Policy Research Working Paper 10563 Savings Facilitation or Capital Injection? Impacts and Spillovers of Livelihood Interventions in Post-Conflict Côte d’Ivoire Alicia Marguerie Patrick Premand Development Economics Development Impact Evaluation Group September 2023 Policy Research Working Paper 10563 Abstract Policy makers grapple with the optimal design of mul- conflict. The interventions had differential effects on the tidimensional strategies to improve poor households’ dynamics of savings and productive asset accumulation. livelihoods. To address financial constraints, are capital The cash grant modalities generated investments in startup injections needed, or is savings mobilization sufficient? This capital, although nearly 30 percent of the grant was saved. paper tests the direct effects and local spillovers of three In contrast, village savings and loan associations did not instruments to relax financial constraints, each combined increase total savings but gradually induced investments, so with micro-entrepreneurship training. “Cash grants” and that productive assets caught up with cash grant recipients “cash grants with repayment” directly inject capital, while after 15 months. Positive local spillovers on savings and “village savings and loan associations” (VSLAs) promote independent activities were also observed. Yet, investments more efficient group saving. The randomized controlled in independent activities were not sufficient to increase trial took place in western regions of Côte d’Ivoire that profits, possibly because they were limited due to high pre- were affected by a post-electoral crisis in 2011 and an earlier cautionary saving motives in the post-conflict study setting. This paper is a product of the Development Impact Evaluation Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at at amarguerie@worldbank.org and ppremand@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Savings Facilitation or Capital Injection? Impacts and Spillovers of Livelihood Interventions ote d’Ivoire in Post-Conflict Cˆ Alicia Marguerie (World Bank) Patrick Premand (DIME, World Bank) JEL classification : D13, D14, O12, O17, I38. Keywords : Cash grant, Savings, Livelihoods, Poverty, Graduation, Economic Inclusion, Field Experiment. Acknowledgments : This paper is the result of a close collaboration with World Bank operational teams, the government of Cˆote d’Ivoire through BCPE, and IRC. We are very grateful to the World Bank team leaders (Steffen Janus and H. A. Wedoud Kamil), BCPE (Hermann Toualy, Adama Bamba, C´ esar G. Toassa, Boubacar Ndiaye, Ismahel A. Barry, Latif Doho, and Germain Kouadio) and IRC (K. Drissa Akou, Hubert K. Kassi, Akouli C. Seri, Mariame Kone, Louis Falcy, Cl´ ement Lorvao and Amaia Bessouet) for their many contributions throughout the program and RCT design and implementation. Horacio Vera Cossio, Eva Lestant and Daniel Corredor Vallejo provided excellent research assistance. We thank Lori Beaman, Christopher Blattman, Bruno Cr´ epon, Rema Hanna, William Parient´ e, Lore Vandewalle, a DIME reviewer, and seminar/conference participants at CREST, CSAE, IZA-WB, IPA, and Development Economics Network Switzerland for their helpful comments. The study was pre-registered in the AEA RCT registry (AEARCTR-0002737) and received IRB 2018-008 from J-PAL Europe. Financial support from the Knowledge for Change Program (KCP) at the World Bank is gratefully acknowledged. DIME analytics verified the computational reproducibility of the results. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank and its affiliated organizations or those of the Executive Directors of the World Bank or the government they represent. 2 1 Introduction Policy makers are looking for a combination of policies to improve the livelihoods, mi- cro and small businesses that prevail in high-poverty settings (Filmer and Fox (2014), Jayachandran (2020)). Integrated interventions such as “graduation” or “economic inclu- sion” programs that provide capital, skill training, consumption support, coaching, and savings facilitation have had robust results.2 Yet there are ongoing debates on how to optimize multi-faceted policies (e.g. Andrews et al. (2021)). In-kind assets, cash grants or microcredit are commonly used to inject capital, but evidence is lacking on which instrument is most effective to address capital constraints. Many multi-faceted programs also facilitate savings accumulation, through various modalities such as savings groups or formal savings accounts. Savings mobilization provides an alternative mechanism to facilitate productive investments, possibly reducing the need for capital injections in the first place. This multiplicity of instruments raises questions about the optimal mecha- nism to address financial constraints, which in turn depends on the nature of the capital or savings constraints that are binding for the poor.3 In this paper, we report results from an RCT testing how three alternative instruments to alleviate financial constraints, combined with training, increase poor households’ invest- ments, improve income-generating activities and influence social outcomes. Specifically, we document the relative effects of relaxing savings constraints through enhanced sav- ings groups (Village Savings and Loans Associations, “VSLAs”) designed to facilitate savings accumulation and provide small credit opportunities without any capital injec- tion, or relaxing capital constraints through a cash grant of 95,000 CFA (USD 384 PPP) (“cash grants”) or the same grant with a 50 percent repayment condition (akin to sub- sidized credit, “cash grants with repayment”).4 These three financial support modalities are combined with a micro-entrepreneurship training on how to develop simple business plans, assess business opportunities and define capital needs. Of note, participants do not 2 E.g. Banerjee et al. (2015b), Bandiera et al. (2017), Bedoya et al. (2019), Bossuroy et al. (2022). Some micro-entrepreneurship interventions have also successfully combined cash or in-kind grants with technical or business training (e.g. Cho and Honorati (2014) and Blattman and Ralston (2017)). 3 Micro-credit has resulted in limited impacts on micro-enterprise growth (Banerjee et al. (2015a), Cr´epon et al. (2015)), especially for households at the bottom of the distribution (Meager (2019)), which has also raised questions on how to optimize financial support for the poorest. 4 Official IMF average exchange rate for 2017: USD 1=CFA 580.657 (IMF)). World Bank PPP con- version factor for Cˆote d’Ivoire in 2017, GDP (LCU per international $): 247.134. 1 receive regular transfers for consumption support, which contrasts with most graduation programs studied in the literature. ote d’Ivoire, The RCT took place in 147 localities in the western regions of post-conflict Cˆ covering both urban and rural areas. Localities were randomly assigned to one of the three interventions (“VSLAs”, “cash grants with repayment” or “cash grants”) or a con- trol group. Within localities, program participants were selected based on a baseline survey designed to calculate a vulnerability proxy measure. We collected follow-up data approximately 22 months after the program started, or 15 months after the injection of cash grants or redistribution of savings. In both treatment and control localities, we surveyed individuals with a baseline proxy score below the program selection cut-off (potentially selected individuals), but also above the cut-off (non-selected individuals). We thus measure direct impacts (for beneficiaries) as well as local spillovers (for non- selected individuals) on economic and social outcomes. The analysis of local spillovers is particularly relevant in a post-conflict setting given concerns over lingering tensions. Our primary contribution is to compare cash grants with savings groups to assess whether an external capital injection is necessary to generate investments in income-generating activities, or whether savings mobilization alone can trigger investments by increasing savings efficiency. Our second contribution is to compare two different types of capital injection modalities, cash grants with or without repayment, to test whether adding a repayment condition alters investment behaviors relative to unconditional capital. Poor households can struggle to access capital for income-generating activities, for exam- ple facing imperfect capital markets, unattractive credit terms or collateral conditions. In credit scarce environments, cash grants can directly address capital constraints and high returns have been documented for micro-entrepreneurs (e.g., de Mel et al. (2008), De Mel et al. (2012) and Blattman et al. (2014)). However, compared with asset transfers, cash is more fungible, raising questions on how much will be used for productive investments. While there is evidence that poor households consume part of grants (e.g. Berge et al. (2015), Haushofer and Shapiro (2016)), unconditional cash grants have shown positive effects on economic diversification and earnings (Blattman et al. (2014), Macours et al. (2022)).5 Cash grants with repayment conditions are akin to highly subsidized credit 5 There is a large literature on the optimal design of small, regular cash transfers, but we focus here on one-off, lumpy cash grants for investments. 2 and can also directly alleviate capital constraints.6 The effect of repayment conditions is theoretically unclear. The reimbursement requirement may introduce an incentive (or possibly a behavioral nudge) to invest, similar to the idea that it is more tempting to con- sume cash in-hand compared to in-kind transfers (Fiala (2018)). Fafchamps et al. (2014) call this a “flypaper” effect, and show that in-kind transfers lead to larger business in- vestments than cash grants. However, repayment conditions may hinder riskier but more profitable investments: Field et al. (2013) show that relaxing repayment requirement and extending grace periods in microcredit contracts increase profits by providing longer-term horizons for riskier investments. To shed light on the net effects between these potential mechanisms, we directly compare cash grants with and without repayment conditions. Addressing savings constraints can also facilitate investments in income-generating ac- tivities, though it is unclear if it is sufficient for lumpy investments. Suboptimal savings can be due to market failures or behavioral constraints (Karlan et al. (2014)). The intro- duction of Village Savings and Loans Associations (VLSAs) offers a commitment device as participants meet with peers (every 1-2 weeks) to make small savings contributions. Evidence on the impact of relaxing savings constraints is mixed. Access to formal savings accounts has increased investments and revenues (Brune et al. (2016), Dupas and Robin- son (2013)). Yet the literature on saving groups rarely reports improvements in business outcomes (Beaman et al. (2014), Ksoll et al. (2016)), with the notable exception of Karlan et al. (2017).7 We are not aware of a direct comparison of instruments to address capital constraints (cash grants) and savings constraints (VSLAs).8 Such a comparison can shed light on whether the underlying financial constraints faced by poor households are driven by a lack of capital or inefficiencies in savings instruments and behaviors. Afzal et al. (2018) show that loans (to be reimbursed) and savings (to be accumulated) are substi- tutes for microcredit clients in Pakistan. However, empirical evidence on the differential effects of saving “up” (using VSLAs) or “down” (using cash grants) is scarce, and there 6 A 50 percent repayment condition on a cash grant is equivalent to an interest free credit on 50% of the grant amount. 7 VSLAs were not used in the first wave of graduation programs studied in Banerjee et al. (2015b) or Bandiera et al. (2017), which included other forms of saving support (e.g. savings accounts, deposits with a collector, or ROSCAs). However, VSLAs have been increasingly considered in economic inclusion programming, adding to the relevance of their analysis. 8 In the context of graduation programs, Sedlmayr et al. (2020) show that there is no additional impact of saving groups on top of cash grants, training and coaching. Bossuroy et al. (2022) find large impacts of a multi-faceted intervention that includes VSLAs, even when the package does not include a cash grant. 3 is also limited evidence on the differential effects of cash grants and savings interventions. Our third main contribution is to add to the literature on integrated interventions ad- dressing financial and skills constraints in fragile settings, including by analyzing both economic and social spillovers. There have been robust impacts of economic inclusion pro- grams, but evidence is more mixed in fragile settings, with strong impacts in Afghanistan (Bedoya et al. (2019)) and South Sudan (Chowdhury et al. (2017)), but weaker results in the Republic of Yemen (Brune et al. (2022)). Our RCT took place in western regions ote d’Ivoire in 2015-2018, a few years after a decade-long conflict and a 2011 post- of Cˆ electoral crisis. In this context, there were concerns over lingering tensions between na- tives, internal migrants and foreign migrants. Our design generates comparable groups of non-selected individuals across treatment and control localities (similar to Angelucci and De Giorgi (2009)) and identifies local spillovers. Employment or entrepreneurship pro- grams can in theory either fuel or crowd out economic activities among non-participants. There is limited empirical evidence on spillovers of economic inclusion programs, for in- stance with both Banerjee et al. (2015b) and Sedlmayr et al. (2020) finding no evidence of spillovers in treated villages.9 Indirect effects may also arise in other domains such as social cohesion. Interventions improving economic outcomes could benefit the entire local economy, increase economic interactions and as such positively affect social cohesion. On the other hand, they could have negative side effects. Social impacts and spillovers may also depend on the nature of the interventions. For instance, VSLAs provide repeated contacts and interactions between participants from various groups over a long period, and as such could reduce tensions in line with the “contact hypothesis”.10 In comparison, grants may induce tensions when allocated to some individuals but not others. We find that the cash grants, cash grants with repayment and VSLAs have differential effects on the dynamics of savings and productive asset accumulation. Without increasing total savings, VSLAs shift savings towards a more efficient instrument and, over time, increase investments in independent activities. Cash grants and cash grants with repay- 9 A larger literature studies spillovers from regular cash transfers. For instance, Angelucci and De Giorgi (2009) and Egger et al. (2022) find positive effects on the untreated. 10 Lowe (2021) shows positive impact of repeated interactions between caste groups in cricket leagues in India. Feigenberg et al. (2013) shows that repeated social interactions in microcredit groups (through more frequent meetings) results in more economic cooperation (willingness to pool risk). 4 ment induce investments in start-up capital, but recipients save a substantial share of the grants (approximately 30%). While start-up investments are twice as large for cash grants than VSLAs, endline impacts on productive assets are similar, suggesting that the VSLA treatment catches up with cash grants over time. This highlights that address- ing savings inefficiencies can generate a similar level of productive asset accumulation (but at a lower cost) than injecting capital. Furthermore, and consistent with findings epon et al. (Forthcoming), adding a repayment in Beaman et al. (Forthcoming) and Cr´ condition to the grant does not alter saving and investment behaviors, and thus does not help overcome behavioral constraints. Despite these changes in savings and investment dynamics, we only find a limited increase in independent activities across treatments, and no significant impact on profits, earnings or household food security. The intervention leads to some increase in the number of economic activities among beneficiaries, but does not reduce economic activities for non-selected individuals. On the contrary, it generates positive spillovers by increasing savings, which leads to marginally higher investments by non-selected individuals, too. An increase in transfers is found among beneficiaries, although without evidence of impacts on broader measures of social cohesion or spillovers on social outcomes for non-selected individuals in the community. Overall, the results in our post-conflict setting are more muted than those found for other economic inclusion or graduation programs in the literature. The large share of the cash grant saved points to high precautionary savings motives, which may limit investments. This is noteworthy as the intervention did not include regular consumption support, which may be needed for individuals to make larger, higher return investments that may also be higher-risk. The paper is structured as follows. Section 2 presents the intervention and data, and Sec- tion 3 the experimental design. Results on economic impacts and spillovers are discussed in Sections 4 and 5. Section 6 presents direct and indirect impacts on social outcomes. Section 7 concludes. Tables and figures are presented in the appendix. 5 2 Intervention and Experimental Design 2.1 Intervention The program we study aimed to promote economic empowerment and social cohesion ote d’Ivoire. These regions were particularly in the Western regions of post-conflict Cˆ affected by two episodes of conflict, and ethnic fragmentation created lingering inter- group tensions.11 The program was implemented by the International Rescue Committee (IRC) under the supervision of a government agency in the Ministry of Youth Employment (BCP Emploi). It took place between 2014 and 2018 in 37 sub-prefectures across four regions. The RCT was embedded in the second implementation phase (between July 2015 to 2017).12 The program was designed to improve livelihoods by addressing both human capital and financial constraints. Three alternative financial support modalities were tested: cash grants, cash grants with repayments and village savings and loan association (VSLA). A similar entrepreneurship training was delivered with each of the three intervention modalities. Table 1 provides a summary of the content of each modality. In each locality, the program created small groups of participants. In cash grant and cash grant with repayment modalities, the groups mostly served as an entry point to deliver entrepreneurship training. In the VSLA intervention, the groups were established as savings groups in addition to being the entry point for entrepreneurship training. Specifically, beneficiaries created an association, elected a committee, and regularly met (weekly or biweekly) to save into a common pot. Facilitators from the locality were trained by the NGO to help participants set up the VSLA and keep books. After a VSLA reached a certain level of savings (usually after four months), participants could request loans (at a rate pre-determined at the start of the cycle). The possibility of taking loans is a key difference with ROSCAs, and one reason VSLAs are considered a more efficient 11 A high fragmentation between natives, other Ivorians and migrants was considered a key driver of fragility and the 2002-07 conflict. Historically, tensions across groups were linked to land ownership. The post-electoral crisis in 2010-11 exacerbated these tensions. At the start of the program, the region had become peaceful again, but tensions remained. 12 Appendix A2 provides additional details. The regions are Tonkpi, Cavally, Gu´ emon and Bafing. The program was rolled out in three phases. The 37 sub-prefectures identified during project preparation were assigned to a specific phase, so that there is no overlap between phases. The RCT was embedded in the second phase, which covered 16 sub-prefectures and had the most participants. 6 Table 1: Intervention content by treatment arm T1 T2 T3 Village Savings and Cash grant Cash Grant Loan Association (VSLA) with repayment Basic training : - Peace building and Social Cohesion X X X - Entrepreneurship 1 (“starting an activity”) Development of business plan X X X Validation of business plan X X Village Savings and Loan Association X Complementary training : - Life skills X X X - Entrepreneurship 2 (“managing an activity”) Cash Grant Disbursement X X Follow-up counselling X Partial repayment X instrument.13 At the end of the cycle (after 9-10 months), the pot is redistributed to participants proportionally to their savings share. The interests paid by borrowers are redistributed as part of the share-out. Hence, while there is no cash injection, participants mobilize savings over time and access a lump-sum transfer at the end of the cycle, which can in turn facilitate investment. They can also request a loan during the cycle to access capital earlier. NGO staff made regular follow-up visits to the VSLA groups during their first cycle, ending with the supervision of the pot redistribution. The entrepreneurship training was common across financial support modalities. It aimed to build entrepreneurship and business skills and was delivered by NGO staff.14 The training took place in two phases, lasting 55 hours over 8 days. A first phase (5 days) focused on how to start a business.15 All participants then prepared a simple business plan for an income-generating activity. This involved searching for relevant information on prices, costs and competitors. A second phase (3 days) covered more advanced topics related to activity management, complemented with a life skills module. In contrast with the VSLA intervention, the “cash grant with repayment” and “cash 13 A ROSCA (known as tontine in Cˆ ote d’Ivoire) is an informal group in which members contribute to a common pot which is awarded to a different member at each meeting. In ROSCAs, the timing to access the pot is constrained and usually predetermined. In contrast, in VSLAs small loans can be taken at any group meeting, and the amount is flexible. 14 A “community expert” was chosen by the NGO in each locality to participate in the training and help explain some of the content to participants beyond the training duration. Those individuals were not in the study sample, and generally had higher levels of education. 15 The basic entrepreneurship training curriculum covers topics such as how to choose a business, how to attract potential clients, how to deal with competition, how to manage costs, how to set a price, etc. A motivational module on peace building and community engagement was also included. 7 grant” modalities provided a direct capital injection, which took place after training completion and approval of the business plans. In these two modalities, business plans were evaluated by a committee including NGO staff and representatives from a local bank or MFI. More than 95% of the business plans were approved. Cash grants were delivered after approval of the business plans, with no additional requirement and limited monitoring. The grants were paid through bank accounts setup with support from the NGO. The exact amount depended on the business plan, but averaged 95,000 CFA per beneficiary (USD 163 nominal, USD 384 PPP in 2017), similar across the two modalities.16 In the “cash grant with repayment” modality, beneficiaries were told they would have to repay half the grant between three and six months later, with a grace period of one month. Following the disbursement, beneficiaries received light follow-up support from NGO staff.17 Collectors went to the field to start recovering funds 4 months after disbursement of the cash grant with repayment modality, but the operation suffered from delays and under- staffing. Less than one-fifth of beneficiaries reimbursed the required amount. On average across beneficiaries, only 20% of the grant amount (instead of the planned 50%) was repaid. Importantly, field visits by the NGO suggested that most of the participants remained convinced that they would have to reimburse half of the funds at some point. Given the clear upfront communication about the repayment condition and beneficiaries’ expectations, we consider the cash grant with repayment modality as different than the (unconditional) cash grants. Table A1 documents the cost of the various treatments. Training, program administration and coordination have a similar cost across arms. In total, the program costs USD 223.4 per beneficiary for the VSLA arm, USD 269 for the cash grant with repayment arm, and USD 305.2 for the cash grant arm. VSLA facilitation costs USD 78.5, approximately half the amount of the cash grant and cash grant with repayment.18 16 The amount depended on the capital needs from the business plan, but in practice 95% of grants were for the maximum amount (100,000 CFA). There is no difference in the amount across cash grant modalities: the average was 95,624 CFA per cash grant with repayment and 94,946 CFA per cash grant. 17 This follow-up lasted between 3 and 8 hours per business, depending on needs. For agricultural activities, experts provided technical support and advice. 18 Only part of the cash grant with repayment was recovered, as mentioned above. However, if 50% had been recovered as announced, then then cost of the cash grant net of repayment would have been similar to the cost of VSLA facilitation. 8 2.2 Experimental design The study was designed as a cluster RCT to test the effectiveness of the overall interven- tion combining financial support and training, as well as to isolate the relative effect of the three alternative instruments to address financial constraints: (i) Village Savings and Loan Associations (VSLA), (ii) cash grant with repayment, and (iii) cash grants. The experiment was designed to identify both direct impacts and local spillovers on economic and social outcomes. Two-stage locality randomization The RCT covered 16 sub-prefectures in four regions. A first lottery sampled 207 out of 354 eligible localities.19 NGO staff then organized an assembly in each locality to provide basic information about the three possible interventions. Interested individuals were invited to enroll and baseline data was collected. A second lottery then assigned the 207 localities to three treatment or control arms. This two-stage process ensured that the same basic information was shared in all localities, which prevented differential enrollment of individuals into the three treatment modalities. The process also identified potential beneficiaries (individuals who would have been selected) and non-beneficiaries (individuals who would not have been selected) in localities ultimately assigned to the control group, which is key for the estimation of treatment effects and spillovers. Through the second lottery, 60 localities were assigned to the control group (C), 53 localities to “VSLAs” (T1), 64 localities to “cash grants with repayment” (T2), and 30 localities to “cash grants” (T3).20 Separate lotteries were held for 13 clusters of sub- prefectures, and for urban and rural areas in each cluster. Table A2 (Panel A) shows the number of localities assigned to each study arm. Approximately 20% of localities were urban. Figure 3 summarizes the experimental design. 19 See Appendix A2 on the eligibility of localities. 20 The lower number of localities (30) in the cash grant modality was due to budget envelopes de- termined at program design stage for each intervention modality, in particular because the cash grant modality was added for the RCT at a late stage in the program design process. To perform the random- ization, localities in nearby sub-prefectures were grouped into clusters (with 13 clusters in total) so that there would be at least 10 localities per cluster. No locality was assigned to the cash grant intervention in a few smaller clusters. 9 Beneficiary selection Across the 207 localities, individuals interested in the program were invited to enroll before the assignment to treatment or control groups. Enrollment data was collected by the implementing NGO. It was then used to check eligibility and screen applicants by level of vulnerability. Out of 14,880 interested individuals, 12,692 met the eligibility criteria as they were either 18-40 years old, a single mother above 15 years old, a widow or disabled individual (up to 60 years old) and were not benefiting from another program. Among the applicants, 5,116 were selected for the program. The selection was based on an individual proxy score computed using the enrollment data.21 Selected individuals are those above the vulnerability cut-off, i.e. those considered the most vulnerable. The cut-off is identical for all localities assigned to the same intervention, but varies across the three interventions based on the budget and target number of beneficiaries determined during the program preparation process. As further detailed below, the identification of treatment and spillover effects uses “common support” cut-offs across treatment modal- ities. 3 Estimation strategy and data 3.1 Estimation strategy We estimate intent-to-treat treatment effects by taking differences in outcomes between treatment and control groups at endline. For an outcome Y , we estimate the following ordinary least squares (OLS) regression by using the main sample to estimate direct effects, and the sample of non-selected individuals to estimate spillover effects: Yi,j = α + βTj + Sj + i,j (1) where i indexes the individual and j the locality. β is the pooled Intention-To-Treat (ITT) estimate of the program’s overall impact. Tj captures treatment assignment in locality j and Sj the randomization strata, i.e. a dummy for rural villages or urban districts within 21 Table A3 details the composition of the score. The score weighs criteria including disability, marital status, education, employment, economic status, assets and economic responsibility in the household. Weights were chosen to maximize the dispersion of the score. 10 each sub-prefecture cluster.22 Robust standard errors are clustered at the locality level. Monetary outcomes are winsorized at the 99th percentile. To estimate the relative effect of each treatment modality, we estimate the following (OLS) regression, again using the main sample to estimate direct effects, and the sample of non-selected individuals to estimate spillover effects: Yi,j = α + β1 ∗ Tj,1 + β2 ∗ Tj,2 + β3 ∗ Tj,3 + Sj + i,j (2) β1 estimates of the effect of the VSLA treatment, β2 of cash grants with repayments and β3 of cash grants. To estimate direct impacts using the main sample, we compare individuals above the vul- nerability cut-off across treated and control localities, as illustrated in Figure 1. Because a different cut-off was used to select beneficiaries in each treatment arm (see dot, plain and dashed lines for each treatment arm in Figure 1.A) due to the budget envelopes for each intervention modality, we use a common support cut-off to pool observations across arms, providing a main sample of 2,936 individuals.23 Another common support cut-off is used to estimate spillovers between comparable groups (below the cut-off) across treatment and control localities. The common support is lower in this case (as illustrated in Figure 1.B) and the spillover sample contains 1,201 individ- uals. Note that spillovers are identified among non-selected individuals ranked below the common support selection cut-off. Consequently, our estimates of spillovers are local and valid for “less vulnerable applicants”, and may not generalize to the whole population of non-applicant in sample localities. Also note that, by design, the spillover sample only contains rural localities. Urban districts were excluded since the data collection budget 22 When estimating spillover effects, the sample does not include urban districts. Therefore, the strata become the sub-prefecture clusters. 23 This approach lowers statistical power in comparison to estimating (1) separately for each treatment arm. As shown in Figure 1, the “common support cut-off” differs from the actual selection cut-off for two of the arms. This common support cut-off is used to estimate differences between the treatment arms and with control. In the control group, we can simulate “who would have been selected” if the locality was assigned to (each) treatment arm, using the actual cut-off. Pairwise comparisons between each treatment arm and the control group can thus be made using the whole sample of individuals above the selection cut-off for that specific arm. Since this does not require using a common cut-off, the approach improves statistical power. In practice these gains are marginal and results are robust. Also note that an alternative strategy using within-village differences between participants and non-participants would be biased due to local spillovers. 11 Figure 1: Selection and common support cut-offs 1.1. A. Intervention arms and pooled sample to estimate direct impacts 1.2. B. Intervention arms and pooled sample to estimate local spillovers 12 was limited and spillovers were mostly expected in rural areas with a relatively higher coverage. For all specifications, we use double-selection Lasso to include relevant baseline controls, starting with the full set of variables in balance Table A5 (for direct effects) or Table A6 (for local spillovers). To account for multiple hypothesis testing, we also calculate p-values controlling for the false discovery rate (FDR) following the step-up approach (Benjamini and Hochberg, 1995). The adjusted p values for outcomes in the main tables are presented in Table A9. Results are generally robust, but we note deviations in the text. Lastly, we analyze treatment heterogeneity by estimating specifications in equations 1 and 2 with an added linear interaction between the treatment variable and the group of interest (and a control for the group of interest). We do so for rural areas (as this dimension is used in the stratified randomization), as well as for key subgroups of program participants such as women, youth and natives. 3.2 Timeline and Surveys Enrollment and Baseline survey. Figure 2 summarizes the study timeline. Enroll- ment took place in the 207 sample localities between December 2015 and January 2016. The NGO visited each locality at least twice. First, it organized a public meeting to de- scribe the program and explain eligibility criteria. Second, the NGO collected enrollment (baseline) data for individuals interested in the program.24 Baseline data was collected for 14,880 (interested) individuals, and 12,692 fulfilled the eligibility criteria. As explained above, and key to the integrity of the experimental design, baseline data was collected prior to assigning localities to the three treatment and control groups. The baseline questionnaire included simple variables on employment, assets, education, and household characteristics. The second public lottery took place in March 2016. The list of the 5,116 selected beneficiaries was publicly released in each locality between July and September 2016.25 24 The baseline instrument was designed by the research team. NGO staff were trained as enumerators. An independent team of experienced enumerators was hired to perform field supervision and data quality checks. A double-blind data-entry process was set-up. 25 The lag between the second lottery and the release of the beneficiary lists was due to delays in the 13 The number of beneficiaries varies by treatment modality and was pre-specified based on the budget for each modality.26 Figure 2: Intervention and experiment timeline Intervention timeline. Beneficiaries of “cash grants with repayment” and “cash grants” participated in the first part of the entrepreneurship training between September and October 2016. Business plans were reviewed between November 2016 and February 2017 and the funds were mostly disbursed between June and July 2017.27 In VSLA localities, savings groups were set-up between October and December 2016 and participated in the first part of the entrepreneurship training between January and April 2017. Although there was no cash injection, beneficiaries had the opportunity to take loans from the groups since February 2017. The first savings cycle ended between July and September 2017, when the pot (including savings and interest) was redistributed to group members. This share-out took place a couple of months after the cash grants delivery. Endline Survey. The endline survey was conducted between July and September 2018, on average 15 months after the end of the program (i.e. between 12 and 18 months after cash grants, or between 10 and 13 months after the end of the first VSLA cycle). The endline sample includes 5,220 individuals (or 4,137 individuals when using the common data entry process and subsequent village committee verifications. 26 There are 1,999 beneficiaries in VSLA localities, 1,870 in cash grant with repayment localities, and 1,247 in cash grant localities. Since VSLAs require a minimum number of 17 participants, a dozen non- selected individuals were invited to join a VSLA together with selected participants. These additional individuals are excluded from the survey sampling frame. 27 A few grants were delivered in March and April 2017, but most were delivered in June and July 2017. A delay occurred for two reasons. First, unrest in the area led to the suspension of payments between April and June 2017. Second, the banking partner encountered difficulties in delivering funds to remote locations. To solve this issue, beneficiaries in remote locations were compensated to cover transportation costs to the closest bank branch. 14 support cut-off). The endline sample size was determined based on power calculation to detect minimum effects between 0.15 standard deviations (for the pooled treatment) and 0.23 standard deviations (for comparisons across arms), for a power level of 80% and significance level of 5% and using baseline data for the intra-cluster correlation (ICC) parameter. Note that spillover effects are likely smaller than direct treatment effects, hence the design is not powered to detect very small spillovers, and we focus on analyzing spillovers for the pooled treatment.28 The main sample includes 2,936 “selected” individuals (above the common vulnerability cut-off in treated and control localities) and is used to estimate direct impacts on economic and social outcomes.29 The spillover sample includes 1,201 “non-selected” individuals (below the common vulnerability cut-off in treated and control localities) and is used to estimate local spillovers.30 Control group individuals were sampled using the same vulnerability cut-off. The follow-up attrition rate was 10.6%, respectively 7.8% (using the common support cut-off), and was balanced between treatment and control groups.31 Table A2 contains the detailed composition of baseline and endline samples. The follow-up questionnaire collects data about household characteristics, employment, assets, food security, well-being, saving and debt, social relationships, community activi- ties and interpersonal trust. A shorter version of the questionnaire was administered to the “non-selected” sample. Appendix A3 provides additional information on the main 28 The endline survey sample was established based on power calculations to detect impacts between treatment arms and the control group. 15 “selected” individuals (respectively 17 and 22) were sampled out of localities in T1 (respectively T2 and T3). Finally, 10 “non-selected” individuals were sampled in each (rural) locality to study spillovers. When it comes to estimating spillovers, the sample size for the non-selected was determined to detect minimum effects between 0.18 standard deviations (for the pooled treatment) and 0.26 standard deviations (for comparisons across arms), for a power level of 80% and significance level of 5% and using baseline data for the intra-cluster correlation (ICC) parameter. Power was recalculated ex-post using the follow-up control group data to compute the mean, standard deviation and intra-cluster correlations for key outcomes. The minimum detectable effects remain between 0.15 and 0.22 standard deviations for the main outcomes. This means that the study is powered to detect a 20% impact (+5,300 CFA) on profits in independent activities, a 37% impact (+5,800 CFA) on starting capital or a 23% impact (+5,800 CFA) on the value of assets in the main activity. 29 The sample includes 3,237 “selected” individuals, and the pooled specification with a common sup- port cut-off contains 2,936 observations. 30 The spillover sample includes 1,983 “non-selected” individuals, and the pooled specification with a common support cut-off contains 1,201 observations. 31 In the sample of “selected” individuals, attrition is 10.56%, 11.30% and 10.44% for T1, T2 respec- tively T3 (using the common support cut-off). In the sample of “non-selected” individuals, attrition is 7.92%, 8.22% and 8.17% across T1, T2 respectively T3 (using the common support cut-off). In the total sample of selected individuals (not using the common support cut-off), attrition is on average 10.8% for “selected individuals” and 8.4% for “non-selected individuals”. 15 outcome variables. 3.3 Descriptive statistics and balance Table A4 summarizes the characteristics of all applicants (column 4), selected individuals (column 2) and non-selected individuals (column 3). Of the applicants, 77% live in rural areas. They are on average 33.7 years old, with 3.3 children, and 62% are women. 80% do not have any diploma, meaning that they have not completed primary school. Nearly all applicants have an activity at baseline (96.3%). They are primarily self-employed in agriculture (54.7%) or non-agricultural (off-farm business) activities (25.4%), while only 9% hold a wage job.32 Access to finance is very limited: only 2% of applicants have a bank account and 20% use mobile money. Half of the applicants report having saved some money over the last 3 months in ROSCAs. Selected individuals differ from non-selected individuals due to the targeting based on the vulnerability proxy score (Table A4, columns 2, 3 and 5). Selected applicants are signif- icantly less educated than non-selected applicants (30pp difference in the share without a school diploma). Selected applicants also tend to be poorer: they have lower baseline earnings (by roughly 30,000 CFA), they are less likely to use mobile money (by 12pp) and they are more likely to report facing constraints for education and health expenditures (by 4pp). Finally, the share of women is larger in the selected group: 71% of selected individuals are women (20pp higher than in the non-selected group). Table A5 and A6 document balance between individuals in treatment and control locali- ties (respectively for the main and spillover samples). A common support cut-off is used, consistent with the identification strategy used to estimate equations 1 and 2. The exper- iment achieved satisfactory balance between the pooled treatment and control group, as well as between treatment modalities. There are few statistically significant differences, and they remain of small magnitude. Table A7 illustrates how the study sample compares with the general population of in- dividuals 18-65 years old in program districts.33 Sample individuals live in larger house- holds, are more likely to be female and not to hold a diploma. The activity rates, types 32 ote d’Ivoire, especially in rural areas (Chris- This is in line with the composition of employment in Cˆ tiaensen and Premand (2017)). 33 This is based on data from a national employment survey. 16 of economic activities and overall asset holdings are otherwise comparable. 3.4 Take-up Table A8 presents take-up rates at various stages of each intervention modality. Take-up is high. 78.9% of selected individuals received the funds in the cash grant with repayment treatment, and 81.1% in the cash grant treatment. The VSLA intervention had a lower take-up rate (69.5%), which is mainly driven by urban areas, where participation is lower.34 Participation in the entrepreneurship training is above 75% in all groups.35 As discussed further below, we focus on intent-to-treat estimates of program impacts among all selected applicants. 4 Results: Direct economic impacts In this section, we discuss direct impacts on employment and independent activities (Table 2), earnings and other welfare proxies (Tables 3-4), savings and investments in independent activities (Tables 5-6). In each table, panel A provides estimates for the pooled treatment based on the specification in equation 1. Panel B presents estimates for each financial support modality based on equation 2 (β1 =“VSLA”, β2 =“cash grant with repayment” and β3 = “cash grant”). We provide p-values for a pairwise test of equality between treatment arms, and for a joint test β1 = β2 = β3 . Recall that treatment effects are estimated approximately 15 months after the end of the interventions. 4.1 Employment and earnings Employment and independent activities The pooled treatment induces small adjustments in participants’ activities towards self- employment. There is no change in the share of individuals employed at the extensive 34 The vulnerability score is positively and significantly associated with program take-up, which means that those who did not fully participate were less vulnerable. There is no significant difference in program participation between the cash grants with repayment and cash grant interventions. Living in rural areas is a strong determinant of participation in VSLAs. 35 Conditional on setting up a VSLA, participation in the entrepreneurship training is 85%. In both cash grant interventions, participation in the second training is conditional on business plan approval. More than 75% of individuals submitted a business plan, got a business plan approved, and participated in the second training. 17 margin (Table 2, Panel A, column 1) as 95% of control individuals already have an activity. A small but significant impact on entry into self-employment is found (+2pp, Table 2, Panel A, column 2). It is offset by a small decrease in participation in wage employment (-3pp, Table 2, Panel A , column 3). Similar results are found for hours worked (Table 2, Panel A, columns 5-7), with no significant change in total hours, and a positive effect on hours in self-employment (not statistically significant, p=0.195) offset by a decrease in hours in wage employment. The pooled treatment induces more substantial changes at the intensive margin by sig- nificantly increasing the number of independent activities by 0.24 per individual (Table 2, Panel A, column 4). This means that, on average, one out of four individuals added a new activity to his/her portfolio, representing a 8% increase relative to the control group. Yet it appears that a substantial share of beneficiaries did not launch the activity they proposed in their business plan. While agricultural activities account for 80% of inde- pendent activities in control, 70% of grant beneficiaries prepared a business plan for a non-agricultural activity (often retail trading). The impact on entry into non-agricultural activity is only 0.08.36 To better understand the dynamics of activity creation, we also compared the business plans developed during the program with the activities listed by respondents in the follow-up survey. Only 40% of the activities in the business plans could be matched to activities reported as active in the 12 months preceding the follow- up survey. This is broadly consistent with the magnitude of the observed impacts on the net number of activities, and suggests that a large share of individuals either did not launch their proposed activity or stopped it quickly. Importantly, we do not find differences in impacts on employment or time worked in wage or self-employment between treatment modalities (Table 2, panel B). We cannot reject that the treatment effects are equal between the three arms for any variable, aside from cash grant modalities leading to fewer hours in wage employment. There is no signifi- cant difference in the impact on the number of independent activities across treatment modalities either (p=0.6, panel B, column 4). 36 The point estimate for the increase in independent activities at endline is 0.15 (not statistically significant) for agriculture, and 0.08 (significant at 10 percent) for non-agricultural activities (Table A10, Panel A, columns 1 and 2). 18 Earnings and Welfare Table 3 documents impacts on earnings, including income from wage employment and profits from independent activities. Overall, earnings are unchanged by the pooled treat- ment. The increase in the number of independent activities does not induce a significant increase in total profits (Table 3, Panel A, column 2). The point estimate on profits has a small magnitude (+1,727 CFA), or less than 0.05 standard deviation. Since the study is well powered and the point estimate is close to zero, we conclude that the pooled treatment does not increase profits from independent activities.37 Consistent with results based on self-reported profits in independent agricultural activities, we do not find im- pacts on revenues from crop or livestock sales over the last 12 months either (Table 3, Panel A, columns 3 and 4).38 Impacts on wage earnings are also very close to zero and not statistically significant (column 1). Similar results are found across the financial support modalities, with no significant im- pact on earnings or difference between treatment arms (Table 3, Panel B). The only exception is a slight reduction in earnings from livestock among cash grant recipients, suggesting lower sales of livestock in that group. To explore potential differences in impacts on self-employment earnings along the distri- bution, Figure 4 displays the distribution of profits in the treatment and control group and Figure 5 shows estimates of quantile treatment effects. The profit distributions are not statistically different from each other at any point (Figure 4). Quantile treatment estimates suggest positive but very small and non-significant effects along almost the full distribution (Figure 5). We conclude that the lack of impact on average profits does not hide large changes in some parts of the distribution. Results on other welfare outcomes are consistent with a lack of impact on earnings. We do not find impact or differences between arms in terms of food security (measured through the food consumption score, which captures dietary diversity, Table 4, column 1), the number of durable goods owned by the household (column 2), or household education 37 No impact is found when analyzing profits separately from agricultural and non-agricultural activities either (Table A10, columns 3 and 4). 38 Since a substantial share of independent activities relates to agriculture, Table A11 contains results for additional agricultural outcomes. We do not find impacts on whether beneficiaries hold livestock, cultivate crops, use fertilizers, or on the area or type of crops they cultivate. The pooled treatment effect is not statistically significant and we cannot reject equality between treatment arms in any case (except that cash grant beneficiaries spend a little less time raising livestock that other modalities). 19 expenditures (column 3).39 There is no change in beneficiaries’ psychological well-being either (column 4). 4.2 Investment and Savings Differential Investment Dynamics The cash grant and cash grant with repayment provide a direct capital injection. Al- though the grants are unconditional, they explicitly aimed to finance investments in the activity outlined in the business plan prepared following the entrepreneurship training. Table 5 documents impacts on start-up capital and productive assets in independent ac- tivities. Results clearly show that interventions with capital injection induce investments in start-up capital (Panel B, column 1). They also increase the value of assets at endline (Panel B, column 2), which is net of potential investments or dis-investments between the launch of the activity and the follow-up survey. Specifically, the start-up capital of inde- pendent activities nearly doubles, increasing by 115% relative to control for cash grants and by 97% for cash grants with repayment (+17,813 CFA, respectively +14,828 CFA) (Table 5, Panel B, column 1). The total value of productive assets at follow-up is also sig- nificantly higher for cash grants (+12,245 CFA, or 31%) and cash grant with repayment (+7,788 CFA, or 20%) relative to control (Table 5, Panel B, column 2). However, the point estimates for the value of productive assets at endline tend to be lower (though not significantly so) than the point estimates for start-up capital, suggesting that the grants provided a one-off increase in capital, without investments continuing over time. In contrast, the VSLA intervention shows different investment dynamics. The VSLA treatment does not provide a capital injection, but instead mobilizes savings which could, over time, be used for investments. VSLAs can contribute to investments by offering small loans from the accumulated group savings, or through the one-off redistribution of savings plus interests at the end of each cycle. Investments in start-up capital are substantially lower for the VSLA modality relative to the cash grants, yet there is no significant difference in the value of productive assets at follow-up. The VSLA treatment increases start-up capital by +7,259 CFA, or 48% relative to control (Table 5, Panel 39 Panel B suggests a decrease in education expenditures among cash grant beneficiaries (p=0.04), but the FDR-adjusted p value is 0.16 in Table A9. 20 B, column 1), which is a significantly lower impact on initial investment than the cash grant with repayment and cash grants (p=0.04, respectively p=0.07). In contrast, the VSLA intervention increases the value of productive assets at endline by 9,614 CFA (24% relative to control, Table 5, Panel B, column 2), which is not statistically different from the cash grant with repayment (p=0.64) or cash grant (p=0.68). This provides evidence that individuals in the VSLA treatment may be on a different trajectory than recipients of cash grants and cash grants with repayment. Also note that the point estimate of the VSLA treatment tends to be larger (though not significantly so) for the value of productive assets at endline than for initial start-up capital, which suggests that VSLA recipients (unlike cash grant recipients) continue investing over time. Differential Effects on Savings We also find important differences in impacts on savings between financial support modal- ities, with the cash grants increasing total savings and the VSLAs mostly enhancing savings efficiency. The cash grant interventions tend to increase savings, by 26,694 CFA for the cash grant and 11,832 CFA for the cash grant with repayment (Table 6, Panel B, column 4). The point estimates correspond to approximately 30% of the grant in both cases (taking into account the 50 percent to be repaid for the cash grant with repayment), or a 20%-45% increase in savings relative to control.40 In absence of formal instruments, savings mostly take place through informal groups such as ROSCAs (“tontines”). For instance, 50% of the saving stock in the control group is held in ROSCAs. The overall increase in savings from cash grants is strongly driven by informal savings groups (+21,466 CFA) (Table 6, Panel B, column 6). While ROSCAs may not be the most efficient saving instrument, they offer a way for cash grant recipients to save in a setting with limited access to financial institutions. The fact that a substantial share of the cash grant is saved also highlights that investments from cash grants, while statistically significant, remain relatively limited in magnitude. This likely contributes to explaining the lack of effects on earnings. Most economic in- 40 Note that despite the magnitude of the coefficients, this is a noisy outcome. The coefficient for cash grant with repayment is not statistically significant, and the statistical significance of the cash grant coefficient is sensitive to multiple hypothesis corrections. 21 clusion or graduation interventions in the literature provide small regular transfers to stabilize consumption and help households make higher-risk, higher-return investments. Regular consumption support was not included in the intervention we study, possibly leading households to use some of the financial resources for precautionary savings rather than investments. Precautionary savings motives may be particularly high in a rural, post-conflict setting with limited access to formal financial services. To further inves- tigate savings behavior, we explore the distribution of savings stock in treatment and control. Precautionary savings motives would imply stronger impacts in the bottom of the distribution. Consistent with this, Figure 6 indicates that the increase in amounts saved arises mostly in the bottom of the distribution (up to approximately 40,000 CFA). The VSLA intervention affects savings in markedly different ways than the cash grants: it increases savings efficiency but not total savings. There is a shift in the instruments used by individuals to save, as intended by the intervention. Among the control individuals, 30% save in ROSCAs, but only 17% in an enhanced savings group such as a VSLA (Table 6, columns 2-3). In contrast, 54% of individuals assigned to the VSLA treatment participate in a VSLA at follow-up (+37pp), with a substitution away from ROSCAs (-12pp) (Table 6, Panel B, columns 2 and 3). The 37pp impact on VSLA participation is noteworthy because it shows sustained effects 15 months after the end of the program. Given that VSLA cycles last 9 months, this means that groups launched new cycles after the program. These results contrast with some of the VSLA interventions studied in the literature that exhibit low take-up.41 However, and importantly, the VSLA intervention does not increase the total amounts saved. The total savings stock remains similar as the increase in the amount saved through VSLAs (+9,505 CFA) is offset by a decrease in savings in ROSCAs (-16,470 CFA) (Table 6, Panel B, columns 5-6). The VSLA intervention thus mostly shifts savings toward enhanced savings instruments. As such, observed impacts on economic activities are not explained by a larger overall mobilization of savings: VSLA beneficiaries save similar amounts than individuals in the control group. However, offering access to enhanced savings mechanisms (combined with micro-entrepreneurship training) fuels investments in income-generating activities. This occurs without any injection of additional capital 41 In constrast, participation is 31.6% across the three countries in Karlan et al. (2017), 37% in Beaman et al. (2014) and 45% in Ksoll et al. (2016). 22 in the community. It is noteworthy that such an improvement in savings efficiency leads to investments that continue over time, as previously documented. The higher participation in VSLAs is also associated with an increase in credits obtained from the groups, which likely contributes to the observed investments. The share of individuals who take any type of credit increases by 13pp (Table A12, Panel B, column 1), more than for cash grant recipients.42 The total amount of credit taken increases by one-third in the VSLA treatment (+7,381 CFA, Table A12, Panel B, column 6), with no change for cash grant recipients. The results are fully driven by credit taken from VSLAs, which more than triple after the VSLA intervention (+9,780 CFA). Of beneficiaries in the VSLA modality, 41 percent have taken a credit with a savings group in the last two years (+31pp, Table A12, column 3). The sustained impact on VSLA participation after the program is also associated with sustained impact on access to credit: 35% of participants in the VSLA treatment have a credit in an ongoing VSLA cycle (+26pp, Table A12, column 5). Finally, we also note a marginal increase in VSLA participation in other treatment modal- ities (+7pp for cash grant with repayment and cash grants, significant only for the former, Table 6, Panel B, column 5), although it did not provide direct support for the estab- lishment of savings groups in other treatment arms. The literature shows that VSLAs can replicate (Beaman et al. (2014), Ksoll et al. (2016) and Karlan et al. (2017)). In our context, the increase in VSLA participation is substantially smaller for cash grants compared to VSLA localities (one-fifth of the coefficient, p=0.00).43 Formally, we identify treatment effects inclusive of spillovers across localities. This means that, when compar- ing the VSLA and cash grant arms, differences might be underestimated. Still, spillovers appear limited and their magnitude not large enough to explain key results such as the lack of impact on profits. 42 Only the coefficient for cash grant with repayment is marginally significant at 10%, and the 5pp difference is lower than the 13pp effect in the VSLA arm (p=0.01). 43 Karlan et al. (2017) have a design to quantify replications and observe a 19.4% participation rate in VSLA in neighboring villages. Ksoll et al. (2016) find that 21% of control households joined a VSLA in treatment localities. We estimate participation rates around 24% in other localities. 23 4.3 Additional interpretation of results on earnings The lack of impact on earnings can be further interpreted in light of the relative effects on savings and investments. On average, the cash grants provide 95,000 CFA. Among cash grant recipients, 18,000 CFA is invested in start-up capital and 27,000 CFA is saved. Among recipients of cash grant with repayment, 15,000 CFA is invested in start-up capi- tal, 12,000 CFA is saved, and 15,000 CFA is repaid. This leaves 50,000 CFA from the cash grant and 53,000 CFA from the cash grant with repayment for other investments (e.g. for business operation) or expenditures. First, the observed investments do not appear sufficient to increase profits.44 We cannot rule out that part of the grants was used for consumption or other household expenditures when they were received, although we do not detect impacts on food security, education expenditures, or household durables at endline. Second, the lack of impact on earnings also seems partly explained by a sub- stantial share of the capital injection being saved. As mentioned above, this may reflect precautionary savings motives, in particular in a post-conflict setting and in the context of an intervention that did not provide regular consumption support. What about the contribution of the cross-cutting training? The pooled treatment induces improvements in an index of business skills and an index of business practices that are statistically significant and in the order of 0.2 standard deviations (Table A13, columns 3-8).45 Most of the literature on traditional entrepreneurship training finds changes in practices that are not sufficient to increase profits (McKenzie and Woodruff (2014)), and our results are consistent with this pattern. Heterogeneity analysis reveals limited differences in impacts between subgroups such as women, youth, native or urban populations (Tables A14 - A17). There is some evidence that youth benefit slightly more in some dimensions compared to older individuals (Table A14), but the patterns are not fully robust. Notably, there is no subgroup exhibiting a statistically significant increase in profits. 44 We also note that the timeline at which the study assesses impacts is similar to other graduation studies that find large effects on earnings (such as Bossuroy et al. (2022)). As such, it is not likely that it is too early for investments to generate profits. Most activities for which business plans were developed are not active at follow-up, which also makes it unlikely that larger returns would arise in the future. 45 The increase in business knowledge is larger for the cash grant interventions, and the increase in business practices is larger for the cash grant with repayment modality that involved some additional post-training follow-up. The magnitude is similar to another experiment in Cˆ ote d’Ivoire where micro- entrepreneurship training was provided to public works participants (Bertrand et al. (2021)). 24 5 Local spillovers In this section, we present estimates of local spillovers based on equations 1 and 2 and the sample of non-selected individuals (with a baseline vulnerability score below the common support vulnerability cut-off) across treated and control (rural) localities. 5.1 Earnings and Employment Economic spillovers on non-selected individuals could theoretically be either positive or negative. As shown above, the three intervention modalities lead to the creation of new independent activities that are more capital-intensive. If these new activities connect to local value chains or increase local labor demand, positive spillovers on non-selected individuals within the same village are possible. In the opposite direction, new activities may crowd-out opportunities for others. Similarly, on the one hand the creation of VSLAs may spur financial activities within the village, but, on the other hand, they may disrupt informal savings mechanisms, for instance by weakening traditional ROSCAs. We document the net effect between these forces. Overall, we do not find evidence that non-selected individuals are economically worse-off than they would have been in absence of the intervention. There is no significant local spillover on the earnings of non-selected individuals (Table 8). The estimated coefficient for earnings from wage employment and self-employment have opposite signs, but are very close to zero and far from statistically significant. If anything, Table 7 reveals some positive local spillovers on independent activities. There is no change in the share of individuals working, the participation in wage work or self- employment, or hours worked. However, the number of independent activities signifi- cantly increases (+0.37 activities, a 11% increase relative to control, column 4). The spillovers are concentrated in independent agricultural activities (+0.33 activities, Table A18, column 2).46 46 This increase in the number of independent activities could hide business closures due to competition induced by the program. However, we find no impact on the number of activities stopped by non-selected individuals in the last 12 months. In addition, a self-reported indicator on the level of local competition also remains unchanged. Overall, we do not find evidence that businesses were negatively affected by the intervention in the same locality. 25 5.2 Capital, Savings and Other Spillover Channels We find evidence of spillovers associated with changes in financial dynamics in study communities, including savings and investments among non-selected individuals. The point estimate for spillover on productive assets in independent activities at endline is substantial (+10,763 CFA, Table 9, column 2), but noisy and not statistically significant. There are positive spillovers on assets in agricultural activities that are also of substantial magnitude (+10,091 CFA, Table A18, column 6). The point estimate for impact on start-up capital is close to zero. Taken together, these results suggest that non-selected individuals increase investments in their agricultural activities over time. As we have shown, the financial support modalities either increase the availability of capital or savings efficiency. The cash grant interventions increase beneficiaries’ savings, with a large share of the grants saved through ROSCAs. The VSLA intervention mobilizes beneficiaries’ savings and offers micro-loans to group members. Over time, a broader circulation of capital or access to enhanced savings technology might also benefit non- selected individuals. Table 10 shows that the propensity to save increases for non-selected individuals (+6pp, Panel A, column 1) and similarly across arms (panel B). Positive spillovers on financial dynamics are observed among non-selected individuals in VSLA treatment villages, with a higher participation in savings groups (+10pp, panel B, column 2).47 The amount saved in VSLA more than doubles (+4,553 CFA, Panel B, column 5), though this is not associated with an increase in total savings, a pattern similar to the one observed for beneficiaries. Point estimates suggest a reduction in the amounts saved in ROSCA (column 6), though standard errors are large and the coefficient is not significant. The same holds for the total savings. As such, we do not find conclusive evidence that the introduction of VSLAs disrupt informal savings arrangements among non-selected individuals. These results are consistent with VSLA expansion, which can take different forms (Beaman et al., 2014).48 Results also show higher access to credit for non-selected individuals in VSLA villages (+8pp, Table A19, Panel B, column 1), 47 We note that the spillover analysis was not powered for each treatment arm, and the results by treatment arm are sensitive to multiple hypothesis corrections. 48 First, existing VSLA groups can expand if new members are added at the start of a new cycle. In our study, participation was restricted to beneficiaries during the first cycle but not after. Second, VSLAs can expand if groups replicate. Both mechanisms are likely to have occurred, since a given VSLA typically cannot expand beyond 30 to 35 members. 26 although there is no significant impact on the amount of loans (Panel B, column 2). Some local spillovers on savings are also observed in villages assigned to cash grant modalities, with a higher share of non-selected individuals saving, an increase in VSLA participation, and an increase in amounts saved in VSLAs (Table 6, Panel B). These findings are consistent with those for cash grant recipients, and suggest that some VSLA replication also took place across localities. We assess several other potential mechanisms for local spillovers. First, we can rule out that new activities were jointly started between beneficiaries and non-selected indi- viduals.49 Second, activities created by beneficiaries might be complementary to those from non-selected individuals in the same locality. We explore a set of proxies on up- stream and downstream economic relations and find no such indication.50 Also note that demand-side effects are unlikely as the program does not significantly increase income, as documented above. Lastly, non-selected individuals in treatment localities could theoret- ically benefit from the diffusion of skills taught during the entrepreneurship training, or imitate adopted practices. Table A20 provides no evidence for this mechanism: the index of business practices is unchanged among non-selected individuals in treated localities (although an improvement is found in the cash grant with repayment localities). 6 Impacts and spillovers on social outcomes This section documents how the interventions affect social outcomes among selected and non-selected individuals. This is directly relevant as the program also had a secondary objective to strengthen social cohesion in the post-conflict study setting. Consistent with the contact hypothesis, the intervention can potentially directly increase interactions between individuals, including those from different ethnic groups. In particular, the VSLAs involve much more frequent interactions relative to the cash grant modalities, with 49 In cash grant localities, program participants could pool resources for activities requiring more expensive equipment. Approximately one-third of beneficiaries prepared a joint activity in their business plan. We do not find an impact on the number of jointly-owned activities at endline (pooled treatment estimate=0.13 ; standard error=0.11). 50 In the control group, respondents declare that 54% of their customers and 31% of their suppliers come from the same locality. Upstream, there is no significant impact on the use of local suppliers among beneficiaries or non-selected individuals. Downstream, half of the market is local, but mostly involves business-to-consumer relations. Less than 15% of control individuals report selling to other businesses, and we do not find significant changes in the likelihood of serving the local market either, with a small exception in cash grant villages. 27 bi-monthly meetings over 9 months. More generally, improved livelihoods or economic outcomes may in turn increase social support and solidarity between community members. We test whether there are lasting effects on social outcomes 15 months after the end of the program. We consider participation in economic or social groups, community activities, as well as financial help received and provided by sample individuals, which may increase either through improvements in economic outcomes or social ties. We also study broader measures of trust, victimization, and perception of safety. Table 11 shows a moderate increase in the number of groups in which beneficiaries par- ticipate (+0.16, Panel A, column 1).51 These effects are observed consistently across intervention modalities (Panel B, p=0.87), and thus are not solely attributable to ad- dditional social contact in VSLAs. Results also show a significant increase in financial help (Table 11, column 2-3), including the number of times beneficiaries have received financial support (+21pp) or have supported someone else (+29pp). Again, we cannot reject equality of treatment effects between intervention arms (Panel B, p=0.23), sug- gesting that these effects are not only due to more frequent social interactions in the VSLA groups or higher availability of capital from cash grant modalities. No effect is found on broader measures of participation in community activities, trust, perception of safety or a victimization index capturing exposure to crime or other violent activities in the locality (Table 11, Panel A, columns 4-8). Table A21 also shows that impacts on social outcomes do not extend to non-selected beneficiaries. Overall, the results thus point to relatively narrow changes in group par- ticipation and financial help between beneficiaries, which do not translate into more widespread changes in social outcomes in the broader community.52 51 This is driven by economic groups. Beneficiaries participate in more group meetings and are also more likely to hold leadership positions. Nearly all additional groups in which individuals participate are mixed ethnic groups. 52 We do not detect much heterogeneity in impacts on social outcomes between subgroups either. Tables A22 - A25 present results for heterogeneity analysis for subgroups including youth, women, rural, and native populations. If anything, impacts on some social outcomes (e.g. receiving financial help) are slightly stronger among non-native groups, which provides a weak indication of improved social inclusion in treatment localities. 28 7 Conclusion In this paper, we report results from the randomized controlled trial of an economic inclusion program providing financial support - through capital injection or enhanced savings groups - combined with micro-entrepreneurship training to vulnerable individuals. We study the overall effectiveness of the intervention, which addresses both financial and skill constraints. We analyze the relative effects of alternative instruments to relax capital or savings constraints in increasing savings and investments in income-generating activities. Of particular relevance for fragile settings, we also document local spillovers on non-selected individuals’ economic and social outcomes. The three financial support modalities increase the number of income-generating ac- tivities as well as productive assets, but not sufficiently to increase profits or welfare. Alternative financial support modalities affect savings and investment dynamics in dif- ferent ways. Individuals appear to face binding constraints to save: the results show that capital injections strongly increase savings and that offering a new saving technology shifts savings and investment behavior. Specifically, cash grants increase investments in start-up capital, but also have large impacts on the total amount saved: recipients save 30% of the grant, which contributes to explaining why impacts on investments are not sufficient to increase earnings. Adding a repayment condition to cash grants does not en- courage further investments, suggesting that repayment conditions do not help overcome behavioral biases. Remarkably, the VSLA intervention leads to similar impacts on pro- ductive assets as the cash grants at follow-up. While it has smaller impacts on start-up capital, beneficiaries invest more continuously over time. The VSLA intervention does not increase total savings, but rather redirects savings to a more efficient technology. In a low-income setting with thin formal financial markets, it is noteworthy that provid- ing access to more efficient savings groups combined with training facilitates investments in income-generating activities, even in the absence of capital injection. However these impacts on investments are not sufficient to increase earnings. We do not find evidence of negative spillovers or crowding-out effects for non-selected individuals in treated villages. This is in line with Banerjee et al. (2015b); Bandiera et al. (2017) and Sedlmayr et al. (2020), who do not detect spillovers from graduation programs. However, our results also suggest that the interventions increase the number of 29 independent activities among the non-selected in treated villages, partly due to increased savings flows. Some replication of VSLAs is also found, consistent with the literature (Ksoll et al. (2016), Karlan et al. (2017), Beaman et al. (2014)). Despite these increases in economic activities and savings flows, we do not observe broader changes in terms of earnings or social outcomes in the communities. Results from the experiment contrast with the positive impacts documented for grad- uation programs that provide asset transfers (Banerjee et al. (2015b); Bandiera et al. (2017); Bedoya et al. (2019)) or cash grants (Bossuroy et al. (2022)). The absence of consumption support and high precautionary saving motives in the post-conflict study setting might partly explain the more muted impacts. Small regular transfers may be key for households to smooth consumption and enter into higher-risk, higher-return ac- tivities. Other studies have found impacts on savings but not earnings in South Sudan (Chowdhury et al. (2017)) and the Republic of Yemen (Brune et al. (2022)), where con- sumption support was disrupted. The results on VSLAs remain encouraging, however, as VSLAs appear to have initiated a dynamic of savings and investments. Yet a combina- tion of capital injection and savings facilitation might ultimately be needed to overcome the constraints highlighted by the results in terms of both the level and the efficiency of savings. Such complementarities between regular cash transfers for consumption support, capital injections, and savings facilitation might be particularly relevant when seeking to optimize multi-dimensional graduation programs in fragile settings. 30 References Afzal, U., G. D’Adda, M. Fafchamps, S. Quinn, and F. Said (2018): “Two Sides of the Same Rupee? Comparing Demand for Microcredit and Microsaving in a Framed Field Experiment in Rural Pakistan,” The Economic Journal, 128, 2161–2190. Andrews, C., A. de Montesquiou, I. Arevalo Sanchez, P. V. Dutta, B. V. Paul, S. Samaranayake, J. Heisey, T. Clay, and S. Chaudhary (2021): “ The State of Economic Inclusion Report 2021: The Potential to Scale.” World Bank: Washington DC. Angelucci, M. and G. De Giorgi (2009): “Indirect effects of an aid program: How do cash transfers affect ineligibles’ consumption?” American Economic Review, 99, 486–508. Bandiera, O., R. Burgess, N. Das, S. Gulesci, I. Rasul, and M. Sulaiman (2017): “Labor markets and poverty in village economies,” Quarterly Journal of Eco- nomics, 132, 811–870. Banerjee, A., E. Duflo, R. Glennerster, and C. Kinnan (2015a): “The Mir- acle of Microfinance? Evidence from a Randomized Evaluation,” American Economic Journal: Applied Economics, 7, 22–53. Banerjee, A., E. Duflo, N. Goldberg, D. Karlan, R. Osei, W. Pariente, J. Shapiro, B. Thuysbaert, and C. Udry (2015b): “A multifaceted program causes lasting progress for the very poor: Evidence from six countries,” Science, 348, 1260799–1260799. Beaman, L., D. Karlan, and B. Thuysbaert (2014): “Saving for a (not so) Rainy Day: A Randomized Evaluation of Savings Groups in Mali,” NBER Working Paper, Cambridge, MA. Beaman, L., D. Karlan, B. Thuysbaert, and C. Udry (Forthcoming): “Self- Selection into Credit Markets: Evidence from Agriculture in Mali,” Econometrica. Bedoya, G., A. Coville, J. Haushofer, M. Isaqzadeh, and J. Shapiro (2019): “No Household Left Behind: Afghanistan Targeting the Ultra Poor Impact Evalua- tion,” NBER Working Paper, Cambridge, MA. 31 Benjamini, Y. and Y. Hochberg (1995): “Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing,” Journal of the Royal Statistical Society. Series B (Methodological), 57, 289–300. ´pon, A. Marguerie, and P. Premand (2021): “Do Work- Bertrand, M., B. Cre fare Programs Live Up to Their Promises? Experimental Evidence from Cote D’Ivoire,” NBER Working Paper 28664, Cambridge, MA. Blattman, C., N. Fiala, and S. Martinez (2014): “Generating Skilled Self- Employment in Developing Countries: Experimental Evidence from Uganda,” The Quarterly Journal of Economics, 129, 697–752. Blattman, C. and L. Ralston (2017): “Generating Employment in Poor and Fragile States: Evidence from Labor Market and Entrepreneurship Programs,” SSRN Working Paper. Bossuroy, T., M. Goldstein, B. Karimou, D. Karlan, H. Kazianga, W. Pari- ente, P. Premand, C. Thomas, C. Udry, J. Vaillant, and K. Wright (2022): “Tackling psychosocial and capital constraints to alleviate poverty,” Nature, 605, 1–7. ´, J. Goldberg, and D. Yang (2016): “Facilitating Savings for Brune, L., X. Gine Agriculture: Field Experimental Evidence from Malawi,” Economic Development and Cultural Change, 64, 187–220. Brune, L., D. Karlan, S. Kurdi, and C. Udry (2022): “Social protection amidst social upheaval: Examining the impact of a multi-faceted program for ultra-poor house- holds in Yemen,” Journal of Development Economics, 155, 102780. Cho, Y. and M. Honorati (2014): “Entrepreneurship programs in developing coun- tries: A meta regression analysis,” Labour Economics, 28, 110–130. Chowdhury, R., E. Collins, E. Ligon, and M. Sulaiman (2017): “Valuing Assets Provided to Low-Income Households in South Sudan,” BRAC University Discussion Paper. ote d’Ivoire Jobs Diagnostic — Em- Christiaensen, L. and P. Premand (2017): “Cˆ ployment, Productivity, and Inclusion for Poverty Reduction.” World Bank: Washing- ton DC. 32 ´ (2015): “Estimating the ´pon, B., F. Devoto, E. Duflo, and W. Pariente Cre Impact of Microcredit on Those Who Take It Up: Evidence from a Randomized Ex- periment in Morocco,” American Economic Journal: Applied Economics, 7, 123–50. ´pon, B., M. El Komi, and A. Osman (Forthcoming): “Is it who you are or what Cre you get? Comparing the impacts of loans and grants for microenterprise development,” American Economic Journal: Applied Economics. de Mel, S., D. McKenzie, and C. Woodruff (2008): “Returns to Capital in Mi- croenterprises: Evidence from a Field Experiment,” Quarterly Journal of Economics, 123, 1329–1372. De Mel, S., D. McKenzie, and C. Woodruff (2012): “Business training and female enterprise start-up, growth, and dynamics: experimental evidence from Sri Lanka,” Journal of Development Economics, 335, 962–966. Dupas, P. and J. Robinson (2013): “Savings constraints and microenterprise devel- opment: Evidence from a field experiment in kenya,” American Economic Journal: Applied Economics, 5, 163–192. Egger, D., J. Haushofer, E. Miguel, P. Niehaus, and M. Walker (2022): “General equilibrium effects of cash transfers: experimental evidence from Kenya,” Econometrica, 90. Fafchamps, M., D. McKenzie, S. Quinn, and C. Woodruff (2014): “Microen- terprise growth and the flypaper effect: Evidence from a randomized experiment in Ghana,” Journal of Development Economics, 106, 211–226. Feigenberg, B., E. Field, and R. Pande (2013): “The Economic Returns to Social The Economic Returns to Social Interaction: Experimental Evidence from Microfi- nance,” The Review of Economic Studies, 80, 1–55. Fiala, N. (2018): “Returns to microcredit, cash grants and training for male and female microentrepreneurs in Uganda,” World Development, 105, 189–200. Field, E., R. Pande, J. Papp, and N. Rigol (2013): “Does the classic microfinance model discourage entrepreneurship among the poor? Experimental evidence from In- dia,” American Economic Review, 103, 2196–2226. 33 Filmer, D. and L. Fox (2014): “Youth Employment in Sub-Saharan Africa,” The World Bank: Washington DC. Haushofer, J. and J. Shapiro (2016): “The Short-term Impact of Unconditional Cash Transfers to the Poor: Experimental Evidence from Kenya,” The Quarterly Jour- nal of Economics, 131, 1973–2042. Jayachandran, S. (2020): “Microentrepreneurship in Developing Countries,” NBER Working Paper January, Cambridge, MA. Karlan, D., A. L. Ratan, and J. Zinman (2014): “Savings by and for the Poor: A Research Review and Agenda,” Review of Income and Wealth, 60, 36–78. Karlan, D., B. Savonitto, B. Thuysbaert, and C. Udry (2017): “Impact of sav- ings groups on the lives of the poor,” Proceedings of the National Academy of Sciences, 114, 3079–3084. Ksoll, C., H. B. Lilleør, J. H. Lønborg, and O. D. Rasmussen (2016): “Impact of Village Savings and Loan Associations: Evidence from a cluster randomized trial,” Journal of Development Economics, 120, 70–85. Lowe, M. (2021): “Types of Contact: A Field Experiment on Collaborative and Adver- sarial Caste Integration,” American Economic Review, 111, 1807–44. Macours, K., P. Premand, and R. Vakis (2022): “Transfers, Diversification and Household Risk Strategies: Can Productive Safety Nets Help Households Manage Cli- matic Variability?” The Economic Journal, 132, 2438–2470. McKenzie, D. and C. Woodruff (2014): “What Are We Learning from Business Training and Entrepreneurship Evaluations around the Developing World?” The World Bank Research Observer, 29, 48–82. Meager, R. (2019): “Understanding the average impact of microcredit expansions: A bayesian hierarchical analysis of seven randomized experiments,” American Economic Journal: Applied Economics, 11, 57–91. Sedlmayr, R., A. Shah, and M. Sulaiman (2020): “Cash-plus: Poverty impacts of alternative transfer-based approaches,” Journal of Development Economics, 144, 102418. 34 Tables and Figures Figure 3: Experimental design 35 Figure 4: Cumulative distribution of self-employment earnings (total profits) Figure 5: Quantile treatment effects for self-employment earnings (total profits) 36 Figure 6: Cumulative distribution for total savings (stock) 37 Table 2: Employment and independent activities (1) (2) (3) (4) (5) (6) (7) Total Employment Self Wage Hours Hours hours (Has an employed employed # Independent worked in worked in worked activity of (at least (at least Activities self wage (last 7 any type) 1 activity) 1 activity) employment employment days) Panel A. Pooled Estimates Treatment (ITT) 0.01 0.02** -0.03** 0.24** 0.78 1.52 -1.10** (0.01) (0.01) (0.01) (0.10) (1.50) (1.17) (0.51) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Mean in Control 95% 91.6% 10.4% 3.13 40.08 10.59 3.08 Observations 2,620 2,620 2,620 2,620 2,620 2,620 2,620 Panel B. Treatment Arm Estimates 38 VSLA (T1) (ITT) 0.01 0.02 -0.01 0.21* 2.35 2.01 -0.44 (0.01) (0.01) (0.02) (0.11) (1.88) (1.42) (0.61) Cash Grant with repayment (T2) (ITT) 0.02 0.02* -0.04** 0.30*** 0.09 1.12 -1.50*** (0.01) (0.01) (0.02) (0.11) (1.68) (1.36) (0.54) Cash Grant (T3) (ITT) 0.01 0.03*** -0.04** 0.16 -0.89 1.40 -1.54** (0.01) (0.01) (0.02) (0.18) (1.82) (1.54) (0.67) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Mean in Control 95% 91.6% 10.4% 3.13 40.08 10.59 3.08 p-value T1=T2 0.65 0.69 0.10 0.41 0.20 0.51 0.03 p-value T2=T3 0.84 0.40 0.93 0.41 0.56 0.85 0.95 p-value T1=T3 0.80 0.22 0.14 0.77 0.09 0.70 0.07 p-value T1=T2=T3 0.90 0.44 0.19 0.60 0.22 0.80 0.06 Observations 2,620 2,620 2,620 2,620 2,620 2,620 2,620 Robust standard errors clustered at locality level. Hours worked are winsorized at the 99th percentile. Hours (columns 5-7) are measured for the last 7 days for the three main activities of the individual. The number of independent activities per individual is based on all activities operating in the last 12 months. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table 3: Earnings (1) (2) (3) (4) Revenues from Revenues from Earnings from Earnings from farming livestock wage employment self employment activities activities (monthly) (Profits, monthly) (last 12 mths) (last 12 mths) Panel A. Pooled Estimates Treatment (ITT) -154.47 1726.78 19909.81 -2091.02 (700.90) (1777.12) (17808.79) (1544.53) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in Control 3057.36 24050.42 124989.51 8013.67 Observations 2,615 2,620 2,620 2,620 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 230.73 859.10 26657.37 -685.07 39 (920.89) (2372.50) (24159.71) (1990.21) Cash Grant with repayment (T2) (ITT) -170.29 2393.22 12659.68 -2162.26 (818.13) (2217.70) (19719.88) (1609.86) Cash Grant (T3) (ITT) -905.88 2000.59 22505.26 -4839.52*** (807.37) (3098.32) (28995.56) (1546.04) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in Control 3057.36 24050.42 124989.51 8013.67 p-value T1=T2 0.67 0.57 0.55 0.36 p-value T2=T3 0.34 0.91 0.74 0.02 p-value T1=T3 0.19 0.74 0.89 0.01 p-value T1=T2=T3 0.38 0.85 0.83 0.01 Observations 2,615 2,620 2,620 2,620 Robust standard errors clustered at locality level. Earnings and revenues are in CFA franc and winsorized at 99%. Earnings are computed over the three main activities of the respondent. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table 4: Welfare indicators (1) (2) (3) (4) Food Psychological # household consumption Education well being durable goods score expenditures index owned (food security) (z-score) Panel A. Pooled Estimates Treatment (ITT) 0.79 0.03 -5388.59 0.07 (0.99) (0.17) (3410.01) (0.06) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in Control 52.57 3.69 48798.88 -0.00 Observations 2,618 2,618 2,617 2,618 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.22 0.04 -2669.95 0.03 (1.16) (0.20) (3816.59) (0.07) Cash Grant with repayment (T2) (ITT) 1.01 -0.05 -6082.24 0.07 (1.19) (0.20) (3974.22) (0.06) Cash Grant (T3) (ITT) 1.48 0.17 -9409.15** 0.14 (1.70) (0.25) (4559.82) (0.10) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in Control 52.57 3.69 48798.88 -0.00 p-value T1=T2 0.50 0.67 0.33 0.50 p-value T2=T3 0.79 0.39 0.43 0.46 p-value T1=T3 0.47 0.62 0.11 0.26 p-value T1=T2=T3 0.69 0.69 0.26 0.50 Observations 2,618 2,618 2,617 2,618 Robust standard errors clustered at locality level. Expenditures are in CFA franc. Expenditures and number of goods are winsorized at 99%. The Food consumption score (FCS) provides a measure of food security that captures the dietary diversity. It is commonly used by humanitarian organizations such as the World Food Programme. The Psychological well-being index includes the CESD Positive Affect scale, the Rosenberg self-esteem scale and a self-assessment of life satisfaction based on a Cantril ladder. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 40 Table 5: Investments and productive assets (1) (2) Total start-up Value of capital productive assets (all activities) (all activities) Panel A. Pooled Estimates Treatment (ITT) 12462.86*** 9320.74*** (2355.42) (3097.54) PDS Lasso selected controls Yes Yes Department X (Urban/Rural) Yes Yes Mean in Control 15260.21 39538.50 Observations 2,620 2,620 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 7259.34** 9613.97*** 41 (2907.06) (3674.09) Cash Grant with repayment (T2) (ITT) 14828.19*** 7788.24** (3066.31) (3755.91) Cash Grant (T3) (ITT) 17812.93*** 12244.49* (5341.51) (6243.08) PDS Lasso selected controls Yes Yes Department X (Urban/Rural) Yes Yes Mean in Control 15260.21 39538.50 p-value T1=T2 0.04 0.64 p-value T2=T3 0.60 0.50 p-value T1=T3 0.07 0.68 p-value T1=T2=T3 0.06 0.77 Observations 2,620 2,620 Robust standard errors clustered at locality level. Start-up capital and value of assets are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table 6: Savings (1) (2) (3) (4) (5) (6) Participate Participate Savings Savings Savings Has Saved in a VSLA in a ROSCA stock stock : stock : (last 2 yrs) (currently) (currently) (Total) VSLA ROSCA Panel A. Pooled Estimates Treatment (ITT) 0.05** 0.19*** -0.04 5713.39 7615.26** -603.14 (0.02) (0.03) (0.03) (10699.54) (3180.46) (9444.95) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 81.8% 16.8% 30.3% 59410.76 4983.39 29913.08 Observations 2,620 2,620 2,620 2,618 2,620 2,618 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.06** 0.37*** -0.12*** -11124.12 9504.65*** -16469.79* 42 (0.03) (0.04) (0.03) (10329.50) (2601.07) (8907.05) Cash Grant with repayment (T2) (ITT) 0.03 0.07** 0.01 11832.10 8143.30 4169.25 (0.03) (0.03) (0.03) (13903.51) (5257.53) (11747.82) Cash Grant (T3) (ITT) 0.07** 0.07 0.02 26693.63* 2496.31 21465.64* (0.03) (0.05) (0.04) (14127.84) (2434.49) (12985.60) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 81.8% 16.8% 30.3% 59410.76 4983.39 29913.08 p-value T1=T2 0.19 0.00 0.00 0.04 0.76 0.02 p-value T2=T3 0.20 0.92 0.74 0.34 0.22 0.20 p-value T1=T3 0.85 0.00 0.00 0.00 0.00 0.00 p-value T1=T2=T3 0.30 0.00 0.00 0.00 0.01 0.00 Observations 2,620 2,620 2,620 2,618 2,620 2,618 Robust standard errors clustered at locality level. Savings stock is in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table 7: Local spillovers on employment and independent activities (1) (2) (3) (4) (5) (6) (7) Total Employment Self Wage Hours Hours hours (Has an employed employed # Independent worked in worked in worked activity of (at least (at least activities self wage (last 7 any type) 1 activity) 1 activity) employment employment days) Panel A. Pooled Estimates Treatment (ITT) 0.01 0.01 -0.01 0.37** -2.10 -0.60 0.19 (0.01) (0.02) (0.02) (0.15) (1.88) (1.37) (0.73) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Mean in Control 97.9% 93.2% 12.4% 3.23 44.11 9.32 2.83 Observations 1,102 1,102 1,102 1,102 1,102 1,102 1,102 Panel B. Treatment Arm Estimates 43 VSLA (T1) (ITT) -0.00 0.00 0.01 0.40** -2.54 -2.76 0.44 (0.01) (0.03) (0.03) (0.18) (2.29) (1.67) (0.96) Cash Grant with repayment (T2) (ITT) 0.01 0.02 -0.03 0.29 -2.60 0.93 -0.24 (0.01) (0.02) (0.03) (0.19) (2.33) (1.83) (0.91) Cash Grant (T3) (ITT) 0.02* 0.01 -0.01 0.50** 0.09 0.01 0.75 (0.01) (0.02) (0.04) (0.22) (2.44) (1.43) (1.30) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Mean in Control 97.9% 93.2% 12.4% 3.23 44.11 9.32 2.83 p-value T1=T2 0.35 0.56 0.20 0.57 0.98 0.06 0.54 p-value T2=T3 0.48 0.77 0.61 0.37 0.29 0.61 0.47 p-value T1=T3 0.17 0.77 0.61 0.66 0.30 0.08 0.82 p-value T1=T2=T3 0.38 0.84 0.44 0.66 0.48 0.12 0.73 Observations 1,102 1,102 1,102 1,102 1,102 1,102 1,102 Robust standard errors clustered at locality level. Hours worked are winsorized at the 99th percentile. Hours (columns 5-7) are measured for the last 7 days for the three main activities of the individual. The number of independent activities per individual is based on all activities operating in the last 12 months. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A6. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table 8: Local spillovers on earnings (1) (2) Earnings in Earnings in wage Employment self Employment (monthly) (Profits, monthly) Panel A. Pooled Estimates Treatment (ITT) -1984.77 1141.09 (1734.89) (3964.52) PDS Lasso selected controls Yes Yes Department X (Urban/Rural) Yes Yes Mean in Control 6382.68 32513.35 Observations 1,099 1,102 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) -1389.97 1347.10 (2032.09) (5791.10) 44 Cash Grant with repayment (T2) (ITT) -2773.07 969.78 (1744.84) (4537.98) Cash Grant (T3) (ITT) -1232.74 1148.59 (2437.79) (6034.99) PDS Lasso selected controls Yes Yes Department X (Urban/Rural) Yes Yes Mean in Control 6382.68 32513.35 p-value T1=T2 0.37 0.95 p-value T2=T3 0.41 0.98 p-value T1=T3 0.94 0.98 p-value T1=T2=T3 0.57 1.00 Observations 1,098 1,102 Robust standard errors clustered at locality level. Earnings and revenues are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A6. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table 9: Local spillovers on investments and productive assets (1) (2) Total start-up Value of capital productive assets (all activities) (all activities) Panel A. Pooled Estimates Treatment (ITT) 1091.27 10762.98 (5120.99) (7751.74) PDS Lasso selected controls Yes Yes Department X (Urban/Rural) Yes Yes Mean in Control 25358.29 59658.83 Observations 1,102 1,102 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) -3209.52 6500.30 (5080.75) (9057.50) 45 Cash Grant with repayment (T2) (ITT) 3823.40 12489.14 (6901.26) (9932.08) Cash Grant (T3) (ITT) 3055.09 15203.18 (7695.99) (9364.66) PDS Lasso selected controls Yes Yes Department X (Urban/Rural) Yes Yes Mean in Control 25358.29 59658.83 p-value T1=T2 0.27 0.54 p-value T2=T3 0.93 0.79 p-value T1=T3 0.38 0.35 p-value T1=T2=T3 0.44 0.63 Observations 1,102 1,102 Robust standard errors clustered at locality level. Start-up capital and value of assets are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A6. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table 10: Local spillovers on savings (1) (2) (3) (4) (5) (6) Participate Participate Savings Savings Savings Has Saved in a VSLA in a ROSCA stock stock : stock : (last 2 yrs) (currently) (currently) (Total) VSLA ROSCA Panel A. Pooled Estimates Treatment (ITT) 0.06** 0.08** -0.04 -10986.98 4770.73*** -7858.09 (0.02) (0.03) (0.03) (14747.08) (1502.82) (10043.35) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 81.8% 14.4% 28.1% 99043.92 3539.34 42099.05 Observations 1,102 1,102 1,102 1,101 1,101 1,102 Panel B. Treatment Arm Estimates 46 VSLA (T1) (ITT) 0.06* 0.10** -0.02 -14206.90 4553.32** -11517.95 (0.03) (0.05) (0.04) (16351.49) (2134.31) (11145.30) Cash Grant with repayment (T2) (ITT) 0.05 0.04 -0.04 -22848.34 4737.46** -18850.16** (0.03) (0.04) (0.03) (15239.54) (1914.75) (9355.21) Cash Grant (T3) (ITT) 0.08*** 0.12** -0.09** 25366.19 5298.25* 27268.92 (0.03) (0.05) (0.04) (28454.78) (2823.74) (25458.08) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 81.8% 14.4% 28.1% 99043.92 3539.34 42099.05 p-value T1=T2 0.71 0.19 0.62 0.51 0.94 0.44 p-value T2=T3 0.28 0.14 0.24 0.08 0.85 0.06 p-value T1=T3 0.45 0.80 0.11 0.16 0.82 0.14 p-value T1=T2=T3 0.53 0.23 0.25 0.20 0.97 0.14 Observations 1,102 1,102 1,102 1,101 1,101 1,102 Robust standard errors clustered at locality level. Savings stock amounts are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A6. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table 11: Social outcomes (1) (2) (3) (4) (5) (6) (7) (8) Participation in Participation in Victimization Insecurity Participation in # times received # times provided Trust community works social activities (conflict) (perception) groups or associations financial support financially support Index (# times in (# times in Index Index (# groups) (last 12 mths) (last 12 mths) (z-score) last 12 mths) last 12 mths) (z-score) (z-score) Panel A. Pooled Estimates Treatment (ITT) 0.16*** 0.21** 0.29** 0.03 0.15 0.03 -0.04 0.04 (0.05) (0.10) (0.12) (0.03) (0.36) (0.06) (0.04) (0.06) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Yes Mean in Control 1.19 0.88 1.31 61.10 6.59 0.00 0.00 0.00 Observations 2,620 2,620 2,620 2,620 2,620 2,374 2,614 2,617 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.18*** 0.15 0.27* 0.04 0.28 0.05 -0.01 0.03 (0.07) (0.13) (0.14) (0.04) (0.45) (0.08) (0.06) (0.07) 47 Cash Grant with repayment (T2) (ITT) 0.14** 0.12 0.31** 0.03 -0.27 0.00 -0.05 0.03 (0.07) (0.11) (0.14) (0.03) (0.41) (0.07) (0.05) (0.07) Cash Grant (T3) (ITT) 0.16** 0.53** 0.28 0.01 0.84 0.04 -0.10* 0.07 (0.08) (0.23) (0.19) (0.04) (0.51) (0.11) (0.05) (0.11) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Yes Mean in Control 1.19 0.88 1.31 61.10 6.59 0.00 0.00 0.00 p-value T1=T2 0.59 0.84 0.78 0.69 0.20 0.53 0.58 0.99 p-value T2=T3 0.81 0.09 0.90 0.64 0.03 0.72 0.35 0.70 p-value T1=T3 0.82 0.13 0.94 0.45 0.28 0.93 0.17 0.71 p-value T1=T2=T3 0.87 0.23 0.96 0.75 0.07 0.81 0.37 0.92 Observations 2,620 2,620 2,620 2,620 2,620 2,374 2,614 2,617 Robust standard errors clustered at locality level. Z-scores are centered in the control group. The Trust index is based on 16 variables measuring different forms of trust (economic, financial, security) and trust in different types of groups or institutions (neighbors, same and other ethnic groups, leaders, foreigners). The Victimization index is based on 2 variables, measuring if the respondent reports having been victim of an attack or involved in a conflict over the past 12 months. The Insecurity (perception) index is based on 6 variables, measuring the insecurity perception of the respondent, including the perceived likelihood of a robbery, attack or shootings, a subjective assessment of the peacefulness of the locality, fear of conflicts, and knowledge of past conflicts in the locality. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Appendix A1 Additional tables 48 Table A1: Costs by treatment arm VSLA Cash Grant Cash Grant Costs per beneficiary with repayment (USD nominal 2014) (T1) (T2) (T3) Training (entrepreneurship and life skills) 32.1 32.1 32.1 Facilitation of Village Savings Loan Associations 78.5 0.0 0.0 (VSLAs) Cash Grants 0.0 155.1 160.4 Repayment of Cash Grants 0.0 -31.0 0.0 Community outreach and targeting 16.0 16.0 16.0 Field implementation (supervision, equipment, 70.5 70.5 70.5 operating costs) Overall project coordination 26.2 26.2 26.2 Total cost per beneficiary 223.4 269.0 305.2 Table A2: Composition of study sample VSLA Cash grant with repayment Cash Grant Control Group Total (T1) (T2) (T3) Number of localities 53 64 30 60 207 Among which rural 43 52 24 49 168 Total Endline sample (main and spillover) 1,297 1,431 825 1,667 5,220 Endline sample (main) restricted to common cut-off 768 919 418 831 2,936 Endline sample (spillover) restricted to common cut-off 426 512 209 836 1,983 Table A3: Vulnerability score (used by the program for targeting) Criteria Weights Score Vulnerability categories 20% =1 if belongs to any category listed - Did not complete primary school - Single mother - Not working - Has a disability Level and source of income A) Employment instability 5% =1 if belongs to any category listed Based on the following employment status : - Family worker or domestic or intern/apprentice ; - Employed with piecework pay or daily pay ; - Not working B) Level of resources (sum of labor incomes and 20% (*) = (B ) money transfers last 30 days) B Financial dependence of the household 10% = C + (1 − D) - Ratio of people contributing to household’s expenses to total household’s members (C ) - Share of financially dependent people in the household (D) Living Standards of the household - Livestock 10% (*) = Livestock heads - Productive assets 10% (*) = Nb of plows + field sprayer + carts + wheelbarrows - Transportation assets 10% (*) = Nb of bikes + motorcycles + cars - Household durables 10% (*) = Nb of fridge + fans + TV - Nb of rooms in the house/Nb of people in the household (E ) 5% (*) = (E ) (*) Indicates that this item enters negatively in the score (i.e. a high score on this item decreases the total vulnerability score) 49 Table A4: Baseline comparison of selected and non-selected individuals (treated localities) (1) (2) (3) (4) (5) Mean Mean p-value among among All Nb. Obs. of test selected non-selected applicants (2)-(3) individuals individuals Personal and Household characteristics Type of locality (1 = village) 9042 0.799 0.803 0.800 0.755 Rural Area 9042 0.764 0.780 0.771 0.240 Female 9042 0.708 0.507 0.621 0.000 Age 9042 35.277 31.735 33.739 0.000 Single mother 9042 0.012 0.005 0.009 0.010 Disabled 9042 0.053 0.029 0.043 0.000 Married 9041 0.285 0.379 0.326 0.000 Lives with a partner (but not married) 9041 0.193 0.252 0.219 0.000 Widowed 9041 0.296 0.106 0.214 0.000 Nb of children 9042 3.622 2.988 3.347 0.000 Has been to school 9042 0.428 0.620 0.511 0.000 No diploma 9037 0.928 0.629 0.798 0.000 Head of the Household 9041 0.567 0.488 0.533 0.000 Spouse of the head of the household 9041 0.250 0.288 0.266 0.007 Household and productive assets Nb of livestock 9042 0.628 1.872 1.168 0.000 Nb of poultry 9042 1.816 4.979 3.190 0.000 Nb of agricultural tools 9042 0.087 0.417 0.231 0.000 Employment Has an activity (last 7 days) 9042 0.945 0.986 0.963 0.000 Main activity is wage work (last 7 days) 8695 0.088 0.077 0.083 0.223 Main activity is non-ag self-empl. 9042 0.250 0.259 0.254 0.443 Main activity is agricultural self-empl. 8695 0.512 0.591 0.547 0.000 Total income (main activity, last month) 9042 22877.624 55339.370 36972.383 0.000 Has a non-ag business 9042 0.719 0.788 0.749 0.000 Savings and Credits Has saved (last 3 months) 6754 0.142 0.127 0.135 0.146 Amount saved (last 3 months) 9040 15297.336 37782.490 25059.972 0.010 Has a mobile money account 9039 0.150 0.275 0.205 0.000 Has a bank account 9039 0.012 0.025 0.018 0.000 Has participated in a ROSCA (Tontine) 9039 0.507 0.500 0.504 0.639 Has debt 9042 0.230 0.220 0.226 0.350 No accounting 9042 0.488 0.609 0.541 0.000 Financial constraints Report strong binding constraints for 9042 0.631 0.590 0.613 0.005 education expenditures Report strong binding constraints for health 9041 0.704 0.659 0.684 0.001 expenditures 50 Table A5: Baseline balance and descriptive statistics (main sample) (1) (2) (3) (4) (5) (6) (7) (8) Mean in Mean in p-value Mean in p-value of treatment Mean in Mean in Obs. control of test Cash Grant joint test group VSLA Cash Grant group (3)=(2) with Repayment (T1=T2=T3) (pooled) Individual and Household characteristics Type of locality (1 = village) 2936 0.806 0.802 0.945 0.803 0.812 0.778 0.935 Rural Area 2936 0.801 0.785 0.780 0.759 0.819 0.756 0.644 Female 2936 0.714 0.703 0.650 0.707 0.702 0.696 0.955 Age 2936 35.208 34.908 0.632 34.479 35.234 34.980 0.584 Youth (< 35 years-old) 2936 0.562 0.565 0.888 0.565 0.569 0.557 0.950 Single mother 2936 0.023 0.011 0.085 0.007 0.011 0.022 0.075 Disabled 2936 0.082 0.052 0.023 0.052 0.062 0.031 0.047 Native ethnic group 2936 0.834 0.812 0.559 0.784 0.839 0.806 0.587 Married 2935 0.266 0.308 0.276 0.331 0.282 0.324 0.537 Lives with a partner (but not married) 2935 0.182 0.190 0.789 0.207 0.190 0.156 0.505 Widowed 2935 0.308 0.285 0.429 0.257 0.303 0.300 0.341 Nb of children 2936 3.575 3.603 0.826 3.574 3.617 3.627 0.946 Has been to school 2936 0.403 0.394 0.798 0.392 0.411 0.359 0.504 No diploma 2935 0.939 0.942 0.809 0.939 0.939 0.955 0.580 Head of the Household 2935 0.589 0.562 0.358 0.539 0.581 0.560 0.555 Spouse of the head of the household 2935 0.241 0.267 0.336 0.305 0.247 0.242 0.175 Nb of other household members 2936 5.304 5.455 0.432 5.275 5.497 5.691 0.350 Assets Nb of livestock 2936 0.543 0.594 0.617 0.604 0.579 0.608 0.976 Nb of poultry 2936 1.646 1.724 0.733 1.975 1.518 1.713 0.314 Nb of agricultural tools 2936 0.049 0.089 0.022 0.089 0.090 0.089 0.997 Employment Has an activity (last 7 days) 2936 0.913 0.946 0.071 0.943 0.938 0.969 0.115 Main activity is non-ag self-empl. 2936 0.236 0.223 0.713 0.224 0.214 0.242 0.859 Main activity is agricultural self-empl. 2745 0.532 0.539 0.880 0.535 0.550 0.523 0.917 Main activity is wage work 2745 0.077 0.083 0.735 0.055 0.098 0.099 0.038 Total income (main activity, last month) 2936 19124.398 21206.601 0.324 20741.276 21659.456 21065.921 0.945 Has a non-ag business 2936 0.717 0.734 0.608 0.725 0.723 0.775 0.379 Savings and Credit Has saved (last 3 months) 2133 0.157 0.159 0.925 0.161 0.172 0.130 0.488 Amount saved (last 3 months) 2935 12041.145 17955.736 0.281 13215.234 23277.503 14965.311 0.661 Has a mobile money account 2936 0.132 0.130 0.932 0.139 0.135 0.103 0.442 Has a bank account 2935 0.013 0.009 0.347 0.010 0.009 0.007 0.841 Has participated in a Tontine (ROSCA) 2934 0.489 0.482 0.836 0.440 0.502 0.516 0.210 Has debt 2936 0.244 0.227 0.541 0.228 0.242 0.194 0.483 Does not do accounting 2936 0.489 0.484 0.893 0.479 0.474 0.514 0.736 Financial constraints Reports constraints for education 2936 0.605 0.600 0.858 0.595 0.607 0.593 0.932 expenditures Reports constraints for health expenditures 2936 0.691 0.706 0.589 0.688 0.709 0.732 0.550 51 Table A6: Baseline balance and descriptive statistics (spillover sample) (1) (2) (3) (4) (5) (6) (7) (8) Mean in p-value Mean in p-value Mean in treatment Mean in Mean in of joint Obs. control of test Cash Grant group VSLA Cash Grant test group (3)=(2) with Repayment (pooled) (T1=T2=T3) Individual and Household characteristics Type of locality (1 = village) 1200 1.000 1.000 . 1.000 1.000 1.000 . Rural Area 1200 0.958 0.939 0.556 0.933 0.967 0.909 0.481 Female 1200 0.406 0.403 0.913 0.408 0.417 0.378 0.737 Age 1200 31.229 30.673 0.280 30.983 30.631 30.332 0.695 Youth (< 35 years-old) 1200 0.717 0.733 0.588 0.715 0.736 0.751 0.740 Single mother 1200 0.002 0.005 0.429 0.000 0.004 0.014 0.229 Disabled 1200 0.018 0.019 0.925 0.015 0.014 0.029 0.453 Native ethnic group 1200 0.817 0.824 0.877 0.831 0.848 0.785 0.719 Married 1200 0.422 0.427 0.910 0.423 0.438 0.416 0.923 Lives with a partner (but not married) 1200 0.263 0.259 0.920 0.255 0.279 0.239 0.790 Widowed 1200 0.083 0.061 0.214 0.082 0.043 0.057 0.301 Nb of children 1200 2.951 2.802 0.315 2.846 2.848 2.684 0.644 Has been to school 1200 0.629 0.637 0.876 0.596 0.638 0.689 0.451 No diploma 1198 0.611 0.557 0.210 0.603 0.545 0.512 0.378 Head of the Household 1200 0.574 0.509 0.050 0.502 0.475 0.565 0.239 Spouse of the head of the household 1200 0.250 0.261 0.743 0.270 0.293 0.206 0.175 Nb of other household members 1200 5.348 5.830 0.093 5.708 5.851 5.957 0.881 Assets Nb of livestock 1200 2.344 2.604 0.570 2.779 2.388 2.665 0.794 Nb of poultry 1200 5.587 5.489 0.922 4.970 5.293 6.411 0.728 Nb of agricultural tools 1200 0.438 0.580 0.068 0.599 0.547 0.598 0.874 Employment Has an activity (last 7 days) 1200 0.993 0.996 0.617 0.996 0.993 1.000 0.215 Main activity is non-ag self-empl. 1200 0.185 0.191 0.863 0.161 0.210 0.206 0.564 Main activity is agricultural self-empl. 1193 0.706 0.703 0.958 0.743 0.675 0.689 0.519 Main activity is wage work 1193 0.040 0.044 0.780 0.030 0.069 0.029 0.205 Total income (main activity, last month) 1200 60978.292 73048.271 0.118 72373.034 67857.609 80765.550 0.565 Has a non-ag business 1200 0.817 0.836 0.511 0.828 0.833 0.852 0.817 Savings and Credit Has saved (last 3 months) 994 0.156 0.127 0.364 0.186 0.104 0.084 0.036 Amount saved (last 3 months) 1198 27752.242 65450.133 0.286 33859.551 119011.775 35075.359 0.671 Has a mobile money account 1200 0.196 0.262 0.083 0.221 0.304 0.258 0.355 Has a bank account 1200 0.025 0.025 0.942 0.019 0.029 0.029 0.653 Has participated in a ROSCA (Tontine) 1200 0.460 0.513 0.216 0.506 0.529 0.502 0.877 Has debt 1200 0.219 0.222 0.923 0.206 0.243 0.215 0.748 No accounting 1200 0.634 0.649 0.717 0.648 0.627 0.679 0.713 Financial constraints Reports constraints for education 1200 0.509 0.537 0.489 0.566 0.525 0.517 0.677 expenditures Reports constraints for health expenditures 1200 0.627 0.668 0.362 0.674 0.688 0.632 0.665 52 Table A7: Study sample and general population in program districts (1) (2) Population 18-65 PRISE in program participants districts (2015-2016) (2013-2014) Age 34.731 36.009 Female 0.481 0.734 Urban 0.163 0.181 Nb of household members 4.729 6.346 No diploma 0.590 0.932 Employment (Has an activity of any type) 0.872 0.904 Self-employed in agriculture (at least 1 activity) 0.679 0.610 Self-employed non-agr. (at least 1 activity) 0.207 0.285 Wage employed (at least 1 activity) 0.115 0.104 HH owns livestock 0.212 0.525 Asset Index (z-score) -0.000 0.001 Asset Index (with livestock, z-score) 0.000 0.018 Sample 3183 743 Column 1 is based on a national employment survey (ENSETE 2013), restricted to the relevant Western districts of Montagnes and Waroba. Column (2) is based on the control group at baseline (except the asset index and livestock ownership, which are from endline). Table A8: Program Take-up VSLA Cash grant with repayment Cash Grant (T1) (T2) (T3) Take up for financial support (*) 69.5% 78.9% 81.1% Training 1 : Entrepreneurship 1 (“starting an activity”) 64.7% 88.4% 91.5% and Peace Building Business plan prepared n. a. 82.4% 84.2% Business plan reviewed and approved n.a. 80.9% 82.0% Training 2 : Entrepreneurship 2 (“managing an activity”) 59.8% 61.8% 64.1% and Life Skills Note : Based on monitoring data. Participation rates are unconditional (i.e. computed over all selected beneficiaries, even if some activities were conditional, e.g. conditional on business plan approval). (*) For VSLA intervention, this means joining a VSLA. For other interventions, this means receiving a cash grant. 53 Table A9: Multiple Hypothesis Testing Corrections (for Tables 1-5) (1) (2) (3) (4) (5) (6) (7) (8) Treatment VSLA (T1) Cash Grant with repayment (T2) Cash Grant (T3) Actual p-value FDR Actual p-value FDR Actual p-value FDR Actual p-value FDR Table 2: Employment and independent activities Employment (has an activity of any type) 0.156 0.218 0.351 0.491 0.152 0.213 0.231 0.405 Self-employed (at least 1 activity) 0.046 0.080 0.265 0.464 0.096 0.167 0.008 0.055 Wage employed (at least 1 activity) 0.042 0.080 0.494 0.494 0.014 0.032 0.030 0.070 # Independent activities 0.016 0.080 0.071 0.464 0.009 0.030 0.386 0.450 Total hours worked (last 7 days) 0.603 0.603 0.213 0.464 0.959 0.959 0.627 0.627 Hours worked in self-employment 0.195 0.228 0.160 0.464 0.412 0.481 0.363 0.450 Hours worked in wage employment 0.033 0.080 0.473 0.494 0.006 0.030 0.023 0.070 Table 3: Earnings 54 Earnings from wage employment (month) 0.826 0.826 0.802 0.802 0.835 0.835 0.263 0.519 Earnings from self-employment (profits, month) 0.332 0.443 0.718 0.802 0.282 0.564 0.519 0.519 Revenues from farming activ. (last 12 mths) 0.265 0.443 0.271 0.802 0.522 0.695 0.439 0.519 Revenues from livestock activ. (last 12 mths) 0.177 0.443 0.731 0.802 0.181 0.564 0.002 0.008 Table 4: Welfare indicators Food consumption score 0.428 0.571 0.853 0.853 0.399 0.532 0.385 0.494 # household durable assets owned 0.876 0.876 0.843 0.853 0.806 0.806 0.494 0.494 Education expenditures 0.116 0.462 0.485 0.853 0.127 0.510 0.040 0.161 Psychological well-being index (z-score) 0.247 0.493 0.641 0.853 0.285 0.532 0.161 0.321 Table 5: Investments and productive assets Total start-up capital (all activities) 0.000 0.000 0.013 0.013 0.000 0.000 0.001 0.002 Value of productive assets (all activities) 0.003 0.003 0.010 0.013 0.039 0.039 0.051 0.051 Table A9: Multiple Hypothesis Testing Corrections (for Tables 6-9) (1) (2) (3) (4) (5) (6) (7) (8) Treatment VSLA (T1) Cash Grant with repayment (T2) Cash Grant (T3) Actual p-value FDR Actual p-value FDR Actual p-value FDR Actual p-value FDR Table 6: Savings Has Saved (last 2 yrs) 0.023 0.047 0.015 0.023 0.233 0.467 0.026 0.155 Participate in a VSLA (currently) 0.000 0.000 0.000 0.000 0.041 0.244 0.176 0.265 Participate in a ROSCA (currently) 0.206 0.309 0.000 0.000 0.768 0.768 0.599 0.599 Savings stock (total) 0.594 0.713 0.283 0.283 0.396 0.594 0.060 0.181 Savings stock : VSLA 0.018 0.047 0.000 0.001 0.123 0.369 0.306 0.368 Savings stock : ROSCA 0.949 0.949 0.066 0.079 0.723 0.768 0.100 0.200 Table 7: Local spillovers on employment and independent activities 55 Employment (has an activity of any type) 0.353 0.688 0.934 0.934 0.188 0.370 0.075 0.261 Self-employed (at least 1 activity) 0.435 0.688 0.877 0.934 0.248 0.370 0.586 0.988 Wage employed (at least 1 activity) 0.492 0.688 0.820 0.934 0.213 0.370 0.706 0.988 # Independent activities 0.014 0.095 0.028 0.196 0.126 0.370 0.025 0.172 Total hours worked (last 7 days) 0.266 0.688 0.269 0.629 0.264 0.370 0.969 0.992 Hours worked in self-employment 0.664 0.775 0.100 0.351 0.612 0.714 0.992 0.992 Hours worked in wage employment 0.793 0.793 0.646 0.934 0.794 0.794 0.562 0.988 Table 8: Local spillovers on earnings Earnings from wage employment (month) 0.254 0.508 0.495 0.816 0.114 0.228 0.614 0.849 Earnings in self-employment (profits, month) 0.774 0.774 0.816 0.816 0.831 0.831 0.849 0.849 Table 9: Local spillovers on investments and productive assets Total start-up capital (all activities) 0.832 0.832 0.528 0.528 0.580 0.580 0.692 0.692 Value of productive assets (all activities) 0.167 0.334 0.474 0.528 0.210 0.421 0.106 0.213 Table A9: Multiple Hypothesis Testing Corrections (for Tables 10-11) (1) (2) (3) (4) (5) (6) (7) (8) Treatment VSLA (T1) Cash Grant with repayment (T2) Cash Grant (T3) Actual p-value FDR Actual p-value FDR Actual p-value FDR Actual p-value FDR Table 10: Local spillovers on savings Has Saved (last 2 yrs) 0.167 0.334 0.474 0.528 0.210 0.421 0.106 0.213 Participate in a VSLA (currently) 0.019 0.038 0.052 0.104 0.136 0.204 0.004 0.024 Participate in a ROSCA (currently) 0.018 0.038 0.023 0.103 0.286 0.286 0.024 0.047 Savings stock (total) 0.131 0.197 0.565 0.565 0.223 0.268 0.015 0.046 Savings stock : VSLA 0.457 0.457 0.386 0.463 0.136 0.204 0.374 0.374 Savings stock : ROSCA 0.002 0.011 0.034 0.103 0.014 0.086 0.062 0.094 56 Table 11: Social outcomes Participation in groups 0.435 0.457 0.303 0.454 0.046 0.137 0.286 0.343 Received financial support 0.004 0.035 0.008 0.062 0.036 0.142 0.048 0.156 Provided financial support 0.035 0.095 0.249 0.540 0.289 0.580 0.023 0.156 Participation in community works 0.013 0.053 0.055 0.221 0.027 0.142 0.144 0.230 Participation in social activities 0.299 0.479 0.270 0.540 0.363 0.580 0.710 0.710 Trust Index (z-score) 0.682 0.682 0.541 0.721 0.509 0.679 0.100 0.200 Victimization Index (z-score) 0.642 0.682 0.512 0.721 0.963 0.963 0.701 0.710 Insecurity (perception) index (z-score) 0.293 0.479 0.821 0.821 0.358 0.580 0.058 0.156 0.502 0.669 0.684 0.781 0.664 0.758 0.498 0.664 Columns labeled ‘Actual’ show the p-values from Tables 2-11. Columns labeled ‘FDR’ show False Discovery Rate-adjusted q-values following the step-up approach of Benjamini and Hochberg which assumes that the p-values within a family are positively correlated. FDR corrects p-values within each treatment arm, within each family of variables. Each table is considered a family of variables. Table A10: Diversification of self-employment between agriculture and non-agricultural activities (1) (2) (3) (4) (5) (6) Earnings in Earnings in Value of Value of # Non-Ag. # Agricultural Self Employment : Self Employment : non agricultural agricultural Independent Independent Non Ag. Activities Ag. Activities productive assets productive assets Activities Activities (Profits, monthly) (Profits, monthly) (all activities) (all activities) Panel A. Pooled Estimates Treatment (ITT) 0.08* 0.15 351.97 1223.17 5883.29** 3977.39** (0.04) (0.10) (1383.47) (1257.04) (2504.71) (1954.54) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 0.67 2.45 13155.21 10793.20 18394.25 20636.95 Observations 2,620 2,620 2,620 2,620 2,620 2,620 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.06 0.14 -14.67 681.05 7514.25** 2052.36 57 (0.06) (0.11) (2070.74) (1316.52) (3111.87) (2116.97) Cash Grant with repayment (T2) (ITT) 0.09* 0.21* 182.68 2139.82 4699.56 3935.57 (0.05) (0.12) (1468.91) (1737.63) (3068.88) (2437.31) Cash Grant (T3) (ITT) 0.10 0.05 1497.98 243.54 5235.69 8058.68 (0.07) (0.20) (2075.77) (2480.95) (4239.79) (5154.83) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 0.67 2.45 13155.21 10793.20 18394.25 20636.95 p-value T1=T2 0.62 0.50 0.92 0.40 0.40 0.46 p-value T2=T3 0.91 0.42 0.53 0.48 0.91 0.45 p-value T1=T3 0.62 0.67 0.55 0.86 0.61 0.27 p-value T1=T2=T3 0.85 0.66 0.79 0.66 0.68 0.47 Observations 2,620 2,620 2,620 2,620 2,620 2,620 Robust standard errors clustered at locality level. Earnings and value of assets are in CFA franc and winsorized at 99%. “Ag.” stands for Agricultural. The number of independent activities per individual is based on all activities operating in the last 12 months. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table A11: Agriculture (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) Total # days Area Production Sales worked in Livestock # assets owned : Is engaged Is engaged in Used fertilizer Has cultivated Has cultivated Hours for Nb of plots cultivated (cash crops) (cash crops) agriculture index agricultural in farming livestock (all types) cash crops food crops livestock (ha) (kg) (kg) (last 12 (TLU) equipment mths) Panel A. Pooled Estimates Treatment (ITT) 0.01 0.00 0.07 0.17 0.05 -0.00 0.03 4122.33 4126.05 -0.43 2.41 -0.08 0.29 (0.02) (0.03) (0.12) (0.25) (0.03) (0.03) (0.02) (4314.23) (4314.09) (0.31) (14.30) (0.06) (0.21) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean in Control 86% 52.5% 3.04 4.09 60.8% 61.7% 78.2% 238.45 230.48 2.25 230.36 0.39 5.26 Observations 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,618 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.00 -0.00 0.03 0.35 0.02 -0.00 0.02 -519.85 -517.17 -0.48 10.91 -0.11* 0.23 58 (0.02) (0.03) (0.14) (0.32) (0.04) (0.03) (0.03) (2565.06) (2564.68) (0.33) (14.94) (0.06) (0.25) Cash Grant with repayment (T2) (ITT) 0.01 0.01 0.21 0.20 0.05 -0.01 0.04* -982.10 -978.20 -0.11 -5.28 -0.03 0.31 (0.02) (0.03) (0.17) (0.29) (0.04) (0.04) (0.02) (2530.11) (2529.85) (0.38) (16.51) (0.08) (0.25) Cash Grant (T3) (ITT) 0.04** -0.02 -0.19 -0.26 0.08 0.04 0.03 25399.19 25404.65 -1.04*** 2.41 -0.15** 0.36 (0.02) (0.04) (0.21) (0.33) (0.05) (0.03) (0.03) (23497.67) (23497.81) (0.34) (21.57) (0.06) (0.32) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean in Control 86% 52.5% 3.04 4.09 60.8% 61.7% 78.2% 238.45 230.48 2.25 230.36 0.39 5.26 p-value T1=T2 0.71 0.57 0.25 0.66 0.43 0.85 0.41 0.85 0.85 0.19 0.21 0.28 0.73 p-value T2=T3 0.26 0.38 0.08 0.18 0.66 0.19 0.79 0.28 0.28 0.00 0.70 0.09 0.90 p-value T1=T3 0.16 0.71 0.30 0.10 0.28 0.17 0.60 0.28 0.28 0.04 0.65 0.40 0.70 p-value T1=T2=T3 0.31 0.67 0.20 0.22 0.52 0.30 0.70 0.56 0.56 0.01 0.45 0.23 0.91 Observations 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,618 Robust standard errors clustered at locality level. All variables are measured over the last 12 months. Surfaces (in ha) and volumes of production and sales (in kg) are winsorized at 99%. Cash crops include hevea, coffee, cocoa, cashew and cotton. Food crops include 39 types of fruits, vegetables, roots and condiments. Column 11 refers to farming only (livestock rearing excluded). The number of assets owned and the livestock headcounts are measured at the time of survey. The livestock index aggregates the number of heads across species (cows, pigs, goats, sheep, poultry and rabbits) in tropical livestock unit (TLU). The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table A12: Credit (1) (2) (3) (4) (5) (6) (7) (8) (9) Has taken Has taken Has taken Credits taken Has taken a credit from Has taken Credits taken Credits taken a credit from a credit from (all) Credits taken a credit oth. informal a credit from from formal from oth. formal sources a savings group (cumulative, from VSLA (last 2 yrs) sources a VSLA sources informal sources (last 2 yrs) (last 2 yrs) last 2 yrs) (last 2 years) Panel A. Pooled Estimates Treatment (ITT) 0.07*** -0.01 0.15*** -0.02 0.12*** 4204.38 -81.09 3863.41** 793.28 (0.02) (0.01) (0.03) (0.03) (0.03) (3116.68) (1228.21) (1536.26) (2394.82) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean in Control 57.2% 2.6% 10.1% 48.9% 8.9% 22736.24 2954.45 3730.34 16051.44 Observations 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 Panel B. Treatment Arm Estimates 59 VSLA (T1) (ITT) 0.13*** -0.01 0.31*** -0.05 0.26*** 7381.10* -461.29 9780.38*** -1493.25 (0.03) (0.01) (0.03) (0.03) (0.04) (3987.30) (1745.11) (2114.22) (2633.89) Cash Grant with repayment (T2) (ITT) 0.05* -0.00 0.04* 0.02 0.02 2156.18 -367.88 3.84 2775.77 (0.02) (0.01) (0.03) (0.03) (0.03) (3274.56) (1241.19) (1649.10) (2629.05) Cash Grant (T3) (ITT) -0.00 0.00 0.07* -0.03 0.04 2326.07 1361.56 430.83 976.75 (0.04) (0.01) (0.04) (0.05) (0.04) (5229.14) (2249.03) (1983.03) (3544.03) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean in Control 57.2% 2.6% 10.1% 48.9% 8.9% 22736.24 2954.45 3730.34 16051.44 p-value T1=T2 0.01 0.21 0.00 0.04 0.00 0.14 0.95 0.00 0.04 p-value T2=T3 0.23 0.56 0.55 0.26 0.60 0.97 0.43 0.84 0.57 p-value T1=T3 0.00 0.15 0.00 0.71 0.00 0.36 0.48 0.00 0.43 p-value T1=T2=T3 0.00 0.25 0.00 0.11 0.00 0.33 0.73 0.00 0.13 Observations 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 2,620 Robust standard errors clustered at locality level. Credit amounts are in CFA franc and winsorized at 99%. Formal credit sources are microcredit institutions, banks, agricultural cooperatives. Other informal credit sources are family / friends, informal lender / pawnbroker, credit from another business. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table A13: Business practices and knowledge (1) (2) (3) (4) (5) (6) (7) (8) Entrepreneurship Separate # employees # paid employees Index of Has done a knowledge Has developed Uses formal regular main main business practices market (quiz) a business plan book-keeping payments activity activity (z-score) assessment (z-score) to self Panel A. Pooled Estimates Treatment (ITT) 0.24 0.40 0.19*** 0.21*** 0.03** 0.02*** 0.03*** 0.04 (0.29) (0.26) (0.06) (0.06) (0.01) (0.01) (0.01) (0.03) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Yes Mean in Control 4.23 2.53 0.00 -0.00 5.9% 0.9% 4.1% 34.4% Observations 2,620 2,620 2,618 2,620 2,620 2,620 2,620 2,620 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.65* 0.73** 0.03 0.12* 0.00 0.01 0.02 0.03 60 (0.37) (0.32) (0.08) (0.07) (0.02) (0.01) (0.01) (0.03) Cash Grant with repayment (T2) (ITT) 0.33 0.42 0.27*** 0.36*** 0.07*** 0.02*** 0.04*** 0.07** (0.35) (0.34) (0.07) (0.07) (0.02) (0.01) (0.01) (0.03) Cash Grant (T3) (ITT) -0.80* -0.32 0.32*** 0.07 0.02 0.02** 0.03* -0.03 (0.46) (0.43) (0.08) (0.09) (0.02) (0.01) (0.02) (0.04) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Yes Yes Mean in Control 4.23 2.53 0.00 -0.00 5.9% 0.9% 4.1% 34.4% p-value T1=T2 0.40 0.40 0.00 0.00 0.00 0.11 0.14 0.15 p-value T2=T3 0.02 0.11 0.57 0.00 0.08 0.82 0.67 0.01 p-value T1=T3 0.00 0.02 0.00 0.52 0.57 0.21 0.51 0.11 p-value T1=T2=T3 0.01 0.07 0.00 0.00 0.01 0.19 0.33 0.03 Observations 2,620 2,620 2,618 2,620 2,620 2,620 2,620 2,620 Robust standard errors clustered at locality level. The main activity is directly identified by the respondent among the list of independent activities undertaken. Z-scores are centered in the control group. The index of business practices is based on four components : market assessment, business plan development, use of formal bookkeeping and separating payments to self. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table A14: Heterogeneous impacts on main economic outcomes: by age (youth) (1) (2) (3) (4) (5) (6) Self Earnings in Total start-up Value of Savings employed # Independent Self Employment capital productive assets stock (at least Activities (Profits, monthly) (all activities) (all activities) (Total) 1 activity) Panel A. Pooled Estimates Treatment (ITT) -0.02 0.19 1539.62 6783.65** 5715.73 -9099.42 (0.02) (0.15) (2651.66) (3141.06) (4305.50) (16056.28) Treatment X (Youth=1) 0.07*** 0.12 179.90 11124.81*** 6685.20 29629.50 (0.03) (0.17) (3701.52) (4277.01) (6222.78) (19078.18) (Youth=1) -0.07*** -0.17 -1606.76 -6215.77* -5013.73 -8162.50 (0.03) (0.17) (4159.25) (3754.21) (6724.79) (18013.22) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control (Youth=0) 95.3% 3.30 24729.67 15799.84 39255.37 68649.38 Total Treatment Effect (Youth=1) 0.05 0.31 1719.51 17908.46 12400.93 20530.08 p-value Total Treatment Effect (Youth=1) 0.00 0.01 0.50 0.00 0.01 0.11 Observations 2,620 2,620 2,620 2,620 2,620 2,618 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) -0.03 0.12 600.09 2931.06 3225.49 -24520.77 (0.02) (0.18) (3528.37) (4340.79) (6361.29) (16388.18) Cash Grant with repayment (T2) (ITT) -0.00 0.34** 1724.90 7475.58** 4404.44 -8939.99 (0.02) (0.17) (2839.11) (3754.76) (4471.08) (16502.23) Cash Grant (T3) (ITT) -0.01 -0.03 2962.86 12722.27* 13701.65* 20749.66 (0.02) (0.21) (3924.35) (6509.18) (8195.03) (21745.05) VSLA (T1) x (Youth=1) 0.09*** 0.23 93.73 8574.80 11394.39 27005.24 (0.03) (0.21) (5328.70) (5732.11) (10174.11) (18126.33) Cash Grant with repayment (T2) x (Youth=1) 0.05* -0.08 1054.25 14448.75** 6409.22 39806.55 (0.03) (0.20) (3911.23) (5817.78) (6121.29) (25248.91) Cash Grant (T3) x (Youth=1) 0.08** 0.36 -1467.32 9692.84 -2404.81 14614.61 (0.03) (0.25) (4502.18) (8169.55) (11268.76) (24592.29) (Youth=1) -0.07*** -0.18 -1694.54 -6597.11* -4976.70 -9298.19 (0.03) (0.17) (4134.46) (3722.33) (6651.19) (18056.02) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in T1 Control (Youth=0) 95.8% 3.50 24563.14 20116.99 42119.15 62488.09 Mean in T2 Control (Youth=0) 94.5% 3.36 25120.66 18900.13 43112.02 57763.89 Mean in T3 Control (Youth=0) 94.9% 3.49 24593.34 18544.51 40999.78 51691.22 Total Effect T1 (Youth=1) 0.05 0.35 693.81 11505.86 14619.88 2484.47 p-value Total Effect T1 (Youth=1) 0.01 0.01 0.85 0.00 0.02 0.83 Total Effect T2 (Youth=1) 0.04 0.27 2779.15 21924.33 10813.66 30866.56 p-value Total Effect T2 (Youth=1) 0.04 0.05 0.36 0.00 0.04 0.14 Total Effect T3 (Youth=1) 0.07 0.34 1495.55 22415.11 11296.84 35364.27 p-value Total Effect T3 (Youth=1) 0.00 0.13 0.69 0.00 0.19 0.03 p-value T1=T2 0.63 0.57 0.60 0.05 0.56 0.14 p-value T2=T3 0.29 0.75 0.75 0.95 0.96 0.85 p-value T1=T3 0.54 0.95 0.86 0.15 0.73 0.04 p-value T1=T2=T3 0.57 0.84 0.86 0.10 0.84 0.06 Observations 2,620 2,620 2,620 2,620 2,620 2,618 Robust standard errors clustered at locality level. Youth defined as up to 35 years old. Earnings, capital, value of assets and savings are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 61 Table A15: Heterogeneous impacts on main economic outcomes : by ethnic group (1) (2) (3) (4) (5) (6) Self Earnings in Total start-up Value of Savings employed # Independent Self Employment capital productive assets stock (at least Activities (Profits, monthly) (all activities) (all activities) (Total) 1 activity) Panel A. Pooled Estimates Treatment (ITT) 0.04 0.33 497.50 23404.39*** 13223.08 -17696.06 (0.03) (0.24) (5323.94) (6644.96) (10410.62) (19448.93) Treatment X (Native Group=1) -0.02 -0.09 1086.23 -13163.16* -5136.61 27589.76 (0.03) (0.29) (5981.05) (7075.88) (11854.75) (22027.58) (Native Group=1) -0.01 -0.16 -12125.71** 5582.04 -12687.85 -30036.28 (0.03) (0.27) (4748.45) (4867.27) (10215.08) (19164.67) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control (Native Gp.=0) 91.3% 3.08 34440.75 14164.66 54827.73 89720.48 Total Treatment Effect (Native Gp.=1) 0.02 0.23 1583.73 10241.23 8086.47 9893.70 p-value Total Treatment Effect (Native Gp.=1) 0.13 0.05 0.42 0.00 0.02 0.40 Observations 2,620 2,620 2,620 2,620 2,620 2,618 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.04 0.26 2151.82 14329.06* 13369.33 -24343.81 (0.03) (0.26) (7111.07) (7918.66) (11801.51) (24920.34) Cash Grant with repayment (T2) (ITT) 0.04 0.45 -4828.68 35382.03*** 6302.71 -22080.37 (0.03) (0.31) (5159.37) (12550.94) (10560.83) (21958.09) Cash Grant (T3) (ITT) 0.02 0.27 6215.71 22006.15** 25152.92 1208.63 (0.03) (0.28) (8732.89) (10543.16) (19158.38) (22009.58) VSLA (T1) x (Native Group=1) -0.03 -0.04 -2348.93 -8494.89 -5358.06 14700.48 (0.04) (0.31) (7605.15) (8632.15) (13639.53) (25974.43) Cash Grant with repayment (T2) x (Native Group=1) -0.02 -0.17 8418.13 -24070.38* 1737.02 39783.71 (0.04) (0.35) (5936.00) (12796.14) (12200.74) (26966.40) Cash Grant (T3) x (Native Group=1) 0.01 -0.14 -6077.98 -4945.50 -16893.78 30019.45 (0.04) (0.35) (9425.37) (9274.19) (21682.14) (26849.67) (Native Group=1) -0.01 -0.16 -12044.11** 5448.29 -12668.57 -29962.46 (0.03) (0.27) (4759.31) (4879.49) (10217.89) (19134.19) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in T1 Control (Native Gp.=0) 93.2% 3.30 35422.56 32917.10 64857.57 82111.29 Mean in T2 Control (Native Gp.=0) 92.6% 3.13 40178.77 26706.07 69083.55 81219.00 Mean in T3 Control (Native Gp.=0) 92.9% 3.16 36749.72 31424.72 65479.71 78087.28 Total Effect T1 (Native Gp.=1) 0.01 0.23 -197.11 5834.17 8011.27 -9643.33 p-value Total Effect T1 (Native Gp.=1) 0.52 0.10 0.93 0.06 0.06 0.35 Total Effect T2 (Native Gp.=1) 0.02 0.28 3589.45 11311.65 8039.73 17703.34 p-value Total Effect T2 (Native Gp.=1) 0.18 0.03 0.15 0.00 0.06 0.26 Total Effect T3 (Native Gp.=1) 0.04 0.13 137.73 17060.64 8259.14 31228.08 p-value Total Effect T3 (Native Gp.=1) 0.02 0.52 0.96 0.00 0.19 0.06 p-value T1=T2 0.60 0.67 0.15 0.12 0.99 0.03 p-value T2=T3 0.25 0.47 0.24 0.27 0.97 0.46 p-value T1=T3 0.12 0.65 0.91 0.04 0.97 0.00 p-value T1=T2=T3 0.26 0.74 0.30 0.08 1.00 0.00 Observations 2,620 2,620 2,620 2,620 2,620 2,618 Robust standard errors clustered at locality level. er´ Native ethnic groups are those originally from the Western region and include Yacouba / Dan / Kla, Toura, Mahouka / Mahou, Wan, Gu´ e, Gourounsi / Groussi and Wobe. Earnings, capital, value of assets and savings are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 62 Table A16: Heterogeneous impacts on main economic outcomes : by gender (1) (2) (3) (4) (5) (6) Self Earnings in Total start-up Value of Savings employed # Independent Self Employment capital productive assets stock (at least Activities (Profits, monthly) (all activities) (all activities) (Total) 1 activity) Panel A. Pooled Estimates Treatment (ITT) 0.01 0.38** 199.31 17066.78*** 15534.51** 12193.87 (0.02) (0.16) (3555.00) (4648.24) (6326.83) (12964.18) Treatment X (Female=1) 0.02 -0.20 1890.61 -6219.25 -8708.23 -9269.98 (0.02) (0.17) (3866.32) (5880.42) (6731.85) (14701.48) (Female=1) -0.06** -0.34** -7135.81** -3154.93 -719.00 3349.58 (0.02) (0.17) (3424.95) (5022.68) (6505.35) (12679.61) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control (Female=0) 92.8% 3.25 28089.48 18227.61 44854.08 52684.34 Total Treatment Effect (Female=1) 0.02 0.19 2089.92 10847.53 6826.28 2923.89 p-value Total Treatment Effect (Female=1) 0.05 0.09 0.28 0.00 0.04 0.81 Observations 2,620 2,620 2,620 2,620 2,620 2,618 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.01 0.35* 1965.55 12084.04** 14900.21* -4374.25 (0.02) (0.20) (4043.90) (6076.30) (8317.61) (11729.62) Cash Grant with repayment (T2) (ITT) -0.01 0.41** -1468.59 23848.75*** 13401.37** 13074.50 (0.02) (0.18) (3882.67) (6299.63) (6499.87) (19566.40) Cash Grant (T3) (ITT) 0.04* 0.40 64.34 12082.10 21711.80 45205.87* (0.02) (0.29) (8683.80) (9156.41) (16486.70) (25607.37) VSLA (T1) x (Female=1) 0.00 -0.19 -2058.61 -6535.27 -7657.11 -10222.37 (0.03) (0.21) (4696.02) (7004.22) (8885.10) (13728.79) Cash Grant with repayment (T2) x (Female=1) 0.04 -0.15 5222.36 -12389.68* -7692.19 -1848.10 (0.02) (0.20) (3822.62) (7358.96) (6676.54) (22687.30) Cash Grant (T3) x (Female=1) -0.01 -0.33 2618.23 8055.31 -13107.35 -25764.76 (0.03) (0.28) (9045.24) (12463.00) (17702.39) (26198.57) (Female=1) -0.06** -0.34** -7056.15** -3080.82 -803.36 3097.72 (0.02) (0.17) (3421.78) (4998.92) (6517.45) (12637.01) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in T1 Control (Female=0) 93.7% 3.50 27591.34 30414.86 55430.22 64648.98 Mean in T2 Control (Female=0) 94.4% 3.45 29169.64 25981.81 56665.74 57216.53 Mean in T3 Control (Female=0) 93.4% 3.50 28716.85 31145.82 56298.52 53803.88 Total Effect T1 (Female=1) 0.01 0.16 -93.06 5548.76 7243.10 -14596.61 p-value Total Effect T1 (Female=1) 0.42 0.20 0.97 0.10 0.05 0.21 Total Effect T2 (Female=1) 0.03 0.26 3753.77 11459.07 5709.18 11226.40 p-value Total Effect T2 (Female=1) 0.03 0.05 0.10 0.00 0.15 0.49 Total Effect T3 (Female=1) 0.03 0.07 2682.57 20137.41 8604.45 19441.10 p-value Total Effect T3 (Female=1) 0.05 0.71 0.30 0.00 0.16 0.18 p-value T1=T2 0.31 0.43 0.19 0.12 0.70 0.06 p-value T2=T3 0.94 0.27 0.70 0.22 0.64 0.61 p-value T1=T3 0.36 0.58 0.37 0.04 0.82 0.00 p-value T1=T2=T3 0.56 0.51 0.42 0.07 0.87 0.01 Observations 2,620 2,620 2,620 2,620 2,620 2,618 Robust standard errors clustered at locality level. Earnings, capital, value of assets and savings are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 63 Table A17: Heterogeneous impacts on main economic outcomes: by urban/rural area (1) (2) (3) (4) (5) (6) Self Earnings in Total start-up Value of Savings employed # Independent Self Employment capital productive assets stock (at least Activities (Profits, monthly) (all activities) (all activities) (Total) 1 activity) Panel A. Pooled Estimates Treatment (ITT) 0.09*** 0.47*** 3141.06 19100.99*** 5473.37 -59370.97* (0.03) (0.18) (3618.09) (6977.27) (8065.38) (34580.63) Treatment X (Village=1) -0.08** -0.27 -1696.65 -8496.20 4470.67 78791.41** (0.03) (0.22) (4191.91) (7435.67) (8801.53) (35583.45) (Village=1) 0.09*** 0.39* -2492.48 9271.09 -6203.12 -87841.55*** (0.03) (0.21) (3652.49) (6994.06) (8383.41) (33489.77) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control (Village=0) 82.2% 2.14 24556.58 18276.67 50731.09 140760.71 Total Treatment Effect (Village=1) 0.01 0.21 1444.42 10604.79 9944.04 19420.44 p-value Total Treatment Effect (Village=1) 0.48 0.07 0.48 0.00 0.00 0.04 Observations 2,620 2,620 2,620 2,620 2,620 2,618 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.07 0.52** 5144.74 17275.74** 17465.81* -80152.91** (0.04) (0.22) (6333.96) (8041.28) (9774.71) (35394.56) Cash Grant with repayment (T2) (ITT) 0.09** 0.68** 576.69 16967.72* -353.84 -58148.19 (0.04) (0.28) (3719.88) (8838.47) (10318.16) (41905.99) Cash Grant (T3) (ITT) 0.11*** -0.01 3302.08 27081.26** -9843.08 -14880.54 (0.04) (0.35) (5266.44) (12443.34) (7194.30) (44750.37) VSLA (T1) x (Village=1) -0.07 -0.34 -5372.92 -13555.86 -10376.70 83335.49** (0.04) (0.26) (6837.90) (8521.12) (10346.50) (36230.56) Cash Grant with repayment (T2) x (Village=1) -0.09** -0.45 2127.16 -2714.49 9799.10 84314.74* (0.04) (0.31) (4531.44) (9425.53) (11084.91) (43652.24) Cash Grant (T3) x (Village=1) -0.09** 0.23 -1593.33 -11816.02 27184.96*** 49264.37 (0.04) (0.40) (6410.82) (13635.96) (10222.91) (46741.86) (Village=1) 0.09*** 0.38* -2391.10 10751.27 -7702.34 -82073.27** (0.03) (0.20) (3576.48) (7149.07) (8221.36) (33064.79) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in T1 Control (Village=0) 88.7% 2.49 25474.42 28413.81 48269.25 110247.19 Mean in T2 Control (Village=0) 87.4% 2.31 29122.41 29727.68 57405.77 101490.51 Mean in T3 Control (Village=0) 87.8% 2.43 27715.15 28900.29 57956.55 90602.68 Total Effect T1 (Village=1) 0.00 0.18 -228.18 3719.88 7089.12 3182.58 p-value Total Effect T1 (Village=1) 0.80 0.17 0.93 0.19 0.06 0.70 Total Effect T2 (Village=1) 0.01 0.23 2703.86 14253.23 9445.27 26166.56 p-value Total Effect T2 (Village=1) 0.59 0.07 0.29 0.00 0.02 0.05 Total Effect T3 (Village=1) 0.02 0.22 1708.75 15265.24 17341.88 34383.83 p-value Total Effect T3 (Village=1) 0.20 0.26 0.64 0.01 0.02 0.01 p-value T1=T2 0.80 0.66 0.30 0.00 0.57 0.05 p-value T2=T3 0.53 0.98 0.79 0.87 0.29 0.61 p-value T1=T3 0.36 0.81 0.62 0.07 0.16 0.01 p-value T1=T2=T3 0.64 0.90 0.59 0.01 0.36 0.01 Observations 2,620 2,620 2,620 2,620 2,620 2,618 Robust standard errors clustered at locality level. Village defined by administrative status used for lotteries. Earnings, capital, value of assets and savings are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 64 Table A18: Local spillovers on diversification of self-employment between agriculture and non-agricultural activities (1) (2) (3) (4) (5) (6) Earnings in Earnings in Value of Value of # Non-Ag. # Agricultural Self Employment : Self Employment : non agricultural agricultural Independent Independent Non Ag. Activities Ag. Activities productive assets productive assets Activities Activities (Profits, monthly) (Profits, monthly) (all activities) (all activities) Panel A. Pooled Estimates Treatment (ITT) 0.04 0.33** -1662.02 2920.22 -507.06 10091.00** (0.05) (0.14) (3378.86) (2458.36) (6182.54) (3899.53) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 0.50 2.72 16957.01 15514.05 30252.69 29296.44 Observations 1,102 1,102 1,102 1,102 1,102 1,102 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) -0.04 0.45*** -391.93 1862.14 -1350.40 6872.00 65 (0.07) (0.17) (4970.00) (3194.00) (7139.71) (5096.88) Cash Grant with repayment (T2) (ITT) 0.09 0.20 -2900.62 4359.50 -659.39 11623.22** (0.07) (0.18) (3266.45) (3694.54) (7808.94) (5552.40) Cash Grant (T3) (ITT) 0.05 0.43** -1157.27 1521.01 1606.67 12850.80** (0.08) (0.21) (4355.10) (3280.81) (6203.42) (5631.20) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 0.50 2.72 16957.01 15514.05 30252.69 29296.44 p-value T1=T2 0.10 0.21 0.56 0.58 0.92 0.48 p-value T2=T3 0.65 0.33 0.64 0.51 0.74 0.86 p-value T1=T3 0.26 0.94 0.88 0.93 0.63 0.35 p-value T1=T2=T3 0.23 0.41 0.79 0.79 0.88 0.61 Observations 1,102 1,102 1,102 1,102 1,102 1,102 Robust standard errors clustered at locality level. Earnings and value of assets are in CFA franc and winsorized at 99%. “Ag.” stands for Agricultural. The number of independent activities per individual is based on all activities operating in the last 12 months. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A6. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table A19: Local spillovers on credit (1) (2) Has taken Credits taken a credit from from VSLA a VSLA Panel A. Pooled Estimates Treatment (ITT) 0.04* 1425.62 (0.02) (1237.66) PDS Lasso selected controls Yes Yes Department X (Urban/Rural) Yes Yes Mean in Control 7.2% 2615.20 Observations 1,102 1,102 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.08*** 2046.59 66 (0.03) (1340.13) Cash Grant with repayment (T2) (ITT) 0.01 841.62 (0.03) (1631.73) Cash Grant (T3) (ITT) 0.05 1606.52 (0.03) (1584.98) PDS Lasso selected controls Yes Yes Department X (Urban/Rural) Yes Yes Mean in Control 7.2% 2615.20 p-value T1=T2 0.03 0.43 p-value T2=T3 0.17 0.67 p-value T1=T3 0.30 0.78 p-value T1=T2=T3 0.08 0.73 Observations 1,102 1,102 Robust standard errors clustered at locality level. Credit amounts are in CFA franc and winsorized at 99%. The specification includes control variables chosen using post double se- lection (PDS) Lasso based on the set in Table A6. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table A20: Local spillovers on business practices and knowledge (1) (2) (3) (4) (5) (6) Separate # employees Index of Has done a Has developed Uses formal regular (main business practices market a business plan book-keeping payments activity) (z-score) assessment to self Panel A. Pooled Estimates Treatment (ITT) -0.03 0.08 -0.01 0.00 0.03* 0.02 (0.45) (0.07) (0.02) (0.01) (0.02) (0.04) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 5.07 -0.00 9.7% 2.1% 9.4% 37.5% Observations 1,101 1,102 1,102 1,102 1,102 1,102 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) -0.16 0.07 -0.03 0.01 0.04* 0.03 (0.52) (0.08) (0.03) (0.01) (0.02) (0.05) Cash Grant with repayment (T2) (ITT) 0.24 0.17* 0.02 0.00 0.04 0.05 (0.59) (0.09) (0.03) (0.01) (0.03) (0.04) Cash Grant (T3) (ITT) -0.43 -0.09 -0.02 -0.01 0.00 -0.06 (0.60) (0.08) (0.03) (0.01) (0.02) (0.04) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 5.07 -0.00 9.7% 2.1% 9.4% 37.5% p-value T1=T2 0.52 0.28 0.15 0.87 0.98 0.67 p-value T2=T3 0.33 0.00 0.22 0.39 0.14 0.02 p-value T1=T3 0.64 0.06 0.84 0.34 0.10 0.08 p-value T1=T2=T3 0.62 0.01 0.32 0.56 0.16 0.05 Observations 1,101 1,102 1,102 1,102 1,102 1,102 Robust standard errors clustered at locality level. The main activity is directly identified by the respondent among the list of independent activities undertaken. Z-scores are centered on the control group. The index of business practices is based on four variables : market assessment, business plan development, use of formal bookkeeping and way of paying oneself. The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A6. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 67 Table A21: Local spillovers on social outcomes (1) (2) (3) (4) (5) (6) Participation in Participation in Participation in # times received # times provided Trust community works social activities groups or associations financial support financial support Index (# times in (# times in (# groups) (last 12 mths) (last 12 mths) (z-score) last 12 mths) last 12 mths) Panel A. Pooled Estimates Treatment (ITT) 0.00 0.04 0.23 -0.00 -0.55 -0.06 (0.06) (0.19) (0.20) (0.02) (0.44) (0.10) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 1.29 1.24 2.06 77.20 8.80 0.00 Observations 1,102 1,102 1,102 1,102 1,102 995 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) -0.05 -0.04 0.38 -0.00 -0.96* -0.11 68 (0.08) (0.30) (0.30) (0.02) (0.53) (0.11) Cash Grant with repayment (T2) (ITT) 0.02 0.18 0.01 -0.01 -0.41 0.00 (0.08) (0.22) (0.22) (0.03) (0.58) (0.11) Cash Grant (T3) (ITT) 0.05 -0.13 0.46 0.02 -0.08 -0.12 (0.10) (0.24) (0.35) (0.02) (0.69) (0.14) PDS Lasso selected controls Yes Yes Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Yes Yes Mean in Control 1.29 1.24 2.06 77.20 8.80 0.00 p-value T1=T2 0.46 0.52 0.25 0.91 0.38 0.23 p-value T2=T3 0.78 0.26 0.21 0.30 0.67 0.31 p-value T1=T3 0.35 0.79 0.84 0.30 0.23 0.91 p-value T1=T2=T3 0.59 0.53 0.33 0.44 0.44 0.40 Observations 1,102 1,102 1,102 1,102 1,102 995 Robust standard errors clustered at locality level. Z-scores are centered in the control group. The Trust index is based on 16 variables measuring different forms of trust (economic, financial, security) and trust in different types of groups or institutions (neighbors, same and other ethnic groups, leaders, foreigners). The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A6. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 Table A22: Heterogeneous impacts on main social outcomes : by age (youth) (1) (2) (3) (4) Participation in # times received # times provided Trust groups or associations financial support financial support Index (# groups) (last 12 mths) (last 12 mths) (z-score) Panel A. Pooled Estimates Treatment (ITT) 0.16** 0.16 0.05 0.08 (0.07) (0.14) (0.16) (0.08) Treatment X (Youth=1) -0.01 0.07 0.42** -0.11 (0.09) (0.21) (0.21) (0.09) (Youth=1) 0.10 -0.18 -0.50** 0.08 (0.10) (0.19) (0.21) (0.10) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in Control (Youth=0) 1.24 0.91 1.44 -0.05 Total Treatment Effect (Youth=1) 0.15 0.23 0.47 -0.03 p-value Total Treatment Effect (Youth=1) 0.04 0.12 0.00 0.71 Observations 2,620 2,620 2,620 2,374 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.15* -0.01 -0.09 0.10 (0.08) (0.15) (0.20) (0.10) Cash Grant with repayment (T2) (ITT) 0.16* 0.10 0.21 0.05 (0.08) (0.16) (0.19) (0.09) Cash Grant (T3) (ITT) 0.19* 0.62* -0.02 0.13 (0.11) (0.37) (0.26) (0.14) VSLA (T1) x (Youth=1) 0.04 0.26 0.61** -0.10 (0.12) (0.27) (0.25) (0.11) Cash Grant with repayment (T2) x (Youth=1) -0.04 0.03 0.18 -0.09 (0.10) (0.23) (0.25) (0.11) Cash Grant (T3) x (Youth=1) -0.07 -0.21 0.57 -0.18 (0.11) (0.37) (0.36) (0.13) (Youth=1) 0.10 -0.18 -0.52** 0.08 (0.10) (0.19) (0.21) (0.10) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in T1 Control (Youth=0) 1.34 1.07 1.48 0.02 Mean in T2 Control (Youth=0) 1.33 1.04 1.38 0.02 Mean in T3 Control (Youth=0) 1.34 0.95 1.45 0.01 Total Effect T1 (Youth=1) 0.20 0.25 0.52 -0.00 p-value Total Effect T1 (Youth=1) 0.04 0.23 0.00 0.97 Total Effect T2 (Youth=1) 0.12 0.13 0.39 -0.04 p-value Total Effect T2 (Youth=1) 0.15 0.44 0.04 0.61 Total Effect T3 (Youth=1) 0.12 0.41 0.55 -0.05 p-value Total Effect T3 (Youth=1) 0.17 0.07 0.05 0.71 p-value T1=T2 0.42 0.59 0.49 0.64 p-value T2=T3 0.98 0.24 0.57 0.99 p-value T1=T3 0.45 0.55 0.92 0.72 p-value T1=T2=T3 0.68 0.49 0.75 0.88 Observations 2,620 2,620 2,620 2,374 Robust standard errors clustered at locality level.Youth defined as up to 35 years old. The Trust index is based on 16 variables measuring different forms of trust (economic, financial, security) and trust in different types of groups or institutions (neighbors, same and other ethnic groups, leaders, foreigners). The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 69 Table A23: Heterogeneous impacts on main social outcomes: by ethnic group (1) (2) (3) (4) Participation in # times received # times provided Trust groups or associations financial support financial support Index (# groups) (last 12 mths) (last 12 mths) (z-score) Panel A. Pooled Estimates Treatment (ITT) 0.24** 0.72*** 0.38 0.18 (0.12) (0.21) (0.28) (0.14) Treatment X (Native Group=1) -0.10 -0.61*** -0.11 -0.19 (0.14) (0.23) (0.31) (0.14) (Native Group=1) 0.18 0.38** -0.06 -0.10 (0.12) (0.15) (0.24) (0.12) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in Control (Native Gp.=0) 1.02 0.47 1.37 0.08 Total Treatment Effect (Native Gp.=1) 0.14 0.11 0.27 -0.01 p-value Total Treatment Effect (Native Gp.=1) 0.02 0.32 0.04 0.92 Observations 2,620 2,620 2,620 2,374 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.35*** 0.89*** 0.39 0.11 (0.13) (0.30) (0.32) (0.16) Cash Grant with repayment (T2) (ITT) 0.19 0.45** 0.73* 0.27* (0.14) (0.19) (0.38) (0.15) Cash Grant (T3) (ITT) 0.09 0.73 -0.30 0.19 (0.15) (0.55) (0.37) (0.19) VSLA (T1) x (Native Group=1) -0.21 -0.91*** -0.16 -0.08 (0.16) (0.32) (0.35) (0.17) Cash Grant with repayment (T2) x (Native Group=1) -0.06 -0.40* -0.49 -0.31* (0.16) (0.22) (0.41) (0.16) Cash Grant (T3) x (Native Group=1) 0.09 -0.24 0.73 -0.18 (0.18) (0.61) (0.46) (0.21) (Native Group=1) 0.18 0.38** -0.06 -0.10 (0.12) (0.15) (0.24) (0.12) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in T1 Control (Native Gp.=0) 1.14 0.85 1.59 0.16 Mean in T2 Control (Native Gp.=0) 1.20 1.04 1.48 0.08 Mean in T3 Control (Native Gp.=0) 1.20 0.98 1.73 0.09 Total Effect T1 (Native Gp.=1) 0.14 -0.02 0.23 0.03 p-value Total Effect T1 (Native Gp.=1) 0.08 0.88 0.13 0.70 Total Effect T2 (Native Gp.=1) 0.13 0.06 0.24 -0.04 p-value Total Effect T2 (Native Gp.=1) 0.08 0.66 0.11 0.56 Total Effect T3 (Native Gp.=1) 0.18 0.49 0.42 0.00 p-value Total Effect T3 (Native Gp.=1) 0.04 0.06 0.06 0.97 p-value T1=T2 0.89 0.59 0.97 0.38 p-value T2=T3 0.56 0.10 0.42 0.70 p-value T1=T3 0.64 0.06 0.42 0.83 p-value T1=T2=T3 0.83 0.18 0.69 0.67 Observations 2,620 2,620 2,620 2,374 Robust standard errors clustered at locality level. Native ethnic groups are those originally from the Western region and include Yacouba / Dan / Kla, Toura, Mahouka / Mahou, Wan, Gu´ er´e, Gourounsi / Groussi and Wobe. The Trust index is based on 16 variables measuring different forms of trust (economic, financial, security) and trust in different types of groups or institutions (neighbors, same and other ethnic groups, leaders, foreigners). The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 70 Table A24: Heterogeneous impacts on main social outcomes : by gender (1) (2) (3) (4) Participation in # times received # times provided Trust groups or associations financial support financial support Index (# groups) (last 12 mths) (last 12 mths) (z-score) Panel A. Pooled Estimates Treatment (ITT) 0.24*** 0.20 0.49** -0.02 (0.08) (0.17) (0.23) (0.11) Treatment X (Female=1) -0.11 0.01 -0.29 0.06 (0.10) (0.19) (0.23) (0.11) (Female=1) 0.12 -0.22 -0.15 -0.31*** (0.10) (0.17) (0.20) (0.10) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in Control (Female=0) 0.99 1.07 1.48 0.17 Total Treatment Effect (Female=1) 0.12 0.21 0.20 0.04 p-value Total Treatment Effect (Female=1) 0.05 0.06 0.08 0.56 Observations 2,620 2,620 2,620 2,374 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.27*** 0.23 0.33 -0.03 (0.10) (0.23) (0.28) (0.13) Cash Grant with repayment (T2) (ITT) 0.21** 0.01 0.69** -0.03 (0.10) (0.18) (0.28) (0.12) Cash Grant (T3) (ITT) 0.22** 0.56 0.36 -0.00 (0.11) (0.40) (0.44) (0.17) VSLA (T1) x (Female=1) -0.14 -0.12 -0.11 0.09 (0.12) (0.24) (0.28) (0.14) Cash Grant with repayment (T2) x (Female=1) -0.10 0.15 -0.53* 0.04 (0.12) (0.21) (0.29) (0.13) Cash Grant (T3) x (Female=1) -0.08 -0.04 -0.10 0.05 (0.15) (0.52) (0.42) (0.16) (Female=1) 0.12 -0.22 -0.16 -0.32*** (0.10) (0.16) (0.20) (0.11) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in T1 Control (Female=0) 1.12 1.16 1.81 0.16 Mean in T2 Control (Female=0) 1.14 1.26 1.65 0.14 Mean in T3 Control (Female=0) 1.16 1.12 1.79 0.15 Total Effect T1 (Female=1) 0.13 0.11 0.22 0.06 p-value Total Effect T1 (Female=1) 0.08 0.42 0.10 0.45 Total Effect T2 (Female=1) 0.11 0.16 0.16 0.01 p-value Total Effect T2 (Female=1) 0.15 0.23 0.24 0.90 Total Effect T3 (Female=1) 0.14 0.52 0.26 0.05 p-value Total Effect T3 (Female=1) 0.18 0.08 0.13 0.67 p-value T1=T2 0.80 0.75 0.62 0.50 p-value T2=T3 0.82 0.25 0.55 0.73 p-value T1=T3 0.98 0.20 0.84 0.88 p-value T1=T2=T3 0.96 0.43 0.80 0.79 Observations 2,620 2,620 2,620 2,374 Robust standard errors clustered at locality level. The Trust index is based on 16 variables measuring different forms of trust (economic, financial, security) and trust in different types of groups or institutions (neighbors, same and other ethnic groups, leaders, foreigners). The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 71 Table A25: Heterogeneous impacts on social outcomes : by urban/rural area (1) (2) (3) (4) Participation in # times received # times provided Trust groups or associations financial support financial support Index (# groups) (last 12 mths) (last 12 mths) (z-score) Panel A. Pooled Estimates Treatment (ITT) 0.12 0.26 0.59** 0.10 (0.08) (0.20) (0.27) (0.17) Treatment X (Village=1) 0.05 -0.04 -0.36 -0.08 (0.10) (0.24) (0.30) (0.18) (Village=1) -0.03 0.05 0.14 0.10 (0.09) (0.26) (0.30) (0.18) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in Control (Village=0) 1.20 0.62 1.33 -0.21 Total Treatment Effect (Village=1) 0.17 0.22 0.23 0.02 p-value Total Treatment Effect (Village=1) 0.01 0.06 0.07 0.78 Observations 2,620 2,620 2,620 2,374 Panel B. Treatment Arm Estimates VSLA (T1) (ITT) 0.32*** 0.47 0.60** 0.20 (0.10) (0.30) (0.28) (0.20) Cash Grant with repayment (T2) (ITT) -0.05 0.10 0.65* 0.12 (0.09) (0.19) (0.36) (0.19) Cash Grant (T3) (ITT) -0.02 0.11 0.47 -0.14 (0.10) (0.21) (0.40) (0.24) VSLA (T1) x (Village=1) -0.18 -0.35 -0.39 -0.17 (0.12) (0.33) (0.33) (0.22) Cash Grant with repayment (T2) x (Village=1) 0.22* 0.03 -0.40 -0.14 (0.11) (0.24) (0.39) (0.20) Cash Grant (T3) x (Village=1) 0.22 0.54 -0.22 0.24 (0.14) (0.34) (0.46) (0.26) (Village=1) -0.05 0.02 0.14 0.07 (0.09) (0.26) (0.30) (0.18) PDS Lasso selected controls Yes Yes Yes Yes Department X (Urban/Rural) Yes Yes Yes Yes Mean in T1 Control (Village=0) 1.18 0.66 1.59 -0.17 Mean in T2 Control (Village=0) 1.31 0.82 1.58 -0.19 Mean in T3 Control (Village=0) 1.27 0.80 1.59 -0.17 Total Effect T1 (Village=1) 0.14 0.12 0.21 0.03 p-value Total Effect T1 (Village=1) 0.07 0.41 0.18 0.71 Total Effect T2 (Village=1) 0.17 0.13 0.24 -0.02 p-value Total Effect T2 (Village=1) 0.02 0.34 0.11 0.81 Total Effect T3 (Village=1) 0.20 0.65 0.25 0.10 p-value Total Effect T3 (Village=1) 0.03 0.02 0.25 0.43 p-value T1=T2 0.64 0.96 0.85 0.56 p-value T2=T3 0.79 0.07 0.95 0.35 p-value T1=T3 0.50 0.07 0.85 0.61 p-value T1=T2=T3 0.78 0.16 0.98 0.61 Observations 2,620 2,620 2,620 2,374 Robust standard errors clustered at locality level. Village defined by administrative status used for lotteries. The Trust index is based on 16 variables measuring different form of trust (economic, financial, security) and trust towards different types of groups or institutions (neighbors, same and other ethnic groups, leaders, foreigners). The specification includes control variables chosen using post double selection (PDS) Lasso based on the set in Table A5. ∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01 72 A2 Additional information on implementation Implementation partners The program, named “PRISE”, was implemented from 2014 to 2017. It was funded by the Japanese Social Development fund through the World Bank and coordinated by the “Office Coordinating Employment Programs” (BCP Emploi), which is part of the Ministry of Youth Employment. The program was implemented by the International ote d’Ivoire Rescue Committee (IRC), a prominent NGO which has offices in Western Cˆ and a long experience with these type of interventions. Implementation phases and selection of sub-prefectures ote d’Ivoire: Tonkpi, Cavally, The program was implemented in four regions of Western Cˆ emon and Bafing. It was rolled out in three phases, from 2014 to 2017. The first Gu´ phase tested implementation, including the targeting procedures, public lotteries, training curricula, and business plans. The RCT was embedded in the second phase, which covered 16 sub-prefectures and had the most participants. The third phase used funds recovered from the first and second phases to reach additional localities. The 37 sub- prefectures (for the three phases) were identified during project preparation by the donor, the coordinating agency (BCP Emploi) and government counterparts. The selection was based on economic needs, vulnerability and displacement levels, as well as the absence of other economic development interventions. Note that each of the 37 sub-prefectures was initially assigned to a specific phase, so that there is no overlap between phases. Selection of eligible localities within sub-prefectures Eligibility criteria for localities included (i) having at least one micro-finance institution within 30km, (ii) a high concentration of vulnerable population and people displaced by the conflict, (iii) a population size suitable for program implementation (i.e. no micro settlements), (iv) not being already covered by a similar program. For the second phase, 354 localities were found eligible out of a total of 415 localities. Sampling lottery The sampling lottery took place in public in August 2015 in the presence of many regional and local officials. For practical matters, lotteries were organized at the same time in two 73 regions. It was stratified by clusters of sub-prefectures and urban/rural areas. The three potential intervention modalities were described in broad terms when the sampling lottery took place. It was explained that potential eligible individuals would be pre-identified in each locality, but only some localities would be selected during the second lottery. Treatment assignment lottery The assignment lottery was public and took place in March 2016. It was again stratified by cluster of sub-prefectures and type of localities. The allocation of localities across interventions required to group sub-prefectures in clusters (so that there would be at least 10 units per cluster). This led to allocating “zero” localities to the cash-grant modality in some sub-prefectures. This was due to budget constraints at the program design stage, which only planned for 30 localities (out of 147 treated localities) in the cash grant modality. Enrollment and selection of beneficiaries within localities Eligibility criteria were carefully explained during the enrollment phase in each locality, with support from village committees composed of respected community members. After randomization, the vulnerability score was calculated from enrollment data, and a first list of beneficiaries was obtained. A first verification of eligibility criteria was done using enrollment data. A second phase of verification subsequently involved consultations with the village committees53 and triangulation with other development partners. 54 Based on the whole process, 2,184 (15%) of applicants were deemed ineligible.55 Entrepreneurship training The training content focuses on building entrepreneurship and business skills. In total, the training lasted 55 hours delivered over 8 days to small groups in each locality (5 days for basic training and 3 days for advanced business training). It was organized and delivered by the implementing agency (IRC). The training content was tailored to 53 Committees in each community verified the list of individuals, which led to the identification of 442 (3.3%) ineligible applicants 54 The list of selected individuals was triangulated with similar programs implemented in the region to confirm whether some individuals had recently benefitted from assistance. Of the remaining applicants, 0.4% were identified as ineligible based on this criterion. 55 Note that, although the two first eligibility checks (using baseline data and cross-validation with lists from other programs) were performed uniformly across the 207 sites, the final check (village committee verification) could not be implemented in the 60 control villages. 74 low-skilled target groups, for instance by relying on pictures and hands-on exercises. The curriculum itself was the output of many tests and adjustments, jointly led by IRC and a consulting firm specialized in training. It was based on a curriculum developed for low- skilled beneficiaries of a Public Works program (Bertrand et al. (2021)) as well as IRC own curriculum “EASE”. The curriculum was revised after the 1st year of the program to be further tailored to the type of beneficiaries and the geographical area. Basic training. First, a basic training covered entrepreneurship fundamentals (focused on starting your own activity) and a motivational module around peace building and community engagement. The basic entrepreneurship training curriculum covers issues related to starting a business, such as how to choose the right business to start, how to attract potential clients, how to deal with competition, how to manage costs, how to set the right price, etc. Business plan development. Second, individuals (or small groups) worked on a busi- ness plan. They received feedback and supervision from the trainers. The business plan development was common across the three intervention modalities, and done in addi- tion to the 55 training hours. This was mostly field-based, and participants had to find relevant information on prices, costs and competitors to fill their business plans. Business plan validation. For localities assigned to ‘cash grants with repayment’ or ‘cash grants’, business plans were evaluated, after which the project was either approved for funding, rejected, or sent for revisions. Three rounds of reviews took place, and ultimately more than 95% of business plans were approved to be funded. In cash grant with repayment localities (respectively cash grant localities), 80.9% of selected individuals submitted a business plan for review (respectively 82%). Most of the business plans were approved to be funded (96.7%, respectively 97.8% of business plans). In the end, 78.9% of selected individuals claimed and received the funds (respectively 81.1%). In the VSLA modality, beneficiaries similarly drafted business plans, but these plans were not evaluated. Advanced training. The third part of the training covered more advanced topics on en- trepreneurship and included a life skills module. The advanced entrepreneurship training curriculum covers the management of existing businesses: managing stocks, monitoring sales, performing basic accounting, monitoring running capital, etc. 75 Local follow-up. A ”community expert” with a higher level of education than bene- ficiaries was designated in each locality to participate in the training and help explain or clarify some of the content. The “community expert” also received training from the NGO on how to animate a group and efficiently deliver a training. He/she had a key role in ensuring that the content of the training had been understood, and to provide some follow-up based on individual needs. Recollection of cash grants with repayment. Only 39.96% of the total target amount was collected. On average across beneficiaries, only 20% of the grant amount (instead of the planned 50%) was repaid. Of beneficiaries, 13.8% fully reimbursed and 28.5% did not reimburse anything. The remaining individuals reimbursed on average 22,109 CFA (23% of the grant instead of 50%) with substantial variation: the 25th percentile reimbursed 11.7% and the 75th percentile 33.4%. On average, 1.8 deposits were made per business plan funded. A3 Measurement of main outcomes We measure a first set of key outcomes related to employment and income-generating activities. Employment is a dummy variable indicating if the individual has worked at least 1 hour in the past 7 days. We decomposed this measure between wage and self-employment (which includes independent activities in both agriculture and non- agricultural self-employment). The Number of independent activities counts all self- employment activities in which the individual was engaged over the last 30 days. Monthly Earnings (in CFA francs ) are reported separately for wage employment and self-employment. In the case of self-employment, the measure captures self-declared profits (across activ- ities and over the last 30 days). Hours worked are computed separately for wage and self-employment and aggregated over the last 7 days. We also measure intermediary outcomes related to savings and investment. Savings Stock (in CFA francs) is calculated at the time of the survey. It accounts for savings in cash, mobile money, micro finance institutions (MFI), bank, farmers’ cooperatives, groups of savings (ROSCA or VSLA). Start-up capital (in CFA francs) is the sum of capital used to launch independent activities that are reported as operating in the 12 months before the follow-up survey Value of productive assets is the sum of (self-reported) value 76 of assets in all independent activities active at the time of the survey (in CFA franc). Entrepreneurship knowledge is the score obtained on a quiz which focuses on the core topics covered during the training including what to include in a business plan and where to get market information. Business practices is an index of the use of business practices for independent activities covering keeping books, having done a market assessment and a business plan, separating household and business accounts. Both indices are demeaned and standardized using the control group. We also measure a range of social outcomes. We capture the total number of groups or associations in which the individual participated over the last 12 months, includ- ing economic and non-economic groups (political, religious, women and youth groups). Transfers received (respectively Transfers given ) is the number of times the participant received financial support from other (and respectively the number of times she/he pro- vided financial support to others). “Financial support” is defined as giving cash for food or health care, school fees, or business inputs. Participation in community works (cleaning, rebuilding public infrastructure) and Participation in social activities (celebra- tions, funerals, festivals) proxy community involvement and is computed as the number of times individuals participated in the related activities over the last 12 months. Trust is a z-score index that captures the level of trust in different groups and institutions.56 The index is demeaned and standardized using the control group (separately for selected and non-selected groups). Victimization is a z-score index that accounts for the reported frequency of robbery, racket, physical assault and armed conflicts.57 We also capture individuals’ Perception of insecurity through questions about whether the participants ever feared to be victim of physical violence in the past 12 months and about how she/he sees the general level of insecurity in the locality. 56 It includes measures of trust in general and trust in economic relations. It also includes different measures of trust towards neighbors, people of same and different ethnic groups, local leaders, and foreigners. Finally, it includes opinion on the way the presence of other ethnic groups in the locality affect the economy and security in the locality. 57 For robbery and racket, the variable takes the value of 1 if those events happen “often”. For physical assault and armed conflicts, it takes the value of 1 when it happens “sometimes” and “often” (and 0 when it never happens). 77