WPS7984 Policy Research Working Paper 7984 Agent Banking in a Highly Under-Developed Financial Sector Evidence from the Democratic Republic of Congo Robert Cull Xavier Gine Sven Harten Anca Bogdana Rusu Development Research Group Finance and Private Sector Development Team February 2017 Policy Research Working Paper 7984 Abstract The paper provides evidence on the number and volume of transactions are higher in low-income, densely popu- financial transactions undertaken by agents (local businesses lated areas with high levels of commercial development. that double as more convenient, lower cost alternatives This finding suggests that the agent network has been to formal branches) of the largest microfinance insti- best at supporting financial transactions among the tution operating in the Democratic Republic of Congo. urban poor. In addition, branding and effective liquid- More important than agents’ personal characteristics, ity management are strongly linked to agent activity. This paper is a product of the Finance and Private Sector Development Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at rcull@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Agent Banking in a Highly Under-Developed Financial Sector: Evidence from the Democratic Republic of Congo Robert Cull, Xavier Gine, Sven Harten, and Anca Bogdana Rusu1 Keywords: microfinance institutions, agent banking, financial inclusion JEL Codes: G21, G23, G29 1 Robert Cull and Xavier Gine are with the World Bank. Sven Harten is with the German Institute for Development Evaluation. Anca Bogdana Rusu is with the International Finance Corporation. This research was funded under the Partnership for Financial Inclusion, a joint initiative of IFC and The MasterCard Foundation to expand microfinance and advance digital financial services in Sub-Saharan Africa. It brings together the intellectual and financial support of the Foundation with IFC's market knowledge, expertise and client base. The partnership is also supported by The Development Bank of Austria, OeEB, and the Bill & Melinda Gates Foundation. 1. Introduction A growing body of evidence indicates that financial services provide substantial benefits to users in terms of managing risks, absorbing financial shocks, accumulating savings, and investing in businesses, all of which can contribute to less volatile income and consumption patterns.2 And as a result, promotion of broader financial inclusion has become an important objective for policy makers across the globe.3 But transaction costs (e.g., time and expenses associated with travel) and actual costs (e.g., fees for account opening and for conducting financial transactions) associated with using financial services can be prohibitively high for many, and the available financial products from formal providers such as banks are often poorly suited to the needs of large segments of the population.4 Those financial institutions that do cater to the poor are still mainly credit driven and do not offer a full set of financial products and services. This is especially true in developing countries where nearly half (46%) of all adults lack a formal banking account.5 We therefore study an alternative model for providing financial services to poorer clients, agent banking. Agents, trusted local retailers selling everything from sundries to automotive parts, can also double as less formal, lower-cost alternatives to bank branches that enable customers to more conveniently make deposits, withdrawals, money transfers, and payments on loans (Lyman et al., 2006; Siedek, 2008; Mas and Kumar, 2008; Flamming et al., 2011). Because of their lower 2 See e.g. World Bank (2014) for an overview of that literature. 3 Two-thirds of bank regulators surveyed in 143 jurisdictions report that they have a mandate to promote financial inclusion (World Bank, 2014). And more than 50 countries have set formal goals and targets for financial inclusion (Demirguc-Kunt, Klapper, Singer, and Oudheusden, 2015; Maya Declaration Commitments, Alliance for Financial Inclusion, http://www.afi-global.org/maya-declaration-commitments). 4 See Collins et al. (2009) for detailed evidence from financial diaries on the vast array of financial arrangements (both formal and informal) used by the poor in developing countries to manage their economic lives, and the lack of fit between many formal products and their needs. 5 Demirguc-Kunt, Klapper, Singer, and Oudheusden (2015). 2 costs and closer proximity (both physical and social) to typically underserved market segments, agents hold the promise of reaching poorer customers living further from formal bank branches, and thus could more cost-effectively expand financial inclusion than traditional banking and microfinance approaches. We rely on data on the number and volume of transactions from the agents of FINCA DRC, a microfinance institution that has achieved rapid growth in the number of clients that it serves within a short time span despite facing a challenging context in which to deliver financial services. In general, agents are small-scale business owners who offer FINCA banking services in addition to operating an already established retail business.6 At an agent, a FINCA client can make transfers to other FINCA accounts, receive a loan payout or do other withdrawals, repay loans or make deposits into an account. While there is no fee for cash-in transactions, there is a fee for withdrawals. Still, customers use the service for withdrawals since agents are often located closer to their own businesses and queues are generally shorter than at bank branches.7 During the first 3 months of their activity as agents, FINCA agents receive a monthly subsistence stipend of $100. From inception onwards, agents are compensated in relation to the number and value of transactions they facilitate and receive coaching on how to increase foot traffic in their locations. In total, it costs FINCA between $2,500 and $3,000 to set up an agent, of which the largest expense ($1,200) is for the Point of Sale (POS) device. In our regressions, we assume that variables describing local market characteristics (average income levels, level of commercial development) are exogenous in that they are not likely to be affected by agent activities. There is, of course, some nonrandom selection into being an 6 FINCA agents cannot also be agents of banks or other microfinance institutions but can be agents for providers of mobile money services such as Tigo, Airtel, and Orange. 7 Other financial services providers also charge fees on transactions. 3 agent, but because the agents had established businesses prior to the opportunity to become an agent, their locations were pre-determined. In other words, none of the owners relocated their main business to become agents. Moreover, in our analysis we compare agent transactions only with those of other agents, so all of our data set is comprised of retailers that selected into this activity.8 At the time of data collection, FINCA’s was the largest agent banking network in DRC, which makes our results, if not necessarily fully generalizable to other contexts, at least highly relevant for the DRC banking sector.9 Specifically, we regress the number and volume of agents’ cash-in, cash-out transactions on variables describing their personal and business characteristics that are taken from their applications to become agents, and on variables describing income levels, population, population density, commercial development, and financial development for the localized markets in which these agents operate. These are taken from official sources or were created by a local economic consulting firm that we retained. Our key market variables are aggregated at the level of 23 municipalities in and around Kinshasa. Our main findings are that market characteristics explain substantially more variation in transaction activity than do characteristics of the agent or his/her business. In addition, and perhaps more importantly, the number and volume of transactions is highest in low income, densely populated markets with high levels of established commercial development. This suggests to us that FINCA DRC’s agent network has been best at generating financial transactions among the poor in the most densely populated areas.10 To our knowledge, 8 Agents were almost all existing clients, so data on account usage helped FINCA to identify business owners who could be viable agents. 9 FINCA’s is largest agent network for a bank/microfinance institution in the DRC, but mobile money providers have larger networks, mostly because the requirements for the two types of agents differ. Banking agents generally have to meet higher standards, especially in terms of security and liquidity. 10 FINCA only operates in urban and peri-urban environments in our sample. It is conceivable that FINCA would have greater impact on rural customers, but their expansion beyond urban areas is just beginning and thus the topic is left for future research. 4 this is the first econometric study of the activities of agents of a microfinance institution in a developing country, certainly the first in a country as poor and financially underdeveloped as the DRC. Conducting such a study in a country that most often lacks reliable data and where local research capacity is limited was challenging, but this also marks our study as an early contribution to better understanding the economy in DRC.11 Our work is closely related to a recent literature that explores how lowering the transaction costs of transferring funds within social networks improves household financial management. Jack and Suri (2014) have shown how mobile telephony has reduced the costs of such within-network transfers in Kenya and find that mobile money has helped households to smooth consumption in the face of economic shocks. Specifically, they find that shocks reduce the consumption of non- users of a widely adopted mobile money service called M-PESA by 7 percent, while shocks have no significant effect on the consumption of households with an M-PESA user. Similarly, Yang and Choi (2007) find that shocks to the incomes of Philippine households are associated with significant increases in the international remittances that they receive, an indication that those remittances could be used to smooth consumption. Both the volume and diversity of remittance senders increase after a shock and the average distance from senders to receivers of remittances also increases substantially (Jack and Suri, 2014), suggesting that M-PESA has enabled households to expand or make fuller use of their social networks at lower cost. Relatedly, Aycinena, Martinez, and Yang (2009) show that reductions in remittance fees increase the frequency with which Salvadoran migrants receive them. Jack and Suri (2013) also find that the purposes of remittances differ between users and non-users of M- 11 There are, however, less formal studies that describe the business model and provide information on the financial viability of agents. See, for example, http://www.cgap.org/data/india-banking-agents-survey-2012 and http://www.helix-institute.com/data-and-insights/agent-network-accelerator-survey-uganda-country-report-2015 for evidence from India and Uganda, respectively. 5 PESA. Users are more likely to receive remittances via M-PESA for credit or in response to an emergency, while the fraction of total M-PESA transactions for regular support declines. The patterns suggest that M-PESA enables households to more easily draw on their social networks for support in trying circumstances.12 Our work is also therefore related to a body of research showing that monetary transfers within networks of family members and friends are pervasive and crucial for household risk mitigation in developing countries. A number of papers have shown that informal financial arrangements provide insurance for households that experience illness (Gertler and Gruber, 2002; DeWeerdt and Dercon, 2006; Genoni, 2012). For example, Genoni (2012) shows that in Indonesia household members who do not personally experience an illness increase their supply of labor, and transfers increase from other households to the affected household. A related literature tries to uncover the motives for participation in informal financial and insurance arrangements and finds evidence consistent with the notion that reciprocity (rather than altruism) motivates most in- network transfers (Fafchamps and Lund, 2003; Blumenstock, Eagle, and Fafchamps, 2011).13 The remainder of the paper is organized as follows. Section 2 describes FINCA DRC and its place with the financial context of the DRC. Section 3 describes our data and our approach to estimation. Section 4 lays out our econometric specification and provides hypotheses about how our explanatory variables are expected to affect agent transactions. We present our main results in Section 5. In Section 6, we perform a series of robustness checks. Section 7 offers conclusions. 12 Relatedly, Blumenstock, Eagle, and Fafchamps (2015) show that Rwandan households affected by earthquake received increased amounts of cellular “airtime” (a simple precursor to mobile money) from members of their social network, especially those with whom they had already established reciprocal relationships. 13 The reciprocity motive for participation in informal financial networks appears to be a pervasive one in the developing world. For example, using historical loan contract data from low-income rural villages in China in the 1930s, Brandt and Hosios (2010) show that informal loans were predominantly used for consumption purposes and that a large share of these loans carried no interest rate. They argue that these arrangements should be interpreted as “long-term reciprocal insurance or patronage relationships between households.” 6 2. FINCA and the DRC Context Even within Africa, which lags other developing regions in financial development (Allen et al., 2014, 2015: Beck and Cull, 2014, 2015), the Democratic Republic of Congo (DRC) has stood out as one of the least banked countries in the region. In 2014, 17% of adults in the DRC had an account with a formal financial institution. In comparison, 34% of adults in Sub-Saharan Africa, 54% of adults in developing economies, and 62% of adults worldwide had formal accounts.14 Moreover, the 17% figure represents a substantial gain in financial inclusion in the DRC, since the 2011 figure stood at less than 5% (Demirguc-Kunt and Klapper, 2012a,b). The country’s long history of conflict has no doubt led to institutional and infrastructural deficiencies that impede economic growth, and thus also impede demand for financial services. However, evidence from financial diaries in other developing countries (Collins et al., 2009) suggests strongly that poor clients have many financial needs and, because their incomes are often irregular and emergencies arise, their financial lives are unpredictable. The low financial usage ratios reported above therefore make it highly likely that demands for financial services for large segments of the Congolese population go unmet by its formal financial services providers such as commercial banks. FINCA DRC is part of FINCA, an international microfinance institution present in 23 countries worldwide, focused on offering financial services and products to small scale businesses and households. Founded in 1985 in Bolivia, FINCA opened its first program in Africa in 1992 in Uganda. In 2009 FINCA introduced branchless banking – ATMs, computerized points-of-sale transactions, and cellular banking – in an effort to expand its outreach while keeping associated costs low. In late 2016, 2016 FINCA had over 643,000 savers in Africa, an average disbursed loan 14 The figures for account ownership are from Demirguc-Kunt, Klapper, Singer, and Van Oudheusden (2015). 7 size of $754 and a gross loan portfolio of $166 million. Of that, FINCA DRC accounted for 256,000 clients and $77 million of the loan portfolio.15 The products offered by FINCA vary to some extent by local context and are adapted to local regulations. Scaling access to financial services for the majority of the population in the uniquely challenging context of DRC required going beyond existing business models and branch- based delivery channels. Branchless banking is a potentially powerful tool to increase outreach given the even higher cost of establishing branch infrastructure in a post-conflict environment. Since 2012, FINCA DRC has been rolling out an agent banking network that had grown swiftly to more than 500 agents by the end of 2015. With agents facilitating transactions worth over $115,000/month, roughly 65% of all FINCA’s transactions are now done with an agent. As part of the application process the potential agents had to provide information regarding the business owner (age, gender, education, nationality) and her/his business (industry, age, location, inventory value, daily turnover, profits, hours of operation, number of employees). In addition, a FINCA branch officer scored the business on a scale of 0-2 on 10 essential characteristics, ranging from the establishment’s location in terms of potential customers, its security and potential for branding, as well as the aforementioned business-related information such as age, operating schedule, stock value and daily turnover. Both the agent application and the score sheet were used to assess a business’s potential for becoming a FINCA agent. For the period that we study, the selection of agents was done in an opportunistic manner. The vast majority of agents as of 2015 were current clients of FINCA, whose businesses had 15 Globally, however, we acknowledge that FINCA faced financial difficulties dating from at least 2015 due to several factors such as the weakening economic climate and increased competition in many of its key markets. FINCA also had some in-house challenges involving its business model as it had a rather large headcount in the US and a reliance on multiple donor funding sources that has not been conducive to a clear strategic focus. FINCA DRC has been one of the more successful outlets, but here also the influence of multiple donors pulling the organization and its overstretched staff into different directions was noticeable. 8 good track records, and who were therefore approached by branch officers about becoming an agent. As such, less than 0.1% of applications to become an agent have been rejected to this point. There were thus too few rejected applicants (less than ten) to permit meaningful analysis differentiating the characteristics of those who became agents from those who did not.16 3. Data and the Estimation Approach Our data come from 190 FINCA DRC agents who began operations at some point from 2012 to 2015. We rely on monthly transactions data for the dependent variables in the regressions that follow. Although we have some data on other transfers to FINCA accounts, those transactions occurred too infrequently to permit meaningful analysis. Similar to Jack and Suri (2013, 2014), we therefore focus on cash-in, cash-out transactions in our analysis. As explanatory variables, we rely on the personal/business characteristics of the agents and the socio-economic characteristics of the agents’ locations. Personal and business characteristics are taken from the applications that agents submitted to FINCA. Our data are from agents located in and around Kinshasa since that was the focus of FINCA DRC in the initial phases of their agent rollout. Agents are located in four major business districts– Funa, Lukunga, Tshangu and Mont Amba –which are further subdivided into 24 so-called municipalities.17 In our regressions, we rely on data from agents operating in 23 of those municipalities. These municipalities vary greatly in 16 In principle, it might be possible to distinguish the characteristics of agents who ceased operations from those that continued throughout our period of study to assess the viability of the agent model in different contexts. But according to FINCA agent network representatives, hardly any agents quit the business in the first few years of developing the network, precisely because those agents were picked based on ongoing business relationships with branch officers. We acknowledge, however, that 14% of our sample is comprised of agents that are missing operational data for some or all of the last five months covered by our study. We cannot say with certainty whether they quit or took a break from agent work, which sometimes happened due to liquidity issues, or whether data were missing for some other reason. 17 About 10 percent of our transactions data come from agents located outside Kinshasa. We were unable to create reliable measures of income and commercial/financial development for those locations, many of which were larger than a municipality, and thus we have excluded them from the regressions that include municipal-level variables that follow. 9 size, and population density, as well as a variety of other characteristics (Appendix A). Municipality-level data therefore offer a relatively fine level of geographic aggregation. Data on population and population density are taken from official sources. Since DRC lacks most common socio-economic and demographic statistics, data on income, commercial development, and financial development at the municipal level had to be created by a local consulting firm. Specifically, in order to measure these municipality characteristics, we consulted Experts SARL, a market research firm with over 15 years of experience working in the DRC. Given their focus on the city of Kinshasa and their extensive experience with fieldwork across all municipalities of the capital city, we asked them to indicate the level of urban, commercial and financial development of each municipality, as well as their average income levels on a scale of low-medium-high. We also asked them to identify the predominant type of financial services provider in each municipality (banks, microfinance institutions, or providers of mobile banking services). While Experts SARL has an extensive network of enumerators, the core team in Kinshasa is composed of 9-10 enumerators. Each person on the core team is assigned a subset of municipalities, on which they have in depth knowledge and therefore have become the go-to experts in the team. The input of the enumerators was taken into account when determining the municipality characteristics used in our analysis. Their assessment was based on prior knowledge accumulated during a variety of market research projects in Kinshasa, as well as direct observation of area characteristics completed under the first phase of a scoping mission to identify potential locations for future rollout of FINCA agents. The ultimate questions that motivate our research are whether and why some clients would prefer to do their banking transactions with an agent rather than at a formal branch, and what types 10 of agents are best able to attract those clients based on their own characteristics and the characteristics of their locations such as population density. One possibility is that clients are more comfortable with, and trusting of, agents than a branch. This could be especially true for poorer clients who lack knowledge of financial products,18 but could also hold for wealthier clients in DRC where trust in institutions is low and business transactions depend on personal relationships. Indeed, at an agent outlet it may be easier to establish trust since it is usually the same person dealing with clients, often an entrepreneur with a long track record in the community. At branches, young professionals from outside the neighborhood deal with clients and are often rotated or promoted from the entry-level client-facing positions to other jobs. The notion of trust, therefore, may be more abstract with respect to branches and linked to an institution rather than to an individual. We lack a survey that would enable us to track levels of income, financial education, and usage of agents versus branches at the household level. We therefore provide indirect tests of the comfort/trust hypothesis by examining whether proxies for an agent’s standing in the local community such as his/her age and the age of his/her business are associated with more transactions. Gender may also play a role in promoting comfort/trust as female clients may be more comfortable doing banking transactions with a female agent.19 Convenience and a lack of alternative providers may also prompt some clients to rely more heavily on agents. Since the agents that we study are located in and around Kinshasa, all of them are situated reasonably close to a FINCA DRC branch. Thus, most (if not all) clients can choose 18 Equity Bank in Kenya, for example, has made a concerted effort to expand financial inclusion by establishing branches beyond urban centers and having their staff speak to clients in their native languages (Allen et al., 2013). 19 About a quarter of our observations on monthly cash in/cash out come from female agents (see summary statistics in Appendix B). We recognize that age and gender may proxy for factors beyond trust. However, we control for additional characteristics of the agent and his/her business and of the market in which he/she operates, and thus the age of the agent and his/her business (which provide information on how well established he/she is in the community) and gender (which, for female agents, could indicate greater ease in serving female clients) are our best proxies for testing hypotheses regarding trust between agents and their clients. 11 to do transactions with either a branch or an agent. In that sense, they have a choice between at least two alternatives. Still, agents are likely to be physically closer to many potential clients, have longer opening hours and transactions may take less time and be less formal than at a branch. Lacking client survey data, however, we again rely on indirect tests of these hypotheses. Controlling for population and population density (both of which should be positively associated with the frequency and volumes of agent transactions), potential clients in less commercially developed municipalities might benefit less from the convenience of, or have less need for, agents. That is, clients in commercially active municipalities might have greater appreciation for the time- saving convenience of an agent, and thus agent transactions might be more frequent in those areas.20 While agent transactions are likely to be heavier in commercially developed municipalities, how the general level of local financial development affects agent transactions is harder to predict. On the one hand, the level of financial development may simply be an indication that the local demand for financial services is high, and thus we would expect more transactions from both agents and other providers in financially developed municipalities. But the demand for financial services is likely to be strongly linked to the level of local commercial activity. Since we already control for the level of local commercial development in our regressions, financial development could be negatively related to agents’ transactions to the extent that the financial services provided by agents are substitutes for those from other providers. 20 We are assuming that population density serves as a proxy for the number of potential clients that are near an agent. In that sense, density provides information on the convenience associated with reduced travel time to agents rather than branches. In a separate ongoing field experiment, we study how the density of the local FINCA agent network (i.e., the number of agents in close proximity) affects the financial behavior of clients and the profitability of the agents themselves. 12 4. Specification and Hypotheses To identify the factors that drive agents’ transaction frequencies and volumes, we estimate the following equation: = + 1 + 2 + 3 + 4 + Y is the number or the volume of cash-in or cash-out transactions for agent i in municipality m in month t.21 Agent represents personal characteristics of the agent including his/her age, education level, and gender. We estimate all models first using robust standard errors and then using standard errors that allow for clustering at the municipality level. While clustering at the municipality level is clearly a more conservative approach, comparing the significance levels across the two sets of estimates provides a way to gauge the robustness and reliability of particular findings. To the extent that older agents are better established and more trusted than younger ones, we expect agent age to be positively linked to transactions. On the other hand, however, younger agents may be more proficient with technology and therefore able to deliver better service, especially when the need arises to troubleshoot problems with the POS device. If female clients are more likely to use an agent if she is also a woman, we expect that the dummy variable for male agent, which appears in the regressions, would be negatively linked to transactions. The relationship between transactions and agent education is harder to predict. Better educated agents may be more trusted or have business acumen that others lack, and thus we would expect the dummy variable for having a primary education (or less) to be negatively related to the number and volume of transactions. At the same time, highly educated agents might be more socially distant from the bulk of their potential clients, and thus less trusted. In that case, the primary level education variable could be positively linked to agent transactions. 21 We look at cash-in and cash-out transactions separately because the fee structure linked to each operation is different. Similarly, the commissions paid to agents are different for the two types of transactions. 13 Business represents characteristics of the business that operates in tandem with FINCA agent activities. Like agent age, we expect that the age of the agent’s business could be positively linked to agent transactions to the extent that more firmly established businesses inspire trust and/or already have experience with a wider number of potential clients. We also hoped to explore how the nature of the business affects agent transactions. To proxy for interactions between an agent’s established and banking businesses, we tried a number of variables including the number of days per week (and hours per day) that the business is open, the number of employees of the business, and the economic sector of the agent’s business.22 We were unable to derive robust results for those variables, and thus rely on the value of stock (meaning inventory) in the retail business as a simple summary indicator of how demanding are an agent’s duties to maintain that business. To the extent that demands from the retail business on an agent’s time are high, it might make it harder (and relatively less financially lucrative) for her/him to fully pursue banking, and thus we would expect a negative relationship between retail stock value and agent transactions. Market represents a set of variables that describe the demographic characteristics of the municipality including population, population density, commercial development, and average income level. Both population and population density are taken from official sources for the most recent year for which data are available.23 We expect both to be positively linked to the number and volume of agent transactions. Average income levels and the level of commercial development were derived by Experts SARL Consulting, as described above. Lacking a census of businesses and a representative survey of households, we asked Experts SARL to sort municipalities into one of three categories for these variables – high, medium, and low. Our hope was to derive coarse, 22 Unfortunately, we were only able to categorize agents’ businesses as providing either services or manufactured goods. A finer delineation of the types of goods and services that agents sell might have produced better insights into the types of businesses that are best suited to accommodate and foster agent transactions. 23 Population figures are taken from the most recent census, which occurred in 2004. 14 but reliable indicators of the market characteristics that might affect agent transactions based on the consulting firm’s long experience in Kinshasa. We include a dummy variable for high income and another for low commercial density in the regressions that follow. To the extent that agents are an effective way to reach low income market segments that are typically financially excluded, we would expect a negative relationship between the high income dummy and agent transactions.24 The level of commercial development is an indicator of both demand for financial transactions and the ease with which many clients can perform them, especially during the course of the business day. We therefore expect the low commercial density dummy variable to be negatively linked to the number and volume of agent transactions. Note that the market characteristics that we include in the regressions are time- invariant because it was not possible to create reliable measures that varied over time. At the same time, however, we argue that variables such as these are slow to change and thus they should do a good job of accurately sorting municipalities into a few bins within the tight window for which we have agent transactions data. Finally, and as described above, Experts SARL created variables to describe the level of financial development in each municipality, including an overall assessment of financial density as being high, medium, or low, and an assessment of the predominant type of financial institution (banks, microfinance institutions, providers of mobile banking services). To the extent that the services provided by FINCA agents are effective in reaching market segments that are underserved by other providers, we would expect agent transactions to be higher in locations where there are fewer alternative service providers. Thus, we would expect a negative relationship between the dummy variable for high financial density and our agent transactions variables. 24 Our assumption is that higher income clients are more likely to access financial services from providers other than FINCA DRC agents. 15 Three dummy variables are used to indicate whether the predominant type of financial services provider in a municipality is a bank, microfinance institution, or mobile financial services provider, respectively. Since FINCA DRC is itself a microfinance institution, the dummy variable for microfinance predominance could be positively linked to agent transactions because those are municipalities in which microfinance products are in high demand. The dummy variables for banking and mobile banking predominance could provide an indication of whether services from those providers are complementary to (or substitutes for) those provided by FINCA DRC agents. If their coefficients are positive, it would provide evidence consistent with complementarities. 5. Empirical Results In models 1-4 of Table 1 we regress the number of monthly cash-in transactions for each agent on his/her personal characteristics and the characteristics of his/her retail business. We find that an agent’s age is positively linked to transactions, which supports the notion that older agents may be more trusted or at least better known among potential clients. As agent age increases by a year, the number of cash-in transactions increases by roughly four. The median (mean) number of monthly cash-in transactions is 117 (202), with a standard deviation of 259. Our model thus predicts that an agent ten years older than another would have a sizable advantage (almost 40) in performing cash-in transactions. The age variable is significant both in the models that employ robust standard errors and in model 4, which used standard errors clustered at the municipality level. Similarly, and also as hypothesized, female agents perform about 30 more cash-in transactions per month, although that coefficient is significant in only the models that use robust standard errors. There is no significant relationship between the primary education variable and the number of cash-in transactions in our base results in Table 1. 16 Business characteristics also explain variation in the number of cash-in transactions that FINCA DRC agents perform in a month. For example, the coefficient for stock value, our proxy for the demands on the agent’s time that are imposed by her/his business, is negative and significant for models using either type of standard error. The coefficient for stock value is -0.287 in models 3 and 4. If stock value increases by one standard deviation ($64,190), the number of cash-in transactions per month decreases by 18.25 We hypothesized that, like owner age, the age of the retail business would also be positively associated with transactions. However, we find a negative relationship in models 3 and 4. One interpretation is that retail business age is a better proxy for the demands of running the business (or an indication of the interest that an agent has in running his/her retail business relative to banking) rather than an indication of the degree of trust that potential clients place in him/her. However, the coefficient is not significant when standard errors are clustered at the municipality level in any of the models in Table 1. In models 5-8 we begin to examine the effects of market characteristics on the number of monthly cash-in transactions. We introduce fixed effects for each business district (models 5 and 6) and municipality (models 7 and 8) to get an initial indication of the variation in agent transactions that local market conditions can potentially explain.26 The overall fit of the regressions improves dramatically when we introduce these fixed effects, especially the municipality fixed effects. The r-squared of the models with only agent and retail characteristics is .02-.03.; with municipality fixed effects, the r-squared improves to .21. This suggests that substantial variation in the number of agent transactions can potentially be explained by variables that describe the characteristics of markets at the municipal level. Note also that the coefficients for owner age and stock value of the established business remain significant when controlling for district or 25 The stock value variable is expressed in thousands of $US. 64.19*-0.287 = 18.42 fewer transactions. 26 Again, the 23 municipalities in our study fall within four business districts. 17 municipality fixed effects. Indeed, the coefficient for owner age becomes larger in magnitude, increasing from 3.7 to 5.5. In models 9 and 10, we replace the municipality fixed effects with variables designed to capture salient characteristics of the local market. We include population (in 10,000s) and population density (population/sq km), both measured at the municipality level. Coefficients for both are positive and significant, as hypothesized, regardless of the type of standard errors that we use. The mean number of residents in a municipality is 368,661, and the standard deviation is 241,783. Increasing population by 10,000 residents is associated with 3.2 more cash-in transactions per month; increasing by one standard deviation yields 78 additional monthly cash-in transactions.27 Similar, if somewhat less pronounced, relationships are found for population density. Increasing that variable by one standard deviation is associated with 42 additional cash- in transactions per month.28 The variables for municipality income levels and commercial density that were constructed by Experts SARL also explain substantial variation in agents’ cash-in transactions, and they provide indications of the markets where conducting banking transactions with an agent is most desired. For example, the negative significant coefficient for the high income dummy variable implies that agents in high income municipalities perform 97 fewer cash-in transactions per month than those in medium- and low-income municipalities. The negative significant coefficient for municipalities with low levels of commercial development implies that agents in less commercially developed municipalities perform 77 fewer cash-in transactions per month than those in municipalities that have medium or high levels of commercial development. 27 The population coefficient is 3.223. Increasing population by one standard deviation implies 3.223*24.1783 = 77.93 additional cash-in transactions per month. 28 2.232*18.91561 = 42.22 additional transactions. 18 In all, the municipality characteristics variables paint a vivid picture of the types of markets where agent services seem to be flourishing. These are low income, highly and densely populated areas with a relatively high level of commercial activity. These patterns suggest strongly that FINCA DRC agents are targeting and effectively reaching the urban poor. Note also that the r- squared of models 9 and 10 with municipality characteristics is 0.136, an indication that these variables can account for a large share of the substantial gain in the overall fit of the regressions when controlling for municipality-level fixed effects. Finally, we include the variables describing financial development at the municipal level in models 11-16. Agents perform 49 fewer cash-in transactions in municipalities with a dense network of financial service providers, an indication that their services compete with, and thus are substitutes for, those of other providers. In models 13-16, we include the dummy variables describing the predominant type of financial services provider in each municipality. As hypothesized, agent cash-in transactions are more prevalent where microfinance institutions are identified as the dominant providers. This could be a reflection that clients have greater demand for and familiarity with the services from that type of provider since FINCA DRC is itself a microfinance institution. While the patterns for the financial development coefficients are suggestive of where FINCA DRC agents are best fitting into the local financial sector, we put less stock in those results since they are generally insignificant when we use standard errors clustered at the municipality level. In Table 2, the number of cash-out transactions replaces cash-in transactions as the dependent variable and we rerun the same set of regressions. Results are very similar to those for cash-in transactions. Two minor differences are that significance levels for some variables, especially those describing agent and retail business characteristics, are lower and the magnitudes 19 of most coefficients are smaller (in absolute value) than they were for cash-in transactions. However, smaller coefficients are to be expected in that the average number of cash-out transactions is 131 per month, which is sixty-five percent of the average number of cash-in transactions. This is, however, above that for the agent banking sector in other contexts, where much larger gaps are the norm. Recall that there is a fee incurred for cash-out but not cash-in transactions which is likely driving the disparity. Still, the results for cash-out transactions reinforce those for cash-in transactions and indicate that market characteristics explain substantially more variation in agent transactions than personal or business characteristics. In Table 3, we rerun the same regressions using (log of) cash-in volumes as the dependent variable. In general, we find the results to be similar to those for the number of cash-in transactions. In models 1-4, focused on the agent’s personal characteristics and the characteristics of his/her retail business, as well as in models 5-8 where fixed effects for business districts (models 5-6) or municipality (7-8) are introduced, we find the sign and significance of the explanatory variables to be very much in line with those observed in Table 1. Specifically, an agent’s age is significantly positively linked to cash-in volumes, while the stock value of his/her retail business is associated with lower cash-in volumes. We do note however that, while those patterns are similar, the r- squared for the cash-in volume models is systematically lower than that of the cash-in transactions models. Results for models 9-10, in which we introduce variables designed to capture market characteristics, are also similar to those observed for cash-in transactions. However, once we account for financial development at the municipal level (models 11-16), we find that variables such as municipal income level and commercial density lose significance. Population is the only market characteristic that is consistently significant in explaining cash-in volumes, regardless of 20 the standard errors that we use. For an agent with the median level of cash-in volume ($14,927 per month), an increase in municipal population of one standard deviation would be associated with a $4,952 increase in cash-in volume. Finally, and similar to the results for the number of transactions, cash-in volumes are 50-70% higher in areas where microfinance institutions are the dominant financial alternative. In Table 4, we use the monthly volume of cash-out transactions as the dependent variable. In models 1-4 we find that, compared to the results for the number of cash-out transactions, or even cash-in volumes, personal and business characteristics are less able to explain variation in cash-out volumes. While introducing fixed effects for business districts does little to increase the explanatory power of our models, including municipality level fixed effects increases r-squared from .016 (models 5-6) to .161 (models 7-8), which is higher than the r-squared levels for the comparable cash-in volumes models (Table 3, models 7-8). For the agent with the median level of cash-out volume ($3,742), a one standard deviation increase in stock value is associated with a $453 decline in monthly cash-out volume. However, all personal and business characteristics lose significance in models 9-16 once we account for municipality characteristics. Unlike for cash-in volumes, results for cash-out volumes show that, in addition to population, municipality characteristics such as income and commercial density explain substantial variation. Interestingly, the financial density level is not significantly associated with cash-out volumes, regardless of the standard errors that we use. However, and similar to the results for cash-in volumes, cash-out volumes are higher in areas where microfinance institutions are the dominant providers of financial services. In contrast, in municipalities where mobile banking is the dominant financial alternative, the volume of cash–out transactions is dramatically lower (67% lower in model 13; 95% lower in models 15-16). We asked the consulting firm to construct the 21 variable to focus on “mobile banking” using a broad definition that captures mobile money services provided by mobile network operators (MNOs) largely to send and receive remittances. These services are much more common in Africa than mobile banking where the customer connects to his bank account via a smartphone or computer. We do not interpret this coefficient as indicating that competition from providers of mobile banking services causes lower cash-out volumes for FINC DRC agents. It is more plausible that, in the handful of municipalities where mobile money is dominant, FINCA DRC agents simply have yet to establish sizable monthly cash-out volumes, likely due to market features that we are unable to observe directly from our data. For example, the MNOs had pre-existing agent networks and expertise in logistics that gave them advantages in providing cash in/cash out services in some areas. Overall, however, the regression model does about as well in explaining cash-in and cash- out volumes as reflected in the respective r-squared statistics. 6. Robustness Checks, Additional Tests To this point, our results indicate that market characteristics explain more variation in agent transactions than personal and business characteristics, and that agents perform more and larger transactions in low income areas that are dense in terms of population and commercial development. In this section, we create sub-samples based on municipal population density, and we test how agents’ personal/business characteristics and market characteristics (other than population density) affect transactions in densely versus sparsely populated areas. This enables us to examine whether different characteristics drive agent transactions in different types of markets. We also exploit data from regular visits by FINCA DRC headquarters staff that assess how well an agent is following FINCA protocols to test whether such monitoring can help agents to adapt to their markets and increase their transactions. 22 6a. High vs. Low Population Density We rerun our base models after separating municipalities into “high” and “low” population density groups to create sub-samples of roughly equal size.29 Agent age has a positive association with the number of cash-in transactions that is of very similar magnitude in municipalities with high or low population density, though that coefficient only achieves significance in models that use robust standard errors (Table 5).30 However, an agent’s level of educational attainment appears to play a different role in explaining the number of cash-in transactions in high and low density environments. Agents with some high school or college-level education perform substantially more cash-in transactions per month (75 more) in low density areas than those with a primary school education (or less). In contrast, in high density areas, agents with a primary school education perform substantially more transactions than those with more education. Although the coefficients for the primary education variable are not significant in the models that use standard errors clustered at the municipality level, the patterns suggest that trust in agents might work differently in densely and sparsely populated areas. In low density areas, where there are likely fewer daily interactions between agents and potential clients, advanced education might help agents signal that they are competent and trustworthy. Also, within Kinshasa, relatively low density could signal higher income residential neighborhoods. Hence, it could be that most agents are relatively well educated in those neighborhoods, and perform more transactions than less educated agents in those areas. In densely populated areas, where daily interactions are frequent, education seems to have an adverse effect on an agent’s business. This could be because the vast majority of potential 29 Because the number of observations differs across municipalities, there was no possible breakdown by municipal population density that would have resulted in sub-samples of equal size. 30 In part, the reduced number of observations in the sub-sample regressions likely contributes to reduced significance for the models that use standard errors clustered at the municipality level. 23 clients in those areas also have low levels of education, and thus there is less social distance between them and less-educated agents. In terms of business characteristics, the negative relationship between the stock value of an agent’s retail business and the number of cash-in transactions is more pronounced in areas with high population density. This suggests that the demands imposed by his/her other business on an agent’s time are likely to reduce the number of transactions performed in densely populated urban areas. However, and perhaps surprisingly, the negative relationship between retail stock value and cash-in volumes is significant only in low density areas (Table 6). Since the stock value coefficient is small and not robustly significant in low density areas in the models that use the number of cash- in transactions as the dependent variable, this suggests that those agents perform about same the number of transactions but that the average value of those transactions is decreasing in retail stock value. It is not clear to us why agents with high retail stock value would attract clienteles that perform cash-in transactions of smaller average size in low density areas.31 Finally, market characteristics other than population density also appear to impact agents’ cash-in transactions differently in high and low density areas. The coefficients for low levels of commercial development and high average income are negative, larger (in absolute value), and more robustly significant in high density areas when we use either the number or volume of cash- in transactions as the dependent variable. The patterns suggest that service provision is more effectively targeted toward low income areas with relatively high levels of commercial development in municipalities that are more densely populated. To date at least, the services 31 One potential explanation may be that these low density areas are in fact relatively high income areas (such as residential or with governmental bureaucracy and embassies) as we noted above, where businesses have a high stock value since they are selling more expensive goods. Wealthier clients may use agent banking services for small convenient transactions while in an agent’s retail business, but they may also be more likely to go to formal financial institutions for their major transactions. 24 provided by FINCA DRC agents could therefore be viewed as a better fit for the urban poor than for other groups. We derive similar results when we replace the cash-in variables with the cash- out variables on the left-hand-side of our sub-sample regressions, and so we do not report those results in the paper. 6b. Monitoring of FINCA Agents FINCA agents receive unannounced visits from agent network officers to assess how well they are following FINCA protocols and to suggest areas for performance improvement. Agents are supposed to be visited on a monthly basis, though the monitoring reports that we have indicate that visits are less frequent than that for a large number of them. Agents are rated in four broad categories: liquidity management, client service, transactions performance, and branding. Within each category, agents receive scores in multiple sub-categories.32 For example, for liquidity management agents receive a score for the management of “e-float,” that is, money in the agent’s account that can be transferred to a client’s account in exchange for cash, and “cash float” (which can be exchanged for e-float). For client service, agents receive scores for state of shop, connectivity/service availability, and operator availability. For branding, scores are given for agent visibility, posting of FINCA images, and the quality of FINCA signage outside the shop. We sum the scores from the sub-categories to derive a score in each of the four broad categories, which we include in Table 7 in the regressions that we presented earlier.33 Although we lose a large number of observations because most agents did not receive visits in all months for which we have transactions data, the total number of observations in the regressions still 32 It should be noted that this assessment of the agent is not in any way linked to the agent’s performance in terms of number of transactions or volumes transacted. These assessments are carried out in the field by personnel that does not have access to the headquarters database on transactions. 33 Because the transactions performance category and its sub-components were not significantly associated with transaction numbers or volumes, we do not include it among the explanatory variables in the models presented in Table 7. 25 approaches 900. Estimated coefficients and patterns of significance are also similar to those for the base results, although we lose significance on a few coefficients in the models that use standard errors clustered at the municipality level. We note that the market characteristic variables (income, commercial density) lose significance when we include the monitoring variables in our regressions. In part, this could merely reflect changes in sample composition due to the reduced number of observations. Another interpretation, however, is that learning through the monitoring process enables agents to adapt to their markets to increase their transactions, and thus the characteristics of those markets matter less for predicting their transactions performance. Indeed, the large and highly significant positive coefficients for the scores for liquidity management, client services, and branding are striking. Those for liquidity management and branding are especially robust across models. A one standard deviation increase in the liquidity management variable is associated with 53-65 additional cash-in transactions per month, depending on the specification; the same increase in the branding variable implies 40-48 more transactions.34 Note also that the overall fit of the regressions (as reflected in the r-squared statistic) increases to around 0.25 when monitoring variables are included, regardless of the dependent variable or the standard errors that we use. Of course, we do not view the correlations in Table 7 as demonstrating a causal link from improved monitoring scores to more agent transactions. Rather, it seems at least as likely that the monitoring scores could be summarizing characteristics that are unobservable to us as researchers that lead to greater (or less) success of agents. We therefore perform additional tests in which we explain variation in transactions using agent- specific fixed effects and the monitoring variables. Agent fixed effects are included to summarize the average transactions performance of each agent. These models are therefore designed to test 34 For example, the coefficient for the liquidity variable in models 3 and 4 is 20.69 and its standard deviation is 3.12 (the score can range from 0 to 10.5). 20.69*3.12 = 64.6 additional transactions. 26 whether improvements (declines) in an agent’s monitoring scores are associated with more (fewer) transactions than his/her average. Table 8 presents agent fixed effects models using the number of cash-in transactions as the dependent variable; Table 9 uses the log of cash-in volume. Again, we find very similar results when the number of cash-out transactions or their volumes is used as the dependent variable, and so we do not present those results here. In Table 8, the total scores for liquidity management and branding remain highly significant and economically meaningful. For example, models 1 and 2 imply that a one standard deviation increase in the liquidity variable is associated with 21 cash-in transactions more than an agent’s typical monthly number. When we de-compose the liquidity and branding variables into their sub-components in models 3 and 4, it appears that most of the explanatory power is loading on cash-float management for the liquidity variable and FINCA signage for the branding variable. The cash-float result points to the importance of having sufficient cash on hand to operate effectively as an agent. Note that, even though we present this result in the context of a cash-in regression, it also holds when we use the number of cash-out transactions as the dependent variable (results not reported to conserve space). Effective liquidity management, therefore, is an essential element of a successful agent’s business strategy, since failure to perform transactions due to a lack of cash on hand undermines clients’ faith in an agent’s reliability. We do not take the result for FINCA signage too literally since it is correlated with agent visibility, which also approaches significance. Our sense is that it is the overall effort to improve 27 branding, rather than improvement on one particular detail that is associated with increased agent transactions.35 In Table 9, we present agent fixed effects models that use (log of) monthly cash-in volumes as the dependent variable. The coefficient for branding is positive and highly significant, though that for liquidity management is not significant (models 1 and 2). However, when the sub- components of those variables enter the regressions (models 3 and 4), the cash float and the FINCA signage variables are positive and highly significant, just as they were in the models that used the number of cash-in transactions as the dependent variable. The coefficients imply sizable gains in cash-in volumes associated with improvement on these dimensions. For example, for an agent with the median cash-in volume ($14,297 per month), increasing the branding total score by one standard deviation would be associated with a $12,072 gain in monthly cash-in volume. In all, the agent fixed effects models suggest that monitoring the performance of agents and suggesting areas for improvements can help them increase the number and volume of their cash-in/cash-out transactions. 7. Conclusions Does agent banking hold the potential for cost-effective provision of financial services to clients that are typically excluded by the formal financial system? We provide evidence on the transaction activities of the agents of a microfinance institution operating in and around Kinshasa, DRC. To our knowledge, this is some of the first empirical research on agents, and certainly the first evidence from a country as financially underdeveloped as the DRC. 35 That said, the FINCA sign is a distinctive and easily recognized trademark that can only be legally provided through official FINCA channels. The sign may therefore provide an important indication to a client that an agent is officially authorized. 28 We focus on explaining variation in the number and volume of agents’ monthly cash- in/cash-out transactions, and find support for the notion that trust in an agent (because of his/her age or the age of his/her retail business) is associated with more transactions. We also find some evidence that the lack of available financial services in a local market and the suitability of microfinance products for that market are associated with greater numbers and volumes of agent transactions. But even after controlling for those factors, basic characteristics of the local market and the agent’s potential clientele explain the lion’s share of variation in cash-in/cash-out transactions. In this context, at least, agents appear to thrive in densely populated, low income areas that have relatively high levels of commercial activity. We also provide evidence suggesting that monitoring visits can help to improve agents’ numbers and volumes of transactions, regardless of the profile of their local markets. Our results suggest that agents can be effective providers of basic financial services among the urban poor who lack suitable alternatives. But a key challenge is to spur financial inclusion among those in more sparsely populated areas, and our results could be viewed as casting doubt on the notion that agents would be feasible providers in those environments, since they rely so heavily on foot traffic in commercially developed, densely populated areas to drive the number and volume of transactions. To be fair, however, our analysis covers only municipalities in and around Kinshasa and thus it is only our conjecture that the agent model might face additional challenges outside those areas.36 Thus it remains to be seen how the agent banking model will play out in less densely populated areas of the DRC.37 36 In addition, relatively low population density within our sample does not imply that these are geographically remote or poor areas relative to the rest of the DRC. 37 In a country of the size of Western Europe, it may be unrealistic to expect agent banking to reach the most scarcely populated areas. 29 However, the fact that demographic variables explain substantial variation in agent transactions in a relatively homogenous set of urban and peri-urban municipalities suggests to us that more sparsely populated settings will likely pose challenges for the agent banking model. Cost-reducing adaptations of the model, incorporating perhaps mobile banking options, could be required. Also important is the question about the external validity of a model calculated with Kinshasa data. While it should be a good estimate of agent transactions in major cities, further research is needed in smaller cities and towns. Here larger challenges in terms of liquidity management and agent quality management may prove difficult obstacles for operators of agent networks such as FINCA. Going forward, better data on agent profitability over time would help in assessing the viability of this model in different types of markets and survey data from users (and nonusers) of agent services would help to assess how the availability of those services affects financial behavior. 30 References Aker, Jenny C. and Isaac M. Mbiti. 2010. “Mobile Phones and Economic Development in Africa.” Journal of Economic Perspectives 24(3): 207-232. Allen, Franklin, Elena Carletti, Robert Cull, Jun Qian, Lemma Senbet, and Patricio Valenzuela. 2013. “Improving Access to Banking: Evidence from Kenya.” World Bank, Policy Research Working Paper 6593. Allen, Franklin, Elena Carletti, Robert Cull, Jun Qian, Lemma Senbet, and Patricio Valenzuela. 2014. “The African Financial Development and Financial Inclusion Gaps.” Journal of African Economies 23(5): 614-642. Allen, Franklin, Elena Carletti, Robert Cull, Jun Qian, Lemma Senbet, and Patricio Valenzuela. 2015. “Resolving the African Financial Development Gap: Cross-Country Comparisons and a Within Country Study of Kenya.” In Sebastian Edwards, Simon Johnson, and David N. Weil, Editors, African Successes, Volume 3: Modernization and Development, Chicago: University of Chicago Press. Alliance for Financial Inclusion, 2012. “Agent Banking in Latin America.” AFI discussion paper, March. Aycinena, Diego Martínez, and Dean Yang. 2010. “The Impact of Remittance Fees on Remittance Flows: Evidence from a Field Experiment among Salvadorian Migrants.” University of Michigan, mimeo. Beck, Thorsten and Robert Cull. 2015. ““Banking in Africa,” In A.N. Berger, P. Molyneux, and J.O.S. Wilson, Editors. Oxford Handbook of Banking, 2nd Edition, Oxford: Oxford University Press. pp. 913-937. Beck, Thorsten and Robert Cull. 2014. “Banking Systems in Sub-Saharan Africa: A Progress Report.” Revue D’Economique Financiere 116(4): 43-56. Blumenstock, Joshua, Nathan Eagle and Marcel Fafchamps, 2011, “Risk and Reciprocity Over the Mobile Phone Network: Evidence from Rwanda,” CSAE Working Paper 2011-19; Oxford, UK. Brandt, Loren and Arthur J. Hosios, 2010. “Interest-free Loans between Villagers.” Economic Development and Cultural Change 58(2): 345-372. Collins, Daryl, Jonathan Morduch, Stuart Rutherford, and Orlanda Ruthven. 2009. Portfolios of the Poor: How the World’s Poor Live on $2 a Day. Princeton, N.J.: Princeton University Press. De Weerdt, Joachim and Stefan Dercon. 2006. “Risk Sharing Networks and Insurance Against Illness.” Journal of Development Economics 81(2): 337-356. 31 Demirgüç-Kunt, Aslı and Leora Klapper. 2012a. “Measuring Financial Inclusion: The Global Findex Database.” World Bank Policy Research Working Paper 6025, Washington, DC: World Bank. Demirgüç-Kunt, Aslı and Leora Klapper. 2012b. “The Global Findex Database: Financial Inclusion in Sub-Saharan Africa.” Findex Note 4. Washington, DC: World Bank. Demirgüç-Kunt, Aslı, Leora Klapper, Dorothe Singer, and Peter Van Oudheusden. 2015. “The Global Findex Database 2014: Measuring Financial Inclusion around the World.” World Bank Policy Research Working Paper 7255, Washington, DC: World Bank. Fafchamps, Marcel, and Susan Lund. 2013. “Risk Sharing Networks in Rural Philippines.” Journal of Development Economics 71(2): 261-287. Flamming, Mark, Claudia McKay, and Mark Pickens. 2011. Agent Management Toolkit: Building a Viable Network of Branchless Banking Agents. Consultative Group to Assist the Poor (CGAP) Technical Guide. CGAP: Washington, D.C. Genoni, Maria Eugenia. 2012. “Health Shocks and Consumption Smoothing: Evidence from Indoensia.” Economic Development and Cultural Change 60(3): 475-506. Gertler, Paul, and Jonathan Gruber. 2002. “Insuring Consumption Against Illness.” American Economic Review 92(1): 51-76. Jack, William, Adam Ray, and Tavneet Suri. 2013. “Money Management by Households and Firms in Kenya.” American Economic Review: Papers and Proceedings 103(3): 1-8. Jack, William and Tavneet Suri. 2014. “Risk Sharing and Transactions costs: Evidence from Kenya’s Mobile Money Revolution.” American Economic Review 104(1): 183-223. Lyman, Timothy, Gautam Ivatury, and Stefan Staschen. 2006. “Use of Agents in Branchless for the Poor: Rewards, Risks, and Regulation.” Consultative Group to Assist the Poor (CGAP), Focus Note No.38. Mas, Ignacio, and Kabir Kumar. 2008. “Banking on Mobiles: Why, How, for Whom?” Consultative Group to Assist the Poor (CGAP), Focus Note No.48. Mbiti, Isaac and David N. Weil. 2015. “Mobile Banking: The Impact of M-PESA in Kenya.” In Sebastian Edwards, Simon Johnson, and David N. Weil, Editors, African Successes, Volume 3: Modernization and Development, Chicago: University of Chicago Press. Siedek, Hannah. 2008. “Extending Financial Services with Banking Agents.” Consultative Group to Assist the Poor (CGAP) Brief. CGAP: Washington, D.C. 32 World Bank. 2014. Global Financial Development Report 2014: Financial Inclusion. Washingotn, DC: World Bank. Yang, Dean, and HwaJung Choi. 2007. “Are Remittances Insurance? Evidence from Rainfall Shocks in the Philippines.” World Bank Economic Review 21(2): 219-248. 33 Table 1. Cash In Number of Transactions Regression P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) 3.630*** 3.630 3.773*** 3.773* 3.747*** 3.747* 5.504*** 5.504** 5.271*** 5.271** 5.412*** 5.412** 5.298*** 5.298** 5.411*** 5.411** Owner age (0.583) (2.146) (0.605) (2.167) (0.617) (2.185) (0.733) (2.381) (0.663) (2.534) (0.661) (2.521) (0.629) (2.300) (0.639) (2.311) characteristics Owner 9.520 9.520 5.625 5.625 -11.81 -11.81 15.87 15.87 4.819 4.819 1.700 1.700 3.754 3.754 2.993 2.993 Primary Education (11.59) (53.80) (12.05) (55.74) (14.08) (68.60) (15.16) (77.35) (13.13) (61.80) (13.74) (64.64) (13.83) (65.52) (13.96) (65.98) Owner Owner Gender -32.60*** -32.60 -31.10*** -31.10 -23.87* -23.87 -13.41 -13.41 -1.808 -1.808 0.223 0.223 -0.701 -0.701 0.298 0.298 (F=0, M=1) (11.46) (35.06) (11.69) (34.81) (12.29) (38.07) (13.44) (42.56) (12.60) (40.51) (12.63) (40.53) (12.37) (40.20) (12.48) (40.49) characteristcs Business -2.040*** -2.040 -2.234*** -2.234 -1.774** -1.774 0.408 0.408 0.0731 0.0731 -0.566 -0.566 -0.366 -0.366 Business Age (0.704) (2.051) (0.671) (1.896) (0.748) (1.645) (0.689) (2.050) (0.709) (2.238) (0.705) (2.212) (0.701) (2.135) -0.287*** -0.287** -0.246*** -0.246** -0.221*** -0.221*** -0.155*** -0.155 -0.153*** -0.153 -0.151*** -0.151 -0.150*** -0.150 Stock Value (0.0344) (0.118) (0.0338) (0.0968) (0.0325) (0.0751) (0.0561) (0.116) (0.0515) (0.108) (0.0506) (0.0957) (0.0503) (0.0950) Commercial -76.91*** -76.91* -96.68*** -96.68** -59.02*** -59.02 -79.90*** -79.90* low density (11.75) (40.67) (12.22) (40.49) (11.41) (35.10) (15.94) (40.75) -96.81*** -96.81** -63.97*** -63.97 -111.8*** -111.8 -102.1*** -102.1 Income high (14.59) (34.99) (18.29) (56.37) (25.37) (82.12) (26.86) (87.97) 3.223*** 3.223*** 2.997*** 2.997*** 3.033*** 3.033*** 3.030*** 3.030*** Municipality characteristics Population (0.377) (0.857) (0.355) (0.859) (0.360) (0.933) (0.359) (0.926) 2.232*** 2.232** 2.262*** 2.262** 1.383*** 1.383 1.538*** 1.538* Density (0.335) (1.020) (0.336) (0.957) (0.331) (0.863) (0.361) (0.847) Financial -48.52*** -48.52 -45.77** -45.77 density high (17.43) (63.62) (23.29) (62.02) 45.62* 45.62 46.26** 46.26 Banks (23.48) (71.07) (23.44) (71.38) 89.64*** 89.64** 51.37** 51.37 Micro finance (16.48) (41.09) (24.59) (61.76) Mobile 23.45* 23.45 -8.875 -8.875 banking (14.09) (36.55) (17.42) (42.91) Mean od 202.6 202.6 206.5 206.5 206.5 206.5 206.5 206.5 210.4 210.4 210.4 210.4 210.4 210.4 210.4 210.4 cashintx (259.7) (259.7) (264.1) (264.1) (264.1) (264.1) (264.1) (264.1) (272.3) (272.3) (272.3) (272.3) (272.3) (272.3) (272.3) (272.3) Observations 2,296 2,296 2,202 2,202 2,202 2,202 2,202 2,202 1,921 1,921 1,921 1,921 1,921 1,921 1,921 1,921 R-squared 0.023 0.023 0.028 0.028 0.053 0.053 0.205 0.205 0.136 0.136 0.140 0.140 0.143 0.143 0.144 0.144 District fixed effects No No No No Yes Yes No No No No No No No No No No Municipality fixed effects No No No No No No Yes Yes No No No No No No No No 34 Table 2. Cash Out Number of Transactions Regression P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) 2.437*** 2.437 2.530*** 2.530 2.363*** 2.363 3.670*** 3.670 3.127*** 3.127 3.136*** 3.136 2.975*** 2.975 3.059*** 3.059 Owner characteristics Owner age (0.506) (1.749) (0.526) (1.835) (0.540) (1.957) (0.639) (2.233) (0.581) (2.223) (0.577) (2.201) (0.528) (2.031) (0.539) (2.067) -9.384 -9.384 -11.89 -11.89 -18.72 -18.72 9.251 9.251 -6.437 -6.437 -6.639 -6.639 -4.453 -4.453 -5.012 -5.012 Owner Primary (9.692) (43.24) (10.10) (45.12) (11.89) (54.72) (12.95) (63.66) (11.34) (52.03) (11.84) (54.29) (11.84) (54.74) (11.98) (55.15) Education -10.46 -10.46 -10.67 -10.67 -7.529 -7.529 -5.153 -5.153 18.33* 18.33 18.46* 18.46 17.39* 17.39 18.12* 18.12 Owner Gender (9.441) (31.52) (9.697) (32.40) (10.31) (36.15) (11.63) (40.35) (10.60) (35.96) (10.64) (36.19) (10.32) (35.79) (10.40) (36.22) (F=0, M=1) -1.041* -1.041 -1.576*** -1.576 -2.044*** -2.044 -0.750 -0.750 -0.772 -0.772 -1.238** -1.238 -1.091** -1.091 characteristcs Business Business Age (0.577) (1.715) (0.549) (1.693) (0.620) (1.649) (0.511) (1.640) (0.534) (1.805) (0.535) (1.779) (0.537) (1.724) -0.137*** -0.137 -0.113*** -0.113* -0.128*** -0.128** -0.0550* -0.0550 -0.0549* -0.0549 -0.0443 -0.0443 -0.0438 -0.0438 Stock Value (0.0286) (0.0900) (0.0255) (0.0646) (0.0268) (0.0591) (0.0298) (0.0720) (0.0296) (0.0720) (0.0270) (0.0609) (0.0268) (0.0608) -77.25*** -77.25* -78.53*** -78.53** -56.37*** -56.37* -71.70*** -71.70* Commercial low density (9.351) (38.41) (9.352) (36.07) (8.195) (28.14) (13.07) (37.23) -81.18*** -81.18** -79.06*** -79.06 -134.3*** -134.3** -127.2*** -127.2* Income high (13.61) (32.97) (15.25) (48.31) (21.19) (62.52) (21.99) (66.83) 2.352*** 2.352*** 2.338*** 2.338*** 2.431*** 2.431** 2.429*** 2.429** Municipality characteristics Population (0.376) (0.832) (0.355) (0.819) (0.359) (0.876) (0.359) (0.870) 0.959*** 0.959 0.961*** 0.961 0.210 0.210 0.324 0.324 Density (0.230) (0.701) (0.229) (0.682) (0.220) (0.583) (0.245) (0.555) -3.142 -3.142 -33.61* -33.61 Financial density high (13.56) (50.54) (19.65) (42.98) 68.87*** 68.87 69.34*** 69.34 Banks (18.83) (52.95) (18.78) (53.13) 49.93*** 49.93* 21.82 21.82 Micro finance (11.97) (24.47) (21.34) (44.04) -1.670 -1.670 -25.41* -25.41 Mobile banking (11.57) (29.11) (13.06) (32.94) 131.0 131.0 133.2 133.2 133.2 133.2 133.2 133.2 131.8 131.8 131.8 131.8 131.8 131.8 131.8 131.8 Mean of cashouttx (215.9) (215.9) (219.9) (219.9) (219.9) (219.9) (219.9) (219.9) (223.0) (223.0) (223.0) (223.0) (223.0) (223.0) (223.0) (223.0) Observations 2,296 2,296 2,202 2,202 2,202 2,202 2,202 2,202 1,921 1,921 1,921 1,921 1,921 1,921 1,921 1,921 R-squared 0.013 0.013 0.015 0.015 0.037 0.037 0.166 0.166 0.102 0.102 0.102 0.102 0.110 0.110 0.110 0.110 District fixed effects No No No No Yes Yes No No No No No No No No No No Municipality fixed effects No No No No No No Yes Yes No No No No No No No No 35 Table 3. Cash In Volume (Log) Regression P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) 0.0175*** 0.0175* 0.0194*** 0.0194* 0.0194*** 0.0194** 0.0272*** 0.0272*** 0.0232*** 0.0232** 0.0247*** 0.0247** 0.0256*** 0.0256** 0.0262*** 0.0262*** Owner characteristics Owner age (0.00430) (0.00966) (0.00430) (0.00962) (0.00466) (0.00810) (0.00533) (0.00855) (0.00429) (0.0107) (0.00428) (0.0104) (0.00454) (0.00927) (0.00456) (0.00919) 0.0879 0.0879 0.0497 0.0497 0.0488 0.0488 0.224* 0.224 0.0633 0.0633 0.0309 0.0309 0.0398 0.0398 0.0355 0.0355 Owner Primary (0.0920) (0.219) (0.0964) (0.223) (0.102) (0.274) (0.121) (0.315) (0.102) (0.239) (0.103) (0.244) (0.104) (0.247) (0.104) (0.248) Education -0.283*** -0.283 -0.294*** -0.294 -0.300*** -0.300 -0.259** -0.259 -0.148 -0.148 -0.127 -0.127 -0.124 -0.124 -0.118 -0.118 Owner Gender (0.0973) (0.193) (0.0981) (0.199) (0.0969) (0.200) (0.104) (0.214) (0.106) (0.180) (0.106) (0.185) (0.107) (0.183) (0.106) (0.182) (F=0, M=1) -0.0173* -0.0173 -0.0225** -0.0225 -0.0242** -0.0242 -0.00157 -0.00157 -0.00494 -0.00494 -0.00880 -0.00880 -0.00768 -0.00768 characteristcs Business Business Age (0.00903) (0.0154) (0.00922) (0.0155) (0.0101) (0.0159) (0.00859) (0.0180) (0.00876) (0.0184) (0.00916) (0.0179) (0.00925) (0.0179) -0.002*** -0.00201* -0.002*** -0.0021** -0.002*** -0.00202 -0.002*** -0.0020** -0.002*** -0.00203* -0.002*** -0.0021** -0.002*** -0.0021** Stock Value (0.000628) (0.00110) (0.000604) (0.000995) (0.000690) (0.00124) (0.000726) (0.000957) (0.000734) (0.00101) (0.000734) (0.000976) (0.000735) (0.000976) -0.194 -0.194 -0.397*** -0.397 -0.140 -0.140 -0.256 -0.256 Commercial low density (0.132) (0.227) (0.138) (0.249) (0.143) (0.225) (0.180) (0.282) -0.410*** -0.410* -0.0750 -0.0750 -0.238 -0.238 -0.184 -0.184 Income high (0.115) (0.220) (0.150) (0.349) (0.205) (0.531) (0.217) (0.574) 0.0123*** 0.0123*** 0.00992** 0.00992* 0.00975** 0.00975* 0.00973** 0.00973* Municipality characteristics (0.00193) (0.00434) * (0.00198) (0.00496) * (0.00200) (0.00499) * (0.00200) (0.00500) Population 0.00610** 0.00610 0.00637** 0.00637 0.000494 0.000494 0.00135 0.00135 Density (0.00283) (0.00769) (0.00282) (0.00692) (0.00361) (0.00738) (0.00380) (0.00749) -0.495*** -0.495 -0.255 -0.255 Financial density high (0.138) (0.382) (0.251) (0.368) -0.000671 -0.000671 0.00317 0.00317 Banks (0.221) (0.476) (0.221) (0.480) 0.700*** 0.700** 0.487* 0.487 Micro finance (0.191) (0.308) (0.287) (0.436) 0.129 0.129 -0.0502 -0.0502 Mobile banking (0.151) (0.287) (0.230) (0.346) Mean of 9.274 9.274 9.285 9.285 9.285 9.285 9.285 9.285 9.263 9.263 9.263 9.263 9.263 9.263 9.263 9.263 cashin_volum (2.047) (2.047) (2.053) (2.053) (2.053) (2.053) (2.053) (2.053) (2.052) (2.052) (2.052) (2.052) (2.052) (2.052) (2.052) (2.052) e (ln) Observations 2,281 2,281 2,187 2,187 2,187 2,187 2,187 2,187 1,913 1,913 1,913 1,913 1,913 1,913 1,913 1,913 R-squared 0.012 0.012 0.019 0.019 0.028 0.028 0.116 0.116 0.042 0.042 0.049 0.049 0.054 0.054 0.054 0.054 District fixed effects No No No No Yes Yes No No No No No No No No No No Municipality fixed effects No No No No No No Yes Yes No No No No No No No No 36 Table 4. Cash Out Volume (Log) Regression P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) Robust SE Cluster(M) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) -0.00206 -0.00206 -0.000641 -0.000641 -0.000438 -0.000438 0.0171*** 0.0171* 0.00153 0.00153 0.00186 0.00186 0.00659 0.00659 0.00756 0.00756 Owner characteristics Owner age (0.00482) (0.0120) (0.00493) (0.0125) (0.00515) (0.0109) (0.00551) (0.00846) (0.00497) (0.0122) (0.00495) (0.0121) (0.00505) (0.0100) (0.00506) (0.00994) -0.0387 -0.0387 -0.0759 -0.0759 -0.0134 -0.0134 0.446*** 0.446 0.0192 0.0192 0.0117 0.0117 0.0304 0.0304 0.0240 0.0240 Owner Primary (0.0996) (0.281) (0.104) (0.291) (0.111) (0.316) (0.122) (0.324) (0.108) (0.314) (0.110) (0.325) (0.109) (0.309) (0.109) (0.310) Education -0.0944 -0.0944 -0.118 -0.118 -0.154 -0.154 -0.0711 -0.0711 0.0727 0.0727 0.0776 0.0776 0.105 0.105 0.113 0.113 Owner Gender (0.110) (0.279) (0.111) (0.284) (0.109) (0.283) (0.109) (0.258) (0.116) (0.235) (0.116) (0.238) (0.115) (0.222) (0.115) (0.222) (F=0, M=1) -0.00146 -0.00146 -0.00974 -0.00974 -0.0205** -0.0205 0.00158 0.00158 0.000766 0.000766 -0.000903 -0.000903 0.000764 0.000764 characteristcs Business Business Age (0.00850) (0.0138) (0.00878) (0.0158) (0.00965) (0.0152) (0.00896) (0.0206) (0.00913) (0.0211) (0.00956) (0.0198) (0.00969) (0.0197) -0.00147* -0.00147 - -0.00173* - -0.00232* -0.00127 -0.00127 -0.00126 -0.00126 -0.00147 -0.00147 -0.00146 -0.00146 (0.000863) (0.00125) 0.00173** (0.000841) (0.00101) 0.00232** (0.000926) (0.00128) (0.00117) (0.00153) (0.00117) (0.00154) (0.00117) (0.00140) (0.00117) (0.00140) Stock Value -0.366** -0.366 -0.413*** -0.413 -0.278* -0.278 -0.458** -0.458 Commercial low density (0.142) (0.486) (0.153) (0.571) (0.145) (0.383) (0.201) (0.522) -0.504*** -0.504 -0.426*** -0.426 -0.780*** -0.780 -0.697*** -0.697 Income high (0.130) (0.436) (0.162) (0.482) (0.226) (0.602) (0.241) (0.631) 0.0158*** 0.0158* 0.0153*** 0.0153 0.0162*** 0.0162 0.0162*** 0.0162 Municipality characteristics Population (0.00238) (0.00904) (0.00250) (0.0103) (0.00250) (0.0104) (0.00250) (0.0104) 0.00733** 0.00733 0.00739** 0.00739 -0.00332 -0.00332 -0.00200 -0.00200 Density (0.00314) (0.0110) (0.00313) (0.0109) (0.00381) (0.00991) (0.00403) (0.00968) -0.116 -0.116 -0.390 -0.390 Financial density high (0.151) (0.472) (0.279) (0.421) 0.0845 0.0845 0.0880 0.0880 Banks (0.228) (0.398) (0.228) (0.404) 0.610*** 0.610* 0.282 0.282 Micro finance (0.188) (0.313) (0.312) (0.468) -0.669*** -0.669 -0.946*** -0.946* Mobile banking (0.165) (0.436) (0.257) (0.517) mean of 7.994 7.994 7.992 7.992 7.992 7.992 7.992 7.992 7.922 7.922 7.922 7.922 7.922 7.922 7.922 7.922 cashout_volum (2.213) (2.213) (2.229) (2.229) (2.229) (2.229) (2.229) (2.229) (2.222) (2.222) (2.222) (2.222) (2.222) (2.222) (2.222) (2.222) e (ln) Observations 2,274 2,274 2,180 2,180 2,180 2,180 2,180 2,180 1,907 1,907 1,907 1,907 1,907 1,907 1,907 1,907 R-squared 0.001 0.001 0.003 0.003 0.016 0.016 0.161 0.161 0.040 0.040 0.040 0.040 0.066 0.066 0.067 0.067 District fixed effects No No No No Yes Yes No No No No No No No No No No Municipality fixed effects No No No No No No Yes Yes No No No No No No No No 37 Table 5. Municipality Density Model (simple) - Cash In Transactions P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Robust SE errors Cluster (municipality) Low density High density Low density High density Owner 4.691*** 4.179*** 4.691 4.179 Characteristics Owner Age (0.910) (0.865) (2.968) (3.312) Owner Primary -74.56*** 82.15*** -74.56 82.15 Education (18.44) (22.32) (82.10) (111.8) Owner Gender -36.02* -1.061 -36.02 -1.061 (F=0, M=1) (18.46) (16.82) (68.75) (39.10) Business -0.532 -0.100 -0.532 -0.100 Characteristics Business Age (1.009) (1.011) (2.505) (2.754) -0.248*** -1.007*** -0.248 -1.007* Stock Value (0.0398) (0.181) (0.129) (0.485) Municipality -122.7*** -130.7*** -122.7 -130.7** Characteristics Commercially Low (25.72) (17.74) (121.0) (52.78) -108.3*** -145.4*** -108.3 -145.4** Income High (26.25) (13.85) (129.2) (59.60) Mean of dependent 186.1 233.2 186.1 233.2 variable (258.4) (270.3) (258.4) (270.3) Observations 927 994 927 994 R-squared 0.081 0.082 0.081 0.082 Table 6. Municipality Density Model (simple) - Log of Cash In Volume P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Robust SE errors Cluster (municipality) Low density High density Low density High density Owner 0.0213*** 0.0225*** 0.0213** 0.0225 Characteristics Owner Age (0.00591) (0.00639) (0.00830) (0.0197) Owner Primary -0.133 0.312* -0.133 0.312 Education (0.128) (0.175) (0.253) (0.510) Owner Gender -0.245* -0.171 -0.245 -0.171 (F=0, M=1) (0.141) (0.158) (0.347) (0.204) Business 0.0195* -0.0217* 0.0195 -0.0217 Characteristics Business Age (0.0117) (0.0131) (0.0268) (0.0262) -0.00296*** -0.000745 -0.00296** -0.000745 Stock Value (0.000781) (0.00180) (0.000836) (0.00333) Municipality -0.142 -0.419** -0.142 -0.419 Characteristics Commercially Low (0.187) (0.188) (0.455) (0.407) -0.300** -0.794* -0.300 -0.794* Income High (0.139) (0.436) (0.507) (0.383) Mean of dependent 9.247 9.336 9.247 9.336 variable (1.977) (2.150) (1.977) (2.150) Observations 922 991 922 991 R-squared 0.031 0.028 0.031 0.028 38 Table 7. Progressive OLS with Monitoring variables P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Cash in Transactions Cash out Transactions Cash in Volume (Log) Cash out Volume (Log) Cluster(M Robust Cluster(M Robust Cluster(M Robust Cluster(M Cluster(M Cluster(M Cluster(M Cluster(M Robust SE ) SE ) SE ) SE ) Robust SE ) Robust SE ) Robust SE ) Robust SE ) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) 5.636*** 5.636 5.574*** 5.574 4.096*** 4.096 3.985*** 3.985 0.0109* 0.0109 0.0127** 0.0127 0.00241 0.00241 0.00517 0.00517 Owner age (1.116) (3.525) (1.091) (3.467) (1.032) (3.458) (0.995) (3.449) (0.00622) (0.0131) (0.00634) (0.0111) (0.00674) (0.0148) (0.00676) (0.0126) -57.77*** -57.77 -55.48** -55.48 -56.4*** -56.40 -55.1*** -55.09 -0.146 -0.146 -0.130 -0.130 -0.0899 -0.0899 -0.0695 -0.0695 characteristics Owner Primary (21.64) (68.29) (21.61) (67.68) (19.87) (67.05) (19.80) (66.40) (0.130) (0.196) (0.131) (0.200) (0.135) (0.250) (0.135) (0.241) Education Owner 12.73 12.73 12.22 12.22 36.21** 36.21 35.15** 35.15 -0.277** -0.277* -0.248* -0.248* -0.0291 -0.0291 0.0158 0.0158 Owner Gender(F=0, M=1) (18.52) (51.10) (17.84) (50.04) (16.12) (50.78) (15.43) (49.72) (0.129) (0.151) (0.128) (0.140) (0.140) (0.229) (0.138) (0.214) -0.503 -0.503 -0.799 -0.799 -2.12*** -2.124 -2.34*** -2.337 -0.00701 -0.00701 -0.00867 -0.00867 -0.0101 -0.0101 -0.0119 -0.0119 characteristcs Business Business Age (0.915) (2.016) (0.920) (2.076) (0.737) (1.780) (0.742) (1.809) (0.00945) (0.0162) (0.00978) (0.0162) (0.00998) (0.0182) (0.0104) (0.0179) -0.145* -0.145 -0.156* -0.156 -0.0544 -0.0544 -0.0533 -0.0533 -0.00109 -0.00109 -0.00126 -0.00126 -0.00107 -0.00107 -0.00125 -0.00125 Stock Value (0.0873) (0.120) (0.0918) (0.126) (0.0631) (0.0865) (0.0623) (0.0846) (0.00137) (0.00137) (0.00133) (0.00124) (0.00201) (0.00194) (0.00200) (0.00178) -87.03*** -87.03** -50.61** -50.61 -80.3*** -80.32* -58.7*** -58.73 0.248 0.248 0.444** 0.444* -0.0956 -0.0956 0.0995 0.0995 Commercial low density (21.34) (39.82) (22.32) (43.96) (16.39) (39.12) (17.02) (37.08) (0.209) (0.205) (0.225) (0.242) (0.216) (0.383) (0.219) (0.340) -43.16* -43.16 -34.69 -34.69 -73.3*** -73.31 -79.1*** -79.06 0.306 0.306 0.402 0.402 -0.0999 -0.0999 -0.105 -0.105 Income high (25.82) (53.00) (29.31) (66.67) (22.44) (51.85) (25.25) (60.51) (0.195) (0.408) (0.249) (0.544) (0.213) (0.590) (0.262) (0.675) 3.733*** 3.733*** 3.629*** 3.629*** 2.996*** 2.996** 2.964*** 2.964** 0.0100** 0.0100** 0.00964** 0.00964* 0.00979** 0.00979 0.0101*** 0.0101 Municipality characteristics (0.561) (0.943) (0.552) (0.985) (0.576) (1.061) (0.565) (1.085) * (0.00253) (0.00426) * (0.00253) * (0.00374) * (0.00316) (0.00905) (0.00317) (0.00900) Population 1.016** 1.016 0.907** 0.907 0.127 0.127 0.0248 0.0248 -0.00232 -0.00232 -0.00545 -0.00545 -0.00734* -0.00734 -0.013*** -0.0131 Density (0.431) (0.995) (0.418) (1.094) (0.328) (0.808) (0.316) (0.955) (0.00353) (0.00752) (0.00406) (0.00690) (0.00403) (0.0113) (0.00424) (0.00898) -82.25*** -82.25 -37.49** -37.49 -0.500*** -0.500 -0.430** -0.430 Financial density high (22.37) (58.18) (17.70) (46.78) (0.172) (0.448) (0.187) (0.548) -26.95 -26.95 3.996 3.996 -0.345 -0.345 -0.293 -0.293 Banks (27.46) (57.94) (22.43) (48.82) (0.282) (0.607) (0.266) (0.598) 71.95*** 71.95 38.41** 38.41 0.559** 0.559 0.637*** 0.637 Micro finance (23.52) (55.29) (18.59) (46.01) (0.232) (0.419) (0.218) (0.450) 63.08** 63.08 36.18* 36.18 0.0217 0.0217 -0.296 -0.296 Mobile banking (24.58) (41.25) (20.51) (35.63) (0.220) (0.356) (0.231) (0.489) 20.27*** 20.27*** 20.69*** 20.69*** 17.00*** 17.00*** 17.17*** 17.17*** 0.135*** 0.135*** 0.136*** 0.136*** 0.141*** 0.141*** 0.141*** 0.141*** Liquidity Total (2.801) (5.669) (2.829) (5.774) (2.365) (5.727) (2.397) (5.808) (0.0211) (0.0186) (0.0210) (0.0196) (0.0214) (0.0249) (0.0212) (0.0267) Monitoring indicators 28.84*** 28.84* 28.46*** 28.46** 9.976 9.976 9.946 9.946 0.230*** 0.230** 0.225*** 0.225** 0.154* 0.154 0.149* 0.149 Client Service Total (8.203) (13.87) (8.149) (13.38) (6.820) (9.768) (6.765) (9.363) (0.0788) (0.0948) (0.0789) (0.0954) (0.0791) (0.0985) (0.0781) (0.0982) 37.52*** 37.52*** 36.86*** 36.86*** 32.10*** 32.10*** 31.84*** 31.84*** 0.514*** 0.514*** 0.505*** 0.505*** 0.580*** 0.580*** 0.568*** 0.568*** Branding Total (5.695) (7.675) (5.757) (8.122) (4.413) (7.172) (4.514) (7.627) (0.0716) (0.0977) (0.0722) (0.0977) (0.0679) (0.106) (0.0672) (0.103) 242.3 242.3 242.3 242.3 162.3 162.3 162.3 162.3 9.482 9.482 9.482 9.482 8.381 8.381 8.381 8.381 Mean of dependent variable (297.3) (297.3) (297.3) (297.3) (258.3) (258.3) (258.3) (258.3) (1.965) (1.965) (1.965) (1.965) (2.081) (2.081) (2.081) (2.081) Observations 886 886 886 886 886 886 886 886 884 884 884 884 884 884 884 884 R-squared 0.276 0.276 0.278 0.278 0.233 0.233 0.233 0.233 0.235 0.235 0.248 0.248 0.229 0.229 0.252 0.252 39 Table 8. Agent Fixed effects- Cash In Transactions P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Robust SE Cluster (M) Robust SE Cluster (M) Liquidity Total (e-float + cash 6.617*** 6.617*** float) (1.662) (2.045) 0.0224 0.0224 E-float (2.577) (3.449) 15.47*** 15.47*** Cash float (3.579) (3.679) Branding Total (visibility + 22.40*** 22.40*** FINCA images +FINCA sign) (5.358) (7.823) 24.70 24.70 Agent visibility (17.38) (15.79) -6.805 -6.805 FINCA Images (9.284) (7.535) 41.32*** 41.32*** FINCA Sign (9.063) (11.68) -3.581 -3.581 0.745 0.745 Client Service Total (5.948) (8.841) (5.754) (7.981) 233.0 233.0 233.1 233.1 mean of dependent variable (289.8) (289.8) (290.2) (290.2) Observations 964 964 960 960 R-squared 0.063 0.063 0.095 0.095 Number of agentid 160 160 160 160 Table 9. Agent Fixed effects- Log of Cash In Volume P-values are in brackets. *, **, *** represent statistical significance at the 10%, 5%, and 1% level respectively. All models are estimated via two methods: robust standard errors OLS and clustering errors by municipality OLS Robust SE Cluster (M) Robust SE Cluster (M) Liquidity Total (e-float + 0.0237 0.0237 cash float) (0.0200) (0.0190) -0.0241 -0.0241 E-float (0.0281) (0.0250) 0.0971*** 0.0971** Cash float (0.0314) (0.0374) Branding Total (visibility + 0.482*** 0.482*** FINCA images +FINCA sign) (0.0738) (0.104) 0.355 0.355 Agent visibility (0.247) (0.270) 0.0141 0.0141 FINCA Images (0.129) (0.117) 0.882*** 0.882*** FINCA Sign (0.164) (0.217) -0.0155 -0.0155 0.0153 0.0153 Client Service Total (0.0721) (0.0901) (0.0698) (0.0857) Mean of dependent 9.423 9.423 9.427 9.427 variable (1.995) (1.995) (1.989) (1.989) Observations 964 964 960 960 R-squared 0.063 0.063 0.095 0.095 Number of agentid 160 160 160 160 40 Appendix A. Municipality level statistics municipality surface municipality population business municipality (sq km) (10,000s) number of agents Bandalungwa 6.82 20.2341 4 Bumbu 5.5 32.9234 2 Kalamu 6.64 31.5342 5 Kasa-Vubu 5.04 15.732 8 Makala 5.6 25.3844 3 Ngiri Ngiri 3.4 17.4843 3 Selembao 23.18 33.5581 2 Barumbu 4.72 15.0319 3 Gombe 29.33 3.2373 8 Kintambo 2.72 10.6772 5 Lingwala 2.88 9.4635 4 Mont Ngafula 358.92 26.1004 8 Ngaliema 244.3 68.3135 20 Kisenso 16.6 38.6151 1 Lemba 23.7 34.9838 15 Limete 67.7 37.5726 14 Matete 4.88 26.8781 8 Ngaba 4 18.065 2 Kimbasenke 237.78 94.6372 13 Maluku 7948.8 6.745 4 Masina 69.73 48.5167 11 N'Djili 11.4 44.2138 12 N'Sele 898.79 14.0929 7 Lubumbashi 747 178.6397 15 Kolwezi 213 45.3147 4 Matadi 110 30.6053 7 Tshofa 1 Soyo 1 41 Appendix B. Variable Descriptive Statistics Variables Definition N mean min max sd variance kurtosis skewness p1 p25 p50 p99 owner's last degree (primary, last_degree highschool or university) 2315 0.796112 0 2 0.591559 0.349942 2.595018 0.089628 0 0 1 2 dummy variable, owner's gender owner_gender (0=female, 1=male) 2324 0.745267 0 1 0.435805 0.189926 2.267477 -1.12582 0 0 1 1 business_number_ number of employees an agent's employees business has 2265 4.326269 0 111 12.40981 154.0035 67.66187 7.917733 1 1 2 111 number of days per week the agent's business_daysperweek location is open for business 2282 6.640228 6 7 0.480039 0.230437 1.341481 -0.58436 6 6 7 7 number of hours per day the agent's hrsopenperday location is open for business 2273 13.09793 8 24 2.587159 6.693389 6.332575 0.914328 8 11.3 13 24 indicates whether the business is business_registered formally registred or not 2313 1 1 1 0 0 . . 1 1 1 1 net profit of business, reported net_profit weekly 1582 1491.597 15 39141 3375.348 1.14E+07 79.82377 7.583774 15 175 560 12000 account_balance current FINCA account balance 1799 1028.397 3 5000 743.1064 552207.1 14.18041 3.026752 10 500 1000 5000 daily_transactions daily transactions amount (USD) 2245 1483.303 35 120000 6932.114 4.81E+07 266.8851 15.70515 60 250 550 10000 stock_value Current inventory value USD 2230 20294.58 110 1000000 64190.54 4.12E+09 149.1209 11.149 1000 4200 7500 200000 min_pos_balance Minimum POS balance 2261 0.813799 0 2 0.633368 0.401155 2.388218 0.174242 0 0 1 2 owner_age business owner's age 2296 43.66202 24 70 10.05934 101.1903 2.790635 0.551494 27 36 42 69 number of years the agent's business business_age has been open 2324 13.71601 1 2010 102.0555 10415.32 378.2991 19.33814 1 4 6 79 municipality_surface municipality surface, in sq km 2032 416.2711 2.72 7948.8 1477.783 2183842 24.61068 4.815137 2.72 5.6 29.33 7948.8 municipality_population municipality population, 2004 data 2032 368660.8 32373 946372 241783.1 5.85E+10 3.162925 0.921581 32373 174843 335581 946372 municipality density, as derived from municipality_density population/surface 1948 18980.62 8 59860 18915.61 3.58E+08 1.994131 0.695122 8 2796 6957 59860 number of administrative divisions municipality_divisions inside the municipality 2032 15.65748 6 30 5.9721 35.66598 3.271088 0.633162 6 13 15 30 level of urban development urban_level (urban/peri-urban/rural) 2032 0.222933 0 1 0.416316 0.173319 2.772542 1.331368 0 0 0 1 indicates level of income in a income_level municipality (high, medium ,low) 2032 1.050689 0 2 0.767951 0.589748 1.700499 -0.08635 0 0 1 2 indicates level of commercial activity comercially_developed in a municipality (high, medium ,low) 1958 0.52094 0 2 0.720047 0.518468 2.6115 1.006969 0 0 0 2 indicates level of density of financial financial_alternatives_ alternatives in a municipality (high, density medium ,low) 1766 0.650623 0 2 0.841365 0.707895 1.806339 0.73067 0 0 0 2 42 indicates the form of financial services financial_alternatives_ predominant in one municipality types (banks, microfinance, mobile banking) 1676 0.738664 0 2 0.832021 0.69226 1.638885 0.519457 0 0 0 2 dummy variable, indicates is business owner is a graduate of primary primary_educ education only 2324 0.296472 0 1 0.4568 0.208666 1.794411 0.891297 0 0 0 1 dummy variable, indicates is business owner is a graduate of secondary secondary_educ education only 2324 0.606282 0 1 0.488679 0.238807 1.189287 -0.43507 0 0 1 1 dummy variable, indicates is business owner is a graduate of tertiary tertiary_educ education 2324 0.093374 0 1 0.291018 0.084691 8.812667 2.795115 0 0 0 1 indicates that a municipality's level of development can be qualified as urban urban, dummy variable 2324 0.679432 0 1 0.466795 0.217898 1.591281 -0.76895 0 0 1 1 indicates that a municipality's level of development can be qualified as peri- peri_urban urban, dummy variable 2324 0.194923 0 1 0.396226 0.156995 3.372359 1.540247 0 0 0 1 indicates high level of income in a income_high municipality, dummy variable 2324 0.236661 0 1 0.425124 0.18073 2.535488 1.239148 0 0 0 1 indicates medium level of income in a income_med municipality, dummy variable 2324 0.356713 0 1 0.479132 0.229568 1.357893 0.598241 0 0 0 1 indicates low level of income in a income_low municipality, dummy variable 2324 0.280981 0 1 0.449575 0.202118 1.949743 0.974547 0 0 0 1 indicates high level of commercial activity in a municipality, dummy comercially_high variable 2324 0.516781 0 1 0.499826 0.249826 1.004511 -0.06716 0 0 1 1 indicates medium level of commercial activity in a municipality, dummy comercially_med variable 2324 0.212565 0 1 0.40921 0.167453 2.974399 1.405133 0 0 0 1 indicates low level of commercial activity in a municipality, dummy comercially_low variable 2324 0.113167 0 1 0.316865 0.100403 6.96411 2.442153 0 0 0 1 indicates high density of financial alternatives in a municipality, dummy financial_density_high variable 2324 0.447935 0 1 0.497389 0.247396 1.043848 0.2094 0 0 0 1 indicates medium density of financial alternatives in a municipality, dummy financial_density_med variable 2324 0.129518 0 1 0.335845 0.112792 5.869719 2.206744 0 0 0 1 indicates low density of financial alternatives in a municipality, dummy financial_density_low variable 2324 0.182444 0 1 0.386293 0.149222 3.70429 1.644473 0 0 0 1 indicates banks as the form of financial services predominant in one banks municipality, dummy variable 2324 0.368331 0 1 0.482456 0.232763 1.29806 0.545948 0 0 0 1 43 indicates microfinance as the form of financial services predominant in one microfinance muncipality, dummy variable 2324 0.172978 0 1 0.378309 0.143118 3.990252 1.729234 0 0 0 1 indicates mobile banking as the form of financial services predominant in mobile_banking one muncipality, dummy variable 2324 0.179862 0 1 0.384155 0.147575 3.779116 1.667068 0 0 0 1 dummy variable for the district of Funa Funa 2324 0.16136 0 1 0.367942 0.135381 4.38974 1.841125 0 0 0 1 dummy variable for the district of Lukunga Lukunga 2324 0.259036 0 1 0.4382 0.192019 2.210059 1.100027 0 0 0 1 dummy variable for the district of Mont_Amba Mont 2324 0.222031 0 1 0.415701 0.172808 2.789274 1.337638 0 0 0 1 dummy variable for the district of Tshangu Tshangu 2324 0.233219 0 1 0.422971 0.178905 2.591976 1.261735 0 0 0 1 Other_Cities dummy variable for other cities 2324 0.124355 0 1 0.330057 0.108937 6.183537 2.276738 0 0 0 1 dummy variable, indicates whether an Commerce agent's registered business is in retail 2324 0.763769 0 1 0.424857 0.180503 2.542447 -1.24195 0 1 1 1 dummy variable, indicates whether an agent's registered business is in Services serivices 2324 0.21988 0 1 0.414254 0.171606 2.829798 1.3527 0 0 0 1 number of cash in transactions, cashintx monthly 2324 201.7818 0 2468 258.6658 66907.98 19.14416 3.335248 1 51 117 1304 volume of cash in transactions, cashinvolume monthly 2324 38138.16 0 2432694 119137.8 1.42E+10 223.0474 13.3504 1 4882.5 14297 296294.7 number of cash out transactions, cashouttx monthly 2324 130.8451 0 2421 215.0843 46261.25 35.334 4.799897 0 23 65 1086 volume of cash out transactions, cashoutvolume monthly 2324 15070.5 0 1404392 52359.43 2.74E+09 367.0128 16.69034 0 965 3741.865 155904 number of transfer transactions, transfertx monthly 2324 1.468158 0 85 4.999554 24.99554 98.05663 8.430969 0 0 0 25 volume of transfer transactions, transfervolume monthly 2324 521.076 0 101225 3337.408 1.11E+07 396.6339 16.27127 0 0 0 13620 totaltx total number of transactions, monthly 2324 334.0951 1 4899 462.6272 214023.9 27.0213 4.062847 2 82 189 2346 totalvolume total volume of transactions, monthly 2324 53729.69 0 3837106 165359.5 2.73E+10 256.5746 14.0736 5 6851.5 18917.75 450320.1 E-float monitoring score, monthly average 964 3.004991 0 5.25 1.919025 3.682659 1.785648 -0.28777 0 1.4465 2.63 5.25 Cash float monitoring score, monthly average 964 3.101465 0 5.69 1.932798 3.735709 1.794103 -0.3257 0 1.753333 2.63 5.25 Liquidity Total monitoring score, monthly average 964 6.10159 0 10.5 3.123703 9.757518 2.219792 -0.31219 0 3.94 6.126667 10.5 Connectivity/Service Availability monitoring score, monthly average 964 1.840497 0 3 0.368283 0.135633 7.911087 -2.26064 0.5 2 2 2 44 State of shop monitoring score, monthly average 964 1.723811 0 2 0.45339 0.205563 3.491746 -1.29915 0.5 1.366667 2 2 Operator availability monitoring score, monthly average 960 1.66386 0 3.833333 0.50318 0.25319 3.751734 -1.09144 0 1 2 2 Client Service Total monitoring score, monthly average 964 5.220825 0 6 0.894344 0.799851 5.533716 -1.3346 2.75 5 5.333333 6 Tenue du Registre des transactions monitoring score, monthly average 964 2.244285 0 2.5 0.636023 0.404526 8.680481 -2.59047 0 2.5 2.5 2.5 Soins Terminal POS monitoring score, monthly average 964 2.346178 0 2.5 0.410989 0.168912 13.09179 -3.06288 0.625 2.5 2.5 2.5 Suivi du solde monitoring score, monthly average 964 2.062023 0 4.791667 0.724041 0.524236 4.494072 -1.46622 0 1.666667 2.5 2.5 Performance total monitoring score, monthly average 964 6.647085 0 7.5 1.238731 1.534456 8.037264 -1.96928 2.5 6.25 7.5 7.5 Agent Visibility monitoring score, monthly average 964 1.871963 0 2 0.413219 0.17075 14.53503 -3.47651 0 2 2 2 FINCA images monitoring score, monthly average 960 1.673878 0 2 0.550039 0.302543 4.929473 -1.65918 0 1.333333 2 2 FINCA sign monitoring score, monthly average 964 1.718525 0 4 0.631125 0.398319 5.740557 -1.98233 0 2 2 2 Branding total monitoring score, monthly average 964 5.267342 0 16.06417 1.275662 1.627313 12.22836 -1.3267 0 5 6 6 Total score monitoring score, monthly average 961 23.26204 0 30 4.288883 18.39452 4.308764 -0.88681 11 20.875 23.76 30 number of quarters in existence as quart FINCA agents 2324 4.267857 0.25 11 2.309173 5.332279 3.007149 0.829085 1 2.5 3.75 11 45