75566 Empowering Adolescent Girls: Evidence from a Randomized Control Trial in Uganda¤ Oriana Bandiera, Niklas Buehren, Robin Burgess, Markus Goldstein, Selim Gulesci, Imran Rasul and Munshi Sulaimany December 2012 Abstract Nearly 60% of Uganda’s population is aged below 20. This generation faces health and economic challenges associated with HIV, early pregnancy and unemployment. Whether these challenges are due to a lack of information and/or vocational skills is however uncertain. We evaluate the impact of a programme that simultaneously provides: (i) vocational training to run small-scale enterprises; (ii) information on health and risky behaviors. The evaluation uses a randomized control trial, tracking 4 800 girls over two years. We …nd the programme positively impacts behaviors on both economic and health margins. On economic margins, the intervention raises the likelihood that girls engage in income generating activities by 32%, mainly driven by increased participation in self-employment. On health related margins, self-reported routine condom usage increases by 50% among the sexually active, and the probability of having a child decreases by 26%. Strikingly, the share of girls reporting sex against their will drops from 21% to almost zero. The …ndings suggest combined interventions might be more e¤ective among adolescent girls than single-pronged interventions aiming to improve labor market outcomes solely through vocational training, or to change risky behaviors solely through education programmes. Keywords: adolescents, ELA programme, risky behaviors, vocational training. JEL Classi…cation: I25, J13, J24, O12. ¤ We thank all those at BRAC Uganda involved in the research programme underlying this intervention. We are especially grateful to Nuarpear Lekfuangfu, Agnes Katushabe, Irene Lumala, and Georgina Shiundu for their re- search assistance. We have bene…ted greatly from discussions with Lucia Corno, Khondoker Ariful Islam, Brendon McConnell, Marcos Vera-Hernández, Abebual Zerihun, and from seminar participants at the JPAL-EBRD Confer- ence and the Barcelona Development Economics Workshop. We are grateful to the Bank Netherlands, Mastercard, Nike, the Gender Action Plan of the World Bank, and the Improving Institutions for Pro-Poor Growth programme of DfID for …nancial support. The views presented in this paper are the authors’ and do not represent those of the World Bank or its member countries, or DfID. All errors remain our own. y Bandiera: London School of Economics and Political Science, o.bandiera@lse.ac.uk; Buehren: University Col- lege London, n.buehren@ucl.ac.uk; Burgess: London School of Economics and Political Science, r.burgess@lse.ac.uk; Goldstein: World Bank, mgoldstein@worldbank.org; Gulesci: Bocconi, selim.gulesci@unibocconi.it; Rasul: Univer- sity College London, i.rasul@ucl.ac.uk; Sulaiman: BRAC International, munshi.slmn@gmail.com. 1 1 Introduction The number of young people in the developing world is increasing: one billion people on the planet are aged between 15 and 24 and reside in a developing country, an increase of 17% since 1995 [World Bank 2009].1 Nowhere is this phenomenon more pronounced than in Sub-Saharan Africa, where 60% of the population is now aged below 25 [World Bank 2009]. Youths face severe economic challenges, as they account for most of the region’s poor and unemployed: in sub-Saharan Africa, 60% of the total unemployed are aged 15-24, and on average 72% of the youth population live on less than $2 per day. For adolescent girls, such economic concerns are compounded by health related challenges such as early wedlock and pregnancy, exposure to STDs and HIV infection.2 These economic and health issues are obviously interlinked: teen pregnancy and early marriage are likely to have a decisive impact on the ability of young girls to accumulate human capital, and limit their future labor force participation [Field and Ambrus 2008, Bruce and Hallman 2008].3 At the same time, a lack of future labor market opportunities can reduce the incentives for young girls to invest in their human capital [Jensen 2012], leading to early marriage and childbearing, and potentially increasing their dependency on older men [Dupas 2011].4 In this paper we evaluate an intervention that attempts to simultaneously tackle the economic and health challenges faced by adolescent girls in Uganda, the Empowerment and Livelihood for Adolescents (ELA) programme, operated by the NGO BRAC. Uganda has the second lowest median age of all countries and the highest child dependency ratio [UN 2010 and Figure 1A]. Uganda also has one of the highest rates of young women out of the labor force: in the late 1990s, 1 The continued rise in the numbers of young people in the global population has led policy makers to consider responses to what has now become termed the ‘youth bulge’ [World Bank 2007]. Indeed, dealing with such demographic trends are at the top of many developing country policy agendas, and was the focus of the World Development Report in 2007. The central policy challenge is to provide increasing numbers of young people the skills and job opportunities to enable them to lead ful…lled and economically self-reliant lives in adulthood. There are a parallel set of concerns, emphasized more in the sociology and political science literatures, is that ever rising numbers and proportions of young people in developing countries will be a key factor driving alienation, social unrest and demands for political reforms, as has been observed throughout North Africa and the Middle East recently [Fuller 1995, Goldstone 2002]. 2 The Demographic and Health Survey data indicates that around 38% of the 52 million women aged 20 to 24 in developing countries were married before the age of 18 [Mensch et al. 2005] and these girls are often subject to unprotected sex. Girls in the 15 to 24 age band are almost eight times more likely than men to be HIV positive in Sub-Saharan Africa [Bruce and Hallman 2008, Dupas 2011, UN 2010]. Unprotected heterosexual intercourse together with the onward transmission of HIV to newborn and breast-fed babies is responsible for the vast majority of new HIV infections in the region. 3 Baird et al. [2011] provide evidence that marriage and schooling are mutually exclusive in Malawi, and Ozier [2011] provides similar evidence from Kenya using a regression discontinuity design to identify the causal impacts of schooling on fertility. In Bangladesh, Field and Ambrus [2008] show using age at menarche as an instrument for age at …rst marriage, that each additional year that marriage is delayed is associated with .3 additional years of schooling and 6.5% higher probability of literacy. 4 UN [2010] estimates that 10% of new HIV infections in 2010 in Uganda are linked to sex work or transactional sex. Luke [2003] provides evidence on age di¤erences across genders in sexual relations involving adolescent girls in Sub-Saharan Africa. She documents that larger age di¤erences lower the probability of condom use. Relatedly, Robinson and Yeh [2011] provide evidence from daily diaries collected from sex workers in Busia, Kenya, to document how they change their behavior in response to health shocks among family members. They …nd women increase their supply of risky, better compensated sex in response to short-term health shocks at home. This is another channel through which economic pressures can exacerbate public and private health problems. 2 the rate was 86% in Uganda, compared to an average of 58% in 14 sub-Saharan countries. For those in the labor force, Figure 1B uses data from the Uganda National Household Survey of 2005/2006 to construct information on unemployment rates by gender and age. The …gure shows that at all ages females have higher unemployment rates than men, and this is especially pronounced in the youngest age cohorts.5 Figure 1C illustrates that relative to females in the same age cohort in richer economies, fertility rates (the number of births per 1,000 women) are three to four times higher in Uganda, and that this gap in fertility rates is most pronounced among the youngest female aged cohort of 15 to 19 year olds. In this context, the ELA programme aims to bolster the cognitive and non-cognitive skills of girls by providing: (i) vocational skills training to enable adolescent girls to start small-scale income generating activities; (ii) life skills to build knowledge and reduce risky behaviors. In contrast to most life skill programmes, the intervention is delivered from designated ‘adolescent development clubs’ rather than in schools, and can thus reach school drop-outs as well as girls currently enrolled in school. We collaborate with BRAC to randomly assign the clubs to 100 out of 150 communities chosen to be part of the evaluation sample, and located in both urban and rural areas of the country. In treatment and control communities we survey a random sample of around 40 eligible girls at baseline and two years later. Eligibility depends solely on gender and age. Our evaluation is thus based on a two-year panel of over 4 800 adolescent girls and we identify the causal e¤ect of the program as the di¤erence in di¤erence in outcomes for eligible girls in treatment and control communities before and after program implementation. Participation is voluntary and unrelated to other BRAC activities. The average take-up rate is 21%, indicating there is latent demand for the services provided. We estimate both ITT and ATT e¤ects using the randomized program placement as an instrument for participation. Our …ndings indicate that the ELA programme increases labor force participation and income, improves health- related knowledge, and reduces risky behaviors. On economic margins, the ITT on the likelihood of an adolescent girl being engaged in some income generating activity increased by 4 percentage points (pp), corresponding to a 32% increase over baseline levels. This result is driven by additional engagement in self-employment activities. Moreover, despite the fact that girls currently enrolled in school are also eligible for the programme, there is no drop in school enrollment rates. Hence there appears to be little trade-o¤ in promoting the empowerment of girls through health and vocational skills based interventions, and their current incentives to invest in human capital. On health related margins, the programme improved HIV and pregnancy related knowledge, and the ITT estimate of self-reported condom usage shows an increases by 13pp among those who are sexually active. While there is no e¤ect on reported infections with a sexually transmitted disease, when infected adolescents are more likely to seek treatment at a health center. Two years after the ELA programme’s initiation, adolescents living in treated villages are 3pp less likely 5 The Uganda National Household Survey (UNHS) is a nationally representative sample of households. A two stage sampling design is used resulting in 783 Enumeration Areas (EAs) being selected. These EAs represented both the general household population and displaced population (UBOS 2005). The survey covers 7246 households. 3 than those in control communities to report having at least one child which corresponds to a 286% decrease in fertility rates among the targeted population. Moreover, the reduction in girls reporting having had sex unwillingly is dramatic: a fall of around 16pp starting from a baseline level of 21%. This is the clearest marker that the combination of life skills and vocational training can successfully improve the quality of relationships that adolescent girls’ …nd themselves in. To provide a richer picture of how the ELA programme transforms the lives of adolescent girls, we consider outcomes not directly related to risky behaviors or income generating activities, but are proxies for welfare, self-con…dence and empowerment that might signal longer term impacts of the programme. On these dimensions we …nd girls to report signi…cant increases in an overall index of gender empowerment, improved satisfaction with earnings/income, and signi…cant reductions in girls reporting worrying about …nding a good job in adulthood. This might suggest the programme is able to reduce stress levels among girls as they transit into the labor market, precisely as intended and part of the underlying motivation for the intervention. Finally, we consider outcomes related to girls opinions and expectations on marriage and childbearing. As emphasized above, the patterns of early marriage and child bearing observed in Uganda and much of Sub-Saharan Africa often lie at the heart of adolescent girls’ restricted opportunities to acquire human capital, engage in productive labor market activities, and increases their exposure to health risks. Here we …nd some hope that the programme is able to set in chain processes and beliefs that delay age at …rst marriage and child-bearing. As such, these girls’ lives might then improve along dimensions that have been shown to be associated with such delays such as improved marriage quality, increased decision-making within households and reduced exposure to domestic violence [Goldin and Katz 2002, Jensen and Thornton 2003, Field and Ambrus 2008]. Our paper contributes to the literature evaluating the impact of life-skills and vocational train- ing interventions, that are two major policy instruments being utilized to address health and eco- nomic challenges for youth around the developing world. In contrast to the intervention we study, and despite the interlinkages between economic and health issues, many policy interventions have focused exclusively on classroom-based education courses designed to reduce risky behaviors, or exclusively on vocational training designed to improve labor market outcomes among youth. As a body of randomized-control trials suggest, these single-pronged programmes have met with, at best, rather mixed success [Gallant and Maticka-Tyndale 2004, Cornish and Campbell 2009, McCoy et al. 2010, Card et al. 2011, Groh et al. 2012].6  7 6 On programmes related to the provision of information on reproductive practices and HIV prevention, Gallant and Maticka-Tyndale [2004] provide a review of eleven, mostly school-based HIV prevention programmes. They conclude that most programmes successfully improved knowledge and attitudes, but that the measurable impacts on behavior were weaker. Cornish and Campbell [2009] also document the generally weak impacts found of peer education HIV campaigns. McCoy et al. [2010], review eleven RCT studies of behavioral change programmes and …nd that only two analyses show signi…cant protective e¤ect on HIV incidence among women, and only three of ten studies that measured behavioral outcomes reduced any measure of HIV-related risk behavior. One a more positive note, Dupas [2011] uses an RCT design to compare the e¤ectiveness of the Kenyan national HIV curriculum to an intervention providing information on the relative risk of HIV infection by the partner’s age. She …nds a signi…cant and large reduction of teen pregnancies only for the informational campaign on relative riskiness. 7 Evaluations of training programs in industrialized countries generally produce mixed results [Blundell et al. 2004]. In developing country settings the results from RCTs of vocational training programmes are also somewhat 4 Our evidence suggests the ELA programme is at least as successful as interventions that have exclusively targeted life skills or vocational skills in similar contexts. This suggests the individual programme elements are complementary to each other: girls are more likely to take-on board health related education in terms of knowledge and behaviors when they are simultaneously o¤ered new income generating skills. At the same time, the expected returns to providing vocational skills training to this target population might be larger when they are simultaneously provided information to help reduce their exposure to activities that involve risky behaviors. Our …ndings complement a small body of research that use large-scale randomized control trials to provide evidence on the interlinkages between economic and health challenges. For example, Du‡o et al. [2011] also investigate a two-pronged intervention: a school-based HIV prevention programme in Kenya coupled with subsidies to attend school. These components are o¤ered sepa- rately as well as jointly. The HIV prevention programme when o¤ered alone is ine¤ective against rates of STI or teen pregnancy although it reduces the number of unwed teen pregnancies. In contrast, the schooling subsidy reduces drop-outs, teen marriage and teen pregnancy. Interest- ingly, the negative e¤ect on teen pregnancy is dampened when the subsidy is bundled with the HIV prevention training, suggesting the two programme components interact. The idea that improvements in income generating activities among targeted girls might feed- back to further reduce risky behaviors is in line with the nascent literature using …eld experiments to estimate income elasticities of risky behaviors. For example, Baird et al. [2010] …nd that a cash transfer of $10 per month conditional on school attendance for adolescent girls in Malawi led to signi…cant declines in early marriage, teenage pregnancy and self-reported sexual activity after one year. This again suggests that twin-pronged programmes are especially e¤ective in reducing risky behaviors.8 As in our study, this might be because the e¢cacy of providing adolescent girls infor- mation on how to reduce their exposure to health risks, is larger when reinforced by programme components that simultaneously empower girls to lead economically independent lives. The paper is organized as follows. Section 2 describes the ELA programme components and their implementation. Section 3 describes the research design, data and estimation strategy. Section 4 presents estimates of the programme’s two-year impacts on income generating activities, engagement in risky behaviors, and the empowerment and aspirations of adolescent girls’. Section 5 provides a cost-bene…t analysis of the programme, and discusses the broader implications of our …ndings for polices designed to address health and economic challenges facing the ever growing numbers of young people in the developing world today. The Appendix presents robustness checks on the baseline results dealing with multiple inference, impact heterogeneity and sample attrition. mixed: Card et al. [2011] …nd little impact of such an intervention in the Dominican Republic; Attanasio et al. [2011] report that the Colombian Jovenes en Accion programme raised earnings as well as the employment rate, especially among women. A separate literature examines the impacts of business training, some elements of which are replicated in the programme we evaluate, such as …nancial literacy training: in contrast to our intervention these interventions are often targeted towards those that have already self-selected into small scale entrepreneurial activities. Despite this, the literature again provides generally disappointing results, as discussed later. 8 In another study, Baird et al. [2011] report that impacts on risky behaviors are larger from unconditional cash transfers rather than cash transfers conditioned on school attendance, an e¤ect purely driven by the impact of unconditional cash transfers on girls that are already out of school. 5 2 The ELA Programme The Empowerment and Livelihood for Adolescents (ELA) programme is designed to improve the cognitive and non-cognitive skills of adolescent girls. The programme is implemented by the NGO, BRAC Uganda.9 In contrast to some other school-based information campaigns on adolescent health, the ELA programme operates through ‘adolescent development clubs’, a …xed meeting place within each community. These clubs are often housed in a single dedicated room which is either donated by the community or rented by BRAC. Typically, the club is open …ve afternoons per week and timed so that girls enrolled full-time in school can attend. Club activities are led by a female mentor. The mentor is selected by programme sta¤ from within the community, tends to be slightly older than the target population of adolescent girls, and is prepared for her supervisory role during a week-long initiation programme, as well as bi-monthly refresher courses. Mentors obtain a small lump-sum payment for their work.10 Club participation is voluntary and unrelated to engagement with other BRAC activities. Eligibility is based on gender and age: only girls are permitted to participate, and the programme is intended for adolescent girls aged between 14 and 20. Given the di¢culties of verifying ages in this setting and the demand for club activities arising from other girls, in practice some girls outside of the 14-20 age range also attend the clubs. The two forms of skills training provided in the ELA programme are life skills training, and vocational skills training, both of which take place within the clubs. In addition, the clubs also host popular recreational activities such as reading, staging dramas, singing, dancing and playing games. As such the clubs serve as a local space in which adolescent girls can meet, socialize, privately further discuss issues of concern and to continue to develop their non-cognitive skills.11 The vocational skills training comprises a series of courses on income generating activities. Although many of the skills are applicable for either wage or self-employment, more focus is placed on the adolescent girls establishing small-scale enterprises of their own. Courses relating to a broad range of income generating activities are provided including hair-dressing, tailoring, computing, agriculture, poultry rearing and small trades operation. Given the range of courses o¤ered, girls self-select into those in which they have an interest or comparative advantage.12 9 BRAC is a non-governmental organization rooted in Bangladesh. The programme design we evaluate in Uganda has been re…ned over a number of years based in part of BRAC’s experiences with similar programmes in Bangladesh. As of December 2006, over 17,000 adolescent development clubs had been established in Bangladesh with over 300,000 of the involved adolescents having formed their own micro…nance groups. BRAC Uganda was established in 2006: starting from a focus on micro…nance, BRAC Uganda has developed and implemented a range of programs including health and adolescent development. The ELA programme in Uganda has now been scaled-up by BRAC to operate in Tanzania and South Sudan. Other donors have also scaled-up the ELA model to countries in Sub-Saharan Africa and South Asia. 10 We also conducted a survey of mentors. Relative to adolescent girl club participants, mentors are signi…cantly older, have signi…cantly more years of completed schooling, but are engaged in income generating activities to the same extent. At baseline they also have better …nancial and self-assessed entrepreneurship skills. 11 There is a nominal fee the girls are encouraged to pay to attend although in practice, this is often waived. Among focus groups pre-intervention, 14% of girls reported not being able to join because of the admission fee. 12 For some of the income generating activities, if the girl decides to take up the activity then they receive additional input supplies, such as seeds in agriculture, or chicks for poultry farming. 6 The vocational training modules are taught by entrepreneurs engaged in the respective activ- ities or by hired professionals as well as BRAC’s own agriculture and livestock programme sta¤. These courses are supplemented by …nancial literacy courses covering budgeting, …nancial services, negotiation and accounting skills. The process of matching girls to income generating activities is partly demand-driven, but account is also taken of the girl’s educational level, the local business environment and demand for such services. The key topics covered in the life skills training sessions include sexual and reproductive health, menstruation and menstrual disorders, pregnancy, sexually transmitted infections, HIV/AIDS awareness, family planning, rape; other sessions covered enabling topics such as management skills, negotiation and con‡ict resolution, and leadership among adolescents; a …nal class of life skills training focused on providing girls with legal knowledge on women’s issues such as bride price, child marriage and violence against women. These life skills training sessions are conducted either by the trained mentors and/or BRAC’s own professional sta¤.13 Two further points are of note. First, given the age range of targeted girls, some of them are enrolled in school, others have graduated from secondary school, while others have dropped out. Although the clubs operate outside of school times, emphasis is still placed on ensuring girls do not drop out of school, or reduce the hours they devote to schooling, in order to take-up membership of the club and engage in its activities. We later examine the evidence of the programme having an impact on the current educational investments of participants. Second, this evaluation examines the impacts of the programme on outcomes related to income generating activities and risky behaviors two years after the initiation of the ELA programme in treated communities. Subsequently, the programme has expanded to include a micro…nance com- ponent, that o¤ers participating adolescents the opportunity to take on credit in order to capitalize on their entrepreneurial potential and newly acquired skills during the vocational training.14 At the time of this evaluation, adolescent club participants were unaware of the potential future o¤er of micro…nance. BRAC sta¤ were unaware of which clubs would be assigned to receive micro…- nance in addition. Thus the evaluation we conduct is based solely on the provision of vocational skills and life skills training over the …rst two years of the programme. As BRAC is best known for its activities in micro…nance, we later con…rm that the anticipation of micro…nance does not drive any of our core …ndings. 13 The life skills provided overlap those studied in previous health-related education programmes [Dupas 2011, Du‡o et al. 2011], and also with those in the separate literature on basic business training (Field et al. 2010, Drexler et al. 2010, Karlan and Valdivia 2010]. We later review our …ndings relative to each of these literatures. 14 In this second round of randomization, 50 out of the 100 treated villages were randomly assigned to additionally receive micro…nance, and 50 were randomly assigned to no receive micro…nance. The 50 originally assigned control communities remained as controls throughout. 7 3 Design, Data and Estimation 3.1 Research Design We evaluate the ELA programme using a randomized control trial. The evaluation takes place during the initial phase of the programme, when it is being rolled-out across Uganda. This expansion uses the service delivery infrastructure BRAC has developed for its earlier programmes. In particular, as part of its earlier programmes, BRAC has established branch o¢ces through the country: ten branch o¢ces were chosen for the ELA evaluation. Five of these branches are located in the urban or semi-urban regions of Kampala and Mukono; the remaining …ve branches are located in the mostly rural region around Iganga and Jinja. In each branch, at least …fteen communities with the potential to host an ELA club were identi…ed. From this list, ten locations within each branch o¢ce were randomly assigned to the receive the treatment, i.e. to set up a club and deliver the ELA programme, with the remaining …ve locations randomly assigned as controls. In each treatment community, a single club was opened up. Hence, the research design delivers 100 treatment and 50 control communities, strati…ed by branch o¢ce.15 The practicalities of programme implementation led to possible non-compliance with the re- search design among those resident in control communities. An adolescent girl resident in a control community wishing to attend a club in a treated community is always able to do so. Non-compliance biases the estimated impact towards zero to the extent that girls from control locations join the ELA clubs. In some urban areas, this can happen because the distance to the nearest club might be quite similar in treatment and control communities. In rural locations, al- though BRAC places much emphasis on …nding suitable club locations in the center of treatment locations, inevitably some clubs are located in more peripheral village locations – due to a lack of available space in the village centre. Again in such cases, if information about the programme spreads across villages, girls from neighboring control villages are in principle able to attend a club located in a treated village. We document the extent of such non-compliance with the research design among control locations below. Of course in treated communities, as club participation is voluntary, not all eligible girls will comply with the design and decide to take-up the o¤er of receiving the ELA programme. We therefore later estimate both ITT and ATE impacts. 3.2 Data, Attrition and Descriptives The primary data used for the analysis is a survey administered to adolescent girls conducted at baseline (pre-randomization), and two years after the ELA programme is initiated. The ques- tionnaire administered to the adolescent respondents covers a range of topics including: (i) those directly related to the vocational skills component, such as …nancial literacy, analytical ability, labor market and income generating activities; (ii) those related directly to the life skills compo- 15 For exposition purposes, we will refer to communities as the unit of randomization. For the rural branches these correspond to villages. For the branches located in urban or semi-urban regions of Kampala and Mukono, the randomized units often correspond to smaller urban areas or slums. 8 nent, such as engagement in risky behaviors, HIV related knowledge, self-con…dence; (iii) other margins such as educational investments, time use, expenditures, as well as further measures of empowerment in social and economic activities. Prior to the baseline survey, in each of the 150 communities in the evaluation sample, a listing was drawn up of all resident adolescent girls eligible to attend the clubs. From this listing 40 girls were randomly selected to be surveyed. Baseline interviews were conducted from March to June 2008. The vast majority of ELA clubs were established between June and September 2008, and the follow-up survey was …elded from March to June 2010. Information from 5,966 randomly survey adolescents was obtained at baseline, corresponding to around 30% of all eligible girls: 3,964 girls reside in treatment communities, and 2,002 girls in controls, with an average of 39.9 (39.7) girls being surveyed in each rural (urban) community. De- spite the high degree of geographic mobility of girls in Uganda in this age range, 4 888 adolescents were tracked to follow-up, corresponding to a two-year tracking-rate of 82%. This is comparable to tracking rates from studies in similar contexts.16 Table A1 shows the correlates of two-year attrition. The dependent variable is a dummy equal to one if the adolescent girl attrits, and we control for her community’s treatment status, her individual characteristics and branch dummies. Column 1 shows that residing in a treatment community does not predict attrition. Column 2 shows this to be robust within branch, and Column 3 shows that the result holds conditioning on individual characteristics at baseline. More- over, none of these characteristics: age, current enrollment in school, being married or having children, themselves predict attrition. Column 4 examines how individual characteristics di¤er- entially relate to attrition between treated and control communities. We …nd some evidence that married adolescent girls in treated locations are less likely to be tracked at follow-up, and this result continues to hold – although with less precision – when attrition is predicted using a probit speci…cation as shown in Column 5. We therefore later present robustness checks that account for attrition. 3.2.1 Baseline Characteristics Table 1 shows the baseline characteristics of adolescent girls, by treatment status. Panel A shows that girls in our sample are on average 16 years old, and 71% of them are enrolled in school. Despite their youth, 6% of girls are married and 11% of all girls have at least one child. All else equal, such life circumstances are likely to reduce their labor market opportunities in the future. Panel B presents evidence on income generating activities, the target of the second component of ELA. The average self-assessed measure of entrepreneurial ability is 70100 and around 65% of girls report being self-employed, and it is this type of income generating activity the ELA programme especially focuses on developing among the adolescents.17 Girls are slightly more 16 Friedman et al. [2011] achieve an e¤ective tracking rate of nearly 80% after four years for a sample of Kenyan females of similar ages to those we study. Du‡o et al. [2011] report an attrition rate of 49% from during their follow-up six to seven years after the initial interview in their sample of female Kenyans (aged 14 at baseline). Second-round intensive tracking lifts the e¤ective tracking rate to 89%. 17 The entrepreneurial index consists of cumulative ranks (scaled from one to ten with ten being the highest) 9 likely to be engaged in wage employment at baseline in treated communities. The …nal two rows of Panel B show that girls supply less than an hour per day to self and wage-employment activities on average.18 Panel C presents evidence on risky behaviors, that the …rst component of the ELA programme aims to shift. At baseline, girls score around 38 on a 0-6 scale of HIV knowledge on average, yet there is considerable variation in this metric: at the tails of the knowledge distribution, 489% of girls correctly answer all the questions and 219% provide no correct answers. 74% of them correctly answer a basic question related to pregnancy knowledge.19 51% of adolescent girls report always using a condom if they are sexually active. Around 14% of the girls report symptoms consistent with an STD, while just over two thirds of them have gone to a health center for treatment.20 Finally, one dimension on which treatment and control groups di¤er is in the likelihood of having had sex unwillingly in the past year: this is 14% in the control group and over 20% among treated girls, although the normalized di¤erence is small. This is a key indicator for how empowered girls are in being able to control their lives. Although on most dimensions the samples are balanced between treatment and control, there are some di¤erences relating to wage employment. Reassuringly, in these cases the normalized di¤erences are all relatively small. As discussed in Imbens and Wooldridge [2009], the fact that none of these di¤erences exceeds .25 suggests a linear speci…cation is appropriate.21 of the following activities: “Run your own business�, “Identify business opportunities to start up new business�, “Obtain credit to start up new business or expand existing business�, “Save in order to invest in future business op- portunities�, “Make sure that your employees get the work done properly�, “Manage …nancial accounts�, “Bargain to obtain cheap prices when you are buying anything for business (inputs)�, “Bargain to obtain high prices when you are selling anything for business (outputs)�, “Protect your business assets from harm by others�, “Collecting the money someone owes you�. We then re-scale the sum of these scores to run from 0 to 100. 18 Throughout, monetary variables are de‡ated and expressed in terms of the price level in January 2008 using the monthly consumer price index published by the Uganda Bureau of Statistics. In January 2008, $1 was worth approximately UGX1,700. For those few observations where month of interview was missing, the interview month most frequently observed for their peers in the same community was assumed. 19 The HIV knowledge index is based on the number of statements correctly identi…ed as true or false. The statements are: (i) “A person who has HIV is di¤erent from a person who is ill with AIDS�; (ii) “During vaginal sex, it is easier for a woman to receive the HIV virus than for a man�; (iii) “Pulling out the penis before a man climaxes keeps a women from getting HIV during sex�; (iv) “A women cannot get HIV if she has sex during her period�; (v) “Taking a test for HIV one week after having sex will tell a person if she or he has HIV�; (vi) “A Pregnant woman with HIV can give the virus to her unborn baby�. The pregnancy knowledge indicator is based on the statement, “A women cannot become pregnant at …rst intercourse or with occasional sexual relations�. 20 To avoid potential biases arising from girls not wishing to report STDs, we asked girls to report if they have experienced symptoms of STDs. A list of typical symptoms were read out loud to respondents during the interview. However, this will miss asymptomatic STDs. In addition, the wording of the question was whether the girls have ever experienced the symptom, and so the time trend in the outcome cannot be negative. This does not a¤ect the ITT estimates that are identi…ed o¤ di¤erences between treatment and controls. 21 In terms of external validity, the rates of self-employment reported in our baseline sample match closely with those we derive for the nationally representative Uganda National Household Survey 2005/2006 based on over 7200 households. There we …nd that among those in the labor force, self-employment rates for 10-12 and 12-20 years olds are 7%. 10 3.2.2 Club Participation Table 2 reports information related to participation in the adolescent development clubs measured at follow-up, two years after baseline. The …rst row shows that the clubs are relatively well known among respondents: as expected, awareness of the existence of the ELA programme is more widespread in treatment communities, but the fact that almost 40% of adolescents resident in control villages have heard about the clubs emphasizes that information about these activities spreads outside their immediate vicinity. As noted earlier, in some urban areas the distance to the nearest club from some pairs of treatment and control communities is similar. As a result we note that 47% of those in control villages have ever participated in ELA club activities in the past, as shown in the second row of Table 2. However, the participation rate in treatment communities is more than four times higher (21%). Given that attendance is entirely voluntary, this suggests there is demand for the services provided by ELA and that, at least for some girls, the informational and skills constraints the programme aims to ease, are binding.22 The remaining rows report statistics conditional on club participation in treated communities. We focus on treatment communities as the number of participants in control communities is negligible, and report information on frequency of attendance and type of training received. As both attendance to the club and participation in training are entirely voluntary, Table 2 provides evidence on the latent demand for these services. The table shows that the majority of adolescents who have ever participated in ELA club activities continued their engagement until the time of the follow-up survey. Nearly half of all participants have attended club meetings, on average, one or two times a week, over the two years of the club’s operation. Hence, over a two-year period, the intervention amounts to a considerable time investment for participants, and it is plausible that such an intense treatment causes permanent shifts in knowledge, skills and behaviors. By the time of the follow-up, a sizeable proportion of club participants have taken part in the training on life skills (847%) and vocational skills (527%). The majority of girls (509%) report having received both forms of training; we can therefore infer that 33% take-up only life skills training, and 1% take-up only vocational skills training. As participation in either or both trainings is entirely voluntary, a revealed preference argument shows that the two programme components are complementary for the majority of participants.23 Table 3 shows how participants di¤er from non-participants in treatment communities. Partic- 22 It is perhaps useful to contrast this take-up rates for those in micro…nance projects. Karlan et al. [2010] provide evidence from two surveys and 13 interventions providing credit, savings or insurance services. They document take-up rates varying from 2 to 84% of eligibles. 23 We have explored the possibility that these take-up rates are driven by supply side constraints rather than demand side heterogeneity (and so under plausible assumptions might be used to separately identify the impact of each component of the ELA programme). This however is not the case: in nearly all treated communities we observe: (i) some eligible girls choosing to take-up a component and other girls not doing so; (ii) the vast majority of eligible girls report life and livelihood skills training as being available even if they don’t themselves take-up the course(s). In addition, we do not …nd school enrolment at baseline to be a signi…cant determinant of enrolment in the vocational training component, that might otherwise have been indicative of implicit supply side constraints operating to prevent all girls from receiving the same package of ELA treatments. 11 ipants are signi…cantly less likely to be married, to report using condoms if engaged in sexual ac- tivity, and having su¤ered from an STD. Moreover, on a host of other characteristics related to de- mographics, risky behaviors and income generating activities related to wage and self-employment, participants do not di¤er much on observables from non-participants. Hence, based on these ob- servables, participants do not appear to be negatively or positively selected on life skills or income generating abilities. The impact of the programme is therefore identi…ed from its e¤ects on girls whose characteristics are similar to all eligible girls’ in their communities, rather than from girls who needed to catch up or who were less disadvantaged to start with. 3.3 Estimation We …rst estimate the intent-to-treat (ITT) e¤ect of the ELA programme. To do so, all girls who reside in treated communities are counted as treated and all those who reside in control communities are assigned to control, regardless of whether they actually participate in the ELA programme. The small degree of non-compliance among girls resident in control communities (Table 2) is likely to bias estimated impacts towards zero because 4.7% of girls in the control group are actually treated. Given random assignment of communities to treatment and control status, estimating the ITT impact of the ELA programme with OLS is straightforward using a standard di¤erence-in-di¤erences (DID) approach. More precisely, ELA’s impact on an outcome ���� measured for adolescent � living in community � at survey wave � is estimated as follows, ���� =  + ����0  + �����  + ������  + (����� £ ������ )  + ���  (1) ����0 controls for the adolescent’s age at baseline, and we also include a series of indicator variables for branch areas as we stratify the sample of communities by branch before randomly assigning them to treatment or control status [Bruhn and McKenzie 2009]. The indicator variable ����� equals one if the observation is measured at follow-up and zero otherwise and captures natural time trends in outcomes; ������ equals one if community � is assigned to be treated and zero otherwise. This captures any potential pre-programme di¤erences between the groups. ��� is a disturbance term that we allow to be clustered by community � because there are likely to be common unobserved factors within communities that determine outcomes related to risky behaviors and income generating activities.  constitutes the coe¢cient of interest as it measures the ELA programme’s ITT impact. The OLS speci…cation in (1) is our base con…guration. However, some outcomes related to dummy variables and other outcomes of interest are left censored, such as the number of hours devoted to self-employment. In such cases we also report ITT estimates based on non-linear Probit and Tobit speci…cations analogous to (1). In such non-linear estimations, the reported coe¢cients correspond to marginal e¤ects. As the variable of interest, ����� £ ������ , is an interaction term, this is computed in accordance with Ai and Norton [2003].24 24 Ai and Norton [2003] report the correction for interaction terms in a Probit speci…cation. In the case of a 12 The speci…cation in (1) e¤ectively treats the data as a repeated cross-section rather than exploiting its panel dimension. The obvious advantage of doing so is to also be able to condition on an individual …xed e¤ect, � , that captures the in‡uence of time-invariant characteristics that drive changes in behaviors over time, such as a girl’s underlying ability, motivation, time and risk preferences. In this case the error term in (1) decomposes as ��� = � + ���� . As the treatment e¤ect cannot be identi…ed if individual …xed e¤ects are included, we estimate a …rst di¤erenced speci…cation as follows, ����1 ¡ ����0 =  + ����0  + ������  + ���1 ¡ ���0  (2) where ����0 is as previously de…ned. This speci…cation allows the correlation between the baseline characteristics ����0 and individual outcomes to vary over time:  captures the change over time in the corresponding  coe¢cient from (1).  then remains as the coe¢cient of main interest as it measures the intention-to-treat impact of the ELA programme. When estimating (2), we necessarily can only include those adolescents that are in the panel sample followed at baseline. In contrast, estimating (1) merely requires exclusion of those ado- lescents for whom no information on the outcome variable exists. Hence to ensure comparability of the  estimates across speci…cations we report results based on the same underlying sample of adolescents, namely those that are in the panel sample. As discussed above and in Table A1, there is little evidence of di¤erential attrition on observables across treatment and control communities. Given that participation into the club is far from universal, the ITT estimates likely underes- timate the impact of the programme on actual club participants. Our next speci…cation attempts to estimate ATE impacts of actual participation on outcomes. Building on the …rst di¤erence speci…cation above, we consider the following estimating equation, ����1 ¡ ����0 =  + ����0  + �������1  + ���1 ¡ ���0  (3) where ������� is a dummy variable which equals one if adolescent � in community � has ever partici- pated in ELA club activities at time �. Since participation is likely to be endogenously determined by unobserved factors that also drive changes in behavior, so that ��� (�������1  ���1 ¡ ���0 ) 6= 0, we follow the standard approach of instrumenting individual club participation with ������ , the Tobit speci…cation, taking into account that both interacted variables are discrete, the interaction e¤ect on the uncensored variable � can be computed by taking the double di¤erence, [ �� � ^ [� j��������� = 1 ���� = 1 �� � = 1 � �  0] ¡ � = (� ^ [�j��������� = 0 ���� = 1 �� � = 0 � �  0]) ¡(� ^ [� j��������� = 1 ���� = 0 �� � = 0 � �  0] ¡ � ^ [� j��������� = 0 ���� = 0 �� � = 0 � �  0]) where, ^ ��������� +  ^ [�j��������� ���� �� � � �  0] =  � ^ �� � + �  ^ ���� +  ^ 1 2 12 " # ^ ��������� +  ©0 (( ^ �� � + � ^ ���� +  ^ ) ^ ) 1 2 12 +:  ©(( ^ ���� +  ^ ��������� +  ^ ) ^ �� � + � ^) 1 2 12 The interaction e¤ect on the probability of being uncensored can be computed analogously. 13 community treatment dummy. As shown in Table 2, girls are more likely to participate in club activities if they live in treatment communities. The impacts are computed using a standard two-stage least squares procedure and, therefore, all coe¢cients are computed using a linear prob- ability model in the …rst stage as favored by Angrist [2001]. In the …rst stage, we regress a dummy variable for participation on the treatment dummy, branch …xed e¤ects and individual baseline characteristics ����0 . The coe¢cient on the treatment dummy is 157 and is signi…cant at the 1% level. The F-statistic from the …rst stage is 974. Under the constant treatment e¤ect assumption, and that treatment assignment has no spillover impacts on non-participants, the IV estimates then produce the population average treatment e¤ect on those that are induced to participate in the presence of a club in their community. 4 Main Results This section presents the impacts of the programme on the two main sets of outcomes targeted by the programme: income generating activities and risky behaviors. For each outcome we show ITT estimates from the standard di¤erence-in-di¤erence OLS speci…cation in equation (1), ITT estimates from the …rst di¤erence OLS speci…cation in equation (2), and ATE estimates from the instrumented …rst di¤erenced speci…cation in equation (3). As indicated above, for each outcome these speci…cations are always based on the same subsample of girls that are tracked to follow-up, the number of adolescents used to estimate each outcome is then shown in all the regression tables. For clarity, the results tables omit all coe¢cients estimated other than the estimated treatment e¤ect, b . 4.1 Income Generating Activities Table 4 shows the programme impacts outcomes related to income generating activities. Again, to benchmark these magnitudes, the …rst column shows the baseline value of each outcome in treated communities. Focusing on the OLS di¤erence-in-di¤erence estimates of the ITT impact of the programme in Column 1, we estimate the programme impact on girls’ self-reported entrepreneurial skills: an index running from 0 to 100 and based on answers to ten questions, as described earlier. This index rises signi…cantly for those that reside in treated communities. This result is in line with girls receiving bene…ts, for example, from the …nancial skill courses that are part of the vocational training component. Figure 2A presents a spider graph showing the ITT impacts (and their associated 95% con…dence interval) for each of the constituent components that go into making up the entrepreneurial skills score. We see that the programme signi…cantly increases self-reported entrepreneurial skills along nearly every dimension. Next, we assess whether the programme impacts real economic activities of adolescents. The second row of Table 4 shows that adolescent girls are 4pp more likely to be engaged in any income generating villages in treated communities. At baseline, 124% of girls were engaged in such activities. Hence the programme impact corresponds to a 32% increase. Sub-dividing income 14 generating activities into those arising from wage employment and self-employment, the next two rows show the increase is entirely driven by the adolescent girls engaging in entrepreneurial self- employment activities. This …nding is in line with expectations given the emphasis placed on developing entrepreneurship skills in the vocational training component of the ELA. Column 2 of Table 4 shows all these results to be robust when estimated by OLS using the …rst di¤erence speci…cation in (2) that allows for time varying impacts of individuals characteristics at baseline. Column 3 also shows that the ITT results for the outcomes that are dummy variables: whether any income generating activity is engaged in, whether self-employed activities are engaged in, and whether wage employment activities are engaged in, are all robust to estimating them using a Probit speci…cation analogous to that in (1). The next few rows focus on the impacts on earned income and labor supply from self and wage employment. To begin with, we …rst report OLS estimates in Column 1 including zero values for earnings and hours. These are informative of the programmes impact on the average adolescent girl in the panel survey, including girls that do not engage in any income generating activities at baseline. The ITT impacts on earnings or hours – averaged over all adolescent girls – are not signi…cantly di¤erent from zero. However, both earnings and hours outcomes are censored at zero and so are observed only if the corresponding latent outcome, ���� ¤ , is strictly positive, namely ¤ ���� = 1(���� ¤  0)���� . Hence these OLS estimates combine the responses of those: (i) uncensored at baseline, weighted by the probability of being uncensored; and, (ii) the change in the probability of being uncensored, weighted by the expected value of the outcome if the latent variable is strictly positive. We therefore use Tobit speci…cations to explore the ITT impacts on income and labor supply hours from wage and self-employment in more detail. For both income and hours the programme can potentially cause: (i) a response at the extensive margin, so that some girls start engaging in such activities in response to the programme; (ii) a response at the intensive margin, so that some girls that were already engaged in some income generating activity at baseline, expand such activities in response to the programme. On income, Column 4 shows the programme signi…cantly increases the likelihood that girls earn income from self-employment by 42pp. This measures the ITT impact on the propensity for latent earnings to be positive, 1(����¤  0). There is no corresponding increase on the extensive margin for income from wage employment as expected. On the intensive margin, Column 5 shows that income from self-employment activities increases by almost ��� 55 000. This measures the ITT impact on expected latent earnings, conditional on these being strictly positive, � (���� ¤ ¤ j����  0). To benchmark this impact we note that among girls that were uncensored at baseline, average earnings from self-employment were ��� 368 394 so that the programme impacts corresponds to a 148% increase relative to this baseline. There is no signi…cant change in income from wage employment although the point estimate is negative, suggesting that some girls might be substituting away from wage-employment towards self-employment. On hours worked, Column 4 shows that on the extensive margin, the programme signi…cantly increases the likelihood that adolescent girls spend any hours engaged in self-employment activities, 15 by 42%. The next row in Table 4 shows that the programme increases the daily hours worked in self-employment by 460, that is also di¤erent from zero at the 5% signi…cance level. This starts at such a low level because, as shown in Table 1, over 70% of the girls are enrolled full-time in school at baseline. For hours of wage employment, the …nal row in Table 4 there are no signi…cant changes on the extensive and intensive margins, although the point estimates are again negative, reinforcing the suggestion that for those 6% of girls that were engaged in wage-employment at baseline, there might be some substitution away from wage-employment towards self-employment as a result of the programme. The …nal two Columns of Table 4 report ATE impacts of actual participation on each outcome related to income generating activities based on the instrumented …rst-di¤erenced speci…cation (3). In Column 6, the ATE estimates are based on club participation being instrumented by the o¤er of treatment. In Column 7 the ATE estimates are based on participation in at least one vocational skills courses being instrumented by the o¤er of treatment. As expected these scale-up the ITT impacts reported in Column 2. However, the estimated impacts of the programme on self-employment income are no longer precisely estimated, as was the case for the earlier OLS estimates, because both extensive and intensive margin responses are pooled together. Given the documented ITT impacts on income generated from self-employment, it is worth probing how such additional income is spent. The next row in Table 4 evaluates the impacts on personal consumption expenditures aggregated across eight items, in the last month.25 Column 1 shows that in treated communities, expenditures signi…cantly rise by around � �� 4 238, corre- sponding to 34% of their baseline value, and that this result is robust to estimation using the …rst di¤erence speci…cation in Column 2. Recall that the intensive margin Tobit results in Column 4, gave an annual increase in income from self-employment of ��� 54 000. If these monthly expen- diture increases of � �� 4 238 are constant for all twelve months of the year, then around 94% of increased income is being consumed in these categories. In comparison with the literature, these results are encouraging as other programmes deliver- ing standalone …nancial and entrepreneurship training and evaluated using RCT methods, have found generally weak impacts – see for example Field et al. [2010], Drexler et al. [2010], Karlan and Valdivia [2010], Fairlie et al. [2012] and Bruhn et al. [2012].26 This is despite the fact that 25 The goods categories are jewelry/ornaments, cosmetics/makeup, clothes, hairdressers, shoes/footwear, going to restaurants/bars/teashop/cafe, talk time for your mobile phone and presents/gifts. 26 Field et al. [2010] report results from an RCT that provides basic …nancial literacy training to a randomly selected group of female entrepreneurs in India. Only a socially restricted sub-group were found to bene…t from the intervention in terms of business income and borrowings. Drexler et al. [2010] …nd that teaching accounting principles to a group of micro-borrowers in the Dominican Republic has no impact on either the way they run their business nor on business outcomes. However, simple rule-of-thumb style training does a¤ect the way in which …nancial records are kept. Karlan and Valdivia [2010] investigate the impact of a relatively intense training intervention of up to two years, that delivered comprehensive training on business practices to clients of a Peruvian Micro…nance institution. Despite improving business knowledge, the intervention failed to impact any major business outcomes. Fairlie et al. [2012] …nd that providing entrepreneurs training has no measurable impact on business operations in the long-run. Finally, more recent evidence in Bruhn et al. [2012] suggests granting small and medium enterprises in Mexico access to consulting services – that are orders of magnitude more costly than the other forms of business intervention described above – does have large positive impacts on …rm pro…ts, although not their employment. 16 other programmes are often targeted towards those who have self-selected to be small-scale en- trepreneurs. Our evidence suggests such outcomes related to business skills and outcomes can be signi…cantly improved even among girls who ex ante, might not consider themselves as being on the verge of setting up a small-scale enterprise. The second natural point of comparison is with the literature evaluating standalone vocational training. Such interventions are often found to have limited impacts in developed [Blundell et al. 2004] and developing country contexts [Card et al. 2011, Groh et al. 2012]. Among those studies that …nd impacts, Attanasio et al. [2012] …nd that for women, the likelihood to be employed increases by 6.1pp, and earnings increase by 22% relative to a randomly-assigned control group. These are comparable impacts to those we …nd for the two-pronged ELA intervention. Finally, we address an important concern behind the introduction of vocational skills training: the potential adverse e¤ect on schooling investments made by bene…ciaries of the ELA programme. More precisely, the potential trade-o¤ is that as the programme targets of school going age, adolescent girls might be encouraged to drop out of school as a result of the programme. The …nal few rows of Table 4 explore whether there are adverse impacts in this dimension. The estimated ITT impact of ELA on an indicator which equals one if the respondent is cur- rently enrolled in a school and zero otherwise is negative is statistically indistinguishable from zero. The programme does not therefore appear to have caused di¤erential school drop-outs between treatment and control communities two years after its introduction. This …nding is con…rmed by the fact that the programme does not impact the number of hours spent on studying/school per week for those currently enrolled. In fact, there are hints that the programme increased the value attached to education in treated villages. The …rst-di¤erenced OLS speci…cation implies the programme seems to have motivated a signi…cantly higher proportion of 111% of those who have dropped-out of school in the past to at least consider going back to school.27 4.2 Risky Behaviors To benchmark the impacts of the programme along each dimension of risky behavior, the …rst Column in Table 5 shows the baseline value of each outcome in treated communities. Rows 1 and 2 of Table 5 show the ITT impact of the programme on knowledge indicators related to risky behaviors. The presence of an ELA club in the community signi…cantly raises the HIV knowledge index by 462 relative to a mean at baseline of 383 in treated communities. On the single question related to pregnancy knowledge, this also signi…cantly rises with the programme, by 67pp relative to a baseline mean of 739%. As such, given that the majority of girls already display sound knowledge on this dimension at baseline, the programme is able to reach those relatively few girls that were uninformed at baseline. Improved knowledge is only a prerequisite to triggering actual behavioral change in adolescents’ lives. The remaining rows of Table 5 show the programme’s impact on …ve such indicators. First, 27 If we focus on those girls that were not in school at baseline, we do not …nd a signi…cant impact of the programme on their enrolment at follow-up. 17 the programme has a strong negative impact on fertility: the di¤erence-in-di¤erence ITT impact in Column 1 shows that the probability of having a child grows at a signi…cant lower rate, by 27pp, in treated communities. Given that at baseline 104% of girls have at least one child, this is a near 26% drop in fertility rates over a two year time window. In line with the reduction in the incidence of child-bearing, treated adolescents report a substantial increase in condom use. Among those who are sexually active, the percentage who always use a condom when having intercourse increases by 125pp.28 The next row shows that this increase in those reporting always using a condom is almost equal and opposite to the reduction in those reporting using condoms often or sometimes. The next row shows that among the sexually active there is little evidence the use of other forms of contraception having increased. This is reassuring given the limited availability of such alternatives, so that the results on condom and contraceptive usage do not seem to re‡ect girls merely repeating what they have been taught in life skills courses, or what they believe enumerators wish to hear. Somewhat surprisingly though, practicing safer sex through increased condom use is not mir- rored in the suspected infection rate with a STD as shown in the next row, nor in signi…cant di¤erences in treatments sought if an STD is suspected. The …rst result is in line with the …ndings in Du‡o et al. [2011] of the standalone provision of a HIV-curriculum, and might in part be driven by the fact that adolescent girls are unable to accurately assess whether that have an STD or not. Finally, the rate of adolescents who report having had sex unwillingly during the past year decreases by 161pp in treated communities. Starting from a baseline of 21% this corresponds to a near 76% reduction in the incidence of such events. This provides a dramatic illustration of the programme enabling girls to become empowered in their relations with men. This impact is likely a direct result of the life skills sessions on negotiation, rape and legal rights, through which girls are sensitized and discuss prevention measures.29 Column 2 of Table 5 shows all these results to be robust when estimated by OLS using the …rst di¤erence speci…cation in (2), that better accounts for the panel nature of the data and allows for the impact on outcomes of individual characteristics at baseline ����0 to vary by time. The last eight outcomes related to risky behaviors in Table 5 are dummy variables. Hence in Column 3 we report ITT estimates based on Probit speci…cations analogous to (1), and report marginal e¤ects 28 As argued in Dupas [2011], childbearing is not a perfect proxy for the incidence of risky sex because: (i) ado- lescent girls in long-term relationships are more likely to get pregnant than girls in several short-term relationships; (ii) teenage girls might be more likely to abort if the father is a teenage boy who cannot provide economic support; (iii) adolescent girls might be more likely to engage in anal sex with partners as a way to avoid pregnancy, and this is especially risky in terms of HIV transmission. The concern that such changes in behavior might be driving fertility drops is partly ameliorated by the increased self-reported condom usage. 29 The coe¢cients reported in Table 4 relate to the coe¢cient of interest,  in (1). However it is also instructive to benchmark this also against the time trends in knowledge and behaviors, as measured by  in (1). We note that on both knowledge indicators, we observed signi…cant reductions in knowledge among girls over time: by ¡379 on the HIV index, and by ¡7�� on the pregnancy knowledge index. Whether this decline in knowledge is due to the spread of misinformation among older girls, or some other reason, is beyond the scope of this study but worthy of future research. In terms of time trends in behaviors, we note that over time, there are signi…cant reductions in the percentage of girls who report using condoms if they are sexually active, signi…cant increases in girls reporting having su¤ered an STD, and signi…cant increases in the percentage of girls reporting having had sex unwillingly in the past year. The program impacts are all the more remarkable given this background. 18 throughout. The earlier results are largely robust to this non-linear speci…cation and remain of similar magnitude and signi…cance. The last two Columns of Table 5 present estimates of the ATE of actual participation on these outcomes related to risky behaviors. To do so, we use the …rst di¤erenced speci…cation in (3). In Column 4, the ATE estimates are based on club participation being instrumented by the o¤er of treatment. In Column 5 the ATE estimates are based on participation in the life skills courses being instrumented by the o¤er of treatment. In both cases, the magnitude of the coe¢cients of interest are ampli…ed by an order of mag- nitude over the ITT estimates. This is an almost mechanical consequence of the di¤erential participation rate between adolescents in treatment and control communities. All the key OLS results discussed before remain intact when estimating the ATE impact as shown in Column 4 in Table 5: namely that on dimensions related to knowledge and practices of risky behaviors, adolescent girl outcomes signi…cantly improve on six out of nine margins.30 One concern is that the outcome measures are all self-reported and so adolescent girls might merely be telling interviewers what they expect them to say. A few results point against this interpretation of the …ndings. First, not all margins signi…cantly improve despite being part of the life skills training girls receive, and as discussed above, girls are not reporting increased use of contraceptive methods that are largely unavailable in this setting. Second, we have explored further outcomes related to behaviors: for example we have estimated the likelihood girls report getting married between the baseline and follow-up. This likelihood is also found to be signi…cantly lower in treated communities relative to control locations. Delaying the onset of marriage is an important channel through which adolescent girls are able to continue to accumulate human capital and improve their long term earnings potential [Field and Ambrus 2008, Baird et al. 2011]. Comparing our …ndings to the literature, we note …rst that meta-analyses of previous interven- tions report generally weak impacts of standalone HIV-education programmes, be they delivered via classroom-based courses [Gallant and Maticka-Tyndale 2004, McCoy et al. 2010, Du‡o et al. 2011] or peer-provided courses [Cornish and Campbell 2009]. There are two recent studies that …nd impacts of standalone education programmes that are worth comparing to. First, Arcand and Wouabe [2010] use a regression discontinuity design to estimate the impacts of a school-based HIV prevention course in Cameroon. Their estimated impacts on childbearing and condom usage 30 We also experimented with augmenting the …rst stage speci…cation by using additional instruments based on distance from the girls’s household to the club. All else equal, one would therefore intuitively expect girls to be more likely to participate in ELA if the nearest club is located close to their home. The instrument validity, however, may be potentially undermined if the distance measure correlates with the outcome variable because, for example, the distance proxies for school and market opportunities, infrastructure or parental background. To counter this concern, we also additionally included a control variable for the distance between the adolescent’s household and the nearest BRAC branch o¢ce. These o¢ces tend to be located at a central point such as the main street in urban settings or the main trade center or town in more rural areas. The inclusion of this variable is thus intended to control for at least one potential dimension of geographical remoteness. The …rst stage regressions then con…rmed that treatment assignment and distance to the club were signi…cantly correlated to participation (distance to the nearest BRAC branch o¢ce had no signi…cant impact on the likelihood of participation). Moreover, in the second stage, the estimated ATE impacts were similar for the majority of life and livelihood skills outcomes documented. However, the distance measures are computed from GPS coordinates and this spatial data is only available for a subset of respondents. Hence we focus on the ATE estimates that use only treatment assignment as an instrument. 19 lie between the ITT and ATE estimates we …nd. Second, Dupas [2011] uses an RCT design to compare the e¤ectiveness of the Kenyan national HIV curriculum to an intervention providing information on the relative risk of HIV infection by the partner’s age. She …nds that exposure to this curriculum causes a 28% reduction in teenage pregnancies over a one-year period, which is in line with the 26% drop in fertility rates due to ELA over a two-year period. In the Appendix we present additional robustness checks on the baseline results from Tables 4 and 5. The checks address issues relating to multiple inference, impact heterogeneity, sample attrition, and the results being driven by the anticipation of micro…nance components in the programme. 4.3 Transforming Lives? To provide a richer picture of how the ELA programme transforms the lives of adolescent girls, we consider outcomes not directly related to risky behaviors or income generating activities, but are proxies for welfare, self-con…dence and empowerment that might signal longer term impacts of the programme. The results are shown in Table 6. The …rst row relates to the programme impacts on an index of ‘gender empowerment’, scaled from 0 to 100. This is based on a series of questions relating to gender roles in various tasks. A higher index value corresponds to girls believing that tasks should be gender neutral or more equally shared among genders.31 The gender empowerment index signi…cantly rises among girls residing in treated communities, in line with the programme’s underlying aims beyond risky behaviors and income generation. Figure 2B presents a spider graph showing the ITT impacts (and their associated 95% con- …dence interval) for each component of the empowerment index. We see that the programme signi…cantly increases self-reported gender empowerment along a few dimensions. The programme has particularly strong impacts on attitudes along the dimension of whether girls agree with the statement that ‘females should earn money for the family’. Among girls resident in treated com- munities, there is an increase of 182pp of girls agreeing with this statement, relative to a baseline of 372% agreeing with it. The next index summarizes work and society related attitudes, again scaled from 0 to 100.32 On this dimension the programme has no impact. The corresponding spider graph for the attitude 31 The empowerment index is a variable that cumulates the number of times a respondent answers “Both/Same� to the following questions: “Who should earn money for the family?�, “Who should have a higher level of education in the family?�, “Who should be responsible for washing, cleaning and cooking?�, “If there is no water pump or tap, who should fetch water?�, “Who should be responsible for feeding and bathing children?�, “Who should help the children in their studies at home?� and “Who should be responsible for looking after the ill persons?� The other possible answers given to the respondent were “Male� and “Female�. The index is then re-scaled such that 100 indicates that the respondent answered that both sexes should be responsible for the mentioned activities. 32 The attitude index is based on the respondent’s rating of the following statements: "If I start working on a task, I de…nitely see the end of it no matter how di¢cult it is", "While doing any task, it is important for me to do it better than others", "If I have the chance, I would make a good leader", "I want to be a respectful person in my village", "I do not care what others think about my success or failure", "I am in control of what happens in my life", "I save regularly", "A person can get rich by taking risks", "I often make plans for the future", "I believe that my future is determined by luck no matter how hard I work". A rating of 10 indicates that the respondent agrees with the statement a lot and 1 that the statement is not true at all for her. The index corresponds to the re-scaled cumulative score. 20 index in Figure 2C, shows the programme has no impact for any of these attitudinal measures, at least as measured two years after the programme’s initiation. This result also need to be balanced against the fact that the baseline scores for the attitudes index are also relatively high and leave less margin for improvement than the gender empowerment and entrepreneurship indices. The next few rows in Table 6 explore impacts on life satisfaction. Here the …ndings are somewhat mixed. On the one hand, the programme appears to have little impact on an overall index of life satisfaction, scaled to run from 0 to 100.33 Although the baseline measure on this index is 451, we do not …nd any signi…cant ITT or ATE impact of the ELA programme on this aggregate index. However more detail is again revealed through the corresponding spider graph for the life satisfaction index, shown in Figure 2D. This highlights that the programme impacts life satisfaction in one life sphere: satisfaction with earnings/income, a margin obviously closely connected to the programme. As reported in Table 7, satisfaction with earnings/income signi…cantly rises by 570 on a 0 to 6 scale from a baseline level of 120. Along similar lines, the next row in Table 6 shows that girls signi…cantly increase reports of never worrying about …nding a good job in adulthood. Among participants to the ELA programme, the ATE estimates in Column 4 shows the likelihood of not worrying about getting a good job rises by 38pp from a baseline of 42%. This might suggest the programme is able to reduce stress levels among girls as they transit into the labor market, precisely as intended and part of the underlying motivation for the intervention. Finally, we consider outcomes related to girls opinions and expectations on marriage and childbearing. As emphasized in the introduction, the patterns of early marriage and child bearing observed in Uganda and much of Sub-Saharan Africa often lie at the heart of adolescent girls’ restricted opportunities to acquire human capital, engage in productive labor market activities, and increases their exposure to health risks. Table 7 shows how the programme impacts adolescent girls’ aspirations related to marriage (Panel A), childbearing (Panel B), and expectations over their own childrens’ lives (Panel C). The …rst row in Panel A relates to the age at which girls anticipate themselves getting married. This was only reliably collected at follow-up.34 The ITT estimate suggests this expectation signi…cantly increases by almost one year for girls residing in treated communities. The last two rows in Panel A relate to girls’ views on ideal ages at marriage for men and women in society as a whole, rather 33 The satisfaction index adds up the responses about the happiness of the respondent about di¤erent aspects of life, namely: education level, family, friends, job, earnings/income, house you live in, the school you go to (if currently studying), school work (if currently studying), and life as a whole. The respondents were asked to indicate their satisfaction on a scale where 1 stands for completely happy and 7 for not at all happy. For our analysis the original coding was reversed and re-scaled so that a cumulative score of 100 stands for completely happy in all aspects and 0 for not at all happy in all aspects. Analogously, the variable capturing satisfaction with earnings/income is reversed such that 6 stands for completely happy and 0 for not happy at all. 34 In our baseline survey, there was some ambiguity as to whether the question referred to the anticipated age at marriage, or the anticipated number of years until they married. We resolved this issue in the follow-up survey and hence only use that data for estimating the ITT impact. When doing so we drop a very small number of observations among girls who responded that they either they would never get married, did not know at what age they expected to marry, or reported an anticipated age above 51. Girls that are married at the time of the follow-up are not asked this question. 21 than in speci…c relation to their own age at marriage. On this dimension girls residing in treated communities report signi…cant increases in these suitable ages for both men and women.35 Panel B focuses on aspirations related to childbearing. We see that there is a signi…cant increase in what girls report being the most suitable age for women to have their …rst child: the OLS estimate in …rst di¤erences is 437, although the ATE is more economically signi…cant at 277 years. At the same time, there is a more robust signi…cant reduction in the preferred number of children: this falls by 368 in our baseline speci…cation, corresponding to a 87% fall from baseline reports on ideal numbers of children. Panel C shows how the adolescent girls responded to questions related to how their own chil- drens’ lives might evolve. To be clear, this referred both to any children alive at the time of the follow-up survey as well as those that were yet unborn. On this dimension, the evidence suggests the potential for the ELA programme’s bene…cial impacts on attitudes to persist across genera- tions: treated girls prefer their daughters to get married at a signi…cantly older age (although no such e¤ect exists for any sons they might have). Taken together, these results on o¤er some hope that the programme is able to set in chain processes and beliefs that delay age at …rst marriage and child-bearing. As such, these girls’ lives might then improve along dimensions that have been shown to be associated with such delays such as improved marriage quality, increased decision-making within households and reduced exposure to domestic violence [Goldin and Katz 2002, Jensen and Thornton 2003, Field and Ambrus 2008]. 5 Discussion 5.1 Cost-Bene…t-Analysis For societies facing the challenge of rising youth populations and social norms whereby girls marry and have children early, increasing their economic dependence on older men at the expense of labor force participation, the ELA programme o¤ers some promise. In this section we document the cost of the programme to benchmark how large the per girl bene…ts would have to be for the programme to be sustainable for a social planner. Table 8 categorizes the …xed and variable costs of the programme. The variable costs depend on the number of participating girls. Depending on whether the costs are incurred once only or recur each month, we list the amounts in Column 1 or 2 respectively. Columns 3 and 4 then split each cost into its …rst and second year component respectively. All costs are in US$. Rows 1 to 3 show the costs associated with the initial programme investment of setting up a programme o¢ce, training of programme sta¤ and programme manual development. The second 35 Adolescent girls were also asked who they thought would be involved in deciding their marriage partners (not shown). We …nd that among treated girls there is a signi…cant reduction in the likelihood they report the choice will be made by them alone, and a corresponding increase of similar magnitude in the likelihood they report decisions over marriage partners will be made in conjunction with their parents. This might be taken as tentative evidence that higher quality marriage partners being sought, as well as changes in the timing of marriage as reported in Table 7. 22 set of …xed costs in Rows 4 to 14 comprise all cost items that are necessary to provide the in- frastructure for the ELA clubs to function (irrespective of the number of actual club participants). Finally, Rows 15 to 19 detail the variable costs of the programme. Summing across all costs in the 100 treated communities, Row 20 shows that in year one, the programme costs $365 690. This falls to $232 240 in year two as some of the set-up costs are sunk and do not recur. As described earlier in Section 3, we conducted a pre-baseline census listing of all households in the communities involved in the evaluation. This revealed that around 130 eligible adolescent girls resided in the average community. Given the bene…ts we document relate to ITT estimates of residing in a community that is o¤ered the ELA programme, we use this number of eligible girls we use to calculate the per girl cost of the programme. Hence in the fourth panel of Table 8, Rows 21 and 22 show the average …xed and variable costs per eligible girl. The overall cost per eligible girl is shown in Row 23. Given our ITT estimates are measured two-years after the baseline, we focus on the second year per-girl incurred cost of $179. To put the cost estimate in context, we note that $179 corresponds to only 54% of household incomes at baseline. Alternatively, this cost corresponds to 210% of the annual expenditures of adolescent girls on the sub-set of consumption goods described in Table 5.36 If the per girl bene…ts to an adolescent girl, of residing in a community that is o¤ered the ELA programme, are larger than this, it would suggest the programme is sustainable from the social planner’s perspective. Our …ndings indicate that the programme increases labor force participation and earnings while reducing risky behaviors. The labor market impacts of the programme are easily monetized. To evaluate these, the …nal row of Table 8 zooms in on those girls engaged in self-employment at baseline and follow-up, and shows the Tobit estimate for earnings conditional on earnings being strictly positive: this increase of $321 is signi…cantly di¤erent from zero and more than o¤sets the per girl programme cost among these girls.37 The impact or reducing risky behaviors such unprotected sex and teen pregnancy as shown in Table 5, or the changes in empowerment, aspirations and beliefs over early marriage and child bearing as shown in Tables 6 and 7, are much more di¢cult to monetize but all obviously adds to these estimated bene…ts. Even if the bene…ts of the programme outweigh its costs, the question of whether the same resources could be spent more e¤ectively remains open. The combined ELA intervention appears to improve outcomes at least as well as single-pronged interventions that have focused exclusively on classroom-based education courses designed to reduce risky behaviors, or exclusively on vocational training designed to improve labor market outcomes among youth. As a body of randomized- control trials suggest, these single-pronged programmes have met with, at best, rather mixed success [Gallant and Maticka-Tyndale 2004, Cornish and Campbell 2009, McCoy et al. 2010, Card et al. 2011]. One class of vocational training programmes that has met with some success are the Jovenes 36 These items include jewelry/ornaments, cosmetics/makeup, clothes, hairdressers, shoes/footwear, going to restaurants/bars/teashop/cafe, talk time for your mobile phone and presents/gifts. 37 The …gure of $321 is taken from the corresponding estimate in UGX shown in Column 5 of Table 4, where we use the exchange rate of UGX1700 to one US dollar. 23 programmes implemented throughout Latin America. For example, Attanasio et al. [2012] …nd that for the Jovenes programme in Colombia, among women, the likelihood to be employed increases by 6.1pp, and earnings increase by 22% relative to a randomly-assigned control group. These are comparable impacts to those we …nd for the two-pronged ELA intervention. However the costs per trainee of the Jovenes programmes vary from $600 to $2000 per participant served [World Bank 2009]. These costs are still an order of magnitude larger than the $179 per eligible girl of the ELA programme, or given a 21% take-up rate, a cost of $85 per participating adolescent girl.38 Moreover, even apart from di¤erences in programme costs, their remains doubt as to whether private sector …rms in Sub-Saharan Africa would be able to currently take part in Jovenes style training programmes for workers. A …nal approach to understanding whether the ELA programme is socially bene…cial is to consider the impacts of providing unconditional cash transfers in a similar setting. This is precisely what is considered in Blattman et al. [2011], who present evidence from the Youth Opportunities Programme (YOP) using a randomized control trial in which youth were given unconditional and unsupervised cash transfers.39 They …nd that almost 80% of youth chose to spend these transfers on acquiring vocational skills and tools, and that the resultant increase in earnings imply an annual return on capital of 35% on average. There are of course many di¤erences between the treated individuals in the ELA and YOP programmes: the YOP targets both genders and those aged 16 to 35; individuals form groups to apply for the unconditional transfers; the per person transfer $374. Although the ELA programme can be thought of as a constrained version of such unconditional cash transfers, even if the rates of return through labor market outcomes alone are half as much, this still compares favorably with regards to other formal sector …nancial investment opportunities available in Uganda in mid-2008 when the ELA programme was initiated.40 5.2 Concluding Remarks Developing countries face enormous challenges stemming from rapid population growth and a rising proportion of young people in the population. This leads naturally to the demand for policies that equip young people to e¤ectively transit into the labor market. Among adolescent girls, such economic concerns are compounded by social contexts characterized by early wedlock and pregnancy, where girls incentives and ability to invest in their human capital is reduced early in life and they are exposed to health risks such as HIV infection. 38 If we also take account of the fact that 52.7% of girls regularly attending the adolescent girl club take part in the vocational training courses, then an upper bound per girl cost estimate is $161 39 Similarly, Baird et al. [2012] report that the provision of unconditional cash transfers via lotteries, to girls aged 13-22 and enrolled in school at baseline in Malawi, signi…cantly reduced the prevalence of HIV and herpes simplex virus 2 [HSV-2] after 18 months. These e¤ects were also supported by self-reported sexual behaviors. To gauge the cost per treated girl, we note that monthly cash transfers valued at between $4 and $10 were provided to girls along with monthly transfers of between $1 and $5 to their guardians. 40 For example, the International Financial Statistics of the IMF state that the deposit rate in the formal sec- tor in Uganda (i.e. the rate paid by commercial banks for savings deposits) was 10.7% in 2008, 9.75% in 2009 and 7.69% in 2010. An alternative investment would have been to buy a two-year Uganda Treasury bond auc- tioned at the end of May 2008. It sold at a discount and yielded 14.45% according to the Bank of Uganda (http://www.bou.or.ug/bou/collateral/tbond_forms/2008/May/tbond_28May2008.html). 24 We have evaluated the ELA programme, operated by the NGO BRAC in Uganda, that simul- taneously provides girls vocational skills to enable them to start small-scale income generating activities, and, life skills to build knowledge and reduce risky behaviors. The …ndings from our randomized control trial on around 4 800 adolescent girls show positive impacts along both mar- gins two years after the programme begins. Girls in treated communities are signi…cantly more likely to engage in self-employment and have higher earnings. Importantly, these gains do not come at the expense of girls’ current investment in human capital, as we observe no reduction in those being full-time enrolled in school. On risky behaviors, girls in treated communities are signi…cantly less likely to be exposed to risks from HIV and other STDs: they are found to be signi…cantly less likely to engage in unprotected sex, less likely to have sex unwillingly, and overall, they are less likely to have been pregnant and had a child relative to girls in control communities. We also document markers that these gains could be long lived, as re‡ected in the signi…cantly changed aspirations and feelings of empowerment that treated girls report on a range of dimensions related to gender relations, and perhaps most crucially, expectations (for themselves and their own children) in relation to ages at marriage and …rst pregnancy. It is precisely these issues that lie at the root of the root causes of poverty and limited economic opportunity the ELA programme is originally designed to address. Taken together, the …ndings provide a promising indication for the e¤ectiveness of combining life skills and vocational skill training in a club environment with voluntary participation. The evidence suggests the ELA programme does at least as well as many interventions that exclusively target either risky behaviors through classroom-based education behaviors, or vocational skills through vocational training courses [Gallant and Maticka-Tyndale 2004, Cornish and Campbell 2009, McCoy et al. 2010, Card et al. 2011].41 The …ndings suggest interventions that simultaneously try to reduce informational constraints related to risky behaviors and reduce constraints on the provision of skills related to income generation, can have bene…cial, quantitatively large and sustained impacts on adolescent girls along both dimensions. The external validity of the these results are currently being researched as ELA-style programmes have been rolled out and are being evaluated in a range of Sub-Saharan countries. The programme o¤ers some promise to policy makers, as being a low cost and scalable intervention that enables adolescent girls to improve their life outcomes. This is especially relevant in developing countries facing a ‘youth bulge’ and a consequent need to skill and provide meaningful opportunities to their youth. The gains to a twin-pronged ELA-style programme are especially acute among adolescent girl populations, who face similarly constrained labor market opportunities as men, but are also impacted by norms of early mar- riage, childbirth and engagement in risky behaviors. Absent early twin-pronged interventions that simultaneously tackle both risky behaviors and the provision of vocational skills, such cohorts may otherwise be resigned to worse life trajectories that become harder or impossible to reverse through later policy interventions. 41 Cornish and Campbell [2009] document how in Africa, such peer education models in HIV related education remain the preferred approach, being used by around 60% of all major HIV prevention NGOs [Kelly et al. 2006]. 25 A Appendix A.1 Multiple Inference The large set of outcomes analyzed on risky behaviors and income generating activities raises concerns about multiple inference, namely the probability of erroneously rejecting at least one null hypothesis of zero impact naturally increases with the number of outcomes considered. We deal with this using two approaches. First, we adjust the p-values on the ITT estimates using the free step-down procedure, detailed in Westfall and Young [1993] and Anderson [2008], to control for the probability of rejecting at least one true hypothesis tested.42 The second approach combines several outcome variables into a single index and, thus, de- creases the number of hypothesis to be tested. This method is discussed in Kling et al. [2007] and Anderson [2008]. We follow Kling et al. [2007] by re-scaling outcome variables such that higher values indicate better outcomes, demeaning the outcomes, and dividing them by the standard deviation measured using the control group at baseline only, and imputing missing values using the mean of the variable in the appropriate survey wave for the adolescent’s treatment status group. We then construct the index as the mean of equally weighted individual outcome variables. One such summary index is created for each of the two broad outcome domains, namely income generating activities and life skills, and then regressed on the usual set of control variables using OLS. The summary index (IGA) is based on the entrepreneurship ability and the indicators for self-employment and wage employment. The summary index (life skills) is based on all outcome variables in the upper panel. The results are reported in Table A2. Column 1 restates the previously estimated ITT impacts on all major outcomes with their corresponding original and standard p-values shown in Column 2. The adjusted p-values using the free step-down re-sampling method intended to account for multiple inference are shown in Column 3. For most of the outcomes considered, the conservative free step-down p-values leave the key …ndings and interpretations intact except for the ITT impacts on pregnancy knowledge, STD proxies and hours spent on self-employment. Finally, Column 4 of Table A2 reports the p-values for the two summary indices. Together with the estimated coe¢cient on the time-treatment interaction term, the p-values indicate positive and signi…cant ITT impacts of the ELA project on the two indices of risky behaviors and income generating activities. A.2 Impact Heterogeneity We next check for whether particular subsamples of the data are driving the main results on risky behaviors and income generating activities. To do so, we present ITT estimates on the 42 Following this procedure, original p-values are compared to a subset of p-values obtained for the entire family of outcome variables in each resampling iteration. This feature makes the corrected p-values more conservative than originally but less conservative than, for example, the Bonferroni or Šidák p-values which are based on an implicit comparison with the entire set of p-values contained in one family. Given that all simulated p-values within one family are based on the same treatment assignment, the dependency structure between the outcomes stays intact which substantially increases the power if outcomes are correlated as detailed in Anderson [2008]. The adjusted p-values are based on 10,000 replications under the null hypothesis of no programme impact. 26 two summary indices, derived for the multiple inference checks above, for outcomes related to income generating activities and risky behaviors. We also check for heterogeneous impacts among outcomes related to current investments in schooling. We explore impact heterogeneity along the following dimensions: (i) rural versus urban households; (ii) rich versus poor households, as de…ned by whether the household’s asset values at baseline are above or below the median for all households; (iii) girls aged above 16 at baseline versus older girls at baseline. Table A3 shows the results. We see that for each dimension, the impacts found are largely homogeneous across rural and urban areas, rich and poor households, and young and old girls. For the summary index related to income generating activities shown in Panel A, the evidence suggests the impacts are weaker among adolescent girls in households that are wealthier at baseline, as shown in Columns 4 and 5. As such the programme reduces the gap in labor market activities between rich and poor households. For the summary index for risky behaviors shown in Panel B, there is also little evidence of impact heterogeneity across these margins. This demonstrates, for example, that behavioral change can be induced with regards to risky behaviors among eligible girls of all ages from 12 to 20. This might not have been the case for older girls if such behavior were habitual for example. Similarly, this might not have been the case for younger girls if they particularly lacked bargaining power or negotiation skills in their relationships with men.43 Panel C of Table A3 shows how for education related outcomes, the previously documented ITT impacts are also similar across the relevant dimensions of heterogeneity. In particular, it remains the case that the programme does not encourage girls to drop out from schooling, and this applies equally to rural and urban areas, rich and poor households, and young and old girls. This is again encouraging: if, for example, girls were especially myopic, the incentives to drop out of school in the presence of the programme might be higher in rural areas where the returns to education are limited due to a lack of labor market opportunities outside of the agricultural section [Kasirye 2011].44 A.3 Attrition To check the robustness of the results to panel attrition, we re-estimate each ITT impact using weights, where these weights are constructed from the inverse of the estimated probability of panel inclusion, as detailed in Wooldridge [2002]. The rationale behind using inverse probability weighting is straightforward: individuals who are very likely to have been re-interviewed receive less in‡uence through weighting them by the inverse of the estimated probability of panel inclusion. On the other hand, adolescents likely to be under-represented in the panel receive higher weights in order to strengthen their contribution to the estimated impacts. 43 The results for younger girls are especially encouraging given the conventional wisdom that girls aged 10-14, particularly those out of school, that face the greatest economic challenges and health challenges arising from unsafe sexual behavior in this context [UNICEF 2003]. 44 Kasirye [2011] reports that four out of …ve women in Uganda are employed in agriculture, according to the 2008 Gender and Productivity Survey in Uganda, and 42% of women in the labor force are unpaid family workers. 27 The key step in this procedure is a Probit regression of the dummy variable capturing panel inclusion on the usual set of control variables included in the main outcome regression as well as baseline variables with decisive in‡uence on the panel status. The latter set of variables contains in our case dummy variables for each surveyor, the number of named friends, a measure for the distance between the village and the BRAC branch o¢ce, the household size, an indicator of whether there is at least one household member who migrated away in the last year as well as a dummy variable indicating membership in an NGO of at least one household member. From this regression predicted probabilities are obtained which are then inverted and used as weights in the actual outcome regressions. The results obtained are presented in Column 5 of Table A2. These are very similar to the baseline ITT estimates shown in Column 1 so that, as suggested in Table A1, selective attrition over time does not appear to be driving the main results. This is the case both for the individual outcomes shown in able A2 as well as the summary indices for all risky behaviors, and for all outcomes related to income generating activities. A.4 Anticipation of Micro…nance This evaluation is based on changes in behavior in the two years after the initiation of the ELA programme. In the year that followed, the programme expands to comprise a micro…nance com- ponent, that o¤ers participating adolescents the opportunity to take on desired credit in order to capitalize on their entrepreneurial potential and newly acquired skills during the vocational training. Micro…nance was randomly assigned to half the treated communities after the two-year period. At the time of the current evaluation, adolescent club participants were unaware of the potential future o¤er of micro…nance. BRAC sta¤ were unaware of which clubs would be assigned to receive micro…nance in addition. However, one concern is that the results presented here may be picking up the e¤ects of participants’ anticipating they will receive credit. To check for this we focus solely on the sample of 100 treated communities and then estimate whether within this sample, the future random assignment to micro…nance predicts outcomes in the …rst two years of the programme. We therefore estimate a speci…cation analogous to (1) where ������ refers to whether the community will in the future receive micro…nance or not: something that is observed to the researchers but not the adolescent girls nor to BRAC workers in the …rst two years of the programme. The coe¢cients on this treatment dummy variable are shown in Column 6 of Table A2 for each outcome of interest. Reassuringly we see that for each of the life skills and income generating activity outcomes, there are no signi…cant ITT impacts of future assignment to micro…nance. References [1] ai.c and e.c.norton (2003) “Interaction Terms in Logit and Probit Models,� Economics Letters 80: 123-29. 28 [2] anderson.m.l (2008) “Multiple Inference and Gender Di¤erences in the E¤ects of Early Intervention: A Reevaluation of the Abecedarian, Perry Pre-school and Early Training Projects,� Journal of the American Statistical Association 103: 1481-95. [3] angrist.j.d (2001) “Estimation of Limited Dependent Variable Models with Dummy En- dogenous Regressors: Simple Strategies for Empirical Practice,� Journal of Business and Economic Statistics 19: 2-16. [4] arcand.j.l and e.d.waoube (2010) “Teaching Training and HIV/AIDS Prevention in West Africa: Regression Discontinuity Design Evidence from the Cameroon,� Health Economics 19: 36-54. [5] attanasio.o, a.kugler and c.meghir (2011) “Subsidizing Vocational Training for Dis- advantaged Youth in Developing Countries: Evidence from a Randomized Trial�, American Economic Journal: Applied Economics 3: 188-220. [6] baird.s, e.chirwa, c.t.mcintosh and b.ozler (2010) “The Short-Term Impacts of a Schooling Conditional Cash Transfer Program on the Sexual Behavior of Young Women�, Health Economics 19: 55-68. [7] baird.s, c.t.mcintosh and b.ozler (2010) Using Cash Transfers to Fight HIV Among Adolescent Girls: Explaining Causal Pathways With a Randomized Experiment, mimeo George Washington University. [8] baird.s, c.t.mcintosh and b.ozler (2011) “Cash or Condition: Evidence from the Cash Transfer Program�, Quarterly Journal of Economics 126: 1709-53. [9] baird.s, r.s.garfield, c.t.mcintosh and b.ozler (2012) “E¤ect of a Cash Transfer Programme for Schooling on Prevalence of HIV and Herpes Simplex Type 2 in Malawi: A Cluster Randomized Trial�, Lancet 379: 1320-9. [10] blattman.c, n.fiala and s.martinez (2011) Employment Generation in Rural Africa: Mid-term Results from an Experimental Evaluation of the Youth Opportunities Program in Northern Uganda, mimeo, Yale University. [11] blundell.r, c.meghir, m.costa dias and j.van reenen (2004) “Evaluating the Em- ployment Impact of a Mandatory Job Search Program�, Journal of the European Economic Association 2: 569-606. [12] bruce.j and k.hallman (2008) “Reaching the Girls Left Behind�, Gender and Development 16: 227-45. [13] bruhn.m, d.karlan and a.schoar (2012) The Impact of Consulting Services on Small and Medium Enterprises: Evidence from a Randomized Trial in Mexico, mimeo Yale University. 29 [14] bruhn.m and d.mckenzie (2009) “In Pursuit of Balance: Randomization in Practice in Development Field Experiments,� American Economic Journal: Applied Economics 1: 200- 32. [15] card.d, p.ibarraran, f.regalia, d.rosas-shady and y.soares (2011) “The Labor Market Impacts of Youth Training in the Dominican Republic,� Journal of Labor Economics 29: 267-300. [16] clark.s (2004) “Early Marriage and HIV Risks in Sub-Saharan Africa,� Studies in Family Planning 35: 149-60. [17] cornish.f and c.campbell (2009) “The Social Conditions for Successful Peer Education: A Comparison of Two HIV Prevention Programs Run by Sex Workers in India and South Africa,� American Journal of Community Psychology 44: 123-35. [18] drexler.a, g.fischer and a.schoar (2010) Keeping It Simple: Financial Literacy and Rules of Thumb, CEPR Discussion Paper 7994. [19] duflo.e, p.dupas and m.kremer (2011) Education, HIV and Early Fertility: Experimental Evidence from Kenya, mimeo MIT. [20] dupas.p (2011) “Do Teenagers Respond to HIV Risk Information? Evidence for a Field Experiment in Kenya,� American Economic Journal: Applied Economics 3: 1-34. [21] fairlie.r, d.karlan and j.zinman (2012) Behind the GATE Experiment: Evidence on E¤ects of and Rationales for Subsidized Entrepreneurship Training, mimeo Yale. [22] field.e and a.ambrus (2008) “Early Marriage, Age of Menarche and Female Schooling Attainment in Bangladesh,� Journal of Political Economy 116: 881-930. [23] field.e, s.jayachandran and r.pande (2010) “Do Traditional Institutions Constrain Fe- male Entrepreneurship? A Field Experiment on Business Training in India,� American Eco- nomic Review Papers and Proceedings 100: 125-9. [24] friedman.w, m.kremer, e.miguel and r.thornton (2011) Education as Liberation?, NBER Working Paper 16939. [25] fuller.g (1995) “The Demographic Backdrop to Ethnic Con‡ict: A Geographic Overview�, in The Challenge of Ethnic Con‡ict to National and International Order in the 1990s, CIA: Washington CIA. [26] gallant.m and e.maticka-tyndale (2004) “School-Based HIV Prevention Programmes for African Youth,� Social Science and Medicine 58: 1337-51. 30 [27] glynn.j.r, m.carael, b.auvert, m.kahindo, j.chege, r.musonda, f.kaona and a.buve (2001) “Why Do Young Women Have a Much Higher Prevalence of HIV than Young Men?,� AIDS 15: S51-60. [28] goldin.c and l.f.katz (2002) “The Power of the Pill: Oral Contraceptives and Women’s Career and Marriage Decisions,� Journal of Political Economy 110: 730-70. [29] goldstone.j.a (2002) “Population and Security: How Demographic Change Can Lead to Violent Con‡ict,� Journal of International A¤air 56: 3-23. [30] groh.m, n.krishnan, d.mckenzie and t.vishwanath (2012) Soft Skills or Hard Cash? The Impact of Training and Wage Subsidies on Female Youth Employment in Jordan, mimeo, World Bank. [31] ilo (2006) Global Employment Trends for Youth 2006, Geneva: ILO. [32] imbens.g.w and j.m.wooldridge (2009) “Recent Developments in the Econometrics of Programme Evaluation,� Journal of Economic Literature 47: 5-86. [33] karlan.d and j.morduch and s.mullainathan (2010) Take-up: Why Micro…nance Take-up Rates Are Low and Why It Matters, Financial Access Initiative, Research Fram- ing Note. [34] karlan.d and m.valdivia (2011) “Teaching Entrepreneurship: Impact of Business Training on Micro…nance Clients and Institutions,� Review of Economics and Statistics 93: 510-27. [35] kasirye.i (2011) Addressing Gender Gaps in the Ugandan Labour Market, Economic Policy Research Centre Policy Brief 12. [36] kelly.j.a, a.m.somlai, e.g.benotsch, y.a.amirkhanian, m.i.fernandez, l.y.stevenson et al. (2006) “Programmes, Resources and Needs of HIV-Prevention Nongovernmental Organizations (NGOs) in Africa, CEntral/Eastern Europe and Central Asia, Latin America and the Carribean,� AIDS Care 18: 12-21. [37] kling.j.r, j.b.liebman and l.f.katz (2007) “Experimental Analysis of Neighborhood Ef- fects,� Econometrica 75: 83-119. [38] jensen.r (2012) “Do Labor Market Opportunities A¤ect Young Women’s Work and Family Decisions? Experimental Evidence from India,� Quarterly Journal of Economics 127: 753-92. [39] jensen.r and r.thornton (2003)“Early Female Marriage in the Developing World,� Gen- der and Development 11: 9-19. [40] luke.n (2003) “Age and Economic Asymmetries in the Sexual Relationships of Adolescent Girls in Sub-Saharan Africa,� Studies in Family Planning 34: 67-86. 31 [41] mccoy.s.l, r.a.kangwende and n.c.padian (2010) “Behavior Change Interventions to Prevent HIV Infection Among Women Living in Low an dMiddle Income Countries: A Sys- tematic Review,� AIDS and Behavior 14: 469-82. [42] mensch.b.s, s.singh and j.b.casterline (2005) “Trends in the Timing of First Mar- riage Among Men and Women in the Developing World,� in C.B.Lloyd, J.R. Behrman, N.P.Stromquist and B.Cohen (eds.) The Changing Transitions to Adulthood in Developing Countries: Selected Studies, Washington DC: NAtional Academies Press, pp118-71. [43] ozier.o (2011) The Impact of Secondary Schooling in Kenya: A Regression Discontinuity Analysis, mimeo UC Berkeley. [44] robinson.j and e.yeh (2011) “Transactional Sex as a Response to Risk in Western Kenya,� American Economic Journal: Applied Economics 3: 35-64. [45] un desa population division (2010) World Population Prospects: The 2010 Revision Population Database. [46] unaids (2010) Global Report: UNAIDS Report on the Global AIDS Epidemic. [47] unicef (2003) Africa’s Orphaned Generations: A Framework to Support Orphaned Children in Sub-Saharan Africa, Geneva: UNICEF. [48] world bank (2007) World Development Report 2007: Development and the Next Genera- tion, Washingtn DC: The World Bank. [49] world bank (2009) Africa Development Indicators 2008/9: Youth and Employment in Africa, Washington DC: The World Bank. [50] westfall.p.h and s.s.young (1993) Resampling-Based Multiple Testing: Examples and Methods for p-value Adjustment, John Wiley and Sons: New York. [51] wooldrige.j.m (2002) Econometric Analysis of Cross Section and Panel Data, MIT Press: Cambridge MA. 32 Table 1: Descriptive Statistics on Adolescent Girls, By Treatment Status Means, standard errors in parentheses, standard deviations in brackets Normalized Treatment Control Difference Difference (1) (2) (3) (4) A. Demographics Age 16.3 16.4 -.094 -.023 [2.80] [2.96] (.168) Currently enrolled [yes=1] .713 .712 .0008 .001 [.452] [.453] (.028) Married [yes=1] .053 .073 -.020 -.058 [.224] [.260] (.014) Has child(ren) [yes=1] .105 .105 .0002 .0006 [.307] [.306] (.018) B. Income 69.6 71.6 -1.94 -.055 Generating Entrepreneurial ability [0-100 score] [24.6] [25.0] (1.78) Activities (IGA) Self-employment [yes=1] .070 .060 .010 .029 [.255] [.237] (.010) Wage employment [yes=1] .057 .036 .021** .071 [.233] [.186] (.010) Individual income from self- 19,665 13,099 6,567 .031 employment in the past year [in UGX] [185,536] [98,101] (4,785) Individual income from wage 22,768 10,197 12,572** .049 employment in the past year [in UGX] [245,364] [76,387] (5,360) Hours spent on self-employment on a .483 .426 .057 .020 typical day [2.06] [1.95] (.076) Hours spent on wage employment on a .704 .493 .211* .060 typical day [2.67] [2.24] (.117) C. Risky Behaviors HIV knowledge [0-6 score] 3.82 3.78 .047 .027 [1.24] [1.24] (.082) Pregnancy knowledge [yes=1] .739 .746 -.006 -.010 [.439] [.436] (.027) Has Child(ren) [yes=1] .105 .105 .0002 .0006 [.307] [.306] (.018) If sexually active, always uses condom .514 .514 .0004 .0006 [yes=1] [.500] [.500] (.042) .137 .141 -.004 -.009 Suffered from STD [yes=1] [.344] [.348] (.017) If suffered from STD, went to health .683 .674 .009 .013 center [yes=1] [.466] [.470] (.057) Had sex unwillingly in the past year .205 .142 .062*** .117 [yes=1] [.404] [.350] (.024) Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors on the differences are estimated from running the corresponding least squares regression allowing for the errors to be clustered by community. The normalized difference is computed following Imbens and Wooldridge (2009). The entrepreneurial ability index is the cumulative and rescaled score aggregating the self-assessed ranks to the following activities (where 10 was the highest rank and 1 the lowest): "Run your own business", "Identify business opportunities to start up new business", "Obtain credit to start up new business or expand existing business", "Save in order to invest in future business opportunities", "Make sure that your employees get the work done properly", "Manage financial accounts", "Bargain to obtain cheap prices when you are buying anything for business (inputs)", "Bargain to obtain high prices when you are selling anything for business (outputs)", "Protect your business assets from harm by others", "Collecting the money someone owes you". The top 1% outliers of the income variables have been removed. The HIV knowledge index is based on the number of statements correctly identified as true or false. The relevant statements are "A person who has HIV is different from a person who is ill with AIDS", "During vaginal sex, it is easier for a woman to receive the HIV virus than for a man", "Pulling out the penis before a man climaxes keeps a women from getting HIV during sex", "A women cannot get HIV if she has sex during her period", "Taking a test for HIV one week after having sex will tell a person if she or he has HIV." and "A Pregnant woman with HIV can give the virus to her unborn baby". The pregnancy knowledge indicator is defined equivalently on a single statement "A women cannot become pregnant at first intercourse or with occasional sexual relations". All monetary variables are deflated and expressed in terms of the price level in January 2008 using the monthly consumer price index published by the Uganda Bureau of Statistics Table 2: Descriptive Statistics on Adolescent Girls, By Treatment Status Means, standard errors in parentheses, standard deviations in brackets Normalized Treatment Control Difference Difference (1) (2) (3) (4) Have heard about club [yes=1] .589 .398 .193*** .275 [.492] [.490] (.030) Have ever participated in club activities, conditional .206 .047 .156*** .348 on having heard about club [yes=1] [.405] [.212] (.016) Continued participation, conditional on ever having .630 participated [yes=1] [.483] Attend(ed) club meetings at least 3 times a week, .273 conditional on ever having participated [yes=1] [.446] Attend(ed) club meetings 1 or 2 times a week, .494 conditional on ever having participated [yes=1] [.500] Received life skills training, conditional on ever .847 having participated [yes=1] [.360] Received livelihood skills training, conditional on ever .527 having participated [yes=1] [.500] Received life and livelihood skills training , .509 conditional on ever having participated [yes=1] [.500] Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors on the differences are estimated from running the corresponding least squares regression allowing for the errors to be clustered by community. The normalized difference is computed following Imbens and Wooldridge (2009). The indicators for having received life skills and/or livelihood skills are elicited from respondents' declarations to having participated in the corresponding training sessions at least very few times. Training area examples mentioned for the life skill training include learning about pregnancy or HIV. Training area examples mentioned for the livelihood training include training in hair-dressing, computer and poultry rearing. Table 3: Descriptive Statistics on Adolescent Girls, By Participation Status Means, standard errors in parentheses, standard deviations in brackets Normalized Participants Non Participants Difference Difference (1) (2) (3) (4) A. Demographics Age 16.2 16.4 -.150 -.038 [2.82] [2.80] (.133) Currently enrolled [yes=1] .716 .712 .004 .006 [.451] [.453] (.025) Married [yes=1] .039 .056 -.017* -.057 [.194] [.231] (.009) Has child(ren) [yes=1] .093 .108 -.015 -.036 [.290] [.311] (.014) B. Income 68.5 69.9 -1.42 -.041 Generating Entrepreneurial ability [0-100 score] [23.9] [24.7] (1.26) Activities (IGA) Self-employment [yes=1] .068 .070 -.002 -.006 [.252] [.256] (.011) Wage-employment [yes=1] .055 .058 -.003 -.010 [.228] [.234] (.009) Individual income from self- 14,614 20,978 -6,364 -.027 employment in the past year [in UGX] [126,634] [197,990] (6,818) Individual income from wage 34,112 19,821 14,290 .032 employment in the past year [in UGX] [399,440] [185,477] (15,537) Hours spent on self-employment on a .448 .492 -.044 -.015 typical day [1.94] [2.09] (.083) Hours spent on wage employment on .714 .702 .013 .003 a typical day [2.65] [2.67] (.114) C. Risky Behaviors HIV knowledge [0-6 score] 3.84 3.82 .022 .012 [1.31] [1.22] (.065) Pregnancy knowledge [yes=1] .754 .735 .019 .031 [.431] [.441] (.021) If sexually active, always uses condom .562 .502 .060* .085 [yes=1] [.497] [.500] (.035) .161 .130 .031* .062 Suffered from STD [yes=1] [.368] [.337] (.017) If suffered from STD, went to health .654 .691 -.037 -.056 center [yes=1] [.479] [.463] (.064) Had sex unwillingly in the past year .177 .212 -.036 -.063 [yes=1] [.382] [.409] (.030) Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors on the differences are estimated from running the corresponding least squares regression allowing for the errors to be clustered by community. The normalized difference is computed following Imbens and Wooldridge (2009). The entrepreneurial ability index is the cumulative and rescaled score aggregating the self-assessed ranks to the following activities (where 10 was the highest rank and 1 the lowest): "Run your own business", "Identify business opportunities to start up new business", "Obtain credit to start up new business or expand existing business", "Save in order to invest in future business opportunities", "Make sure that your employees get the work done properly", "Manage financial accounts", "Bargain to obtain cheap prices when you are buying anything for business (inputs)", "Bargain to obtain high prices when you are selling anything for business (outputs)", "Protect your business assets from harm by others", "Collecting the money someone owes you". The top 1% outliers of the income variables have been removed. The HIV knowledge index is based on the number of statements correctly identified as true or false. The relevant statements are "A person who has HIV is different from a person who is ill with AIDS", "During vaginal sex, it is easier for a woman to receive the HIV virus than for a man", "Pulling out the penis before a man climaxes keeps a women from getting HIV during sex", "A women cannot get HIV if she has sex during her period", "Taking a test for HIV one week after having sex will tell a person if she or he has HIV." and "A Pregnant woman with HIV can give the virus to her unborn baby". The pregnancy knowledge indicator is defined equivalently on a single statement "A women cannot become pregnant at first intercourse or with occasional sexual relations". Table 4: The Impact of the ELA Programme on Income Generating Activities Coefficients, standard errors in parentheses, standard deviations in brackets in Column on Baseline Levels ITT Estimates ATE Estimates (6) IV, Club (7) IV, Livelihood Baseline (3) Probit, (4) Tobit (5) Tobit Outcome N (1) OLS, Level (2) OLS, FD Participation, FD Training, FD Levels Marginal Effects [Pr(y*>0)] [E[y*|y*>0] Treatment as IV Treatment as IV 69.7 4,765 7.77** 7.22*** 45.6*** 52.7*** Entrepreneurial ability [0-100 score] [24.5] (3.17) (1.56) (11.4) (13.2) Engaged in any IGA [yes=1] .124 4,831 .040* .043** .278** .321** [.330] (.021) (.020) (.123) (.143) Self-employment [yes=1] .070 4,831 .050*** .053*** .051*** .336*** .389*** [.255] (.015) (.015) (.014) (.093) (.109) Wage employment [yes=1] .057 4,831 -.013 -.011 -.010 -.073 -.084 [.231] (.013) (.010) (.010) (.068) (.079) Individual total income past year from self- 19,312 4,826 -503 -2,915 .042*** 54,503*** -18,697 -21,604 employment [in UGX] [185,563] (11,620) (11,128) (.011) (17,838) (71,793) (82,927) Individual total income past year from wage 22,376 4,828 -8,837 -7,055 -.004 -18,341 -45,156 -52,165 employment [in UGX] [244,837] (8,896) (7,528) (.007) (23,021) (48,338) (56,045) .478 4,824 .196 .185 .042*** .460** 1.18 1.37 Hours spent on self-employment on a typical day [2.04] (.121) (.117) (.014) (.202) (.740) (.859) .703 4,824 -.192 -.131 -.016 -.385 -.836 -.966 Hours spent on wage employment on a typical day [2.66] (.137) (.111) (.011) (.299) (.720) (.834) 12,327 4,752 4,238*** 3,741*** .078*** 3,155*** 23,980*** 27,653*** Expenditure on goods in the last month [in UGX] [18,904] (1,515) (1,195) (.026) (895) (7,537) (8,756) Currently enrolled [yes=1] .716 4,831 -.013 -.006 -.014 -.039 -.046 [.451] (.025) (.021) (.029) (.135) (.156) If enrolled, hours spent on going to and attending 62.3 2,423 .855 -.170 -1.20 -1.45 school, homework and study per week [18.1] (2.65) (1.39) (9.75) (11.8) If dropped out, plan to start/go back to school .515 739 .081 .111** .083 .704** .807** [yes=1] [.500] (.051) (.047) (.052) (.332) (.380) Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors are clustered by community. The control variables include the adolescent age and a series of indicator variables for branch areas. The entrepreneurial ability index is the cumulative and rescaled score aggregating the self-assessed ranks to the following activities (where 10 was the highest rank and 1 the lowest): "Run your own business", "Identify business opportunities to start up new business", "Obtain credit to start up new business or expand existing business", "Save in order to invest in future business opportunities", "Make sure that your employees get the work done properly", "Manage financial accounts", "Bargain to obtain cheap prices when you are buying anything for business (inputs)", "Bargain to obtain high prices when you are selling anything for business (outputs)", "Protect your business assets from harm by others", "Collecting the money someone owes you". The top 1% outliers of the income variables have been removed. For the expenditure variable, the goods categories are jewelry/ornaments, cosmetics/makeup, clothes, hairdressers, shoes/footwear, going to restaurants/bars/teashop/cafe, talk time for your mobile phone and presents/gifts. All monetary variables are deflated and expressed in terms of the price level in January 2008 using the monthly consumer price index published by the Uganda Bureau of Statistics. In Column 6, the ATE estimates are based on club participation being instrumented by the offer of treatment. In Column 7 the ATE estimates are based on participation in at least one livelihood skills courses being instrumented by the offer of treatment. Table 5: The Impact of the ELA Programme on Risky Behaviors Coefficients, standard errors in parentheses, standard deviations in brackets in Column on Baseline Levels ITT Estimates ATE Estimates (4) IV, Club (5) IV, Life Skill Baseline (3) Probit, Marginal Outcome N (1) OLS, Level (2) OLS, FD Participation, FD Training, FD Levels Effects Treatment as IV Treatment as IV HIV knowledge [0-6 score] 3.83 4,831 .462*** .418*** 2.67*** 4.88*** [1.23] (.109) (.069) (.504) (.910) Pregnancy knowledge [yes=1] .739 4,550 .067** .061** .071** .381** .680** [.439] (.030) (.028) (.032) (.176) (.318) Has child(ren) [yes=1] .104 4,670 -.027** -.028** -.024** -.179** -.325** [.305] (.013) (.012) (.011) (.079) (.142) If sexually active, always uses .457 816 .125** .144*** .135** .747*** 1.64** condom [yes=1] [.499] (.055) (.050) (.059) (.280) (.682) If sexually active, uses often or .262 816 -.112** -.124** -.117** -.646** -1.42** sometimes condom [yes=1] [.440] (.053) (.049) (.054) (.272) (.654) If sexually active, uses other .183 816 .054 .042 .057 .217 .477 contraceptives [yes=1] [.387] (.060) (.041) (.059) (.214) (.459) Suffered from STD [yes=1] .140 3,870 -.003 -.007 -.004 -.041 -.075 [.347] (.026) (.022) (.026) (.128) (.233) If suffered from STD, went to .714 120 .044 .107 .026 .537 1.00 health center [yes=1] [.454] (.113) (.102) (.121) (.515) (.973) Had sex unwillingly [yes=1] .212 922 -.161*** -.160*** -.160*** -.648*** -1.31*** [.409] (.041) (.039) (.040) (.182) (.405) Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors are clustered by community. The control variables include the adolescent age and a series of indicator variables for branch areas. The HIV knowledge index is based on the number of statements correctly identified as true or false. The relevant statements are "A person who has HIV is different from a person who is ill with AIDS", "During vaginal sex, it is easier for a woman to receive the HIV virus than for a man", "Pulling out the penis before a man climaxes keeps a women from getting HIV during sex", "A women cannot get HIV if she has sex during her period", "Taking a test for HIV one week after having sex will tell a person if she or he has HIV." and "A Pregnant woman with HIV can give the virus to her unborn baby". The pregnancy knowledge indicator is defined equivalently on a single statement "A women cannot become pregnant at first intercourse or with occasional sexual relations". In Column 4, the ATE estimates are based on club participation being instrumented by the offer of treatment. In Column 5 the ATE estimates are based on participation in the life skills courses being instrumented by the offer of treatment. Table 6: The Impact of the ELA Programme on Other Measures Coefficients, standard errors in parentheses, standard deviations in brackets in Column on Baseline Levels ITT Estimates ATE Estimates Baseline (3) Probit, (4) IV, Club Participation, FD N (1) OLS, Level (2) OLS, FD Levels Marginal Effects Treatment as IV Gender empowerment index [0,100] 28.6 4,831 5.76** 6.23*** 39.9*** [24.2] (2.51) (1.53) (10.7) Attitude Index [0,100] 75.6 4,727 1.86 1.70 10.7 [19.4] (2.51) (1.49) (9.70) Satisfaction index [0,100] 45.1 4,766 .127 -.035 -.223 [17.9] (1.31) (.916) (5.82) 1.20 1,191 .570* .499*** 3.70** Satisfaction with earnings/income [0,6] [1.68] (.330) (.189) (1.44) Never worry to get a good job in .418 4,637 .067** .062** .068** .384** adulthood [yes=1] [.493] (.033) (.027) (.035) (.170) Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors are clustered by community. The control variables include the adolescent age and a series of indicator variables for branch areas. The gender empowerment index is a variable that cumulates the number of times a respondent asswers "Both/Same" to the following questions: "Who should earn money for the family?", "Who should have a higher level of education in the family?", "Who should be responsible for washing, cleaning and cooking?", "If there is no water pump or tap, who should fetch water?", "Who should be responsible for feeding and bathing children?", "Who should help the children in their studies at home?" and "Who should be responsible for looking after the ill persons?" The other possible answers given to the respondent were "Male" and "Female". The index is then rescaled such that 100 indicates that the respondent answered that the female should be (at least partly) be responsible for all the activities. For the attitudes index, the adolescents were asked to rate how much they think that the following statements are true or false for them. On a scale of 1-10, where 1 is "not at all" and 10 is "a lot" (clockwise, begining with the spoke on top): "If I start working on a task, I definitely see the end of it no matter how difficult it is", "While doing any task, it is important for me to do it better than others", "If I have the chance, I would make a good leader", "I save regularly", "I do not care what others think about my success or failure", "I am in control of what happens in my life", "I want to be a respectful person in my village", "A person can get rich by taking risks", "I often make plans for the future", and "I believe that my future is determined by luck no matter how hard I work." The index is then rescaled to run from 0 to 100. The satisfaction index adds up the responses about the happiness of the respondent about different aspects of life, namely: "Your education level", "Your family", "Your friends", "Your job", "Your earnings/income", "The house you live in", "(If currently studying) The school you go to", "(If currently studying) Your school work", "Life as a whole". Originally, the respondents were asked to indicate their satisfaction on a scale where 1 stands for completely happy and 7 for not at all happy. For our analysis the original coding was reversed and rescaled so that a cumulative score of 100 stands for completely happy in all aspects and 0 for not at all happy in all aspects. Analogously, the variable capturing satisfaction with earnings/income is reversed such that 6 stands for completely happy and 0 for not happy at all. Table 7: The Impact of the ELA Programme on Aspirations Related to Marriage, Childbearing and Childrens' Lives Coefficients, standard errors in parentheses, standard deviations in brackets in Column on Baseline Levels ITT Estimates ATE Estimates Baseline (3) IV, Club Participation, FD N (1) OLS, Level (2) OLS, FD Levels Treatment as IV A. Marriage Own marriage: Anticipated age at marriage 25.1 3,024 .943*** (measured at follow-up only) [2.96] (.157) Suitable age for marriage for a female 24.0 4,730 .733** .653*** 4.17*** [3.09] (.287) (.182) (1.25) Suitable age for marriage for a male 27.9 4,705 .910*** .821*** 5.32*** [3.71] (.306) (.191) (1.36) B. Child Bearing Suitable age for women to have the first 24.0 4,701 .443 .437** 2.77** baby [3.16] (.305) (.173) (1.14) Preferred number of children 4.22 4,700 -.368*** -.382*** -2.43*** [1.55] (.119) (.080) (.573) C. Childrens' Lives Preferred age at which daughter(s) get 25.0 4,603 .587*** .511*** 3.26*** married [2.82] (.207) (.167) (1.08) 28.5 4,504 .034 -.039 -.249 Preferred age at which son(s) get married [3.24] (.228) (.156) (.991) Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors are clustered by community. The control variables include the adolescent age and a series of indicator variables for branch areas. All variables indicating ages were trimmed at 15 years or younger. Indicated ages of 51 years or older are removed from the variable "Anticipated age at marriage". The first outcome variable of Panel A refers to the adolescent's anticipation concerning her own marriage. The remaining outcome variables of Panel A, Panel B and the first two outcome variables of Panel C refer to the adolescent's preferences in the domains of marriage, child bearing and her children's lives. Table 8: Cost-Benefit Analysis, in US$ (1) Non- (2) Recurring (3) Year (4) Year Recurring Monthly One Two Fixed (1) Office Space & Equipment 10 Branch Offices 4,000 4,000 Costs (2) Programme Assistant Training 10 Assistants 2,250 2,250 Training & Operational Material (3) 2 Manuals 4,000 4,000 Development Programme Management (4) 2 Coordinators 780 9,360 9,360 Compensation (5) Programme Assistant Compensation 10 Assistants 1,690 20,280 20,280 (6) Adolescent Leader Compensation 100 Adolescent Leaders 1,200 14,400 14,400 (7) Adolescent Leader Training 100 Adolescent Leaders 22,500 22,500 Adolescent Leader Training (for (8) 20 Adolescent Leaders 4,500 4,500 Replacements) (9) Adolescent Leader Refreshers 100 Adolescent Leaders 400 4,800 4,800 (10) Club Rent 100 Clubs 1,000 12,000 12,000 (11) Club Materials 100 Clubs 42,000 42,000 (12) Club Materials (Replenishment) 100 Clubs 16,800 16,800 (13) Branch Office Overhead 10 Branch Offices 800 9,600 9,600 (14) Country Office Overhead 1 Country Office 4,000 48,000 48,000 Variable (15) Financial Literacy Courses 2,500 Members 12,500 12,500 12,500 Costs (16) Livelihood Training (Year 1) 2,000 Members 100,000 100,000 (17) Livelihood Training Inputs (Year 1) 2,000 Members 60,000 60,000 (18) Livelihood Training (Year 2) 1,000 Members 50,000 50,000 (19) Livelihood Training Inputs (Year 2) 1,000 Members 30,000 30,000 Total (20) ELA Programme Costs for the 100 studied Communities 365,690 232,240 Costs Yearly Assuming 130 potential girl (21) Fixed Costs 14.9 10.7 Per Unit attendees per club Average (22) Variable Costs 13.3 7.12 Costs (23) Total Costs 28.1 17.9 Yearly ITT Impact of ELA on Individual Total Income in the Past Year from Self-Employment (24) 32.1 Benefits (Tobit, E[y*|y*>0]) Notes: *** indicates that the coefficient/marginal effect on which the estimate is based is significant at 1%, ** at 5%, and * at 10%. The exchange rate used to convert monetary values is based on January 2008 at which point $1 was worth approximately UGX1,700. The yearly costs shown in Columns 3 and 4 are obtained by multiplying column 2 times 12 (months) and adding Column 1 for all fixed and variable cost categories applicable to the respective year of operation. The yearly total cost of the ELA Programme stated in row 20 is the summation of all individual cost items applicable to the respective year. The yearly benefits shown in row 24 to 25 are based on the ITT impact estimates shown before. Table A1: Correlates of Two-Year Attrition Dependent Variable: In Panel Sample [yes=1] OLS estimates in Columns 1-4, Probit estimate in Column 5 Standard errors clustered by community (1) (2) (3) (4) (5) Probit Treatment .025 .026 .026 -.019 -.042 (.029) (.027) (.027) (.100) (.093) Age -.0002 -.0009 -.002 (.002) (.004) (.004) Currently enrolled [yes=1] .015 -.013 -.028 (.015) (.029) (.030) Married [yes=1] -.0002 .059 .056 (.023) (.037) (.037) Has child(ren) [yes=1] .025 -.005 -.014 (.020) (.033) (.035) Treatment x age .0010 .002 (.005) (.005) Treatment x currently .041 .046 enrolled [yes=1] (.038) (.039) -.097** -.110* Treatment x married [yes=1] (.048) (.068) Treatment x has child(ren) .045 .036 [yes=1] (.043) (.039) Branch Dummies No Yes Yes Yes No Observations 5,661 5,661 5,661 5,661 5,661 Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The dependent variable is a dummy equal to one if the adolescent girl attrits between the baseline survey and the two-year follow-up survey. The standard errors are clustered by community. Columns 1 to 4 are estimated using a linear probability model. Column 5 is estimated using a probit model, where marginal effects are reported. There are ten branch dummies controlled for in Columns 2 to 4. Table A2: Robustness Checks Marginal effects, standard errors in parentheses, p-values Anticipation of P-Value Correction for Multiple Inference Sample Attrition Microfinance (5) Weighted ITT (6) ITT Estimate on (1) Baseline ITT (4) Summary (2) Original (3) Step-Down Estimates Adjusting Future Assignment to Estimates Index for Attrition Microfinance A. Income Entrepreneurial ability [0-100 7.77** .015 .082 8.42*** -1.98 Generating score] (3.17) (3.06) (3.88) Activities (IGA) Self-employment [yes=1] .051*** .0003 .006 .052*** -.016 (.014) (.015) (.020) Wage employment [yes=1] -.010 .335 .362 -.009 -.004 (.010) (.011) (.014) Individual total income past year 54,503*** .002 .043 60,486*** 1,792 from self-employment [in UGX] (17,838) (18,644) (20,472) Individual total income past year -18,341 .426 .544 -20,021 3,048 from wage employment [in UGX] (23,021) (23,982) (28,277) Hours spent on self-employment .460** .023 .180 .453** -.159 on a typical day (.202) (.216) (.240) Hours spent on wage employment -.385 .197 .363 -.378 -.133 on a typical day (.299) (.316) (.349) .150*** .005 .162*** -.051 Summary Index (IGA) (.052) (.053) (.068) B. Risky HIV knowledge [scale 0,6] .462*** .00004 .0005 .470*** .107 Behaviors (.109) (.111) (.134) Pregnancy knowledge [yes=1] .071** .025 .150 .057* .011 (.032) (.032) (.038) Has child(ren) [yes=1] -.024** .031 .017 -.025** .009 (.011) (.011) (.015) If sexually active, always uses .135** .021 .00000 .126** -.024 condom [yes=1] (.059) (.063) (.075) Suffered from STD [yes=1] -.004 .870 .956 -.006 .028 (.026) (.028) (.031) If suffered from STD, went to .026 .830 .145 .143 -.229* health center [yes=1] (.121) (.139) (.136) -.160*** .00006 .0005 -.160*** -.023 Had sex unwillingly [yes=1] (.040) (.044) (.049) .252*** .00000 .247*** -.004 Summary Index (Risky Behaviors) (.023) (.024) (.028) Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors are clustered by community. The control variables include the age of the respondents as well as dummies for being enrolled, having at least one child and being married. The entrepreneurial ability is an index consisting of cumulative ranks (scale from one to ten with ten being the highest) of the following activities: "Run your own business", "Identify business opportunities to start up new business", "Obtain credit to start up new business or expand existing business", "Save in order to invest in future business opportunities", "Make sure that your employees get the work done properly", "Manage financial accounts", "Bargain to obtain cheap prices when you are buying anything for business (inputs)", "Bargain to obtain high prices when you are selling anything for business (outputs)", "Protect your business assets from harm by others", "Collecting the money someone owes you". The top 1% outliers of the income variables have been removed. The HIV knowledge index is based on the number of statements correctly identified as true or false. The relevant statements are "A person who has HIV is different from a person who is ill with AIDS", "During vaginal sex, it is easier for a woman to receive the HIV virus than for a man", "Pulling out the penis before a man climaxes keeps a women from getting HIV during sex", "A women cannot get HIV if she has sex during her period", "Taking a test for HIV one week after having sex will tell a person if she or he has HIV." and "A Pregnant woman with HIV can give the virus to her unborn baby". The pregnancy knowledge indicator is defined equivalently on a single statement "A women cannot become pregnant at first intercourse or with occasional sexual relations". The summary index (life skills) is based on all outcome variables in the upper panel. The summary index (IGA) is based on the entrepreneurship ability and the indicators for self-employment and wage employment. All monetary variables are deflated and expressed in terms of the price level in January 2008 using the monthly consumer price index published by the Uganda Bureau of Statistics. Table A3: Impact Heterogeneity Marginal effects, standard errors in parentheses, p-values (4) Above (5) Below (1) Baseline ITT (6) Younger (7) 16yrs or (2) Rural (3) Urban Median HH Median HH Estimates than 16yrs older Asset Value Asset Value A. Income Generating .150*** .139** .152** .126** .170*** .150** .149** Summary Index (IGA) Activities (IGA) (.052) (.060) (.066) (.063) (.062) (.059) (.063) B. Risky Behaviors .252*** .221*** .285*** .283*** .219*** .280*** .230*** Summary Index (Risky Behaviors) (.023) (.025) (.036) (.031) (.025) (.031) (.027) C. Education Currently enrolled [yes=1] -.014 -.021 -.010 -.015 -.009 -.0002 -.028 (.029) (.038) (.041) (.030) (.044) (.019) (.045) If enrolled, hours spent on going to and attending .855 -2.15 3.72 1.62 .228 -.894 3.78 school, homework and study per week (2.65) (4.73) (2.49) (3.11) (3.13) (2.82) (3.07) If dropped out, plan to start/go back to school .083 .065 .107 -.009 .136** .255 .073 [yes=1] (.052) (.066) (.084) (.098) (.060) (.222) (.055) Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The standard errors are clustered by community. The control variables include the age of the respondents as well as dummies for being enrolled, having at least one child and being married. The entrepreneurial ability is an index consisting of cumulative ranks (scale from one to ten with ten being the highest) of the following activities: "Run your own business", "Identify business opportunities to start up new business", "Obtain credit to start up new business or expand existing business", "Save in order to invest in future business opportunities", "Make sure that your employees get the work done properly", "Manage financial accounts", "Bargain to obtain cheap prices when you are buying anything for business (inputs)", "Bargain to obtain high prices when you are selling anything for business (outputs)", "Protect your business assets from harm by others", "Collecting the money someone owes you". The top 1% outliers of the income variables have been removed. The HIV knowledge index is based on the number of statements correctly identified as true or false. The relevant statements are "A person who has HIV is different from a person who is ill with AIDS", "During vaginal sex, it is easier for a woman to receive the HIV virus than for a man", "Pulling out the penis before a man climaxes keeps a women from getting HIV during sex", "A women cannot get HIV if she has sex during her period", "Taking a test for HIV one week after having sex will tell a person if she or he has HIV." and "A Pregnant woman with HIV can give the virus to her unborn baby". The pregnancy knowledge indicator is defined equivalently on a single statement "A women cannot become pregnant at first intercourse or with occasional sexual relations". The summary index (life skills) is based on all outcome variables in the upper panel. The summary index (IGA) is based on the entrepreneurship ability and the indicators for self-employment and wage employment. All monetary variables are deflated and expressed in terms of the price level in January 2008 using the monthly consumer price index published by the Uganda Bureau of Statistics. The rural sample comprises respondents from BRAC's Njeru, Iganga, Busia, Bugembe and Buwenge areas. Protein intake is derived from the indicated household consumption of eggs, fish and meat. Figure 1B: Unemployment Rates (%), by Age and Gender, Uganda Figure 1A: Female Population by Age, 2010 2005/6 Notes: The data stems from the 2010 UN World Population Prospects data base. More developed Notes: Unemployment is defined as those who actively wanted a job but did not participate in any regions comprise Europe, Northern America, Australia/New Zealand and Japan. employment activities, inclusively self-employment and agricultural works). Figure 1C: Age-Specific Fertility Rate, 1995-2010 Notes: The data stems from the 2010 UN World Population Prospects data base. The fertility rate is measured by the number of births per 1,000 women. More developed regions comprise Europe, Northern America, Australia/New Zealand and Japan. Figure 2A: The Impact (ITT) of the ELA Programme on Figure 2B: The Impact (ITT) of the ELA Programme on Entrepreneurship Measures Gender Empowerment Measures Notes: The control variables include the adolescent age and a series of indicator variables for Notes: The control variables include the adolescent age and a series of indicator variables for branch areas. The adolescents were asked to rank their ability on how well they can do the branch areas. The adolescents were asked to indicate whether the following taks should be following activities on a scale of 1 to 10, 1 means they cannot do this activity and 10 is they performed by the "Male", "Female" or “Both/Same�: “Who should earn money for the family?�, “Who definitely can (clockwise, begining with the spoke on top): "Run your own business", "Identify should have a higher level of education in the family?�, “Who should be responsible for washing, business opportunities to start up new business", "Obtain credit to start up new business or expand cleaning and cooking?�, “If there is no water pump or tap, who should fetch water?�, “Who should existing business", "Save in order to invest in future business opportunities", "Make sure that your be responsible for feeding and bathing children?�, “Who should help the children in their studies at employees get the work done properly", "Manage financial accounts", "Bargain to obtain cheap home?� and “Who should be responsible for looking after the ill persons?�. The outcome variable prices when you are buying anything for business (inputs)", "Bargain to obtain high prices when you equals 1 if the adolescent indicated that the respective task should be performed by"Both/Same" are selling anything for business (outputs)", and "Protect your business assets from harm by and 0 otherwise. othersCollecting the money someone owes you". Figure 2C: The Impact (ITT) of the ELA Programme on Figure 2D: The Impact (ITT) of the ELA Programme on Satisfaction Attitude Measures Measures Notes: The control variables include the adolescent age and a series of indicator variables for branch areas. The adolescents were asked t rate how much they think that the following statements Notes: The control variables include the adolescent age and a series of indicator variables for are true or false for them. On a scale of 1-10, where 1 is "not at all" and 10 is "a lot" (clockwise, branch areas. The adolescents were asked to indicate their satisfaction on a 7 point scale, where 1 begining with the spoke on top): "If I start working on a task, I definitely see the end of it no matter is completely happy and 7 is not at all happy. For this exposition, the ranking has been reversed. how difficult it is", "While doing any task, it is important for me to do it better than others", "If I have Clockwise, begining with the spoke at the top, the satisfaction domains are: "Your education level", the chance, I would make a good leader", "I save regularly", "I do not care what others think about "Your family", "Your friends", "Your job", "Your earnings / income", "The house you live in", "(If my success or failure", "I am in control of what happens in my life", "I want to be a respectful person currently studying) The school you go to", "(If currently studying) Your school work", and "Life as a in my village", "A person can get rich by taking risks", "I often make plans for the future", and "I whole". believe that my future is determined by luck no matter how hard I work."