66523 S P D I S C U S S I O N PA P E R NO. 1120 Employment Generation in Rural Africa: Mid-term Results from an Experimental Evaluation of the Youth Opportunities Program in Northern Uganda Christopher Blattman Nathan Fiala Sebastian Martinez December 2011 Employment Generation in Rural Africa: Mid-term Results from an Experimental Evaluation of the Youth Opportunities Program in Northern Uganda Christopher Blattman Nathan Fiala Sebastian Martinez Yale University DIW Berlin IADB1 December 2011 1 Christopher Blattman (corresponding author): Yale University, Departments of Political Science & Economics, 77 Prospect Street, New Haven, CT 06511, (203) 432-3347, christopher.blattman@yale.edu; Nathan Fiala: German Institute for Economic Research, 10108 Berlin, Germany, nfiala@diw.de; Sebastian Martinez: Inter American De- velopment Bank, Office of Strategic Planning and Development Effectiveness, 1300 New York Avenue, NW, Washington DC 20577, (202) 623-1000, smartinez@iadb.org. Abstract: Can cash transfers promote employment and reduce poverty in rural Africa? Will lower youth unemployment and poverty reduce the risk of social instability? We experimentally evaluate one of Uganda’s largest development programs, which provided thousands of young people nearly unconditional, unsupervised cash transfers to pay for vocational training, tools, and business start-up costs. Mid-term results after two years suggest four main findings. First, despite a lack of central monitoring and accountability, most youth invest the transfer in vocational skills and tools. Second, the economic impacts of the transfer are large: hours of non-household employ- ment double and cash earnings increase by nearly 50% relative to the control group. We estimate the transfer yields a real annual return on capital of 35% on average. Third, the evidence suggests that poor access to credit is a major reason youth cannot start these vocations in the absence of aid. Much of the heterogeneity in impacts is unexplained, however, and is unrelated to conven- tional economic measures of ability, suggesting we have much to learn about the determinants of entrepreneurship. Finally, these economic gains result in modest improvements in social stabili- ty. Measures of social cohesion and community support improve mildly, by roughly 5 to 10%, especially among males, most likely because the youth becomes a net giver rather than a net tak- er in his kin and community network. Most strikingly, we see a 50% fall in interpersonal aggres- sion and disputes among males, but a 50% increase among females. Neither change seems re- lated to economic performance nor does social cohesion—a puzzle to be explored in the next phase of the study. These results suggest that increasing access to credit and capital could stimu- late employment growth in rural Africa. In particular, unconditional and unsupervised cash trans- fers may be a more effective and cost-efficient form of large-scale aid than commonly believed. A second stage of data collection in 2012 will collect longitudinal economic impacts, additional data on political violence and behavior, and explore alternative theoretical mechanisms. JEL Classification: I38, J24, D13, O12 Keywords: poverty alleviation, cash transfers, vocational training, Uganda Acknowledgements: We thank Uganda’s Office of the Prime Minister, the management and staff of the Northern Uganda Social Action Fund, and Patrick Premand and Suleiman Namara of the World Bank for their contributions and collaboration. For comments we also thank Bernd Beber, Kelly Bidwell, Pius Bigirimana, Ariel Fiszbein, Louise Fox, Julian Jamison, Robert Limlim, Mattias Lundberg, David McKenzie, Suresh Naidu, Obert Pimhidzai, Patrick Premand, Josefina Posadas, Sam Sakwa, Alexandra Scacco, and numerous seminar participants. We gratefully acknowledge fund- ing from the World Bank’s Spanish Impact Evaluation Fund (SIEF), Gender Action Plan (GAP), the Bank Netherlands Partnership Program (BNPP), Yale University’s ISPS, and appreciate sup- port from the Africa Impact Evaluation Initiative, the Office of the Chief Economist for Human Development and the SIEF Active Labor Market Cluster. Finally, Filder Aryemo, Mathilde Emeriau, Lucy Martin, Ben Morse, Doug Parkerson, Pia Raffler, and Alexander Segura provided superb research assistance through Innovations for Poverty Action (IPA). All findings and inter- pretations in this paper are those of the authors, and do not necessarily represent the views of the Government of Uganda or the World Bank, Executive Directors or the governments they repre- sent. 1 Introduction In the U.S. and Europe, governments channel huge sums towards employment programs to re- lieve poverty, spur growth, and bolster political support. In developing countries, governments invest in employment and anti-poverty programs with additional motives in mind: to strengthen the sense of citizenship and civic action, and to lessen the risk of social instability. Roughly two billion people, nearly a third of the world population, are between the ages of 15 and 34 and live in a developing nation.1 This proportion is continuing to rise and will peak in coming years, creating a global “youth bulge� (World Bank 2007). Fears are bulging even faster. A shortage of educational and job opportunities may heighten inequality and slow poverty alle- viation. Moreover, policymakers, the media, and many social scientists worry this bulge of un- deremployed youth will weaken community and societal bonds and heighten social unrest, in- cluding (in extreme cases) crime, riots, and even armed conflict and terrorism.2 To reduce both poverty and instability, policymakers and pundits commonly propose gov- ernment or aid-funded employment interventions, from finance to skills (e.g. Kristof 2010; World Bank 2010). A new breed of decentralized, participatory development programs provides cash or other resources to communities and groups, and allows them to decide how to best use funds. These programs go by different names—social action funds, or community-driven devel- opment programs—but are an increasingly common tool of governments and aid agencies. Some of the best known disburse aid to communities for infrastructure or other projects, but uncondi- tional cash transfers are an increasingly common means of spurring employment and enterprise development among the poor. This paper describes the impacts of a participatory state-supported employment program in Uganda, the Youth Opportunities Program (YOP) component of the Northern Uganda Social Ac- tion Fund (NUSAF), which provided relatively unconditional cash transfers to small groups of young men and women to help them start new vocations and enterprises. In the least developed nations, where firms are rare, aid-based employment interventions commonly provide inputs into self-employment—cash, microfinance, or in-kind skills training or business assets. Such pro- 1 Based on U.S. Census Bureau international population data: http://www.census.gov/ipc/www/idb/worldpop.php. 2 (Kaplan 1994; Fuller 1995; Goldstone 2002; Heinsohn 2003) 1 grams are rooted in at least three assumptions. First, poor people have agency and are capable of making informed economic decisions. Second, the poor have high returns to human and physical capital, often because of a market failure, such as credit constraints. Third, anti-poverty pro- grams, especially participatory ones, will produce more engaged, less alienated and less violent citizens. Evidence for all three propositions remains limited. Take the first belief: From a purely prac- tical standpoint, giving a group of young people a lump sum of cash worth several times their annual earnings, with limited supervision, and expecting them to invest it wisely, is at best a risky development strategy. It is a policy approach criticized both generally and in the case of Uganda (Golooba-Mutebi and Hickey 2010; Hickey 2010). A growing body of research in be- havioral economics highlighting time inconsistency and limited rationality heightens concern. There is some evidence for the second belief. While the number of data points remains small, there is a growing sense that the poor have high returns to cash and in-kind physical capital. Economic theory and some experimental evidence suggest that these returns go unrealized be- cause the poor have little capital of their own to invest and limited access to credit (Banerjee and Duflo 2005; Udry and Anagol 2006; de Mel et al. 2008; Banerjee et al. 2010). There are two rea- sons to be cautious, however. One is that the evidence on high returns and market failures is pre- liminary: the number of studies is small; they deal with particular populations; and the evidence comes largely from observational analysis of heterogeneous treatment responses. While optimism pervades research on physical capital, the research on returns to human capi- tal investments is less encouraging. Job training programs in developed nations have generally low impacts.3 Business skills and financial literacy training, which are more common in develop- ing countries, appear to yield only modest returns (Field et al. 2010; Karlan and Valdivia 2011). Technical and vocational training is even more common, representing almost $3 billion in devel- opment assistance from 1990 to 2005—about 7.5% of all education-related aid (World Bank 2010). Here the evidence is especially thin. Three evaluations of job training programs in mid- dle-income Latin American nations suggests mixed impacts overall and little impact on poor 3 After dozens of evaluations, meta-analyses conclude that job training programs have modest impacts, are some- times harmful, and seldom pass an economic cost-benefit test (Heckman et al. 1999; Betcherman et al. 2007; Card et al. 2009). Nearly all the underlying studies concern industrial economies, few are experimental, few try to explain heterogeneity in performance, and almost none explore social-political impacts and related externalities. 2 males—the most worrisome population from the perspective of social stability.4 Almost none of these studies examine programs to self-employment, however, which is the main basis of em- ployment in the least developed countries. To our knowledge, there have been no rigorous evalu- ations of vocational training and employment programs in the least developed nations. Finally, the theory and evidence on the third belief—from poverty to lower alienation and ag- gression—is especially uncertain, though not for lack of theory. We review competing theories that argue for a link from employment programs (and higher incomes) to greater social cohesion, reduced alienation, and lower aggression and potential for instability. Instrumentalist and eco- nomic theories of crime and conflict argue that higher incomes and employment raise the oppor- tunity cost of aggression and predatory activities. A large body of psychology, sociology and po- litical science emphasizes that aggression arises from stress, adversity and frustrated ambitions, each of which may be accentuated by poverty, inequality, and economic marginalization (and hence mitigated by successful employment programs). Field evidence for any of these theories, however, is scarce. We look at the evidence for all three propositions through a randomized evaluation of a state development program in northern Uganda, a region just emerging from economic stagnation and political insecurity, including insurgency, banditry, and wars in neighboring states. This paper reports the mid-term (two year) results, with final longitudinal results to become available in 2012. In 2008 the program provided cash transfers to thousands of young men and women for in- vestment in skills training and capital for self-employment. The focus of the program was voca- tional training and employment, and applicants were required to form a group of roughly 15 to 25 young adults interested in a vocation and submit a proposal for purchasing skills training, tools, and other materials required to start an enterprise. On average, successful groups received a lump sum cash transfer of $7,108 to a jointly held bank account—roughly $374 per group member, at market exchange rates. Groups were otherwise free of supervision or oversight in the actual spending. Not surprisingly, demand for the program far outstripped supply of funds: hun- dreds of groups, representing tens of thousands of young adults, applied. 4 Rigorous evaluations in Argentina and Colombia found significant impacts on female employment only (Aedo and Nuñez 2004; Attanasio et al. 2008) while the positive impacts of a training program in the Dominican Republic comes mainly from highly-educated workers (Card et al. 2007). 3 Given excess demand for program funds, we worked with the Government to allot 535 groups randomly to treatment (the transfer) or control. We follow a random subset of treatment and con- trol group members over two years. The mid-term economic and social impacts are substantial. Our results show that program beneficiaries make good use of the transfers. Groups spend the majority of their transfer on skills training fees and durable assets, with the remainder for materials, consumption, transfers and savings. Nearly 80 percent of the ‘treated’—those in groups who receive the government cash transfer—enroll in vocational training, and they acquire and grow business assets. We see some evidence of capture of transfers by the group leaders, but the capture is small and not very signif- icant, and non-leaders still earn substantial returns. Moreover, the program has large and significant effects on employment and income. Both men and women increase their hours in employment outside the home—by about 25% among males and by 50% among females. Two years after the transfer, roughly two-thirds of the treated are engaged in skilled work, compared to just over one-third of controls. Finally, economic re- turns are almost uniformly positive, and are relatively high for a majority of beneficiaries. The average beneficiary increases net income by about $9 per month, a nearly 50% increase over the control group, representing real returns of roughly 35% per annum. These returns are higher than the real prime lending rate (5%) and higher than real commercial lending rates to small and me- dium enterprises (15 to 25% per annum) but lower than the 200% annualized rate available from microfinance institutions or moneylenders. Why were these returns not realized without the program? A growing body of theory and evi- dence argues that poor entrepreneurs are constrained by imperfect markets (especially inadequate access to credit, alongside fixed start-up costs to self-employment) and imperfect decision- making (such as self-control problems in spending and saving, or an absence of future focus in general). We develop a simple model that predicts how, under severe credit constraints, YOP- like investments and returns should vary with starting capital, entrepreneurial ability, patience levels, and existing vocations. We have detailed pre-intervention data on ability, access to credit, starting capital, and existing enterprises. The resulting patterns of heterogeneity are consistent with the idea that investments and returns increase with patience, and that the impacts of cash transfer programs are greatest for the poorest and those without existing vocations. We see no evidence that cognitive ability or formal schooling influence success, however, suggesting that, if “entrepreneurial ability� exists, it is made of different matter. 4 Finally, this increase in income and wealth leads to modest improvements in community par- ticipation, social integration, and male aggression. The results are most consistent with psycho- logical and anthropological accounts of market success and alienation and aggression. Program participation leads to lower levels of psychological stress, as well as increased wealth and ability to provide transfers within and outside the household. Social status increases, stress diminishes, and aggression falls, at least among males. Our analysis of aggression and social alienation also produces puzzles, however, such as elevated female aggression, and the absence of a correlation between actual economic performance and aggression for either gender. Both are to be explored in future research, including the 2012 round of data collection. Overall, the results support a strong role for public and aid-based financing for poor entrepre- neurs and employment creation, and suggest that relatively unsupervised and unconditional cash grants, which are cheaper to implement, will also be effectively and responsibly used. 2 Context: Northern Uganda Uganda is a small, landlocked East African nation. While once a classic example of the dys- functional African state, growth took off in the late 1980s with the end to a major civil war, a stable new government, and reforms that freed markets and political competition. The economy grew an average of 7% per year from 1990 to 2009. By the end of this period national income per capita was 8.5% ahead of the sub-Saharan average (World Bank 2009). Growth, however, has concentrated in southern and central Uganda. The north, home to roughly a third of the population, has lagged behind. Northern Uganda was once the home of the nation’s political and military elite, as well as a bread basket for the country, and hence wealthy relative to the rest of the country (Omara-Otunnu 1994). Since the 1980s, however, northern Uganda has held less political influence, received fewer public investments, and has been plagued by insecurity. In the north-central region, an insurgency displaced millions and de- stroyed assets and production from 1987 to 2006. The northwest and northeast were less affected by rebels, but were subject to other dangers. Conflicts in neighboring south Sudan and Demo- cratic Republic of Congo (DRC) fostered insecurity in the northwest, while cattle rustling and heavily armed banditry persisted in the northeast (Lomo and Hovil 2004). In 2003 peace came to Uganda’s neighbors, South Sudan and (to some extent) the DRC. Their demand for Ugandan products boomed. The Government of Uganda also accelerated efforts to 5 pacify, control, and develop the north. By 2006, the Ugandan military pushed the rebels out of the country, began to disarm northeastern cattle-raiders, and increased security and political con- trol. The north was peaceful, but sustained peace would require catch-up with the rest of the country in terms of economic opportunities and infrastructure.. With the arrival of democracy in the 1990s and multiparty competition the following decade, the government also began to build political coalitions with northern leaders, encouraged reconciliation with and reintegration of the disaffected and increased public spending. A national Peace, Recovery and Development Plan (PRDP) set ambitious economic and secu- rity goals in the north (Government of Uganda 2007). The centerpiece of this plan was a decen- tralized development program, NUSAF, the country’s second-largest development program at the time. Starting in 2003, communities and groups could apply for government transfers for in- frastructure construction or income support and livestock for the ultra-poor. Increasing the num- ber, size and productivity of informal enterprises was also a major policy priority, since the growth of the labor force greatly exceeds the absorption capacity of Uganda’s formal sector (World Bank 2009). To stimulate such employment growth, in 2006 the government announced a new NUSAF component: the Youth Opportunities Program (YOP), which provided cash trans- fers to groups of young adults for self-employment in trades. 3 The intervention and experiment 3.1 The intervention With YOP the government had two main aims: raise youth incomes and employment; and improve community reconciliation and reduce conflict. The program required young adults from the same town or village to organize into groups and submit a proposal for a cash transfer to pay for: (i) fees at a local technical or vocational training institute of their choosing, and (ii) tools and materials for practicing the craft. The program was targeted to poor and underemployed “youth�—roughly ages 16 to 35 in lo- cal terms. Since technical and vocational schools typically require some education and aptitude, YOP targeted poor youth who had the minimum capacity to benefit from vocational training, and so are not the very poorest. On average, applicants were just slightly wealthier and more educat- 6 ed than the average Ugandan5, but are still poor by any reasonable standard: the average appli- cant reported weekly cash income of 7,806 Ugandan Shillings (UGX), about US$4 at 2008 mar- ket exchange rates (1,800 UGX to the dollar), or almost exactly at the PPP$1.25 international poverty line.6 More than a quarter had not finished primary school. A fifth were engaged in semi- skilled or capital intensive employment and more than two-fifths reported no income or em- ployment in the past month. Like many participatory development programs, the objective was not only to enrich but also to empower young adults. Groups were responsible for selecting a management committee of five members, choosing the skills and schools, and budgeting, allocating, and spending all funds. Groups self-organized, or were spurred by a facilitator. Such facilitators, often a community leader or local government employee, helped groups identify projects and trainers, budget, and apply, but played no formal role after the proposal was submitted. The group management com- mittee and members were wholly responsible for disbursement and purchases, accountable only to one another. If a group was selected, the government transferred cash in a single tranche to a bank account in the names of the group leadership, with no further supervision. Thousands of groups applied and hundreds were funded to YOP from 2006 to 2008. Roughly half the groups existed prior to the NUSAF program, as sports or religious or community youth clubs. The rest were formed in response to the call for proposals, organized by group executives or community facilitators. In 2008, the government determined that it had funding for 265 of 535 eligible groups. The average group had 22 members, and 80% of groups ranged from 13 to 31 members in size, ac- cording to pre-intervention group rosters (Table 1). Group cash transfers averaged nearly UGX 12.8 million ($7,108), and varied not only by group size but by group request (i.e. transfers were 5 We compare 2008 baseline data on the eligible population of youth (described below) to representative household surveys: the 2004 Northern Uganda Survey (NUS), the 2006 Demographic Health Survey (DHS), and the 2006 Uganda National Household Survey (UNHS). Among youth eligible for the program, 93% had completed some primary school, 45% completed some secondary, and only 7% had no education. Compared to their age cohort in Uganda, they were four times more likely to have had some secondary and 15 times less likely to have no education. They are also more likely to own assets like mobile phones and radios, implying greater wealth. 6 The application and review process was inherently selective. Youth who self-selected into the program may be more motivated than the average youth, and may have above average aptitude for skilled vocations. The local and district officials who selected the projects may have been influenced by political or personal ties to the community or the group members, or opportunities for financial gain. These sources of selection are unobserved, but important for understanding external validity. In general, the program reached a huge number of youth with a breadth of skills, means, and war experiences, and impacts and patterns probably apply quite broadly . 7 not uniform). The average transfer size was UGX 673,026 ($374) per member—more than 20 times the average monthly income of the youth at the time of the baseline survey. Given the vari- ation in group size and requests, however, transfer size per official group member varied from UGX 200,000 to more than 2 million across groups. Figure 1 displays the distribution of trans- fers in US dollar equivalents. Assuming no additional persons were added after the transfer, the majority received between UGX 350,000 ($200) and 800,000 ($450). 3.2 Experimental design 3.2.1 Treatment assignment NUSAF received many times more applications than could be funded, and so the government decided to allocate final disbursements randomly among eligible groups.7 Funding was stratified by district, and 13 of 18 districts had sufficient YOP funds to participate in an experimental study.8 Unfortunately, non-participating districts include the three most civil war-affected dis- tricts that may have benefited most from the program: Gulu, Kitgum and Pader. Other districts affected less intensely by the insurgency were included. The central government asked district governments to sift through their (usually vast) pool of existing applications and nominate two to three times as many group applications as there was funding. From this pool the central government screened and audited applications, including physical verification of the groups, to confirm existence and eligibility.9 The authors received a 7 We also attempted to design a second randomization, one that treated a third of the treatment groups with an addi- tional cash balance (worth 2% of the total grant) to hire back their facilitator (or another of their choosing) to help them plan and manage the grant. In another third of groups, the funds would be transferred to the district govern- ments and they would be asked to provide those extension services directly. Our data indicate that this additional design was not properly implemented, and there is no difference in the use of post-grant facilitation across the two types of treatment and the control group. We omit further discussion of this element of the design from this paper. 8 We use the original 2003 NUSAF districts. Many districts were subdivided after 2003. 9 Applications were screened by several levels of government. A village or town leader had to approve and pass along applications to the District authorities, sometimes executively and other times through a participatory commu- nity process. District authorities reviewed applications and nominated projects to the central government. The cen- tral NUSAF office verified the existence of the group and reviewed proposals for completeness and compliance. At the central level, applicant groups were eligible if members were mainly of this age range, at least one-third female, had roughly 15 to 30 members, and if their application was accurate and complete. 8 list of the 535 screened groups and randomly assigned 265 groups (5,460 individuals) to treat- ment and 270 groups (5,828 individuals) to control, stratified by district.10 Despite the scale of the program, we judge spillovers to be unlikely. The 535 eligible groups were spread across 454 towns and villages, in a population of more than 5.4 million. 3.2.2 Treatment compliance We define treatment compliance fairly narrowly: all individuals in the group are coded as treated if the group received a funds transfer (according to administrative records) and if those funds were not diverted or stolen by district officials (according to a post-treatment survey of group members). We consider other forms of “compliance�, such as using the funds for skills training, or equitable distribution, to be intermediate outcomes of study rather than treatment in- dicators, and discuss them in the results section. In total, 30 groups did not receive funds, for a treatment compliance of 89%. 22 of these groups could not access government funds due to un- satisfactory accounting, complications with their bank account, or delays in collecting the funds. 8 groups reported that they never were given access to the funding due to the intercession of a local official. To our knowledge, no “ghost� groups—fictional groups invented by local leaders used to steal funds—were funded. 3.2.3 Average treatment effect (ATE) estimation Given the small and unsystematic treatment non-compliance, our preferred ATE estimator is a treatment-on-the-treated (TOT) estimate using assignment to treatment, Aij, as an instrument for treatment Tij for individual i in community j: Y1ij = θTij + �Y0ij + βXij + αij + εij (1a) Tij = πAij + �Y0ij + δXij + αij + εij (1b) where Y1ij denotes an outcome variable and Y0ij is its baseline level. This approach (the AN- COVA estimate) is more efficient than a difference-in-difference estimator (Frison and Pocock 10 Each district had a fixed budget. The 535 groups were sorted using a pseudo-random number generator in Mi- crosoft Excel 2003, stratified by district. Applicant groups were awarded funding until the pools of available re- sources for that district were exhausted. All other projects remained unfunded and were assigned to the control group. Within districts, 30 to 60% of applications were assigned to treatment. All analysis includes district dummies. 9 1992; McKenzie 2011). Xij is a pre-specified (optional) set of baseline covariates (principally used to correct for covariate imbalance after random assignment), αij is a stratum fixed effect, and εij denotes the error term. The ATE estimate is θ. In the end, different estimators—an inten- tion-to-treat estimate, or one calculated by differences-in-differences—have little material effect on the findings and conclusions (results not shown). 4 Economic theory and intended impacts 4.1 When will transfers boost employment and income, and for whom? The simplest interpretation of the intervention is that it provides cash to entrepreneurs for in- vestment in human and physical capital. To understand why transfers might boost employment and incomes (and for whom), it’s useful to remember that, when credit and insurance markets function reasonably well, transfers to the poor will reduce poverty but they will not lead to in- vestment, enterprise, and earnings. 4.1.1 Cash transfers and the unfettered entrepreneur Consider a simple model of household (entrepreneurial) production with entrepreneurs who can borrow freely and are either risk neutral or can insure themselves against risk (See Bardhan and Udry 1999 for simple examples). These unfettered entrepreneurs will choose their stock of capital (human or physical) so that the marginal return to capital equals the market interest rate. Further investment would push the marginal return below the market interest rate. Given a cash windfall, the entrepreneur would consume some now and save the rest for future consumption. As for employment, labor levels might even decrease—if leisure is a normal good, wealthier en- trepreneurs will consume more of it. If the windfall arrives as in-kind capital, or on the condition that it is invested, entrepreneurs would be forced to invest below the market rate of return. In the short run, earnings and em- ployment would rise. But rational entrepreneurs would be worse off than if they received cash, and over time they would draw down their investment until they reach the earlier equilibrium. 4.1.2 Imperfect markets Of course, in developing countries, markets seldom function so smoothly. A growing body of literature suggests that poor people the world round have high potential returns to investment, 10 especially physical capital, but are unable to realize them because they have few assets and inad- equate access to credit (Banerjee and Duflo 2005). Access to credit was especially poor in northern Uganda in the years after the war. At the time the NUSAF YOP program began, few large public or private lenders had a presence in the re- gion, in part because of insecurity, but also because of constraints on the Ugandan finance sector more generally. Moneylenders and village savings and loan associations were relatively com- mon, but loan terms seldom extend more than one to two months. These small lenders typically loaned funds at rates of 10% per month, or more than 200% per annum (Levenson 2011).11 As a result, at the time of the baseline, just 11% of the sample had saved funds in formal or informal institution in the previous 6 months, with a median level of savings of 40,000 UGX (or $22). A third of respondents had borrowed funds in the previous 12 months, but these were gen- erally small loans (10,500 UGX, or $5.83 at the median) and mainly from friends and family. Less than one in ten borrowed from an institution, with the median loan just 30,000 UGX ($17). About 37% said they believed they could get a loan of 100,000 UGX ($55), with 60% saying it would come from family and 40% from institutions. Just eleven percent said they believed they could obtain a loan of 1 million UGX ($555), 20% from family and 80% from institutions.12 4.1.3 Imperfect entrepreneurs Entrepreneurs, moreover, are not always forward-looking, time-consistent, and disciplined decision-makers. A growing behavioral economics literature emphasizes the difficulties people have in making complex economic decisions, including bounded rationality, overconfidence bi- as, time inconsistency, or other self-control problems (Bertrand et al. 2004). And some people are simply less patient than others, and will tend to consume windfalls. Interventions like YOP will not yield high private or social returns if high-return investments are available but not seized. Fafchamps et al. (2011) find some evidence of such self-control problems in a microen- 11 Commercial prime lending rates were approximately 20% per annum in 2008-09, or roughly 5% in real terms, accounting for inflation of approximately 15% (CIA 2011). Our informal assessment suggests that commercial lend- ing rates for small to medium firms were roughly 15% to 25% in real terms. 12 Over the course of the study, both the security environment and the level of financial development improved in northern Uganda, undoubtedly increasing the availability of credit. The level of financial development remains poor, however, and security (especially peace in neighboring southern Sudan, and the massive boom in trading opportuni- ties) probably raised the returns to capital faster than the availability of internal and external credit. Hence NUSAF ought to provide an excellent example of the returns to grants in a constrained credit environment. 11 terprise program in Ghana, especially among the poor, women, and those who received cash in- stead of in kind assistance. Indeed, a qualitative study of the NUSAF components that provided cash to support livestock and community infrastructure, concluded that beneficiaries often did not manage the funding well (Golooba-Mutebi and Hickey 2010). Interviews suggested that projects were not well re- searched, funds were mismanaged, and intra-group disagreements were commonplace. The study argued it is unrealistic to expect poor people to be responsible for their own recovery, and that the program actually had disempowering effects. This study did not focus on the YOP program, but ut our own observations and interviews with YOP beneficiaries before and during the evalua- tion revealed many failures and concerns akin to those identified by the qualitative study. At the same time, the group organization of YOP, with planning support from facilitators, was partly intended to provide some form of commitment and help overcome self-control problems. Banerjee and Mullainathan (2009) suggest that, in theory, the poor might exhibit more self- control with large lump sums rather than small savings (although there is little empirical evi- dence to suggest this is the case). 4.2 A simple model of occupational choice and cash transfers To structure our thinking and predictions we turn to a simple two-period occupational choice model with imperfect markets (no borrowing ability) and “imperfect� individuals (patient and impatient types).13 The model not only illustrates why cash transfer programs can spur business development and raise incomes, but produces predictions for impact heterogeneity that help il- lustrate whether these market and behavioral imperfections are binding in the Uganda case. Individuals start with initial wealth w. Each can choose to be a laborer, earning an income of y each period, or to be an entrepreneur, and earning f(A, K), where f is a production function in- creasing in inherent ability, A, and the stock of capital, K. Entrepreneurs can use their wealth and current income to invest in capital, but becoming an entrepreneur has a one-time fixed cost F ≥ 0, 13 The model was developed by the authors along with Julian Jamison, for a series forthcoming s of collaborative projects and papers. It could be considered a two-period version of the one-period entrepreneurial investment choice model proposed by Mel et al. (2008), or a cash transfer version of the two-period microcredit model proposed by Banerjee et al. (2010). Credit constraints are not the only potential market imperfection. One is risk and imperfect insurance. de Mel et al. (2008) examine a model where households are risk averse and insurance markets are imper- fect, and show that the gap between the market interest rate and the marginal return to capital are increasing in the level of risk in business profits and in the level of risk aversion displayed by the household. More risk averse indi- viduals should benefit more from cash transfers. 12 which does not go into productive capital. Existing entrepreneurs have already paid the fixed cost and are in business with initial capital, K0 ≥ 0. Anyone can save amount s at interest rate r. To simplify matters, and to reflect actual condi- tions in places like Uganda, we assume r = 0. For similar reasons, we also assume that individu- als are unable to borrow.14 In this setup, everyone chooses s and K to maximize their (concave) utility function: U = u(c1) + δu(c2) where ct is consumption in period t and δ is the individual’s discount rate for period 2. Laborers solve U subject to: c1 + s = y + w c2 = y + s while budding entrepreneurs solve U subject to: c1 + s – F – K = y + w c2 = f(A, K) + s and existing entrepreneurs solve U subject to: c1 + s – K = f(A, K0) + w c2 = f(A, K + K0) + s We illustrate the major implications of the model in Figures 2 to 4. We start in Figure 2 by ig- noring existing entrepreneurs and looking at initially poor individuals (with low w, or wL) who are laborers in period 1 and must choose whether to be laborers or entrepreneurs in period 2. Point E represents their starting endowment at (y + wL, y). Saving corresponds to the -45 line extending from E to the vertical axis. If they choose to start an enterprise, they lose F and invest K, which pays f(A, K) in period 2. We assume f() is concave (decreasing returns) and is increas- ing in both arguments.15 The stylized example in Figure 2 depicts a relatively high-ability entre- preneur with consequently high potential returns (a steep production function). 14 Indeed, real interest rates in village savings association are generally negative, due to fees and inflation. Allowing short-term borrowing at high rates, as we see in Uganda, would not change the model’s conclusions. 15 Production could easily be linear without changing conclusions. If the slope of the production function falls below one, the entrepreneur would switch to savings instead of capital investment. This is not a necessary assumption but it seems reasonable given the stylized facts that (i) poor people often have high returns to small amounts of capital, but (ii) very few microenterprises ever increase beyond a small scale, even with access to credit. In our stylized example no entrepreneur optimally hits such a region, and hence we can take s = 0 for entrepreneurs. 13 Still focusing on the wL case, we can see that different indifference curves (corresponding to different high and low discount rates, δH and δL) will lead to different choices between labor and enterprise, with more patience making entrepreneurship more likely. If δ and w are low enough, individuals will consume and produce at E rather than a point of tangency. The larger is A (or the smaller is F), the more attractive is entrepreneurship. This case reasonably applies to the majority of YOP applicants, who are either petty laborers or traders at the outset or, if they are small en- trepreneurs, they are not engaged in vocations (and their capital stock is not easily transferred). Next consider the higher wealth case, wH, to the right, representing receipt of a cash transfer (though it could also represent any source of liquid wealth or windfall). It is clear from the graph that, fixing A, there is a smaller range of δ for which the agent will choose to be a laborer: patient or ability would have to be relatively low. Intuitively, everyone wants to smooth their consump- tion (concave utility) unless they're very impatient. The higher is w, the more asymmetric the ini- tial endowment, and hence the more individuals want to smooth. Given that they smooth, capital investment typically gives a better return than saving (depending, of course, on A). We assume the initial fixed cost F is small relative to the change in wealth, and F is less important as w (and, indeed, the scale of everything) grows. Figure 3 illustrates the difference between high and low ability (AH and AL) individuals. While magnitudes depend on the shape of the production and utility functions, we can nevertheless see a few relatively general patterns. In this illustration, we see it is possible even for patient individ- uals to remain laborers if the returns to their ability are lower than the return from saving (in this case zero). Given a cash transfer, there will be threshold values of w, A and δ below which indi- viduals will remain laborers after a cash windfall, though in general these threshold values be- come lower and lower as the transfer increases. Generally, higher ability and more patient people should see a larger increase in period 2 earnings and consumption. Finally, Figure 4 considers existing versus budding entrepreneurs, focusing on relatively high ability individuals only. Existing entrepreneurs have paid F and so their production function is shifted to the right, even at initially low wealth levels. The effect of a cash transfer on period 2 earnings and consumption will tend to be greater for budding rather than existing entrepreneurs, especially less patient individuals who would not have chosen to start an enterprise in the ab- sence of the cash transfer. 14 4.3 What is the role of groups in group-based transfers? YOP transfers funds to groups rather than individuals. From the Government and World Bank perspective, there were several motivations for the group design. Administratively it is simpler and cheaper to disburse funds to thousands of groups than tens of thousands of people. Designers also viewed the group organization as intrinsically and ideologically important. The NUSAF program more broadly was designed to promote decentralized, participatory decision-making. It is representative (and indeed modeled after) other “Community-Driven Development� (CDD) initiatives in other countries, initiatives which spend in the tens or even hundreds of billions of dollars globally (Mansuri and Rao 2011). While the most common CDD programs grant cash to communities for community projects, transfers to groups within communities are not uncommon. The intention of the group and participatory approach is to improve the success of targeting, build social capital, and strengthen accountability—specifically, in the YOP case, the likelihood that cash transfers are invested rather than consumed. Based on these theories and our qualitative observation of groups before and after the treat- ment, we see four main hypotheses. First, groups may act as a form of commitment device. For instance, payments for training and some tools are commonly made by the leadership on behalf of all members, and individuals may feel more peer pressure or encouragement to invest rather than consume the transfer. In our model above, this would lead to higher levels of period 1 in- vestment even among low ability and low patience types. In a multi-period setting, these low types might disinvest and return to laboring or less capital intensive entrepreneurship, but in the interim earnings of low patience types would be higher than otherwise. Second, there may be increased availability of capital. Most post-program YOP enterprises are individual rather than group-based, so individual production functions probably remain the right framework for thinking about program impacts.16 But some groups share tools and physical capital (e.g. a building, or high-value tools). It is not obvious whether this increased the potential 16 Among the treated, only 14% report coming together for income-generating activities on a daily basis, while 30% report coming together once a week for this purpose. Of the 14% that come together on a daily basis, 75% report sharing tools, while of the 30% that report coming together once a week, 85% report that tools are shared in the group. 15 returns to capital but, in general, the sharing of high-return capital with high fixed costs probably raises rather than lowers expected returns. Third, low ability types may benefit from high ability peers. This positive effect is not as- sured; social psychological research on small groups suggests that group-based decision-making and learning can enhance or detract from group performance, depending on context and a large number of characteristics (Levine and Moreland 1998). But our qualitative observation suggests that there exist opportunities to learn and observe from peers, increasing the returns of low abil- ity people (and narrowing the performance gap between high and low ability persons). Fourth, observers of CDD programs in general, and NUSAF in particular, fear the potential for elite or leader capture, leading to unequal distributions, possibly positively correlated with ability. If so, we would observe higher average returns among pre-specified leaders. Only this last hypothesis is directly testable with our research design, as leaders and executive committees were pre-specified in both treatment and control groups. As long as endline meas- urement error (e.g. underreporting) of investments and earnings is uncorrelated with baseline leader status, we can test for the presence of leader capture. The other three hypotheses are not directly testable, as NUSAF programs rules didn’t allow for individual transfers. But we can look for indirect evidence based on baseline data on group quality, cohesion and composition. In particular, we hypothesize that the extent to which groups act as effective commitment devices, effectively share tools and raise shared capital (and re- turns), and raise the performance of low ability types is increasing in levels of group cohesion and quality. Low types are more likely to benefit from heterogeneous groups (those with higher ability people). We return to these concepts and tests below. 5 Impacts on social cohesion, alienation, and instability: A conceptual framework YOP, like many development programs, explicitly aims to promote social cohesion and stabil- ity. The underlying logic, however, is seldom as explicit. We highlight six bodies of social theo- ry, each of which plausibly links cash transfers and higher incomes and employment to socio- political outcomes. We are not aware of efforts to discuss or analyze each of these competing theories together, and identify the empirical predictions that can distinguish between them. A comprehensive attempt is well beyond the scope of this paper, and thus this discussion is stylized and preliminary. 16 5.1 The “participatory� view: Group formation and participatory decision making in- crease social support and cohesion The first is an assumption underlying most community-driven and participatory development programs, implicitly and explicitly: group decision making, especially in combination with eco- nomic empowerment, promotes social cohesion, social and community participation, and notions of citizenship. The belief is consistent with sociological theories that associational life is a cru- cial form of social capital and well-being (Putnam 2001), though the application to development programs assumes that this associational life and cohesion can be induced by state development programs and incentives. Mansuri and Rao (2011) review the theory and evidence of communi- ty-driven development programs akin to NUSAF and argue that the rhetoric often exceeds reali- ty. Their findings are consistent with a large body of social-psychological research that suggests that group work and decision-making can have highly heterogeneous impacts depending on con- text, composition and other factors (Levine and Moreland 1998) as well as recent evidence from CDD program evaluations (Casey et al. 2011). This mechanism suggests we should observe increases in social cohesion and, possibly, other forms of community participation as a consequence of treatment. We do not see a clear reason for aggression to be affected through this channel. These effects do not necessarily increase with economic success. They may be greater in groups where the initial quality of the dynamic is greater. 5.2 The “social role� view: Increased incomes elevate social position and cohesion Throughout agrarian societies, and perhaps especially in contemporary rural Africa, commu- nities and social groups act as a mutual insurance system, and the kin system in particular works as a form of mutual assistance among members of an extended family, traditionally from the old- er to the younger.17 In such societies—and northern Uganda is no exception—the transition from being a “youth� to becoming an “adult�, from disregard to social esteem and support, is in part determined by one’s ability to give rather than receive gifts and transfers. To the extent that par- ticipation in a YOP-like program increases relative wealth, and the ability to increase net trans- 17 See Hoff and Sen (2005) for a review. 17 fers to kin or the community, we may expect an increase in social support, respect, and opportu- nities for community leadership and engagement.18 Conversely, African anthropological literature stresses that youth who are alienated from this system, and have little means of being net givers at the age when they ought to be “adults� in the social sense of the term, are more likely to engage in anti-social behavior and even insurrection (e.g. Richards 1996; Peters and Richards 1998). This mechanism suggests we should observe increases in social cohesion and support, and that these changes should be correlated with higher economic success and (perhaps most of all) evidence of transfers. To the extent that lower alienation reduces anti-social behavior, we may also expect to see lower aggression as a result. 5.3 The “materialist� view: Higher incomes raise the opportunity cost of predatory activ- ities A third, more materialist view, argues that those with low earnings, or nothing to lose, have a lower opportunity cost of aggression, crime and insurrection, and hence are more easily mobi- lized into predation. By this account, employment programs reduce predatory activities to the extent that they raise incomes and either crowd out or raise the opportunity cost of these activi- ties. This employment-predation link comes from classic economic theories of crime: poverty lowers the opportunity cost of peaceful production, providing incentives for predatory activities (Becker 1968; Freeman 1999). Economists have extended this logic to insurrection, arguing that youth unemployment and adverse economic shocks raise the risk of conflict in developing coun- tries, and a growing body of evidence from cross-country studies is emerging to confirm this (Blattman and Miguel 2010). This mechanism makes no predictions about alienation or cohesion per se. With respect to an- ti-social behavior or violence, the materialist view would only apply to predatory or anti-social activities with an opportunity cost of time or funds. None of the measures in the present study have such a cost, and so we will not speak to this view in this paper. 18 Hoff and Sen (2005) also note, however, that with a large enough gain, individuals might have an incentive to excise themselves from their kin group, to avoid the financial obligations and protect their YOP transfer. There is thus the potential for reduced social support and cohesion. 18 5.4 The “frustration-aggression� view: Anti-social behavior and conflict are a function of frustrated ambitions, especially relative deprivation A fourth, more psychological and sociological view is that poverty produces aggression and alienation through frustrated ambitions. Some follow sociologists Durkheim (1893) and Merton (1938) and see poverty and blocked goals as producing strain on the social system, leading to deviance, delinquency and crime. Political scientists also emphasize how, throughout history, these frustrations have been mobilized and led to insurrection, especially where poverty is une- qual and unjust, leading some individuals to find intrinsic value in the act of aggression or insur- rection itself (Gurr 1971; Scott 1976; Wood 2003). This belief is rooted in early psychological research that argues that aggression is a reaction to external conditions frustrating a desirable outcome (Dollard et al. 1939). This mechanism makes no obvious predictions about social cohesion or community participa- tion. If receipt of the program rectifies a perceived injustice or inequality, and reduces frustra- tions, then we might expect to see lower aggression. This may or may not be associated with the degree of economic success. A randomized control trial might not be the ideal test of this view, however, since receipt and non-receipt of the program is widely understood to be random. 5.5 The “psychological stress� view: Employment and income reduce anti-social behavior due to reduced stress Frustration-aggression theories of violence and anti-social behavior take a fairly narrow view of psychology and aggression, one that is rooted in psychological research from the 1960s and even 1930s (Dollard et al. 1939). As important as injustice and frustrated ambitions may be to violence, a larger body of psychological research since this time emphasizes that aggression is also a highly charged emotional state suggests that aggression and anti-social behavior can be reactions to a variety of adverse stimuli or stress (Berkowitz 1993). This is hypothesized to be one reason for the association between low socio-economic status and anti-social behaviors in developed nations. This mechanism predicts that treatment and the degree of economic success should be associated with lower aggression. 5.6 The “situationalist� view: Violence is the product of circumstance, which may be (spuriously) associated with poverty A final view sees violence as the product of circumstance, not calculations or impulses (Collins 2008). For instance, the poor may have less access to justice and security, and so be 19 more vulnerable to victimization or mobilization (Scacco 2008). The view is particularly com- mon in urban settings and communal violence. We do not see a clear role for this mechanism on the northern Ugandan setting, where there was little obvious variation in risk of insecurity and access to justice within the sample. 6 Data and measurement 6.1 Survey data The 535 eligible groups contained nearly 12,000 official members, and we follow a panel of five members per group, or 2675 persons. We achieve an effective tracking rate of 87%.19 A baseline survey was conducted in February and March 2008. Enumerators located 522 of the 535 groups and mobilized all available group members—about 95% on average—to com- plete a group survey that collected demographic data on all members, present or not, as well as group characteristics.20 Five of the members present were randomly selected for an in-depth questionnaire in their local language. (Appendix Table 1 displays summary statistics for key baseline variables and also demonstrates the degree of treatment-control balance.) The government disbursed YOP funds between July and September 2008, 5 to 7 months after the baseline survey. Groups typically began training shortly thereafter and most had completed training by mid-2009. We conducted a follow-up survey between August 2010 and March 2011, roughly 24 to 30 months after disbursement and 12 to 18 months after most completed training. We attempted to track and interview all 5 members of the 522 groups found at baseline, plus members of the 13 unfound groups. At least one (and often several) attempts were made to find each individual, and we selected a random sample of migrants and other unfound individuals for intensive tracking, often in another district.21 Overall the effective attrition rate is 13%. Attrition 19 All estimates in the paper are within-sample predictions, and we do not weight for differential selection from the population of 12,000. 20 In two survey rounds we were unable to locate 12 of the 13 missing groups on follow-up attempts, suggesting that these 12 groups may have been fraudulent “ghost� groups that slipped through the auditing process. Unusually, all 13 missing groups had been assigned to the control group and so received no funding. For logistical reasons related to program operations, treatment had to be randomized prior to baseline, but assignment was only known to the re- searchers and the central government director. District officials and enumerators also did not know the treatment status of the groups. 21 We conducted tracking in three phases. In Phase 1, from August to September 2010, we drew a random 75 percent sample of the groups for tracking. Enumerators sought respondents in their original villages, and located 61%. In 20 does vary by treatment status—approximately 15% among those assigned to treatment and 9% among controls. But analysis of attrition patterns using baseline data suggests that attrition is rel- atively unsystematic.22 6.2 Key outcomes Primary outcomes are described in Table 2, grouped into eight “families� based on pre- specified conceptual linkages. 6.2.1 Economic outcomes Investments in vocational skills and capital. We first examine investments in human and physical capital—both those made upon receiving the cash transfer as well as stocks at the time of the endline survey. In some sense, these investments represent a form of treatment compli- ance, although we feel they are more properly regarded as “intermediate� outcomes of the treat- ment—especially because the cash, once received, is relatively unmonitored and unconditional. Respondents self-report the Hours of training received since baseline, the value of Tools and machines acquired since baseline (in thousands of Ugandan Shillings, or UGX), and the value of their total Stock of raw materials, tools and machines. We censor all UGX-denominated varia- bles at 99th percentile (the size but not the significance of treatment effects are sensitive to this censoring, as discussed below). Unfortunately, we do not know the exact distribution of the transfer within groups, or specific amounts spent on training, raw materials, or start-up costs. Groups divided and disbursed funds among members in diverse and difficult-to-observe ways, sometimes paying for training on behalf of the group, sometimes making bulk tool purchases, Phase 2, from October to November 2010, we selected a 54% random sample of unfound Phase 1 respondents for tracking, wherever that may be in the country. This sampling technique was designed to use scarce resources to min- imize attrition bias from long panels, providing a lower effective attrition weight and reducing potential for bias (Thomas et al. 2001; Gerber et al. 2011). Those selected for Phase 2 tracking are representative of all respondents not found in the first phase, and receive greater weight in all analysis. Those not selected for tracking in the second phase receive zero weight. Enumerators made at least three attempts to track Phase 2 respondents to their current location and found 76%, for an effective attrition rate of 90.6% in the first two phases. Finally, new resources at the end of Phase 2 made possible a Phase 3, from January to February 2011, where enumerators sought the 25% of groups randomly dropped at the outset. Enumerators found 79% of those targeted in Phase 3 22 We assess the probability of being unfound on treatment status, 16 demographic characteristics and indices of lagged dependent variables. Collectively the explanatory power is low (an r-squared of 0.06). We observe three sub- stantive and statistically significant differences: Males were four percentage points less likely to be found; urban persons were 8 percentage points more likely to go unfound; and a standard deviation increase in wealth led to a 2.6 percentage point greater likelihood of not being found. 21 and sometimes dispensing cash to members. Groups seldom kept records, and members could not reliably estimate the value of any in-kind transfers. Hence hours of training and durables ac- quired and owned represent our best (albeit incomplete) investment estimates. Income, consumption and employment. To measure employment levels and occupational choice, we measure total Hours on all economic activities the past four weeks, which excludes household work and chores but includes subsistence work (e.g. hunting, farming, charcoal mak- ing) as well as Hours on market activities (all business activities, vocational employment, pro- fessions, wage work, or other market work).23 To calculate incomes, we ask respondents to esti- mate their profits from business activities and wages or earnings from other activities in the pre- vious four weeks by activity, and calculate Total cash earnings in past month (`000s of UGX).24 Finally, to approximate consumption levels we calculate an Index of wealth z-score using 7 measures of housing quality, 55 household and business assets, 5 types of landholdings, and 3 measures of personal appearance. The index is the score from the first principal component of these assets—shown to be a relatively reliable proxy for full consumption aggregates (Filmer and Scott 2008). 6.2.2 Alienation and aggression Participation and engagement. One facet of integration (or, conversely, alienation) is public participation, social and political. We ask respondents about their Number of group memberships in the community, whether they Attend community meetings, and whether they Speak out at community meetings. We also ask whether they are a Community leader of any form, or a Com- 23 The distinction between subsistence and market work is based on occupation type, and activities were classified as subsistence if less than 15% of persons reported cash earnings from the activity. 24 Net income is one of the most important measures but also one of the most difficult. While subject to recall and other potential forms of bias, some experimental evidence from microenterprise profit measurement suggests self- reported profits and earnings may be the least biased measure of income, imperfect as it may be (de Mel et al. 2007). In addition to measurement, income (like all our UGX-denominated outcomes) has a long upper tail to which any measure of central tendency, including average treatment effects, is sensitive. Outliers are particularly influential. After accounting for outliers beyond the 95th or 99th percentile, the distribution of net income is roughly log-normal. But a quarter of respondents report zero net income in the past four weeks, and non-zero earnings are more likely among the treated. We take four steps to conservatively estimate statistics on UGX-denominated measures, especial- ly income. First, we truncate the variable at the 99th percentile to account for outliers. Second, we examine both the linear effect and a non-linear transformation, the inverse hyperbolic sine, which is similar to a log transformation but defined at zero (Burbidge et al. 1988). Third, because both the level and logged values of income may be mislead- ing, we examine the median and other major quantiles, including treatment effects. We also explore sensitivity to all these assumptions. 22 munity mobilizer, which is a position commonly filled by youth, who help to organize meetings, gather members, or spread messages. We also ask four questions about their perceived Locus of control—a psychological construct that attempts to measure the extent to which individuals be- lieve that they can control events that affect them. Unfortunately, due to an impending election (and a desire not to politically charge the survey at a sensitive time), we were asked by the government partner to exclude questions on civic atti- tudes or participation beyond the community level, including political knowledge and attitudes. As a result we can only evaluate participation impacts at the community level, although broader participation data will be gathered in the 2012 round of data collection. Social integration. To assess alienation, we also examine interpersonal relationships and in- tegration. We have an indicator for whether respondents indicated their Families are very caring towards them. We also calculate a more general Index of social support, an additive index run- ning from 0 to 16 based on responses to 8 self-reported questions about concrete forms of social support received in the past four weeks.25 We also construct a Neighbor relations index running from 0 to 8 based on four perceptions (each a 0-2 scale) about the quality of neighbor support, relations, esteem, and trust. Finally, respect for and quality of relations with elders in the com- munity is an important indicator of social and community integration in rural Uganda, and we construct a Reverence for elders index running from 0 to 9 based on three questions (each a 0-3 scale) on self-reported helpfulness to and respect for elders, and their authority over youth. Depression and distress symptoms. We adapt an additive index of psychological distress that runs from 0 to 21, using 7 self-reported symptoms of depression and anxiety, each rated 0 to 3 by frequency.26 Aggression and hostility. We have three main types of aggression and hostility measures. The first measures the frequency of angry disputes on a 0 to 3 scale (for never, rarely, some- times, or often) with particular parties, giving us an Index of disputes with neighbors, an Index of 25 Each is measured on a 0-2 scale from “no support received� to “yes, often�). Examples include whether or not someone: looked after a family member or the possessions of the respondent while they were away, or sat with the respondent when they were feeling distressed or lonely. 26 Symptoms include feelings of isolation, nightmares, difficulty sleeping, hyper-arousal, etc. We adapt our 7-item scale from the 19-item distress scale used by the Survey of War Affected Youth in northern Uganda (Blattman and Annan 2010). All 19 symptoms were collected at baseline, and for the 7-item endline scale we took the 7 most influ- ential items from the rotated first factor of all 19. 23 disputes with family, an Index of disputes with community leaders, an Index of disputes with po- lice, and an Index of physical fights. The second type measures the aggression of their peer group (on the same scale), including whether Peers have disputes with local leaders or police, and Peers involved in physical fights. Finally, we ask about three self-reported behaviors associated with hostile behavior in the psychological literature, including scales for how frequent they are Quarrelsome, Take things without permission, Use abusive language, or Threaten to hurt oth- ers.27 As with political participation, we did not collect data on attitudes towards political vio- lence, or on participation in crime, protests, riots, or communal or armed violence, but more ex- tensive data will be collected in the 2012 round. 6.2.3 Subjective well-being Finally, we measure current subjective well-being by asking respondents to place them- selves (relative to other community members) on 9-step ladders of Wealth, Community respect, Power in community, Access to basic services, and Asked for advice (an important social role of respect in northern Uganda). For future subjective well-being, we also asked each respondent to give us their expected place on the ranking in 5-years for wealth, respect and power. We also asked a general question on Optimism, specifically, on a 0-3 scale, whether they “believe good things will happen in your life�. 7 Results Table 3 displays treatment effects for each outcome family, for the full sample and by gender. Each family is represented by a mean standardized outcome (a z-score) calculated as the stand- ardized sum of each of the outcomes in the family (themselves mean standardized). The main reason to look at these aggregates is to guard against the heightened probability of rejecting a true null hypothesis when testing multiple outcomes (Duflo et al. 2007). The economic impacts of the program are large. On average, being part of a treated group re- sults in a standard deviation increase in investments in vocational skills and capital and a 0.28 standard deviation increase in economic success. Male and female impacts are nearly identical. 27 Aggression and dispute questions were developed by the authors after extensive pretesting, and the aggression measures are similar in content to psychometric hostility measures used in developed countries, but locally adapted by the authors to the Ugandan context. We are not aware of a validated measure of aggression for Africa. 24 The impacts on alienation and aggression are smaller and the effects more mixed. On average, there are small improvements in community participation, social integration, distress symptoms, and aggression but these are statistically insignificant. When we differentiate by gender, howev- er, we see that these small average results conceal heterogeneous, divergent impacts. Males show small but significant improvements in social integration (0.11 s.d.), distress symptoms (-0.15 s.d.) and aggression (-0.20 s.d.). Females, on the other hand, generally show small and statistical- ly significant increases in alienation, and a significant increase in self-reported aggression (0.20 s.d.). Here and in all future ATE tables, female ATEs are calculated at the base of the table as the sum of the male ATE and the interaction term. Finally, overall subjective well-being increases by 0.15 s.d. Anticipated changes in subjective well-being, however, are lower among the treated by a nearly equal amount. The mechanical rea- son is that members of treated groups estimate the same levels of future relative well-being no differently than controls. We discuss potential reasons for the finding below. While important for testing multiple hypotheses, these standardized family treatment effects conceal a great deal of important variation. There is no theoretical reason, for instance, why eco- nomic outcomes like employment levels, earnings and consumption ought to move in the same direction. They do not help us calculate returns to investment, or the determinants of heteroge- neity in returns. And patterns of specific forms of aggression and alienation may illuminate the general patterns we see above. The remainder of the paper focuses on individual outcomes. 7.1 Investments in vocational skills and capital Overall—and rather remarkably—the vast majority of beneficiaries make the investments they proposed: most engage in vocational training and approximately two-thirds of the transfer appears to be spent on fees and durable assets (not including other startup costs or materials), suggesting that the fears over funds mislaid and misspent are confined to a minority of benefi- ciaries. 7.1.1 Skills training Table 4 displays the average treatment effects (ATE) for self-reported investments. As in Ta- ble 3, we calculate female ATEs at the base of the table. To provide a sense of magnitude, we also report control group means and (except in the case of non-linear transformation) calculate the treatment effect as a proportion of the control group mean. 25 Seventeen percent of the control group enrolled in some form of vocational training, suggest- ing a degree of demand and ability to invest. 70 percent of the members of treated groups en- rolled in vocational school since baseline. Enrolment levels are similar for men and women.28 The most common types of vocational training were tailoring (32%), carpentry (21%), metal- working (10%), and hairdressing (6%). On average, this translates to 405 more hours of training than controls, more than 10 weeks of full time training (Table 4, columns 1 and 2). A small few used YOP funds to enroll in secondary school, technically a “misuse� of the funds.29 7.1.2 Asset acquisition and stocks The average control group member reports acquiring business assets worth UGX 136,500 ($62) since baseline, and value their stock of tools, machines and raw materials at UGX 348,000 ($158). Treated individuals report an additional 656,016 UGX ($298) in acquisitions and UGX 523,318 ($238) in asset stock, a 481% increase in acquisitions and 150% increase in asset stock relative to the control group. The impact on asset stocks is sensitive to the upper tail and any censoring, however. A log transformation would be less sensitive to outliers, but would treat ze- ros as missing, introducing selection bias since the probability of acquisitions is affected by treatment. An alternative transformation with similar properties to the log, but defined at zero, is the inverse hyperbolic sine, or IHS (Burbidge et al. 1988). While the coefficient has no easy in- terpretation, it indicates to what extent the linear treatment effects are sensitive to outliers and functional form. In this case, the treatment effect seems to be robust but the male-female gap is not robust to the IHS transformation. We will see the same with income, below. This suggests that any male-female gap is driven by the upper tails and outliers, and perhaps not so salient. Because of these potential biases, we also turn to quantile analysis. Figure 5 maps the quantile treatment effects (QTEs) for business assets owned. The median control group member owned just UGX 14,000 ($6) of business assets at endline. Below the 30th percentile, treated group members report virtually zero business assets as well, but the two groups diverge sharply from that point onwards. At the median, the QTE for assets acquired is UGX 164,000 ($75) for assets 28 See Appendix Table 2. Those who dropped out with fewer than 16 hours of training were not counted as enrolled. 29 10 percent of the control and 13% of those in treated groups re-enrolled in formal schooling (usually secondary school) since baseline—small in absolute terms but proportionally-speaking a large (30%) increase. See Appendix Table 2. 26 owned, and at the 70th and 90th percentiles the QTE rises to more than UGX 300,000 and 1,400,000—each one many multiples of the corresponding control quantile. What proportion of the transfer is used for vocational investments? Treated groups reported that approximately 35% of any YOP transfer was spent on training fees (Table 1). The asset QTE, above, moreover, suggest that the median treated individual spent approximately 26 per- cent of the transfer on assets. This suggests that nearly 61% of transfers were spent on skills training and durable assets alone. While some of the remainder was undoubtedly consumed or transferred, some was likely invested in working capital (such as materials and stock purchases), operating expenses, or held as savings. These results suggest that either self-control issues are not a major constraint on investment (at least with large transfers) or that the design of the pro- gram—specification of a proposal, auditing prior to disbursal, and group organization and con- trol over funds—may have acted as a commitment device. We return to the role of the group, below, and find little evidence that group quality affected investments. 7.1.3 Group dynamics and investment The group-based disbursement of funds implies that investment may not have been solely an individual decision. Do group characteristics matter? To what extent do better quality or more homogenous groups differ in investments and performance? Is there any evidence that the group disbursement acted as a commitment device? Table 5 looks at treatment heterogeneity on key investment and economic outcomes, by group characteristics measured at baseline. We interact treatment with: an indicator for whether the Group previously existed for other purposes, before they applied for YOP funding; a standard- ized index of the Quality of the group dynamic (based on the average response in a group to five opinion questions, such as trust in group members, the quality of cooperation, or whether they would work with the group again); the Group size; the Proportion female; and finally a Group heterogeneity index (a standardized additive index of the standard deviation of characteristics within the group, including education, starting capital, and age). If the group plays a large role in investment decisions, commitment to investments, or sharing information and tools, we hypothesize that investments and economic performance should be increasing in group cohesion and quality, indicated by previous existence and the dynamic. The effect of group size and heterogeneity is theoretically ambiguous, but effectiveness is potentially decreasing in both. 27 We see only weak evidence for any effect of group characteristics on investments or perfor- mance. The coefficient on the Group previously existed interaction is positive across all out- comes, but in general small relative to the treatment effect and not statistically significant. In- vestments and earnings are both increasing in the quality of the group dynamic, but the effect is only statistically significant for capital acquired. This is consistent with the idea that groups op- erate as commitment devices, but the magnitude is only moderate relative to the treatment effect, and is not reflected in significantly higher earnings or wealth. We see little relationship between group size and performance—an unusual result, which we return to below, since smaller groups tended to receive larger per capita transfers. Treated groups with a higher proportion of females are more likely to invest in training hours. Strikingly, how- ever, these groups are much less profitable and wealth levels are also much lower. Finally, treat- ed members of more heterogeneous groups do more poorly on average, but the impacts are small and not robust. Finally, given the absence of upward accountability after the cash transfer, a reasonable con- cern is that transfers may have been captured by some members, particularly the group executive committee in charge of finances and planning. We see little evidence that transfers were captured by leaders. First, less than 2% of groups assigned to treatment reported that a group leader ap- propriated most or all of the funds. Second, most group members remain satisfied with their group: more than 90% still work with the group and more than 80% feel the group cooperates well (Table 1). Third, we test for heterogeneous impacts among leaders, but see few significant differences. We look at how self-reported investments vary by leadership position—whether a member of the full executive, or one of the two most senior positions—the committee chair or vice-chair, controlling for ability and wealth. The coefficient on an interaction between treatment and leadership indicates how leaders responded or benefited disproportionately from the transfer. Results are displayed in Appendix Table 3. The sign on the leader interaction is generally posi- tive, implying leaders received more training and capital than the average member. The differ- ence in training hours is large (about a one quarter increase over other group members) and sta- tistically significant. But the coefficients on capital acquired and stocks are closer to zero and not robust. Coefficients on the group chair interaction are actually negative for capital investments. 28 7.2 Economic impacts 7.2.1 Income, consumption and employment Our model predicts a shift from unskilled to skilled employment, and an increase in earnings and consumption. Table 6 reports average treatment effects for the full sample and by gender. First, we see a substantial increase in skilled or somewhat capital-intensive work. A third of the control group is engaged in such enterprises at endline, but this rate doubles among treated indi- viduals. The impact is slightly greater for women than men, but the difference is not robust. Second, we see a substantial increase in net income, both from the linear and IHS transfor- mation. On average, the treated report UGX 19,515 ($9) greater incomes in the last 4 weeks at endline. While seemingly small, the impact is huge relative to the counterfactual—a 45% in- crease over the control group mean. The size (but not the significance) of linear estimates is sen- sitive to the upper tail, and so we look at IHS results as well and find them similarly robust. Linear and IHS average treatment effects differ in one crucial respect: returns by gender. The linear income results suggest that women earn significant less than men, with the female treat- ment effect just UGX 5,992, and not significantly different from zero. The non-linear results, however, suggest that women’s average treatment effect is similar (and if anything, greater) than that of men, though the difference is not significant. As with assets acquired and owned, the in- consistency appears to be due to the long upper tail in earnings. When we turn to quantile treatment effects, we see that women benefit significantly from the program, bolstering the IHS results. Figure 6 shows the QTE for men and women. The treatment effect at and below the median is similar for both genders: positive and generally significant after the 10th percentile, and nearly equal at the median at UGX 10,500 for males and 10,000 for fe- males. Above the median, male QTEs diverge, jumping to roughly 20,000 at the 70th and 80th percentiles and 78,000 at the 90th. The female QTE is relative steady until the 80th and 90th per- centiles, in the latter case only reaching 33,000. Third, we use a standardized household wealth index to proxy for consumption. The treated exhibit a 0.13 standard deviation increase in housing quality and durable assets, with the increase concentrated primarily among men. The change in wealth for women is positive but close to zero and not robust. Cash savings show a similar pattern—the treatment effect is large among men, significant at the 1% level, and small and not significant for women (Appendix Table 2). 29 The model assumes full labor utilization. If we were to relax this assumption, theoretically the cash transfer has an ambiguous effect on employment. Labor hours should increase to the extent that labor and capital are complements, and decrease to the extent that labor is a normal good. Few of our sample, however, are fully employed—the average control group member is engaged in market and household employment just 4.3 hours a day. Hence we expect employment to in- crease on balance. Indeed, hours in all activities—subsistence and market based—increase among men and women by nearly 20 hours per month. This is principally an increase in market- based activity; treated individuals report 22 more hours of market employment. (The differ- ence—time spent on subsistence activities—changes little for both men and women.) While in absolute terms this amount may seem small—less than an extra hour per day—it represents a 32% increase over the control group. Among women, who tend to engage in less market based work in the absence of treatment, the ATE represents a 49% increase.30 7.2.2 Returns on investment The average transfer amount was UGX 673,026 ($374) per group member (Table 1), and the median transfer was 545,642 ($303). The monthly earnings ATE is 19,515 ($9) and the QTE is 10,000 ($5). Ignoring heterogeneity in transfer amounts received and earnings (and any correla- tion between the two), and assuming earnings in the most recent month are representative of past and future real earnings (i.e. ignoring inflation and any change in enterprise size and productivi- ty) the ATE represents a return of 2.9% per month (35% per annum, non-compounded) and the QTE represents a return of 1.8% per month (22% per annum). These returns reflect added inputs, especially added labor. We can calculate an “adjusted� earnings measure that subtracts from each individual’s earnings a wage for each of their hours employed. We do not have data on wages, and so predict wages using control group endline da- ta: we use baseline education and demographic data to predict a wage level for each individual and subtract the sum from their earnings. We obtain nearly identical returns: the ATE on these adjusted earnings is UGX 16,614 ($8) (Table 6) and the QTE is UGX 9,185 ($4) (regressions not shown). These figures correspond to annual rates of return of 30% and 20%. 30 The amount of time spent at household work and chores falls by 23% among the treated, by 9 hours in the past four weeks (Appendix Table 2). The absolute fall in hours is much larger for women (a fall of 18 hours over the past four weeks compared to a fall of 5 among men). 30 Do these returns exceed market interest rates? Are they “high�? This depends largely on the real interest rate used. In 2008-09, Uganda’s real prime lending rate to banks was just 5%. Short- term microfinance rates, on the other hand, are roughly 200% per annum. While detailed data are not available, real commercial lending rates of 10 to 20% appear to be common among small firms. The average returns to capital above also approach the “high� returns of 40 to 60% recorded for microenterprises in Sri Lanka, for firms with moderate amounts of capital in Mexico or for farmers producing traditional crops in Ghana (Udry and Anagol 2006; de Mel et al. 2008; McKenzie and Woodruff 2008). These results suggest that the average beneficiary possesses moderate to high returns to capi- tal, even when those investments are somewhat constrained to vocational training and tools. The- se estimates, moreover, focus on earnings alone and ignore any non-pecuniary impacts on physi- cal and mental health, social status or other impacts valued by the beneficiary, and discussed be- low. Another means of evaluating returns is to ask a hypothetical question: given the earnings ob- served, how many months (N) would be needed to repay a loan the size of the average NUSAF cash transfer (T) based on a real interest rate r and a constant payment level P? We calculate the number of months to repay for different T and r in Appendix Table 4. At the median profit level, payback is never reached at high real commercial lending rates (25%) or at typical rural money- lender rates (200%). At the lower end of real commercial rates (15%), payback is reached in 12 years. It may be that the “social� rate of interest is lower (e.g. because a social planner has a low- er cost of capital, or lower discount rate in general) payback is achieved in about 6 years at rates of 0 to 5%. Payback times are faster at the mean profit level—roughly 3 years at the hypothetical “social rate of interest�, 4 and 5 years at the low and high commercial rates, and never at money- lender rates. Finally, if individuals or social planners value non-pecuniary benefits of the inter- vention, or externalities, “payback� is considerably faster. In these scenarios the transfer is “re- paid� in as much as half the time. 7.2.3 Economic impacts and transfer size As we saw in Figure 1, per capita transfers vary widely across groups: the majority received between UGX 350,000 ($200) and 800,000 ($450). This is principally because some groups were smaller than others, but tended to request transfers of similar aggregate size. Our model, and 31 common sense, implies that those receiving larger transfers should invest more and earn higher earnings (in absolute terms, even if it is optimal to consume a higher proportion of larger trans- fers in period 1). Of course, per capita transfer size is unlikely to be exogenous—in principle, more savvy or more selfish applicants may engineer larger transfers. If correlated with entrepre- neurial ability, this would exacerbate the disparity in investment and profit levels. We regress our key investment and economic outcomes on (potentially endogenous) transfer size in Table 7, for treatment groups only. Strikingly, the correlation between transfer size and both investments and performance is nearly zero. The relationship is positive, but only slightly (and not statistically significantly) so. This finding presents a puzzle. One possible answer is that de facto group size and distribu- tion was greater than their de jure size. This could be because, once the transfer was obtained, smaller groups tended to attract new members or supplicants. Alternatively, the community lead- ers who helped the groups receive funding (and was perhaps complicit in the high per capita benefit) extracted rents. We do not have data on either phenomenon, but have an opportunity to collect it retrospectively in the 2012 round of data collection. 7.3 Testing the model: Impact heterogeneity Our theory is rooted in two related models of credit constraints: a single-period entrepreneuri- al model with grants from de Mel et al. (2008) and a two-period model of microfinance by Banerjee et al. (2010). Each paper finds some support for their predictions in experimental im- pact heterogeneity. The former finds that, among the treated, the returns to capital are decreasing in initial household assets and increasing in a measure of cognitive ability (a digit span test) though not in education. The latter finds that, among the treated, microfinance is more likely to be invested among non-existing business owners who have high entrepreneurial potential (calcu- lated from literacy and wage labor of the wife of the household head, the number of prime-aged women in the household, and whether the household owns land). The YOP experiment has three advantages: a large sample size, an out-of-sample test of exist- ing theories (and ex-ante predictions), and rich data on initial ability, working capital, and pa- tience. Our model, adapted to a two-period cash transfer context, makes related predictions: I. Levels of investment, earnings and consumption are increasing in patience, ability, and initial wealth (or working capital); 32 II. Cash transfers should have a greater impact on investment, earnings and consumption when ability and patience are high; III. Ability and patience are complements; and IV. Cash transfers should have a lower impact on investment, earnings and consumption among those with high levels of initial working capital or an existing vocation. Tables 8 and 9 look at impact heterogeneity on four key investment and economic outcomes. Table 8 looks only at individuals without an existing vocation at baseline (those with non- vocational microenterprises are not excluded). We look at heterogeneity along three main dimen- sions: a standardized Ability index31, Working capital index32, and Patience index.33 For each outcome, the first column displays the coefficient on the index for the treatment group alone (prediction I). The second column looks at the full sample, and interacts treatment with each in- dex to look for disproportionate effects of treatment based on these baseline characteristics (pre- diction II).34 Looking at the treatment group alone, the coefficient on initial working capital is generally small relative to the treatment effect, changes sign from outcome to outcome, and is not statisti- cally significant. The same is generally true for the treatment and working capital interaction co- 31 The index of ability is a weighted average of baseline measures of educational attainment, a literacy indicator, an indicator for prior vocational training, performance on a digit recall test, a measure of physical disabilities, and a measure of emotional distress and depression. For weights, we use each variable’s predictive power of economic success in the control group. We regress a composite measure of the economic impacts on the baseline measures of ability using the control group only. We use the estimated coefficients to predict a “score� for all treatment and con- trol individuals, and standardize the score to have mean zero and unit standard deviation. Hence in the heterogeneity regressions, the level Index is correlated with the dependent variable by construction, but our interest is in the inter- action between the Index and treatment. 32 The index of working capital is a weighted average of baseline measures of savings, loans outstanding, cash earn- ings, perceived access to a 100,000 UGX loan, perceived access to a 1 million UGX loan, and indices of housing quality and assets (similar to the index of wealth endline measure). Weights are obtained in the same manner as abil- ity. 33 The patience index is a weighted average of endline measures of 10 self-reported measures of impulsiveness and patience, including self-reported willingness to wait long periods for material goods, to spend money “too quickly�, to put off hard or costly tasks, or to resist temptation. Weights are obtained in the same manner as ability. Endline measures are used as no baseline data are available, on the assumption that preferences are time-invariant and are not affected by treatment. As seen in Appendix Table 1, there is no appreciable difference in patience levels between treatment and control groups. 34 The human and working capital indices are each a weighted average of baseline survey variables, where the weights are not equal but rather depend on each variable’s relative predictive power over endline economic out- comes among the control group alone. Hence in the heterogeneity regressions (where the control group is included) the level of each index is correlated with the dependent variable by construction. We are mainly interested in the interaction between the index and treatment in the full regression. 33 efficient, except in the case of IHS(earnings), where the coefficient has the expected negative sign and is significant at the 10 percent level. Treated members with higher ability engage in significantly more training hours (equivalent to roughly half the treatment effect) but have no consistent or significant relationship with capital investment, earnings or wealth. Treatment and ability interact positively for hours of training, but the coefficient on the interaction is negative or zero for capital investment, earnings and wealth. Since ability undoubtedly affects returns, this suggests that our baseline components of the abil- ity index—education and literacy, working memory (digit recall), and physical and mental health—are not robust determinants of entrepreneurial success (in contracst to the evidence from de Mel et al. 2008). Heckman et al. (2006) and others stress “non-cognitive� skills, and Bruhn et al. (2010) emphasize “managerial capital�, but we unfortunately have no baseline data on either. The patience index is the largest and most robust predictor of capital investments, earnings and wealth, but the interaction between patience and treatment is typically negative and not sta- tistically significant. Table 9 looks at all individuals, but splits the sample into those with a patience index above and below the median (i.e. δH and δL). Within each δ subsample, we regress each outcome the Working capital index, Ability index, an indicator for an Existing vocation at baseline. Treatment effect for high patience individuals should be greater overall (prediction II), and should positive- ly interact with ability (prediction III). Treatment should interact negatively with existing voca- tions (prediction IV). Consistent with Table 8, treatment effects are no higher among patient than non-patient individuals. Nor do we see the predicted relationship between ability and high pa- tience individuals (although, again, this may be because we have the “wrong� measure of abil- ity). Of those with an existing vocation, however, the signs and magnitudes are all in the ex- pected directions, and are significant for earnings: existing entrepreneurs have high profit levels (because they are larger) but the effect of treatment is lower amongst these existing entrepreneurs (because they are less constrained to begin with). 7.4 Impacts on subjective well being Consistent with these income and wealth gains, treated subjects perceive themselves as doing economically better than fellow community members. They report a 14% increase in perceived wealth levels relative to the control group (Table 11) and a similarly large and significant in- crease in access to basic services in their community. They do not perceive themselves to receive 34 more respect, have more power, or be sought out for advice relative to others in the community. Their gains seem to be purely economic. These perceived economic gains, moreover, are signifi- cant only for men. For women the treatment effect is lower by about half, and not significant at conventional levels. Respondents were asked to rank their position 5 years from now, and we can calculate treat- ment effects on the future level or the change from today to 5 years from now. Treated individu- als, especially males, do not see their relative gains as persistent. Or the untreated are optimistic about their future. There is no substantive or significant difference in reported level of expected economic well-being between the two groups. Mechanically, this means that treatment is associ- ated with a lower expected change in future well-being than controls. 7.5 Impacts on alienation and aggression 7.5.1 Participation and social integration Tables 11 and 12 display treatment effects for our measures of community engagement and social integration and participation (or, conversely, alienation). In general, we see modest in- creases, of the order of 0 to 10%, in common community participation and other indicators of social and community support. We focus on percentage impacts relative to the control mean, cal- culated at the base of the table for each outcome. In terms of community participation and engagement, treated individuals are engaged in 9.3% more community groups than controls, an effect unlikely to be a mechanical effect of funding, since the majority of control group members still consider themselves a part of their NUSAF group (Table 1). Treated individuals are 4.4% more likely to attend community meetings and 7.7% more likely to speak out at meetings (though only the latter is statistically significant, at the 10% level). Treated individuals are 3.3% more likely to be a community leader and 8.9% more likely to be a (more junior) community mobilizer (again only the latter impact is significant). Turning to social integration of a more interpersonal nature, we see little significant difference in an indicator of family connectedness (the sign is actually negative), nor do we see any differ- ence in an index of community relations or an index of reverence for elders. However, treated individuals do report 4.7% more social support compared to the control group, and an index measure of depression symptoms is 5% lower among the treated (although only significant for males). 35 It is difficult to say whether this is a direct consequence of economic success or a result of other program impacts, such as the group and participatory process. Economic success is un- doubtedly a part of the impact, but not necessarily all. First, while social support and economic success are closely correlated in the overall sample, adding measures of economic success to the treatment regressions (not shown) diminishes social treatment effects by just a third, suggesting other channels of impact are present. Moreover, in northern Uganda, as youth’s most important transition is from being a recipient of transfers and assistance to a patron, especially among males, contributions to the household and kin are crucial to social support and status. Indeed, males in the control group are net recipients and treated males are net contributors, but the treat- ment effect is small in absolute terms (just 11,000 UGX, or $5, in the past 12 months) and not significant (Appendix Table 2). 7.5.2 Aggression In Table 3 we saw that collectively our aggression measures decreased by approximately 0.2 standard deviations among males and increased nearly 0.4 standard deviations among females. ATEs for individual dependent variables are reported in Table 13. The first five dependent variables (and 10 columns) ask about disputes with different parties. The steepest and most significant declines for males are with community leaders and police— both in substantive terms and statistical robustness. The largest and most significant increase for females, meanwhile, is in physical fights. Physical fights are less common among females than males in absolute terms (5% of males versus 3% of females in the control group) but treated fe- males are twice as likely as control females to report a physical fight, bringing them to roughly the same level of physical fights as males. Males also report significantly lower disputes with leaders and police, or physical fights, among their peers. Females do the opposite. It is not clear whether this represents a change in the composition of the peer group, or the fact that the peers referred to are fellow group members reacting in similar fashion. The final four dependent variables look at self-reported hostile behaviors, based on questions asked in the psychosocial section of the questionnaire (along with measures of distress and de- pression). The signs are consistently negative for males and positive for females. The largest male decline, and female increase, are seen for quarrelsomeness and threatening others—two of the more serious forms of hostile behavior we measure. 36 We should note that these treatment effects appear quite large, but in absolute terms the change is relatively small. Overall levels of self-reported hostile behaviors are low; if we add all four hostile behavior measures, for instance, we have an index of 0 to 12 (representing four be- haviors and four levels of severity ranging from 0 to 3). The control group mean is just 0.71— implying that the average person says that they “rarely� engage in one of the four behaviors. This level is unsurprising, given that aggression is typically rare. Figure 7 displays a histogram of the hostility measure for males and females and treatment and control separately. Males in the con- trol group, for instance, report an index value of 3 at the 90th percentile, 4 at the 95th, and 6 at the 99th—the latter value corresponding to a response of “often� committing two of the behaviors or “sometimes� committing all four. The effect of treatment is thus to push the average from rarely committing one of the transgressions to even more rarely committing them. We see a similar pat- tern for an additive index of disputes (not displayed). In absolute terms the treatment effect is small—it suggests moving from a very rare dispute to one even rarer, but the proportional impact is large. Overall, the proportional effect of treatment appears to hold relatively steadily throughout the distribution. If we combine all three measures additively and create an indicator for being in the highest 5% of self-reported aggression, for instance, 6% of control males are in this top tier but only 4% of treated males are there (regressions not shown). Similarly, 3% of control females are in this top tier but 7% of treated ones are. These differences are highly statistically significant. The results suggest that treatment reduces aggression both among the least and most aggressive males, and increases aggression among females across the distribution. The reduction in aggression is also greatest for those with the highest initial levels of aggres- sion, and the most exposure to war. We look at impact heterogeneity on the aggregate aggression family index in Table 14. The interaction between treated and baseline aggression levels and ex- posure to war violence is negative and significant. The least risk averse individuals respond to treatment with higher reported levels of aggression, however. The reduction in aggression among males is consistent with our predictions, especially those that emphasize reduced psychological stress. The results among females, however, present an unexpected puzzle. One possibility is that women’s increase in disputes, quarrels and threats are a consequence of greater market engagement, interaction outside the home, and hence opportuni- ties for aggression. Women in the marketplace, or who make money, may also be targets of un- 37 wanted male attention, such as officials or police seeking bribes of a financial or intimate nature. This too will be explored in the next round of data collection. 7.6 Implications for theories of alienation and aggression We cannot experimentally distinguish between competing theories and mechanisms but, as outlined in the theory section above, certain patterns in the data would be more consistent with some mechanisms over others. The patterns are not strongly consistent with any one view, but the evidence seems to be most consistent with two claims. First, increases in wealth are associated with greater transfers and higher social support among men, consistent with the “social roles� view. Males especially show a modest increase in social support and community relations (Table 11). These gains are also correlated with econom- ic success. Table 15 displays correlations between our major outcome family indices. Those with higher economic outcomes are more likely to have higher social outcomes. This relationship is not causally identified, but it is consistent with the pattern. Perhaps most important of all, treated males are much more likely to make transfers to others for health and education expenditures—a 31% increase over the control group for education transfer and a 46% increase for health expend- itures (Appendix Table 2). The increase is significant for transfers within and outside the house- hold. Second, evidence is somewhat consistent with the psychological approaches to aggression, through reduced stress or perhaps frustrated ambition. Aggression is strongly and positively cor- related with emotional distress symptoms and negatively correlated with social support (Table 15). It is not at all correlated, however, with economic performance. The negative impact on ag- gression and the positive impact on subjective well-being are in principle consistent with the frustration-aggression hypothesis. There is only a weak correlation between aggression and sub- jective well-being, however, and virtually no association between actual economic performance and aggression. It may simply be that the act of inclusion, and receipt of a government transfer, is enough to ameliorate feelings of frustrated ambition. The determinants of aggression, howev- er, will be explored in the longitudinal study with more extensive 2012 data on a wider variety of aggression outcomes, as well as more detail on the acts and actors. Meanwhile, it is worth noting that the patterns are not particularly consistent with the partici- patory view underlying so many community driven development programs—that participation in a group empowers individuals and therefore leads to social engagement. We see little change in 38 community engagement and participation, little change in self-perceived power or respect, and group performance is only weakly correlated with group cohesion, longevity, and the quality of the group dynamic. 8 Discussion and conclusions The principles that drive NUSAF are common to social action funds and community-driven development programs around the world: a preference for market-based approaches to develop- ment; a marginalist view of poverty and poverty alleviation; a sense that individuals or groups are capable of making good, even better decisions, than a planner (and hence favoring decentral- ized and participatory programs over centralized or paternalistic ones); the idea that this deci- sion-making and its success may even be empowering; and a sense that higher incomes and em- ployment themselves may also directly reduce the risks of aggression or conflict. This optimism is largely borne out in the YOP case, though in different proportions: the economic impacts are generally large, while the social ones are relatively modest. The results suggest that the relatively unconditional, decentralized cash transfer programs tar- geted at poor entrepreneurs can translate to high levels of investment. It is possible that the group organization acted as a disciplinary device, and further research on the use of group organization as a commitment device emerges as an important area for future experiments. Consistent with other studies, we see that many of the poor, especially males, have reasona- bly high returns to investment when capital is made available and without close supervision or conditionality from the donor. The findings are also consistent with the prevalence of high un- deremployment, and suggest that earnings from household production could be increased by simply increasing more hours of work without need for raising productivity or reallocating time from subsistence agriculture. The results also suggest that, whatever the structural or institutional constraints on poverty in northern Uganda, the poor can make substantial gains on the margin. Nevertheless, this is not to say that the program helped the poor reach their full capacities. No matter the returns we ob- serve, these were still relatively inexperienced and uneducated youth making decisions over more cash than they have seen in their lifetimes. Information on market opportunities or assis- tance with project planning and budgeting is probably an important but underexplored input into efficient production. This too is an important area for further research. 39 If individuals are capable of the same discipline and returns as the youth in NUSAF YOP groups, the results also suggest that credit constraints and the lack of financial development are a substantial impediment to poverty alleviation. To the extent that the poor have access to finance, it is for short horizons and at absurdly high rates, in excess of 200% per annum. There are un- doubtedly gains from improved access to finance. The results also suggest that economic success leads to increased engagement in the commu- nity, social support, and (among males) lower levels of aggression. We admittedly cannot disen- tangle the contribution of higher incomes and employment from the symbolic importance of re- distribution, or the experience of planning and engaging with a community group. That there are non-pecuniary private benefits of employment and higher incomes, however, seems clear. The aggression results suggest positive externalities as well, in terms of social stability, a topic to be explored in future research as well. The presence of non-pecuniary private gains, or externalities, could help explain underin- vestment by poor entrepreneurs without the program. If the cost of capital is 20 or 30%, the me- dian entrepreneur in our sample would not earn sufficient earnings to pay back the investment, and the average entrepreneur would just barely be able to repay. The private returns to employ- ment clearly go beyond earnings, however, and so cash transfers or subsidized credit may be a means to achieve higher levels of stability and freedoms than otherwise available to the poor. 9 References Aedo, C. and S. Nuñez (2004). "The Impact of Training Policies in Latin America and the Caribbean: The Case of Programa Joven." Inter-American Development Bank RES Working Pape 3175. Attanasio, O., A. D. Kugler, et al. (2008). "Training disadvantaged youth in Latin America: evidence from a randomized trial." NBER Working Paper 13931. Banerjee, A., E. Duflo, et al. (2010). "The Miracle of Microfinance? Evidence from a Randomized Evaluation." Unpublished working paper, MIT. Banerjee, A. V. and E. Duflo (2005). Growth theory through the lens of development economics. Handbook of Economic Growth. P. Aghion and S. Durlauf. Amsterdam, Elsevier Press. Banerjee, A. V. and S. Mullainathan (2009). "The Shape of Temptation: Implications for the Economic Lives of the Poor." Unpublished working paper, MIT. 40 Bardhan, P. and C. Udry (1999). Development Microeconomics. Oxford, Oxford University Press. Becker, G. S. (1968). "Crime and Punishment: An Economic Approach." The Journal of Political Economy 76(2): 169-217. Berkowitz, L. (1993). Aggression: Its causes, consequences, and control. New York, Mcgraw- Hill Inc. Bertrand, M., S. Mullainathan, et al. (2004). "A Behavioral-Economics View of Poverty." The American Economic Review 94(2): 419-423. Betcherman, G., M. Godfrey, et al. (2007). "A Review of Interventions to Support Young Workers: Findings of the Youth Employment Inventory." World Bank Social Protection Discussion Paper 0715(October). Blattman, C. and J. Annan (2010). "The Consequences of Child Soldiering." Review of Economics and Statistics 92(4): 882-898. Blattman, C. and E. Miguel (2010). "Civil War." Journal of Economic Literature 48(1): 3-57. Bruhn, M., D. Karlan, et al. (2010). "What Capital is Missing in Developing Countries?" American Economic Review: Papers & Proceedings. Burbidge, J. B., L. Magee, et al. (1988). "Alternative Transformations to Handle Extreme Values of the Dependent Variable." Journal of the American Statistical Association 83(401): 123-127. Card, D., P. Ibarraran, et al. (2007). "The Labor Market Impacts of Youth Training in the Dominican Republic: Evidence from a Randomized Evaluation." NBER Working Paper 12883. Card, D., J. Kluve, et al. (2009). "Active labor market policy evaluations: A meta-analysis." CESifo Working Paper 2570. Casey, K. P., R. Glennerster, et al. (2011). "Reshaping Institutions: Evidence on External Aid and Local Collective Action." NBER Working Paper 17012. CIA. (2011). "Uganda." CIA World Factbook Retrieved 5/26/2011, from https://www.cia.gov/library/publications/the-world-factbook/geos/ug.html. Collins, R. (2008). Violence: A Micro-sociological Theory. Princeton, Princeton University Press. de Mel, S., D. McKenzie, et al. (2007). "Measuring microenterprise profits: Don't ask how the sausage is made." World Bank Policy Research Working Paper Series 4229. de Mel, S., D. J. McKenzie, et al. (2008). "Returns to Capital in Microenterprises: Evidence from a Field Experiment." Quarterly Journal of Economics 123(4): 1329-1372. Dollard, J., N. E. Miller, et al. (1939). "Frustration and aggression." 41 Duflo, E., R. Glennerster, et al. (2007). Using Randomization in Development Economics Research: A Toolkit. Handbook of Development Economics. T. P. Schultz and J. Strauss, Elsevier Science Ltd. 4: 3895-3962. Durkheim, E. (1893). The division of labor in society. Fafchamps, M., D. McKenzie, et al. (2011). "When is capital enough to get female microenterprises growing? Evidence from a randomized experiment in Ghana." Unpublished working paper. Field, E., S. Jayachandran, et al. (2010). "Do Traditional Institutions Constrain Female Entrepreneurship? A Field Experiment on Business Training in India." American Economic Review 100(2): 125–129. Filmer, D. and K. Scott (2008). "Assessing asset indices." World Bank Policy Research Working Paper Series 4605. Freeman, R. B. (1999). The Economics of Crime. Handbook of Labor Economics. O. Ashenfelter and D. Card, Elsevier. 5: 3529-3572. Frison, L. and S. Pocock (1992). "Repeated measures in clinical trials analysis using mean summary statistics and its implications for design." Statistics in Medicine 11: 1685-1704. Fuller, G. (1995). The Demographic Backdrop to Ethnic Conflict: A Geographic Overview. The Challenge of Ethnic Conflict to National and International Order in the 1990s. CIA. Washington CIA: 151-154. Gerber, A., D. Green, et al. (2011). "Addressing missing outcome data in randomized experiments: A design-based approach." Unpublished working paper. Goldstone, J. A. (2002). "Population and Security: How Demographic Change Can Lead to Violent Conflict." Journal of International Affairs 56(1): 3-23. Golooba-Mutebi, F. and S. Hickey (2010). "Governing chronic poverty under inclusive liberalism: The case of the Northern Uganda Social Action Fund." Journal of Development Studies 46(7). Government of Uganda (2007). National Peace, Recovery and Development Plan for Northern Uganda: 2006-2009. Kampala, Uganda, Government of Uganda. Gurr, T. R. (1971). Why Men Rebel. Princeton, Princeton University Press. Heckman, J. J., R. J. LaLonde, et al. (1999). The economics and econometrics of active labor market programs. Handbook of Labor Economics. O. Ashenfelter and D. Card, Elsevier. 3: 1865-2095. Heckman, J. J., J. Stixrud, et al. (2006). "The effects of cognitive and noncognitive abilities on labor market outcomes and social behavior." Journal of Labor Economics 24(3): 411-482. Heinsohn, G. (2003). Söhne und Weltmacht, Orell Füssli. Hickey, S. (2010). "The Government of Chronic Poverty: From Exclusion to Citizenship?" Journal of Development Studies 46(7): 1139 - 1155. 42 Hoff, K. and A. Sen (2005). "The kin system as a poverty trap?" World Bank Policy Research Working Paper 3575. Kaplan, R. D. (1994). The Coming Anarchy. Atlantic Monthly: 44–76. Karlan, D. and M. Valdivia (2011). "Teaching entrepreneurship: Impact of business training on microfinance clients and institutions." Review of Economics and Statistics 93(2): 510- 552. Kristof, N. D. (2010). Here's a Woman Fighting Terrorism. With Microloans. The New York Times. New York: WK10. Levenson, R. (2011). Northern Uganda – Rural Interest Rates. New Haven, Innovations for Poverty Action. Levine, J. M. and R. L. Moreland (1998). Small groups. The Handbook of Social Psychology. S. T. Fiske, D. T. Gilbert and G. Lindzey. Oxford, Oxford University Press. 2: 415-469. Lomo, Z. and L. Hovil (2004). Behind the Violence: Causes, Consequences, and the Search for Solutions to the War in Northern Uganda. The Refugee Law Project Working Paper Series. Kampala, Uganda. Mansuri, G. and V. Rao (2011). Localizing Development: Does Participation Work? Washington DC, World Bank. McKenzie, D. (2011). "Beyond Baseline and Follow-up: The Case for More T in Experiments." World Bank Policy Research Working Paper 5639. McKenzie, D. J. and C. Woodruff (2008). "Experimental evidence on returns to capital and access to finance in Mexico." The World Bank Economic Review 22(3): 457. Merton, R. K. (1938). "Social Structure and Anomie." American Sociological Review 3(5): 672- 682. Omara-Otunnu, A. (1994). Politics and the Military in Uganda, 1890-1985. London, Macmillan in association with St. Antony's College, Oxford. Peters, K. K. B. and P. Richards (1998). ""Why We Fight": Voices of Youth Combatants in Sierra Leone." Africa 68(2). Putnam, R. D. (2001). Bowling alone: The collapse and revival of American community. New York, Simon and Schuster. Richards, P. (1996). Fighting for the Rain Forests: War, Youth, and Resources in Sierra Leone. Portsmouth, New Hampshire, Heinemann. Scacco, A. (2008). Who Riots? Explaining Individual Participation in Ethnic Violence, Unpublished manuscript, New York University. Scott, J. C. (1976). Moral Economy of the Peasant: Rebellion and Subsistence in South East Asia. New Haven, Yale University Press. 43 Thomas, D., E. Frankenberg, et al. (2001). "Lost but Not Forgotten: Attrition and Follow-up in the Indonesia Family Life Survey." The Journal of Human Resources 36(3): 556-592. Udry, C. and S. Anagol (2006). "The return to capital in Ghana." The American Economic Review 96(2): 388-393. Wood, E. J. (2003). Insurgent Collective Action and Civil War in El Salvador. New York, Cambridge University Press. World Bank (2007). World Development Report 2007: Development and the Next Generation. Washington DC, The World Bank. World Bank (2009). Sharing the Growth in Uganda: Recent Labor Market Outcomes and Poverty Reduction. Kampala, Uganda, World Bank Uganda Office. World Bank (2009). World Development Indicators 2009, The World Bank. World Bank (2010). Concept Note for a Study of TVET or broader skills development programs. Washington DC, The World Bank. World Bank (2010). World Development Report 2011: Conflict Security and Development. Washington, The World Bank. 44 Figure 1: Distribution of transfer size per group member (in US dollars) 30 20 10 0 0 1000 2000 3000 Average grant size within groups (USD) 45 c2 f(A, K) 2y + wH δH 2y + wL δL δHu’(c1)/u’(c2) E y F δLu’(c1)/u’(c2) F y + wL y + wH c1 Cash transfer = �w Figure 2: Impact of cash transfers on occupational choice (No existing entrepreneurs) At wL, more patient and higher ability people become entrepreneurs while others remain laborers. Highly impatient laborers will have a corner solution at E. For small F (relative to �w) patient and impatient cash transfer recipients become entrepreneurs. But investment and period 2 income are generally increasing with patience. c2 AH AL δH δH AH AL δ δH δL L δH y δL y + wL y + wH c1 Figure 3: High versus low ability individuals (No existing entrepreneurs) The impact of a cash transfer is larger among higher ability and more patient individuals. Ability and patience positively interact. Only highly impatient or very low ability individuals (those who do not have high return earning opportunities) would remain laborers after a cash transfer. c2 “Existing� “Budding� δH δH 2y + wL δH δL δL δH y δL δL y + wL + K0 y + wH + K0 c1 Figure 4: Existing versus budding entrepreneurs, with equal levels of starting capital For illustrative simplicity we assume first period entrepreneur income is equal to labor income: f(A,K0) = y. The impact of cash transfers on investment and profits is larger among budding entrepreneurs than existing entrepreneurs. The larger thee fixed cost of becoming an entrepreneur, the more impactful the transfer will be on profits (relative to existing entrepreneurs) Figure 5: Quantile treatment effects for business assets acquired and owned Business assets owned and QTE at each quantile 500 Biz assets owned (000s of UGX) 400 300 200 100 0 10 20 30 40 50 60 70 80 90 Quantile QTE QTE 95% CI QTE 95% CI Control level 49 Figure 6: Quantile treatment effects for monthly income, by gender Males: Control income and QTE at each quantile 100 80 Income (000s of UGX) 60 40 20 0 10 20 30 40 50 60 70 80 90 Quantile QTE QTE 95% CI QTE 95% CI Control income level Females: Control income and QTE at each quantile 100 Income (000s of UGX) 80 60 40 20 0 10 20 30 40 50 60 70 80 90 Quantile QTE QTE 95% CI QTE 95% CI Control income level 50 Figure 7: Distribution of self-reported hostile behaviors Males 80 60 40 20 0 0 2 4 6 8 Index of aggressive behaviors Control Treatment 80 60 40 20 0 0 2 4 6 8 10 Index of aggressive behaviors Control Treatment 51 Table 1: Group summary statistics Mean Std. Dev. Group characteristics Group existed prior to NUSAF 0.463 [0.0] Age of group at baseline 3.814 [0.1] Size of group 21.8 [7.0] Proportion female 0.40 [0.25] Grant size (UGX) 12,794,279 [3258832] Grant size (USD) 7,108 [2080] Grant size per member (UGX) 673,026 [371697] Grant size per member (USD) 374 [206] Group members (all) Age 24.1 [26.8] Committee Member 0.38 [0.24] Officer 0.14 [0.12] Treasurer 0.05 [0.04] Secretary 0.04 [0.04] Vice Chair 0.01 [0.01] Chair 0.04 [0.04] Muslim 0.10 [0.09] Literate 0.76 [0.18] Speak some English 0.31 [0.21] Disabled 0.04 [0.04] Treatment Proportion of funds spent on training 0.35 [0.77] Mean [Std. Dev.] Do you… Treatment Control Still consider yourself a part of the group? 0.952 0.981 [0.21] [0.14] Still work with this group? 0.91 0.96 [0.28] [0.19] Feel the group cooperates well 0.82 0.85 [0.39] [0.36] Table 2: Key outcomes and summary statistics Assigned to control Assigned to treatment Mean Std. Dev. Mean Std. Dev. Obs Investments in vocational skills and capital Hours of training received 50 [214] 389 [480] 2,005 Tools and machines acquired since baseline ('000s of UGX) 136 [909] 774 [2218] 2,006 Stock of raw materials, tools and machines ('000s of UGX) 348 [1296] 858 [2143] 1,999 Income, poverty and employment Cash earnings from past 4 weeks ('000s of UGX) 43.5 [94.6] 61.6 [114.0] 2,006 Monthly cash earnings adjusted for hourly earnings 24.6 [89.6] 39.6 [107.6] 2,006 Index of wealth z-score (Poverty/consumption proxy) -0.020 [0.998] 0.094 [1.042] 2,000 Hours spent on all economic activities in past 4 weeks 120.6 [108.2] 138.8 [112.6] 2,006 Hours spent on market activities in past 4 weeks 70.4 [102.7] 90.9 [100.1] 2,006 Community participation and engagemennt Number of group memberships 3.7 [2.8] 3.8 3.0 2,006 Attends community meetings (indicator) 0.67 [0.47] 0.70 [0.46] 2,000 Speaks out at community meetings (indicator) 0.61 [0.49] 0.66 [0.47] 1,997 Is a community leader (indicator) 0.40 [0.49] 0.41 [0.49] 2,000 Is a community mobilizer (indicator) 0.51 [0.50] 0.60 [0.49] 1,996 Locus of control index (1-4) 2.18 [0.31] 2.17 [0.33] 2,000 Social integration Family very caring (indicator) 0.75 [0.43] 0.71 [0.46] 2,003 Index of social support (0-16) 9.18 [3.69] 9.81 [3.47] 2,006 Community/neighbor relations index (0-8) 6.81 [1.32] 6.75 [1.27] 2,006 Reverence for elders index (0-9) 6.30 [0.97] 6.33 [0.97] 2,006 Depression and distress symptoms Index of depression and distress symptoms (0-19) 6.85 [3.95] 7.07 [3.70] 2,006 Aggressive and hostile behaviors Index of disputes with neighbors (0-3) 0.20 [0.57] 0.20 [0.60] 1,995 Index of disputes with family (0-3) 0.29 [0.63] 0.27 [0.63] 1,996 Index of disputes with community leaders (0-3) 0.08 [0.36] 0.06 [0.30] 1,996 Index of disputes with police (0-3) 0.05 [0.30] 0.03 [0.20] 1,992 Involved in physical fights (0-3) 0.04 [0.23] 0.05 [0.27] 1,995 Peers have disputes with local leaders or police (0-3) 0.37 [0.75] 0.35 [0.77] 1,980 Peers involved in physical fights (0-3) 0.34 [0.72] 0.30 [0.70] 1,987 Quarrelsome (0-3) 0.30 [0.62] 0.29 [0.62] 1,986 Takes things without permission (0-3) 0.14 [0.51] 0.12 [0.44] 1,998 Uses abusive language (0-3) 0.12 [0.41] 0.12 [0.43] 1,998 Threatens to hurt others (0-3) 0.15 [0.47] 0.13 [0.45] 1,999 Community participation and engagemennt Wealth: Current position (0-9) 2.73 [1.55] 3.05 [1.65] 1,997 Community respect: Current position (0-9) 4.52 [2.40] 4.44 [2.16] 1,988 Community power: Current position (0-9) 4.45 [2.23] 4.45 [2.12] 1,967 Access to basic services: Current position (0-9) 3.72 [2.11] 4.13 [2.02] 1,984 Asked for advice: Current position (0-9) 5.04 [2.29] 4.86 [2.20] 1,995 Subjective well being (expected future change) Expected 5-year change in wealth position 2.85 [1.82] 2.55 [1.73] 1,987 Expected 5-year change in respect position 1.98 [1.76] 1.93 [1.70] 1,976 Expected 5-year change in power position 1.82 [2.00] 1.68 [1.89] 1,951 Optimism index (0-3) 2.57 [0.78] 2.51 [0.82] 1,978 All UGX-denominated outcomes were censored at the 99th percentile to contain potential outliers Table 3: Average treatment effect by outcome family (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) Outcome family z-score Investments in Community Employment, cash Depression and distress Aggressive and hostile Perceived ranking in Expected future change vocational skills and participation & Social integration earnings and poverty symptoms behaviors community in community ranking capital engagement Male/ Male/ Male/ Male/ Male/ Male/ Male/ Male/ All All All All All All All All Female Female Female Female Female Female Female Female Treated 1 025 1 047 0 286 0 290 0 084 0 075 0 035 0 107 -0 086 -0 151 -0 067 -0 198 0 154 0 199 -0 118 -0 164 [0 054]*** [0 063]*** [0 054]*** [0 067]*** [0 053] [0 060] [0 049] [0 053]** [0 051]* [0 058]*** [0 052] [0 064]*** [0 054]*** [0 064]*** [0 048]** [0 057]*** Treated × Female -0 066 -0 012 0 027 -0 216 0 196 0 396 -0 135 0 138 [0 104] [0 113] [0 109] [0 105]** [0 115]* [0 104]*** [0 108] [0 110] Female -0 023 0 004 -0 254 -0 249 -0 354 -0 365 -0 208 -0 119 0 087 0 006 -0 006 -0 169 -0 000 0 055 -0 057 -0 114 [0 052] [0 058] [0 059]*** [0 075]*** [0 057]*** [0 073]*** [0 055]*** [0 070]* [0 063] [0 081] [0 062] [0 077]** [0 053] [0 065] [0 053] [0 065]* Observations 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 1983 1983 1984 1984 R-squared 0 356 0 356 0 196 0 196 0 243 0 242 0 177 0 178 0 200 0 202 0 203 0 208 0 163 0 163 0 192 0 192 Female Treatment Effect 0 981 0 277 0 102 -0 109 0 044 0 198 0 064 -0 026 p-value 0 000*** 0 002** 0 287 0 253 0 655 0 021** 0 487 0 779 Robust standard errors in brackets, clustered by group and stratified by district. Where an index includes a UGX-denominated dependent variable, the inverse hyperbolic sine (IHS) rather than the linear transformation is used in the index. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. *** p<0.01, ** p<0.05, * p<0.1 Table 4: Average treatment effects on investments in vocational skills and capital (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Tools and machines acquired since IHS(Tools and machines acquired Stock of raw materials, tools, and IHS(Stock of raw materials, tools, Hours of training received baseline ('000s of UGX) since baseline) machines ('000s of UGX) and machines) Treated 404 904 400 264 656 016 791 904 2 219 2 375 523 318 658 554 1 576 1 651 [24 353]*** [25 162]*** [95 806]*** [130 305]*** [0 182]*** [0 215]*** [103 228]*** [141 476]*** [0 164]*** [0 207]*** Treated × Female 13 996 -409 800 -0 467 -408 071 -0 225 [46 693] [171 343]** [0 367] [191 037]** [0 351] Female 33 220 27 474 -218 079 -49 611 -0 424 -0 231 -313 111 -145 331 -0 371 -0 278 [24 509] [25 389] [90 247]** [85 262] [0 171]** [0 193] [87 703]*** [103 627] [0 158]** [0 208] Observations 1985 1985 1986 1986 1986 1986 1985 1985 1985 1985 R-squared 0 278 0 278 0 131 0 135 0 232 0 234 0 114 0 117 0 186 0 186 Control means All 49 77 136 5 1 904 348 0 3 628 Males 41 80 159 8 1 987 414 2 3 783 Females 63 34 96 71 1 763 234 9 3 364 Female Treatment Effect 414 3 382 1 1 908 250 5 1 426 p-value 0 000 0 001 0 000 0 046 0 000 ATE as % of control mean All 814% 481% 150% Males 958% 496% 159% Females 654% 395% 107% Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Table 5: Investments and performance by group characteristics (1) (2) (3) (4) (5) (6) (7) (8) IHS(Tools and machines acquired VARIABLES Hours of training received IHS(Cash earnings) Index of wealth (z-score) since baseline) Treated All Treated All Treated All Treated All Treated 179.9 2.745 0.715 0.477 [96.3]* [0.697]*** [0.489] [0.216]** Treated × Group existed prior to YOP 83.5 0.254 0.315 0.098 [50.0]* [0.380] [0.247] [0.108] Group existed prior to YOP (indicator) 58 3 -3.8 0.230 0.011 0.074 -0.200 0.075 -0.023 [38.3] [23.3] [0.288] [0.201] [0.170] [0.160] [0.072] [0.075] Treated × Group dynamic index 24.4 0.484 0.136 -0.062 [18.5] [0.151]*** [0.105] [0.047] Group dynamic index -2.1 -17.0 0.191 -0.215 -0.020 -0.131 -0.087 -0.015 [16.6] [9.7]* [0.122] [0.097]** [0.076] [0.083] [0.035]** [0.033] Treated × Group size 3.2 -0.028 0.004 -0.009 [3.4] [0.025] [0.018] [0.008] Group size -1.8 -1.5 -0.025 0.014 0.004 0.005 -0.003 0.004 [2.7] [1.6] [0.020] [0.015] [0.012] [0.012] [0.006] [0.005] Treated × % of group female 274.1 -0.076 -0.804 -0.547 [96.2]*** [0.726] [0.443]* [0.225]** % of group female 150.3 -82.0 0.303 0.345 -0.392 0.248 -0.509 -0.009 [93.8] [45.6]* [0.705] [0.406] [0.348] [0.307] [0.152]*** [0.154] Treated × Group heterogeneity index -40.3 -0.251 -0.186 -0.053 [20.8]* [0.182] [0.116] [0.055] Group heterogeneity index -24.0 5.8 -0.026 0.161 -0.066 0.109 0.013 0.051 [17.9] [10.4] [0.131] [0.130] [0.084] [0.086] [0.036] [0.038] R-squared 0.1 0.3 0.181 0.241 0.136 0.126 0.337 0.294 Observations 899 1847 900 1848 900 1848 900 1848 Control Mean 49.77 49.77 1.904 1.904 2.704 2.704 -0.0196 -0.0196 Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Table 6: Average treatment effects on income, poverty and employment (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) Cash earnings in last 4 weeks Poverty/Consumption proxy Employment levels Adjusted for hourly Hours spent on market Hours spent on all economic Currently engaged in VARIABLES Level (000s of UGX) IHS(Cash earnings) Index of wealth (z-score) earnings ('000s of UGX) activities in past 4 weeks activities in past 4 weeks skilled work (indicator) Treated 19 515 26 225 0 675 0 664 16 614 23 435 0 131 0 182 22 239 20 473 19 705 17 596 0 340 0 314 [5 319]*** [7 326]*** [0 119]*** [0 143]*** [5 157]*** [7 066]*** [0 055]** [0 067]*** [5 708]*** [7 118]*** [6 009]*** [7 287]** [0 029]*** [0 035]*** Treated × Female -20 234 0 033 -20 571 -0 156 5 328 6 362 0 078 [11 317]* [0 256] [10 698]* [0 106] [11 293] [12 330] [0 057] Female -17 859 -9 547 -0 383 -0 397 -11 013 -2 557 -0 071 -0 006 -24 913 -27 102 -26 073 -28 686 -0 092 -0 124 [6 083]*** [7 379] [0 124]*** [0 165]** [5 618]* [6 917] [0 052] [0 066] [6 160]*** [7 736]*** [6 849]*** [8 207]*** [0 030]*** [0 036]*** Observations 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 R-squared 0 130 0 133 0 110 0 110 0 102 0 105 0 280 0 282 0 108 0 108 0 109 0 109 0 168 0 169 Control means All 43 45 2 704 24 61 -0 0196 70 36 120 6 0 343 Males 50 01 2 907 27 66 -0 00328 80 69 132 9 0 404 Females 32 27 2 359 19 42 -0 0476 52 76 99 60 0 241 Female Treatment Effect 5 992 0 697 2 864 0 0261 25 80 23 96 0 392 p-value 0 447 0 00121 0 700 0 762 0 00435 0 0187 0 ATE as % of control mean All 45% 68% 32% 16% 99% Males 52% 85% 25% 13% 78% Females 19% 15% 49% 24% 163% Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capita All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Table 7: Grant size per person as treatment (1) (2) (3) (4) (5) (6) IHS(Tools and machines IHS(Stock of raw materials, IHS(Cash earnings in the acquired since baseline) tools, and machines) past 4 weeks) VARIABLES Grant size per group member 0.001 0.001 -0.000 [0.000]** [0.000]* [0.000] IHS(Grant size per group member) 0.449 0.274 0.108 [0.335] [0.293] [0.189] R-squared 0.226 0.224 0.206 0.204 0.118 0.119 Control Mean 1.904 1.904 3.628 3.628 2.704 2.704 Obs 835 835 835 835 835 835 Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Table8: Impact Heterogeneity among those without an existing vocation (1) (2) (3) (4) (5) (6) (7) (8) IHS(Tools and machines IHS(Cash earnings in last 4 Hours of training received Index of wealth (z-score) acquired) weeks) Treated All Treated All Treated All Treated All Treated 401.1 2.329 0.786 0.173 [25.5]*** [0.190]*** [0.125]*** [0.059]*** Working capital index -58.9 -26.8 0.182 0.027 -0.076 0.191 0.085 0.166 [53.9] [22.8] [0.435] [0.202] [0.304] [0.137] [0.155] [0.077]** Treated × Working capital -35.7 0.086 -0.195 -0.045 [24.6] [0.222] [0.144] [0.081] Ability index 191.8 -19.7 -0.745 -0.567 -0.380 -0.014 -0.142 -0.185 [85.4]** [57.6] [0.649] [0.521] [0.467] [0.332] [0.234] [0.189] Treated × Ability 61.0 -0.040 -0.106 0.001 [27.2]** [0.237] [0.175] [0.082] Patience index 20.3 13.1 0.851 0.546 0.370 0.280 0.165 0.236 [35.4] [20.7] [0.260]*** [0.188]*** [0.162]** [0.136]** [0.082]** [0.067]*** Treated × Patience index -19.6 -0.014 0.003 -0.056 [33.9] [0.260] [0.172] [0.087] R-squared 0.2 0.3 0.204 0.243 0.135 0.123 0.317 0.290 Obs 862 1769 863 1770 863 1770 863 1770 Control Mean 49.77 49.77 1.904 1.904 2.704 2.704 -0.0196 -0.0196 Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Table9: Impact Heterogeneity by Patience (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) Hours of training received IHS(Tools and machines acquired) IHS(Cash earnings in last 4 weeks) Index of wealth (z-score) Low patience High patience Low patience High patience Low patience High patience Low patience High patience Treated All Treated All Treated All Treated All Treated All Treated All Treated All Treated All Treated 424 7 370 3 2 352 2 168 0 834 0 691 0 238 0 189 [41 6]*** [38 5]*** [0 313]*** [0 289]*** [0 213]*** [0 204]*** [0 096]** [0 085]** Working capital index -73 2 -21 4 -60 5 -26 5 -0 422 -0 237 0 736 0 336 -0 286 -0 218 -0 127 0 329 0 346 0 222 -0 055 0 149 [90 3] [23 9] [55 1] [34 2] [0 628] [0 213] [0 518] [0 292] [0 508] [0 158] [0 338] [0 195]* [0 254] [0 059]*** [0 144] [0 118] Treated × Working capital -45 5 -16 0 -0 157 -0 022 -0 276 -0 267 0 103 -0 230 [33 0] [29 9] [0 329] [0 261] [0 215] [0 179] [0 100] [0 094]** Ability index -771 3 -256 6 261 2 -33 7 -2 297 -0 039 0 269 0 026 -2 091 -0 410 0 140 -0 178 -0 819 -0 606 0 144 -0 098 [391 1]* [28 1]*** [90 5]*** [81 5] [1 479] [0 212] [0 603] [0 821] [0 811]** [0 153]*** [0 458] [0 502] [0 510] [0 066]*** [0 195] [0 309] Treated × Ability 78 5 51 6 0 053 0 099 -0 092 0 038 -0 045 0 059 [35 0]** [34 7] [0 334] [0 308] [0 210] [0 212] [0 085] [0 097] Existing vocation indicator -46 4 53 4 -7 6 73 2 -0 918 0 783 0 485 0 915 -0 219 0 888 -0 156 0 934 -0 049 -0 122 0 020 0 499 [96 6] [63 4] [69 4] [33 2]** [0 732] [0 570] [0 490] [0 595] [0 463] [0 430]** [0 349] [0 391]** [0 214] [0 227] [0 153] [0 193]** Treated × Existing vocation -32 9 -78 9 -1 196 -0 460 -1 384 -1 123 -0 028 -0 420 [112 4] [73 2] [0 943] [0 788] [0 616]** [0 520]** [0 303] [0 244]* R-squared 02 03 02 02 0 296 0 293 0 200 0 222 0 156 0 137 0 173 0 118 0 409 0 296 0 290 0 274 Obs 492 967 458 969 492 967 459 970 492 967 459 970 492 967 459 970 Control Mean 49 77 49 77 49 77 49 77 1 904 1 904 1 904 1 904 2 704 2 704 2 704 2 704 -0 0196 -0 0196 -0 0196 -0 0196 Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Table 10: Impacts on Subjective well-being (1) (2) (3) (4) (5) (6) (7) (8) (9) Current position relative to community in Future expected changes Believes good 5-year change on 5-year change on 5-year change on Power in Respect in Access to basic Being sought things will Wealth wealth ladder (-8 respect ladder (-8 power ladder (-8 community community services for advice happen in their to 8) to 8) to 8) life (0 to 3) Treated 0.486 0.176 0.204 0.470 0.003 -0.415 -0.168 -0.141 -0.024 [0.105]*** [0.145] [0.141] [0.130]*** [0.143] [0.104]*** [0.107] [0.128] [0.048] Treated x Female -0.270 -0.223 -0.161 -0.125 -0.192 0.326 0.249 -0.055 0.051 [0.176] [0.267] [0.247] [0.228] [0.254] [0.199] [0.208] [0.231] [0.083] Female 0.176 0.141 -0.088 0.095 0.046 -0.121 -0.207 0.049 -0.105 [0.112] [0.179] [0.157] [0.155] [0.153] [0.127] [0.128] [0.148] [0.052]** Observations 1983 1974 1953 1970 1981 1973 1962 1937 1964 R-squared 0.125 0.134 0.111 0.110 0.116 0.113 0.120 0.101 0.206 Control means All Males 2.750 4.608 4.615 3.817 5.124 2.868 1.971 1.736 2.600 Females 2.685 4.363 4.178 3.552 4.883 2.815 1.983 1.970 2.510 Female Treatment Effect 0.217 -0.0473 0.0432 0.345 -0.189 -0.0889 0.0811 -0.196 0.0262 p-value 0.132 0.836 0.840 0.0812 0.379 0.601 0.643 0.303 0.711 ATE as % of control mean All Males 18% 4% 4% 12% 0% -15% -9% -8% -1% Females 8% -1% 1% 10% -4% -3% 4% -10% 1% Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. *** p<0.01, ** p<0.05, * p<0.1 Table 11: Average treatment effects on community participation & engagement (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Number of group Attends community meetings Spoke out at a community Is a community leader Is a community mobilizer Locus of control index: 1 to memberships (indicator) meeting (indicator) (indicator) (indicator) 3.4 Treated 0 348 0 407 0 030 0 021 0 048 0 039 0 013 0 034 0 045 0 044 -0 019 -0 030 [0 150]** [0 189]** [0 027] [0 030] [0 026]* [0 029] [0 027] [0 033] [0 026]* [0 029] [0 017] [0 021] Treated x Female -0 176 0 026 0 027 -0 062 0 003 0 036 [0 304] [0 058] [0 056] [0 055] [0 055] [0 034] Female -0 418 -0 346 -0 160 -0 170 -0 179 -0 190 -0 117 -0 092 -0 113 -0 114 -0 015 -0 029 [0 150]*** [0 192]* [0 030]*** [0 038]*** [0 030]*** [0 037]*** [0 028]*** [0 037]** [0 030]*** [0 038]*** [0 017] [0 021] Observations 1986 1986 1986 1986 1983 1983 1986 1986 1982 1982 1986 1986 R-squared 0 224 0 225 0 099 0 099 0 149 0 149 0 146 0 147 0 192 0 192 0 121 0 121 Control means All 3 742 0 669 0 614 0 402 0 507 2 208 Male 3 980 0 741 0 702 0 459 0 571 2 224 Female 3 337 0 546 0 464 0 305 0 398 2 180 Female Treatment Effect 0 231 0 0470 0 0656 -0 0279 0 0473 0 00522 p-value 0 337 0 357 0 191 0 532 0 318 0 849 ATE as % of control mean All 9% 4% 8% 3% 9% -1% Males 10% 3% 6% 7% 8% -1% Females 7% 9% 14% -9% 12% 0% Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. *** p<0.01, ** p<0.05, * p<0.1 Table 12: Average treatment effects on social integration (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Index of depression and Family very caring Index of social support: 0 Index of Community/ Index of Reverence for Elders: distress symptoms (additive (indicator) to 16 Neighbor Relations: 0 to 8 0 to 7 bad): 0 to 19 Treated -0.036 -0.030 0.430 0.498 0.015 0.127 0.035 0.091 -0.312 -0.551 [0.026] [0.028] [0.180]** [0.210]** [0.066] [0.076]* [0.050] [0.057] [0.194] [0.220]** Treated x Female -0.017 -0.205 -0.339 -0.169 0.719 [0.053] [0.378] [0.147]** [0.117] [0.437] Female -0.054 -0.047 -0.680 -0.596 -0.183 -0.044 -0.064 0.005 0.345 0.050 [0.027]** [0.034] [0.197]*** [0.249]** [0.073]** [0.096] [0.056] [0.071] [0.238] [0.304] Observations 1986 1986 1986 1986 1986 1986 1986 1986 1986 1986 R-squared 0.116 0.116 0.176 0.176 0.135 0.137 0.101 0.100 0.210 0.211 Control means All 0.748 9.179 6.808 6.296 6.845 Males 0.780 9.449 6.822 6.307 6.767 Females 0.693 8.721 6.783 6.276 6.979 Female Treatment Effect -0.0471 0.293 -0.212 -0.0781 0.168 P Values 0.324 0.365 0.0953 0.443 0.655 ATE as % of control mean All -5% 5% 0% 1% -5% Males -4% 5% 2% 1% -8% Females -7% 3% -3% -1% 2% Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. *** p<0.01, ** p<0.05, * p<0.1 Table 13: Average treatment effects on aggressive and hostile behaviors (1) (2) (3) (4) (6) (7) (8) (9) (10) (11) (12) Intensity and frquency of disputes Peers Hostile behaviors Curses/uses With Have disputes Involved in Threatens to With neighbors With family (0 With police (0 Physical Quarrelsome (0 abusive community with leaders or physical fights Steals (0 to 3) hurt others(0 (0 to 3) to 3) to 3) fights(0 to 3) to 3) language (0 to leaders (0 to 3) police (0 to 3) (0 to 3) to 3) 3) Treated -0 041 -0 033 -0 049 -0 042 -0 012 -0 089 -0 110 -0 047 -0 030 -0 028 -0 097 [0 040] [0 041] [0 023]** [0 021]** [0 016] [0 053]* [0 048]** [0 037] [0 031] [0 028] [0 031]*** Treated x Female 0 067 0 111 0 057 0 017 0 071 0 212 0 149 0 150 0 053 0 081 0 197 [0 057] [0 075] [0 039] [0 025] [0 032]** [0 082]*** [0 073]** [0 070]** [0 055] [0 050] [0 054]*** Female -0 024 -0 016 -0 052 -0 039 -0 034 -0 145 -0 155 0 028 -0 009 0 017 -0 068 [0 037] [0 051] [0 028]* [0 023]* [0 018]* [0 052]*** [0 050]*** [0 045] [0 047] [0 034] [0 034]* Observations 1981 1982 1982 1978 1981 1966 1973 1972 1984 1984 1985 R-squared 0 229 0 076 0 055 0 055 0 054 0 218 0 236 0 093 0 069 0 112 0 104 Control means All Males 0 222 0 287 0 103 0 0655 0 0514 0 423 0 389 0 289 0 146 0 110 0 176 Females 0 169 0 285 0 0540 0 0294 0 0293 0 272 0 259 0 324 0 133 0 127 0 109 Female Treatment Effect 0 0261 0 0778 0 00801 -0 0248 0 0597 0 123 0 0390 0 103 0 0234 0 0533 0 0996 p-value 0 547 0 202 0 794 0 139 0 0341 0 0753 0 497 0 0957 0 641 0 212 0 0414 ATE as % of control mean All Males -18% -11% -48% -64% -23% -21% -28% -16% -20% -25% -55% Females 15% 27% 15% -85% 204% 45% 15% 32% 18% 42% 91% Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. *** p<0.01, ** p<0.05, * p<0.1 Table 14: Aggression Heterogeneity (1) (2) Aggression and Hostility Family (z- score) Treated Full sample Treated -0.086 [0.050]* Treated X Aggressive behaviors index -0.372 [0.116]*** Aggressive behaviors index 0.018 0.332 [0.072] [0.079]*** Treated X War violence index -0.282 [0.141]** War violence index -0.077 0.155 [0.105] [0.092]* Treated X Risk index 0.719 [0.188]*** Risk index 0.180 -0.416 [0.180] [0.136]*** Treated X Patience index -0.106 [0.124] Patience index 0.346 0.420 [0.074]*** [0.077]*** R-squared 0.310 0.252 Obs 921 1881 Control Mean 0.0358 0.0358 Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. *** p<0.01, ** p<0.05, * p<0.1 Table 15: Correlation Matrix of Outcome Families Employment and Community Depression and Aggression and A. MALES Investments Social integration Perceived ranking income participation distress hostility Investments 1.000 p-value Employment and income 0.317 1.000 p-value 0.000 Community participation 0.225 0.212 1.000 p-value 0.000 0.000 Social integration 0.136 0.096 0.173 1.000 p-value 0.000 0.001 0.000 Depression and distress -0.034 -0.016 -0.020 -0.147 1.000 p-value 0.212 0.550 0.473 0.000 Aggression and hostility -0.065 0.028 0.006 -0.238 0.346 1.000 p-value 0.017 0.310 0.829 0.000 0.000 Current subjective well being 0.157 0.195 0.256 0.214 -0.068 0.046 1.000 p-value 0.000 0.000 0.000 0.000 0.013 0.096 Employment and Community Depression and Aggression and B. FEMALES Investments Social integration Perceived ranking income participation distress hostility Investments 1 p-value Employment and income 0.342 1.000 p-value 0.000 Community participation 0.234 0.287 1.000 p-value 0.000 0.000 Social integration 0.091 0.040 0.156 1.000 p-value 0.019 0.302 0.000 Depression and distress -0.026 -0.033 0.018 -0.144 1.000 p-value 0.506 0.402 0.637 0.000 Aggression and hostility -0.022 0.042 0.048 -0.305 0.357 1.000 p-value 0.576 0.282 0.216 0.000 0.000 Current subjective well being 0.095 0.212 0.217 0.201 -0.014 0.003 1.000 p-value 0.014 0.000 0.000 0.000 0.720 0.939 Appendix Table 1: Baseline summary statistics and test of balance (1) (2) (3) Difference (controlling for Treatment Control district) Age 25 1 24 8 -0 006 [5 3] [5 3] [-0 021] Female 0 317 0 361 -0 032 [ 465] [ 481] [-1 1] Educational attainment 80 80 0 098 [3 1] [3 0] [0 577] Literate 0 723 0 741 -0 012 [ 448] [ 438] [-0 517] Prior vocational training 0 08 0 07 0 021 [ 276] [ 263] [1 7]* Activities of Daily Living Index (additive bad; 0-32) 86 87 -0 203 [2 3] [2 7] [-1 3] Index of emotional distress (additive bad; 0-43) 18 9 18 4 -0 249 [8 0] [8 0] [-0 613] Human capital index (z-score) -0 010 0 023 -0 032 [1 0] [ 947] [-0 541] Index of housing quality (-1 1-2 4) 0 023 0 000 0 007 [1 0] [1 0] [0 119] Index of assets (-2 7-3 5) 0 038 0 010 0 046 [1 1] [1] [0 785] Indicator for loans 0 350 0 327 0 014 [ 477] [ 469] [0 569] Total value of outstanding loans (UGX) 18731 19872 -188 [90713] [90068] [-0 046] Savings indicator 0 133 0 107 0 012 [ 340] [ 310] [0 786] Total savings in past 6 months 22092 15297 6,788 [113374] [92338] [1 4] Can obtain a 100000 UGX loan if needed 0 405 0 340 0 046 [ 491] [ 474] [1 9]* Can obtain a 1m UGX loan if needed 0 122 0 091 0 020 [ 328] [ 288] [1 3] Working capital index (z-score) 0 041 -0 001 0 031 [1 1] [ 977] [0 514] Total revenue in past 4 weeks 30284 26031 4,547 [63201] [53111] [1 4] Days of household work in past 4 weeks 66 59 0 722 [11 4] [11 0] [1 2] Days of nonhousehold work in past 4 weeks 17 1 16 3 0 933 [16 0] [16 3] [0 909] Total hours spent on non-household activities in past week 10 5 10 6 -0 104 [19 5] [20 1] [-0 103] Patience index (z-score) -0 017 0 023 -0 065 [1 0] [ 965] [-1 0] Had vocation at baseline (indicator) 0 085 0 074 0 008 [ 2796] [ 262] [0 606] Aggressive behaviors index (z-score) 0 00 0 02 -0 018 [1 0] [ 978] [-0 377] War violence index (z-score) -0 004 -0 001 0 001 [1 0] [ 965] [0 013] Observations 1323 1278 2,599 Standard errors in brackets, clustered in column 3 by group and stratified by district. *** p<0.01, ** p<0.05, * p<0.1 Appendix Table 2: Impacts on other (secondary) outcomes Skill investments Other employment Savings and credit Transfers Business formality Other transfers Enrolled in Returned to Hours spent on Total education Total health received from vocational Hours spent on Net household Index of school since subsistence IHS(Current Access to credit expenditures in expenditures in Number of Govt/NGOs training since chores in past 4 transfers ('000s business baseline work in past 4 savings) index past 12 months past 12 months employees since baseline baseline weeks of UGX) formality (indicator) weeks ('000s of UGX) ('000s of UGX) ('000s of UGX) (indicator) Treated 0 026 0 607 -5 1 -2 3 0 611 0 109 -11 099 105 650 16 169 0 395 -0 199 94 466 [0 021] [0 030]*** [2 4]** [4 4] [0 183]*** [0 049]** [7 007] [51 976]** [5 005]*** [0 206]* [0 093]** [30 652]*** Treated x Female 0 015 0 033 -12 5 0 393 -0 563 -0 097 13 163 -106 643 -10 093 -0 714 -0 059 -23 984 [0 034] [0 046] [8 1] [7 7] [0 311]* [0 088] [10 471] [78 197] [7 160] [0 283]** [0 132] [44 385] Female -0 062 -0 014 67 8 -2 7 0 095 -0 045 -7 226 74 759 3 379 -0 238 0 192 -47 153 [0 023]*** [0 031] [5 0]*** [4 6] [0 185] [0 060] [6 968] [54 312] [4 107] [0 179] [0 094]** [29 956] Observations 1985 1985 1986 1986 1984 1986 1986 1986 1986 1986 1986 1986 R-squared 0 118 0 389 0 380 0 138 0 188 0 112 0 039 0 126 0 083 0 052 0 085 0 052 Control means Males 0 124 0 169 11 53 9 2 456 0 904 8 785 345 6 35 08 1 753 5 634 31 74 Females 0 0663 0 157 88 5 47 2 2 153 0 726 3 385 324 9 33 20 1 312 5 841 21 60 Female Treatment Effect 0 0407 0 640 -17 6 -1 9 0 0472 0 0125 2 064 -0 993 6 076 -0 320 -0 258 70 48 p-value 0 138 0 0 0229 0 769 0 848 0 866 0 785 0 987 0 253 0 122 0 0126 0 0747 ATE as % of control mean Males 21% 359% -46% -4% 12% -126% 31% 46% 23% -4% 298% Females 61% 407% -20% -4% 2% 61% 0% 18% -24% -4% 326% Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Appendix Table 3: Leader Heterogeneity (1) (2) (3) (4) (5) (6) IHS(Tools and machines IHS(Stock of raw materials, Hours of training received acquired since baseline) tools, and machines) Treated 368.6 387.2 2.141 2.269 1.509 1.609 [28.6]*** [25.7]*** [0.204]*** [0.192]*** [0.192]*** [0.173]*** Treated X Member of executive committee 92.7 0.317 0.286 [42.4]** [0.336] [0.340] Member of executive committee -8.3 -0.102 -0.008 [16.5] [0.197] [0.202] Treated X Group chair or vice-chair 67.3 -0.313 -0.158 [50.5] [0.449] [0.478] Group chair or vice-chair 1.5 0.221 0.416 [21.0] [0.260] [0.296] Treated X Human capital index 43.2 47.8 -0.082 -0.051 -0.271 -0.252 [24.3]* [23.9]** [0.206] [0.205] [0.187] [0.187] Human capital index -16.1 -15.0 -0.018 -0.031 0.440 0.433 [20.5] [20.7] [0.203] [0.204] [0.190]** [0.191]** Treated X Working capital index -44.7 -42.7 -0.042 -0.026 -0.165 -0.145 [21.3]** [21.2]** [0.201] [0.200] [0.191] [0.189] Working capital index -26.1 -29.1 0.141 0.137 -0.050 -0.058 [35.8] [35.6] [0.274] [0.275] [0.251] [0.251] R-squared 0.3 0.3 0.235 0.234 0.193 0.194 Obs 1985 1985 1986 1986 1985 1985 Control Mean 49.77 49.77 1.904 1.904 3.628 3.628 Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employmnet and human and working capital. All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Appendix Table 4: Payback / Return on investment analysis Months to repay (N) Real annual interest rate (r) A. Cash earnings QTE - Median None Commercial prime Other comercial Other comercial Moneylender Per person cost of NUSAF grant 673,026 0% 5% 15% 25% 200% QTE, real monthly cash earnings 10,000 67.3 79.1 148.2 inf inf 0.1 61.2 70.8 116.5 inf inf Nonpecuniary value as % of cash earnin 0.5 44.9 49.8 66.2 132.4 inf 1 33.7 36.3 43.9 58.6 inf B. Cash earnings ATE Per person cost of NUSAF grant 673,026 ATE on real monthly cash earnings 19,515 34.5 37.3 45.4 61.5 inf 0.1 31.4 33.7 40.0 51.4 inf Nonpecuniary value as % of cash earnin 0.5 23.0 24.2 27.3 31.6 inf 1 17.2 17.9 19.5 21.6 inf C. All, but including estimated program costs of 30% Per person cost of NUSAF grant 874,933 ATE on real monthly cash earnings 19,515 44.8 49.7 66.2 131.9 inf 0.1 40.8 44.8 57.3 1.1 inf Nonpecuniary value as % of cash earnin 0.5 29.9 32.0 37.7 1.5 inf 1 22.4 23.6 26.5 2.0 inf Notes: Panel A considers the median transfer and QTE for all beneficiaries for five different real interest rates: 0, 5, 15, 25 and 200%. Panel B does the same for mean profits. Finally, Panel C considers the case where program implementation costs 30% of the transfer itself. A zero interest rate may be relevant from the perspective of a social planner who does not discount future welfare over present welfare. The 5% rate corresponds to the real prime lending rate, and could also be considered a social or state discount rate. Higher interest rates are closer to those available on the commercial market, up to the microfinance rate of 200%. Payback equation: N = -log[1 - (r /12 × A /P )] / log(1 + r /12), where N is the number of months, r is the real interest rate, A is the loan amount and P is the repayment. Appendix Table 5: Sensitivity analysis (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Cash earnings in last 4 weeks Cash earnings in last 4 weeks Cash earnings in last 4 weeks (no (with full list of individual ln(Cash earnings) IHS(Cash earnings) (without individual covariates) censoring) covariates) Treated 20 260 26 608 20 003 27 034 32 462 52 351 0 874 0 825 0 675 0 664 [5 831]*** [7 960]*** [5 320]*** [7 434]*** [11 296]*** [21 260]** [0 153]*** [0 180]*** [0 119]*** [0 143]*** Treated × Female -19 118 -21 164 -59 865 0 147 0 033 [11 685] [11 390]* [34 758]* [0 330] [0 256] Female -23 135 -15 217 -13 834 -5 115 -22 284 2 379 -0 442 -0 502 -0 383 -0 397 [5 093]*** [6 671]** [5 587]** [7 173] [10 615]** [10 549] [0 158]*** [0 213]** [0 124]*** [0 165]** Observations 2011 2011 1986 1986 1986 1986 1986 1986 1986 1986 R-squared 0 051 0 053 0 122 0 124 0 069 0 072 0 099 0 099 0 110 0 110 Control means All 43 45 43 45 49 04 8 419 2 704 Males 50 01 50 01 57 53 8 658 2 907 Females 32 27 32 27 34 56 8 013 2 359 Female Treatment Effect 7 489 5 870 -7 514 0 971 0 697 p-value 0 356 0 450 0 651 0 001 0 001 ATE as % of control mean All 47% 46% 66% Males 53% 54% 91% Females 23% 18% -22% Robust standard errors in brackets, clustered by group and stratified by district. Omitted regressors include an age quartic, district indicators, and baseline measures of employment and human and working capital. All UGX denominated variables censored at the 99th percentile. All inverse hyperbolic sine (IHS) variables are calculated as ln(x + ((x^2) + 1)^.5). *** p<0.01, ** p<0.05, * p<0.1 Social Protection Discussion Paper Series Titles No. Title 1201 Micro-Determinants of Informal Employment in The Middle East and North Africa Region by Diego F. Angel-Urdinola and Kimie Tanabe, January 2012 (online only) 1120 Employment Generation in Rural Africa: Mid-term Results from an Experimental Evaluation of the Youth Opportunities Program in Northern Uganda by Christopher Blattman, Nathan Fiala and Sebastian Martinez, December 2011 (online only) 1119 Measuring Governance and Service Delivery in Safety Net Programs by Gloria M. Rubio, September 2011 (online only) 1118 Assessing Safety Net Readiness in Response to Food Price Volatility by Margaret Grosh, Colin Andrews, Rodrigo Quintana, Claudia Rodriguez-Alas, September 2011 1117 Social Safety Nets in Fragile States: A Community-Based School Feeding Program in Togo, August 2011 (also available in French) 1116 Strengthening Governance of Social Safety Nets in East Asia by Sara Giannozzi and Asmeen Khan, August 2011 (online only) 1115 International Portability of Health-Cost Coverage: Concepts and Experience by Martin Werding and Stuart McLennan, July 2011 (online only) 1114 Liberia’s Cash For Work Temporary Employment Project: Responding to Crisis in Low Income, Fragile Countries by Colin Andrews, Prospère Backiny-Yetna, Emily Garin, Emily Weedon, Quentin Wodon and Giuseppe Zampaglione, July 2011 1113 Employability and Productivity among Older Workers: A Policy Framework and Evidence from Latin America by Edmundo Murrugarra, July 2011 (online only) 1112 Cash Transfers, Children and the Crisis: Protecting Current and Future Investments by Ariel Fiszbein, Dena Ringold, Santhosh Srinivasan, June 2011 (online only) 1111 Severance Pay Programs around the World: History, Rationale, Status, and Reforms by Robert Holzmann, Yann Pouget, Milan Vodopivec and Michael Weber, May 2011 (online only) 1110 Portability of Pension, Health, and other Social Benefits: Facts, Concepts, Issues by Robert Holzmann and Johannes Koettl, May 2011 (online only) 1109 Disability and Poverty in Developing Countries: A Snapshot from the World Health Survey by Sophie Mitra, Aleksandra Posarac and Brandon Vick, April 2011 1108 Advancing Adult Learning in Eastern Europe and Central Asia by Christian Bodewig and Sarojini Hirshleifer, April 2011 (online only) 1107 Results Readiness in Social Protection & Labor Operations by Laura Rawlings, Maddalena Honorati, Gloria Rubio and Julie Van Domelen, February 2011 1106 Results Readiness in Social Protection & Labor Operations: Technical Guidance Notes for Social Service Delivery Projects by Julie Van Domelen, February 2011 1105 Results Readiness in Social Protection & Labor Operations: Technical Guidance Notes for Social Safety Nets Task Teams by Gloria Rubio, February 2011 1104 Results Readiness in Social Protection & Labor Operations: Technical Guidance Notes for Social Funds Task Teams by Julie Van Domelen, February 2011 1103 Results Readiness in Social Protection & Labor Operations: Technical Guidance Notes for Labor Markets Task Teams by Maddalena Honorati, February 2011 1102 Natural Disasters: What is the Role for Social Safety Nets? by Larissa Pelham, Edward Clay and Tim Braunholz, February 2011 1101 North-South Knowledge Sharing on Incentive-based Conditional Cash Transfer Programs by Lawrence Aber and Laura B. Rawlings, January 2011 1008 Social Policy, Perceptions and the Press: An Analysis of the Media’s Treatment of Conditional Cash Transfers in Brazil by Kathy Lindert and Vanina Vincensini, December 2010 (online only) 1007 Bringing Financial Literacy and Education to Low and Middle Income Countries: The Need to Review, Adjust, and Extend Current Wisdom by Robert Holzmann, July 2010 (online only) 1006 Key Characteristics of Employment Regulation in the Middle East and North Africa by Diego F. Angel-Urdinola and Arvo Kuddo with support from Kimie Tanabe and May Wazzan, July 2010 (online only) 1005 Non-Public Provision of Active Labor Market Programs in Arab-Mediterranean Countries: An Inventory of Youth Programs by Diego F. Angel-Urdinola, Amina Semlali and Stefanie Brodmann, July 2010 (online only) 1004 The Investment in Job Training: Why Are SMEs Lagging So Much Behind? by Rita K. Almeida and Reyes Aterido, May 2010 (online only) 1003 Disability and International Cooperation and Development: A Review of Policies and Practices by Janet Lord, Aleksandra Posarac, Marco Nicoli, Karen Peffley, Charlotte McClain-Nhlapo and Mary Keogh, May 2010 1002 Toolkit on Tackling Error, Fraud and Corruption in Social Protection Programs by Christian van Stolk and Emil D. Tesliuc, March 2010 (online only) 1001 Labor Market Policy Research for Developing Countries: Recent Examples from the Literature - What do We Know and What should We Know? by Maria Laura Sanchez Puerta, January 2010 (online only) 0931 The Korean Case Study: Past Experience and New Trends in Training Policies by Young-Sun Ra and Kyung Woo Shim, December 2009 (online only) 0930 Migration Pressures and Immigration Policies: New Evidence on the Selection of Migrants by Johanna Avato, December 2009 (online only) 0929 Ex-Ante Methods to Assess the Impact of Social Insurance Policies on Labor Supply with an Application to Brazil by David A. Robalino, Eduardo Zylberstajn, Helio Zylberstajn and Luis Eduardo Afonso, December 2009 (online only) 0928 Rethinking Survivor Benefits by Estelle James, December 2009 (online only) 0927 How Much Do Latin American Pension Programs Promise to Pay Back? by Alvaro Forteza and Guzmán Ourens, December 2009 (online only) 0926 Work Histories and Pension Entitlements in Argentina, Chile and Uruguay by Alvaro Forteza, Ignacio Apella, Eduardo Fajnzylber, Carlos Grushka, Ianina Rossi and Graciela Sanroman, December 2009 (online only) 0925 Indexing Pensions by John Piggott and Renuka Sane, December 2009 (online only) 0924 Towards Comprehensive Training by Jean Fares and Olga Susana Puerto, November 2009 0923 Pre-Employment Skills Development Strategies in the OECD by Yoo Jeung Joy Nam, November 2009 0922 A Review of National Training Funds by Richard Johanson, November 2009 0921 Pre-Employment Vocational Education and Training in Korea by ChangKyun Chae and Jaeho Chung, November 2009 0920 Labor Laws in Estern European and Central Asian Countries: Minimum Norms and Practices by Arvo Kuddo, November 2009 (online only) 0919 Openness and Technological Innovation in East Asia: Have They Increased the Demand for Skills? by Rita K. Almeida, October 2009 (online only) 0918 Employment Services and Active Labor Market Programs in Eastern European and Central Asian Countries by Arvo Kuddo, October 2009 (online only) 0917 Productivity Increases in SMEs: With Special Emphasis on In-Service Training of Workers in Korea by Kye Woo Lee, October 2009 (online only) 0916 Firing Cost and Firm Size: A Study of Sri Lanka's Severance Pay System by Babatunde Abidoye, Peter F. Orazem and Milan Vodopivec, September 2009 (online only) 0915 Personal Opinions about the Social Security System and Informal Employment: Evidence from Bulgaria by Valeria Perotti and Maria Laura Sánchez Puerta, September 2009 0914 Building a Targeting System for Bangladesh based on Proxy Means Testing by Iffath A. Sharif, August 2009 (online only) 0913 Savings for Unemployment in Good or Bad Times: Options for Developing Countries by David Robalino, Milan Vodopivec and András Bodor, August 2009 (online only) 0912 Social Protection for Migrants from the Pacific Islands in Australia and New Zealand by Geoff Woolford, May 2009 (online only) 0911 Human Trafficking, Modern Day Slavery, and Economic Exploitation by Johannes Koettl, May 2009 0910 Unemployment Insurance Savings Accounts in Latin America: Overview and Assessment by Ana M. Ferrer and W. Craig Riddell, June 2009 (online only) 0909 Definitions, Good Practices, and Global Estimates on the Status of Social Protection for International Migrants by Johanna Avato, Johannes Koettl, and Rachel Sabates-Wheeler, May 2009 (online only) 0908 Regional Overview of Social Protection for Non-Citizens in the Southern African Development Community (SADC) by Marius Olivier, May 2009 (online only) 0907 Introducing Unemployment Insurance to Developing Countries by Milan Vodopivec, May 2009 (online only) 0906 Social Protection for Refugees and Asylum Seekers in the Southern Africa Development Community (SADC) by Mpho Makhema, April 2009 (online only) 0905 How to Make Public Works Work: A Review of the Experiences by Carlo del Ninno, Kalanidhi Subbarao and Annamaria Milazzo, May 2009 (online only) 0904 Slavery and Human Trafficking: International Law and the Role of the World Bank by María Fernanda Perez Solla, April 2009 (online only) 0903 Pension Systems for the Informal Sector in Asia edited by Landis MacKellar, March 2009 (online only) 0902 Structural Educational Reform: Evidence from a Teacher’s Displacement Program in Armenia by Arvo Kuddo, January 2009 (online only) 0901 Non-performance of the Severance Pay Program in Slovenia by Milan Vodopivec, Lilijana Madzar, Primož Dolenc, January 2009 (online only) To view Social Protection Discussion papers published prior to 2009, please visit www.worldbank.org/sp. Summary Findings Can cash transfers promote employment and reduce poverty in rural Africa? Will lower youth unemployment and poverty reduce the risk of social instability? We experimentally evaluate one of Uganda’s largest development programs, which provided thousands of young people nearly unconditional, unsupervised cash transfers to pay for vocational training, tools, and business start- up costs. Mid-term results after two years suggest four main findings. First, despite a lack of central monitoring and accountability, most youth invest the transfer in vocational skills and tools. Second, the economic impacts of the transfer are large: hours of non-household employment double and cash earnings increase by nearly 50% relative to the control group. We estimate the transfer yields a real annual return on capital of 35% on average. Third, the evidence suggests that poor access to credit is a major reason youth cannot start these vocations in the absence of aid. Much of the heterogeneity in impacts is unexplained, however, and is unrelated to conventional economic measures of ability, suggesting we have much to learn about the determinants of entrepreneurship. Finally, these economic gains result in modest improvements in social stability. Measures of social cohesion and community support improve mildly, by roughly 5 to 10%, especially among males, most likely because the youth becomes a net giver rather than a net taker in his kin and community network. Most strikingly, we see a 50% fall in interpersonal aggression and disputes among males, but a 50% increase among females. Neither change seems related to economic performance nor does social cohesion—a puzzle to be explored in the next phase of the study. These results suggest that increasing access to credit and capital could stimulate employment growth in rural Africa. In particular, unconditional and unsupervised cash transfers may be a more effective and cost- efficient form of large-scale aid than commonly believed. A second stage of data collection in 2012 will collect longitudinal economic impacts, additional data on political violence and behavior, and explore alternative theoretical mechanisms. HUMAN DEVELOPMENT NETWORK About this series... Social Protection Discussion Papers are published to communicate the results of The World Bank’s work to the development community with the least possible delay. The typescript manuscript of this paper therefore has not been prepared in accordance with the procedures appropriate to formally edited texts. The findings, interpretations, and conclusions expressed herein are those of the author(s), and do not necessarily reflect the views of the International Bank for Reconstruction and Development / The World Bank and its affiliated organizations, or those of the Executive Directors of The World Bank or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. For more information, please contact the Social Protection Advisory Service, The World Bank, 1818 H Street, N.W., Room G7-703, Washington, D.C. 20433 USA. Telephone: (202) 458-5267, Fax: (202) 614-0471, E-mail: socialprotection@worldbank.org or visit the Social Protection website at www.worldbank.org/sp.