i.e., both on willingness to receive the voucher and to share it, has to be answered within 24 hours to be admissible. Messages received after this deadline are ignored.8 This deadline ensures that each square follows a similar sequencing—similar to what happens in a lab experiment. Using four separate phone numbers—one for each of the four receiving and four sending decisions—makes it possible for the experimenter to identify the sender and intended recipient of each of the messages received on our switchboard. Payoffs are paid on the mobile money account of each subject at the end of the game. There are two variants of this baseline intervention: a no-information variant N and a partial- information variant P. Six of the 12 squares of 16 individuals are assigned to the no-information variant N and the other half to variant P. Since subjects are randomly assigned to squares, this guarantees random assignment to variants. In the no-information variant, no information on the other participants is provided to either sender or receiver: all the sender knows is that another participant in the study will receive an SMS voucher like the 8 At the time of the study, it was extremely unlikely to lose phone service for more than an hour in the part of Mozambique where our study took place. Very few attempts were made to redeem after the 24-hour window expired. The World Bank Economic Review 865 one they received; similarly, all that the receiver knows is that another study participant has instructed the experimenter to send them an SMS voucher. Individuals in the previous or following row are referred to as “Person p” with p = 1, …, 4. In the partial-information variant, the sender is told about some characteristics of the receiver—namely Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 gender, age, schooling, and income category. The receiver is given analogous information about the sender. Information on gender is implicitly conveyed through the first name of the sender or receiver (which is spelt out in the message), age is given in years, education is given in years of completed schooling (up to 12th grade) or as the type of post-secondary education (e.g., bachelors or masters), and income is given as one of seven possible categories of monthly income. In general we expect subjects to empathize more with senders and receivers for whom they have some information that enables them to “put themselves in their shoes” (e.g., Kirman and Teschl 2010). As a result, we expect more redeeming and more sending in the informed variant. In addition, subjects may empathize more with individuals with characteristics similar to themselves and with whom they identify (e.g., Akerlof and Kranton 2000; Bauer et al. 2018). If this is true, we expect more sending of messages to individuals with shared characteristics. 2.3. Sampling and Implementation We implemented the design as a field experiment in Mozambique from May to July 2015. Participants were recruited among heads of households or their spouses who had taken part in an RCT entitled “Project mKesh” on the introduction of mobile money in rural Mozambique—a study that took place from June to August 2012 and is described in Batista and Vicente (2013, 2022).9 Both the 2012 RCT and this study were conducted by NOVAFRICA, a development economics research center located at the Nova University in Lisbon, Portugal. The “mKesh NOVAFRICA” label was used throughout this study to build upon the confidence already gained from respondents in the anterior RCT study. The sample for our field experiment is drawn from a representative sample of rural enumeration areas with mobile phone coverage in the Mozambican provinces of Northern Maputo, Gaza, and Inhambane. Within each of the 102 enumeration areas sampled for that study, an average of 19 households per enu- meration area was selected through a random walk process—i.e., by walking from the center of the enumeration area in different directions and inviting each nth house along the way to participate in the study. The original sample was selected in 2012 and was followed as a panel, with several survey rounds (the last of which was in mid-2014). In half of the sample, i.e., in 51 randomly chosen enumeration areas, mobile money was introduced through the recruitment of a local agent and the organization of various dissemination activities at the enumeration area level. Within these locations, a random subsam- ple was targeted for individual dissemination of mobile money. By design, participants in the experiment thus more knowledgeable than the average Mozambican about mobile phone communication and mobile money services. In this paper we focus on individually treated households from the original sample. This ensures that all participants had previously been introduced to mobile money, had used the service, and had an active mobile money account on their mobile phone at the time of the experiment.10 Most of the 192 individuals in our study were recruited by phone or SMS message. Some were recruited through face-to-face contact. At the time of their recruitment for the study, subjects were told that, over a period of three weeks, they would receive SMS messages from Project mKesh and that paying attention to these messages would enable them to earn money on their mKesh mobile money account. They were also told that they would receive a participation fee paid at the end of the experiment. To someone accustomed to paying or being 9 mKesh is the name of a commercial mobile-money service provider in Mozambique. 10 Batista and Vicente (2020) show that mobile money adoption and usage over time among treated individuals in this sample does not depend on age, gender, or expenditure. Mobile money users are, however, likely to be better educated than non-users within this sample of treated individuals. 866 Batista, Fafchamps, and Vicente paid with mobile money, like our subjects are, the idea of receiving a transfer of funds to their mobile money account is a natural one: this is how our respondents use their mobile money account on a regular basis, e.g., to send money to a distant relative or to pay for goods in a shop. Informed consent was obtained at the time of recruitment. Subjects were reminded of their participation in the experiment by an SMS Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 message just before starting the base game. Because the mobile phone operator sends marketing SMS’s to individual subscribers on a regular basis, our messages may be misconstrued as spam. To minimize this concern, we label all our messages to sub- jects as coming from “Project mKesh NOVAFRICA.” This identifier is associated with the 2012 “Project mKesh” RCT in our study area and it was also used to recruit subjects into this study. All the SMS mes- sages sent to participants are short, to stay within the character limit of a single SMS in Mozambique. The actions required of subjects are reduced to their simplest expression, namely reply with “yes/sim” to our message.11 It remains that the experiment takes place in a real-life environment where unsolicited SMS’s are common. While this increases the possibility that our messages are seen as spam, it also raises the external validity of our findings, since any information campaign using SMS services in Africa is bound to face the same problem. The split of the 192 participants into 12 squares follows a random procedure that ensures that no two subjects from the same enumeration area are allocated to the same square. This is done to avoid the possibility of direct communication between subjects. The last survey round held in mid-2014 is the source of the information on individual characteristics that is used in the partial-information variant P of the baseline intervention. Funding for this research was provided by the International Growth Center. The experiment was implemented in collaboration with Carteira Móvel/Mkesh and the NOVAFRICA office in Mozambique. All SMS messages were sent and relayed by research assistants recruited for the project. Key characteristics of the sample are presented in table 1. Approximately 59 percent of participants are female, and the average participant is 40 years old. Non-college-educated participants constitute 96 percent of our sample and have 6 years of education on average. Average monthly income is 3,445 meticais, which is approximately equal to USD 98 per month. Table 1. Sample Characteristics and Balance Years Post- Income in Age of 0–12 secondary ’000 meticais/ Female in years education education month Sample characteristics Sample mean 0.589 39.963 6.175 0.042 3.445 Sample standard error (0.036) (1.003) (0.235) (0.015) (0.420) Balance across squares Proportion of pairwise comparisons between squares 2/66 2/66 7/66 8/66 0/66 that are significant at the 10% level Joint F-test of balance across all squares p-value 0.762 0.818 0.195 0.126 0.934 Joint F-test of balance across the partial-information p-value 0.189 0.358 0.126 0.481 0.963 and no-information treatments Joint F-test that games 1–2–3 = games 4–5–6 p-value 0.662 0.632 0.813 0.481 0.417 Source: Authors’ calculations based on the experimental data collected by the authors. Note: Pairwise comparison tests are obtained by regressing the variable of interest on a square dummy, using only two squares at a time, and counting how many times the dummy is significant. There are 66 (i.e., N(N − 1)/2) possible pairs of 12 squares, where N is the number of squares. Using a 10 percent significance level, there should on average be 10 percent significant dummies (i.e., 6.6) if the null of perfect balance across all squares is true. Balance across all squares is tested by regressing the characteristic of interest on square dummies and performing a joint F-test of all dummies. Balance between games 1–2–3 and games 4–5–6 is tested by regressing the characteristic of interest on a games 4–5–6 dummy. Balance across the no-information and partial-information treatments is tested by regressing the characteristic of interest on the partial-information dummy. The p-values from these tests are reported in the table. 11 Entering “y” or “s” is considered as a yes, entering “no” or “n” is regarded as a no. The World Bank Economic Review 867 Table 1 also presents balance tests across experimental treatments. It begins by comparing each pair of squares in terms of demographic characteristics. Across the 330 differences we tested (66 pairwise tests × 5 variables), we find a total of 19 that are statistically significant at the 10 percent level—well below what would be expected to occur by chance (10 percent). We additionally test for the joint significance of Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 square dummies to check for systematic differences between squares, and we compare subjects in partial- information and no-information treatment squares. All these tests fail to reject the null hypothesis of no difference for each of the observable characteristics considered. Randomization thus appears to have achieved balance on key individual characteristics across squares. In supplementary online appendix table S2.1 we present power calculations for all the main tests presented in the empirical analysis. Although we do not reach the maximum achievable power allowed by our design due to the fact that not all subjects in later rounds receive a voucher, we nonetheless have sufficient power to detect effects of the magnitude uncovered by our analysis. 2.4. Experimental Results in the Baseline Intervention Figure 2 shows the average behavior of the experimental subjects in the baseline intervention. We find that the probability of a participant redeeming the voucher is 26 percent, while the probability of sending the voucher to any of the four subjects in the next row is 24 percent. We interpret the 26 percent probability of redeeming vouchers as evidence that a large proportion of participants do not accept what is essentially a “free lunch”: by replying to the SMS voucher offer with a “yes/sim” SMS message at a cost of 1–2 meticais, they would have received 35 meticais. Given that subjects are selected because of their familiarity with mobile phones and active usage of mobile money, this cannot be due to lack of familiarity. Furthermore, the research team secured explicit agreement from each individual subject to participate in the experiment, and reminded each participant individually, shortly before the baseline intervention was implemented, that messages would follow containing opportunities to earn money. These efforts attenuate the possibility that lack of trust may arise purely from an insufficient understanding of the messages, conditional on reading them attentively. It is nonetheless conceivable that Figure 2. Redeeming and Sharing Behavior in the Baseline Intervention Base Game .5 .4 Proportion of Individuals .1 .2 0 .3 Redeeming Sharing All Subjects Partial information No information Source: Authors’ calculations based on the experimental data collected by the authors. Note: Redeeming the voucher means responding with a “yes/sim” SMS to our switchboard. Sending the voucher means responding with a “yes/sim” SMS to an SMS invitation to share information about the voucher with another randomly selected subject. The height of the bars represents the proportion of experimental subjects redeeming/sending mobile money vouchers. In the no-information treatment, no information is given about the recipient or the sender. In the partial-information treatment, some limited information (i.e., gender, age, schooling, and income category) is given about the recipient or the sender, but never their identity or phone number. Confidence intervals are plotted for a 10% significance level. 868 Batista, Fafchamps, and Vicente subjects read the message in a distracted manner and thus failed to grasp its meaning. Choosing not to pay attention is more likely if SMS messages are treated with suspicion to start with, e.g., because subjects do not believe the message is likely to benefit them. Alternatively, subjects may fear that, in order to receive the money, they will be asked to do more than answer “si” to the first SMS. This would indicate that they Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 believe the message to be disingenuous, discarding what they were told in our information campaign. Given the very low cost of redeeming, we therefore suspect that subjects who do not redeem the voucher are a priori skeptical about our SMS messages and, for this reason, they either dismiss them outright, do not read them attentively, or disbelieve what they have read.12 In contrast, the propensity to share vouchers appears relatively high, given the cost of sending messages and the absence of a material benefit for the sender. One possible interpretation is that sending follows a “warm glow” motivation: subjects seem keen to share with others a valuable opportunity, even if they themselves are skeptical about it. Some evidence to this effect comes from observing that, among the players given the opportunity to both redeem and send vouchers, 11 percent send at least one voucher but do not redeem themselves. Together they represent 33 percent of the subjects who send any voucher. These findings are reminiscent of Allcott and Gentzkow (2017): fake news stories circulated widely on social media during the 2016 US election even though at least half of those who read them did not believe them. Turning to the difference between the partial-information and no-information variants of the base- line intervention, we find that, contrary to our hypothesis, there is more redeeming and sending in the no-information variant. Although this difference is not statistically significant in the baseline interven- tion taken in isolation, it becomes significant when we include observations from the other treatments introduced below. This point is revisited in detail later. 3. Exploring the Reasons for Low Redeeming and Sharing The results from the baseline intervention show that most subjects do not take the mobile money vouchers seriously enough to redeem them, even though they share these vouchers with others. In addition, infor- mation sharing is reduced when subjects receive information on the characteristics of voucher senders and recipients. As a result, information diffusion fails to spread to all subjects in row 4 of each square. These findings demonstrate that simply allowing the transmission of valuable but non-rival information is insufficient to trigger an information cascade in our setting. For this reason, we introduce a series of treatments intended to vary credibility and the cost of informa- tion sharing. These treatments are introduced to our subjects as additional games of the experiment, which are played in random order. We first present the experimental design and sequencing of these treatments, before discussing our testing strategy and examining our empirical results. 3.1. Experimental Design and Sequencing of Treatments T1/T2/T3 Half of our experimental subjects were invited to three additional games after the baseline intervention.13 Each of these games shares many common features with the baseline intervention, but we vary the cost of sending SMS vouchers. If information sharing is hindered by cost considerations, we expect a dramatic drop in information circulation once we increase the cost of sharing vouchers. We also vary the type of cost—monetary or psychological—that subjects incur for sending (or not) SMS vouchers to other subjects. 12 Lack of interest may also arise because subjects do not want to be seen, e.g., for religious or cultural reasons, as someone who is in need of receiving gifts. This is conceivable but unlikely given that respondents knew that participating would enable them to earn money and their informed consent about this was solicited and obtained before their participation in the study. Moreover, all participants received a participation fee at the end of the study and no one returned the money to us. 13 The other half were assigned to a different batch of treatments, discussed below. The World Bank Economic Review 869 Lab experiments have suggested that subjects can be induced to undertake actions based on self-image considerations—e.g., whether an action is the “right thing to do” based on shared social norms (e.g., Tirole 2002; Akerlof and Kranton 2005; Johansson-Stenman and Svedsäter 2012). To see whether this mechanism can be used to induce subjects to share valuable information, we introduce a treatment in Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 which subjects are “shamed” anonymously for not sharing the voucher. In addition, we introduce a treatment in which subjects pay a lower cost for sharing misleading infor- mation. If subjects care about their social image (e.g., Tirole 2002; Andreoni and Bernheim 2009; Bursztyn and Jensen 2017)—i.e., they want to be seen to do the “right thing” in the eyes of others—but are not altruistic towards them, sending a seemingly generous message to others may appeal to them. Senders may also circulate misleading messages if they have invidious or rival preferences—or are mischievous. In contrast, if information sharing is motivated primarily by altruism, we do not expect the sharing of untrue messages. By varying these experimental parameters, we aim to throw light on the role of cost and lack of credibility in the imperfect message transmission observed in the baseline intervention. We now describe the design of each of the three treatments. As in the baseline intervention, each of them is played in four rounds within a square with 16 subjects as depicted in fig. 1. Treatment T1 (the “variable sending cost” treatment) introduces an additional cost of sending the voucher to another subject. This cost takes four possible values: 0 (as in the baseline intervention), 5, 10, or 15 meticais per shared message. It is paid on top of the 1–2 meticais that is charged per SMS by the phone provider. Each subject faces each of the four different cost levels in a randomized order, in each of the subsequent rounds. Incurred costs are deducted from the payoff sent to the subject’s mobile money accounts at the end of the game. In all other respects, this treatment is the same as the baseline intervention. Varying the cost of sending the voucher allows us to infer subjects’ willingness to pay for sending valuable information to others. Treatment T2 (the “fixed sending cost and shaming” treatment) presents subjects with a different default option when sending vouchers to others. In the baseline intervention and in treatment T1, if the subject does not respond to the initial SMS sent by the experimenter, no action is taken—i.e., no voucher or message is sent to the potential recipient. In contrast, in treatment T2 the default is that, in the event that the subject takes no action (i.e., responds “no” or does not reply), the experimenter sends a message to the recipient revealing that the sender was given an opportunity to pass on the voucher but failed to do so—as a consequence of which the recipient is unable to win 35 meticais. In this treatment, the cost of sending is set to 5 meticais—in addition to the phone operator’s cost per SMS. The rest of the design is the same as in the baseline intervention. The purpose of this treatment is to increase the psychological cost of not sending the voucher to others. To put it more bluntly, it shames the sender for failing to send the voucher. As a result we expect it to increase sharing. To the extent that shame is related to social image within a group sharing a similar identity, we expect this treatment to be particularly effective in the partial-information variant when experimental subjects know each other’s characteristics. Treatment 3 (the “fixed sending cost and erroneous code message” treatment) adds a second default option to treatment T2 when subjects are asked about the sending of vouchers. In the same way as in treatments T0 and T1, if the sender does not reply “yes/sim” to the offer to share the voucher, no further action is taken by the experimenter. Similar to treatment T2, if the sender responds “yes/sim” to the initial message sent by the system, the SMS voucher is sent to the recipient and a fixed price of 5 meticais is deducted from the sender’s payoff. If the sender responds “no,” the receiver gets an SMS containing an erroneous code that cannot be redeemed for money.14 The remainder of the design is as in the baseline intervention. The purpose of this treatment is to disentangle an explicit decision not to share—e.g., mo- tivated by rival or invidious preferences—from simple inaction. In treatment T2, these two motives are confounded. In treatment T3, if the sender sends an incorrect voucher to the recipient by responding “no” 14 To avoid deceiving the subject, this is made clear in the message sent to the recipient—see the supplementary online appendix for details. 870 Batista, Fafchamps, and Vicente Figure 3. Sequencing of Treatments Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Source: Authors. Note: T0 to T6 refer to the type of treatment, T0 being the baseline intervention. Block 1 focuses on treatments T1–2–3 and block 2 on treatments T4–5–6. The baseline intervention is only played in period 1. In periods 2 to 4, the order of treatments is varied within blocks to ensure an equal number of squares playing each combination of treatments. The symbol N refers to the no-information treatment in which no information is given about the recipient or the sender, and P refers to the partial-information treatment in which information on the gender, age, schooling, and income category is given about the recipient or the sender, but never their identity or phone number. (at the small cost of sending an SMS), this clearly manifests a desire not to share with the recipient—as opposed to inattention or inaction. In the experiment, all the squares—i.e., groups of 16 subjects—are first subjected to the baseline inter- vention. Half of the squares are then subjected to the three treatments T1, T2, and T3 in three separate games played in a prespecified order. This allows us to achieve identification within subjects. The six squares are divided into two sets of three squares: one set always plays the no-information variant N, the other always plays the variant where the characteristics of senders and recipients (gender, age, education, and income range) are provided. Within each of these groups of three squares, the order of treatments T1/T2/T3 is varied systematically, ensuring balance in the order in which they are played. The assignment structure of treatments to squares is depicted in the top panel of fig. 3, where Ti stands for treatment i and N/P stand for the no-information and partial-information variants, respectively. 3.2. Testing Strategy We split our analysis between the decision to receive mobile money from others, and the decision to send mobile money to others. In each case, we test for differences across treatments, whether no information is provided on sender and receiver, and whether sending and receiving vary systematically with subject characteristics when information on these characteristics is provided. In addition to reporting average choices for each treatment, we report results from a regression analysis. For receiving or redeeming vouchers, we use the following core specification: Ri jrt = α + β1 G1 2 3 i jrt + β2 Gi jrt + β3 Gi jrt + γ Pi + δr + εi jrt , (1) where the dependent variable Rijrt is a binary variable taking value 1 in the case that subject i redeemed a voucher opportunity sent by subject j in round r and period t. Regressors are as follows: Gk i jrt is a treatment k dummy variable, Pi is a binary variable equal to 1 in the partial-information variant, and δ r is a vector The World Bank Economic Review 871 of round and period dummies, included to control for the possibility that experimental fatigue or loss of attention affects our findings. We also estimate a specification that adds prior redeeming in earlier games to see whether a positive experience with redeeming in an earlier game spurs more confidence in voucher messages. Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 To test the role of empathy due to a shared identity, we estimate a model that includes absolute differ- ences |Xi − Xj | in individual characteristics X between subject i and the subject j from whom i received the voucher.15 We only use the four characteristics Xi that are revealed to i about j—and vice versa. Since pair- wise characteristics are only revealed to subjects in the partial-information variant, |Xi − Xj | is interacted with the partial-information dummy Pi . When estimating this regression we also include characteristics Xi and absolute differences |Xi − Xj | as additional controls.16 The estimated regression is thus of the form Ri jrt = α + β1 G1 2 3 i jrt + β2 Gi jrt + β3 Gi jrt + γ Pi +, θ |Xi − X j |Ii + μXi + λ|Xi − X j | + δr + εi jrt . Empathy towards similar people implies θ < 0—i.e., the more dissimilar i and j are, the less i is willing to redeem a voucher from j.17 When estimating regression (2), we only include redeeming decisions that apply to SMS vouchers received from another subject—i.e., we drop observations from round 1 subjects who receive the voucher from the experimenter. To examine sending behavior, the baseline specification for treatments T1/T2/T3 takes the form Si jrt = α + β1 G1 2 3 i jrt + β2 Gi jrt + β3 Gi jrt + θ Ci jt + γ Pi + δr + εi jrt , (3) where the dependent variable Sijrt is a dummy equal to 1 in the case that subject i sends a voucher opportu- nity to subject j in round r and period t. Variable Cijt is the cost of sending the voucher to another subject which, in treatments T0/T1/T2/T3, varies exogenously by subject pair ij. We also estimate a specification that includes the redeeming decision as additional control, and a specification that adds |Xi − Xj |, and controls Xi , to test for empathy towards similar subjects in sending decisions. All the econometric speci- fications are estimated using a linear probability model and the reported standard errors are clustered at the individual level (i.e., across games/periods). 3.3. Empirical Results on Treatments T1/T2/T3 3.3.1. Treatment Averages Table 2 reports the average behavior of the subjects in the baseline intervention and in each of treatment T1/T2/T3. Columns (2)–(4) of table 2 present average redeeming and sending decisions in treatments T1/T2/T3. As explained earlier, the order of the treatments varies randomly across squares, i.e., they are not necessarily played in the order in which they appear in table 2—and hence the order in which treatments T1/T2/T3 were played should not drive the results. As in the baseline intervention, the num- ber of redeeming observations is less than 192, the number of individuals in the squares, because many subjects in rounds 2–3–4 never receive any voucher they could redeem.18 Since links are assigned exoge- nously, whether a subject receives a message or not from another subject cannot be correlated with any 15 To facilitate interpretation, when Xi is a dichotomous variable—e.g., gender—we replace the absolute difference with a dummy equal to 1 if i and j have the same gender, and 0 otherwise. 16 For instance, |Xi − Xj | may be systematically larger when Xi is large. 17 When the regressor is a dummy equal to 1 if i and j share a characteristic—e.g., gender—the interpretation is reversed. 18 This happens even though several (up to four) vouchers could potentially be redeemed by each subject in rounds 2–4. The number of sending observations is higher than the number of redeeming observations because each subject who receives a voucher in rounds 1–3 is automatically given the option to send it to four other subjects, while these subjects can only redeem one voucher. Note that the variation in the number of observations is a consequence of our experimental design, which is aimed at investigating how far information diffuses among strangers. 872 Batista, Fafchamps, and Vicente Table 2. Average Choices Made by Subjects in Treatments T0/T1/T2/T3 Treatment 3: Treatment 1: Treatment 2: erroneous message Baseline variable cost shaming and fixed and fixed cost intervention of sending cost of sending of sending Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (1) (2) (3) (4) Redeeming the voucher All subjects 25.9% 13.3% 15.8% 18.8% (0.037) (0.051) (0.060) (0.070) Round 1 only (1) 27.1% 12.5% 12.5% 16.7% (0.065) (0.069) (0.069) (0.078) Rounds 2–4 (2) 25.3% 14.3% 21.4% 25.0% (0.045) (0.078) (0.114) (0.164) Partial information 21.7% 10.7% 12.5% 14.3% (0.043) (0.060) (0.069) (0.067) No information 33.3% 17.6% 21.4% 50.0% (0.067) (0.095) (0.114) (0.289) Number of observations 143 45 38 32 Sending the voucher All subjects 24.2% 14.3% 10.1% 6.3% (0.022) (0.029) (0.026) (0.021) Partial information 21.7% 7.0% 0.0% 6.3% (0.029) (0.034) (0.000) (0.030) No information 26.8% 18.9% 14.7% 6.3% (0.032) (0.041) (0.037) (0.030) Number of observations 392 147 139 128 Source: Authors’ calculations based on the experimental data collected by the authors. Note: Each number in the table is the percentage of decisions falling in the relevant category, with standard errors reported in parentheses. Redeeming the voucher means responding with a “yes” SMS to our switchboard. Sending the voucher means responding with a “yes” SMS to an SMS invitation to share information about the voucher with another randomly selected subject. In treatment 3, the zero value includes both alternatives to sending. Only two subjects sent the erroneous voucher. (1) In round 1 the voucher SMS is sent by the experimenter. (2) In rounds 2–4 the voucher SMS is sent at the request of a subject. unobservable characteristic of the potential recipient—and hence can be regarded as random for the pur- pose of inference. We observe a dramatic drop in both redeeming and sending behavior in treatments T1/T2/T3 relative to the baseline intervention. The voucher redemption rate falls by between 27 (T3) and 49 (T1) percent, even though the cost of redeeming it is the same across treatments. Sending in treatments T1/T2/T3 falls relative to the baseline intervention by an even larger percentage (between 41 percent in T1 and 74 percent in T3), possibly because the cost of sending is higher in these treatments relative to the baseline intervention. Contrary to expectations, sending is more common in T1 than in T2 and T3, even though the cost of sending is, on average, highest in T1. The propensity to send is lower in T2 than in T1—suggesting that changing the no-reply default action to a shaming message did not create a psychological pressure to give. This is reminiscent of situations (e.g., DellaVigna et al. 2012) in which individuals give because they perceive a moral pressure to do so but feel exonerated if a device (in our case, a default erroneous message) takes an action for them. In T3, subjects could either pay 5 meticais to send an SMS voucher to the receiver, send an erroneous voucher message, or do nothing. In practice, we only observe two cases of a subject sending an erroneous voucher message, making this treatment similar to T1 with a slightly lower cost of sending on average. We nonetheless observe a further decrease in the sending probability, which now falls to 6 percent. One possible explanation is that the introduction of an irrelevant but selfish The World Bank Economic Review 873 Table 3. The Decision to Redeem the Voucher in Treatments T0/T1/T2/T3 (1) (2) (3) Treatment dummies (T0 is omitted category) Treatment 1 dummy (variable cost) −0.192*** −0.284*** −0.244** Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (0.069) (0.092) (0.110) Treatment 2 dummy (shaming and fixed cost of sending) −0.194*** −0.286*** −0.302* (0.063) (0.083) (0.167) Treatment 3 dummy (erroneous message and fixed cost of sending) −0.128 −0.199* −0.164 (0.088) (0.104) (0.120) Partial information dummy −0.162* −0.149* −0.343 (0.087) (0.083) (0.262) Dummy=1 if subject redeemed a voucher in a previous period — 0.309*** — (0.115) Pairwise differences in individual characteristics times the partial-information dummy Same gender — — 0.194 (0.136) Same post-secondary education dummy — — −0.018 (0.163) Absolute difference in age — — 0.007 (0.008) Absolute difference in income (in ’000 meticais/month) — — −0.006 (0.012) Round dummies Yes Yes Yes Period dummies Yes Yes Yes Treatment sequence dummies Yes Yes Yes Individual characteristics No Yes Yes Pairwise differences in individual characteristics (uninteracted) No No Yes Intercept 0.467*** 0.613*** 0.996** (0.104) (0.193) (0.409) Adjusted R-squared 0.017 0.146 0.045 Number of observations 258 244 117 Joint coefficient tests Test that T1 (β 1 ) = T2 (β 2 ) p-value 0.960 0.975 0.688 Test that T1 (β 1 ) = T3 (β 3 ) p-value 0.407 0.293 0.427 Test that T2 (β 2 ) = T3 (β 3 ) p-value 0.263 0.144 0.440 Source: Authors’ calculations based on the experimental data collected by the authors. Note: All regressions are estimated by ordinary least squares (OLS). The dependent variable is a binary variable defined as 1 if, when given the chance, the subject sends an SMS accepting the voucher. In column 3 we only include observations from rounds 2–3–4 since, in round 1, all SMS’s originate from the experimenters and thus differences in individual characteristics are not defined; we also include as controls the pairwise differences in individual characteristics uninteracted with the partial-information dummy. Individual characteristics include a female dummy, age, a post-secondary education dummy, and income in meticais/month. Clustered standard errors, at the level of the individual, reported in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. alternative prompts subjects to act selfishly. Similarly to the baseline intervention, no-information variants of treatments T1/T2/T3 yield higher redeeming and sending rates than their partial-information variant. 3.3.2. Redeeming the Voucher To fully assess the determinants of redeeming vouchers in treatments T0/T1/T2/T3, we regress the re- deeming decision as specified in the testing strategy section. The dependent variable is a binary variable taking value 1 if the subject sends a “yes/sim” SMS in response to a voucher offer, and 0 otherwise. The results are shown in table 3. Column (1) reports the results from regression model (1).19 In column (2) 19 For the analysis of redeeming, the partial-information dummy takes value 1 (partial information) in round 1 since subjects know that vouchers originate from the experimenter. 874 Batista, Fafchamps, and Vicente we add a dummy variable with value 1 if the subject redeemed a voucher in a previous game: subjects who trust the SMS enough to redeem it in one game should also be more likely to trust it in a subsequent game. Column (3) reports the estimates for model (2) that tests for the effect of shared characteristics. In addition to regression coefficients, at the bottom of table 3 we report test statistics of the null hypothesis Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 that there is no difference between pairs of treatments. Regression analysis confirms that the probability of redeeming decreases between the baseline interven- tion and the other three treatments although, for T3, this is only significant in column (2). The reduction in redeeming is large relative to the counterfactual probability of redeeming in the baseline intervention: the probability of redeeming drops by 18 to 30 percentage points in T1 and T2 relative to T0, and by 21 percentage points in T3. Pairwise comparisons reported at the bottom of table 3 nonetheless indicate that we cannot reject the hypothesis that redeeming is equally likely under treatments T1, T2, and T3. As already observed in table 2, we find a large reduction in redeeming in the partial-information variant: this difference is about 15 to 16 percentage points and is statistically significant in the main specification (columns 1 and 2). This confirms that subjects are more likely to redeem a voucher that comes from an individual on which they have no information. We also observe more redeeming in round 1, that is, when the voucher originates from the experimenter, than when the voucher comes from another subject. We do not find systematic treatment order effects. Since payoffs are deposited on subjects’ mobile money accounts at the end of each game, subjects who redeem in a given game receive the voucher money at the end of that game. This should make them more confident of receiving the voucher money in subsequent games. We therefore expect redeeming behavior to be persistent. This is indeed what we find: there is a strong positive correlation between redeeming now and redeeming in a previous game. We cannot, however, rule out the possibility that this captures differences in trusting behavior across subjects. When adding pairwise regressors (column 3), point estimates suggest that subjects are more likely to redeem a voucher received from a person of the same gender and education level. But none of these effects is statistically significant.20 From this we conclude that there is no conclusive evidence that shared charac- teristics matter in redeeming decisions. Perhaps this is not too surprising given that there is on average less trust in the partial-information variant. From the estimated coefficients of individual characteristics Xi , we also note that older subjects redeem less and richer participants redeem more. This could be because individuals who are younger and richer are more familiar with mobile phones and more willing to risk 1–2 meticais for the prospect of receiving 35 meticais. 3.3.3. Sending the Voucher We report in table 4 a similar analysis for the decision to send the voucher to another participant in treatments T0/T1/T2/T3. The dependent variable is a binary variable taking value 1 if the subject sends an SMS instructing the experimenter to send the mobile money voucher to another subject. Recall that there are four such decisions per voucher recipient, one for each of four possible recipients in the following round (i.e., to the next row in fig. 1). We control for the cost of sending the SMS, which varies between 0/5/10/15 meticais across subject pairs ij in T1. This cost is set at 5 meticais in T2 and T3, and 0 meticais in the baseline intervention. Column (1) of table 4 reports coefficient estimates for specification (3). In column (2) we add two redeeming dummies—one for the previous game, as in table 3, and one for the current game, just before the decisions to send. The purpose of including these control variables is to test whether subjects are more likely to send a voucher that they themselves redeem—as would be the case if sharing is done primarily 20 Similar results (not shown here) are obtained if we estimate an individual fixed effect model that compares redeem- ing behavior across different senders for the same receiver. Because the number of subjects who receive multiple SMS vouchers is relatively small, the number of observations is small and statistical power is limited. The World Bank Economic Review 875 Table 4. The Decision to Send the Voucher in Treatments T0/T1/T2/T3 (1) (2) (3) Treatment dummies (T0 is omitted category) Treatment 1 dummy (variable cost) −0.110* −0.085 −0.152** Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (0.056) (0.053) (0.075) Treatment 2 dummy (shaming and fixed cost of sending) −0.170*** −0.133** −0.166** (0.063) (0.066) (0.072) Treatment 3 dummy (erroneous message and fixed cost of sending) −0.189*** −0.184*** −0.222** (0.072) (0.065) (0.087) Partial information dummy −0.050 0.025 −0.115 (0.057) (0.040) (0.084) Additional cost of sending the voucher −0.001 0.001 0.000 (0.004) (0.003) (0.004) Dummy=1 if subject redeemed a voucher in the current period — 0.450*** — (0.068) Dummy=1 if subject redeemed a voucher in a previous period — 0.137** — (0.059) Pairwise differences in individual characteristics times the partial-information dummy Same gender — — 0.036 (0.037) Same post-secondary education dummy — — 0.058 (0.058) Absolute difference in age — — 0.000 (0.002) Absolute difference in income (in ’000 meticais/month) — — −0.001 (0.004) Round dummies Yes Yes Yes Period dummies Yes Yes Yes Treatment sequence dummies Yes Yes Yes Individual characteristics No Yes Yes Pairwise differences in individual characteristics (uninteracted) No No Yes Intercept 0.159 0.249*** 0.524*** (0.109) (0.092) (0.163) R-squared 0.070 0.391 0.168 Number of observations 806 770 731 Joint coefficient tests Test that T1 (β 1 ) = T2 (β 2 ) p-value 0.066 0.280 0.806 Test that T1 (β 1 ) = T3 (β 3 ) p-value 0.016 0.050 0.035 Test that T2 (β 2 ) = T3 (β 3 ) p-value 0.611 0.162 0.286 Source: Authors’ calculations based on the experimental data collected by the authors. Note: All regressions are estimated by ordinary least squares (OLS). The dependent variable is a binary variable defined as 1 if, when given the chance, the subject sends an SMS giving the voucher to another subject. In game 3, sending the false message (only 2 observations) is assimilated to not sending the voucher. The additional cost of sending the voucher is 0 in the baseline intervention, 5 meticais in treatments 2 and 3, and varying between 0/5/10/15 meticais in treatment 1. There is no sending in round 4. In column 3, we also include as controls the pairwise differences in individual characteristics uninteracted with the partial-information dummy. Individual characteristics include a female dummy, age, a post-secondary education dummy, and income in meticais/month. Clustered standard errors, at the level of the individual, reported in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. by those who trust the message enough to redeem it. Column (3) includes |Xi − Xj |Ii and related controls as additional regressors to test for the role of shared characteristics in sending choices. As already noted when discussing table 2, we observe a strong reduction in sending probability be- tween the baseline interventions and treatments T1/T2/T3. These differences are all large in magnitude and statistically significant, ranging between 9 and 22 percentage points depending on the specification. Given that sending is more costly in treatments T1/T2/T3 than in T0, these findings suggest that sharing 876 Batista, Fafchamps, and Vicente information is cost sensitive. However, the cost of sending a message, which varies randomly in T1, has no significant effect on the probability of sending a voucher, casting some doubt on the hypothesis that cost difference is the only cause for the difference in sending probability between T0 and treatments T1/T2/T3. The results further indicate that sending the voucher is less likely in T2 and T3 than in T1. In T2, when Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 the sender chooses not to send the voucher, the recipient receives a message saying that the sender had the option to send something but did not. This can be interpreted as shaming the sender (for not sending valuable information) in the hope of increasing information sharing. This attempt appears to backfire: if anything, this treatment reduces sharing. The difference between T1 and T2 is, however, only statistically significant in column (1), as shown at the bottom of table 4. But we do find that, in all columns, sending the voucher is significantly less likely in T3 than in T1. To recall, treatment T3 is when the sender has the opportunity to alert the recipient that they chose not to share the voucher. While this almost never happens, senders may anticipate that information is less likely be trusted (even though there is no evidence of this in table 3) and decide not to incur the cost of sending it. Alternatively, they may find the choices confusing and, perhaps, objectionable and opt not to participate. In any case, this treatment significantly reduces information sharing. In column (2) we see that individuals who have redeemed a voucher in the past or current game are also significantly more likely to send it. The estimated coefficient is largest for those who redeem in the current game. Since subjects only find out whether the promised transfer was deposited in their account at the end of the game, this correlation cannot be driven by having received the voucher. Rather, it suggests either that those who redeem are more attentive to the experiment, or that those who trust our message are more likely to both redeem and share it. We find that sending is less likely in the partial-information variant, but this effect is not statistically significant—unlike what happens with redeeming behavior in table 3. The magnitude of the effect is non- negligible however: a 5 percentage point reduction in information sharing in column (1), compared to the no-information probability of sharing of 30 percent in the baseline T0. This suggests that participants are more willing to share information in a fully anonymous setting. Because redeeming is also lower in the partial-information variant, controlling for past and current redeeming behavior in column (2) absorbs the effect of the partial-information dummy. To investigate the role of anonymity further, we reestimate specification (3) with additional regressors to test for empathy towards similar people. If the reluctance to share information comes from the sender real- izing that the prospective recipient is different from them, the partial-information treatment effect should vanish for subject pairs who have similar characteristics. This is not what we find: differences or similar- ities between sender and receiver are never statistically significant although, as in table 3, point estimates for same gender and same education are large in magnitude. If the reduction in information sharing is not due to a reluctance to share with dissimilar individuals, then it might be due to the sender’s reluctance to have their characteristics revealed to the recipient—i.e., the fear of being recognized. This may be particu- larly problematic if senders are unsure of the value of the message. In any event, subjects seem more willing to share valuable information with complete strangers while remaining fully anonymous themselves. Finally, we note that sending is more common among younger, better educated, and richer participants—consistent with these subjects being more familiar with the mobile phone technology, and being less concerned about the cost of sending a message to benefit others. 4. Incentivizing Information Transfer We have established that information transmission by SMS is imperfect. Apart from demonstrating a sensitivity to the monetary cost of sending the SMS, none of the treatments introduced so far managed to improve information diffusion—either through self-image, social image, or empathy considerations. Since subjects react (negatively) to an increase in the cost of sending, we now seek to incentivize senders for sharing valuable information. The World Bank Economic Review 877 We could try simply paying subjects for sending the voucher. Such intervention, however, is likely to be costly for the policy maker, and subject to manipulation. It is also difficult to decentralize. Instead, our objective is to identify a suitable bargaining mechanism by which the broadcaster of a valuable message can seed a population and then offer a structured bargaining mechanism to encourage peer-to-peer trans- Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 mission among strangers. This would enable the broadcaster of the message to reach a larger audience without the need to provide direct incentives to senders. To this effect, we investigate different forms of decentralized peer-to-peer transfers by which the sender can be rewarded directly by the recipient. We first present the experimental design and sequencing of these new treatments, before discussing the testing strategy and the empirical results obtained with these additional treatments. 4.1. Experimental Design and Sequencing of Treatments T4/T5/T6 It has often been noted that sharing valuable information with others generates a sense of gratefulness, and triggers a desire for the recipient to reciprocate. To capture these ideas in a stylized manner, we introduce treatments that allow the sender to impose, solicit, or receive a payment. We hypothesize that, if these payments are accepted by recipients on the basis of reciprocity, incentivizing senders should improve the dissemination of valuable information. To test this hypothesis, we introduce three additional treatments—labeled T4/T5/T6—and we apply them to the half of the sample that did not take treatments T1/T2/T3. Treatments T4/T5/T6 allow transfers between the sender and receiver of the voucher. To do this in a structured way over anonymous links, we adapt the standard dictator, ultimatum, and reverse dictator games to our setting. To streamline SMS communication, all three treatments have a default option that is implemented if the sender does nothing. We examine whether the type of default option matters. The details are as follows. Treatment T4 (the “dictator game with a default option”) adapts a standard dictator game to our setting. Although it is perhaps the mechanism with the lowest intuitive appeal for information sharing, it is also the easiest to implement in our setting and it offers a well-known benchmark: in general, laboratory subjects allocate around half of the “pie” to the other player (e.g., Camerer 1997; Andreoni and Bernheim 2009). In this treatment, a subject is asked to share a 35 meticais voucher between themselves and one subject in the subsequent row of the square. Each row 1 subject does this four times, once for each subject in row 2. In other words, each subject in row 1 receives 35 meticais four times and can share this amount with one subject from row 2. These decisions are then combined to calculate the total payoff of the sender. If the sender does not respond to one of the four messages, this is treated as equivalent to sending nothing, in which case the sender keeps the 35 meticais. This is different from a standard dictator game where there is no default option and the subject is forced to pick a division of the pie. If the subject does not respond to any of the four messages, they receive 35 × 4 = 140 meticais. The same decision structure is repeated in round 2: the experimenter sends 35 meticais four times to each round 2 subject, and each time the round 2 subject can share part of it with a round 3 subject. The same is again repeated in round 3. Subjects in row 4 do not decide anything; they just receive what row 3 subjects choose to send them. As in the baseline intervention, subjects in rounds 2 to 4 do not receive any message if nothing is sent to them by previous participants. The idea behind this aspect of the design is again to investigate how far information diffuses. In this treatment, the sender is given the opportunity to appropriate the entire value of each voucher. The purpose of this is to determine the extent to which subjects are willing to share something valuable— at their own cost—instead of simply appropriating it. If the subject does nothing, this is treated as not sharing. Furthermore, if the sender does nothing, the recipient is not informed that the sender had an opportunity to share. These differences with the standard dictator game are introduced into our design to capture the fact that, in practice, sharing information requires a deliberate action—doing nothing is the default—and if someone does not share valuable information, potential recipients typically do not 878 Batista, Fafchamps, and Vicente learn about it. Whether T4 induces more or less sharing is unclear a priori. The fact that not sharing is financially attractive may reduce sharing, especially given that it is the default option. But allowing subjects to appropriate part of the voucher also rewards them for sharing the rest of the voucher, which may encourage sharing. Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Treatment T5 (the “ultimatum game with a default option”) adapts an ultimatum game to our frame- work. It lets the sender set a price that the receiver has to pay in order to receive the voucher. If the receiver refuses, the sender receives nothing. In terms of implementation, this treatment is similar to treatment T4: each subject in rounds 1 to 3 is asked four times to share 35 meticais between themselves and one subject in the next row. The difference is that, in this treatment, the designated recipient can refuse the share of the 35 meticais that is proposed by the sender. If the recipient refuses the sender’s offer, both sender and receiver get nothing. Each receiver has to make this decision each time they receive an offer to share 35 meticais. If the sender does not make any offer to a particular recipient—i.e., does nothing—this is treated as a rejection and both subjects receive nothing. This introduces an important difference relative to T4: in order for the recipient to have an opportunity to reject an offer, an offer has to be made. If the recipient does not agree with an offer—or does nothing—this is treated as a rejection by the recipient, and both subjects also receive nothing. Since this treatment is likely to create an entitlement effect in the mind of the sender (e.g., Camerer 2003, Chapter 2), it mimics a market for information in which the seller sets a take-it-or- leave-it price: if the potential buyer refuses the offer, the seller forfeits their profit. This design offers the advantage that it gives the recipient of the information a veto: if the recipient does not believe or value the information provided, there is no reason to accept the offer. Treatment T6 (the “reverse dictator game with a default option”) is similar to T4 except that it is the recipient who unilaterally decides how much to send back to the sender. This treatment mimics a “pay- what-you-want” approach, used for instance by shareware providers and, offline, used by many religious and philanthropic organizations. It has not really caught on as a method for selling non-rival content, however, probably because sellers can make a higher profit from direct sales (i.e., treatment T5).21 But it still provides some incentive to the sender and may prove more egalitarian (e.g., Gneezy et al. 2010). In this treatment, round 1 is exactly the same as in the baseline intervention: subjects choose whether to redeem the voucher and whether to send vouchers to each row 2 subject. Subjects in round 4 only decide how much to send back. Subjects in rounds 2 and 3 first decide how much to send back to the sender from the previous row, and then whether to send a voucher to each of the receivers in the subsequent row. Unlike in the baseline intervention, subjects do not have to respond “yes/sim” to the SMS voucher in order to receive it—they are only asked to determine how much they wish to send back. If a subject does not respond, they are assumed to send back nothing—which is the mirror image to the sender’s decision in T4: doing nothing is equivalent to appropriating the whole voucher. As in the baseline intervention, a subject in rows 2 to 4 only participates if at least one subject from the previous row decided to send them a voucher. Importantly, T6 is not entirely equivalent to a standard reverse dictator game in the sense that the receiver knows that the voucher was sent by the sender. We hypothesize that this distinction may create a reciprocity effect. Each treatment is played on a square—i.e., group of 16 subjects—as for the baseline intervention. We have already noted that 6 of the 12 squares that played the baseline intervention were randomly as- signed to treatments T1, T2, and T3 for the subsequent three games. The other 6 are similarly assigned to play treatments T4, T5, and T6, in random order, over three games. These 6 squares are further di- vided into two groups of 3: one is always assigned to the no-information variant and the other to the 21 In 2007, rock band Radiohead famously let people download their In Rainbows album for free, inviting them to “pay what you want.” It is believed that 1.2 million people downloaded the free album but gave little in return. The band never used this approach again, although In Rainbows remains one of their best selling albums in terms of CD sales. The World Bank Economic Review 879 partial-information variant. The assignment structure of treatments to squares is depicted in the bot- tom panel of fig. 3, where Ti stands for treatment i and N/P stands for the no-information and partial- information variants, respectively. Table 1 compares the two halves of our sample, namely those playing treatments T1/T2/T3 and those Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 playing treatments T4/T5/T6. Within each of the two halves of the sample, balance across treatments is achieved by experimental design. All the tests that we performed fail to reject the null hypothesis of no difference for each of the observable characteristics. From this we conclude that randomization achieved balance on key individual characteristics across squares and treatment blocks. 4.2. Experimental Results on Treatments T4/T5/T6 4.2.1. Treatment Averages In treatments T4/T5/T6 the primary emphasis is on sending decisions. Recall that in T4 and T5 senders decide an amount to be sent. In T6 they decide whether to send the voucher or not. In T4, receivers do nothing. In T5, receivers can either accept or reject the take-it-or-leave-it offer. In T6, receivers decide whether to redeem a voucher from the experimenter in round 1 and then whether and how much to send back to the sender. We report the average behavior of the subjects on all these choices in table 5. Note that some actions are not relevant in some treatments, e.g., receiving is automatic in T4, and sending back is an action only possible in T6. Table 5. Average Choices Made by Subjects in Games T0/T4/T5/T6 Baseline Treatment 4: Treatment 5: intervention dictator game ultimatum game Treatment 6: reverse dictator (1) (2) (3) (4) (5) Sending the voucher Sender sent Sender sent Sender sent Sender sent Receiver sent back All subjects 24.2% 14.8% 17.9% 24.2% 11.8% (0.022) (0.033) (0.036) (0.029) (0.046) Partial information 21.7% 10.9% 15.9% 19.4% 9.5% (0.029) (0.042) (0.044) (0.038) (0.066) No information 26.8% 18.3% 20.8% 28.8% 13.3% (0.032) (0.050) (0.059) (0.043) (0.063) Share sent — 3.9% 4.3% — 11.6% Share sent conditional on sending — 26.5% 23.9% — 98.6% Number of observations 392 115 117 219 51 Redeeming/accepting the voucher Receiver redeemed Receiver accepted Sender redeemed All subjects 25.9% — 57.1% 37.5% — (0.037) (0.202) (0.101) Round 1 only (1) 27.1% — n.a. 37.5% — (0.065) n.a. (0.101) Rounds 2–4 (2) 25.3% — 57.1% n.a. — (0.045) (0.202) n.a. Number of observations 143 — 7 24 — Source: Authors’ calculations based on the experimental data collected by the authors. Note: Except for the “Share sent” and the “Share sent conditional on sending,” each number is the proportion of decisions for which the relevant decision was made, e.g., sending something. Standard errors are reported in parentheses. In treatment T4, senders can send up to 35 meticais to receivers. “Sender sent” is the proportion of senders sending positive amounts. The “share sent” is the average amount sent divided by 35, the value of the voucher. Receiving is automatic in this game. Treatment T5 is analogous, except that receivers decide whether to accept offers sent by senders. “Receiver accepted” is the proportion of accepted take-it-or-leave-it offers. In treatment T6, senders in round 1 have the choice of redeeming the voucher sent by the experimenter by responding with a “yes” SMS to our switchboard. “Sender redeemed” shows the proportion of senders doing so. In this treatment senders can send vouchers to receivers like in the baseline intervention: “sender sent” is the proportion of vouchers sent. Receiving after round 1 is automatic. Receivers can then send back to senders up to the full amount of the voucher received (35 meticais). “Receiver sent back” is the proportion of receivers sending back positive amounts. The “share sent” is the average amount sent back divided by 35, the value of the voucher. (1) In round 1 the voucher SMS is sent at the initiative of the experimenter. (2) In rounds 2–4 the voucher SMS is sent at the request of another subject. 880 Batista, Fafchamps, and Vicente In T4 the sender appropriates the full value of the voucher by doing nothing. We see that introducing this possibility leads to a fall in the propensity to send something to the receiver: from 24 percent in the baseline intervention to 15 percent in T4. This difference is statistically significant. It suggests that when senders cannot appropriate the voucher, they are willing to spend some of their own money to benefit Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 someone else, and when they can appropriate the voucher, many prefer doing so instead of sharing even a fraction of it. We also note that, even when they send something, subjects only give 27 percent of the average voucher value. Across all subjects and decisions, senders retain more than 96 percent of the voucher value. This suggests that adding the possibility of appropriating the value of the information crowds out altruistic motives, and that most subjects choose to do nothing when it is to their material advantage. In T5, sharing the value of information entails the risk of rejection: the receiver may refuse the offer made—something that occurs in 43 percent of the cases. We observe an overall 18 percent probability of sending money to the receiver, lower than in T0 and only slightly higher than T4. This is a priori surprising because, in T4, the sender appropriates everything if no offer is made, while in T5 the sender receives something only if making an offer. This suggests that subjects are reluctant to make an offer that can be rejected. We also note that the amount sent does not increase relative to T4, which may explain why many offers are rejected. This evidence indicates that introducing squabbling among subjects over how to share the value of information is detrimental to information diffusion. In T6, the sender can only elect to send or not the full voucher value to the receiver, as in the baseline intervention. We find that the probability of sending in T6 is identical to that in T0. This suggests that the prospect of receiving something back from the receiver does not incentivize senders to send more. In 12 percent of the cases, the receiver elects to send something back, i.e., at a rate that is broadly similar to what senders do in T4. But when they do, they send back a much higher proportion of the voucher value—typically almost all of it, suggesting, among these subjects, a reciprocity motive. Senders in round 1 are also given the choice to redeem or not the voucher sent by the experimenter. In fact, 38 percent of subjects do so. Finally we note that, as in table 2, the no-information variants of the treatments T4 to T6 cause higher sending rates. 4.2.2. Transfers We now estimate a model of the decision to transfer any amount, i.e., employing as a dependent variable a binary variable taking value 1 if the sender sends a positive amount to the recipient, and 0 otherwise. For the decision to send or send back money in treatments T4/T5/T6, we estimate the following specification: Si jrt = α + β5 G5 6 6b i jrt + β6 Gi jrt + β6b Gi jrt + γ Ii + δr + εi jrt , (4) where the treatment dummy G superscript 6 refers to the decision to send in treatment T6 while 6b refers to the decision to send back in that same treatment. The specification is similar to (3), except that we do not include the cost of sending since it is constant. We also estimate a specification that adds absolute difference terms |Xi − Xj | and controls Xi , again to test for the role of shared characteristics. These econometric specifications are estimated using linear probability models and, as before, reported standard errors are clustered across games at the individual level. The amount sent is examined in a separate regression. Results for the decision whether to transfer are shown in table 6. Column (1) follows specification (4); column (2) adds pairwise characteristics. Note that treatment T6 has two sending decisions, one made by the sender and another one made by the receiver. From table 5, we already know that sending is on average less frequent in T4 and T5 than in the baseline intervention. The exception is T6, where the likelihood of sending money is higher. By comparing point estimates for T4 and T6, we see that the difference between them is large in magnitude: 16 to 18 percentage points. This makes sense: of the four sending actions taken in treatments T4/T5/T6, sharing by the sender in T6 is the one that is most similar The World Bank Economic Review 881 Table 6. The Decision to Send in Treatments T4/T5/T6 (1) (2) Treatment dummies (T4 is omitted category) Treatment T5 dummy (ultimatum) 0.051 0.050 Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (0.041) (0.045) Treatment T6 dummy (reverse dictator sender) 0.159*** 0.180*** (0.058) (0.061) Treatment T6 dummy (reverse dictator receiver) 0.072 0.094 (0.070) (0.068) Partial information dummy −0.049 0.026 (0.075) (0.110) Pairwise differences in individual characteristics times partial-information dummy Same gender dummy — 0.113 (0.073) Absolute difference in age — 0.002 (0.004) Absolute difference in income (in ’000 meticais/month) — −0.015 (0.013) Round dummies Yes Yes Period dummies Yes Yes Treatment sequence dummies Yes Yes Individual characteristics No Yes Pairwise differences in individual characteristics (uninteracted) No Yes Intercept 0.216* 0.517*** (0.114) (0.169) R-squared 0.062 0.147 Number of observations 502 465 Joint coefficient tests Test that T5 (β 5 ) = T6 sender (β 6 ) p-value 0.073 0.033 Test that T5 (β 5 ) = T6 receiver (β 6b ) p-value 0.776 0.524 Test that T6 sender (β 6 ) = T6 receiver (β 6b ) p-value 0.095 0.091 Source: Authors’ calculations based on the experimental data collected by the authors. Note: All regressions are estimated by ordinary least squares (OLS). The dependent variable is a binary variable defined as 1 if, when given the chance, the subject sends an SMS sharing the voucher with another subject. In column 2 we also include as controls the pairwise differences in individual characteristics uninteracted with the non-anonymous dummy. The absolute difference in education level is omitted due to multicollinearity. Individual characteristics include a female dummy, age, a post-secondary education dummy, and income in meticais/month. Standard errors, clustered at the level of the individual, are reported in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. to sending in T0. The fact that propensities to send are similar in both cases indicates that giving the sender an opportunity to receive something in return does not, by itself, increase willingness to send. In contrast, in T4, not sending anything lets the sender appropriate the full value of the voucher. This likely explains the significant difference between the two treatments. Treatment T5 is similar to T6 regarding senders’ decisions: not sending anything means forfeiting the voucher. We should thus observe a similar propensity to send in both T5 and T6. This is however not what we observe: the frequency of sending in T5 is similar to T4, where the sender appropriates the voucher by not sending anything, and lower than in T6 (sender’s decision). This suggests that subjects prefer sending the information and letting the recipient decide whether to send something back, rather than making a take-it-or-leave-it offer to the recipient and risking rejection: indeed, 43 percent of offers are rejected in T5. It follows that the fear of rejection seems to serve as a disincentive to share. We also observe that the probability of sending back in T6 is not statistically different from sending in T4: sender and receiver are equally likely to appropriate everything. This arises even though, in T6, the 882 Batista, Fafchamps, and Vicente recipient knows that the sender is aware that the recipient could send something back while, in T4, the potential recipient is not aware that the sender could have sent anything. This suggests the absence of a reciprocity motive, at least in terms of sending anything at all as we discuss further below. We also note that in both T4 (sender) and T6 (receiver) the probability of sending is lower than what is often observed Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 in dictator games.22 This difference may be due to the fact that, in both cases, appropriating everything can be achieved by picking the default option, which is doing nothing. This exonerates subjects from the moral pressure that is present in a standard dictator or reverse dictator game, where there is no default option. Column (1) also shows that the likelihood of sending in the no-information variant is 5 percentage points higher than in the partial-information variant (albeit not significant). Turning to column (2), we again find no statistical evidence that shared characteristics affect sending behavior—even if the point estimate on same gender is a large 11 percentage points. These results are similar to those we reported in table 5. Taken together, this evidence confirms subjects’ reluctance to share information in the partial- information setting. Regarding other coefficient estimates (not reported in the table to save space), we find that subjects who are male, young, educated, and poorer are more likely to send something. Table 6 focused on the effect of treatment on the extensive margin—the likelihood of sending some- thing. We complement these results by showing in table 7 the effect of treatment on the intensive mar- gin. To this effect, we present a regression of the amount sent (conditional on sending) as a function of treatment. Given the small number of non-missing observations, we only include treatment dummies as regressors. The results show that, conditional on giving, the amount given is far larger for subjects who send something back in T6, suggestive of a reciprocity motive among the 12 percent of subjects who choose to send anything back. Table 7. Amount Sent in Treatments T4/T5/T6, Conditional on Sending (1) Treatment variables (T4 is omitted category) Treatment 5 dummy (ultimatum) −0.911 (4.448) Treatment 6 dummy (reverse dictator receiver) 25.214*** (4.577) Intercept 9.286 (4.557) R-squared 0.646 Number of observations 44 Joint coefficient tests Test that game 5 (β 5 ) = game 6 receiver (β 6b ) p-value 0.000 Source: Authors’ calculations based on the experimental data collected by the authors. Note: All regressions are estimated by ordinary least squares (OLS). The dependent variable is the amount sent to another subject in meticais, conditional on an amount being sent. This decision is only relevant in T4 (sender), T5 (sender), and T6 (receiver). Due to the small number of observations, other regressors are not included. Standard errors, clustered at the level of the individual, are reported in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. 22 Batista et al. (2015) conducted a lab-in-the-field experiment among urban residents in Maputo, the capital city of Mozambique. In this experiment the average fraction of cash shared was 40 percent. The counterparts receiving the dictator’s transfer were close relations from outside the own household, and hence not anonymous. The World Bank Economic Review 883 5. Robustness Before concluding, we investigate the robustness of our findings to the possibility that some subjects simply ignore all the messages originating from the experiment. This may still arise in spite of our efforts to the contrary: all the subjects are familiar with the research team, having participated in an earlier randomized Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 controlled trial by the same researchers; we selected subjects who were already familiar and actively using text messages and mobile money; and we secured explicit informed consent from all the subjects shortly before the experiment began. We start by noting that 31 percent of the subjects assigned to rounds 2–3–4 were never sent any voucher by subjects in earlier rounds. As a result, they never had the opportunity to redeem or send vouchers to other subjects. These subjects have already been omitted from the analysis. Of the remaining participants, 55 percent never actively participated in the experiment either by accepting a voucher or by sending a message to another subject. Our concern is that some of these subjects may have failed to participate for lack of understanding or for reasons beyond their control—e.g., they permanently lost access to the phone number that was used to contact them. We wish to ensure that our findings—e.g., low redeeming of vouchers—are not mechanically driven by their non-activity. To this effect, we repeat the analysis of tables 3, 4, and 6 using only subjects who responded to at least one of our messages. We focus on the main specifications of the previous tables, i.e., with a full list of controls, and with previous redeeming behavior when considering treatments T0/T1/T2/T3. We omit the specifications with shared characteristics since they are never significant. Results are shown in table 8 for treatments T0/T1/T2/T3 and in table 9 for treatments T4/T5/T6. Not surprisingly, estimated treatment effects are larger in magnitude—given that inactive subjects are omitted. But otherwise the findings are qualitatively similar to those reported in tables 3, 4, and 6. In particular, the roles of anonymity and previous redeeming have the same sign. There are some small differences however. We now find that sending back in T6 is significantly more likely than in T4 (see table 9), consistent with reciprocity on the part of receivers in that treatment. We also find that sending in T2 is significantly lower than in T1 (see table 8) and that high-income subjects are less likely to send information to others across all treatments. 6. Concluding Remarks In this paper we followed a sample of rural Mozambicans with access to mobile money services. We investigated (a) their willingness to act upon valuable information they receive and (b) their willingness to share this valuable information with others. To this effect, we randomly assigned to subjects four other participants to whom they could send a voucher, and we tested a number of experimental settings implemented through SMS messages containing vouchers redeemable for mobile money. By assigning links exogenously, we avoid endogeneity issues that arise in experiments on information sharing that rely on preexisting social links that are context specific. We find that subjects have a relatively low propensity to redeem the voucher, but a comparatively high propensity to send it to others. This pattern of behavior is hard to reconcile with a lack of attention or a poor understanding of our messages. A more convincing explanation is that people are skeptical about the value of the message they receive, but this does not stop them from incurring a small cost to share it with others. Many subjects indeed share information that they do not use themselves, a behavior that can be interpreted as consistent with a warm glow motive. We nonetheless observe that both redeeming and sending are higher among subjects who previously redeemed the voucher. This behavior is consistent with the idea that subjects are more likely to share information if they find it trustworthy. Contrary to expectations, anonymity increases both receiving and sending, and there is no evidence that shared characteristics increase sharing. Why this is the case is unclear. One possibility is that senders are unsure of the value of the message and may worry others may think poorly of them for passing it on. 884 Batista, Fafchamps, and Vicente Table 8. The Decisions to Redeem and Send in Treatments T0/T1/T2/T3—Omitting Inactive Subjects Redeem Send (1) (2) Treatment dummies (T0 is omitted category) Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Treatment 1 dummy (variable cost) −0.484** −0.367*** (0.231) (0.135) Treatment 2 dummy (shaming and fixed cost of sending) −0.466** −0.485*** (0.185) (0.135) Treatment 3 dummy (erroneous message and fixed cost of sending) −0.287 −0.568*** (0.254) (0.127) Partial information dummy −0.255 0.106 (0.174) (0.087) Additional cost of sending the voucher 0.004 (0.008) Dummy=1 if subject redeemed a voucher in the current period 0.300*** (0.078) Dummy=1 if subject redeemed a voucher in a previous period 0.091 0.267*** (0.147) (0.087) Round dummies Yes Yes Period dummies Yes Yes Treatment sequence dummies Yes Yes Individual characteristics Yes Yes Intercept 1.035*** 0.437** (0.362) (0.210) R-squared 0.125 0.341 Number of observations 107 337 Joint coefficient tests Test that T1 (β 1 ) = T2 (β 2 ) p-value 0.899 0.187 Test that T1 (β 1 ) = T3 (β 3 ) p-value 0.297 0.040 Test that T2 (β 2 ) = T3 (β 3 ) p-value 0.197 0.338 Source: Authors’ calculations based on the experimental data collected by the authors. Note: All regressions are estimated by ordinary least squares (OLS). In “Redeem,” the dependent variable is a binary variable defined as 1 if, when given the chance, the subject sends an SMS accepting the voucher. In “Send,” the dependent variable is a binary variable defined as 1 if, when given the chance, the subject sends an SMS giving the voucher to another subject. In treatment 3, sending the false message (only 2 observations) is assimilated to not sending the voucher. The additional cost of sending the voucher is 0 in the baseline intervention, 5 meticais in treatments 2 and 3, and varying between 0/5/10/15 meticais in treatment 1. Individual characteristics include gender, age, a post-secondary education dummy, and monthly income. Standard errors, clustered at the level of the individual, are reported in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. In terms of behavioral variation between treatments, we find that the sharing of information falls when we introduce an explicit cost of sharing—but we do not find that subjects respond to variation in that cost. We find no evidence that shaming helps information transfer: sharing falls when we reveal that senders send nothing, and subjects do not like to reveal that they sent nothing. We also observe less sharing in treatments that allow subjects to appropriate the value of the shared information—irrespective of the system put in place to allow transfers between subjects. Allowing information recipients to send anything back to the sender achieves just the same amount of information diffusion as the baseline intervention without this option. Taken together, these findings indicate that sharing information is not motivated by the hope of reciprocation—at least in our setting. In terms of policy, this research reveals the difficulty of using mobile phone messages to diffuse valu- able information in a developing country. Even when participants have been sensitized beforehand and a substantial amount of money is at stake, many individuals fail to make use of the valuable information they receive. Our take-home lessons for policy makers are that you can reach a lot of people cheaply via SMS, but do not think of it as a perfect substitute for other forms of information dissemination. When The World Bank Economic Review 885 Table 9. The Decision to Send in Treatments T4/T5/T6—Omitting Inactive Subjects Send any amount Treatment dummies (T4 is omitted category) Treatment 5 dummy (ultimatum) 0.087 Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (0.069) Treatment 6 dummy (reverse dictator sender) 0.350*** (0.088) Treatment 6 dummy (reverse dictator receiver) 0.277* (0.140) Partial information dummy 0.072 (0.137) Round dummies Yes Period dummies Yes Treatment sequence dummies Yes Individual characteristics Yes Intercept 0.431 (0.281) R-squared 0.151 Number of observations 245 Joint coefficient tests Test that T5 (β 5 ) = T6 sender (β 6 ) p-value 0.007 Test that T5 (β 5 ) = T6 receiver (β 6b ) p-value 0.195 Test that T6 sender (β 6 ) = T6 receiver (β 6b ) p-value 0.577 Source: Authors’ calculations based on the experimental data collected by the authors. Note: All regressions are estimated by ordinary least squares (OLS). The dependent variable is a binary variable defined as 1 if, when given the chance, the subject sends an SMS sharing the voucher with another subject. Individual characteristics include gender, age, a post-secondary education dummy, and monthly income. Clustered standard errors, at the level of the individual, reported in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. using SMS communication, think twice about doing it in a personalized manner, do not attempt to shame participants into sharing with others, and do not spend energy trying to reward information sharing. Keep it simple. References Abebe, G., B. Tekle, and Y. Mano. 2018. “Changing Saving and Investment Behavior: The Impact of Financial Literacy Training and Reminders on Micro-Businesses.” Journal of African Economies, 27 (5): 1–25. doi:10.1093/jae/ejy007. Afzal, U., G. d’Adda, M. Fafchamps, S. Quinn, and F. Said. 2018. “Two Sides of the Same Rupee? Comparing Demand for Microcredit and Microsaving in a Framed Field Experiment in Rural Pakistan.” Economic Journal, 128 (614): 2161–90. Afzal, U., G. d’Adda, M. Fafchamps, S. Quinn, and F. Said. 2022. “Demand for Commitment in Credit and Saving Contracts: A Field Experiment.” Oxford University, June (mimeo). Aker, J. C., P. Collier, and P. Vicente. 2017. “Is Information Power? Using Mobile Phones and Free Newspapers during an Election in Mozambique.” Review of Economics and Statistics 99(2): 185–200. Akerlof, G., and R. Kranton. 2000. “Economics and Identity.” Quarterly Journal of Economics 115(3): 715–53. Akerlof, G. A., and R. E. Kranton. 2005. “Identity and the Economics of Organizations.” Journal of Economic Per- spectives 19(1): 9–32. Alatas, V., A. G. Chandrasekhar, M. Mobius, B. A. Olken, and C. Paladines. 2022. “Designing Effective Celebrity Mes- saging: Results from a Nationwide Twitter Experiment Promoting Vaccination in Indonesia.” Stanford University, February (mimeo). Allcott, H., and M. Gentzkow. 2017. “Social Media and Fake News in the 2016 Election.” Journal of Economic Perspectives 31(2): 211–36. 886 Batista, Fafchamps, and Vicente Alsan, M., A. Banerjee, E. Breza, A. G. Chandrasekhar, E. Duflo, P. Goldsmith-Pinkham, and B. A. Olken. 2020. “Mes- sages on COVID-19 Prevention in India Increased Symptoms Reporting and Adherence to Preventive Behaviors among 25 Million Recipients with Similar Effects on Non-recipient Members of Their Communities.” Presentation at NBER SI 2020 Development Economics, Stanford University (mimeo). Andreoni, J. 1990. “Impure Altruism and Donations to Public Goods: A Theory of Warm-Glow Giving.” Economic Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Journal 100(401): 464–77. Andreoni, J., and B. D. Bernheim. 2009. “Social Image and the 50–50 Norm: A Theoretical and Experimental Analysis of Audience Effects.” Econometrica 77(5): 1607–36. Bandiera, O., R. Burgess, E. Deserranno, R. Morel, I. Rasul, and M. Sulaiman. 2022. “Social Incentives, Delivery Agents and the Effectiveness of Development Interventions.” Journal of Political Economy (forthcoming). Bandiera, O., and I. Rasul. 2006. “Social Networks and Technology Adoption in Northern Mozambique.” Economic Journal 116(514): 869–902. Banerjee, A., E. Breza, A. G. Chandrasekhar, E. Duflo, M. O. Jackson, and C. Kinnan. 2021a. “Changes in Social Net- work Structure in Response to Exposure to Formal Credit Markets.” Review of Economic Studies (forthcoming). Banerjee, A., A. G. Chandrasekhar, S. Dalpath, E. Duflo, J. Floretta, M. O. Jackson, and H. Kannan et al., 2021b. “Selecting the Most Effective Nudge: Evidence from a Large-Scale Experiment on Immunization.” NBER Working Paper 28726. Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. O. Jackson. 2013. “The Diffusion of Microfinance.” Science 341 (6144), 1236498-1–7. Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. O. Jackson. 2019. “Using Gossips to Spread Information: Theory and Evidence from Two Randomized Controlled Trials. ” Review of Economic Studies 86(6), 2453–90. Batista, C., D. Silverman, and D. Yang. 2015. “Directed Giving: Evidence from an Inter-household Transfer Experi- ment.” Journal of Economic Behavior and Organization 118(C): 2–21. Batista, C., and P. C. Vicente. 2013. “Introducing Mobile Money in Rural Mozambique: Evidence from a Field Exper- iment.” Universidade Nova de Lisboa, mimeograph. Batista, C., and P. C. Vicente. 2020. “Adopting Mobile Money: Evidence from an Experiment in Rural Africa.” AEA Papers and Proceedings 110: 594–98. Batista, C., and P. C. Vicente. 2022. “Is Mobile Money Changing Rural Africa? Evidence from a Field Experiment.” Universidade Nova de Lisboa, mimeograph. Bauer, M., N. Fiala, and I. Levely. 2018. “Trusting Former Rebels: An Experimental Approach to Understanding Reintegration after Civil War.” Economic Journal 128(613): 1786–1819. Beaman, L., A. BenYishay, J. Magruder, and A. M. Mobarak. 2021. “Can Network Theory-Based Targeting Increase Technology Adoption?” American Economic Review 111(6): 1918–43. Beaman, L., and J. Magruder. 2012. “Who Gets the Job Referral? Evidence from a Social Networks Experiment.” American Economic Review 102(7): 3574–93. Berg, E., M. Ghatak, R. Manjula, D. Rajasekhar, and S. Roy. 2019. “Motivating Knowledge Agents: Can Incentive Pay Overcome Social Distance?” Economic Journal 129(617): 110–42. Bloch, F., G. Genicot, and D. Ray. 2008. “Informal Insurance in Social Networks.” Journal of Economic Theory 143(1): 36–58. Blumenstock, J., M. Callen, and T. Ghani. 2016. “Why Do Defaults Affect Behavior? Experimental Evidence from Afghanistan.” American Economic Review 108(10): 2868–901. Bursztyn, L., and R. Jensen. 2017. “Social Image and Economic Behavior in the Field: Identifying, Understanding, and Shaping Social Pressure.” Annual Review of Economics 9(1): 131–53. Cai, J., and A. Szeidl. 2018. “Interfirm Relationships and Business Performance.” Quarterly Journal of Economics 133 (3): 1229–82. Camerer, C. F. 1997. “Progress in Behavioral Game Theory.” Journal of Economic Perspectives 11(4): 167–88. Camerer, C. F. 2003. Behavioral Game Theory: Experiments in Strategic Interaction, Princeton University Press, Princeton. The World Bank Economic Review 887 Carter, M. R., R. Laajaj, and D. Yang. 2021. “Subsidies and the African Green Revolution: Direct Effects and So- cial Network Spillovers of Randomized Input Subsidies in Mozambique.” American Economic Journal: Applied Economics 13(2): 203–29. Centola, D. 2010. “The Spread of Behavior in an Online Social Network Experiment.” Science 329(5996): 1194–97. Chandrasekhar, A. G., H. Larreguy, and J. P. Xandri. 2020. “Testing Models of Social Learning on Networks: Evidence Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 from Two Experiments.” Econometrica 88(1): 1–32. Cole, S., and N. Fernando. 2021. “Mobile’izing Agricultural Advice: Technology Adoption, Diffusion and Sustain- ability.” Economic Journal 131(633): 192–219. Comola, M., and S. Prina. 2021. “Treatment Effects Accounting for Network Changes.” Review of Economics and Statistics 103(3): 1–8. Deck, C., and S. Jahedi. 2015. “The Effect of Cognitive Load on Economic Decision Making: A Survey and New Experiments.” European Economic Review 78: 97–119. DellaVigna, S., J. A. List, and U. Malmendier. 2012. “Testing for Altruism and Social Pressure in Charitable Giving.” Quarterly Journal of Economics 127(1): 1–56. Drexler, A., G. Fischer, and A. Schoar. 2014. “Keeping It Simple: Financial Literacy and Rules of Thumb.” American Economic Journal: Applied Economics, 6.2 (2014): 1–31. Drichoutis, A. C., and R. M. Nayga Jr. 2020. “Economic Rationality under Cognitive Load.” Economic Journal 130(632): 2382–409. Fafchamps, M., A. Islam, A. Malek, and D. Pakrashi. 2020a. “Can Referral Improve Targeting? Evidence from a Vocational Training Experiment.” Journal of Development Economics 144(C), 10.1016/j.jdeveco.2019.102436. Fafchamps, M., and B. Minten. 2012. “Impact of SMS-Based Agricultural Information on Indian Farmers.” World Bank Economic Review 26(3): 383–414. Fafchamps, M., and S. Quinn. 2018. “Networks and Manufacturing Firms in Africa: Results from a Randomized Field Experiment.” World Bank Economic Review 32 (3): 656–75. Fafchamps, M., M. Soderbom, and Boogaart M. vanden. 2022. “Adoption with Social Learning and Network Exter- nalities.” Oxford Bulletin of Economics and Statistics (forthcoming). Fafchamps, M., A. Vaz, and P. C. Vicente. 2020b. “Voting and Peer Effects: Experimental Evidence from Mozambique.” Economic Development and Cultural Change 68 (2): 567–606. Fafchamps, M., and P. C. Vicente. 2013. “Political Violence and Social Networks: Experimental Evidence from a Nigerian Election.” Journal of Development Economics 101: 27–48. Foster, A., and M. Rosenzweig. 1995. “Learning by Doing and Learning from Others: Human Capital and Technical Change in Agriculture.” Journal of Political Economy 103(6): 1176–209. Garlick, R., K. Orkin, and S. Quinn. 2020. “Call Me Maybe: Experimental Evidence on Using Mobile Phones to Survey African Microenterprises.” World Bank Economic Review 34 (2): 418–43. Gneezy, A., U. Gneezy, L. D. Nelson, and A. Brown. 2010. “Shared Social Responsibility: A Field Experiment in Pay-What-You-Want Pricing and Charitable Giving.” Science 329(5989): 325–27. Golub, B., and M. O. Jackson. 2010. “Naïve Learning in Social Networks and the Wisdom of Crowds.” American Economic Journal: Microeconomics 2(1): 112–49. Granovetter, M. 1974. Getting a Job: A Study of Contacts and Careers, University of Chicago Press, Chicago. Greif, A. 1993. “Contract Enforceability and Economic Institutions in Early Trade: The Maghribi Traders’ Coalition.” American Economic Review 83(3): 525–48. Griliches, Z. 1957. “Hybrid Corn: An Exploration in the Economics of Technological Change.” Econometrica 25(4): 501–22. Hjort, J., D. Moreira, G. Rao, and J. F. Santini. 2021. “How Research Affects Policy: Experimental Evidence from 2,150 Brazilian Municipalities.” Americal Economic Review 111(5): 1442–80. J-PAL (Abdul Latif Jameel Poverty Action Lab). . “Improving Learning Outcomes through Providing Information to Students and Parents.” J-PAL Policy Insights. Last modified July 2020. https://doi.org/10.31485/pi.2756.2020. Jackson, M., and A. Wolinsky. 1996. “A Strategic Model of Social and Economic Networks.” Journal of Economic Theory, 71 (1): 44–74. Jackson, M. O. 2010. Social and Economic Networks, Princeton University Press, Princeton. Jackson, M. O., T. Rodriguez-Barraquer, and X. Tan. 2012. “Social Capital and Social Quilts: Network Patterns of Favor Exchange.” American Economic Review 102(5): 1857–97. 888 Batista, Fafchamps, and Vicente Johansson-Stenman, O., and H. Svedsäter. 2012. “Self-Image and Valuation of Moral Goods: Stated versus Actual Willingness to Pay.” Journal of Economic Behavior & Organization 84(3): 879–91. Kandori, M. 1992. “Social Norms and Community Enforcement.” Review of Economic Studies 59(1): 63–80. Karlan, D., M. McConnell, S. Mullainathan, and J. Zinman. 2016a. “Getting to the Top of Mind: How Reminders Increase Saving.” Management Science 62 (12): 3393–411. Downloaded from https://academic.oup.com/wber/article/36/4/857/6711586 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Karlan, D., M. Morten, and J. Zinman. 2016b. “A Personal Touch in Text Messaging Can Improve Microloan Repay- ment.” Behavioral Science and Policy 1 (2): 25–31. Kelley, E., J. Magruder, and C. Ksoll. 2022. “How Do Online Job Portals Affect Employment and Job Outcomes? Evidence from India.” UC Berkeley (mimeo). Kirman, A., and M. Teschl. 2010. “Selfish or Selfless? The Role of Empathy in Economics.” Philosophical Transactions of the Royal Society of London. Series B, Biological sciences 365(1538): 303–17. McKenzie, D. 2021. “Small Business Training to Improve Management Practices in Developing Countries: Re- assessing the Evidence for ‘Training Doesn’t Work.” Oxford Review of Economic Policy 37(2): 276–301. Mobius, M., T. Phan, and A. Szeidl. 2015. “Treasure Hunt: Social Learning in the Field.” NBER, Working Paper No. 21014. Naeher, D., and M. Schündeln. 2021. “The Demand for Advice: Theory and Empirical Evidence from Farmers in Sub-Saharan Africa.” World Bank Economic Review 36(1): 91–113. Obermayer, J. L., W. T. Riley, O. Asif, and J. Jean-Mary. 2004. “College Smoking-Cessation Using Cell Phone Text Messaging.” Journal of American College Health 53(2): 71–8. Okyere, C. Y., E. H. Pangaribowo, and N. Gerber. 2019. “Household Water Quality Testing and Information: Identify- ing Impacts on Health Outcomes and Sanitation- and Hygiene-Related Risk-Mitigating Behaviors.” 43(6): 370–95. Patrick, K., F. Raab, M. Adams, L. Dillon, M. Zabinski, C. Rock, W. Griswold, and G. Norman. 2009. “A Text Message-Based Intervention for Weight Loss: Randomized Controlled Trial.” Journal of Medical Internet Research 11(1): e1. Raifman, J. R. G., H. E. Lanthorn, S. Rokicki, and G. Fink. 2014. “The Impact of Text Message Reminders on Adher- ence to Antimalarial Treatment in Northern Ghana: A Randomized Trial.” PloS ONE 9, 10: e109032. Ryan, B., and N. C. Gross. 1943. “The Diffusion of Hybrid Seed Corn in Two Iowa Communities.” Rural Sociology 8: 15–24. Tirole, J. 2002. “Rational Irrationality: Some Economics of Self-Management.” European Economic Review 46(4-5): 633–55. Vega-Redondo, F. 2007. Complex Social Networks, Cambridge University Press, Cambridge. Vilela, I. 2019. “Diffusion of Rival Information in the Field.” Universidade Nova de Lisboa, mimeograph.