IFPRI Discussion Paper 01312 December 2013 Don’t Tell on Me Experimental Evidence of Asymmetric Information in Transnational Households Kate Ambler Markets, Trade and Institutions Division INTERNATIONAL FOOD POLICY RESEARCH INSTITUTE The International Food Policy Research Institute (IFPRI), established in 1975, provides evidence-based policy solutions to sustainably end hunger and malnutrition and reduce poverty. The Institute conducts research, communicates results, optimizes partnerships, and builds capacity to ensure sustainable food production, promote healthy food systems, improve markets and trade, transform agriculture, build resilience, and strengthen institutions and governance. Gender is considered in all of the Institute’s work. IFPRI collaborates with partners around the world, including development implementers, public institutions, the private sector, and farmers’ organizations, to ensure that local, national, regional, and global food policies are based on evidence. IFPRI is a member of the CGIAR Consortium. AUTHORS Kate Ambler (k.ambler@cgiar.org) is a postdoctoral fellow in the Poverty, Health, and Nutrition Division of the International Food Policy Research Institute (IFPRI), Washington, DC. Notices 1. IFPRI Discussion Papers contain preliminary material and research results. They have been peer reviewed, but have not been subject to a formal external review via IFPRI’s Publications Review Committee. They are circulated in order to stimulate discussion and critical comment; any opinions expressed are those of the author(s) and do not necessarily reflect the policies or opinions of IFPRI. 2 The boundaries and names shown and the designations used on the map(s) herein do not imply official endorsement or acceptance by the International Food Policy Research Institute (IFPRI) or its partners and contributors. Copyright 2013 International Food Policy Research Institute. All rights reserved. Sections of this material may be reproduced for personal and not-for-profit use without the express written permission of but with acknowledgment to IFPRI. To reproduce the material contained herein for profit or commercial use requires express written permission. To obtain permission, contact the Communications Division at ifpri-copyright@cgiar.org. mailto:k.ambler@cgiar.org iii Contents Abstract v Acknowledgments vi 1. Introduction and Motivation 1 2. Theoretical Framework 4 3. Project Design 6 4. Data and Estimation Strategy 11 5. Results 19 6. Conclusion 28 Appendix A: Supplementary Tables 29 Appendix B: Theoretical Framework 30 Appendix C: Text Used in Migrant Experiment 35 Appendix D: Text Used in Recipient Experiment 36 References 37 iv Tables 4.1 Baseline summary statistics 11 4.2 Balance tests: Migrant experiment 13 4.3 Balance tests: Recipient experiment 15 5.1 Impact of monitoring treatment on migrant remittance decision 20 5.2 Impact of monitoring treatment on amount sent by migrant: By proxies for punishment ability 23 5.3 Mean amounts allocated to spending groups by recipients and migrants: Recipient experiment 24 5.4 Impact of monitoring and communication treatments on recipient allocation decision 25 A.1 Comparison of migrants in study with DC-area Salvadorans and Hispanics in the American Community Survey 29 A.2 Impact of monitoring and communication treatments on recipient allocation decision: By information quality measure 29 Figures 3.1 Project timeline 6 3.2 Migrant experiment: Treatments 7 5.1 Cumulative distribution of amount sent by migrant by treatment group 19 v ABSTRACT Although most theoretical models of household decisionmaking assume perfect information, empirical studies suggest that information asymmetries can have large impacts on resource allocation. In this study, I demonstrate the importance of these asymmetries in transnational households, where physical distance between family members can make information barriers especially acute. I implement an experiment among migrants in Washington, DC and their families in El Salvador that examines how information asymmetries can have strategic and inadvertent impacts on remittance decisions. Migrants make an incentivized decision over how much of a cash windfall to remit and recipients decide how to spend a remittance. Migrants strategically send home less when their choice is not revealed to recipients but only when recipients can punish migrants for deviation from remittance agreements. Recipients make spending choices closer to migrants’ preferences when those preferences are revealed, suggesting that recipients’ choices are inadvertently affected by imperfect information. Keywords: remittances, intrahousehold allocation, information asymmetries, transnational households JEL Codes: D13, D82, F24, O15 vi ACKNOWLEDGMENTS Diego Aycinena and Dean Yang contributed as co-authors on a companion experiment. Jessica Snyder provided excellent research assistance. I thank Sarah Burgard, David Lam, Rebecca Thornton, Dean Yang, Josh Hyman, Morgen Miller, Caroline Theoharides, and participants at numerous seminars and conferences for comments and guidance. This work would not have been possible without the support of the Salvadoran consulates and the dedicated work of the survey team. I also thank Alan de Brauw and Angelino Viceisza for advice on project design. This project was funded by the Tokyo Foundation and the Inter-American Development Bank and was supported in part by an NICHD training grant to the Population Studies Center at the University of Michigan (T32 HD007339). 1 1. INTRODUCTION AND MOTIVATION Although the implications of asymmetric information have been well documented in the study of important economic institutions such as labor, credit, and insurance markets, theoretical models of intrahousehold resource allocation have largely assumed perfect information (Chiappori 1988, 1992; Manser and Brown 1980; McElroy and Horney 1981; Lundberg and Pollack 1993). 1 Despite this, a growing body of empirical literature has shown that information asymmetries do exist in households and, further, that household members take strategic advantage of opportunities to use these asymmetries to alter the allocation of resources in the household (Ashraf 2009; Ashraf, Field, and Lee 2013; Schaner 2013). This paper brings the study of how information asymmetries affect intrahousehold resource allocation to a different setting: transnational households, defined as households composed of international migrants and their family members in the home country, in this case El Salvador. Using experimental methods, I examine the effects of a set of information imperfections on remittance decisions made by both migrants and their family members. The context of transnational households is significant because migrants and their family members are making financial decisions in a situation where information asymmetries are especially acute. Because of the physical distance separating family members, families with a migrant living away from the household are precisely those where information asymmetries may be the most pronounced. A number of studies have documented the existence of these asymmetries in households with migrants. For example, De Laat (2014) shows that domestic migrants in Kenya spend resources on costly monitoring of their wives. Chen (2006, 2013) finds that in China, wives with migrant husbands exhibit noncooperative behavior more often for activities that are more difficult to monitor, and Seshan and Yang (2013) find suggestive evidence that Indian migrants underestimate how much their wives at home are saving. McKenzie, Gibson, and Stillman (2013) find that potential Tongan migrants underestimate earnings in New Zealand, a fact the authors partly attribute to underreporting of earnings by current migrants. However the empirical analysis in these papers is largely observational. This is the first study to causally examine how information asymmetries directly affect behavior, specifically decisions about the sending and spending of remittances. The importance of understanding how information asymmetries affect decisions in transnational households is heightened by the fact that migrants and their family members are financially linked through the sending of remittances, a large financial flow. Global aggregate international remittances to the developing world were US$332 billion in 2010, more than any other kind of resource flow with the exception of foreign direct investment (Ratha and Silwal 2012). In El Salvador specifically, remittances received were 16 percent of gross domestic product in 2010 (Ratha and Silwal 2012). In the same year, 21 percent of households in El Salvador received remittances from abroad and average monthly remittances were $166 for families that received them—a figure that is almost 50 percent of average monthly household expenditures for remittance recipients (Dirección General de Estadística y Censos 2010). Additionally, the receipt of remittances has been shown to have large, positive impacts on a variety of measures of well-being, underscoring their importance as a tool for development (Cox-Edwards and Ureta 2003; Adams and Page 2005; Yang and Martinez 2005; Woodruff and Zenteno 2007; Yang 2008; Adams and Cuecuecha 2010). Given the importance of remittances for development, a more complete understanding of how these decisions are made is crucial for policymakers who are hoping to maximize their economic impact. This paper addresses two types of information asymmetries that may affect decisions about the sending and spending of remittances. The first are asymmetries that can lead to strategic behavior, meaning that migrants and recipients recognize that the asymmetry exists and use it for their benefit. The specific asymmetries considered here are the limited abilities of remittance recipients to observe migrant income and of migrants to observe recipient spending. The second type are those that can have 1 Exceptions include Bloch and Rao (2002) and Chen (2013). 2 inadvertent impacts, defined as asymmetries that unintentionally affect decisions. These asymmetries are represented here as communication barriers that result in recipients having an incomplete understanding of migrant preferences for how the remittances they send should be spent. Communication barriers should be interpreted broadly as any obstacle—social, financial, or logistical—to understanding these preferences. I first describe a simple framework that shows how these two types of information asymmetries can affect remittance decisions, and then I test these predictions using experiments conducted with a matched sample of migrants from El Salvador and their family members at home. The framework views the decisions made by migrants and remittance recipients as being driven both by altruism and by contracts that dictate how much of their income migrants should send to recipients and what the money should be spent on when received by the recipients. The contracts are enforced through the threat of punishment for noncompliance, and migrants and recipients may strategically deviate from these agreements when it is not likely their behavior will be observed. However, in pairs where the potential for punishment is low, decisions will be motivated mostly by altruism, and strategic effects will therefore be less important. Additionally, communication barriers, specifically in regard to migrants’ preferences over recipients’ spending habits, may lead to inadvertent deviation from migrant preferences by recipients. The experiments explicitly test for both strategic and inadvertent responses to information asymmetries. They were designed to mimic real-life decisions about remittances made by migrants and their family members; and by randomly assigning treatments, I am able to causally identify the impacts of the informational conditions being tested. The first experiment was conducted among Salvadoran migrants recruited in the Washington, DC, area. The migrants were asked how much of a potential $600 prize they wished to keep and how much they wished to send to a family member in El Salvador. The decision was incentivized, meaning that participants had the chance to win the allocation that they chose. To test whether migrants strategically react to changes in the observability of their income, they were randomly allocated into two treatment groups: those who were told their decision would not be revealed to their family and those who were told that their decision would be revealed. These family members then participated in a second experiment. They made an incentivized decision about how to spend a potential $300 remittance prize. To test for strategic reactions to the observability of their spending choices, as in the migrant experiment, half of the recipients were told that their choice would not be revealed to the migrant, and the other half were told that their choice would be revealed. In a second, cross-randomized treatment addressing the inadvertent effects of communication barriers, half of the recipients were informed of the migrant’s preferences for how the money should be spent, and the others were not. I find that migrants remit $20 more on average out of the possible $600 (an increase of 5 percent over the control group mean of $440 sent) when their decisions are revealed. This effect is concentrated (and larger) in subsamples where the recipient’s ability to punish the migrant for deviation is plausibly high. There is no corresponding evidence of strategic behavior in the recipient experiment. However, reducing communication costs by revealing migrant preferences to recipients does have an impact, resulting in a 10 percent reduction in the difference between migrant preferences and recipient choices. This paper is related to a set of field experiments that have examined the effects of offering migrants varying degrees of control over remittances. The idea behind these experiments is that offering control to migrants will mitigate a moral hazard problem in how recipients spend remittances. Ashraf et al. (2011) show that savings levels in bank accounts in El Salvador increase when migrants are given greater control over these accounts. Chin et al. (2011) find that the impacts of an experiment that offered migrants assistance in opening bank accounts in the United States are concentrated among migrants who report having no control over how their remittances are spent. However, in a lab experiment, Torero and Viceisza (2011) find little evidence that Salvadoran migrants send more when they are able to control how remittances are spent but attribute this to the fact that the control offered by their experiment was too limiting. 3 The main limitation of these papers is that while their premise is that migrants might have difficulty controlling the spending of remittances, they do not consider that information problems might run in both directions. The observational studies documenting information asymmetries in migrant households have also focused on migrant monitoring of recipient behavior (Chen 2006, 2013; de Laat 2008). One of the principal contributions of this paper is that it examines the impacts of information asymmetries on both sides of the migrant–recipient relationship. In fact, in this experiment, it is only migrants—and not recipients—who react strategically to whether or not their choices will be monitored. This demonstrates that recipients have important influence in the migrant–recipient relationship, something that has not previously been demonstrated empirically. This paper also fits into a growing, broader literature on how information asymmetries affect resource allocation in families. Ashraf (2009) shows that in the Philippines, men whose wives are the household financial managers hide income from their wives when that decision is private. When their decision is public, men choose to divert income to committed consumption that cannot be undone. Only when spouses communicate about their choices before they make them do men choose to share the income with their wives. Schaner (2013) finds that spouses are more likely to choose to save in individual (as opposed to joint) savings accounts when they are not well informed about each other’s finances. 2 Also related, Jakiela and Ozier (2012) find that women in Kenya sacrifice investment returns to keep income secret from family members outside their household and avoid the pressure to share that income. This study builds on the literature in several ways. First, while these studies focus on strategic behavior; I study responses to different types of information asymmetries, strategic and inadvertent, allowing me to evaluate the relative effects in the same population. Second, these papers largely focus on a single choice in the resource allocation process (whether or not to share income); the present experiment considers how information asymmetries can affect two different decisions made by families about economic resources. Finally, this study documents that information asymmetries can be important outside of the husband–wife pair, which has been the context of most previous experimental work in this area. 3 People in developing countries often transfer resources within extended families, and therefore decisions about resource allocation consequently are likely to involve people beyond just the husband and wife. The results show that information asymmetries can have important impacts in extended families, but because migrants react to being monitored only when the recipients’ ability to punish them is high, the results also indicate that asymmetries may not matter in all families where resources are shared. The paper proceeds as follows: Section 2 describes a framework for understanding the empirical results. Section 3 explains the experiment. Section 4 describes the data and the empirical strategy. Section 5 presents and discusses the results, and Section 6 concludes. 2 In another experiment in Zambia, women are more likely to take advantage of vouchers for contraception and use concealable forms of contraception when these vouchers are given to them outside of the presence of their husbands, showing that strategic reactions to information asymmetries extend beyond simply the allocation of funds in the household (Ashraf, Field and Lee 2012). 3 The dynamics of transfer arrangements in extended family networks has been studied (Foster and Rosenzweig 2001), but little is known about how information imperfections affect behavior in these arrangements. The notable exception is Jakiela and Ozier (2012), but the main focus of that paper is investment decisions, as opposed to sharing and spending decisions. 4 2. THEORETICAL FRAMEWORK In this section I describe the intuition behind a simple model that frames my experimental results. A detailed exposition of the model can be found in Appendix B. The intent of the model is to show that the probability that recipients observe migrant income and the probability that migrants observe recipient spending can impact migrant sending and recipient spending decisions. The structure of the model is based on Chen’s (2013) description of how male migrants in China monitor their wives’ behavior. Specifically, Chen shows that when a migrant has imperfect information about his wife’s actions and incomplete information about her preferences, the contract offered to the wife may not always be incentive compatible. I adapt a simplified version of this model to describe the outcomes considered in this paper. In this framework, the migrant’s decision about how much to send home in remittances is determined by a combination of altruism and a contract with the recipient that is enforced through the threat of punishment. One example of such a cost is substandard care for or attention to people (children or elderly relatives) or possessions (land, livestock, or new investments) left by the migrant in the care of his family. Another is social sanctions against the migrant: many migrants come from areas with high rates of migration and strong social norms and expectations regarding the amount of money that migrants send home. Particularly for migrants who wish to return home one day, a damaged reputation may be seen as quite costly. Finally, migrants who refuse to send home as much money as their families expect may damage their relationships with their families, relationships that migrants with tenuous positions in foreign countries may view as important. 4 Many of these potential punishments are related to the social closeness of migrants and recipients, and indeed, in a qualitative study of Ghanaian migrants in the Netherlands, Mazzucato (2009) emphasizes the importance of the social proximity of migrants and recipients for the effective enforcement of remittance agreements. 5 The amount that the migrant is compelled to send through the contract depends on the size of the punishment: The greater the ability of the recipient to punish, the more the recipient can push the migrant past the amount the migrant would choose to send voluntarily. These contracts are defined based on the migrants’ earnings, which vary over time. Recipients will expect different amounts depending on how much a migrant earns in any given period. Because of the geographical distance between migrants and recipients, recipients observe migrant income only with a certain probability, and that probability varies from period to period. 6 I model this probability as being either high or low. Migrants always know the value of this probability, but recipients know only its distribution. If migrant income is not observed by the recipient, the recipient must rely only on information from the migrant. Therefore, migrants can deviate from the contract by underreporting their income to the recipients and will be discovered and punished only if that income is actually observed. Recipients must set up these contracts to be incentive compatible: Migrants’ utility from compliance must be greater than or equal to the expected utility from deviating. This expected utility will be higher when the probability that income will be observed is low; therefore, the cooperation-inducing remittance is lower when that probability is low than when it is high. Because recipients do not know the 4 A similar description of enforcement mechanisms can be found in Rapoport and Docquier (2006). Brown (1997), Hoddinott (1994), Lucas and Stark (1985), and Poirine (1997) all describe remittance contracts enforced through the discussed mechanisms. Additionally, in studies of dictator games within social networks, Leider et al. (2009) and Ligon and Schecter (2012) document the importance of the expectation of reciprocity in motivating giving. 5 Additionally, in focus groups done prior to the start of the project, migrants repeatedly cited high levels of pressure from family members as a key reason why they sent remittances home. Relatedly, in their work on Tongan migration to New Zealand, McKenzie, Gibson, and Stillman (2013) find suggestive evidence that migrants underreport earnings to avoid pressures from family members to remit. 6 This observation of income will likely occur through the migrant’s social and family network. Given that the networks within which migrants in the United States live and work are often closely related to their home country networks (Munshi 2003), instances when the probability that income will be observed is quite high are likely. 5 probability of observing migrant income, they must set one contract that satisfies incentive compatibility for all migrants and will lose potential surplus when that probability turns out to be high. However, one other option is available to the recipients. They can set a contract that binds when the probability of observing migrant income is high but that is not incentive compatible when that probability is low. Because recipients have to lower the remittance amount so much to induce compliance when the probability is low, they may be better off receiving the higher amount only some of the time instead of the lower amount all the time. The model leads to two predictions for migrant behavior: First, migrants will act strategically and send less money home when the probability that recipients will observe their income is low. Second, this will happen only when recipients have power to punish migrants for not sending home enough money. If recipients cannot punish migrants, then migrants will send only the amount they choose to send altruistically. This altruistic amount will not depend on whether the migrant’s income can be observed. The decision made by recipients about how to spend the remittances that they receive is modeled in a parallel manner. Recipients choose the extent to which they follow the migrants’ preferences for how remittances should be spent. Their choice is motivated by altruism (in this case simply the extent to which recipients want to follow migrant preferences) and a contract with migrants enforced by the migrant’s ability to punish the recipient. The probability that migrants will observe actual recipient spending can be either high or low. This probability is known to the recipient, while the migrant knows only its distribution. As with the migrant remittance decision, this leads to a situation where migrants may offer contracts that are incentive compatible only when the probability that recipient spending will be observed is high. The recipient decision may additionally be complicated by barriers to communication that result in confusion over what the migrant’s preferences actually are. I will refer to these barriers as communication costs, but the concept is broader than just the cost of a telephone call. With distance, specificity about preferences may become difficult, migrants may feel uncomfortable expressing what they want, and recipients may sometimes have to make decisions without time to directly consult with migrants. Family members may also incorrectly assume that they know what the migrant would prefer. If the recipient does not know the migrant’s preferences, the preferences will not be followed regardless of the recipient’s intentions. This setup leads to three predictions for the recipient decision: First, recipients will strategically spend less money according to migrant preferences when the probability their spending will be observed is low. Second, when migrants cannot punish recipients, the probability that spending will be observed will not affect spending decisions because the recipient choice will be strictly voluntary. Finally, if recipients do not have full information, they may be inadvertently deviating from migrant preferences. Improved information can increase the extent to which recipients follow those preferences. The main point of this discussion is that strategic behavior can be a part of the relationship between migrants and their family members, but the extent to which it is important will depend on their ability to pressure and punish each other. At the same time, communication costs can lead to inadvertent deviation when recipients make remittance spending decisions. 6 3. PROJECT DESIGN Given that information asymmetries in transnational families are difficult to measure and may be correlated with a number of unobserved characteristics, I implemented a randomized experiment to test the predictions of the framework discussed in the previous section. This experiment is conducted within the context of survey work for a separate field experiment on remittances and education among Salvadoran migrants in Washington, DC, and their families in El Salvador (Ambler, Aycinena, and Yang 2013). Specifically, I exploit an unusual feature of this data collection exercise; it involves surveys with matched pairs of migrants and family members, allowing me to investigate the preferences and choices of both. In the experiment, I randomly vary (1) whether migrant income and recipient spending are observed and (2) the size of communication costs, allowing me to identify the causal impacts of both of these factors on migrant and recipient remittance behavior. Demographic survey data are used to explore how impacts vary by punishment ability. Migrants were recruited in the Washington, DC, metro area at the two area locations of the Salvadoran consulate and were interviewed while they were waiting for consular services. 7 The migrant survey was conducted between late September 2011 and late February 2012. Surveyors in the consulate approached migrants and invited them to participate. Because the focus of the companion experiment was remittances and education, participants were required to name a high school or college-aged relative or acquaintance in El Salvador. 8 Those who qualified and agreed to participate were administered a baseline survey. 9 The experiment described in this paper was conducted at the end of the survey. During the survey migrants identified a high school or college-aged student in El Salvador. Although the migrants were not required to select a family member as the student, in practice 97 percent did. Interviews were subsequently conducted with the student or a household member. If the student was 18 years of age or older, the student was to be interviewed; and if under 18, a guardian was identified to be interviewed. If the indicated person was not available, an alternative adult in the household was interviewed instead. The El Salvador survey was conducted by phone in the days following the migrant survey in the United States; the median number of days between the US and El Salvador surveys was eight. The El Salvador surveys concluded in mid-March 2012, two weeks after fieldwork ended in the United States. The experiment in the El Salvador survey was also at the end of the survey. Figure 3.1 describes the phases of the project in the order that they occurred for each pair of participants. Figure 3.1 Project timeline Source: Author. 7 The most common reason to go to the consulate is passport renewal, but other services include renewal of temporary protected status, registry of births and deaths, and notarization of documents. 8 Among the migrants approached, 24 percent participated. Of those that did not participate, 77 percent did not know an eligible student in El Salvador, 14 percent refused, 7 percent were not from El Salvador, and 2 percent had other reasons. 9 This was followed by the randomized offer of a product designed to facilitate the sending of remittances for education to El Salvador. This was a randomized intervention, and migrants received offers of different versions of the product depending on their assigned treatment group. Migrants in a control group received only information and no product offer. Migrant recruited Migrant survey administered Migrant experiment performed Recipient survey administered Recipient experiment performed Migrant and recipient: information revelations Migrant and recipient: Lotteries performed 7 Migrant Experiment The migrant experiment consisted of an incentivized remittance-sending decision. Migrants were told that they were being given the chance to win $600 and would have to decide how much of the prize to keep for themselves and how much to remit to their family member in El Salvador. Migrants could split the $600 as they wished but were restricted to using $100 intervals for simplicity. 10 The prize was awarded through a lottery. 11 Although budgetary restrictions did not allow for all participants to win the prize, the use of the lottery incentivized participants to treat this as a real decision. In the Ashraf et al. (2011) study of a similar population of Salvadoran migrants, migrant median monthly income was $2,080. Consequently, $600 represents a significant increase in monthly income. The question text can be found in Appendix C. Migrants were randomly allocated into two groups: those who were told that their choice would be revealed to their family member, and those who were told that their choice would not be revealed. The family member referred to in the question was the person to be surveyed in El Salvador. A description of the treatments is presented in Figure 3.2. Because the treatment varies the ability of the recipient to monitor the migrant’s actions, I refer to this treatment as the migrant-monitoring treatment. This experiment exogenously varies the probability that recipients will observe migrant income and measures the extent to which migrants strategically take advantage of asymmetric information. The framework described in Section 2 predicts that migrants will send less money home when the probability their income will be observed is low. In the context of the experiment, this means that migrants whose choice is not revealed to the recipient should send less than those whose choice is revealed. However, following the predications of the framework, migrants should act strategically only when recipients can punish them for noncompliance. Migrant responses from the baseline survey will be used to proxy for recipients’ ability to punish migrants. Differences between the two groups—choice revealed and choice not revealed—should be evident only when those proxies indicate that the ability to punish exists. Figure 3.2 Migrant experiment: Treatments Source: Author. Recipient Experiment The recipient experiment consisted of an incentivized remittance spending decision. The respondents in the El Salvador phone survey were told that because their family member in the United States participated in the study, they now had the chance to win a remittance worth $300. They had to decide what to spend the remittance on and were asked to split the $300 among four spending categories: restaurant meals, education, daily expenses, and health expenses. Recipient choices were limited to four categories for simplicity in the context of a phone survey. If among the winners, recipients would receive exactly the 10 In pilot surveys where migrants were not given restrictions, almost all chose to split the money in $100 intervals. 11 Two prizes were awarded. If asked, surveyors told migrants the number of prizes and the date of the drawing. The first prize was awarded midway through survey work, and the second when survey work had concluded. Migrants were eligible for only one drawing. 8 allocations that they requested.12 Prizes were awarded in kind. The median monthly remittance in the Ashraf et al. (2011) study was $325, so a $300 remittance is a standard amount for many recipients. The question text can be found in Appendix D. Two separate treatments were administered to recipients: the recipient-monitoring treatment and the recipient communication treatment. The two recipient treatments were cross-randomized, also allowing for the analysis of their interaction. They are depicted in Figure 3.3. Figure 3.3 Recipient experiment: Treatments Source: Author. Recipient Monitoring Treatment The recipient-monitoring treatment is parallel to the migrant-monitoring treatment. Recipients were randomly allocated into two groups: those who were told that their choice would be revealed to the migrant, and those who were told that their choice would not be revealed to the migrant. This treatment randomly varied the probability that recipient spending would be observed and measures the extent to which recipients strategically take advantage of this asymmetric information. The framework predicts that recipients are more likely to strategically deviate by spending less according to the migrant’s preferences when the probability that spending will be observed is low. Therefore, the results of the experiment should show that recipients make choices closer to migrant preferences when those choices are revealed to the migrant. However, this effect will be present only in pairs where the migrant can punish the recipient for noncompliance. Recipient Communication Treatment During the US survey, migrants were told about the lottery for recipients and asked what their preferences were for how the recipients would spend the money. Again, recipients were randomly allocated into two groups: those for whom the migrant’s preferences were revealed and those for whom the migrant’s preferences were not revealed. Making these preferences clear is a proxy for improving communication, 12 Four prizes were awarded. If asked, surveyors told recipients the number of prizes and the date of the drawing. Two prizes were awarded midway through survey work and the other two when survey work had concluded. Recipients were eligible for only one drawing. Communication treatment Migrant preference not revealed to recipient Migrant preference revealed to recipient M o n it o ri n g tr ea tm en t Recipient choice not revealed to migrant N = 314 N = 324 N = 638 Recipient choice revealed to migrant N = 327 N = 333 N = 660 N = 641 N = 657 9 so the experiment exogenously improves communication about migrant preferences for expenditures. This treatment is therefore a test of whether or not communication costs can lead to inadvertent deviation from migrant preferences by the recipient. Revealing migrant preferences to the recipient should decrease the difference between the recipients’ choices and the migrants’ preferences when communication problems exist. Experiment Logistics For the experiment to work as intended, respondents must have believed that the threat of revealing their choices to their family members was credible. Because the interviewer collected contact information for the recipient families from the migrants and allowed the migrants to use a project phone during the interview to call their family members and tell them about the study, migrants were aware that their family members could indeed be contacted. Similarly, because recipients being interviewed knew that they had been contacted through the migrant, they also knew that their migrant family members could be contacted. Although it has no impact on the results of the experiment, for all respondents in the choice- revealed treatment groups of the monitoring treatments, an effort was made to inform their family member of the choice made by the participant. The randomization in this study was performed at the participant level. Surveys were preassigned treatment status, and migrant and recipient treatments were randomized separately. Because remittance behavior can vary by season, it was important to ensure that treatments were balanced over time. I achieved this by stratifying the randomization for all treatments within groups of 16 surveys and by the treatment offered in the companion experiment. The recipient treatments were additionally stratified by the migrant treatment. Threats to Interpretation Although the experimental methodology used in this paper allows for the causal identification of the effects of information asymmetries that are otherwise difficult to isolate, several aspects of the design could lead to arguments that participants’ behavior in the experiment is not the same as it would be in their day-to-day lives. The first of these is that the experiments, particularly the migrant experiment, ask participants to make decisions about windfall income that is given to them rather than earned, and therefore migrants may be more generous with this income than they would be with other income. There are several responses to this. First, although it is true that the income in the experiment is transitory and not permanent, many of the migrants in this study work in jobs where income is highly variable from month to month, making transitory versus permanent income a less important distinction. Second, studies that have examined earned versus unearned windfall income have found that people are more generous with unearned winnings but that the effect is small (Jakiela 2009). Finally, the focus of this paper is not on the total amount sent by the migrants but on the effect of the monitoring treatment on the amount sent. The issues discussed here should apply equally to each treatment group. If anything, if migrants in the choice-not-revealed group do indeed have a greater ability to keep the funds for themselves, the impact of the monitoring treatment should increase as migrants feel more ownership over the winnings, meaning that the results in this paper can be considered a lower bound on the true effect. A second potential issue is that in both the migrant and the recipient experiments, prizes were awarded by lottery, meaning that the expected value of the prize for each participant is much lower than the value of the actual prize. Participants may be more generous or less likely to make controversial decisions because they know they are unlikely to win the lottery prize. Little research has been done into how subjects react to lottery prizes, but some evidence does exist. Laury (2005) conducts a laboratory experiment in which respondents make the same choices when payoffs are random as when payoffs are guaranteed. Additionally, because the questions in this paper relate to the differences between the treatment groups, if the lottery does impact participants’ decisions, the estimated effect can be considered to be a lower bound. 10 The final issue is that of the fungibility of choices made during the experiment. Both migrants and recipients could potentially undo their choices during the experiment through their actions afterward. For example, migrants could choose to not send a remittance that they would have sent otherwise. Although some of this behavior may be occurring, it is not necessarily important for the interpretation of the treatment. If the results show differences between the two treatments then that is evidence that people are reacting to variations in information. 13 13 Even without evidence of differences between treatments, if participants wish to do something that their partner would disagree with, it makes sense for them to take advantage of the experiment to do so, when the probability of keeping that action secret is high. 11 4. DATA AND ESTIMATION STRATEGY Data The migrant baseline survey collected information on demographics, characteristics of migrant family relationships both in the United States and in El Salvador, detailed information on remittances sent by the migrant to the recipient household and to other households, and a set of questions to assess the quality of the migrant’s relationship with the recipient household and the migrant’s involvement in household affairs. The recipient survey, administered by phone, was shorter and contained demographic information, some limited questions on remittances received from the migrant, and information on the education of children in the household. Table 4.1 shows summary statistics from both surveys: 1,581 migrant surveys and 1,298 recipient surveys were performed, a completion rate of 82 percent for the recipient surveys. For the migrant survey, summary statistics are shown for both the full sample and the sample with completed recipient surveys. No meaningful differences are evident between the two samples; therefore, I limit the analysis sample to the 1,298 migrant–recipient pairs with completed El Salvador surveys. Results from the migrant experiment do not change significantly between the two samples. Additionally, I show that attrition from the full sample of migrant surveys to the sample of completed recipient surveys is not related to treatment status (Tables 4.2 and 4.3). The breakdown of participants in the final analysis sample into the different treatment groups can be seen in Figure 3.2 (migrant experiment) and Figure 3.3 (recipient experiment). Table 4.1 Baseline summary statistics All observations Observations with completed recipient survey Mean SD N Mean SD N Baseline variables from migrant survey Migrant is female 0.50 0.50 1,581 0.51 0.50 1,298 Migrant age 36.83 9.41 1,538 36.92 9.29 1,264 Migrant can read and write 0.96 0.20 1,554 0.96 0.20 1,275 Migrant’s years of education 9.08 4.67 1,560 9.01 4.67 1,282 Migrant’s years in the United States 11.31 6.38 1,577 11.13 6.27 1,295 Migrant is married 0.62 0.48 1,575 0.63 0.48 1,294 Migrant lives with spouse 0.49 0.50 1,579 0.50 0.50 1,296 Migrant’s total number of children 2.28 1.69 1,579 2.34 1.69 1,296 Migrant’s number of children in El Salvador 1.01 1.43 1,577 1.07 1.47 1,294 Migrant’s number of children in United States 1.26 1.32 1,575 1.25 1.29 1,293 Migrant’s hh size in United States 4.32 1.98 1,581 4.36 1.96 1,298 Migrant has child 22 or under in El Salvador 0.32 0.47 1,581 0.34 0.47 1,298 Recipient is migrant’s close relative 0.29 0.45 1,574 0.31 0.46 1,291 Migrant has worked in last 12 months 0.89 0.31 1,581 0.89 0.31 1,298 Migrant in lowest income bracket 0.52 0.50 1,429 0.53 0.50 1,181 Migrant sent remittances to recipient hh 0.85 0.36 1,580 0.87 0.34 1,297 Migrant’s annual regular remittances to recipient hh ($) 2,298 2,907 1,565 2,440 2,998 1,283 Migrant’s annual irregular remittances to recipient hh ($) 337 706 1,575 344 707 1,293 Migrant’s annual total remittances to recipient hh ($) 2,629 3,199 1,563 2,777 3,284 1,281 Migrant’s annual total remittances to other hhs ($) 1,097 1,905 1,567 1,123 1,944 1,284 Migrant communicates with recipient hh at least weekly 0.69 0.46 1,578 0.71 0.45 1,295 12 Table 4.1 Continued All observations Observations with completed recipient survey Mean SD N Mean SD N Baseline variables from recipient survey Recipient is target student 0.45 0.50 1,298 Recipient is student’s guardian 0.40 0.49 1,298 Recipient is female 0.68 0.47 1,298 Recipient age 34.20 15.84 1,295 Recipient is married 0.36 0.48 1,298 Recipient’s years of education 9.37 5.27 1,292 Recipient lives in urban area 0.43 0.50 1,298 Recipient’s hh size 4.99 2.04 1,296 Annual remittances received from migrant ($) 1,522 1,916 1,203 Baseline comparison variables Migrant and recipient report same hh budget priorities 0.48 0.50 1,231 Migrant and recipient report same student GPA 0.24 0.43 1,041 Migrant and recipient report same student mode of transport 0.43 0.50 1,107 Source: Author’s calculations. Notes: SD = standard deviation; N = number; hh = household; GPA = grade point average. The “All Observations” sample is respondents with nonmissing data for questions in the migrant experiment. The “Completed Recipient Survey” sample additionally conditions on completion of the recipient survey and nonmissing migrant and recipient information for questions in the recipient experiment. Number of observations varies slightly with missing values. Recipient is defined as close relative if migrant reports recipient to be spouse, parent, or child. Migrants in the lowest income bracket chose $400 or less as the weekly income of themselves plus their coresident spouses. The other categories were $401–$600, $601–$800, and $801 and above. Annual regular remittances were collected by asking for the frequency of remittances sent and the average amount sent each time. Annual irregular remittances are remittances sent for special occasions or emergencies. The recipient variables in all cases refer to the person completing the recipient survey. The baseline comparison variables were asked on both surveys and are equal to 1 if the migrant and recipient responses match. Both respondents were asked to choose the three most important budget priorities for the recipient hh from a list of seven categories. Student refers to the student identified by the migrant during the baseline survey. GPA and mode of transport were only asked when student was reported to be in school. The migrants are half female with an average age of 38. 85 percent sent remittances to the recipient household in the last 12 months, indicating that most pairs in the sample have an established remittance relationship. Average annual remittances to the recipient household (reported by the migrant) are $2,629. Average annual remittances to other households in El Salvador are $1,059. The $1,600 difference between average remittances to the recipient household and those to other households suggests that most recipient households are the migrant’s primary remittance recipient. The mean number of years in the United States is 11, so the migrants are largely established in the United States. In my sample, 32 percent of migrants report having a son or daughter aged 22 or under in El Salvador, and 69 percent report communicating with the recipient household at least weekly. The sample is also low income; half of the migrants report earning $400 a week or less. Because of the structure of the project, the interviewed recipients are either the student identified by the migrant (45 percent) or the student’s guardian if the student is under 18 (40 percent). The remaining 15 percent of interviews were done with a different adult in the household if the student or guardian could not be reached. The recipient sample is heavily female (68 percent) because identified student guardians tend to be female. Because migrants were recruited in the Salvadoran consulate and screened into the study on the basis of having a young adult relative in El Salvador, a concern may be that the respondents are not representative of the larger migrant community and that the results are therefore not indicative of what might be found in a more representative sample. In Table A.1 in the Appendix, I compare characteristics of the migrants from the baseline survey (gender, age, time in the United States, household size, and education) to migrants in the 2008–2010 American Community Survey (ACS). I restrict the ACS sample 13 to non-US citizens aged 18 to 65 who live in the Washington, DC, metro area who are either Salvadoran- born or Hispanic. The study participants are quite similar to the ACS samples, in particular to the Salvadoran-born sample, suggesting that study participants are not overly different from the greater migrant population. The random assignment of the treatments in this experiment allows for the causal identification of their impacts. Randomization should provide treatment groups that are the same on average so that any difference between the groups can be attributed to the treatment and not to some preexisting difference between groups. Tables 4.2 and 4.3 test whether the treatment groups are balanced on observed characteristics from the baseline survey for the treatment groups for the migrant experiment and the recipient experiments, respectively. In Table 4.2 the means for both treatment groups in the migrant- monitoring treatment are presented in the first two columns, and the p-value of the hypothesis test of whether or not those means are equal is in the third column. Overall, the treatment groups are well balanced: Only 2 of 34 differences are significantly different from zero at the 10 percent level. Table 4.3 shows the means by treatment group for the two recipient treatments and p-values for differences in those means. Only 3 of the 34 p-values for the recipient-monitoring treatment and one for the recipient communication treatment are less than 0.10. Some differences between treatment groups may occur by chance, and these few small differences are not cause for concern. However, to allay any concerns of an unbalanced sample affecting results, I include regression specifications with control variables. Table 4.2 Balance tests: Migrant experiment Treatment group means: P-value for difference of means: Choice not revealed and choice revealed Migrant choice not revealed to recipient Migrant choice revealed to recipient Attrition Recipient survey completed 0.82 0.83 0.819 Baseline variables from US survey Migrant is female 0.53 0.49 0.165 Migrant age 36.90 36.94 0.941 Migrant can read and write 0.95 0.97 0.150 Migrant’s years of education 9.01 9.00 0.966 Migrant’s years in the US 10.90 11.37 0.178 Migrant is married 0.61 0.65 0.151 Migrant lives with spouse 0.50 0.50 0.956 Migrant’s total number of children 2.34 2.34 0.956 Migrant’s number of children in El Salvador 1.03 1.10 0.365 Migrant’s number of children in US 1.28 1.22 0.410 Migrant’s hh size in US 4.34 4.38 0.720 Migrant has child 22 or under in El Salvador 0.32 0.37 0.059 Recipient is migrant’s close relative 0.29 0.33 0.178 Migrant has worked in last 12 months 0.90 0.89 0.943 Migrant in lowest income bracket 0.53 0.53 0.886 Migrant sent remittances to recipient hh 0.87 0.86 0.586 Migrant’s annual regular remittances to recipient hh ($) 2,494 2,386 0.520 Migrant’s annual irregular remittances to recipient hh ($) 354 334 0.627 Migrant’s annual total remittances to recipient hh ($) 2,828 2,726 0.579 Migrant’s annual total remittances to other hhs ($) 1,059 1,185 0.245 Migrant communicates with recipient hh at least weekly 0.73 0.69 0.057 14 Table 4.2 Continued Treatment group means: P-value for difference of means: Choice not revealed and choice revealed Migrant choice not revealed to recipient Migrant choice revealed to recipient Attrition Recipient survey completed 0.82 0.83 0.819 Baseline variables from recipient survey Recipient is target student 0.45 0.45 0.907 Recipient is student’s guardian 0.42 0.38 0.160 Recipient is female 0.69 0.67 0.331 Recipient age 35.09 33.31 0.043 Recipient is married 0.36 0.36 0.941 Recipient’s years of education 9.21 9.54 0.285 Recipient lives in urban area 0.43 0.44 0.649 Recipient’s hh size 4.90 5.08 0.111 Annual remittances received from migrant ($) 1,491 1,553 0.580 Baseline comparison variables Migrant and recipient report same hh budget priorities 0.48 0.48 0.926 Migrant and recipient report same student GPA 0.25 0.24 0.709 Migrant and recipient report same student mode of transport 0.44 0.42 0.573 Source: Author’s calculations. Notes: GPA = grade point average; hh = household. Sample is observations with nonmissing values for the experiment questions and completed recipient survey. Attrition is measured from sample of all migrants who completed the survey and the migrant experiment to sample with completed recipient survey and recipient experiment. Sample size for each comparison of means varies slightly by missing values for each variable. The percentage of missing values for each variable is also tested for balance across treatment groups with no significant differences. Other notes on variable construction are as in Table 3.1. P-values come from a regression of each variable on treatment, with standard errors adjusted for heteroskedasticity. 15 Table 4.3 Balance tests: Recipient experiment Monitoring treatment Communication treatment Treatment group means: P-value for difference of means: Choice not revealed and choice revealed Treatment group means: P-value for difference of means: Preference not revealed and preference revealed Recipient choice not revealed to migrant Recipient choice revealed to migrant Migrant preference not revealed to recipient Migrant preference revealed to recipient Attrition Recipient survey completed 0.81 0.83 0.315 0.82 0.83 0.730 Baseline variables from US Survey Migrant is female 0.52 0.50 0.532 0.49 0.53 0.186 Migrant age 36.56 37.27 0.176 36.90 36.95 0.922 Migrant can read and write 0.95 0.96 0.461 0.96 0.95 0.295 Migrant’s years of education 9.02 9.00 0.947 8.97 9.04 0.798 Migrant’s years in the US 11.18 11.08 0.774 11.13 11.13 0.993 Migrant is married 0.65 0.61 0.175 0.63 0.63 0.952 Migrant lives with spouse 0.51 0.49 0.543 0.50 0.50 0.957 Migrant’s total number of children 2.30 2.38 0.352 2.37 2.31 0.560 Migrant’s number of children in El Salvador 1.01 1.12 0.206 1.04 1.09 0.557 Migrant’s number of children in US 1.27 1.24 0.725 1.31 1.20 0.105 Migrant’s hh size in US 4.43 4.29 0.183 4.43 4.29 0.214 Migrant has child 22 or under in El Salvador 0.33 0.35 0.366 0.34 0.34 0.885 Recipient is migrant’s close relative 0.30 0.32 0.539 0.34 0.29 0.059 Migrant has worked in last 12 months 0.89 0.90 0.401 0.89 0.89 0.950 Migrant in lowest income bracket 0.51 0.54 0.229 0.53 0.53 0.934 Migrant sent remittances to recipient hh 0.86 0.88 0.510 0.87 0.87 0.802 Migrant’s annual regular remittances to recipient hh ($) 2,435 2,444 0.953 2,315 2,561 0.141 Migrant’s annual irregular remittances to recipient hh ($) 382 308 0.062 353 335 0.655 Migrant’s annual total remittances to recipient hh ($) 2,802 2,752 0.786 2,648 2,903 0.165 Migrant’s annual total remittances to other hhs ($) 1,137 1,110 0.804 1,068 1,177 0.314 Migrant communicates with recipient hh at least weekly 0.73 0.69 0.192 0.70 0.72 0.585 Baseline variables from recipient survey Recipient is target student 0.44 0.46 0.402 0.46 0.44 0.495 Recipient is student’s guardian 0.42 0.38 0.239 0.39 0.41 0.319 Recipient is female 0.68 0.68 0.998 0.68 0.68 0.726 Recipient age 34.44 33.97 0.589 34.29 34.11 0.835 Recipient is married 0.41 0.32 0.001 0.35 0.38 0.243 Recipient’s years of education 9.22 9.53 0.294 9.30 9.45 0.622 Recipient lives in urban area 0.41 0.46 0.061 0.42 0.45 0.312 Recipient’s hh size 5.04 4.95 0.471 5.06 4.93 0.271 Annual remittances received from migrant ($) 1,534 1,510 0.825 1,484 1,559 0.497 16 Table 4.3 Continued Monitoring treatment Communication treatment Treatment group means: P-value for difference of means: Choice not revealed and choice revealed Treatment group means: P-value for difference of means: Preference not revealed and preference revealed Recipient choice not revealed to migrant Recipient choice revealed to migrant Migrant preference not revealed to recipient Migrant preference revealed to recipient Baseline comparison variables Migrant and recipient report same hh budget priorities 0.46 0.50 0.189 0.47 0.49 0.401 Migrant and recipient report same student GPA 0.25 0.24 0.844 0.24 0.24 0.952 Migrant and recipient report same student mode of transport 0.41 0.45 0.228 0.43 0.42 0.671 Source: Author’s calculations. Notes: GPA = grade point average; hh = household. Sample is observations with nonmissing values for the experiment questions and completed recipient survey. Attrition is measured from sample of all migrants who completed the survey and the migrant experiment to sample with completed recipient survey and recipient experiment. Sample size for each comparison of means varies slightly by missing values for each variable. The percentage of missing values for each variable is also tested for balance across treatment groups with no significant differences. Other notes on variable construction are as in Table 3.1. P-values come from a regression of each variable on treatment, with standard errors adjusted for heteroskedasticity. 17 The first row of Tables 4.2 and 4.3 also test whether attrition from the full sample of migrants to the estimation sample of migrant–recipient pairs with completed recipient surveys is related to treatment. Attrition is not significantly related to treatment for migrants or recipients. Estimation Strategy: Migrant Experiment The results of the migrant experiment can be analyzed by estimating the following regression using ordinary least squares: (1) where is the dependent variable indicating the amount that the migrant chose to send to the recipient. is the treatment indicator for the monitoring treatment, and it is equal to 1 when the migrant’s choice is revealed to the recipient. The coefficient is the averagdifference between how much migrants choose to send when their decisions are not revealed and when they are revealed. If is positive, migrants send more money to the recipients when equals 1. are stratification cell fixed effects. There are 111 survey group stratification cells in all regressions. The inclusion of baseline control variables is not necessary for causal inference, but I will show specifications with controls to show that they do not affect the results. The control variables include migrant age, gender, education, household size, years in the United States, remittances to recipient household, and other migrant background characteristics. is the error term, which I adjust for heteroskedasticity. Estimation Strategy: Recipient experiment Unless migrant and recipient preferences are different on average, regressions examining the impact of the treatment on the amounts allocated to the four different categories by the recipients will be uninformative. However, because the US survey collected the migrant’s preferences for the recipient’s choices for all participants, it is possible to examine the exact parameter described in the framework guiding the experiment: the extent to which the recipient’s choices match the migrant’s preferences. I implement this by calculating the absolute value of the difference between the recipient’s choice and the migrant’s preference in each of the four categories. I also create a summary measure across the four categories by summing the four difference variables and dividing by 2 to scale the total to 300. I refer to this as the total difference, and it is the primary dependent variable of interest. It is a measure of the number of dollars out of the $300 on which the migrant and recipient match. For example, a total difference of 100 would mean the recipient’s choices matched the migrant’s preferences on $200 of the $300, but that they allocated the remaining $100 to different categories. The results of the recipient experiment can be analyzed by estimating the following regression: (2) where is the difference between migrant preferences and recipient choices in each of the four spending categories or the total difference. is the treatment indicator for the recipient-monitoring treatment and is equal to 1 when the recipient’s choice is revealed to the migrant. is the treatment indicator for the communication treatment and is equal to 1 when the migrant’s preferences are revealed to the recipient before the recipient decides how to allocate the remittance funds. The coefficient is the average difference in the difference between migrant preferences and recipient choices when the recipient choice is not revealed compared with when it is revealed. Similarly, is the average difference in the difference between migrant preferences and recipient choices when the migrant’s preferences are revealed to the recipient compared with when they are not revealed. If, as predicted, revealing the recipients’ choices to the migrants and communicating the migrants’ preferences to the recipients causes the recipients to make choices more similar to the migrants’ preferences, then the difference variable will be smaller in the choice-revealed and preference-revealed 18 treatment groups, and and should be negative. are stratification cell fixed effects (survey group and migrant treatment). Again, I also present specifications with control variables. I use the same variables as in the migrant experiment as well as recipient gender, age, education, household size, and the number of days between the migrant and recipient surveys. is the error term, which I adjust for heteroskedasticity. 19 5. RESULTS Migrant Experiment I first analyze the results of the migrant experiment in which migrants make an incentivized decision over how much of a potential $600 windfall to send to the recipient and how much to keep. Figure 5.1 shows the cumulative distribution of the amount sent by migrants, separately by treatment group. Because the experimental protocol limited migrants to splitting the money in $100 increments, the distributions are discrete. The first observation to be made from this figure is that the migrants send large amounts: More than half of the migrants in both treatment groups chose to send the entire $600 to the recipient. The other, smaller concentration in both distributions is at $300, where migrants decide to split the money equally between themselves and the recipient. Despite the fact that the two distributions follow the same basic shape, differences are evident. The percentage of migrants sending everything is smaller when choices are not revealed (53 percent versus 58 percent) and the percent of migrants choosing to send $400 or less is higher (44 percent versus 38 percent). It is also easy to see that the distribution of the choices in the choice-revealed treatment group is always below the distribution of choices in the choice-not-revealed group; that is, the choice-revealed distribution stochastically dominates the choice-not-revealed distribution. 14 Figure 5.1 Cumulative distribution of amount sent by migrant by treatment group Source: Author’s calculations. Notes: Sample is observations with nonmissing values for experiment questions and completed recipient survey. Choice not revealed: N = 648. Choice revealed: N = 650. The fact that almost all migrants in the choice-not-revealed treatment group chose to send something is consistent with the model presented in Section 2, where migrants who deviate when the probability their income will be observed is low still send positive amounts in remittances. The fact that most migrants in this group choose to send the entire $600 is suggestive that the altruistic component of remittances is high. However, the differences between the two distributions are evidence that information asymmetries also play a role. Migrants whose choices are not revealed are choosing to send less home. 14 The p-value on the two-sample Wilcox rank-sum for equality of distributions is 0.034. 0 .2 .4 .6 .8 1 C u m u la ti v e P r o b a b il it y 0 200 400 600 Amount sent Choice not revealed Choice revealed 20 These results are formalized in Table 5.1, which presents the results of estimating regression equation 1. Column 1 is a simple regression of the dependent variable on treatment status, and column 2 adds the demographic control variables. All regressions include stratification group fixed effects. The results are robust to the inclusion of control variables. Migrants send $20 more when their choice will be revealed, which represents an approximate 5 percent increase over the choice-not-revealed group mean. Table 5.1 Impact of monitoring treatment on migrant remittance decision Dependent variable: Amount sent by migrant Variable (1) (2) Migrant choice revealed to recipient 20.40** 19.50* [10.27] [10.26] Migrant is female –26.46** [11.09] Migrant age –0.487 [0.741] Migrant’s years of education –0.119 [1.225] Migrant’s years in the US 1.968* [1.071] Migrant lives with spouse –28.75** [11.83] Migrant’s hh size in US 1.293 [2.800] Migrant has child 22 or under in El Salvador 0.984 [12.41] Recipient is migrant’s close relative –0.675 [12.74] Migrant in lowest income bracket –21.73* [12.65] Migrant’s annual total remittances to recipient hh 0.00319* [0.00192] Migrant communicates with recipient hh at least weekly –1.122 [12.68] Observations 1,298 1,298 R-squared 0.133 0.159 Mean in treatment = Migrant choice not revealed to recipient 441.4 Source: Author’s calculations. Notes: hh = household. Robust standard errors in brackets. Sample is observations with nonmissing values for all experiment questions and completed recipient survey. Amount sent by migrant is the amount that migrants chose to send when splitting $600 between themselves and recipients. All regressions include stratification group fixed effects: dummy variables for the groups of survey numbers within which randomization was stratified. Recipient is defined as close relative if migrant reports recipient to be his spouse, parent, or child. Migrants in the lowest income bracket chose $400 or less as the weekly income of themselves plus their co-resident spouses. The other categories were $401–$600, $601– $800 and $801 and above. Annual total remittances are the combination of regular and irregular remittances. Annual regular remittances were collected by asking for the frequency of remittances sent and the average amount sent each time. Annual irregular remittances are remittances sent for special occasions or emergencies. *** p < 0.01, ** p < 0.05, * p < 0.1 21 Table 5.1 also reports the coefficients on the demographic control variables included in column 2. Five characteristics predict the migrant’s choice. Female migrants send on average $26 less than male migrants. Although women keep more on average than men, the effect of the treatment does not vary by gender (results not shown). Migrants who have been in the United States longer send more, although the effect is small. Migrants who live with a spouse send $29 less than those who do not. This is possibly because they are more likely to have their immediate family with them. Migrants in the lowest income bracket are estimated to send $22 less on average than those in the other income brackets. Finally, total annual remittances sent are positively correlated with amount sent in the experiment. The coefficient is small, but it suggests that migrant behavior in the experiment is related to real-world migrant behavior. The results in Table 5.1 show that information asymmetries can affect migrants’ remittance decisions and that at least some migrants take strategic advantage of a situation where the probability that their income will be observed is very low. The size of the effect (a $20 increase in amount sent) is not large, but it is similar to the size of the correlations with the demographic variables in Table 5.1. The effect size is also comparable to other experimental studies in families (Hoel 2013) and social networks (Leider et al. 2009; Ligon and Schechter 2012) that study the effects of making choices in dictator games known to the recipient. For example, Ligon and Schechter find that 91 percent of sharing in their experiment is related to altruistic motives. However, they also find that strategic behavior in their games predicts real-world strategic behavior, while altruistic behavior in the games does not predict any real- world activity. This suggests that strategic behavior may in fact be even more important outside of the experimental context than within it. The effect size can also be compared with other studies with experimental designs that are not as similar. Jakiela and Ozier (2012) estimate a 4 percent kin tax on income in an experiment where participants sacrifice returns on income in order to keep it secret. Goldberg (2011) estimates a 7 percent sharing tax on income in an experiment where she compares the spending of the winners of public lotteries with that of the winners of private lotteries. The framework presented in Section 2 predicts that if recipients cannot threaten to punish migrants, no differences between treatment groups should be observed. Several variables from the baseline survey can plausibly be thought to proxy for punishment costs, and I examine how the treatment effect varies by these variables. I do not have a perfect measure of these potential punishment costs (and certainly one would be hard to obtain); but by showing a consistent pattern with all five of these variables the argument that ability to punish is important is convincing. The five variables, the predicted relationship with punishment ability, and the rationale for choosing them are described below.  Migrant years in the United States (negative correlation): A migrant’s reputation at home is important for migrants who wish to return, and the probability of return may decline with length of time in the United States. With time it is also more likely that the migrant has paid off any debts related to his initial migration costs. The median number of years in the United States is 10.  Migrant has a child 22 or under in El Salvador (positive correlation): Migrants may have left children in the care of the recipient, and the possibility of child care that does not meet the migrant’s preferences could be a powerful tool; 34 percent of migrants have a son or daughter aged 22 or under in El Salvador. 15  Migrant and recipient are closely related (positive): This is defined as spouses or parent and child. Migrants may have entrusted recipients with the care of things that are important to them, and positive relationships with the recipients may be valuable to the migrants; 31 percent of migrants and recipients are closely related.  Migrant communicates with recipient household weekly (positive correlation): Frequent communication is a sign that migrants value their relationships with recipients; 71 percent of migrants report communicating weekly with the recipient household. 15 The 22-and-under cutoff is used because it was available on the survey that measured the number of young relatives up to college age the migrant had in El Salvador. 22  Remittances sent by migrant to recipient household (positive correlation): Because remittance relationships where recipients induce migrants to send money result in higher payments, higher remittances may indicate high punishment costs. The median annual remittance total to the recipient household reported by the migrant is $1,800. Although these variables are all plausible proxies for punishment costs, given that they generally indicate a stronger or closer relationship between migrant and recipient, arguably they may also be proxies for higher levels of altruism. It is true that in general punishment costs and altruism may be correlated, but as described in the framework, altruistic remittances should not be affected by variations in monitoring. In other words, if these variables were proxies for only altruism and not punishment ability, the treatment effect of monitoring in the high-altruism subgroups should not be higher than in low- altruism subgroups. Additionally, the mean amount sent in the group where decisions are not revealed is in every case lower in the high-punishment-cost subgroups than in the low-punishment-cost subgroups. Because payments in the choice-not-revealed group should be largely motivated by altruism, this is evidence that the variables chosen are representing more than just higher levels of altruism. Table 5.2 presents regression results by subsamples of these variables. For the continuous variables (years in the United States and remittances) the sample is split at the sample median, and for the binary variables (child in El Salvador, close relationship, and weekly communication) the sample is split according to the two values of the variable. The results are striking in that for each of these variables, the treatment effects are almost entirely concentrated in the subsample where punishment costs should be higher (columns 2, 4, 6, 8, and 10). These treatment effects are larger and more precisely estimated than in the full sample: Depending on the subsample, coefficients range from $32 to $59 more sent when the choice is revealed. These numbers are about 7–14 percent of the average amount sent in the choice-not- revealed group. In the subsamples where punishment costs should be low (columns 1, 3, 5, 7, and 9), the coefficients are mostly small and do not approach statistical significance. The table also reports the p- values on the test for equality of the treatment effects in the two subsamples for each of the five proxy variables. Two of the five coefficient pairs in both panels are statistically significantly different from each other. Alternate specifications that use the first principal component of the five proxy variables as a summary measure or interact the proxy variables with the treatment indicator in the full sample yield similar results. These results are consistent with the model’s prediction that when punishment costs are low, variation in the observability of migrant income will not affect migrant remittance decisions. An alternative explanation of the results in the migrant experiment is that migrants simply care about being perceived as altruistic and use the choice-revealed treatment to signal that altruism to their family. This concept of signaling altruism was developed by Bénabou and Tirole (2006). I cannot definitively rule this out, but the results suggest that it is unlikely. The strong patterns of heterogeneity by subgroup are directly connected to the framework presented here, but it is not obvious how they would relate to a story about signaling. To be consistent with the subgroup results, the variables that describe recipient punishment ability would also have to represent groups to which migrants care about appearing altruistic. However, across all subgroups, the migrants’ allocations in the not-revealed treatment group are high, suggesting that most migrants are altruistic. If the signaling story was true, it would then have to be the case that true altruism and the desire to signal altruistic behavior were not correlated. 16 16 Although the context is different, in an experiment studying the social networks of Harvard students, Leider et al. (2009) are able to definitely rule out the signaling explanation for non-anonymous giving in favor of one based on reciprocity and future interactions. 23 Table 5.2 Impact of monitoring treatment on amount sent by migrant: By proxies for punishment ability (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Dependent variable: amount sent by migrant Years in the United States Migrant has child 22 or under in El Salvador Recipient is close relative of migrant Migrant communicates with recipient household weekly Migrant’s annual remittances to recipient household Above sample median Below sample median No Yes No Yes No Yes Below sample median Above sample median Migrant choice revealed to recipient 19.83 32.75** 10.45 59.12*** 12.21 36.29* 9.414 24.97** –1.532 36.98** [15.33] [15.32] [12.79] [20.73] [12.78] [20.11] [22.03] [12.47] [15.91] [14.84] P-value for equality of treatment effect 0.551 0.040 0.318 0.573 0.077 Observations 639 656 853 445 889 402 376 919 611 670 R-squared 0.221 0.201 0.154 0.299 0.159 0.314 0.306 0.152 0.236 0.190 Mean in treatment = Migrant choice not revealed to recipient 457.8 426.1 449.3 424.3 444.4 433.9 449.4 438.4 447.0 437.5 Source: Author’s calculations. Notes: Robust standard errors in brackets. Sample is observations with nonmissing values for all experiment questions, completed recipient survey, and nonmissing values for variables used for division into subsamples. Amount sent by migrant is the amount that migrants chose to send when splitting $600 between themselves and recipients. All regressions include stratification group fixed effects: dummy variables for the groups of survey numbers within which randomization was stratified. Recipient is defined as close relative if migrant reports recipient to be his spouse, parent, or child. Annual total remittances are the combination of regular and irregular remittances. Annual regular remittances were collected by asking for the frequency of remittances sent and the average amount sent each time. Annual irregular remittances are remittances sent for special occasions or emergencies. The median years in the United States is 10 and the median remittances sent to the recipient household are $1,800. *** p < 0.01, ** p < 0.05, * p < 0.1 Recipient Experiment Analysis of the migrant experiment found that migrants react strategically to variations in the ability of recipients to monitor their income. Previous literature examining information asymmetries in remittance behavior has suggested that migrant monitoring of recipients should also be important (de Laat 2008; Ashraf et al. 2011; Chen 2013). To look for these effects in the context of this experiment, I now turn to analysis of the recipient experiment in which recipients allocated a potential $300 remittance prize among four different spending categories. Mean amounts allocated to different spending categories by recipients and migrants are presented in Table 5.3. The first two columns show the mean amounts allocated by recipients broken down by the recipient-monitoring treatment, and columns 3 and 4 show mean recipient allocations by the recipient communication treatment. The fifth column shows the means of the preferences reported by the migrant. Across both recipients and migrants, education is the most popular choice. 17 Daily expenses are the next 17 The preference for education may be partly due to the fact that participants answered this question at the conclusion of a survey that was rather heavily focused on questions about education, meaning that they may have been primed to consider education. This is not necessarily a problem as there is no reason to believe that either migrants or recipients were more primed 24 most popular category, followed closely by health and finally restaurant meals. Migrants allocate less to education than recipients and more to daily expenses, health expenses, and restaurant meals, but no strong evidence indicates that migrants on average prefer different expenditure categories than recipients. Table 5.3 Mean amounts allocated to spending groups by recipients and migrants: Recipient experiment Means of recipient choices by treatment group: Means of migrant preferences Monitoring treatment Communication treatment Recipient choice not revealed to migrant Recipient choice revealed to migrant Migrant preferences not revealed to recipient Migrant preferences revealed to recipient Amount allocated to Restaurant meals 6.11 5.46 5.38 6.17 11.74 Education 175.54 166.22 170.97 170.64 141.41 Daily expenses 66.05 75.59 72.85 68.99 76.56 Health expenses 52.30 52.73 50.80 54.20 70.28 Observations 638 660 641 657 1298 Source: Author’s calculations. Notes: Sample is observations with nonmissing values for all experiment questions and completed recipient survey. Means in columns 1 through 4 are from responses by recipients when asked to allocate $300 across four spending categories. Means in column 5 are responses from migrants when asked how they would like the recipient to allocate the funds. I utilize the data collected from both the migrant and the recipient to analyze how the treatments affect the pair-level differences between their choices. Table 5.4 shows the results from estimating regression equation 2. The dependent variables in columns 1 through 4 are the migrant–recipient differences in restaurant spending, education spending, spending on daily expenses, and health spending, respectively. The dependent variable in columns 5 and 6 is the total migrant–recipient difference. Column 6 adds demographic control variables. The control variables are the same as those presented in Table 5.1 with the addition of recipient gender, age, years of education, household size, and a control for number of days between the migrant and the recipient surveys. than the other. 25 Table 5.4 Impact of monitoring and communication treatments on recipient allocation decision (1) (2) (3) (4) (5) (6) Dependent variable: migrant–recipient difference in… Dependent variable: Treatment Restaurant spending Education spending Daily expenses spending Health spending Total migrant–recipient difference Recipient choice revealed to migrant –0.999 4.287 4.001 –1.251 3.019 3.330 [2.141] [4.922] [4.206] [4.234] [4.775] [4.827] Migrant preference revealed to recipient –2.741 –14.18*** –6.154 –4.137 –13.61*** –13.69*** [2.119] [4.936] [4.228] [4.196] [4.753] [4.818] Observations 1,298 1,298 1,298 1,298 1,298 1,298 R-squared 0.101 0.104 0.092 0.091 0.105 0.122 Mean in recipient choice not revealed 15.9 107.3 78.0 75.0 138.1 Mean in migrant preference not revealed 16.7 116.3 83.0 76.6 146.2 Control variables NO NO NO NO NO YES Source: Author’s calculations. Notes: Robust standard errors in brackets. Sample is observations with nonmissing values for all experiment questions and completed recipient survey. Dependent variables are the absolute difference between the recipient’s choice and the migrant’s preferences in each category. The total difference is the sum across the four difference variables for each observation, divided by 2. All regressions include stratification group fixed effects: dummy variables for the groups of survey numbers within which randomization was stratified and treatment status in the migrant experiment. Control variables are migrant and recipient gender, age, years of education, and household size. Controls also include migrant years in the United States, whether migrant lives with spouse, whether migrant has a child 22 or under in El Salvador, whether the migrant and recipient are close relatives, if the migrant is in the lowest income bracket, annual total remittances to recipient household, whether the migrant and recipient communicate at least weekly, and the number of days between migrant and recipient surveys. *** p < 0.01, ** p < 0.05, * p < 0.1. For both the monitoring and the communication treatments, the prediction is that the difference will be smaller in the revealed treatment group. When the probability is high that spending choices will be observed or when recipients are well informed about migrant preferences, recipients should more greatly adhere to those preferences. This prediction is not borne out for the monitoring treatment. For all spending categories and the total difference across specifications, the coefficients for the monitoring treatment are small in magnitude and statistically insignificant. However, differences are evident for the communication treatment. For all spending categories, the coefficient on the variable indicating that the migrant preference was revealed to the recipient is negative. This means that the differences between recipient choices and migrant preferences are smaller when the migrants’ preferences are revealed than when they are not. Of all the spending categories, only the difference for education (column 3) is statistically significant, but importantly, so is the total difference (columns 5 and 6), implying that migrant and recipient choices are getting closer together overall. Across specifications the results indicate that there is a $14 reduction in the total difference when the migrant’s preferences are revealed. This represents about a 10 percent reduction relative to the mean in the preference-not-revealed group. The effect is driven by the difference in education spending, with the corresponding reductions in the differences in other categories being split between daily and health expenses and, to a lesser extent, spending on restaurant meals. 26 I also examine the interaction of the two treatments (results not shown) and find that the impact of the communication treatment does not vary with the monitoring treatment. There is also no consistent pattern of variation when the results of the monitoring treatment are examined by punishment ability (results not shown). This is not surprising given the very small estimated coefficients in the in the full sample. The impact of the communication treatment suggests that migrant preferences do matter to recipients and that some deviation from those preferences may be inadvertent. However, the lack of impact of the monitoring treatment further implies that recipients do not react strategically to changes in the probability of detection. The framework presented in Section 2 proposes an explanation for why recipients may not take advantage of the opportunity to hide their spending choices from migrants. Migrants simply may have limited ability to punish the recipients for not following their preferences. While the results in the migrant experiment are that recipient ability to punish varies across recipients, these results from the recipient experiment suggest that, in this context, migrant ability to punish is low across the population. In practice, this would result in a situation where the migrants have very little power to compel recipients to spend the remittances as they wish. Although a limited ability to punish is the explanation for the lack of effect of the monitoring treatment that is suggested by the framework, it is important to consider other possible explanations. The first alternative explanation is that migrant monitoring of recipients is essentially perfect and that recipients know that their choices will be discovered if they win. However, some information collected on the baseline survey suggests that this is not the case. Table 4.1 provides summary statistics of some questions that aimed to measure migrant knowledge of the recipient household. Only 24 percent of migrants could correctly report student grade point average (GPA), and 43 percent correctly report how students travel to school; therefore it does not seem plausible that existing monitoring is good enough across the board to render the experimental variation irrelevant. A second explanation is that migrants and recipients have the same preferences for spending, and therefore they make the same choices regardless of punishment ability. This may be true for some families, but if it were true for most, there should be no impact of the communication treatment. Additionally, only 48 percent of migrant–recipient pairs report the same three budget priorities (Table 4.1), further evidence of heterogeneity in preferences. A potential criticism of the results in the communication experiment is that recipients respond to the information about migrant preferences not because they necessarily want to follow migrant preferences but because they are reacting to being given a suggested allocation for the choice they are making. In other words, recipients may have reacted in the same way even if the suggested preferences were attributed to someone besides the migrant. I can address this concern by examining heterogeneity in the effects of the communication treatment by proxies for the quality of information in the relationship. Specifically, in Table A.2 in the Appendix, I estimate regression equation 2 separately by whether or not the migrant can correctly report the student’s GPA and mode of transport to school. Although these variables are not direct representations of recipient knowledge of migrant preferences, they are likely to indicate low information quality in general. If recipients are reacting to a lack of knowledge of the migrants’ preferences, the effects of the communication treatment should be concentrated among pairs where these variables suggest that information quality is low. I find that this is indeed the case: Effects of the communication treatment are evident only among migrants who do not know students’ GPAs or modes of transportation to school. Discussion Economic studies of information asymmetries in households with migrants have until now focused on migrant monitoring of recipient behavior (Chen 2013; de Laat 2008) and the impacts of offering migrants greater control over how remittances are spent (Ashraf et al. 2011; Chin et al. 2011; Torero and Viceisza 2011). This is the first study that explicitly looks at the effect of information asymmetries on both sides of the remittance relationship—migrants’ sending of remittances as well as recipients’ spending of those remittances. Despite the previous emphasis on migrant monitoring, the results of the two monitoring 27 treatments presented in this paper are that, in this context, only migrants, and not recipients, strategically react to variations in the probability that their actions will be monitored. This is an important finding, not only because it shows that information asymmetries have an important impact on the remittance sending decision, but also because this implies that recipients have important power in the migrant–recipient relationship. Although this influence has been considered in the extensive literature on the motivations to send remittances, it has not previously been rigorously documented empirically. Policymakers who seek to design tools to facilitate the sending of remittances and enhance their impacts18 should consider the role of the recipient in determining remittance amounts. Policies that assume that migrants have complete autonomy over the manner in which remittances are sent may fall short of their full potential. The analysis in this paper indicates that because migrants are responding to the opportunity to hide income, some of them are already sending home more than they would choose to voluntarily. This implies that programs that seek to increase remittances further will face difficulties within this group. Policymakers should also consider the welfare implications of such a policy. The pressure that migrants face to send remittances is related to a growing literature on sharing pressures in family networks in developing countries. Recent work by di Falco and Bulte (2011) and Jakiela and Ozier (2012) suggests that the expectation that resources will be shared with extended family may inhibit individual economic progress. In this context, particularly given the low income status of the migrants, the results indicate the possibility that any potential extra remittance funds may be more efficiently used by the migrants in the United States than by their family members at home. Additionally, the fact that the monitoring treatment had no effect on the recipients’ spending decision adds a new angle to the recent work on the impact of control on remittance behavior (Ashraf et al. 2011; Chin et al. 2011; Torero and Viceisza 2011). To varying degrees, these studies offer migrants direct control over money sent to family members at home. Viewed through the framework presented in this paper, control over remittances improves both the monitoring and enforcement of remittance- spending contracts, but the existing studies are not able to distinguish between the two channels. The results of this study—that migrant monitoring of recipient remittance spending does not seem to matter— suggest that if migrants do desire control (and the literature is mixed on whether or not they do), this desire may not be due to an inability to monitor the recipients but rather to an inability to effectively punish recipients and therefore compel recipients to spend remittances in a certain way. In the absence of punishment ability, the ability to control would act as the enforcement mechanism in the migrant– recipient contract. Overall, the findings that information asymmetries can affect both the sending and spending of remittances suggest that interventions or technological innovations that improve communication in transnational households could have important effects on financial decisions made by both migrants and recipients. In particular, the results of the communication experiment imply that for migrants who wish to change the spending behavior of their family members, policies that improve communication about spending preferences may be an inexpensive way to achieve a higher level of compliance with their preferences. Although this study addressed only the inadvertent impacts of information asymmetries in the context of preferences for the spending of remittances, it is possible that improving communication could also alleviate other possible inadvertent effects of information asymmetries in the migrant–recipient relationship. For example, if recipients do not have a full understanding of migrants’ cost of living in the United States, improving that knowledge could lead recipients to expect lower remittance payments. 18 See Yang (2011) for a discussion. 28 6. CONCLUSION This paper analyzes a set of experiments designed to test for the effects of information asymmetries in transnational households. First, an experiment among Salvadoran migrants in the Washington, DC, area examines the extent to which the probability that recipients will observe migrant income is a factor in remittance decisions. The migrant’s remittance decision is modeled as a combination