Serving the Underserved: Microcredit as a Pathway to Commercial Banks

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Serving the Underserved: Microcredit as a Pathway to Commercial Banks Sumit Agarwal, Thomas Kigabo, Camelia Minoiu, Andrea F. Presbitero, Andr´e F Silva 2021-041 Please cite this paper as: Agarwal, Sumit, Thomas Kigabo, Camelia Minoiu, Andrea F. Presbitero, and Andr´e F Silva (2021). “Serving the Underserved: Microcredit as a Pathway to Commercial Banks,” FinanceandEconomicsDiscussionSeries2021-041. Washington: BoardofGovernorsofthe Federal Reserve System, https://doi.org/10.17016/FEDS.2021.041. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Serving the Underserved: Microcredit as a Pathway to Commercial Banks∗ Sumit Agarwal Thomas Kigabo Camelia Minoiu Andrea F. Presbitero André F. Silva Review of Economics and Statistics, forthcoming Abstract: A large-scale microcredit expansion program—together with a credit bureau accessible to all lenders—can enable unbanked borrowers to build a credit history, facilitating their transition to commercial banks. Loan-level data from Rwanda show the programimprovedaccesstocreditandreducedpoverty. Asizableshareoffirst-timeborrowers switched to commercial banks, which cream-skim less risky borrowers and grant them larger, cheaper, and longer-maturity loans. Switchers have lower default risk than non-switchersandarenotriskierthanotherbankborrowers. Switchersalsoobtainbetter loan terms from banks compared with first-time bank borrowers without a credit history. JEL Codes: G21, O12, O55 Keywords: Access to credit, microfinance, unbanked, credit bureau, bank loans ∗ Date: June 2021. We are grateful to Raymond Fisman (the Editor), three anonymous referees, Shashwat Alok, Sofia Amaral, Haelim Anderson, Meghana Ayyagari, Thorsten Beck, Martin Brown, Miriam Bruhn, Fernando Chague, Ralph De Haas,JensHagendorff,MartinKanz,TumerKapan,AmartyaLahiri,W.BlakeMarsh,MariaSoledadMartinezPeria,Jonathan Morduch, Felix Noth, Steven Ongena, Ben Roth, Miri Stryjan, Sergio Vicente, Paolo Volpin, Nan Yang, and participants at numerous conferences and seminars for useful comments and suggestions. We also thank the staff at the National Bank of Rwanda, the TransUnion Credit Reference Bureau, the Rwanda Cooperative Agency, and the National Institute of Statistics of Rwanda for providing the data used in this study, assistance with numerous queries, and helpful discussions. This paper waspreviouslyentitled“FinancialAccessUndertheMicroscope” andispartofaprojectonMacroeconomicResearchinLow- Income Countries (project ID 60925) supported by the U.K.’s Foreign, Commonwealth & Development Office (FCDO). The views expressed herein are those of the authors and should not be attributed to the National Bank of Rwanda, International MonetaryFund,FCDO,FederalReserveBoard,orthoseaffiliatedwiththeFederalReserveSystem. Thispaperisdedicatedto thememoryofThomasKigabo,ChiefEconomistoftheNationalBankofRwanda,andawonderfulco-authorandcolleague,who passedawayinJanuary2021. SumitAgarwal,NationalUniversityofSingapore: ushakri@yahoo.com;CameliaMinoiu,Federal Reserve Board: camelia.minoiu@frb.gov; Andrea F. Presbitero, International Monetary Fund & CEPR: apresbitero@imf.org, AndréF.Silva,FederalReserveBoard: andre.f.silva@frb.gov. 1

1 Introduction Much of the progress toward financial inclusion in developing countries, where weak institutions and scarce collateral make information frictions in credit markets particularly costly (Morduch, 1999; Kaboski and Townsend, 2012), is attributed to the microfinance sector. A key but understudied question is to what extent and under which conditions microfinance facilitates access to credit from commercial banks. In this paper, we examine the effects of a large-scale microcredit expansion program on financial access and on the transition of unbanked borrowers to commercial banks. We show that the microfinance institutions (MFIs) created by the program—in the presence of a credit bureau that monitors individual borrowing activities and is accessible to all lenders—enable first-time borrowers to build a credit history and signal their quality to other lenders, mitigating information frictions. Commercial banks cream-skim the less risky borrowers, offering them better loan terms. Thus, for unbanked borrowers, microfinance can serve as a pathway to commercial banks. We analyze the effect of a nationwide government-subsidized microcredit expansion program that created an extensive network of community-focused savings and credit cooperatives (Umurenge SACCOs, henceforth U-SACCOs) across the 416 municipalities in Rwanda. The program resulted in more than 90 percent of Rwandans residing within three miles of a U-SACCO (AFI, 2014). Despite an official launch in 2009, U-SACCOs initiated lending operations in different months starting in late 2011, giving rise to a staggered implementation of the program. Our identification strategy exploits time-series variation in the opening of U-SACCOs across municipalities, coupled with administrative microdata on the lending activities of all financial institutions. Specifically, we use a comprehensive credit register with detailed information on the universe of loans to individuals in the entire country for a total of nine years around the implementation of the program (2008–16). Our empirical strategy hinges on the identifying assumption that the program rollout is orthogonal to local unobserved factors, including credit demand. In fact, our results could be spurious if U-SACCOs were more likely to start lending operations in more dynamic and 2

economically developed areas. To mitigate potential concerns about nonrandom variation in program rollout, we first show that timing of U-SACCO openings is not predicted by a wide range of ex ante municipality characteristics. These include proxies for economic development (for example, nightlights and nightlight growth, poverty rate, conflicts, population size), the size and financial positions of U-SACCOs before being licensed to lend, and distance to capital to capture the cost of reaching U-SACCOs for training and inspections. We find that the program significantly raised the probability of access to credit for unbanked individuals, particularly in less developed areas. While this effect is mostly driven by U-SACCOs, we also find a positive—albeit delayed—spillover effect on bank lending about one year after the program rollout. We then trace first-time U-SACCO borrowers over time and identify those who subsequently obtain loans from commercial banks. The number of these switchers grows rapidly after the program and, overall, represents eight percent of all first-time U-SACCO borrowers and five percent of new bank borrowers. U-SACCOs are unable to keep these borrowers in part due to tight regulatory and balance sheet constraints. To examine the link between switching from U-SACCOs to banks and financial outcomes, we compare switchers with other borrowers using a nearest-neighbor matching estimator that accounts for numerous loan, borrower, and municipality characteristics (Abadie and Imbens, 2011). In all tests, we compare loans granted in the same month and also control for borrowers’ ex ante risk by matching loans on payment history, credit capacity, and length of credit history. We find that switchers obtain larger, lower interest rate, and longer-maturity loans than non-switchers. Controlling for credit history and loan contract terms, switchers also have lower ex post default risk—both before and after they switch—compared with nonswitchers. Furthermore, they are no riskier than existing bank borrowers. Together, these results suggest that banks cream-skim the less risky borrowers from the U-SACCO borrower pool. Using data on banks’ geographic footprint across municipalities, we also show that banks increase their branch presence more in initially less developed areas, suggesting that banks’ lending activities in those areas may be spurred by the opening of U-SACCOs. 3

The transition of U-SACCO borrowers to banks is possible not only because of U- SACCOs’ screening, but also because of the availability of a credit information sharing mechanism that allows borrowers to signal their quality to other lenders. When unbanked individuals take their first loan at a U-SACCO, their information is submitted to the credit register, which tracks all lending activities and is available to all lenders.1 We present evidence suggesting that credit history at U-SACCOs is indeed valuable to banks. Comparing loan contract terms offered by banks to switchers with those offered by banks to first-time borrowers with no credit history, we find that banks systematically offer lower interest rates and longer-maturity loans to borrowers whose first lending relationship was at a U-SACCO. Finally, we show improved access to credit through the program translates into better economic outcomes. Poverty rates in municipalities where U-SACCOs start granting loans decline, especially in initially less developed areas, and in areas with more switching activity from U-SACCOs to banks. Furthermore, the program rollout is associated with an increase in small-firm creation and employment. Overall, our results suggest MFIs play a key, and thus far under-documented, role in alleviating information frictions in thin credit markets wherecommercialbanksfinditdifficulttoservetheunbankedpopulationthroughtraditional lending technologies (Castellanos et al., 2020) Our paper builds on an influential literature analyzing the effects of bank expansion programs on financial inclusion and economic development. Burgess and Pande (2005) show that a large state-led banking expansion program in India significantly reduced rural poverty through increased savings mobilization and credit provision. Agarwal et al. (2017) analyze thelargestfinancialinclusionprograminIndia(JanDhanYojana)andfindthatregionsmore exposed to the program saw an increase in credit availability, with households substituting 1The credit register is maintained by a private credit reference bureau that supplies borrower information on payment history and defaults—that is, both positive and negative information—against a fee. According to the World Bank’s 2013 Global Financial Development Report, 77 percent of countries in sub-Saharan Africa have a public credit register or private credit bureau, with almost two-thirds of these collecting both positive and negative borrower information. However, where credit reporting for microfinance clients exists, bank and microfinance credit registries are often separate (CGAP, 2011). 4

informal loans for less expensive bank credit. Focusing on the branch network expansion by Banco Azteca in Mexico, Bruhn and Love (2014) show that access to finance boosts labor market activity and incomes, particularly among the poor and in areas with lower bank presence. Brown et al. (2016) show that the expansion of an East European commercial microfinance bank to low-income areas increased the share of banked households. Finally, Allen et al. (2021) examine the branch expansion of Equity Bank in Kenya to low-income and underserved regions and show that it increased the likelihood of households having bank accounts and obtaining loans.2 A common feature of previous studies is their reliance on survey data to measure access to and usage of financial services, and economic outcomes. The data are often aggregated at the district or state level, inviting questions on whether the outcomes are being driven by a particular financial institution or its competitors. Notable exceptions include Azevedo et al. (2019), who exploit a novel screening technology that identifies loan-eligible individuals in Paraguay, Castellanos et al. (2020) who study the credit card market in Mexico, and Arraiz et al. (2020) who analyze free-riding in screening efforts in the context of small-to-medium enterprise (SME) lending in Peru. Our credit register data covers the lending activities of all microfinance institutions and commercial banks in a country, overcoming challenges related to aggregation and potential reporting biases. In addition, the data enable us to gauge which credit institutions are driving gains in access to credit, track individuals’ borrowing activities over time and across lenders, and characterize borrowers’ credit risk profile based on credit history and past loan performance. Finally, the data extend several years into the program 2Problems of financial inclusion are not exclusive to developing countries. According to the most recent estimates, 5.4 percent of U.S. households—or 7.1 million households—do not have access to basic banking services such as a checking or savings account. In addition, 23 percent of households do not have any credit product, although one in every five of such householdsseeksaccess(FDIC,2020). Furthermore,severalstudiesdocumentpositiveeffects of increased bank branch density on financial inclusion and economic outcomes in advanced economies (e.g.,Nguyen, 2019). Inparticular, Célerier andMatray (2019)show that theU.S. interstate bank branching deregulation increased financial inclusion and improved economic conditions for low-income households through asset accumulation and financial security. 5

so we can examine not only short-term, but also medium-term effects of the program on financial access.3 Our paper also relates more broadly to a long-standing literature on banks and financial development as key drivers of economic growth (e.g., King and Levine, 1993; Jayaratne and Strahan, 1996; Beck et al., 2000). Despite broad agreement that access to credit is critical for economic development, the evidence on the size of the effect is mixed. Randomized control trials (RCTs) generally reveal “a consistent pattern of modestly positive, but not transformative effects” (Banerjee et al., 2015; Meager, 2019), while studies of aggregated household survey data show somewhat larger effects (Bruhn and Love, 2014; Brown et al., 2016; Allen et al., 2021). A potential explanation for these mixed results is that the RCT literature has only “scratched the surface of identifying spillover and general equilibrium effects” (Banerjee et al., 2015). From this perspective, our analysis with administrative microdata on all loans in a country complements the RCT literature, as we are able to document the positive spillovers of a microcredit expansion program on bank lending, and emphasize the role of microfinance as a pathway for previously unbanked borrowers to commercial banks. In a related paper, Breza and Kinnan (2021) examine the real effects of a contraction in the supply of microcredit in India using district-level variation in exposure to a negative financial shock, and show that microcredit reduces labor earnings and consumption in more exposed areas. Contrary to that paper, we examine the effects of a positive shock—an expansion of the microfinance sector—on access to credit from both newly set up credit cooperatives as well as commercial banks. Our results suggest that the expansion of U- SACCOs can foster local development not only directly through the provision of financial 3It is important to note that our data do not cover informal credit. While a large share of the population have access to informal credit (61 percent in 2016 according to the FinScope (2016) survey compared with only 17 percent with access to formal credit), informal credit typically bears sizable interest rates and transaction costs (Giné, 2011). Therefore, the presence of a first-time borrower in the credit register indicates the first time that individual takes a loan from a formal creditor—whether or not they had access to informal credit before. In addition, our analysis of the FinScope survey data does not show that access to formal finance crowds out informal lenders (Appendix A). 6

services, but also indirectly by allowing lower-risk unbanked individuals to build a credit history and to obtain, in time, more attractive loan terms from banks. Our analysis of switchers relates to an influential paper by Ioannidou and Ongena (2010). Using data from the Bolivian credit register, the authors show that firms that switch across lenders initially obtain lower loan rates, but these rates subsequently increase, suggesting adverse selection and a hold-up problem (Sharpe, 1990; Rajan, 1992; von Thadden, 2004). We extend this line of research by documenting the transition of individual borrowers from credit cooperatives to commercial banks and hence emphasizing the screening role of MFIs. In addition, we analyze the terms of consumer loans (as opposed to business loans) and compare the default risk of switchers with non-switchers. Our analysis highlights potential risks and benefits associated with the transition of individuals from MFIs to banks. Given that switchers are, on average, less risky than non-switchers, U-SACCOs face the challenge of an increasingly riskier pool of borrowers over the medium run. At the same time, access to better loan contracts at commercial banks can spur entrepreneurship and income growth. Finally, our paper contributes to the literature on the positive effects of mandatory sharing of borrower information among financial institutions.4 Liberti et al. (2016) show that the expansion of the credit register in Argentina improved the efficiency of bank credit allocation by easing lending standards for previously excluded high-quality borrowers. Bos et al. (2021) find that the introduction of a credit register in Bosnia and Herzegovina reduced loan defaults, particularly among first-time borrowers, and that repeated borrowers receive better loan terms due to their ability to signal quality to competing lenders. Our results support the view that credit bureaus mitigate information frictions, with positive effects on financial inclusion and credit supply (Pagano and Jappelli, 1993; Padilla and Pagano, 1997). 4Cross-country evidence indicates that information sharing is associated with improved availability and lower cost of credit (Jappelli and Pagano, 2002; Brown et al., 2009), as well as lower bank risk-taking (Houston et al., 2010). 7

2 Background Rwandan Economy and Financial Sector. Rwanda is a landlocked country in East Africa with a population of around 13 million, largely in rural areas, and few natural resources. Following business-friendly reforms in the early 2000s, Rwanda experienced competitiveness gains, strong economic growth, and poverty reduction. Annual GDP growth averaged 7.8 percent and per capita income doubled between 2008 and 2016 (IMF, 2017a,b), and the 2020 World Bank’s Doing Business survey ranks Rwanda second in Africa and 38th in the world according to the ease of doing business. Rwanda’s commercial banking sector has developed rapidly in recent years. Total bank assets grew from 22 percent to 39 percent of GDP between 2008 and 2016, while bank credit to the private sector grew at an annual average of 13 percent in real terms over the same period (IMF, 2017a). Commercial banks represent about two-thirds of total banking sector assets. The banking sector is relatively concentrated, with the three largest commercial banks (out of 17) accounting for more than half of total bank assets, loans, and deposits.5 Most banks are foreign owned, but the majority of bank funding comes from local deposits, limiting the banking system’s exposure to external shocks. There are also 521 microfinance institutions (MFIs), including the 416 municipal credit cooperatives (U-SACCOs) set up through the microcredit expansion program examined in this paper—that is, one U-SACCO in each municipality. Taken together, MFIs account for about 10 percent of total banking sector assets. Over the past decade, Rwanda also made notable strides towards financial inclusion. The share of individuals with access to formal financial services, defined as those who have or 5There are 17 banks in total: 11 commercial banks (one of which only obtained regulatory approval in December 2016), one development bank, one cooperative bank, and four microfinance banks. As we have data until December 2016, we observe the lending activities of 16 active banks. In this paper, we refer to all banks as “commercial banks.” We include microfinance banks in this list, since, in contrast to microfinance institutions, microfinance banks have similar legal status to commercial banks and are also supervised by the National Bank of Rwanda. 8

use products or services from regulated financial institutions (bank or nonbank), increased from 21 percent to 68 percent of the adult population between 2008 and 2016, and access to formal credit from five percent to 17 percent over the same period (FinScope, 2012, 2016). According to statistics across 26 developing countries, primarily in Africa, where FinScope surveys measure financial access and use of financial products, Rwanda is ranked second in terms of the share of the adult population with access to formal financial services.6 These developments are partly the result of policies and regulations aimed at expanding financial access for the unbanked population, such as the nationwide microcredit program we analyze. Microcredit Expansion (U-SACCO) Program. We examine the effects of the U- SACCO program, which set up one “savings and credit cooperative” (SACCO) in each of Rwanda’s 416 municipalities.7 The program aimed to provide financial services at a low cost to all individuals but, in practice, targeted the unbanked population prevalent in rural communities. The program was launched in March 2009 and initially focused on providing access to savings accounts, with U-SACCOs granting their first loans starting in late 2011 upon licensing by the National Bank of Rwanda (NBR). In 2008, before the launch of U-SACCOs, there were no formal lending institutions in more than half of the 416 municipalities. The program significantly improved the availability of financial services across the country, with 1.6 million new customers and more than 90 percent of Rwandans residing within three miles of a U-SACCO branch (AFI, 2014), a larger share than in countries such as Kenya (86 percent) or Uganda (77 percent). 6In terms of financial inclusion, Rwanda fares well relative to its regional peers. The share of adult population with access to formal financial services (68 percent in 2016) places Rwanda above its East African peers such as Kenya (67 percent in 2013), Tanzania (57 percent in 2013), Uganda (54 percent in 2013), and Mozambique (24 percent in 2014). The Economist Intelligence Unit’s Global Microscope, which ranks countries based on policies for financial inclusion, ranked Rwanda eighth among 55 countries in 2016. 7Municipalities (translated in Kinyarwanda as “Umurenge”) are administrative subdivisions of the 30 districts that make up five provinces. In Rwanda, there are also several non- Umurenge SACCOs that already existed before the Umurenge program and where members come from the same profession. In the analysis, non-Umurenge SACCOs are therefore part of “other MFIs”. 9

U-SACCOs are owned by their members. They are legally set up as microfinance institutions that provide credit and savings facilities exclusively to members, and are financed mainly from their own resources. U-SACCOs operate according to the Finance and Cooperative laws and are supervised by the Rwanda Cooperative Agency and the NBR. They are located in both rural and urban areas, with each U-SACCO having one branch with membership drawn from the local community. Although set up as private cooperatives, U- SACCOs received public subsidies before reaching the break-even point. By the end of 2013, 85 percent of U-SACCOs were profitable and stopped receiving external funds (AFI, 2014). It is generally believed that the U-SACCO program significantly raised the share of the population with access to bank accounts, increasing financial inclusion especially in economically underprivileged areas. The FinScope surveys show that the fraction of adults who were granted loans increased from 4.6 percent in 2012 to 8.1 percent in 2016 (Table A1). Data from the credit register we use portray a similar picture, as the share of adults with an outstanding loan increased substantially across the country. In 2010, before the program was rolled out, 54 percent of loans were granted by commercial banks and the rest by other MFIs. When the program started, the number of loans granted by U-SACCOs increased dramatically, accounting to 34 percent of total loans in the credit register by the end of 2016 (Figure C1). These statistics suggest that the program coincided with significant gains in financial inclusion and are consistent with government and news reports (e.g., Randall, 2014). In this paper, we examine whether the microcredit expansion program had effects beyond increasing access to basic financial services. Specifically, we are not only interested in the program’s effect on unbanked individuals’ ability to take up loans from U-SACCOs and in the terms of those loans, but also in borrowers’ ability to build a credit history at U-SACCOs and to transition to commercial banks. 10

3 Data Our study employs detailed loan-level data from all credit institutions operating in Rwanda. The country has a detailed credit register maintained by the TransUnion Credit Reference Bureau, aprivateproviderwithglobaloperations, supervisedbytheNBR.Thecreditregister collects data on the loans granted by deposit-taking institutions—that is, commercial banks, U-SACCOs, and other MFIs—on a monthly basis with no threshold for loan size—a crucial element when examining microloans—and is highly representative of total banking sector loans, as shown in Figure C2.8 Our period of analysis is January 2008 to December 2016. The credit bureau data set covers 190,138 unique borrowers with active loans in any given month across 337 municipalities, corresponding to 4.9 million observations at the borrowerlender-monthlevel. Weaugmentthiscreditregisterwithaseparatedatasetwithinformation on individuals who never received a loan from any financial institution during our sample period. Specifically, we add to our sample 127,143 individuals (accounting for 40 percent of the total number of individuals) who either (i) applied for a loan during the sample period but were rejected or (ii) obtained their first loan after the end of the sample period (January 2017–June 2020). Together, our main data set contains 34.3 million observations in a balanced panel at the borrower-month level, corresponding to 317,281 unique individuals residing in 337 municipalities. For each loan, we have information on its amount (principal and outstanding), interest rate, maturity, and whether the loan is in arrears. Individuals are identified with a unique numericalcodewhichallowsustotracktheirlendingactivityovertimeandacrosslenders. In addition, forall317,281individualswithandwithoutloans, wehavedataonthemunicipality 8The figure compares total bank credit in billions of Rwandan francs (RWF) for all commercial banks operating in Rwanda from the credit register with aggregate statistics from the banks’ balance sheets. Aggregate bank balance sheet figures, representing total credit to individuals and firms, are available for the 16 active commercial banks operating in Rwanda over our sample period at quarterly frequency. To ensure comparability between the two series, we compute total bank credit in the credit register using loans to both individuals and firms in each quarter from the same set of banks. 11

where they reside9, as well as other characteristics such as age, gender, marital status, and sector of employment (government or non-government).10 Summary statistics for key variables are reported in Table 1. On average, 13.3 percent of individuals had a loan from any financial institution in a given month—3.4 percent from U- SACCOs, 6.3 percent from commercial banks, and 4.5 percent from other MFIs. The average loan balance amounts to 2.9 million Rwandan francs (RFW) (approximately USD 3,000) and the average interest rate on outstanding loans is 19.5 percent. U-SACCOs provide smaller, shorter-term, and more expensive loans than other credit institutions—see Table C1, which shows descriptive statistics separately for U-SACCOs, commercial banks, and other MFIs. We also collect several indicators of local economic and financial development measured at the municipality level in the pre-program period. These include bank presence (number of bank branches per capita), economic growth (change in nightlights between 2008 and 2010), economic development (nightlights in 2010), poverty (percentage of population below the poverty line), adult population, conflicts (number of violent events between 1997 and 2010), distance to capital (the travel time by car of the shortest route from Kigali to the municipality centroid, in hours) and rural (a dummy that identifies rural municipalities).11 9To identify the location accurately, we drop from the sample individuals residing in (i) municipalities that have the same name as districts (for example, Nyarugenge); (ii) municipalities that have the same name as provinces (for example, Kigali); and (iii) municipalities that are not uniquely assigned to a district (for example, Murambi). We also drop foreign currency loans, which account for less than one percent of the total number of loans. 10The non-government employee category refers to individuals who do not work in the public sector, including the unemployed. 11Since Rwanda does not collect data on output or consumption at the municipality level, we measure local economic activity with the share of population living under the national poverty line and with nightlights, a widely used measure of economic activity at the national and sub-national levels (Henderson et al., 2012). Nightlights are computed using data from satellite images from the National Oceanic and Atmospheric Administration of the U.S. Department of Commerce and exhibit a great degree of spatial heterogeneity. We acknowledge its limitations in measuring economic activity, and especially its variation over time (Jean et al., 2016). These problems are compounded by the change in the satellites used to capture the images around the time of the U-SACCO program and, as a result, we do not use that measure when we examine the real effects of the program. However, we validate the crosssectional use of nightlights—in our sample, there is a negative and significant correlation 12

4 U-SACCO Program and Access to Credit Timing of the Program Rollout. We identify the effect of the microcredit expansion program on access to credit by exploiting its staggered implementation, with different U- SACCOs granting their first loans in different months starting in 2011. U-SACCOs were granted licenses to start lending operations by the NBR if they met several requirements. Specifically, licensing formally required the U-SACCOs to collect a minimum of 10 million RWF in deposits (around 10,000 USD), have sufficient trained staff (one manager, accountant, loan officer, loan recovery officer, and cashiers, all typically hired locally), develop a saving and credit policy manual, and pass onsite inspections from the NBR. This cautious and protracted process resulted in a staggered opening of U-SACCOs between November 2011 and April 2016, as illustrated in Figure 1, which depicts the number of institutions that granted their first loan each month. During our sample period (January 2008–December 2016), U-SACCOs started lending operations in 301 out of 337 municipalities. The key identification assumption is that the timing of the program rollout across municipalities is unrelated to (unobserved) loan demand. A systematic correlation between the timing of the implementation of the program and loan demand could arise if U-SACCOs were set up earlier in relatively more dynamic and economically developed areas. To the extent more developed areas also have higher loan demand, the endogeneity bias would generate a spurious positive correlation between the opening of a U-SACCO and access to credit. To check this identifying assumption, we examine the correlation between the time elapsed before U-SACCOs start granting loans and a wide range of variables capturing economic development and other municipality characteristics before the program, as well as the pre-program financial position of each U-SACCO—including size (total number of members, totalassets, andtotaldeposits)andfinancialratios(liquidity-to-assetsandcapital-to-assets). We estimate a survival model (Cox, 1972) in which the dependent variable is the waiting time (in months) until the granting of the first loan (the “event”) in a given municipality. (−0.73) between nightlights and the poverty headcount, even within districts. 13

Figure 1: Staggered Implementation of the U-SACCO Program 30 20 10 0 naoL tsriF eht gnidivorP sOCCAS-U fo rebmuN 2008m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 2016m1 Date of First U-SACCO loan Note: The figure plots the number of U-SACCOs that granted their first loan in a given month during our sample period (January 2008–December 2016). Source: Rwandan Credit Reference Bureau. As reported in Table 2, the results reveal no systematic association between any of these variables and the timing of the program rollout, both when each variable is included independently(columns1–13)aswellaswithallvariablestogether(column14). Importantly, lending activities at U-SACCOs did not start earlier where U-SACCOs had acquired more members, capital, or more deposits.12 These results suggest that pre-program local economic and financial conditions do not predict the timing the program’s rollout. This finding is robust to an alternative approach by which we split municipalities between early and late starters (the first 75 percent versus the last 25 percent municipalities) and compare the two groups along the same set of potential confounders (Table C3).13 12Controlling for total deposits is particularly important given the minimum deposit requirement for U-SACCOs licensing. This requirement, however, had become not binding for most U-SACCOs by the time they were licensed to grant loans. Balance sheet data show that 61 percent of all U-SACCOs met the minimum deposit requirement by the end of 2010 (on average, by a multiple of 2.2) and all but five U-SACCOs did so by the end of 2011 (on average, by a multiple of 5.4). The summary statistics for the pre-program balance sheet characteristics of U-SACCOs—that is, as of end-2010—are reported in Table C2. 13Our identifying assumption is also supported by anecdotal evidence regarding the de- 14

Identification Strategy. To identify the effect of the microcredit expansion program on access to credit, we exploit its staggered implementation over time across municipalities in a balanced borrower-month panel. The baseline specification we estimate is as follows P(Loan ) = βPost U-SACCO +δ(cid:48)X +α +φ +υ t+ε , (1) imt mt i m t m imt where Loan is a dummy variable equal to one if individual i residing in municipality m has imt an outstanding loan with any (or alternatively, a particular type of) financial institution at time (year:month) t, and zero otherwise. Our main variable of interest is Post U-SACCO , mt a dummy variable equal to one after a U-SACCO starts its lending activities in a given municipality m at time t, and zero beforehand. The coefficient of interest β is identified by comparing, before and after the U-SACCO in municipality m starts operating, the probability of borrowers in municipality m having a loan relative to individuals in municipalities without yet an active U-SACCO. X are time-invariant individual characteristics, including i gender, marital status, age, and sector of employment. Municipality fixed effects α control m for unobserved time-invariant spatial factors that might correlate both with the timing of U-SACCO openings and with financial access. Time (year:month) fixed effects φ absorb t common time-varying shocks, such as changes in economic conditions affecting all municipalities at the same time. We also include municipality-specific time trends υ t to minimize m the possibility that our estimates are confounded by differential trends in financial access across municipalities or other unobserved time-varying municipality attributes such as credit demand. WeestimateEquation1asalinearprobabilitymodelwithstandarderrorsclustered at the municipality level. ployment of the program, which suggests that the timing of U-SACCO openings was largely related to idiosyncratic logistical and administrative issues, rather than local economic conditions. Given the novelty, scale, and risks associated with the program, NBR staff were cautious in licensing U-SACCOs compared with other financial institutions. Once the legal requirements were in place (including minimum deposits and staff), the NBR scheduled field visits by inspectors to each U-SACCO to assess if procedures and staffing were adequate. The NBR then examined each recommendation before issuing the license to issue loans. 15

Figure 2: Probability of Obtaining a Loan Before and After the U-SACCO Program )naoL(rP 61. 21. 80. 40. 0 40.- ]63-t<[ ]13-t,63-t[ ]52-t,03-t[ ]91-t,42-t[ ]31-t,81-t[ ]7-t,21-t[ ]1-t,6-t[ ]6+t,t[ ]21+t,7+t[ ]81+t,31+t[ ]42+t,91+t[ ]03+t,52+t[ ]63+t,13+t[ ]63+t>[ Months Relative to Implementation (a) Any Lender )naoL(rP 21. 1. 80. 60. 40. 20. 0 20.- ]63-t<[ ]13-t,63-t[ ]52-t,03-t[ ]91-t,42-t[ ]31-t,81-t[ ]7-t,21-t[ ]1-t,6-t[ ]6+t,t[ ]21+t,7+t[ ]81+t,31+t[ ]42+t,91+t[ ]03+t,52+t[ ]63+t,13+t[ ]63+t>[ Months Relative to Implementation Loan from U-SACCO Loan from Bank (b) U-SACCOs vs Commercial Banks Note: The figure shows the effect of the U-SACCO program on the probability of an individual obtaining a loan from any financial institution (U-SACCOs, other MFIs, or commercial banks, Panel A) and separately for U-SACCOs and commercial banks (Panel B). The charts plot the estimated coefficients and the associated 90 percent confidence intervals of model 1 with interaction terms between Post U-SACCO and a set of six-month time interval dummt mies centered on the municipality-specific month of program rollout, as reported in column 3 (Panel A) and columns 4–5 (Panel B) of Table 3. The (lack of) effects of the U-SACCO program on the probability of an individual obtaining a loan from other MFIs are shown in Figure C3. Source: Rwandan Credit Reference Bureau. BaselineResults. ThebaselineresultsreportedinTable3showapositiveandstatistically significant effect of the U-SACCO program on the probability that an individual obtains a loan. Columns 1–3 refer to loans granted by any financial institution—that is, U-SACCOs, commercial banks, and other MFIs. The estimates in column 2—our preferred specification withmunicipalitytimetrends—indicatethattheU-SACCOprogramraisedtheprobabilityof anindividualhavinganoutstandingloanby2.1percentagepoints. Thiseffectiseconomically sizable given that the average share of individuals with an outstanding loan before the program is 6.3 percent. Figure C4 illustrates this result by plotting the share of individuals with loans, by municipality, before and after the program’s rollout. To rule out potential anticipation effects, which could undermine our identification strategy, we explore the dynamic effects of the U-SACCO program. Specifically, we split the β 16

coefficient according to the time that elapsed before and after the implementation of the program, using six-month intervals centered on the timing of U-SACCO openings. The estimated coefficients and associated confidence intervals are shown in Figure 2 (Panel A) and reported in column 3 of Table 3. The estimates indicate that the likelihood of having a loan increases after the program was implemented and that this effect becomes stronger over time. The increasing magnitude of the effect over time suggests that the program had sustained effects on financial access as opposed to a one-off (transitory) effect. Figure 2 also shows that the parallel trends assumption holds in our setting—the point estimates before the program are close to zero and statistically insignificant in the entire pre-program period. After the implementation of the program, U-SACCOs likely competed for new borrowers with local lenders such as commercial banks and other MFIs. U-SACCOs also started screening potential borrowers from a population of lower income-earners, who were largely unknown to the financial sector. Their credit and repayment information was submitted to the credit bureau—available to all lenders. Therefore, we next examine if the overall effect of the program is entirely driven by the U-SACCOs or if it spills over to other lenders as well. To this end, we analyze loans from U-SACCOs, commercial banks, and other MFIs separately, with the dependent variables equal to one for individuals who have a loan at each type of lender in a given month. The results in columns 4–6 of Table 3 show that while the rise in access to credit is mostly driven by the U-SACCOs, there is a delayed positive spillover effect at commercial banks. Instead, other MFIs do not expand access to credit as a result of the U-SACCO program. We plot our estimates on the probability of having an outstanding loan at U-SACCOs and, separately, at commercial banks in Figure 2 (Panel B).14 The figure confirms two key results. First, the positive effect of the program on access to credit is mostly driven by U-SACCOs, with the likelihood of an individual having an outstanding loan at a U-SACCO rising over time to reach almost 10 additional percentage points than before the program. 14Figure C3 plots the same results for other MFIs and shows no significant effects after the program as well as no preexisting trends. 17

Second, there are spillover effects of the program to commercial banks, which catch up with a lag. In fact, starting in the second year of the program, the probability of obtaining a loan from a commercial bank increases up to three percentage points compared with before the program. This lagged positive spillover effect on bank lending suggests that banks may be tapping into a new customer base after U-SACCO openings. In fact, previously unbanked individuals can build a credit history at newly opened U-SACCOs and may be able to signal their quality to other lenders through the credit bureau. We explore this question explicitly as well as the transition of U-SACCO borrowers to commercial banks in Section 5. Spatial Heterogeneity. The positive effect of the U-SACCO program on access to credit likely varies across municipalities depending on the level of economic and financial development before the program. In line with the large literature on the importance of finance for growth and economic convergence (e.g., Beck et al., 2007; Bruhn and Love, 2014; Agarwal et al., 2017), we expect the increase in access to credit to be more pronounced in less developed areas. To test this hypothesis, we exploit variation across municipalities in several ex ante indicators of development, including the number of bank branches per capita, the level and growth rate of nightlights, and the poverty rate. We also examine if there are differential effects for municipalities that are geographically remote and located in rural areas. For all variables except the rural dummy, we follow Bruhn and Love (2014) and split each continuous variable around the 75th percentile of its distribution to create a dummy variable for less versus more developed municipalities. Across all measures, we find that the program’s average effect on access to credit is stronger in municipalities with lower initial levels of development (Table 4), as indicated by the positive and statistically significant coefficients on the interaction terms between Post U-SACCO and each of the dummy variables. 4.1 Robustness and Falsification Tests Weconductaseriesofrobustnessandfalsificationteststogaugethesensitivityofourbaseline results to a range of methodological choices. 18

AlternativeDifference-in-differencesEstimator. Thestandarddifference-in-differences estimator, which exploits variation across municipalities that receive treatment in different periods (Equation 1), recovers a weighted average of all possible two-group/two-period difference-in-differences estimators in our sample (Goodman-Bacon, 2021). De Chaisemartin and d’Haultfoeuille (2020) demonstrate that when treatment effects are heterogeneous across groups and periods, some of these weights can be negative, and they propose an alternative estimator that provides unbiased estimates under treatment effect heterogeneity. The results when implementing their approach confirm that the parallel trends assumption holds in our setting and that the program boosted access to credit—not only directly through U-SACCOs but also indirectly through positive spillover effects to commercial banks (Table C4).15 Alternative Data Structures. The large share of zeros in the dependent variable in the balanced borrower-month panel we use in Table 3 may cause a bias in our coefficient of interest. Asaresult, weconsideralternativeaggregationsofthedata, including(i)collapsing the data over the time dimension and constructing a balanced panel at the borrower-quarter or borrower-year levels, and (ii) collapsing the data over the municipality dimension such that the dependent variable is the number of individuals (in absolute terms or as a share of total population) with an outstanding loan in a given municipality-month combination. The latter approach is particularly relevant, as it leads to a reduction in the share of zeros in the dependent variable from 87 percent to eight percent. The results show that the baseline effects of the U-SACCO program are robust to these alternative data structures (Tables C5 and C6).16 15We use the same set of controls and fixed effects as in columns 3–6 of Table 3, including municipality time trends. Standard errors are cluster bootstrapped at the municipality level with 1,000 replications. Note that the test for pre-trends of De Chaisemartin and d’Haultfoeuille (2020) differs from the standard event study pre-trend tests (for example, Figure 2), which would be invalid when treatment effects are indeed heterogeneous. 16Several additional methodological choices have no material effect on our baseline results. First, we replace the borrower-level controls with individual fixed effects (Table C7, columns 1–5). Second, we ensure our baseline results are not driven by the inclusion of borrowers who do not obtain a loan during the sample period (Table C7, columns 6–10). Third, we include 19

Randomized Treatment Dates. We conduct a falsification test to rule out the potential concernthatourresultsaredrivenbycoincidenteventsotherthantheopeningofU-SACCOs. For this purpose, for each municipality we randomly assign the U-SACCO opening date in the interval January 2008–December 2016 and repeat this exercise 1,000 times. Table C9 reports the average coefficients across simulations corresponding to the random assignments of the Post U-SACCO variable across municipalities. The estimated coefficients are very close to zero and statistically insignificant throughout, suggesting that our main findings are not driven by a spurious correlation between the rollout of the program and access to credit. External Validity with Survey Data. We assess the external validity of our results using survey data on financial access from the 2012 and 2016 FinScope surveys. The main differences from the baseline analysis is that we only have two time periods (2012 and 2016) and we observe the district where the borrower resides instead of the municipality. We define as dependent variables a dummy equal to one for survey respondents who report having access to any type of loan and then distinguish between informal or formal loans and between loans from U-SACCOs or banks. We also allow for changes in access to credit before and after the program to differ across districts with different initial bank presence, following the evidence in Table 4. The results confirm our baseline findings that access to formal loans increased in the post-program period, especially in low bank presence areas, and mostly because of U-SACCOs. Similar to our baseline results, we also find a positive and statistically significant expansion of bank lending, albeit with a smaller coefficient than that for U-SACCOs. Furthermore, the results support the view that the U-SACCO program did nothavecrowding-outeffectsoninformallending,echoingseveralpreviousstudies(Angelucci et al., 2015; Karlan and Zinman, 2019). We discuss these results further and relate them to the literature in Appendix A. quadratic (instead of linear) municipality-specific time trends, which could pick up more complex convergence dynamics (Table C8, columns 1–5). Finally, we cluster the standard errors at the district (instead of at the municipality) level (Table C8, columns 6–10). 20

5 Borrower Transition to Commercial Banks The first part of our analysis documents the effect of the microcredit expansion program on financial inclusion. We show that the program increased access to credit, particularly in less developed municipalities, and mainly through the newly set up U-SACCOs. Importantly, we also find evidence of positive spillovers to bank lending starting about one year after the program rollout. This lagged effect suggests that banks may be attracting new customers after U-SACCO openings. With a credit information sharing system in place, unbanked individuals can take loans from U-SACCOs, build a credit history, and be able to signal their type to other lenders. To test this hypothesis, we zoom in on the lending activities of first-time U-SACCO borrowers and, by tracing their borrowing history in the credit register over time, examine the extent to which they are able to subsequently obtain loans from commercial banks. Switching Activity. We focus on individuals that were previously unbanked, had a loan at a U-SACCO, and obtained a new loan at a commercial bank during the period of analysis. Using this definition, we identify 2,910 switchers from U-SACCOs to commercial banks.17 In relative terms, of all borrowers who obtained their first loan through a U-SACCO, eight percent switched to a commercial bank starting in June 2012—of those who obtained more than one loan, 15 percent switched to commercial banks.18 From the banks’ perspective and considering all new bank borrowers since the program, five percent previously had a loan with a U-SACCO. This share increased over time, from less than one percent in 2012 to nine 17About half of the switchers start lending exclusively from banks, while the rest keep the relationship with the U-SACCO. The number of borrowers moving from commercial banks to U-SACCOs is negligible. 18While these numbers may not appear strikingly large, it is important to note that our analysis focuses on a particular type of switching—from only a subset of MFIs to commercial banks—and that switching across lenders rarely occurs in both developing and developed economies. Forinstance,onlyfourandeightpercentofU.K.andU.S.borrowers,respectively, switch banks in any given year (The Economist, 2019). Switching rates are comparable for nonfinancial firms, estimated at 4.5 percent per year in Bolivia (Ioannidou and Ongena, 2010) and 5.9 percent per year in Portugal (Bonfim et al., 2021) 21

percent in 2016.19 The rapid increase in switching activity from MFIs—that is, U-SACCOs and other MFIs—to banks over the sample period is shown in Figure C5. We plot the coefficients from a regression similar to the baseline approach but on a municipality-month balanced panel, where the dependent variable is the number of switching loans from MFIs to commercialbanks. Consistentwiththeevidenceshownonthespillovereffectsoftheprogram to commercial banks, switching activity estimated with this alternative data structure starts in the second year after the start of the U-SACCO program. Importantly, the analysis does not show any pre-existing trend in the intensity of switching across municipalities. We start by characterizing the borrowers who switch from U-SACCOs to commercial banks, particularly their ex ante credit risk profile. For this purpose, we exploit the richness of the credit register to build several measures of risk based on the borrower’s credit history. These measures mimic the ingredients of a consumer credit risk measure such as the FICO score in the United States and include payment history (dummy variables for nonperforming loans at any lender within one or two years or any time before the new loan), credit capacity (number of lending relationships, number of loans, total outstanding loan volume before the new loan), and length of credit history (number of years since first loan). As the result of the lack of data on income, we use the employment status dummy (which identifies government employees)asaproxy, aspublic-sectoremploymentisassociatedwithhigherandmore-stable wage income (Teal, 2011). Descriptive statistics in columns 1–2 of Table C11 show that switchers differ along many characteristicsfromnon-switchers—thatis,U-SACCOborrowerswhodonotswitchtobanks. For instance, they have a relatively longer credit history and greater credit capacity (more borrowing relationships and higher outstanding loan volume). This comparison suggests that switchers have higher credit quality than other U-SACCO borrowers. By contrast, 19The characteristics of first-time U-SACCO borrowers have little common support with thosefirstservedbybanks, withstatisticallysignificantdifferencesbetweenthetwogroupsof first-time borrowers along multiple dimensions, including individual characteristics, location, loan type, and loan terms (Table C10). 22

they are ex ante riskier than other bank borrowers—with shorter credit history, fewer loans, and markedly lower total outstanding loan volume (columns 1 versus 3). These systematic differences emphasize the importance of controlling for ex ante credit risk profile of switchers and borrowers in the comparator groups. Empirical Strategy. Similar to Ioannidou and Ongena (2010), we consider two control groups to estimate the link between switching and financial outcomes: (i) new loans granted by U-SACCOs to similar borrowers who do not switch (“non-switchers”); and (ii) new loans extendedbybankstosimilarexistingbankborrowers(“otherbankborrowers”). Furthermore, loancontracttermsareanequilibriumoutcomethatreflectsborrowerriskandeconomicconditions. To control for these factors, we use the Abadie and Imbens (2011) bias-corrected nearest-neighbor matching estimator and match loans granted in the same month. Then, we select the nearest-neighbors of each switching loan based on a wide range of borrower, loan, and location characteristics. These include (i) gender, marital status, age, and employment status, (ii) loan type (whether the loan is a mortgage or is collateralized), (iii) credit history (number of years since first loan, number of lending relationships, number of loans, and total outstanding loan amount in the month before the new loan), (iv) pre-program municipality characteristics (bank presence, nightlights and nightlight growth, poverty rate, population size, number of past conflicts, distance from the capital, and a dummy for rural municipalities) and, critically, (v) proxies for payment history—a key determinant of credit scores—computed as dummies for NPLs with any lender within one or two years or any time before the new loan.20 Switching and Loan Terms. We start the analysis by comparing loan terms across switchersandnon-switchers. Asreportedincolumns1–3ofTable5, switchingloansatbanks are relatively larger, cheaper, and longer term compared with loans granted by U-SACCOs in the same month to non-switchers. These effects are economically meaningful—switching 20The results are robust to propensity score matching based on the same set of matching characteristics as in our baseline approach (Table C12). 23

loans are on average larger by 0.812 million RWF (around 800 USD), cheaper by 648 basis points, and have maturity that is longer by almost 10 months. Second, we compare loan terms for switchers with those of new loans granted by banks in the same month to other similar borrowers. As reported in columns 4–6 of Table 5, switching loans at banks have similar amounts and interest rates compared with loans to existing bank borrowers, but have a slightly longer maturity (three months). Switching and Default Risk. Several studies emphasize the risks associated with rapid expansions of microcredit (e.g., Schicks, 2014; Zinman, 2014). On the one hand, if switchers are more likely to default than bank borrowers, rapid growth in access to credit can increase the likelihood of loan losses at banks, with implications for financial stability. On the other hand, banks may take advantage of the screening done by U-SACCOs and cream-skim from the pool of microfinance borrowers, which would leave microlenders with riskier borrowers and pose challenges to the sustainability of their business model. We examine default risk usinginformationonnon-performingloans(atleast90dayspastdue)withinoneortwoyears after loan origination, or over the lifetime of the loan. In addition, we add loan contract terms (amount, interest rate, and maturity) to the set of matching variables because they can influence the probability of repayment. Comparing the performance of switching and non-switching loans, and conditional on loan contract terms and ex ante credit risk, we find that switching loans at banks are less likely to become non-performing than loans granted by U-SACCOstoborrowersthandidnotswitch(Table6, columns1–3).21 Inaddition, switching loans are not ex post riskier than loans granted by banks to other borrowers (columns 4–6).22 21It is important to note that our analysis covers a period of high and sustained economic growth. Asaresult, wedonotobservehowU-SACCOborrowers’repaymentcapacityevolves overaperiodofseverefinancialstress. Borrowerswhoaremorevulnerabletonegativeshocks could see significantly weakened ability to repay and become excluded from the loan market altogether. This situation highlights the perils of improved credit access after MFI entry for individuals with early exposure to greater risk. 22The results in Tables 5–6 are robust to a within-lender analysis, by which we narrow the control groups to loans granted by the same U-SACCO the switcher left (columns 1–3 of Tables C13–C14, and respectively, to loans issued by the same bank that took the switcher 24

Figure 3: Commercial Banks’ Branch Expansion After the U-SACCO Program 120 115 110 105 100 sehcnarB knaB .oN htworG 125 120 115 110 105 100 t=0 t=1 t=2 t=3 t>3 Years Relative to Implementation Low Bank Presence Municipalities High Bank Presence Municipalities (a) Low vs high bank presence sehcnarB knaB .oN htworG t=0 t=1 t=2 t=3 t>3 Years Relative to Implementation Low Bank Presence Municipalities with Low NPLs High Bank Presence Municipalities with Low NPLs Low Bank Presence Municipalities with High NPLs High Bank Presence Municipalities with High NPLs (b) Low vs high default rates Note: ThefigureinPanel(a)showsthegrowthofcommercialbankbranchesinmunicipalities with low (solid line) versus high (dotted line) pre-program bank presence. The number of branches is calculated relative to the year the U-SACCO started granting loans in a given municipality—normalized at 100 for both low and high bank presence municipalities. Low bank presence municipalities are below the 75th percentile of the distribution of bank branches per capita before the program (as of 2011). The figure in Panel (b) further unpacks the sample of municipalities into those with low and high non-performing loans (NPLs). Low NPL municipalities have an NPL-to-total loans ratio below the 75th percentile of the sample distribution. Source: National Bank of Rwanda. Finally, we look for direct evidence of banks’ cream-skimming from U-SACCOs’ borrower pool. To this end, we examine the ex post default risk of switchers on their loans at U- SACCOs before switching. As reported in Table C16, we find that switchers are indeed less likely to default compared with non-switchers at U-SACCOs conditional on loan contract terms and ex ante credit risk, supporting the idea that banks cream-skim the less risky borrowers from the pool of U-SACCO borrowers. Bank Branch Expansion. The analysis of switching loans suggests that switchers obtain better loan terms at banks, which cream-skim the best borrowers from U-SACCOs. Does the (columns 4–6 of Tables C13–C14). Furthermore, for the same switcher, we compare the terms of the first loan with subsequent loans at the bank and find no significant changes in lending terms over time (Table C15). 25

increased lending activity targeting U-SACCO borrowers translate in more bank presence on the ground? To provide suggestive evidence on this question, we use annual data on the geographic footprint of bank branches across municipalities during 2011–16. Panel A in Figure 3 depicts the growth in bank branches in the years when U-SACCOs open. We see that banks expand their branch network relatively more in municipalities with lower initial bank presence, bringing borrowers and banks closer. Furthermore, panel B in Figure 3 shows that this expansion was concentrated in low-risk municipalities—where risk is proxied by the overall share of NPLs. This evidence is consistent with the FinScope surveys that show that the share of adults located within 30 minutes to the nearest commercial bank increased from 21 percent in 2012 to 25 percent in 2016 (FinScope, 2012, 2016). 6 Mechanisms and Real Effects Our analysis so far shows that the U-SACCO program increased access to credit and facilitated the transition of an economically significant number of previously unbanked borrowers tocommercialbanks. Inthisfinalsection, weprovideevidenceonthemechanismsthatmight explain our results, focusing on U-SACCOs’ lending constraints and screening technology, and document the effects of the program on the local economy. Lending Constraints at U-SACCOs. We show a lagged spillover effect of the program on commercial bank lending and that switchers obtain better loan terms at banks. What might explain the fact that U-SACCOs are unable to offer loan terms on par with those from banks? Put differently, why are they letting go of their good clients? To explore this question, we conjecture that lending capacity related to balance sheet and regulatory constraints plays an important role. U-SACCOs must meet specific requirements set up by the NBR in its regulation of the microfinance sector. However, in light of their heavy reliance on callable deposits as a source of funds (with average maturity around five months) and low capacity for treasury 26

and loan portfolio management, U-SACCOs are subject to tighter prudential regulation not only compared with banks—as MFIs often are—but also relative to international practices. Initially, the liquidity ratio was set at 80 percent for all U-SACCOs. By December 2013, the liquidity threshold was lowered to 30 percent, yet in practice it remained elevated at 72.7 percent (AFI, 2014). By contrast, commercial banks, subject to the Basel II regulatory framework, did not have to meet a minimum liquidity ratio. U-SACCOs were also subject to steep provisioning rates (for example, 50 percent for loans 90+ days past due, and 100 percent for loans 180+ days past due), and were prohibited from giving out loans if they had an overall NPL ratio above 10 percent—a regulation that was seen as potentially disruptive to lending relationships and service delivery (MINECOFIN, 2013).23 For a direct test of lending constraints at U-SACCOs, we assemble data on U-SACCOs and commercial bank balance sheets before the program rollout, and relate lender size and financial ratios (measured at end-2010) to loan terms, zooming in on this relationship for U-SACCOs. Specifically, we regress loan terms (amount, interest rate, and maturity) on lender balance sheet characteristics and a dummy equal to one if the lender is a U-SACCO and zero if it is a bank. The results indicate that U-SACCOs offer smaller, more expensive, and shorter-duration loans than banks (Table C17, columns 1, 3, and 5), with statistically significant coefficients at the one percent level. However, once we control for balance sheet size (total deposits), capital ratio, and liquidity ratio, the coefficient estimates on the U- SACCO dummy become statistically insignificant (Table C17, columns 2, 4, and 6). These results show that balance sheet characteristics explain why U-SACCOs offer worse loan terms than banks do, supporting the idea of lending constraints at U-SACCOs. We also test for the presence of constraints indirectly, by exploiting the notion of relationship lending. Our results, discussed in Appendix B, show that borrowers obtain better loan terms as 23There were also constraints on maximum credit exposures to a single firm or individual set at five percent of total net worth in general, 10 percent for the savings and credit cooperatives with NPL ratios below five percent. Furthermore, maximum single-loan exposures could not exceed 2.5 percent of total deposits. 27

relationship length increases and this effect is weaker for U-SACCOs than for banks. U-SACCOs’ Lending Technology. Our findings suggest that U-SACCOs play a critical screening role for the unbanked population. Coupled with the presence of the credit bureau, which tracks individuals’ borrowing activities and is accessible by all lenders, U-SACCOs’ lending technology—by gathering community-level information—might facilitate the transition of borrowers to commercial banks.24 How do U-SACCOs screen their borrowers? Does theirapproachdifferfrombanks? Toanswerthesequestions, wefirstdiscusstheU-SACCOs’ lending technology while comparing it with banks, and then turn to the data to pin down U-SACCOs’ relative advantage in screening. There are two key differences in the U-SACCOs’ lending technology compared with a commercial bank. First, U-SACCO borrowers are also members of the U-SACCO cooperative, which means borrowers know each other, the staff, and the loan officers—all of whom are local hires. Strong community ties, cultural proximity, and social enforcement of norms increase access to finance and the likelihood of loan repayment (Guiso et al., 2004; Feigenberg et al., 2013; Fisman et al., 2017). Second, U-SACCO members have access to a “finance forum” comprising members of the local community, who disseminate information about financial products and educate potential borrowers about the importance of repayment. This community effort, unique to U-SACCOs, creates awareness about financial services usage, cash flow management, and financial planning.25 24Mitra et al. (2018) features an application of a similar mechanism in the context of agricultural markets. Recent studies also examine new mechanisms by which borrowers can signal their type through means other than the traditional repayment of existing loans. For instance, Azevedo et al. (2019) highlight a screening mechanism by which financial outcomes improve if the individual is merely found eligible for a loan. Arraiz et al. (2020) document free-riding in lenders’ screening efforts in the context of SME lending by showing that loan approvalbyonelendercausesotherlenderstoapprovethesameborrowers. Thisphenomenon can cause underinvestment in screening new borrowers and hinder financial inclusion. 25The 2012 FinScope survey revealed that U-SACCO members had a collective sense of ownership, often referring to U-SACCOs as “our bank”—a bank managed by people they knew, and was local and accessible. Half of the respondents chose U-SACCOs as the most trusted lender and 31 percent of respondents chose banks. In the 2016 survey, these shares changed to 57 percent and 26 percent, respectively, signaling the success of U-SACCOs in 28

In Table 7 we compare loan terms offered by banks with U-SACCO-to-bank switchers with those offered by banks to first-time borrowers (columns 1–3). While we cannot control for ex ante credit risk, as the variables measuring it are unavailable by definition for the control group, we match the loans on all borrower, loan, and location characteristics. Additionally, we compare loan terms offered by banks to U-SACCO-to-bank switchers with those offered by banks to other-MFI-to-bank switchers (columns 3–6)—the summary statistics and unconditional differences across the different groups are reported in Table C18. Other MFIs serve as a useful control group because, like commercial banks, they target wealthier individuals and are concentrated in urban areas. In both comparisons, we find that banks systematically offer lower interest rates and longer-maturity loans to borrowers who had their first lending relationship at the U-SACCO and then switched to banks. Overall, while we cannot fully rule out that the transition to commercial banks may be driven by the income effect of borrowing from U-SACCOs, our results suggest that the U-SACCO credit history is valuable to banks and is consistent with U-SACCOs’ lending technology playing a key role in the transition of first-time borrowers to commercial banks. It is important to highlight that the availability of a credit bureau accessible to all lenders iscriticalforthistransitiontooccur. WhilenodataareavailableonU-SACCOs’information requests to the credit bureau, such data have been collected for commercial banks since 2018 (outside our sample period). These data allow us to explore the extent to which banks use information from the credit bureau in their lending decisions. As shown in Table C19, there is a strong and positive correlation between the number of information requests and the number of new loans extended by banks—even within bank and a municipality-quarter pair—suggesting an “information channel” in banks’ lending decisions. A complementary building trust. Focus group participants reported U-SACCOs as having faster service, better customer care, and lower account fees than commercial banks (AFI, 2014). Hakizimfura et al.(2020)conductedafieldexperimentinvolving200U-SACCOsinRwandatotestmodelsof local financial education delivery and found improvements in financial planning skills for the participants from U-SACCOs that were able to select the trainers from the local community (as opposed to the U-SACCO management). 29

explanation of borrowers’ transition from U-SACCOs to banks is that access to microcredit raises incomes, making borrowers more attractive for banks. We turn to this point in the next section, where we test for the program’s real effects. Real Effects. To analyze the effect of the program on economic activity, we assembled two types of microdata—poverty indicators from nationally representative household surveys and information on small-firm formation from the establishment census. Forthepovertyanalysis, weusedatafromthe2010–11and2016–17IntegratedHousehold Living Conditions Surveys, which identify surveyed households as poor or extremely poor based on the $1.09/day international (consumption) poverty line and 0.65× of this poverty line, respectively. For each household, we observe poverty status, district, and characteristics such as size and rural status. We stack the surveys and examine the link between the probability that a household is in poverty or extreme poverty after the program implementation (captured by “Post” dummy variable equal to one for the 2016–17 survey), allowing for differential effects based on the pre-program level of financial development—that is, high versus low bank presence districts. The results, reported in Table 8, show that households were 5.6 percent less likely to be poor (and 4.1 percent less likely to be extremely poor) after the program implementation (columns 1 and 4). The poverty reduction effect of the program is stronger for households in districts with low bank presence (columns 2 and 5). We also examine the real effects of access to commercial bank loans by U-SACCO borrowers who transitioned to banks. We capture this effect with the fraction of switchers in each district (relative to all U-SACCO borrowers) over the 2012–16 period, computed from the credit register. Specifically, we unpack the program’s effect in low bank presence municipalities by level of switching intensity. We expect the program’s poverty-reducing effect to be further boosted by a higher share of switchers because switchers obtain loans on relatively more advantageous terms from banks. As reported in columns 3 and 6 of Table 8, this is indeed the case—the poverty reduction effects of the program are driven by districts with both low pre-program bank presence and a high fraction of switchers during the program. 30

We complement these results with evidence on the program’s effect on firm formation using the 2014 Establishment Census, which provides the municipality and date when each firmwascreatedbetweenJanuary2008andDecember2014. Werestrictoursampletoprivate sector firms and examine three outcome variables that are tightly related to the availability of microcredit: the rate at which new microenterprises and sole proprietorships are formed, and the number of jobs created by new firms (regardless of size). Given that we observe firm creation and new employment at the municipality-month level, we can exploit the staggered rollout of the program for identification. The results, shown in Table C20, indicate that municipalities where U-SACCOs start granting loans earlier tend to have a higher rate of firm formation (columns 1–4)—a result that speaks directly to the positive link between microcredit and entrepreneurship (Bruhn and Love, 2014; Banerjee et al., 2015, 2019). The magnitude of the effects is larger in municipalities with ex ante low bank presence. We also find that the program boosted private-sector employment (columns 5–6). Overall, these positive effects of the U-SACCO program on the local economy are inconsistent with the idea of negative local general-equilibrium effects caused by crowding-out of other types of lending—such as informal lending—and are therefore consistent with the results based on the FinScope surveys (Appendix A). They also increase our confidence that the financial access boosting effect of U-SACCOs translated into positive local real effects. 7 Conclusion We exploit the staggered implementation of a large-scale microfinance expansion program to analyzeitseffectsonaccesstocreditandonthetransitionofpreviouslyunbankedindividuals to commercial banks. Using data from a unique administrative dataset comprising the universe of individual loans granted by all financial institutions in Rwanda between 2008 and 2016, we show that the program raised the likelihood of access to loans for the previously unbanked population. While this effect is mostly driven by newly set up U-SACCOs, we 31

show that there were also positive spillover effects on bank lending starting one year after implementation. The program also generated positive effects on the local economy through poverty reduction and firm creation. A key result of our paper is that an economically sizable share of U-SACCO borrowers switch to commercial banks, which grant them larger, cheaper, and longer-term loans than the loans they might have received from U-SACCOs. The borrowers that transition from U-SACCOs to banks are less risky than non-switchers but have comparable ex post risk of default, adjusted for ex ante credit risk, to other commercial bank borrowers. A credit historyataU-SACCOandtheabilitytosignalqualitythroughthecreditbureau—accessible to both microfinance institutions and banks—is key to the program enabling this transition. Borrowers at U-SACCOs who switch to banks receive better loan terms than first-time bank borrowers with no credit history. While our results point to the importance of the U-SACCO lending technology to facilitate access to bank credit, more research is needed to isolate this channel from the potential income effect of U-SACCO borrowing. Overall, our analysis supports the notion that microfinance institutions play an important screening role for the unbanked population in markets where scarce collateral and weak institutions impede commercial bank activities. Together with a credit bureau that is universally accessible, the microfinance sector can support the expansion of commercial banks in traditionally underserved markets, boosting access to financial services and supporting local development. However, since banks cream-skim the best borrowers, microlenders may be left with a pool of relatively riskier borrowers, which may raise questions about the feasibility of their business models and create risks for financial stability. Indeed, aggregate statistics show that the NPLs at U-SACCOs increased between December 2014 and June 2016 (from 7.1 to 12.5 percent), while NPLs at commercial banks remained flat (NBR, 2016). As the program was funded through government subsidies, our results have important implications for aggregate welfare and efficiency. Previous theoretical work shows that government interventions in credit markets where lending is inefficiently low due to information 32

asymmetries can increase welfare (Mankiw, 1986). Our evidence suggests that the program acted as an indirect subsidy to commercial banks, which took advantage of U-SACCOs’ screening to expand credit to good borrowers in underserved but promising markets. These findings show that accounting for the general equilibrium effects can change the cost-benefit calculus of microfinance subsidies (Cull et al., 2009). In particular, the transition of good borrowers from microfinance to banks can be a goal of the social planner, justifying subsidies up to a point. We believe that the interplay between the microfinance and commercial banking sectors, and the government’s role in this area, deserve further research. References Abadie, Alberto, and Guido W Imbens (2011) ‘Bias-corrected matching estimators for average treatment effects.’ Journal of Business & Economic Statistics 29(1), 1–11 AFI (2014) ‘Rwanda’s financial inclusion success story: Umurenge SACCOs.’ Alliance for Financial Inclusion Case Study Agarwal, Sumit, Shashwat Alok, Pulak Ghosh, Soumya Ghosh, Tomasz Piskorski, and Amit Seru (2017) ‘Banking the Unbanked: What Do 255 Million New Bank Accounts Reveal about Financial Access?’ Working Paper Allen, Franklin, Elena Carletti, Robert Cull, Jun Qian, Lemma W Senbet, and Patricio Valenzuela (2021) ‘Improving Access to Banking: Evidence from Kenya.’ Review of Finance Angelucci, Manuela, Dean Karlan, and Jonathan Zinman (2015) ‘Microcredit Impacts: Evidence from a Randomized Microcredit Program Placement Experiment by Compartamos Banco.’ American Economic Journal: Applied Economics 7(1), 151–182 Arraiz, Irani, Miriam Bruhn, Benjamin N. Roth, and Claudia Ruiz-Ortega (2020) ‘Free Riding in Loan Approvals: Evidence from SME Lending in Peru.’ Working Paper 20-079, Harvard Business School Azevedo, Viviane, Jeanne Lafortune, Liliana Olarte, and José Tessada (2019) ‘Does Formal Credit Lead to more Financial Inclusion or Distress? Results Using a Strict Scoring Rule among Marginal Clients in Paraguay.’ Technical Note 20, IDB Invest, Washington, DC, November Banerjee, Abhijit, Dean Karlan, and Jonathan Zinman (2015) ‘Six randomized evaluations of microcredit: Introduction and further steps.’ American Economic Journal: Applied Economics 7(1), 1–21 33

Banerjee, Abhijit, Emily Breza, Arun Chandrasekhar, Esther Duflo, Matthew Jackson, and Cynthia Kinnan (2021) ‘Changes in social network structure in response to exposure to formal credit markets.’ NBER Working Paper 28365, National Bureau of Economic Research, jan Banerjee, Abhijit, Emily Breza, Esther Duflo, and Cynthia Kinnan (2019) ‘Can microfinance unlock a poverty trap for some entrepreneurs?’ NBER Working Paper 26346, National Bureau of Economic Research, oct Beck, Thorsten, Asli Demirguc-Kunt, and Maria Soledad Martinez Peria (2007) ‘Reaching out: Access to and use of banking services across countries.’ Journal of Financial Economics 85(1), 234–266 Beck,Thorsten,RossLevine,andNormanLoayza(2000)‘Financeandthesourcesofgrowth.’ Journal of Financial Economics 58(1), 261–300 Besley, T. J., K. B. Burchardi, and M. Ghatak (2012) ‘Incentives and the De Soto Effect.’ The Quarterly Journal of Economics 127(1), 237–282 Bonfim, Diana, Gil Nogueira, and Steven Ongena (2021) ‘“Sorry, We’re Closed” Bank Branch Closures, Loan Pricing, and Information Asymmetries.’ Review of Finance Boot, Arnoud W. A., and Anjan V. Thakor (1994) ‘Moral Hazard and Secured Lending in an Infinite Repeated Credit Market Game.’ International Economic Review 35(4), 899–920 Bos, Jaap W.B., Ralph de Haas, and Matteo Millone (2021) ‘Information sharing in a competitive microcredit market.’ Journal of Money, Credit and Banking Breza, Emily, and Cynthia Kinnan (2021) ‘Measuring the equilibrium impacts of credit: Evidence from the Indian microfinance crisis.’ Quarterly Journal of Economics Brown, Martin, Benjamin Guin, and Karolin Kirschenmann (2016) ‘Microfinance Banks and Financial Inclusion.’ Review of Finance 20(3), 907–946 Brown, Martin, Tullio Jappelli, and Marco Pagano (2009) ‘Information sharing and credit: Firm-level evidence from transition countries.’ Journal of Financial Intermediation 18(2), 151–172 Bruhn, Miriam, and Inessa Love (2014) ‘The Real Impact of Improved Access to Finance: Evidence from Mexico.’ The Journal of Finance 69(3), 1347–1376 Burgess, Robin, and Rohini Pande (2005) ‘Do Rural Banks Matter? Evidence from the Indian Social Banking Experiment.’ American Economic Review 95(3), 780–795 Castellanos, Sara G., Diego Jimenez-Hernandez, Aprajit Mahajan, and Enrique Seira (2020) ‘Expanding Financial Access Via Credit Cards: Evidence from Mexico.’ Working Paper Célerier, Claire, and Adrien Matray (2019) ‘Bank-branch supply, financial inclusion, and wealth accumulation.’ The Review of Financial Studies 32(12), 4767–4809 34

CGAP(2011)‘CreditReportingattheBaseofthePyramid-KeyIssuesandSuccessFactors. FORUM No. 1. International Finance Corporation and CGAP, Washington, DC.’ Cox, David R. (1972) ‘Regression models and life-tables.’ Journal of the Royal Statistical Society, Series B 34(2), 187–220 Cull, Robert J, Asli Demirgüç-Kunt, and Jonathan Morduch (2009) ‘Microfinance Meets the Market.’ Journal of Economic Perspectives 23(1), 167–192 De Chaisemartin, Clément, and Xavier d’Haultfoeuille (2020) ‘Two-way fixed effects estimators with heterogeneous treatment effects.’ American Economic Review 110(9), 2964–96 FDIC (2020) ‘2019 FDIC Survey of Household Use of Banking and Financial Services’ Feigenberg, Benjamin, Erica M. Field, and Rohini Pande (2013) ‘The economic returns to social interaction: Experimental evidence from microfinance.’ Review of Economic Studies 80(4), 1459–1483 FinScope (2012) ‘Financial Inclusion in Rwanda 2008-2012’ (2016) ‘Financial Inclusion in Rwanda 2016’ Fisman, Raymond, Daniel Paravisini, and Vikrant Vig (2017) ‘Cultural proximity and loan outcomes.’ American Economic Review 107(2), 457–92 Giné, Xavier (2011) ‘Access to capital in rural Thailand: An estimated model of formal vs. informal credit.’ Journal of Development Economics 96(1), 16–29 Goodman-Bacon, Andrew (2021) ‘Difference-in-differences with variation in treatment timing.’ Journal of Econometrics Guiso, Luigi, Paola Sapienza, and Luigi Zingales (2004) ‘The role of social capital in financial development.’ The American Economic Review 94(3), 526–556 Hakizimfura, Emmanuel, Douglas Randall, and Bilal Zia (2020) ‘Decentralized delivery of financial education: Experimental evidence from Rwanda.’ Journal of Development Economics 144, 102439 Henderson, J. Vernon, Adam Storeygard, and David N. Weil (2012) ‘Measuring economic growth from outer space.’ American Economic Review 102(2), 994–1028 Hoff, Karla, and Joseph E. Stiglitz (1990) ‘Introduction: Imperfect Information and Rural Credit Markets—Puzzles and Policy Perspectives.’ The World Bank Economic Review 4(3), 235–250 Houston, Joel F, Chen Lin, Ping Lin, and Yue Ma (2010) ‘Creditor rights, information sharing, and bank risk taking.’ Journal of Financial Economics 96(3), 485–512 IMF (2017a) ‘Rwanda: Selected issues.’ IMF Country Report No. 17/214 35

(2017b) ‘Staff Report for the 2017 Article IV Consultation, Seventh Review under the Policy Support Instrument, and Second Review under the Standby Credit Facility.’ IMF Country Report No. 17/214 Ioannidou, Vasso, and Steven Ongena (2010) ‘“Time for a Change”: Loan Conditions and Bank Behavior when Firms Switch Banks.’ The Journal of Finance 65(5), 1847–1877 Jappelli, Tullio, and Marco Pagano (2002) ‘Information sharing, lending and defaults: Crosscountry evidence.’ Journal of Banking & Finance 26(10), 2017–2045 Jayaratne, Jith, and Philip E Strahan (1996) ‘The finance-growth nexus: Evidence from bank branch deregulation.’ The Quarterly Journal of Economics 111(3), 639–670 Jean, N., M. Burke, M. Xie, W. M. Davis, D. B. Lobell, and S. Ermon (2016) ‘Combining satellite imagery and machine learning to predict poverty.’ Science 353(6301), 790–794 Kaboski, Joseph P, and Robert M Townsend (2012) ‘The Impact of Credit on Village Economies.’ American Economic Journal: Applied Economics 4(2), 98–133 Karlan, Dean, and Jonathan Zinman (2011) ‘Microcredit in theory and practice: Using randomized credit scoring for impact evaluation.’ Science 332(6035), 1278–1284 Karlan, Dean, and Jonathan Zinman (2019) ‘Long-Run Price Elasticities of Demand for Credit: Evidence from a Countrywide Field Experiment in Mexico.’ The Review of Economic Studies 86(4), 1704–1746 King, Robert G, and Ross Levine (1993) ‘Finance and growth: Schumpeter might be right.’ The Quarterly Journal of Economics 108(3), 717–737 Liberti, Jose Maria, Amit Seru, and Vikrant Vig (2016) ‘Information, credit, and organization.’ Working Paper, London Business School Mankiw, N.Gregory(1986)‘TheAllocationofCreditandFinancialCollapse.’The Quarterly Journal of Economics 101(3), 455 Meager, Rachael (2019) ‘Understanding the Average Impact of Microcredit Expansions: A Bayesian Hierarchical Analysis of Seven Randomized Experiments.’ American Economic Journal: Applied Economics 11(1), 57–91 Menkhoff, Lukas, and Ornsiri Rungruxsirivorn (2011) ‘Do Village Funds Improve Access to Finance? Evidence from Thailand.’ World Development 39(1), 110–122 MINECOFIN (2013) ‘National Microfinance Policy Implementation Strategy 2013–2017, Ministry of Finance and Economic Planning (MINECOFIN), Government of Rwanda’ Mitra, Sandip, Dilip Mookherjee, Maximo Torero, and Sujata Visaria (2018) ‘Asymmetric Information and Middleman Margins: An Experiment with Indian Potato Farmers.’ The Review of Economics and Statistics 100(1), 1–13 36

Mookherjee, D., and A. Motta (2016) ‘A theory of interactions between MFIs and informal lenders.’ Journal of Development Economics 121, 191–200 Morduch, Jonathan (1999) ‘The Microfinance Promise.’ Journal of Economic Literature 37(4), 1569–1614 NBR (2016) ‘Annual Financial Stability Report.’ Technical Report, National Bank of Rwanda, Kigali Nguyen, Hoai-Luu Q (2019) ‘Are Credit Markets Still Local? Evidence from Bank Branch Closings.’ American Economic Journal: Applied Economics 11(1), 1–32 Padilla, A Jorge, and Marco Pagano (1997) ‘Endogenous communication among lenders and entrepreneurial incentives.’ The Review of Financial Studies 10(1), 205–236 Pagano, Marco, and Tullio Jappelli (1993) ‘Information sharing in credit markets.’ The Journal of Finance 48(5), 1693–1718 Rajan, Raghuram G (1992) ‘Insiders and outsiders: The choice between informed and arm’slength debt.’ Journal of Finance 47(4), 1367–400 Randall, Douglas (2014) ‘Financial Inclusion Up Close in Rwanda.’ World Bank Private Sector Development Blog Schicks, Jessica (2014) ‘Over-indebtedness in microfinance – an empirical analysis of related factors on the borrower level.’ World Development 54, 301–324 Shapiro, D.A. (2015) ‘Microfinance and dynamic incentives.’ Journal of Development Economics 115(C), 73–84 Sharpe, Steven A (1990) ‘Asymmetric information, bank lending, and implicit contracts: A stylized model of customer relationships.’ Journal of Finance 45(4), 1069–87 Teal, Francis (2011) ‘The price of labour and understanding the causes of poverty.’ Labour Economics 18, S7–S15 Tedeschi, GwendolynAlexander(2006)‘Heretoday, gonetomorrow: Candynamicincentives make microfinance more flexible?’ Journal of Development Economics 80(1), 84 – 105 The Economist (2019) ‘Special report: Banking. The banking revolution is great for customers. May 2019’ von Thadden, Ernst-Ludwig (2004) ‘Asymmetric information, bank lending and implicit contracts: the winner’s curse.’ Finance Research Letters 1(1), 11–23 Zinman, Jonathan (2014) ‘Consumer credit: Too much or too little (or just right)?’ The Journal of Legal Studies 43(S2), S209–S237 37

Table 1: Descriptive Statistics The table presents summary statistics for the main variables in our sample. The sample period is January 2008–December 2016 and includes 317,281 individuals in 337 municipalities who borrow from commercial banks, U-SACCOs, and other MFIs, or who do not receive any loans during this period. The access to credit variables (Panel A) are dummies equal to one if the borrower in a given month has an outstanding loan from any financial institution, a U-SACCO, a commercial bank, or other MFIs. In Panel B, loan exposure and principal amounts are expressed in millions of Rwandan Francs (RWF). Non-performing loan is a dummy equal to one if the loan is in arrears for more than 90 days. In Panel C, Female is equal to one for females and zero for males; Young takes value one for individuals below 30 years of age and zero otherwise; Single is equal to one for single individuals and zero otherwise; and Government Employee is equal to one for government employees and zero for any other occupation and the unemployed. Panel D lists the municipality characteristics measured in the pre-program period. Bank Presence is the number of bank branches per capita, Economic Growth is the change in nightlights between 2008 and 2010, Economic Development is nightlights in 2010, Poverty is the poverty headcount ratio (percent population below the poverty line), Adult Population is the number of adults (in thousands), Conflicts is the number of violent events between 1997 and 2010, Distance to Capital is the travel time (in hours) by car of the shortest route from Kigali to the municipality centroid, and Rural is a dummy that identifies rural municipalities. Sources: Rwandan Credit Reference Bureau, National Institute of Statistics of Rwanda, National Oceanic and Atmospheric Administration, OpenStreetMaps, and Armed Conflict Location & Event Data. N Mean Median SD (1) (2) (3) (4) A. Access to Credit LoanfromAnyInstitution 34,266,348 0.133 0.000 0.340 LoanfromU-SACCO 34,266,348 0.034 0.000 0.181 LoanfromBank 34,266,348 0.063 0.000 0.243 LoanfromotherMFI 34,266,348 0.045 0.000 0.207 B. Loan Characteristics LoanExposure(RWFmn) 4,936,650 2.913 0.558 18.95 LoanPrincipalAmount(RWAmn) 4,936,650 4.132 1.000 25.85 InterestRate(%) 4,102,615 19.52 18.00 12.88 Maturity(months) 4,936,650 31.01 24.00 27.76 Non-PerformingLoan 4,936,650 0.122 0.000 0.327 C. Borrower characteristics Female 317,281 0.414 0.000 0.493 Single 317,281 0.091 0.000 0.287 Young 317,281 0.414 0.000 0.493 GovernmentEmployee 317,281 0.089 0.000 0.285 D. Municipality characteristics BankPresence 337 0.046 0.000 0.074 EconomicGrowth 337 0.993 0.000 3.547 EconomicDevelopment 337 2.570 0.000 9.910 Poverty 337 0.420 0.444 0.125 AdultPopulation 337 13.85 12.92 5.075 Conflicts 337 0.852 0.000 3.436 DistancetoCapital 337 1.716 1.589 0.899 Rural 337 0.700 0.000 0.459 38

snoitidnoC lacoL dna margorP OCCAS-U eht fo gnimiT ehT :2 elbaT eht si elbairav tnedneped eht erehw )2791 ,xoC( ledom drazah lanoitroporp a fo setamitse tneicffieoc eht stneserp elbat ehT eht roF .)”tneve“ eht( ytilapicinum nevig a ni OCCAS-U a yb dednetxe naol tsrfi eht fo ecnerrucco eht litnu )shtnom ni( emit ”tneve“ eht sekat ledom eht ,doirep elpmas eht gnirud snaol dnetxe ton od taht elpmas ruo ni seitilapicinum )103−733 =( 63 .doirep margorp-erp eht ni derusaem era dna level ytilapicinum eht ta denfied era selbairav yrotanalpxe ehT .gnirrucco ton sa nevig a ni gnitarepo OCCAS-U hcae rof ,erusaem selbairav OCCAS-U ehT .elbairav hcae fo noitinfied eht rof 1 elbaT eeS stessa-ot-latipac eht ,)FWR snoillim ni( stisoped latot ,)FWR snoillim ni( stessa latot ,srebmem fo rebmun eht ,ytilapicinum *** .sesehtnerap ni era level ytilapicinum eht ta deretsulc srorre dradnatS .0102 fo sa oitar stessa-ot-ytidiuqil eht dna ,oitar .1.0<p * ,50.0<p ** ,10.0<p )41( )31( )21( )11( )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( 640.0- 520.0htworGcimonocE )360.0( )120.0( 300.0 800.0tnempoleveDcimonocE )520.0( )800.0( 029.0- 620.0 ytrevoP )326.0( )105.0( 477.0- 493.1ecneserPknaB )091.1( )339.0( 240.0 480.0 srebmeMOCCAS-U )960.0( )250.0( 530.0 300.0 stessAOCCAS-U )030.0( )400.0( 340.0- 200.0 stisopeDOCCAS-U )430.0( )500.0( 050.1- 582.0oitaRlatipaCOCCAS-U )556.0( )263.0( 893.0- 382.0oitaRytidiuqiLOCCAS-U )462.0( )813.0( 810.0 300.0noitalupoPtludA )510.0( )110.0( 320.0- 520.0stciflnoC )120.0( )610.0( 310.0- 340.0 latipaCotecnatsiD )170.0( )460.0( 321.0 171.0 laruR )761.0( )731.0( 733 733 733 733 733 733 733 733 733 733 733 733 733 733 snoitavresbOfo .oN 39

Table 3: Effect of the U-SACCO Program on Access to Credit The table presents coefficient estimates of model 1 where the dependent variable is a dummy equal to one for individuals who, at time t, have an outstanding loan with any financial institution (columns 1-3) or only U-SACCOs (column 4), commercial banks (column 5), or other MFIs (column 6). Post U-SACCO is a dummy equal to one after a U-SACCO starts its lending activity in a given municipality and month, and zero otherwise. Borrower controls include a set of dummies for gender, marital status, age, and employment status. See Table 1 for the definition of each variable. In columns 3-6 the Post U-SACCO dummy is split in a set of dummies for six-month intervals around the date when theU-SACCO starts its lending activity in the municipality. The dummy variables for the pre-program period are not shown for reason of space, but they are plotted in Figures 2 (columns 3–5) and C3 (column 6). The data are at the borrower-municipality-month level. The sample period is 2008:M1 to 2016:M12. Standard errors clustered at the municipality level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Dummy=1ifindividualhasaLoanin AnyInstitution U-SACCO Bank OtherMFI (1) (2) (3) (4) (5) (6) PostU-SACCO 0.032*** 0.021*** (0.006) (0.005) PostU-SACCO[t,t+6] 0.014*** 0.013*** 0.002 0.002 (0.003) (0.003) (0.001) (0.002) PostU-SACCO[t+7,t+12] 0.041*** 0.042*** 0.004 0.003 (0.007) (0.006) (0.003) (0.003) PostU-SACCO[t+13,t+18] 0.068*** 0.069*** 0.008* 0.004 (0.010) (0.008) (0.004) (0.004) PostU-SACCO[t+19,t+24] 0.083*** 0.080*** 0.014** 0.004 (0.012) (0.009) (0.006) (0.006) PostU-SACCO[t+25,t+30] 0.091*** 0.084*** 0.018** 0.005 (0.014) (0.010) (0.008) (0.007) PostU-SACCO[t+31,t+36] 0.102*** 0.088*** 0.024** 0.008 (0.016) (0.010) (0.010) (0.008) PostU-SACCO[>t+36] 0.116*** 0.095*** 0.031** 0.009 (0.019) (0.010) (0.012) (0.010) MunicipalityFE Y Y Y Y Y Y Time(Year:month)FE Y Y Y Y Y Y BorrowerControls Y Y Y Y Y Y MunicipalityTimeTrends N Y Y Y Y Y No. ofObservations 34,266,348 34,266,348 34,266,348 34,266,348 34,266,348 34,266,348 No. ofMunicipalities 337 337 337 337 337 337 No. ofIndividuals 317,281 317,281 317,281 317,281 317,281 317,281 AdjustedR-squared 0.193 0.200 0.201 0.121 0.078 0.156 40

Table 4: Access to Credit—Cross Sectional Heterogeneity The table presents coefficient estimates of model 1 where the dependent variable is a dummy equal to one for individuals who, at time t, have an outstanding loan with a U-SACCOs. Post U-SACCO is a dummy equal to one after a U-SACCO starts its lending activities in a given municipality and month, and zero otherwise. The Post U-SACCO dummy is interacted with a set of dummy variables which identify municipalities: (i) with low bank presence; (ii) with low growth; (iii) with low economic development; iv) with high poverty; (v) distant from the capital Kigali; and vi) in a rural area. These dummy variables are defined from the continuous ones (as defined in Table 1) by splitting the continuous variable around the 75th percentile of the sample distribution. Specifically, Low Bank Presence is equal to one if the number of bank branches per capita is below the 75th percentile, and zero otherwise; Low Growth is equal to one if the change in nightlights between 2008 and 2010 is below the 75th percentile, and zero otherwise; Low Nightlights is equal to one if nightlights in 2010 is below the 75th percentile, and zero otherwise; High Poverty is equal to one if the inverse of the poverty headcount ratio in the pre-period is below the 75th percentile, and zero otherwise; Long Distance to Capital is equal to one if the inverse of the travel time by car from Kigali to a municipality centroid in the pre-period is below the 75th percentile, and zero otherwise. Borrower controls include a set of dummies for gender, marital status, age, and employment status. See Table 1 for the definition of each variable. The data are at the borrower-municipality-month level. The sample period is 2008:M1 to 2016:M12. Standard errors clustered at the municipality level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Dummy=1ifindividualhasaLoaninaU-SACCO (1) (2) (3) (4) (5) (6) PostU-SACCOxLowBankPresence 0.024*** (0.008) PostU-SACCOxLowGrowth 0.027*** (0.008) PostU-SACCOxLowDevelopment 0.026*** (0.008) PostU-SACCOxHighPoverty 0.022*** (0.008) PostU-SACCOxLongDistancetoCapital 0.018** (0.008) PostU-SACCOxRural 0.025*** (0.005) PostU-SACCO 0.005 0.002 0.003 0.005 0.009 0.010* (0.008) (0.008) (0.008) (0.008) (0.008) (0.005) MunicipalityFE Y Y Y Y Y Y Time(Year:month)FE Y Y Y Y Y Y BorrowerControls Y Y Y Y Y Y MunicipalityTimeTrends Y Y Y Y Y Y No. ofObservations 34,266,348 34,266,348 34,266,348 34,266,348 34,266,348 34,266,348 No. ofMunicipalities 337 337 337 337 337 337 No. ofIndividuals 317,281 317,281 317,281 317,281 317,281 317,281 41

enilesaB—snaoL gnihctiwS fo sisylanA :5 elbaT tnedneped eht erehw rotamitse gnihctam detcerroc-saib )1102( snebmI dna eidabA eht fo setamitse stneicffieoc stroper elbat ehT ohw sreworrob ot snaol wen fo stsisnoc puorg tnemtaert ehT .ytirutam naol ro ,etar tseretni naol ,tnuoma naol eht si elbairav -U yb snaol wen fo desirpmoc si puorg lortnoc ehT .snaol gnihctiws ,si taht—sknab laicremmoc ot sOCCAS-U morf hctiws eW .)6–4 snmuloc( sreworrob rehto ot sknab laicremmoc yb snaol wen ro )3–1 snmuloc( sreworrob rehctiws-non ot sOCCAS reworrob no desab naol gnihctiws hcae fo srobhgien-tseraen eht tceles dna htnom emas eht ni detnarg era taht snaol hctam ,yrotsih tiderc fo sraey fo .on( scitsiretcarahc yrotsih tiderc etna xe ,)eeyolpme tnemnrevog ,gnuoy ,elgnis ,elamef( scitsiretcarahc ,naol wen eht erofeb htnom eht ni reworrob a fo tnuoma naol gnidnatstuo latot dna ,snaol fo .on ,spihsnoitaler gnidnel fo .on dna egagtrom( scitsiretcarahc naol ,)naol wen eht erofeb emit yna ro sraey owt ro eno nihtiw rednel yna htiw sLPN sa llew sa si tnuoma naol ehT .margorp eht erofeb derusaem sediser reworrob eht erehw ytilapicinum eht fo scitsiretcarahc dna ,)laretalloc elbairav deliated ehT .shtnom ni ytirutam naol eht dna ,stniop egatnecrep ni etar tseretni naol eht ,FWR snoillim ni desserpxe .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni srorre dradnatS .1 elbaT ni elbaliava era snoitinfied :puorGlortnoC :puorGlortnoC sreworrobrehtootsknabybsnaolweN srehctiws-nonotsOCCAS-UybsnaolweN naoL tseretnI naoL naoL tseretnI naoL ytirutaM etaR tnuomA ytirutaM etaR tnuomA )6( )5( )4( )3( )2( )1( ***438.2 753.0- 600.0 ***456.9 ***184.6- ***218.0 snaoLrehtO-naoLgnihctiwS )394.0( )062.0( )011.0( )454.0( )815.0( )580.0( :selbairaVgnihctaM Y Y Y Y Y Y ecnaussInaoLfohtnoM:raeY Y Y Y sOCCAS-U Y Y Y sknaBlaicremmoC Y Y Y Y Y Y elameF Y Y Y Y Y Y elgniS Y Y Y Y Y Y gnuoY Y Y Y Y Y Y eeyolpmEtnemnrevoG Y Y Y Y Y Y naoLweNerofeByrotsiHtiderCfosraeYfo .oN Y Y Y Y Y Y naoLweNerofeBspihsnoitaleRgnidneLfo .oN Y Y Y Y Y Y naoLweNerofeBsnaoLfo .oN Y Y Y Y Y Y naoLweNerofeBtnuomAnaoLgnidnatstuOlatoT Y Y Y Y Y Y naoLweNerofeBraey1nihtiwredneLynahtiwLPN Y Y Y Y Y Y naoLweNerofeBsraey2nihtiwredneLynahtiwLPN Y Y Y Y Y Y naoLweNerofeBredneLynahtiwLPN Y Y Y Y Y Y egagtroM Y Y Y Y Y Y laretalloC Y Y Y Y Y Y ytilapicinuMniecneserPknaB Y Y Y Y Y Y ytilapicinuMnihtworGcimonocE Y Y Y Y Y Y ytilapicinuMnitnempoleveDcimonocE Y Y Y Y Y Y ytilapicinuMniytrevoP Y Y Y Y Y Y ytilapicinuMninoitalupoPtludA Y Y Y Y Y Y ytilapicinuMnistciflnoC Y Y Y Y Y Y ytilapicinuMmorflatipaCotecnatsiD Y Y Y Y Y Y ytilapicinuMlaruR 019,2 019,2 019,2 019,2 019,2 019,2 )detaerT(srehctiwSfo .oN 479,032 479,032 479,032 215,67 215,67 215,67 sreworroBdetaertnUfo .oN 42

ecnamrofreP naoL tsop xe—snaoL gnihctiwS fo sisylanA :6 elbaT rotamitse gnihctam robhgien-tseraen detcerroc-saib )1102( snebmI dna eidabA eht fo setamitse stneicffieoc stroper elbat ehT litnu ro sraey owt ro eno nihtiw gnimrofrep-non semoceb naol a fi eno ot lauqe ymmud a si elbairav tnedneped eht erehw laicremmoc ot sOCCAS-U morf hctiws ohw sreworrob ot snaol fo stsisnoc puorg tnemtaert ehT .esiwrehto orez dna ,ytirutam sreworrob rehctiws-non ot sOCCAS-U lla yb snaol wen :fo desirpmoc si puorg lortnoc ehT .snaol gnihctiws ,si taht—sknab snaol hctam eW .)6–4 snmuloc( ot dehctiws reworrob OCCAS-U eht knab laicremmoc emas eht yb snaol wen ro ,)3–1 snmuloc( scitsiretcarahc reworrob no desab naol gnihctiws hcae fo srobhgien-tseraen eht tceles dna htnom emas eht ni detnarg era taht gnidnel fo .on ,yrotsih tiderc fo sraey fo .on( scitsiretcarahc yrotsih tiderc etna xe ,)eeyolpme tnemnrevog ,gnuoy ,elgnis ,elamef( sLPN sa llew sa ,naol wen eht erofeb htnom eht ni reworrob a fo tnuoma naol gnidnatstuo latot dna ,snaol fo .on ,spihsnoitaler ,tnuoma ,laretalloc ,egagtrom( scitsiretcarahc naol ,)naol wen eht erofeb emit yna ro sraey owt ro eno nihtiw rednel yna htiw ehT .margorp eht erofeb derusaem sediser reworrob eht erehw ytilapicinum eht fo scitsiretcarahc dna ,)ytirutam ,etar tseretni .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni srorre dradnatS .1 elbaT ni elbaliava era snoitinfied elbairav deliated :puorGlortnoC :puorGlortnoC sreworrobrehtootsknabybsnaolweN srehctiws-nonotsOCCAS-UybsnaolweN ytirutamlitnu sraey2nihtiw raey1nihtiw ytirutamlitnu sraey2nihtiw raey1nihtiw )6( )5( )4( )3( )2( )1( 310.0 310.0 000.0 ***040.0- ***440.0- ***930.0snaoLrehtO-naoLgnihctiwS )900.0( )900.0( )900.0( )010.0( )010.0( )900.0( :selbairaVgnihctaM Y Y Y Y Y Y ecnaussInaoLfohtnoM:raeY Y Y Y sOCCAS-U Y Y Y sknaBlaicremmoC Y Y Y Y Y Y elameF Y Y Y Y Y Y elgniS Y Y Y Y Y Y gnuoY Y Y Y Y Y Y eeyolpmEtnemnrevoG Y Y Y Y Y Y gnihctiwSerofebyrotsiHtiderCfosraeYfo .oN Y Y Y Y Y Y gnihctiwSerofebspihsnoitaleRgnidneLfo .oN Y Y Y Y Y Y gnihctiwSerofebsnaoLfo .oN Y Y Y Y Y Y gnihctiwSerofebtnuomAnaoLgnidnatstuOlatoT Y Y Y Y Y Y gnihctiwSerofebraey1nihtiwredneLynahtiwLPN Y Y Y Y Y Y gnihctiwSerofebsraey2nihtiwredneLynahtiwLPN Y Y Y Y Y Y gnihctiwSerofebredneLynahtiwLPN Y Y Y Y Y Y egagtroM Y Y Y Y Y Y laretalloC Y Y Y Y Y Y tnuomAnaoL Y Y Y Y Y Y etaRtseretnInaoL Y Y Y Y Y Y ytirutaMnaoL Y Y Y Y Y Y ytilapicinuMniecneserPknaB Y Y Y Y Y Y ytilapicinuMnihtworGcimonocE Y Y Y Y Y Y ytilapicinuMnitnempoleveDcimonocE Y Y Y Y Y Y ytilapicinuMniytrevoP Y Y Y Y Y Y ytilapicinuMninoitalupoPtludA Y Y Y Y Y Y ytilapicinuMnistciflnoC Y Y Y Y Y Y ytilapicinuMmorflatipaCotecnatsiD Y Y Y Y Y Y ytilapicinuMlaruR 019,2 019,2 019,2 019,2 019,2 019,2 )detaerT(srehctiwSfo .oN 479,032 479,032 479,032 215,67 215,67 215,67 sreworroBdetaertnUfo .oN 43

sreworroB knaB emit-tsriF htiw nosirapmoC—snaoL gnihctiwS fo sisylanA :7 elbaT rotamitse gnihctam robhgien-tseraen detcerroc-saib )1102( snebmI dna eidabA eht fo setamitse stneicffieoc stroper elbat ehT fo stsisnoc puorg tnemtaert ehT .ytirutam naol dna ,etar tseretni naol ,ezis naol ,ylevitanretla ,si elbairav tnedneped eht erehw desirpmoc si puorg lortnoc ehT .snaol gnihctiws ,si taht—sknab laicremmoc ot sOCCAS-U morf hctiws ohw sreworrob ot snaol snmuloc( ot dehctiws reworrob OCCAS-U eht knab laicremmoc emas eht yb dednetxe sreworrob emit-tsrfi ot snaol wen :fo dehctiws reworrob OCCAS-U eht knab laicremmoc emas eht yb dednetxe srehctiws knab-ot-IFM-rehto ot snaol wen dna ;)3-1 naol gnihctiws hcae fo srobhgien-tseraen eht tceles dna htnom emas eht ni detnarg era taht snaol hctam eW .)6-4 snmuloc( ot fo .on( scitsiretcarahc yrotsih tiderc etna xe ,)eeyolpme tnemnrevog ,gnuoy ,elgnis ,elamef( scitsiretcarahc reworrob no desab eht ni reworrob a fo tnuoma naol gnidnatstuo latot dna ,snaol fo .on ,spihsnoitaler gnidnel fo .on ,yrotsih tiderc fo sraey naol ,)naol wen eht erofeb emit yna ro sraey owt ro eno nihtiw rednel yna htiw sLPN sa llew sa ,naol wen eht erofeb htnom erofeb derusaem sediser reworrob eht erehw ytilapicinum eht fo scitsiretcarahc dna ,)laretalloc dna egagtrom( scitsiretcarahc ytirutam naol eht dna ,stniop egatnecrep ni etar tseretni naol eht ,FWR snoillim ni desserpxe si tnuoma naol ehT .margorp eht * ,50.0<p ** ,10.0<p *** .sesehtnerap ni srorre dradnatS .1 elbaT ni elbaliava era snoitinfied elbairav deliated ehT .shtnom ni .1.0<p :puorGlortnoC :puorGlortnoC srehctiwsknab-ot-IFMotsknabybsnaolweN sreworrobemit-tsrfiotsknabybsnaolweN ytirutaM etaRtseretnI tnuomAnaoL ytirutaM etaRtseretnI tnuomAnaoL )6( )5( )4( )3( )2( )1( ***572.4 ***554.1- 801.0 ***894.2 ***096.0- 220.0snaoLrehtO-naoLgnihctiwS )266.0( )992.0( )461.0( )334.0( )722.0( )741.0( :selbairaVgnihctaM Y Y Y Y Y Y ecnaussInaoLfohtnoM:raeY Y Y Y Y Y Y sknaBlaicremmoC Y Y Y Y Y Y elameF Y Y Y Y Y Y elgniS Y Y Y Y Y Y gnuoY Y Y Y Y Y Y eeyolpmEtnemnrevoG Y Y Y gnihctiwSerofebyrotsiHtiderCfosraeYfo .oN Y Y Y gnihctiwSerofebspihsnoitaleRgnidneLfo .oN Y Y Y gnihctiwSerofebsnaoLfo .oN Y Y Y gnihctiwSerofebtnuomAnaoLgnidnatstuOlatoT Y Y Y gnihctiwSerofebraey1nihtiwredneLynahtiwLPN Y Y Y gnihctiwSerofebsraey2nihtiwredneLynahtiwLPN Y Y Y gnihctiwSerofebredneLynahtiwLPN Y Y Y Y Y Y egagtroM Y Y Y Y Y Y laretalloC Y Y Y Y Y Y ytilapicinuMniecneserPknaB Y Y Y Y Y Y ytilapicinuMnihtworGcimonocE Y Y Y Y Y Y ytilapicinuMnitnempoleveDcimonocE Y Y Y Y Y Y ytilapicinuMniytrevoP Y Y Y Y Y Y ytilapicinuMninoitalupoPtludA Y Y Y Y Y Y ytilapicinuMnistciflnoC Y Y Y Y Y Y ytilapicinuMmorflatipaCotecnatsiD Y Y Y Y Y Y ytilapicinuMlaruR 918,2 918,2 918,2 898,2 898,2 898,2 )detaerT(srehctiwSfo .oN 527,5 527,5 527,5 957,38 957,38 957,38 sreworroBdetaertnUfo .oN 44

Table 8: Real Effects of the U-SACCO Program on Household Poverty The table presents coefficient estimates for the effects of the U-SACCO program on household poverty. The dependent variables are dummies for households in poverty (columns 1–3) and households in extreme poverty (columns 4–6). The data come from the 2010–11 and 2016–17 Integrated Household Living Conditions Surveys (EICV3 and EICV5), obtained from the data portal of the National Institute of Statistics Rwanda. There are nationally representative household surveys conducted for purposes of poverty monitoring. Households are identified as being in poverty (extreme poverty) status if the consumption level is below the $1.09 per day international poverty line (0.65x of the line). The poverty lines are 159,375 RWF and 105,064 RWF per adult equivalent per year (at January 2014 prices). The dummy variable Post takes value one for the (post-program) 2016–17 survey, and zero for the (pre-program) 2010–11 survey. The bottom two rows report p-values from one-sided tests of the null hypothesis: (i) the coefficient on the Post dummy in low bank presence municipalities is higher than that in high bank presence municipalities and (ii) the coefficient of Post×Low Bank Presence in high switching municipalities is higher than in low switching municipalities. Low bank presence municipalities are those below the 75th percentile of the distribution of the number of bank branches per capita before the program. High switching municipalities are above the 25th percentile of the distribution of share of switchers over 2012–16 relative to the total number of U-SACCO borrowers in 2012. Household controls include size and a rural dummy. Regressions are weighted by household survey sampling weight. Standard errors clustered at the district level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. (1) (2) (3) (4) (5) (6) Householdinpoverty Householdinextremepoverty Post -0.0563*** -0.0407*** (0.017) (0.014) Post×LowBankPresence -0.0634*** -0.0462*** (0.019) (0.016) Post×LowBankPresence× -0.0559*** -0.0363** HighSwitching (0.018) (0.015) Post×LowBankPresence× -0.0927 -0.0846 LowSwitching (0.056) (0.050) Post×HighBankPresence -0.0242 -0.0261 -0.0160 -0.0171 (0.041) (0.042) (0.034) (0.034) No. ofObservations 43,307 43,307 43,307 43,307 43,307 43,307 DistrictFE Y Y Y Y Y Y Householdcontrols Y Y Y Y Y Y AdjustedR-squared 0.117 0.117 0.118 0.070 0.070 0.072 Meanofdependentvariable 0.364 0.364 0.364 0.149 0.149 0.149 45

Internet Appendix A U-SACCO Program and Access to Credit: Evidence from Household Survey Data Here we test whether the microcredit expansion program increased financial access using survey data from the 2012 and 2016 rounds of the FinScope surveys run by Access to Finance Rwanda as part of a cross-country project developed by FinMark Trust. The purpose of the FinScope surveys is to describe levels of access to and take-up of financial products and services in the formal and informal sectors. Summary statistics are shown in Table A1 and illustrate the prevalence of informal lending and the increase in formal lending, especially among U-SACCOs.26 We employ a slightly different identification strategy than in the baseline analysis for two reasons: (i) We only have two cross-sections of data; and (ii) borrower location is available at the district (not municipality) level. Since we cannot exploit the staggered rollout of the program across municipalities as we did in the baseline analysis, we take the 2012 survey data as the pre-program period and the 2016 survey data as the post-program period and compare changes in access to credit before and after the program across districts.27 We estimate the following specification, which controls for observable borrower characteristics, as in the loan-level analysis, and absorbs unobserved local heterogeneity with district fixed effects: Pr(Access) = βPost +δ(cid:48)X +α +ε , (A-I) idt t i d idt where the dependent variable is the probability that individual i in district d has a bank loan in year t (where t = 2012 or t = 2016), and α are district fixed effects. X are individuald i level characteristics controlling for gender, age (with a dummy equal to one for 35-year old individuals or younger), marital status, food security status (with a dummy equal to one for individuals who had to skip a meal many times in the 3 months before the survey) and lack of formal education. The β coefficient of the key explanatory variable Post is then t 26Individuals with access to formal financial services are defined as those who have or use products or services from financial institutions that are regulated through an Act of Law. Therefore, formal lending includes commercial banks, U-SACCOs and other MFIs. Informal loans include moneylenders, community and savings groups and borrowing from employers and farmers associations. 27Ideally, we would have liked to use the 2008 survey as baseline, but the microdata are unavailable. It is important to note, however, that using 2012 as the benchmark will likely underestimate the effects of the program given that its implementation started in 2011. 46

split between district with low and high bank presence before the rollout of the U-SACCO program Main results. The results in Table A2 show that the likelihood of individuals having a loan increased significantly between 2012 and 2016, especially in areas with low initial bank presence. In addition, the expansion of access to formal credit is mostly driven by U-SACCOs: The probability of having a loan at a U-SACCO increased between 2012 and 2016 by 120 percent, while the same probability increased by 40 percent at commercial banks. Notably, the point estimate for the effect of the microcredit expansion program on the likelihood of having a loan at a U-SACCO (2.4 percentage points, column 9) is similar to that in the loan-level analysis (2.1 percentage points, see Table 3, column 2). Formal and informal lending. A key advantage of using survey data from Finscope is that we can examine the interaction between formal and informal lending. On the one hand, one would expect crowding-out effects on informal lending because MFIs can provide an outside option to low-income borrowers that should effectively reduce the level of exploitation by informal lenders (Besley et al., 2012; Mookherjee and Motta, 2016). A recent analysis by Banerjee et al. (2021), for instance, finds that the introduction of microfinance crowds out social network relationships and informal lending. On the other hand, if there is segmentation across loan markets (Hoff and Stiglitz, 1990), or if loans from informal lenders are used for different purposes than are microfinance loans, one can expect substitution to be limited. In the context of the Village Funds (VF) microcredit program in rural Thailand, Menkhoff and Rungruxsirivorn (2011) show that households that borrow from informal lenders tend to be poorer in terms of assets and less educated than households taking VF credit (suggesting segmentation), and that informal credit is more often used for consumption than for productive purposes. Our results show that the increase in access to credit is common among formal and informal lenders, supporting the view that the U-SACCO program did not have crowdingout effects on informal lending. In fact, our results suggest crowding-in effects of access to microcredit for informal loans that echo those in several previous studies. Angelucci et al. (2015) and Karlan and Zinman (2019), for instance, study the case of a microlender in Mexico and find that households prefer informal loan sources over formal institutions because they are perceived as faster easier to access and because they are not reported to the credit bureau in case of default. Similarly, Karlan and Zinman (2011) show that access to microcredit in the Philippines strengthens community ties and encourage informal credit arrangements by providing households with more resources and liquidity. These studies 47

suggest that microcredit may have positive effects through channels that are different from the traditional ones, and informal credit is an understudied potential channel. Table A1: Descriptive Statistics on Financial Inclusion—Household Survey Evidence The table presents population-weighted descriptive statistics (mean and standard deviation, SD) for a set of variables used in A2. The first five rows report statistic for a set of dummy variables equal to one for individuals with a loan with 1) any (formal or informal) lender, 2) only informal lenders, 3) only formal lenders, 4) commercial banks, and 5) U-SACCOs. Low bank presence is a dummy equal to one for the districts with a low number of bank branches per capita before the program and zero for the top five districts in terms of number of bank branches per capita. Post is a dummy equal to one for observations in the 2016 survey and zero for those in the 2012 survey. Woman, Young, and Married, are indicator variables equal to one for women, individuals who are less than 36 year-old, and married individuals, respectively. Food security status is a dummy for individuals who had to skip a meal many times in the three months before the survey. No formal education is a dummy identifying individuals without any formal education. The dataset consists of repeated cross-sections of borrowers in the 2012 and 2016 FinScope surveys. Finscope2012 Finscope2016 Finscope2012and2016 (n=6,150) (n=12,480) (n=18,630) Mean SD Mean SD Mean SD AnyLoan 0.256 0.437 0.499 0.500 0.323 0.468 InformalLoan 0.204 0.403 0.405 0.491 0.255 0.436 FormalLoan 0.053 0.223 0.094 0.292 0.068 0.252 BankLoan 0.028 0.164 0.037 0.189 0.031 0.174 U-SACCOLoan 0.019 0.138 0.046 0.209 0.025 0.157 LowBankPresence 0.788 0.409 0.809 0.393 0.811 0.392 Post 0.000 0.000 1.000 0.000 0.418 0.493 Woman 0.577 0.494 0.103 0.304 0.340 0.474 Young 0.463 0.499 0.527 0.499 0.497 0.500 Married 0.564 0.496 0.516 0.500 0.553 0.497 FoodSecurityStatus 0.260 0.439 0.263 0.440 0.262 0.439 NoFormalEducation 0.265 0.441 0.209 0.406 0.246 0.431 48

ecnedivE yevruS dlohesuoH—tiderC ot sseccA no margorP OCCAS-U fo tceffE :2A elbaT eno ot lauqe selbairav ymmud a si elbairav tnedneped eht hcihw ni I-A noitauqe fo setamitse tneicffieoc stneserp elbat ehT )3 ,)4-3 snmuloc( srednel lamrofni ylno )2 ,)2-1 snmuloc( rednel )lamrofni ro lamrof( yna )1 htiw naol a htiw slaudividni rof si ymmud tsoP ehT .)01-9 snmuloc( sOCCAS-U )5 dna ,)8-7 snmuloc( sknab laicremmoc )4 ,)6-5 snmuloc( srednel lamrof ylno srotacidni ymmud owt era ecneserp knab hgih dna wol eht dna yevrus 2102 eht rof orez dna yevrus 6102 eht rof eno ot lauqe reworroB .ylevitcepser ,580.0 evoba dna erofeb margorp eht erofeb atipac rep sehcnarb knab fo rebmun a htiw stcirtsid rof .1A elbaT ot eton eht ni denfied era )noitacude fo level eht dna sutats ytiruces doof ,sutats latiram ,ega ,redneg( scitsiretcarahc troper swor mottob ehT .syevrus epocSniF 6102 dna 2102 eht ni sreworrob fo snoitces-ssorc detaeper fo stsisnoc tesatad ehT smret noitcaretni eht fo stneicffieoc eht fo ytilauqe eht rof tset-t eht dna yevrus 2102 eht ni selbairav tnedneped eht fo naem eht laudividni ’syevrus gnisu dethgiew era snoissergeR .seimmud ecneserP knaB hgiH dna woL eht dna ymmud tsoP eht neewteb .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni level tcirtsid eht ta deretsulc era srorre dradnatS .sthgiew )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( naolOCCAS naoLknaB naoLlamroF naoLlamrofnI naoLynA naolagnivaH :.V.D ***420.0 *010.0 ***930.0 ***981.0 ***722.0 tsoP )400.0( )500.0( )800.0( )810.0( )020.0( ***520.0 ***210.0 ***440.0 ***202.0 ***542.0 ecneserPknaBwoL×tsoP )500.0( )400.0( )800.0( )910.0( )220.0( ***020.0 100.0- 810.0 ***731.0 ***451.0 ecneserPknaBhgiH×tsoP )700.0( )800.0( )110.0( )630.0( )630.0( ***800.0- ***800.0- 400.0- 400.0- **310.0- **310.0- **130.0- **030.0- ***440.0- ***340.0namoW )300.0( )300.0( )500.0( )500.0( )500.0( )500.0( )210.0( )210.0( )110.0( )110.0( ***610.0- ***610.0- ***410.0- ***510.0- ***030.0- ***130.0- 410.0- 410.0- ***440.0- ***540.0gnuoY )300.0( )300.0( )400.0( )400.0( )500.0( )500.0( )900.0( )900.0( )010.0( )010.0( ***420.0 ***420.0 ***920.0 ***920.0 ***950.0 ***950.0 ***980.0 ***980.0 ***741.0 ***841.0 deirraM )300.0( )300.0( )500.0( )500.0( )500.0( )500.0( )900.0( )900.0( )900.0( )010.0( ***810.0- ***810.0- ***810.0- ***810.0- ***340.0- ***340.0- 300.0 200.0 ***140.0- ***240.0sutatSytiruceSdooF )300.0( )400.0( )300.0( )300.0( )400.0( )400.0( )310.0( )310.0( )110.0( )110.0( ***020.0- ***020.0- ***620.0- ***620.0- ***150.0- ***150.0- *610.0- *610.0- ***860.0- ***860.0noitacudElamroFoN )400.0( )400.0( )400.0( )400.0( )500.0( )600.0( )800.0( )800.0( )900.0( )900.0( 036,81 036,81 036,81 036,81 036,81 036,81 036,81 036,81 036,81 036,81 snoitavresbOfo .oN seY seY seY seY seY seY seY seY seY seY EFtcirtsiD 120.0 120.0 730.0 730.0 840.0 840.0 290.0 290.0 121.0 911.0 derauqs-RdetsujdA 020.0 020.0 520.0 520.0 050.0 050.0 602.0 602.0 652.0 652.0 2102ni.V.DfonaeM 995.0 980.0 340.0 211.0 630.0 )eulav-p(tsetytilauqE 49

B Relationship Lending at U-SACCOs Here we test for the presence of lending constraints, by exploiting the notion of relationship lending. The idea is that loan terms should improve as borrowing relationships matures, especially for informationally opaque borrowers (Boot and Thakor, 1994; Bos et al., 2021). Better loan terms would also reflects dynamic repayment incentives in microfinance (Tedeschi, 2006; Shapiro, 2015). However, such returns would be lower at U-SACCOs if the regulatory environment or balance sheet size acted as a constraint. We bring this conjecture to the data by examining the patterns of repeated borrowing at U-SACCOs to test whether loan terms (loan size, interest rate, and maturity) improve with the length of the lender-borrower relationship and how these terms compare with loan terms at commercial banks. First, inthesampleofswitchers, weconductawithin-borroweranalysisbycomparingthe first loan at U-SACCOs with all the subsequent loans before switching (taking the average of loan terms when there is more than one loan). We find that a longer lending relationship is associated with a significant increase in loan size and maturity (Table B1, Panel A). Then, comparing the last loan obtained by switchers at the U-SACCO with the first loan that the same borrower obtains from a bank, we find that the premium offered by banks in terms of loan size and maturity is larger than what borrowers can obtain through repeated borrowing at U-SACCOs. In addition, the first loan at the commercial bank is 3.5 percentage points cheaper than the last at the U-SACCO (Table B1, Panel B). Second, in the full sample we relate loan terms to a measure of relationship length (defined as the number of months since the first loan was granted to a given borrower by a given financial institution), a dummy variable for loans granted by U-SACCOs, and their interaction term. This specification allows us to test whether returns to a longer lender-borrower relationship differ between U-SACCOs and banks. As reported in Table B2, longer borrower relationships are associated, on average and across all lender types, with better loan terms. However, the benefits of relationship lending are lower at U-SACCOs than they are at banks (Panel A). The same results hold if the dependent variables identify large loans and those with low interest rates and long maturity (Panel B). We interpret the evidence in Tables B1–B2 on lower returns to relationships lending at U-SACCOs than at banks suggesting lending constraints at U-SACCOs. 50

Table B1: Analysis of Switching Loans—Within Borrower and Across Lender Types The table presents within-borrower mean comparison tests of loan terms (loan amount, interest rate spread, and loan maturity). The interest rate spread is the interest rate on the loan minus the repo rate. In panel A, we consider individuals with multiple loans at the U-SACCO before they switch to commercial banks and compare the subsequent loan (or the average of loan terms of the subsequent loans) at the U-SACCO with the first one. In panel B we consider the same individuals who switched from U-SACCOs to commercial banks and compare the switching loan at the bank with the last loan obtained at the U-SACCO before switching. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. LoanAmount Interest Rate Maturity Spread (1) (2) (3) Panel A.SubsequentLoanofSwitcheratU-SACCO- 0.531*** -0.403 5.388*** FirstLoanofSwitcheratU-SACCOBeforeSwitching (0.065) (0.480) (0.385) No. ofSubsequentLoansofSwitchersatU-SACCOBeforeSwitch- 5,181 2,790 5,181 ing No. ofFirstLoansofSwitchersatU-SACCOBeforeSwitching 2,910 1,772 2,910 Panel B.SwitchingLoanatCommercialBank- 0.729*** -3.554*** 7.618*** LastLoanatU-SACCOBeforeSwitching (0.075) (0.507) (0.413) No. ofFutureLoansofSwitchers 2,910 1,694 2,910 No. ofSwitchingLoans 2,910 1,694 2,910 51

Table B2: Lending Constraints at U-SACCOs: The Role of Relationship Length The table presents coefficient estimates of a linear model in which the dependent variable is a measure of loan terms (amount, interest rate, and maturity). In Panel A, loan amount is expressed in millions RWF, interest rate in percent (divided by 100), and maturity in months. In Panel B the dependent variables are defined as follows: (i) a dummy variable equal to one if loan size is larger than the 75th percentile of the sample distribution, and zero otherwise; (ii) a dummy variable equal to one if the interest rate on the loan is lower than the 25th percentile of the sample distribution, and zero otherwise; and (iii) a dummy variable equal to one if loan maturity is larger than the 75th percentile of the sample distribution, and zero otherwise. U-SACCO is a dummy equal to one if the borrower has a loan with a U-SACCO and zero otherwise. Relationship length measured the length of the bankborrower relationship, in months. Borrower controls include a set of dummies for gender, marital status, age, and employment status. See Table 1 for the definition of each variable. The data are at the borrower-municipality-month level, conditional on individuals having an outstanding loan. The sample period is 2008:M1 to 2016:M12. Standard errors clustered at the municipality level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. (1) (2) (3) Panel A Loanamount Interestrate Maturity RelationshipLength×U-SACCO -0.828*** 0.296* -2.294*** (0.108) (0.157) (0.247) RelationshipLength 1.007*** -0.506*** 7.339*** (0.122) (0.037) (0.144) U-SACCO -0.960*** 5.256*** -9.991*** (0.230) (0.507) (0.496) AdjustedR-squared 0.027 0.115 0.262 Panel B Loanamount Interestrate Maturity >75thpercentile <25thpercentile >75thpercentile RelationshipLength×U-SACCO -0.011*** -0.074*** -0.042*** (0.002) (0.005) (0.004) RelationshipLength 0.032*** 0.063*** 0.087*** (0.002) (0.002) (0.002) U-SACCO -0.141*** -0.010 -0.121*** (0.008) (0.015) (0.008) MunicipalityxTimeFE Y Y Y BorrowerControls Y Y Y No. ofObservations 2,816,928 2,816,928 2,816,928 No. ofMunicipalities 337 337 337 No. ofIndividuals 129,523 129,523 129,523 AdjustedR-squared 0.188 0.168 0.193 52

C Additional Figures and Tables Figure C1: Number of Loans of U-SACCOs, Other MFIs, and Commercial Banks 120000 100000 80000 60000 40000 20000 0 snaoL fo rebmuN 2008m1 2010m1 2012m1 2014m1 2016m1 U-SACCOs Other MFIs Commercial Banks Note: The figure shows the evolution over time of the number of outstanding loans by U- SACCOs, other MFIs, and commercial banks at a monthly frequency. Source: Rwandan Credit Reference Bureau. 53

Figure C2: Credit Register Representativeness 1500 1000 500 0 )FWR snoillib( emuloV naoL 2008q3 2009q3 2010q3 2011q3 2012q3 2013q3 2014q3 2015q3 2016q3 Credit Register Banks B/S (aggregate) Note: ThefigureshowstotalbankcreditinbillionsofRwandanfrancs(RWF)forallcommercial banks operating in Rwanda from the credit register (solid line) and aggregate statistics from the banks’ balance sheets (dashed line). Source: Rwandan Credit Reference Bureau, National Bank of Rwanda. 54

Figure C3: Probability of Getting a Loan Before and After the U-SACCO Program—Other MFIs )naoL(rP 21. 1. 80. 60. 40. 20. 0 20.- ]63-t<[ ]13-t,63-t[ ]52-t,03-t[ ]91-t,42-t[ ]31-t,81-t[ ]7-t,21-t[ ]1-t,6-t[ ]6+t,t[ ]21+t,7+t[ ]81+t,31+t[ ]42+t,91+t[ ]03+t,52+t[ ]63+t,13+t[ ]63+t>[ Months Relative to Implementation Loan from Other MFIs Note: The figure shows the effect of the U-SACCO program on the probability of an individual obtaining a loan from other MFIs. The charts plot the estimated coefficients and the associated 90 percent confidence intervals of model 1 with interaction terms between Post U-SACCO and a set of six-month time interval dummies centered on the mt municipality-specific month of program rollout, as reported in column 6 of Table 3. Source: Rwandan Credit Reference Bureau. 55

Figure C4: Share of Individuals with a Loan Before and After the U-SACCO Program (a) Pre-program (b) Post-program Note: The figure shows the share of individuals with an outstanding loan (relative to total adult population, in thousands), by municipality, before and after the U-SACCO program. The pre- and post-program period are defined using as a threshold the month in which the U-SACCO grants its first loan in a given municipality. Source: Rwandan Credit Reference Bureau, National Bank of Rwanda. 56

Figure C5: Probability of Switching, Before and After the U-SACCO Program sreworroB 0001 rep sehctiwS knaB ot sIFM rehto & sOCCAS-U 5.2 2 5.1 1 5. 0 5.- 1- ]63-t<[ ]13-t,63-t[ ]52-t,03-t[ ]91-t,42-t[ ]31-t,81-t[ ]7-t,21-t[ ]1-t,6-t[ ]6+t,t[ ]21+t,7+t[ ]81+t,31+t[ ]42+t,91+t[ ]03+t,52+t[ ]63+t,13+t[ ]63+t>[ Months Relative to Implementation Note: ThefigureshowstheeffectoftheU-SACCOprogramontheprobabilityofswitchingto commercial banks. The chart plots the estimated β coefficients and the associated 90 percent confidence intervals of the following regression in a balanced panel at the municipality-time level: +7 (cid:88) P(Switching ) = β Post U-SACCO ×Time Interval +α +φ +υ t+ε , mt T mt T m t m mt T=−7 where Switching is the fraction of switchers from all microfinance institutions—that is, mt U-SACCOs and other MFIs—to commercial banks (per 1000 borrowers) in municipality m at time (year:month) t. The effect of the dummy variable Post U-SACCO (equal to mt one after a U-SACCO starts its lending activities in a given municipality m at time t, and zero otherwise) is decomposed in six-month time intervals T centered on the municipalityspecific month of program rollout. Municipality fixed effects α and time (year:month) fixed m effects φ control for unobserved time-invariant spatial factors and common time-varying t shocks. υ t are municipality-specific time trends and standard errors are clustered at the m municipality level. Source: Rwandan Credit Reference Bureau. 57

epyT redneL yb ,scitsitatS yrammuS :1C elbaT yb ,elpmas ruo ni selbairav niam eht rof )DS—noitaived dradnats dna ,naidem ,naem( scitsitats yrammus stneserp elbat ehT ohw seitilapicinum 733 ni slaudividni 831,091 sedulcni dna 6102 rebmeceD–8002 yraunaJ si doirep elpmas ehT .epyt rednel .secruos atad dna elbairav hcae fo noitinfied eht rof 1 elbaT eeS .sIFM rehto dna ,sOCCAS-U ,sknab laicremmoc morf worrob emas eht sa ,elpmas eht ni sreworrob fo rebmun latot eht naht regral si epyt rednel yb sreworrob fo rebmun eht fo mus ehT .epyt rednel eno naht erom morf worrob yam laudividni )501=n(sIFMrehtO )61=n(sknablaicremmoC )103=n(sOCCAS-U DS naideM naeM N DS naideM naeM N DS naideM naeM N )21( )11( )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( tiderC ot sseccA .A 000.0 000.1 000.1 345,245,1 000.0 000.1 000.1 659,051,2 000.0 000.1 000.1 697,561,1 noitutitsnIynamorfnaoL 322.0 000.0 350.0 345,245,1 702.0 000.0 540.0 659,051,2 000.0 000.1 000.1 697,561,1 OCCASmorfnaoL 662.0 000.0 770.0 345,245,1 000.0 000.1 000.1 659,051,2 672.0 000.0 380.0 697,561,1 knaBmorfnaoL 000.0 000.1 000.1 345,245,1 822.0 000.0 550.0 659,051,2 552.0 000.0 070.0 697,561,1 IFMrehtomorfnaoL scitsiretcarahC naoL .B 324.4 974.0 604.1 098,565,1 48.72 849.0 402.5 825,712,2 789.0 003.0 455.0 232,351,1 erusopxEnaoL 382.5 009.0 421.2 098,565,1 70.83 876.1 062.7 825,712,2 380.1 005.0 348.0 232,351,1 tnuomAlapicnirPnaoL 75.11 00.41 07.71 786,582,1 51.01 00.81 59.81 699,941,2 78.91 00.02 38.42 239,666 )%(etaRtseretnI 75.32 39.22 15.62 098,565,1 31.23 00.83 49.04 825,712,2 02.31 00.21 10.81 232,351,1 )shtnom(ytirutaM 923.0 000.0 321.0 098,565,1 343.0 000.0 631.0 825,712,2 192.0 000.0 390.0 232,351,1 naoLgnimrofreP-noN scitsiretcarahc reworroB .C 654.0 000.0 492.0 474,75 984.0 000.0 793.0 300,09 524.0 000.0 732.0 661,67 elameF 813.0 000.0 411.0 474,75 903.0 000.0 701.0 300,09 992.0 000.0 001.0 661,67 elgniS 374.0 000.0 833.0 474,75 684.0 000.0 483.0 300,09 974.0 000.0 853.0 661,67 gnuoY 964.0 000.0 723.0 474,75 533.0 000.0 921.0 300,09 672.0 000.0 380.0 661,67 eeyolpmEtnemnrevoG scitsiretcarahc ytilapicinuM .D 745.3 000.0 399.0 733 745.3 000.0 399.0 733 830.3 000.0 108.0 103 htworGcimonocE 019.9 000.0 075.2 733 019.9 000.0 075.2 733 668.8 000.0 490.2 103 tnempoleveDcimonocE 521.0 444.0 024.0 733 521.0 444.0 024.0 733 221.0 844.0 624.0 103 ytrevoP 470.0 000.0 640.0 733 470.0 000.0 640.0 733 070.0 000.0 240.0 103 ecneserPknaB 570.5 29.21 58.31 733 570.5 29.21 58.31 733 569.4 88.21 17.31 103 noitalupoPtludA 634.3 000.0 258.0 733 634.3 000.0 258.0 733 768.2 000.0 827.0 103 stciflnoC 998.0 985.1 617.1 733 998.0 985.1 617.1 733 898.0 766.1 467.1 103 latipaCotecnatsiD 954.0 000.0 007.0 733 954.0 000.0 007.0 733 844.0 000.0 427.0 103 laruR 58

Table C2: Summary Statistics: U-SACCOs’ Balance Sheet Characteristics The table reports the average values of U-SACCO characteristics, measured in 2010, before the program. U-SACCO members is the number of members with shares. Total assets and total deposits are measured in millions RWF. Capital is the ratio of capital over total assets. Liquidity is the ratio of cash over total assets. Data source: Rwandan Cooperative Agency. N Mean Median SD (1) (2) (3) (4) U-SACCO Members 337 1.723 1.470 1.056 U-SACCO Assets 337 20.74 16.67 15.08 U-SACCO Deposits 337 15.64 11.90 13.73 U-SACCO Capital Ratio 337 0.289 0.258 0.147 U-SACCO Liquidity Ratio 337 0.948 1.000 0.182 59

Table C3: The Timing of the U-SACCO Program: Early versus Late Starters The table report the mean values of a set of municipality-level characteristics measured before the U-SACCO program, separately for municipalities that are early and late starters in the program, depending on the time when the U-SACCO opened. Late starters are the last 25 percent of municipalities in granting the first U-SACCO loan. See Tables 1 and 2 for the definition of each variable. Late Early starter starter (n=83) (n=254) Economic Growth 1.389 0.863 Economic Development 3.571 2.243 Poverty 0.419 0.420 Bank Presence 0.059 0.041 U-SACCO Members 1.586 1.768 U-SACCO Assets 19.38 21.18 U-SACCO Deposits 14.80 15.92 U-SACCO Capital Ratio 0.286 0.290 U-SACCO Liquidity Ratio 0.952 0.947 Adult Population 13.89 13.83 Conflicts 1.482 0.646 Distance to Capital 1.750 1.704 Rural 0.663 0.713 60

Table C4: Access to Credit—Alternative Difference-in-differences Estimator The table presents two-way fixed effect estimates of model 1 following De Chaisemartin and d’Haultfoeuille (2020), where the dependent variable is a dummy equal to one for individuals who, at time t, have an outstanding loan with any financial institution (column 1) or only U-SACCOs (column 2), commercial banks (column 3), or other MFIs (column 4). Post U-SACCO is a dummy equal to one after a U-SACCO starts its lending activity in a given municipality and month, and zero otherwise. Borrower controls include a set of dummies for gender, marital status, age, and employment status. See Table 1 for the definition of each variable. The Post U-SACCO dummy is split in a set of dummies for six-month intervals around the date when the U-SACCO starts its lending activity in the municipality. The data are at the borrower-municipality-month level. The sample period is 2008:M1 to 2016:M12. Bootstrapped (1,000 replications) standard errors clustered at the municipality level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Dummy=1ifindividualhasaLoanin AnyInstitution U-SACCO Bank OtherMFI (1) (2) (3) (4) PostU-SACCO[t-36,t-31] -0.0002 0.0000 -0.0001 0.0000 (0.0002) (0.0000) (0.0001) (0.0001) PostU-SACCO[t-30,t-25] -0.0003 0.0000 -0.0003 -0.0001 (0.0004) (0.0000) (0.0003) (0.0001) PostU-SACCO[t-24,t-19] -0.0002 0.0000 -0.0002 -0.0001 (0.0003) (0.0000) (0.0003) (0.0002) PostU-SACCO[t-18,t-13] 0.0002 0.0000 0.0001 0.0000 (0.0003) (0.0000) (0.0002) (0.0003) PostU-SACCO[t-12,t-7] 0.0001 0.0000 0.0000 0.0001 (0.0004) (0.0000) (0.0002) (0.0003) PostU-SACCO[t-6,t-1] 0.0004 0.0000 0.0003 0.0002 (0.0004) (0.0000) (0.0003) (0.0003) PostU-SACCO[t,t+6] 0.0128*** 0.0143*** 0.0012*** -0.0002*** (0.0016) (0.0014) (0.0008) (0.0007) PostU-SACCO[t+7,t+12] 0.0415*** 0.0460*** 0.0039*** -0.0005*** (0.0036) (0.0030) (0.0022) (0.002) PostU-SACCO[t+13,t+18] 0.0728*** 0.0793*** 0.0082*** -0.0015*** (0.0064) (0.0046) (0.0043) (0.0042) PostU-SACCO[t+19,t+24] 0.0905*** 0.0960*** 0.0155*** -0.0048*** (0.0093) (0.0054) (0.0062) (0.0068) PostU-SACCO[t+25,t+30] 0.1016*** 0.1083*** 0.0201*** -0.0085*** (0.0125) (0.0056) (0.0082) (0.0092) PostU-SACCO[t+31,t+36] 0.1125*** 0.1143*** 0.026*** -0.0082*** (0.0154) (0.0059) (0.0108) (0.0105) MunicipalityFE Y Y Y Y Time(Year:month)FE Y Y Y Y BorrowerControls Y Y Y Y MunicipalityTimeTrends Y Y Y Y No. ofObservations 34,272,396 34,272,396 34,272,396 34,272,396 No. ofMunicipalities 337 337 337 337 No. ofIndividuals 317,337 317,337 317,337 317,337 61

erutcurtS ataD evitanretlA—tiderC ot sseccA no margorP OCCAS-U eht fo tceffE :5C elbaT ot level htnom-ytilapicinum-reworrob eht ta tesatad lanigiro eht gnispalloc 1 ledom fo setamitse tneicffieoc stneserp elbat ehT slaudividni rof eno ot lauqe ymmud a si elbairav tnedneped ehT .ycneuqerf )01–6 snmuloc( ylraey ro )5–1 snmuloc( ylretrauq a 3 snmuloc( sOCCAS-U ni yllacfiiceps ro )7–6 dna 2–1 snmuloc( snoitutitsni yna htiw naol gnidnatstuo na evah ,t emit ta ,ohw retfa eno ot lauqe ymmud a si OCCAS-U tsoP .)01 dna 5 snmuloc( sIFM rehto ro )9 dna 4 snmuloc( sknab laicremmoc ,)8 dna tes a edulcni slortnoc reworroB .esiwrehto orez dna ,htnom dna ytilapicinum nevig a ni ytivitca gnidnel sti strats OCCAS-U a snmuloc nI .elbairav hcae fo noitinfied eht rof 1 elbaT eeS .sutats tnemyolpme dna ,ega ,sutats latiram ,redneg rof seimmud fo strats OCCAS-U eht nehw etad eht dnuora slavretni htnom-xis rof seimmud fo tes a ni tilps si ymmud OCCAS-U tsoP eht 5–2 htnom-21 rof seimmud fo tes a ni tilps si ymmud OCCAS-U tsoP eht ,01–7 snmuloc nI .ytilapicinum eht ni ytivitca gnidnel sti eht rof selbairav ymmud ehT .ytilapicinum eht ni ytivitca gnidnel sti strats OCCAS-U eht nehw etad eht dnuora slavretni doirep elpmas ehT .level emit-ytilapicinum-reworrob eht ta era atad ehT .ecaps fo nosaer rof nwohs ton era doirep margorp-erp .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni level ytilapicinum eht ta deretsulc srorre dradnatS .21M:6102 ot 1M:8002 si ylraeY ylretrauQ ninaoLasahlaudividnifi1=ymmuD IFMrehtO knaB OCCAS-U noitutitsnIynA IFMrehtO knaB OCCAS-U noitutitsnIynA )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( ***910.0 ***120.0 OCCAS-UtsoP )600.0( )500.0( 200.0 100.0 ***210.0 ***210.0 ]6+t,t[OCCAS-UtsoP )200.0( )100.0( )300.0( )300.0( 300.0 400.0 ***140.0 ***930.0 ]21+t,7+t[OCCAS-UtsoP )300.0( )300.0( )600.0( )700.0( 400.0 300.0 ***220.0 ***420.0 ]21+t,t[OCCAS-UtsoP )300.0( )400.0( )600.0( )700.0( 400.0 *800.0 ***760.0 ***660.0 ]81+t,31+t[OCCAS-UtsoP )400.0( )500.0( )800.0( )010.0( 300.0 **410.0 ***770.0 ***080.0 ]42+t,91+t[OCCAS-UtsoP )500.0( )600.0( )900.0( )210.0( 400.0 900.0 ***760.0 ***860.0 ]42+t,31+t[OCCAS-UtsoP )500.0( )800.0( )210.0( )510.0( 400.0 **710.0 ***180.0 ***680.0 ]03+t,52+t[OCCAS-UtsoP )700.0( )800.0( )010.0( )410.0( 600.0 **420.0 ***480.0 ***790.0 ]63+t,13+t[OCCAS-UtsoP )800.0( )010.0( )010.0( )610.0( 500.0 *120.0 ***570.0 ***880.0 ]63+t,52+t[OCCAS-UtsoP )700.0( )210.0( )510.0( )020.0( 010.0 **830.0 ***580.0 ***811.0 700.0 **130.0 ***190.0 ***111.0 ]63+t>[OCCAS-UtsoP )900.0( )810.0( )810.0( )620.0( )010.0( )310.0( )110.0( )910.0( Y Y Y Y Y Y Y Y Y Y EFytilapicinuM N N N N N Y Y Y Y Y EF)retrauq:raeY(emiT Y Y Y Y Y N N N N N EF)raeY(emiT Y Y Y Y Y Y Y Y Y Y slortnoCreworroB Y Y Y Y Y Y Y Y Y Y sdnerTemiTytilapicinuM 925,558,2 925,558,2 925,558,2 925,558,2 925,558,2 611,224,11 611,224,11 611,224,11 611,224,11 611,224,11 snoitavresbOfo .oN 733 733 733 733 733 733 733 733 733 733 seitilapicinuMfo .oN 182,713 182,713 182,713 182,713 182,713 182,713 182,713 182,713 182,713 182,713 slaudividnIfo .oN 241.0 570.0 031.0 291.0 191.0 451.0 870.0 221.0 002.0 991.0 derauqs-RdetsujdA 62

leveL htnoM-ytilapicinuM eht ta lenaP—tiderC ot sseccA no margorP OCCAS-U eht fo tceffE :6C elbaT gnikat dna level htnom-ytilapicinum eht ta tesatad lanigiro eht gnispalloc 1 ledom fo setamitse tneicffieoc stneserp elbat ehT ni( noitalupop tluda eht yb delacs ro )5–1 snmuloc( rebmun etulosba ni—slaudividni fo rebmun eht elbairav tnedneped sa snmuloc( sOCCAS-U ;)7–6 dna 2–1 snmuloc( snoitutitsni yna htiw naol gnidnatstuo na evah ohw—)01–6 snmuloc ,sdnasuoht eno ot lauqe ymmud a si OCCAS-U tsoP .)01 dna 5 snmuloc( sIFM rehto ro ;)9 dna 4 snmuloc( sknab laicremmoc ;)8 dna 3 edulcni slortnoc reworroB .esiwrehto orez dna ,htnom dna ytilapicinum nevig a ni ytivitca gnidnel sti strats OCCAS-U a retfa nI .elbairav hcae fo noitinfied eht rof 1 elbaT eeS .sutats tnemyolpme dna ,ega ,sutats latiram ,redneg rof seimmud fo tes a nehw etad eht dnuora slavretni htnom-xis rof seimmud fo tes a ni tilps si ymmud OCCAS-U tsoP eht ,01–7 dna 5–2 snmuloc rof nwohs ton era doirep margorp-erp eht rof selbairav ymmud ehT .ytilapicinum eht ni ytivitca gnidnel sti strats OCCAS-U eht dradnatS .21M:6102 ot 1M:8002 si doirep elpmas ehT .level emit-ytilapicinum-reworrob eht ta era atad ehT .ecaps fo nosaer .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni level ytilapicinum eht ta deretsulc srorre ytilapicinuMnislaudividnIfonoitcarF ytilapicinuMnislaudividnI .oN :morfgniworroB :morfgniworroB rehtO knaB OCCAS-U noitutitsnIynA rehtO knaB OCCAS-U noitutitsnIynA IFM IFM )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( ***711.1 ***529.11 OCCAS-UtsoP )122.0( )853.3( 550.0 921.0 ***907.0 ***717.0 036.1 547.0 ***465.7 ***019.7 ]6+t,t[OCCAS-UtsoP )970.0( )001.0( )741.0( )471.0( )955.1( )340.2( )957.1( )817.2( 760.0 713.0 ***446.2 ***305.2 934.2 580.3 ***762.03 ***773.92 ]21+t,7+t[OCCAS-UtsoP )351.0( )991.0( )962.0( )833.0( )460.3( )167.3( )141.3( )241.5( 840.0 **436.0 ***383.4 ***432.4 972.2 046.7 ***414.15 ***621.15 ]81+t,31+t[OCCAS-UtsoP )212.0( )603.0( )463.0( )184.0( )782.4( )144.5( )812.4( )604.7( 230.0- **420.1 ***760.5 ***101.5 291.1 *995.31 ***453.06 ***880.36 ]42+t,91+t[OCCAS-UtsoP )472.0( )324.0( )234.0( )126.0( )885.5( )611.7( )580.5( )775.9( 520.0- **362.1 ***923.5 ***125.5 105.1 **133.91 ***639.36 ***323.17 ]03+t,52+t[OCCAS-UtsoP )923.0( )345.0( )874.0( )757.0( )797.6( )228.8( )236.5( )117.11( 731.0 **127.1 ***535.5 ***362.6 762.4 ***722.82 ***550.66 ***247.38 ]63+t,13+t[OCCAS-UtsoP )973.0( )866.0( )935.0( )409.0( )519.7( )547.01( )381.6( )599.31( 771.0 **750.2 ***060.6 ***501.7 195.5 **279.33 ***434.17 ***272.59 ]63+t>[OCCAS-UtsoP )944.0( )678.0( )685.0( )111.1( )877.9( )208.31( )174.6( )124.71( Y Y Y Y Y Y Y Y Y Y EFytilapicinuM Y Y Y Y Y Y Y Y Y Y EF)htnom:raeY(emiT Y Y Y Y Y Y Y Y Y Y sdnerTemiTytilapicinuM 693,63 693,63 693,63 693,63 693,63 693,63 693,63 693,63 693,63 693,63 snoitavresbOfo .oN 733 733 733 733 733 733 733 733 733 733 seitilapicinuMfo .oN 229.0 149.0 768.0 849.0 449.0 929.0 469.0 088.0 769.0 469.0 derauqs-RdetsujdA 63

slaudividnI gnipporD dna stceffE dexiF reworroB gniddA—tiderC ot sseccA no margorP OCCAS-U eht fo tceffE :7C elbaT worroB reveN ohw ohw sreworrob gnidulcxe ro )5–1 snmuloc( stceffe dexfi reworrob gnidda 1 ledom fo setamitse tneicffieoc stneserp elbat ehT rof eno ot lauqe ymmud a si elbairav tnedneped ehT .)01–6 snmuloc( doirep elpmas eht gnirud naol a detnarg reven erew dna 3 snmuloc( sOCCAS-U ;)7–6 dna 2–1 snmuloc( snoitutitsni yna htiw naol gnidnatstuo na evah ,t emit ta ,ohw slaudividni a retfa eno ot lauqe ymmud a si OCCAS-U tsoP .)01 dna 5 snmuloc( sIFM rehto ro ;)9 dna 4 snmuloc( sknab laicremmoc ;)8 fo tes a edulcni slortnoc reworroB .esiwrehto orez dna ,htnom dna ytilapicinum nevig a ni ytivitca gnidnel sti strats OCCAS-U 5–2 snmuloc nI .elbairav hcae fo noitinfied eht rof 1 elbaT eeS .sutats tnemyolpme dna ,ega ,sutats latiram ,redneg rof seimmud OCCAS-U eht nehw etad eht dnuora slavretni htnom-xis rof seimmud fo tes a ni tilps si ymmud OCCAS-U tsoP eht ,01–7 dna fo nosaer rof nwohs ton era doirep margorp-erp eht rof selbairav ymmud ehT .ytilapicinum eht ni ytivitca gnidnel sti strats srorre dradnatS .21M:6102 ot 1M:8002 si doirep elpmas ehT .level emit-ytilapicinum-reworrob eht ta era atad ehT .ecaps .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni level ytilapicinum eht ta deretsulc naoLadetnarGreveNereWtahtsreworroBporD stceffEdexiFreworroB :ninaoLasahlaudividnifi1=ymmuD rehtO knaB OCCAS-U noitutitsnIynA rehtO knaB OCCAS-U noitutitsnIynA IFM IFM )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( ***230.0 ***120.0 OCCAS-UtsoP )700.0( )500.0( 300.0 200.0 ***120.0 ***120.0 200.0 200.0 ***310.0 ***410.0 ]6+t,t[OCCAS-UtsoP )300.0( )300.0( )400.0( )500.0( )200.0( )100.0( )300.0( )300.0( 500.0 600.0 ***760.0 ***560.0 300.0 400.0 ***240.0 ***140.0 ]21+t,7+t[OCCAS-UtsoP )500.0( )500.0( )800.0( )110.0( )300.0( )300.0( )600.0( )700.0( 600.0 310.0 ***011.0 ***801.0 400.0 *800.0 ***960.0 ***860.0 ]81+t,31+t[OCCAS-UtsoP )800.0( )800.0( )110.0( )510.0( )400.0( )400.0( )800.0( )010.0( 600.0 **320.0 ***721.0 ***231.0 400.0 **410.0 ***080.0 ***380.0 ]42+t,91+t[OCCAS-UtsoP )010.0( )110.0( )310.0( )910.0( )600.0( )600.0( )900.0( )210.0( 700.0 **130.0 ***431.0 ***541.0 500.0 **810.0 ***480.0 ***190.0 ]03+t,52+t[OCCAS-UtsoP )210.0( )410.0( )410.0( )320.0( )700.0( )800.0( )010.0( )410.0( 110.0 **240.0 ***931.0 ***361.0 800.0 **420.0 ***880.0 ***201.0 ]63+t,13+t[OCCAS-UtsoP )410.0( )710.0( )410.0( )620.0( )800.0( )010.0( )010.0( )610.0( 210.0 **550.0 ***941.0 ***681.0 900.0 **130.0 ***590.0 ***611.0 ]63+t>[OCCAS-UtsoP )810.0( )220.0( )510.0( )230.0( )010.0( )210.0( )010.0( )910.0( Y Y Y Y Y Y Y Y Y Y EFytilapicinuM Y Y Y Y Y Y Y Y Y Y EF)htnom:raeY(emiT N N N N N Y Y Y Y Y EFreworroB Y Y Y Y Y - - - - slortnoCreworroB Y Y Y Y Y Y Y Y Y Y sdnerTemiTytilapicinuM 409,435,02 409,435,02 409,435,02 409,435,02 409,435,02 843,662,43 843,662,43 843,662,43 843,662,43 843,662,43 snoitavresbOfo .oN 733 733 733 733 733 733 733 733 733 733 seitilapicinuMfo .oN 831,091 831,091 831,091 831,091 831,091 182,713 182,713 182,713 182,713 182,713 slaudividnIfo .oN 271.0 521.0 661.0 562.0 462.0 014.0 224.0 862.0 314.0 214.0 derauqs-RdetsujdA 64

gniretsulC dna sdnerT emiT—tiderC ot sseccA no margorP OCCAS-U eht fo tceffE :8C elbaT srorre dradnats eht gniretsulc ro )5–1 snmuloc( sdnert emit citardauq gnidda 1 ledom fo setamitse tneicffieoc stneserp elbat ehT na evah ,t emit ta ,ohw slaudividni rof eno ot lauqe ymmud a si elbairav tnedneped ehT .)01–6 snmuloc( level tcirtsid eht ta dna 4 snmuloc( sknab laicremmoc ;)8 dna 3 snmuloc( sOCCAS-U ;)7–6 dna 2–1 snmuloc( snoitutitsni yna htiw naol gnidnatstuo ytivitca gnidnel sti strats OCCAS-U a retfa eno ot lauqe ymmud a si OCCAS-U tsoP .)01 dna 5 snmuloc( sIFM rehto ro ;)9 ,sutats latiram ,redneg rof seimmud fo tes a edulcni slortnoc reworroB .esiwrehto orez dna ,htnom dna ytilapicinum nevig a ni OCCAS-U tsoP eht ,01–7 dna 5–2 snmuloc nI .elbairav hcae fo noitinfied eht rof 1 elbaT eeS .sutats tnemyolpme dna ,ega ni ytivitca gnidnel sti strats OCCAS-U eht nehw etad eht dnuora slavretni htnom-xis rof seimmud fo tes a ni tilps si ymmud eht ta era atad ehT .ecaps fo nosaer rof nwohs ton era doirep margorp-erp eht rof selbairav ymmud ehT .ytilapicinum eht ytilapicinum eht ta deretsulc srorre dradnatS .21M:6102 ot 1M:8002 si doirep elpmas ehT .level emit-ytilapicinum-reworrob .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni level )01-6 snmuloc( tcirtsid ro )5–1 snmuloc( level-tcirtsiDehttaderetsulCsrorrEdradnatS sdnerTemiTcitardauQ ninaoLasahlaudividnifi1=ymmuD rehtO knaB OCCAS-U noitutitsnIynA rehtO knaB OCCAS-U noitutitsnIynA IFM IFM )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( ***120.0 ***020.0 OCCAS-UtsoP )500.0( )500.0( 200.0 200.0 ***310.0 ***410.0 200.0 100.0 ***310.0 ***310.0 ]6+t,t[OCCAS-UtsoP )200.0( )100.0( )300.0( )300.0( )100.0( )100.0( )300.0( )300.0( 300.0 400.0 ***240.0 ***140.0 300.0 400.0 ***140.0 ***040.0 ]21+t,7+t[OCCAS-UtsoP )300.0( )300.0( )700.0( )700.0( )300.0( )300.0( )600.0( )700.0( 400.0 *800.0 ***960.0 ***860.0 400.0 *800.0 ***760.0 ***760.0 ]81+t,31+t[OCCAS-UtsoP )500.0( )400.0( )110.0( )110.0( )400.0( )400.0( )800.0( )010.0( 400.0 **410.0 ***080.0 ***380.0 400.0 **410.0 ***870.0 ***180.0 ]42+t,91+t[OCCAS-UtsoP )600.0( )600.0( )310.0( )410.0( )500.0( )600.0( )900.0( )210.0( 500.0 **810.0 ***480.0 ***190.0 500.0 **810.0 ***180.0 ***980.0 ]03+t,52+t[OCCAS-UtsoP )800.0( )700.0( )310.0( )510.0( )700.0( )800.0( )900.0( )310.0( 800.0 ***420.0 ***880.0 ***201.0 800.0 **420.0 ***480.0 ***990.0 ]63+t,13+t[OCCAS-UtsoP )900.0( )800.0( )410.0( )610.0( )800.0( )900.0( )010.0( )610.0( 900.0 **130.0 ***590.0 ***611.0 900.0 **130.0 ***190.0 ***211.0 ]63+t>[OCCAS-UtsoP )110.0( )110.0( )510.0( )020.0( )010.0( )210.0( )010.0( )910.0( Y Y Y Y Y Y Y Y Y Y EFytilapicinuM Y Y Y Y Y Y Y Y Y Y EF)htnom:raeY(emiT Y Y Y Y Y Y Y Y Y Y slortnoCreworroB Y Y Y Y Y N N N N N sdnerTemiTytilapicinuM N N N N N Y Y Y Y Y sdnerTemiTcitardauQ 843,662,43 843,662,43 843,662,43 843,662,43 843,662,43 843,662,43 843,662,43 843,662,43 843,662,43 843,662,43 snoitavresbOfo .oN 733 733 733 733 733 733 733 733 733 733 seitilapicinuMfo .oN 182,713 182,713 182,713 182,713 182,713 182,713 182,713 182,713 182,713 182,713 slaudividnIfo .oN 651.0 870.0 121.0 102.0 002.0 651.0 870.0 121.0 102.0 002.0 derauqs-RdetsujdA 65

Table C9: Effect of the U-SACCO Program on Access to Credit—Falsification Test (Randomizing Treatment Dates) The table presents coefficient estimates of model 1 where the dependent variable is a dummy equal to one for individuals who, at time t, have an outstanding loan with any financial institution (columns 1 and 2); U-SACCOs (column 3); commercial banks (column 4); or other MFIs (column 5). Post U-SACCO is a dummy constructed by randomly assigning the treatment across municipalities and over time. Specifically, for each municipality we randomly assign the program implementation date in the interval 2008:M1–2016:M12 and repeat this exercise 1,000 times. The table reports the average coefficients of the simulation. Borrower controls include a set of dummies for gender, marital status, age, and employment status. See Table 1 for the definition of each variable. In columns 2–5 and 7–10, the Post U-SACCO dummy is split in a set of dummies for six-month intervals around the date when the U-SACCO starts its lending activity in the municipality. The dummy variables for the pre-program period are not shown for reason of space. The data are at the borrower-municipality-time level. The sample period is 2008:M1 to 2016:M12. Standard errors clustered at the municipality (columns 1–5) or district (columns 6-10) level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. AnyInstitution U-SACCO Bank OtherMFI (1) (2) (3) (4) (5) PostU-SACCO 0.0001 (0.0039) PostU-SACCO[t,t+6] 0.0000 0.0000 0.0000 0.0000 (0.0019) (0.0014) (0.0010) (0.0008) PostU-SACCO[t+7,t+12] 0.0001 0.0001 0.0000 0.0000 (0.0034) (0.0025) (0.0017) (0.0014) PostU-SACCO[t+13,t+18] 0.0001 0.0002 0.0000 0.0000 (0.0048) (0.0035) (0.0023) (0.0019) PostU-SACCO[t+19,t+24] 0.0001 0.0002 0.0000 0.0000 (0.0061) (0.0045) (0.0029) (0.0025) PostU-SACCO[t+25,t+30] 0.0001 0.0003 -0.0001 0.0000 (0.0075) (0.0055) (0.0035) (0.0030) PostU-SACCO[t+31,t+36] 0.0001 0.0003 -0.0001 0.0000 (0.0088) (0.0065) (0.0041) (0.0036) PostU-SACCO[>t+36] -0.0001 0.0003 -0.0002 -0.0001 (0.011) (0.0081) (0.0052) (0.0043) MunicipalityFE Y Y Y Y Y Time(Year:month)FE Y Y Y Y Y BorrowerControls Y Y Y Y Y MunicipalityTimeTrends Y Y Y Y Y No. ofObservations 34,266,348 34,266,348 34,266,348 34,266,348 34,266,348 No. ofMunicipalities 337 337 337 337 337 No. ofIndividuals 317,281 317,281 317,281 317,281 317,281 AdjustedR-squared 0.200 0.200 0.117 0.078 0.156 66

Table C10: New Loans to First-time Borrowers at U-SACCOs versus Commercial Banks The table presents average loan and borrower characteristics for new loans of first-time borrowers at U-SACCOs and commercial banks. The detailed variable definitions are available in Table 1. Loan amount is expressed in millions RWF, interest rate in percentage points and maturity in months. NewLoansbyU-SACCOs NewLoansbyBanks toFirst-timeBorrowers toFirst-timeBorrowers (n=33,491) (n=83,759) LoanAmount 0.536 2.383*** InterestRate 25.36 21.13*** LoanMaturity 14.50 23.41*** Female 0.237 0.400*** Single 0.080 0.103*** Young 0.188 0.217*** GovernmentEmployee 0.015 0.139*** EconomicGrowthinMunicipality 0.827 3.793*** EconomicDevelopmentinMunicipality 2.157 10.49*** PovertyinMunicipality 0.426 0.331*** BankPresenceinMunicipality 0.042 0.091*** AdultPopulationinMunicipality 13.70 18.21*** NumberofConflictsinMunicipality 0.669 2.186*** DistancetoCapitalfromMunicipality 1.645 1.409*** RuralMunicipality 0.249 0.600*** 67

Table C11: Analysis of Switching Loans—Summary Statistics The table presents average loan and borrower characteristics for the treatment (column 1) and control groups (columns 2–3) in the switching analysis described in Section 5. The detailed variable definitions are available in Table 1. Column 1 refers to loans of borrowers who switch from a U-SACCO to a commercial bank, column 2 to loans of U-SACCO borrowers who do not switch, and column 3 to loans of commercial bank borrowers. All variables except borrower characteristics are measured ex ante, in the pre-program period. Columns 2 and 3 report the result of a t-test of equality of means with column 1. *** p<0.01, ** p<0.05, * p<0.1. U-SACCO to NewLoansbyU- New Loans by Banks SACCOs Banks SwitchingLoans to Non- to Other Bor- Switchers rowers (n=2,910) (n=76,512) (n=230,974) (1) (2) (3) Female 0.269 0.227*** 0.266 Single 0.114 0.092*** 0.088*** Young 0.174 0.191** 0.194*** GovernmentEmployee 0.115 0.087*** 0.348*** No. ofYearsofCreditHistorybeforeSwitching 0.787 0.740** 1.097*** No. ofLendingRelationshipsbeforeSwitching 0.613 0.441*** 0.583*** No. ofLoansbeforeSwitching 0.685 0.530*** 0.781*** TotalOutstandingLoanAmountbeforeSwitching 0.435 0.443 5.257*** EconomicGrowthinMunicipality 0.991 1.006 4.216*** EconomicDevelopmentinMunicipality 2.573 2.674 11.64*** PovertyinMunicipality 0.410 0.420*** 0.311*** BankPresenceinMunicipality 0.049 0.044*** 0.097*** AdultPopulationinMunicipality 14.542 13.854*** 19.16*** NumberofConflictsinMunicipality 0.855 0.675*** 2.487*** DistancetoCapitalfromMunicipality 1.642 1.623 1.283*** RuralMunicipality 0.688 0.730*** 0.332*** 68

gnihctaM erocS ytisneporP—snaoL gnihctiwS fo sisylanA :21C elbaT naol ,ylevitanretla ,si elbairav tnedneped eht erehw gnihctam erocs ytisneporp gnisu setamitse stneicffieoc stroper elbat ehT sOCCAS-U morf hctiws ohw sreworrob ot snaol wen fo stsisnoc puorg tnemtaert ehT .ytirutam naol ro ,etar tseretni naol ,ezis rehctiws-non ot sOCCAS-U yb snaol wen fo desirpmoc si puorg lortnoc ehT .snaol gnihctiws ,si taht—sknab laicremmoc ot detnarg era taht snaol hctam eW .)6–4 snmuloc( sreworrob rehto ot sknab laicremmoc yb snaol wen ro )3–1 snmuloc( sreworrob ,elgnis ,elamef( scitsiretcarahc reworrob no desab naol gnihctiws hcae fo erocs ytisneporp eht etupmoc dna htnom emas eht ni ,spihsnoitaler gnidnel fo .on ,yrotsih tiderc fo sraey fo .on( scitsiretcarahc yrotsih tiderc etna xe ,)eeyolpme tnemnrevog ,gnuoy rednel yna htiw sLPN sa llew sa ,naol wen eht erofeb htnom eht ni reworrob a fo tnuoma naol gnidnatstuo latot dna ,snaol fo .on fo scitsiretcarahc dna ,)laretalloc dna egagtrom( scitsiretcarahc naol ,)naol wen eht erofeb emit yna ro sraey owt ro eno nihtiw eht ,FWR snoillim ni desserpxe si tnuoma naol ehT .margorp eht erofeb derusaem sediser reworrob eht erehw ytilapicinum eht elbaT ni elbaliava era snoitinfied elbairav deliated ehT .shtnom ni ytirutam naol eht dna ,stniop egatnecrep ni etar tseretni naol .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni srorre dradnatS .1 :puorGlortnoC :puorGlortnoC sreworrobrehtootsknabybsnaolweN srehctiws-nonotsOCCAS-UybsnaolweN ytirutaM etaRtseretnI tnuomAnaoL ytirutaM etaRtseretnI tnuomAnaoL )3( )2( )1( )3( )2( )1( ***889.2 201.0- 231.0 ***71.01 ***931.6- ***428.0 snaoLrehtO-naoLgnihctiwS )745.0( )852.0( )911.0( )834.0( )094.0( )080.0( )gnihctamerocsytisneporp( ***766.3 **136.0 ***751.2- ***323.9 ***470.5- ***188.0 snaoLrehtO-naoLgnihctiwS )184.0( )662.0( )214.0( )622.0( )293.0( )120.0( )gnihctamtuohtiw( :selbairaVgnihctaM Y Y Y Y Y Y ecnaussInaoLfohtnoM:raeY Y Y Y sOCCAS-U Y Y Y sknaBlaicremmoC Y Y Y Y Y Y elameF Y Y Y Y Y Y elgniS Y Y Y Y Y Y gnuoY Y Y Y Y Y Y eeyolpmEtnemnrevoG Y Y Y Y Y Y gnihctiwSerofebyrotsiHtiderCfosraeYfo .oN Y Y Y Y Y Y gnihctiwSerofebspihsnoitaleRgnidneLfo .oN Y Y Y Y Y Y gnihctiwSerofebsnaoLfo .oN Y Y Y Y Y Y gnihctiwSerofebtnuomAnaoLgnidnatstuOlatoT Y Y Y Y Y Y gnihctiwSerofebraey1nihtiwredneLynahtiwLPN Y Y Y Y Y Y gnihctiwSerofebsraey2nihtiwredneLynahtiwLPN Y Y Y Y Y Y gnihctiwSerofebredneLynahtiwLPN Y Y Y Y Y Y egagtroM Y Y Y Y Y Y laretalloC Y Y Y Y Y Y ytilapicinuMnihtworGcimonocE Y Y Y Y Y Y ytilapicinuMnitnempoleveDcimonocE Y Y Y Y Y Y ytilapicinuMniytrevoP Y Y Y Y Y Y ytilapicinuMniecneserPknaB Y Y Y Y Y Y ytilapicinuMninoitalupoPtludA Y Y Y Y Y Y ytilapicinuMnistciflnoC Y Y Y Y Y Y ytilapicinuMmorflatipaCotecnatsiD Y Y Y Y Y Y ytilapicinuMlaruR 019,2 019,2 019,2 019,2 019,2 019,2 )detaerT(srehctiwSfo .oN 479,032 479,032 479,032 215,67 215,67 215,67 sreworroBdetaertnUfo .oN 69

redneL-nihtiW—sreworroB gnihctiwS fo sisylanA :31C elbaT tnedneped eht erehw rotamitse gnihctam detcerroc-saib )1102( snebmI dna eidabA eht fo setamitse stneicffieoc stroper elbat ehT ohw sreworrob ot snaol wen fo stsisnoc puorg tnemtaert ehT .ytirutam naol ro ,etar tseretni naol ,tnuoma naol eht si elbairav eht yb snaol wen fo desirpmoc si puorg lortnoc ehT .snaol gnihctiws ,si taht—sknab laicremmoc ot sOCCAS-U morf hctiws dehctiws rehctiws eht taht sknab laicremmoc eht yb snaol wen ro )3–1 snmuloc( sreworrob rehctiws-non ot OCCAS-U s’rehctiws srobhgien-tseraen eht tceles dna htnom emas eht ni detnarg era taht snaol hctam eW .)6–4 snmuloc( sreworrob rehto ot tub ot yrotsih tiderc etna xe ,)eeyolpme tnemnrevog ,gnuoy ,elgnis ,elamef( scitsiretcarahc reworrob no desab naol gnihctiws hcae fo tnuoma naol gnidnatstuo latot dna ,snaol fo .on ,spihsnoitaler gnidnel fo .on ,yrotsih tiderc fo sraey fo .on( scitsiretcarahc erofeb emit yna ro sraey owt ro eno nihtiw rednel yna htiw sLPN sa llew sa ,naol wen eht erofeb htnom eht ni reworrob a fo sediser reworrob eht erehw ytilapicinum eht fo scitsiretcarahc dna ,)laretalloc dna egagtrom( scitsiretcarahc naol ,)naol wen eht dna ,stniop egatnecrep ni etar tseretni naol eht ,FWR snoillim ni desserpxe si tnuoma naol ehT .margorp eht erofeb derusaem *** .sesehtnerap ni srorre dradnatS .1 elbaT ni elbaliava era snoitinfied elbairav deliated ehT .shtnom ni ytirutam naol eht .1.0<p * ,50.0<p ** ,10.0<p :puorGlortnoC :puorGlortnoC knabgniviecerehtybsnaolweN srehctiws-nonotOCCAS-Us’rehctiwsybsnaolweN ytirutaMnaoL etaRtseretnI tnuomAnaoL ytirutaMnaoL etaRtseretnI tnuomAnaoL )3( )2( )1( )6( )5( )4( ***059.2 712.0- 731.0 ***968.7 ***329.6- ***236.0 snaoLrehtO-naoLgnihctiwS )493.0( )532.0( )011.0( )866.0( )948.0( )080.0( :selbairaVgnihctaM Y Y Y Y Y Y ecnaussInaoLfohtnoM:raeY Y Y Y rehctiwSfoOCCAS-UsuoiverP Y Y Y rehctiwSfoknaBlaicremmoC Y Y Y Y Y Y elameF Y Y Y Y Y Y elgniS Y Y Y Y Y Y gnuoY Y Y Y Y Y Y eeyolpmEtnemnrevoG Y Y Y Y Y Y gnihctiwSerofebyrotsiHtiderCfosraeYfo .oN Y Y Y Y Y Y gnihctiwSerofebspihsnoitaleRgnidneLfo .oN Y Y Y Y Y Y gnihctiwSerofebsnaoLfo .oN Y Y Y Y Y Y gnihctiwSerofebtnuomAnaoLgnidnatstuOlatoT Y Y Y Y Y Y gnihctiwSerofebraey1nihtiwredneLynahtiwLPN Y Y Y Y Y Y gnihctiwSerofebsraey2nihtiwredneLynahtiwLPN Y Y Y Y Y Y gnihctiwSerofebredneLynahtiwLPN Y Y Y Y Y Y egagtroM Y Y Y Y Y Y laretalloC Y Y Y Y Y Y ytilapicinuMniecneserPknaB Y Y Y Y Y Y ytilapicinuMnihtworGcimonocE Y Y Y Y Y Y ytilapicinuMnitnempoleveDcimonocE Y Y Y Y Y Y ytilapicinuMniytrevoP Y Y Y Y Y Y ytilapicinuMninoitalupoPtludA Y Y Y Y Y Y ytilapicinuMnistciflnoC Y Y Y Y Y Y ytilapicinuMmorflatipaCotecnatsiD Y Y Y Y Y Y ytilapicinuMlaruR 978,2 978,2 978,2 143,1 143,1 143,1 )detaerT(srehctiwSfo .oN 479,032 479,032 479,032 215,67 215,67 215,67 sreworroBdetaertnUfo .oN 70

OCCAS-U htiw pihsnoitaleR dereveS susrev deunitnoC yB—snaoL gnihctiwS fo sisylanA :41C elbaT rotamitse gnihctam robhgien-tseraen detcerroc-saib )1102( snebmI dna eidabA eht fo setamitse stneicffieoc stroper elbat ehT litnu ro sraey owt ro eno nihtiw gnimrofrep-non semoceb naol a fi eno ot lauqe ymmud a si elbairav tnedneped eht erehw retfa OCCAS-U eht htiw pihsnoitaler gnidnel eht deunitnoc taht srehctiws senimaxe A lenaP .esiwrehto orez dna ,ytirutam gnihctiws retfa OCCAS-U eht htiw pihsnoitaler gnidnel eht dereves taht srehctiws no sesucof B lenaP elihw ,sknab ot gnihctiws snaol wen si ti 6–4 snmuloc ni elihw ,srehctiws-non ot sOCCAS-U yb snaol wen era 3–1 snmuloc ni puorg lortnoc ehT .sknab ot desab naol gnihctiws hcae fo srobhgien-tseraen eht tceles dna htnom emas eht ni detnarg era taht snaol hctam eW .sknab yb fo sraey fo .on( scitsiretcarahc yrotsih tiderc etna xe ,)eeyolpme tnemnrevog ,gnuoy ,elgnis ,elamef( scitsiretcarahc reworrob no erofeb htnom eht ni reworrob a fo tnuoma naol gnidnatstuo latot dna ,snaol fo .on ,spihsnoitaler gnidnel fo .on ,yrotsih tiderc scitsiretcarahc naol ,)naol wen eht erofeb emit yna ro sraey owt ro eno nihtiw rednel yna htiw sLPN sa llew sa ,naol wen eht ehT .margorp eht erofeb derusaem sediser reworrob eht erehw ytilapicinum eht fo scitsiretcarahc dna ,)laretalloc dna egagtrom( ehT .shtnom ni ytirutam naol eht dna ,stniop egatnecrep ni etar tseretni naol eht ,FWR snoillim ni desserpxe si tnuoma naol .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni srorre dradnatS .1 elbaT ni elbaliava era snoitinfied elbairav deliated :puorGlortnoC :puorGlortnoC sreworrobrehtootsknabybsnaolweN srehctiws-nonotOCCAS-UybsnaolweN -utamlitnu-LPN 2 nihtiw - LPN 1 nihtiw - LPN -utamlitnu-LPN 2 nihtiw - LPN 1 nihtiw - LPN ytir sraey raey ytir sraey raey )6( )5( )4( )3( )2( )1( 010.0 210.0 400.0- ***830.0- ***240.0- ***140.0snaoLrehtO-naoLgnihctiwS )310.0( )310.0( )210.0( )410.0( )310.0( )210.0( gnidnelehtdeunitnoctahtsrehctiwS( ]335,1[ ]335,1[ ]335,1[ ]335,1[ ]335,1[ ]335,1[ )sOCCAS-Uehthtiwpihsnoitaler 610.0 410.0 200.0 ***940.0- ***250.0- ***140.0snaoLrehtO-naoLgnihctiwS )210.0( )210.0( )110.0( )410.0( )410.0( )310.0( gnidnelehtderevestahtsrehctiwS( ]773,1[ ]773,1[ ]773,1[ ]773,1[ ]773,1[ ]773,1[ )sOCCAS-Uehthtiwpihsnoitaler :selbairaVgnihctaM Y Y Y Y Y Y ecnaussInaoLfohtnoM:raeY Y Y Y sOCCAS-U Y Y Y sknaBlaicremmoC Y Y Y Y Y Y elameF Y Y Y Y Y Y elgniS Y Y Y Y Y Y gnuoY Y Y Y Y Y Y eeyolpmEtnemnrevoG Y Y Y Y Y Y gnihctiwSerofebyrotsiHtiderCfosraeYfo .oN Y Y Y Y Y Y gnihctiwSerofebspihsnoitaleRgnidneLfo .oN Y Y Y Y Y Y gnihctiwSerofebsnaoLfo .oN Y Y Y Y Y Y gnihctiwSerofebtnuomAnaoLgnidnatstuOlatoT Y Y Y Y Y Y gnihctiwSerofebraey1nihtiwredneLynahtiwLPN Y Y Y Y Y Y gnihctiwSerofebsraey2nihtiwredneLynahtiwLPN Y Y Y Y Y Y gnihctiwSerofebredneLynahtiwLPN Y Y Y Y Y Y egagtroM Y Y Y Y Y Y laretalloC Y Y Y Y Y Y tnuomAnaoL Y Y Y Y Y Y etaRtseretnInaoL Y Y Y Y Y Y ytirutaMnaoL Y Y Y Y Y Y ytilapicinuMniecneserPknaB Y Y Y Y Y Y ytilapicinuMnihtworGcimonocE Y Y Y Y Y Y ytilapicinuMnitnempoleveDcimonocE Y Y Y Y Y Y ytilapicinuMniytrevoP Y Y Y Y Y Y ytilapicinuMninoitalupoPtludA Y Y Y Y Y Y ytilapicinuMnistciflnoC Y Y Y Y Y Y ytilapicinuMmorflatipaCotecnatsiD Y Y Y Y Y Y ytilapicinuMlaruR 71

Table C15: Analysis of Switching Loans—Subsequent Loans at the Bank The table presents within-borrower mean comparison tests of loan terms—loan amount, interest rate spread, and loan maturity—between the new loans at the same commercial banksthattheU-SACCOborrowerswitchedto, andtheoriginalswitchingloan. Theinterest rate spread is the interest rate on the loan minus the repo rate. When there is more than one loan after the switching one, these loans are averaged to compute the average amount, interest rate spread, and maturity. The loan amount is expressed in millions RWF, the loan interest rate in percentage points, and the loan maturity in months. The sample period is 2008:M1 to 2016:M12. Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. LoanAmount InterestRateSpread LoanMaturity (1) (2) (3) NewLoans-SwitchingLoan -0.099 0.212 -0.561 (0.148) (0.524) (0.660) No. ofFutureLoansofSwitchers 1,050 1,019 1,050 No. ofSwitchingLoans 2,910 2,910 2,910 72

Table C16: Analysis of Switchers’ Default Risk At U-SACCOs Before Switching The table reports coefficients estimates of the Abadie and Imbens (2011) bias-corrected nearest-neighbor matching estimator where the dependent variable is a dummy equal to one if a loan becomes non-performing within one or two years or until maturity, and zero otherwise. The treatment group consists of loans to U-SACCO-to-bank switcher at the U- SACCO before switching. The control group is comprised of new loans by U-SACCOs to non-switcher borrowers. We match loans that are granted in the same month and select the nearest-neighbors of each switching loan based on borrower characteristics (female, single, young, government employee), ex ante credit history characteristics (no. of years of credit history, no. of lending relationships, no. of loans, and total outstanding loan amount of a borrower in the month before the new loan, as well as NPLs with any lender within one or two years or any time before the new loan), loan characteristics (mortgage, collateral, amount, interest rate, maturity), and characteristics of the municipality where the borrower resides measured before the program. The detailed variable definitions are available in Table 1. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. NPL - within 1 NPL - within 2 NPL - until mayear years turity (1) (2) (3) LoanofSwitcheratU-SACCOBeforeSwitching- -0.015** -0.017** -0.014* LoanofNon-SwitcheratU-SACCO (0.008) (0.008) (0.008) Matching Variables: Year:MonthofLoanIssuance Y Y Y U-SACCOs Y Y Y Female Y Y Y Single Y Y Y Young Y Y Y GovernmentEmployee Y Y Y No. ofYearsofCreditHistorybeforeSwitching Y Y Y No. ofLendingRelationshipsbeforeSwitching Y Y Y No. ofLoansbeforeSwitching Y Y Y TotalOutstandingLoanAmountbeforeSwitching Y Y Y NPLwithanyLenderwithin1yearbeforeSwitching Y Y Y NPLwithanyLenderwithin2yearsbeforeSwitching Y Y Y NPLwithanyLenderbeforeSwitching Y Y Y Mortgage Y Y Y Collateral Y Y Y LoanAmount Y Y Y LoanInterestRate Y Y Y LoanMaturity Y Y Y BankPresenceinMunicipality Y Y Y EconomicGrowthinMunicipality Y Y Y EconomicDevelopmentinMunicipality Y Y Y PovertyinMunicipality Y Y Y AdultPopulationinMunicipality Y Y Y ConflictsinMunicipality Y Y Y DistancetoCapitalfromMunicipality Y Y Y RuralMunicipality Y Y Y No. ofLoansofSwitchersatU-SACCOsBeforeSwitching 3,615 3,615 3,615 No. ofUntreatedBorrowers 76,512 76,512 76,512 73

Table C17: Lending Constraints at U-SACCOs The table presents coefficient estimates from a regression of loan terms (amount, interest rate, and maturity) on a U-SACCO dummy variable (equal to one if the loan if granted from a U-SACCO and 0 from a commercial bank) and balance sheet characteristics. The data are at the borrower-municipality-month level. Balance sheet characteristics are measured at 2010 year-end and refer to total deposits, capital-to-asset ratio and liquidity-to-asset ratio (total cash divided by total assets). The data refer to loans from U-SACCOs or commercial banks (loans from other MFIs are excluded). The sample period is 2008:M1 to 2016:M12. Standard errors clustered at the municipality level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. (1) (2) (3) (4) (5) (6) Loanamount Interestrate Maturity U-SACCO -2.777*** -0.775 5.487*** -0.131 -10.192*** 4.271 (0.265) (2.783) (0.613) (5.044) (0.425) (7.953) Deposits 1.248*** -1.742*** 6.440*** (0.165) (0.146) (0.189) CapitalRatio 3.653* 0.661 25.558*** (1.941) (2.463) (2.394) LiquidityRatio 7.405*** -7.747 33.291*** (2.705) (5.102) (8.192) Municipality×Year:monthFE Y Y Y Y Y Y BorrowerControls Y Y Y Y Y Y No. ofObservations 308,982 308,982 308,982 308,982 308,982 308,982 No. ofMunicipalities 337 337 337 337 337 337 No. ofIndividuals 129,523 129,523 129,523 129,523 129,523 129,523 AdjustedR-squared 0.001 0.004 0.146 0.160 0.156 0.209 74

srehctiwS knaB-ot-IFM dna sreworroB knaB emiT-tsriF htiw nosirapmoC—snaoL gnihctiwS fo sisylanA :81C elbaT snmuloc( spuorg lortnoc dna )3 dna 1 snmuloc( tnemtaert eht rof scitsiretcarahc reworrob dna naol egareva stneserp elbat ehT a morf hctiws ohw sreworrob fo snaol ot refer 3 dna 1 snmuloC .5 noitceS ni debircsed sisylana gnihctiws eht ni )4 dna 2 wen ot 3 nmuloc dna ,sreworrob emit-tsrfi ot sknab laicremmoc yb snaol wen ot 2 nmuloc ,knab laicremmoc a ot OCCAS-U naoL .1 elbaT ni elbaliava era snoitinfied elbairav deliated ehT .sIFM rehto morf srehctiws ot sknab laicremmoc yb snaol eht troper 4 dna 2 snmuloC .shtnom ni ytirutam dna stniop egatnecrep ni etar tseretni ,FWR snoillim ni desserpxe si tnuoma ,50.0<p ** ,10.0<p *** .)snaol gnihctiws( ylevitcepser ,4 dna 2 snmuloc ni detroper seulav eht htiw ytilauqe fo tset-t a fo tluser .1.0<p * -remmoC yb snaoL weN sknaB ot OCCAS-U -remmoC yb snaoL weN sknaB ot OCCAS-U srehctiwS ot sknaB laic snaoLgnihctiwS emit-tsriF ot sknaB laic snaoLgnihctiwS sIFMrehtomorf sreworroB )527,5=n( )918,2=n( )957,38=n( )898,2=n( ***883.2 755.1 ***383.2 545.1 tnuomAnaoL ***75.12 46.02 ***31.12 95.02 etaRtseretnI 539.12 08.12 ***14.32 40.22 ytirutaMnaoL 350.0 750.0 ***770.0 750.0 egagtroM ***980.0 120.0 ***370.0 020.0 laretalloC ***792.0 362.0 ***004.0 862.0 elameF ***911.0 890.0 301.0 111.0 elgniS **141.0 161.0 ***712.0 471.0 gnuoY ***052.0 301.0 ***931.0 111.0 eeyolpmEtnemnrevoG 977.0 887.0 587.0 gnihctiwSerofebyrotsiHtiderCfosraeYfo .oN ***825.0 116.0 216.0 gnihctiwSerofebspihsnoitaleRgnidneLfo .oN 286.0 186.0 486.0 gnihctiwSerofebsnaoLfo .oN ***908.0 534.0 334.0 gnihctiwSerofebtnuomAnaoLgnidnatstuOlatoT ***790.0 940.0 ***190.0 940.0 ytilapicinuMniecneserPknaB ***898.4 699.0 ***397.3 299.0 ytilapicinuMnihtworGcimonocE ***81.31 185.2 ***94.01 675.2 ytilapicinuMnitnempoleveDcimonocE ***103.0 014.0 ***133.0 014.0 ytilapicinuMniytrevoP ***74.91 65.41 ***12.81 35.41 ytilapicinuMninoitalupoPtludA ***093.2 868.0 ***681.2 358.0 ytilapicinuMnistciflnoC ***371.1 446.1 ***904.1 346.1 ytilapicinuMmorflatipaCotecnatsiD ***396.0 413.0 ***006.0 213.0 ytilapicinuMlaruR 75

Table C19: Information Requests at Credit Bureau and New Bank Loans The table presents coefficient estimates of a regression of the logarithm of the number of new loans as a function of the logarithm of the number of information requests in the credit register (columns 1–3) or the logarithm of the number of information requests in the credit register that successfully identified the individual being queried (columns 4–6). The data are organized at the bank-municipality-quarter level and cover the period from January 2018 to December 2019. Standard errors clustered at the bank-municipality pair level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. (1) (2) (3) (4) (5) (6) No. NewLoans(ln) No. ofInformationRequests(ln) 0.0234*** 0.0571*** 0.1391*** (0.007) (0.007) (0.017) No. ofInformationRequestswith 0.0258*** 0.0594*** 0.1361*** SuccessfullyIdentifiedIndividuals(ln) (0.007) (0.007) (0.017) Municipality×Time(Year:quarter)FE N Y Y N Y Y BankFE N N Y N N Y No. ofObservations 26,864 26,864 26,864 26,864 26,864 26,864 AdjustedR-squared 0.002 0.226 0.500 0.003 0.227 0.500 76

noitaerC tnemyolpmE dna mriF no margorP OCCAS-U eht fo stceffE laeR :02C elbaT ehT .noitaerc tnemyolpme dna mrfi-llams no margorp OCCAS-U eht fo stceffe eht rof setamitse tneicffieoc stneserp elbat ehT elos wen fo rebmun eht ,)seeyolpme xis naht erom on htiw smrfi( sesirpretneorcim wen fo rebmun eht era selbairav tnedneped llA .ezis fo sseldrager ,smrfi demrof ylwen ta srekrow fo rebmun latot eht dna ,)eeyolpme eno htiw smrfi( spihsroteirporp eht fo latrop atad eht morf deniatbo ,susneC tnemhsilbatsE 4102 eht morf emoc atad ehT .sgol ni era selbairav tnedneped .demrof saw mrfi hcae nehw )htnom:raey( etad eht dna ytilapicinum eht sedivorp hcihw ,adnawR scitsitatS fo etutitsnI lanoitaN .esiwrehtoorezdna ,htnomdnaytilapicinumneviganiytivitcagnidnelstistratsOCCAS-UaretfaenootlauqeymmudasitsoP rebmeceD dna 8002 yraunaJ neewteb detaerc smrfi ot elpmas eht tcirtser eW .level htnom-ytilapicinum eht ta era atad ehT ecneserp knab wol ni ymmud tsoP eht no tneicffieoc eht taht tset dedis-eno a morf eulav-p eht stroper swor mottob ehT .4102 ecneserP knaB woL×tsoP fo tneicffieoc eht )ii( taht dna seitilapicinum ecneserp knab hgih ni taht naht rehgih si seitilapicinum woleb esoht era seitilapicinum ecneserp knab woL .seitilapicinum gnihctiws wol ni naht rehgih si seitilapicinum gnihctiws hgih ni deretsulc srorre dradnatS .margorp eht erofeb atipac rep sehcnarb knab fo rebmun eht fo noitubirtsid eht fo elitnecrep ht57 eht .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni level ytilapicinum eht ta )6( )5( )4( )3( )2( )1( )nl( seeyolpmE weN fo .oN )nl( spihsroteirporp eloS weN fo .oN )nl( sesirpretneorciM weN fo .oN ***9352.0 ***7312.0 ***9142.0 tsoP )720.0( )020.0( )220.0( ***0272.0 ***1232.0 ***2952.0 ecneserP knaB woL × tsoP )820.0( )020.0( )320.0( ***0512.0 ***8861.0 ***6102.0 ecneserP knaB hgiH × tsoP )660.0( )250.0( )450.0( 803,82 803,82 803,82 803,82 803,82 803,82 snoitavresbO .oN Y Y Y Y Y Y EF ytilapicinuM Y Y Y Y Y Y dnert emiT 9253.0 7153.0 4173.0 5963.0 0204.0 5004.0 derauqs-R detsujdA 0327.0 0327.0 0224.0 0224.0 7265.0 7265.0 elbairav tnedneped fo naeM 77