The World Bank Economic Review, 36(3), 2022, 734–756 https://doi.org10.1093/wber/lhac007 Article Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Mobile Money and Economic Activity: Evidence from Kenya Raissa Fabregas and Tite Yokossi Abstract This paper estimates the impact of access to mobile money services on local economic activity. The analysis combines data from the early expansion of the mobile agent network in Kenya with a local-level measure of economic performance proxied by the intensity of nighttime lights. Leveraging the variation in areas that gained access to mobile money services at different times and the high resolution of the data, the paper shows that access to mobile money services increased local economic activity and that these effects increased over time. The positive effects are more pronounced for areas that were initially more affluent, urban, and better connected to infrastructure. These results suggest that mobile money can complement, rather than just substitute for, other alternatives that enable people to connect, trade, and allocate investments within their networks. JEL classification: G29, O16, O17, O47 Keywords: mobile money, night lights, financial inclusion, Kenya 1. Introduction The rapid expansion of digital technologies worldwide has enabled millions of people in developing and emerging markets to access a range of financial services for the first time. Mobile money services—which allow people to store and transfer money through a mobile phone—have played a central role in driv- ing this financial inclusion, with over 866 million registered accounts in 90 countries (GSMA 2019). By reducing transaction costs, improving transparency, providing a saving technology, and allowing individ- uals to share risk, mobile money has been heralded as a leading financial innovation with the potential to improve efficiency and spur economic growth. This paper estimates the effects of access to Kenya’s M-PESA on local economic activity in the years following its launch. M-PESA, one of the most successful mobile money services globally, enables indi- viduals to transfer money through short messaging services (SMS). Existing literature has documented positive effects of mobile money in mitigating risk (Jack, Ray, and Suri 2013; Jack and Suri 2014; Raissa Fabregas (corresponding author) is an assistant professor at the University of Texas at Austin, Austin, USA; her email address is rfabregas@utexas.edu. Tite Yokossi is a data scientist at QuantCo, Cologne, Germany; his email address is tite.yokossi@gmail.com. Tite acknowledges financial assistance from the MIT Department of Economics, where he was affiliated at the time of this study. The authors thank Abhijit Banerjee, Stacy Carlson, Esther Duflo, John Firth, Gabriel Kreindler, Ernest Liu, Matthew Lowe, Benjamin Marx, Benjamin Olken, Tavneet Suri, MIT Development Lunch participants, and the African School of Economics SIER conference participants for comments and discussions. Special thanks to Tavneet Suri for helping with access to the mobile money network expansion data. Felipe Lima provided excellent research assistance. A supplementary online appendix is available with this article at The World Bank Economic Review website. © The Author(s) 2022. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com The World Bank Economic Review 735 Riley 2018), increasing household consumption (Suri and Jack 2016; Munyegera and Matsumoto 2016), and improving business outcomes (Aggarwal, Brailovskaya, and Robinson 2020). Yet, to judge the over- all impacts of mobile money, one would want to consider effects on aggregate economic activity, which would account for these various mechanisms and the potential spillovers that could attenuate or am- Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 plify the direct impacts to users (Aron and Muellbauer 2019; Bateman, Duvendack, and Loubere 2019). This article asks whether the new possibilities generated by M-PESA translate into a tangible increase in aggregate local economic activity and, if so, which areas benefited the most. Following recent papers measuring economic growth in Africa, local economic activity is proxied using light density derived from nighttime satellite imagery. In a variety of contexts, light density at night has been shown to reflect economic activity at the national level (Chen and Nordhaus 2011; Henderson, Storeygard, and Weil 2012; Bundervoet, Maiyo, and Sanghi 2015) and at subnational lev- els (Michalopoulos and Papaioannou 2013; Storeygard 2016). In Kenya, hardly any other measure of local economic activity exists. Therefore, the paper uses the grid underlying light density images to divide Kenya into pixel cells. A pixel cell represents a geographic area slightly under 1 km2 , and it is the primary unit of analysis in this study. One important advantage of using nighttime light data is that it provides a highly disaggregated measure of economic performance on an annual basis (Donaldson and Storeygard 2016).1 To confirm that the night-light measure is adequate for this context, the article also shows that light density at night is highly correlated with household wealth as measured by the Kenya Demographic and Health Survey (DHS). The empirical strategy used in this paper exploits the variation in local access to M-PESA over time. After its launch in 2007, M-PESA was adopted by about 70 percent of all households in Kenya in just four years.2 Essential to this fast adoption was the growth of a network of agents, small business outlets that provide cash-in and cash-out services. Agents exchange cash for an electronic balance sent by SMS from one mobile money account to another. In 2012 there were over 49,000 mobile money agents in Kenya (CCK 2014), a staggering number considering that the country had about 1,000 bank branches at the time.3 To measure access to mobile money services, the analysis relies on the ever-decreasing distance between a local area (pixel cell) and the agent network. To ensure that the study is based on a comparison of cells with plausibly similar underlying growth trends, the analysis only focuses on the variation in the timing of initial access to M-PESA during the first five rollout years. All specifications in the analysis include cell fixed effects to address any potential endogeneity arising from agents targeting specific areas because of fixed characteristics, e.g., more affluent, urban, or densely populated. Sublocation-year fixed effects are also included in the regressions. A sublocation is the lowest administrative level in the country (on average 9 km × 9 km), and they generally consist of 2 to 3 villages in rural areas or one large neighborhood within a city (Mbiti and Weil 2015).4 These controls effectively remove any bias stemming from knowledge of the future growth trajectory of a sublocation, which might have induced more agents to operate in these areas. While cells that gained access to agents in earlier waves were richer and better connected to infrastructure, there is limited evidence of systematic differences in their pre-trends. However, to further address concerns about the validity of the results and account for 1 Additionally, since it is automatically collected, it is less prone to recall and attrition bias relative to a household survey (Pinkovskiy 2013). This is a particular advantage for the study of mobile services since survey attrition in urban areas has limited the study of these populations. For instance, for the seminal studies of Jack and Suri (2014) and Suri and Jack (2016), the sample of people in Nairobi was dropped due to high attrition. Even in the non-Nairobi sample, overall attrition was around 35 percent. 2 SIM card registration data from the telecommunications firm Safaricom. 3 Kenya National Bureau of Statistics, Economic Survey 2012. 4 Sublocations come after provinces, districts, divisions, and locations. Until the 2009 national census, the country was divided into 8 provinces, 70 districts, 506 divisions, 2456 locations, and 6631 sublocations. These subdivisions are used throughout the article. 736 Fabregas and Yokossi secular trends in outcomes, the study also shows regression results that include cell-specific linear time trends. Therefore, the results’ causal interpretation relies on the assumption that any deviations of cell outcomes from their trends in the post-expansion period are due to the existence of M-PESA agents. Overall, the study shows that access to mobile money led to an increase in local economic activity. The Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 intensity of nighttime lights increased about 0.04 standard deviations (s.d.) one year after receiving access to M-PESA and up to 0.39 s.d. six years later. However, there is substantial heterogeneity in impacts: if cells that accessed agents in the first rollout wave are excluded from the analysis, the estimated effects are in the 0.02–0.10 s.d. range. The analysis also finds consistent evidence that the effects are stronger in areas closer to cities and with better access to infrastructure, such as roads and banks. While there might be several unobserved factors that bolster the effects of M-PESA in better-off areas, this result is consistent with the idea that mobile money affects overall economic growth rather than just redistributing income from wealthier to impoverished regions. This paper make a number of contributions. First, there has been little empirical investigation of the aggregate economic effects of mobile money.5 Closest to this study, Suri and Jack (2016) investigate the effects of M-PESA on household consumption and poverty. Using household distance to agents and con- trolling for location fixed effects, the authors find positive consumption growth, particularly for female- headed households. The authors suggest that this result is likely driven by shifts in occupational choice, with moves away from farming and into business and retail. They estimate that poverty in Kenya declined by 2 percentage points as a result of access to mobile money. The benefits of mobile money from microstudies could be understated, for instance, because of pos- itive externalities related to network growth and an increased formalization of the economy (Aron and Muellbauer 2019). In contrast, these studies could be ignoring adverse effects in equilibrium. As an ex- ample, skeptics have argued that if mobile money facilitates the move from subsistence agriculture into microenterprises, other small local businesses might be negatively affected since this would just redis- tribute overall demand among an increasing number of suppliers (Bateman, Duvendack, and Loubere 2019). This paper shows that at least at the local level, there are net positive economic effects from mo- bile money access. Second, rural households have generally been expected to benefit more from mobile money since they are less likely to have access to banks and other formal financial services. Several studies have exclusively focused on the impacts of mobile money for these populations (Munyegera and Matsumoto 2016; Wieser et al. 2019).6 Yet these effects might miss broader gains that occur in regions with complementary factors that can amplify the direct impacts of mobile services. In fact, the analysis in this paper suggests that the effects of M-PESA are larger in areas that were better connected to banks, roads, cell towers, and electricity transmission lines at baseline. Access to mobile money services appears to complement, rather 5 Existing evidence has explored the role of mobile money on specific outcomes. For instance, on sharing risks and smoothing shocks (Jack, Ray, and Suri 2013; Jack and Suri 2014; Blumenstock, Eagle, and Fafchamps 2016; Riley 2018), increasing household consumption (Munyegera and Matsumoto 2016; Suri and Jack 2016), increasing savings (Demombynes and Thegeya 2012; Mbiti and Weil 2015), improving food security (Murendo and Wollni 2016; Wieser et al. 2019), and affecting some business outcomes (Aggarwal, Brailovskaya, and Robinson 2020). Several descriptive studies have analyzed the adoption and use of mobile money services in developing countries and particularly in Kenya (Mas 2009; Mas and Morawczynski 2009; Morawczynski and Pickens 2009; Jack and Suri 2011; Higgins, Kendall, and Lyon 2012). A second complementary literature has been concerned with the effects of mobile money on macroe- conomic policy, particularly monetary policy (Adam and Walker 2015; Ndirangu and Nyamongo 2015; Mawejje and Lakuma 2017). Finally, other authors have investigated the role that digital services play when used to deliver public programs (Aker et al. 2016). 6 A broader literature has found positive impacts from bringing financial services into rural and remote areas, for instance, Burgess and Pande (2005), Bruhn and Love (2009), and Dupas and Robinson (2013). The World Bank Economic Review 737 than just substitute for, other alternatives that enable people to connect, trade, and allocate investments within their networks. Finally, this paper makes a modest contribution to the long-standing debate about the importance of financial innovations in spurring economic growth (Robinson 1979; Lucas 1988; Miller 1998; Levine Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 2005). Existing literature has focused on the development of banks and securities markets, both of which are institutions that are hardly accessible to the majority of people in developing countries. This article shows that a financial innovation generated outside the sphere of classical financial institutions can affect economic activity. The rest of the article proceeds as follows. The following section describes the context of M-PESA use and adoption in Kenya. The construction of the data set is discussed in the Data section. The empirical strategy is explained in the Sample and Empirical Strategy section. The Results section presents the results and robustness checks. Unpacking the Effects of Mobile Banking shows how the results vary based on pre-M-PESA characteristics. The final section concludes. 2. Background of Mobile Money in Kenya In March 2007, Safaricom, the largest mobile network operator in Kenya, launched a mobile-phone- based payment and money transfer service called M-PESA. The service allows users to exchange cash for electronic money (e-money) and make deposits into a mobile money account stored in their cellphones. Customers can also send electronic balances to other users, including sellers of goods and services, via SMS, and convert e-money back into cash. Since 2007, M-PESA has spread quickly in Kenya and has become one of the most successful mobile- phone-based financial services in the world. The average number of daily registrations reached 10,000 in December 2007 and increased in the subsequent years. In just four years, M-PESA reached the 14 million accounts mark, representing about 70 percent of households in the country. Key to the rapid adoption of M-PESA was its network of agents, small business outlets that provide cash-in and cash-out services to users. Agents typically run small businesses—airtime distribution stores, cellphone retail shops, grocery stores, gas stations, etc.—and receive a commission for both M-PESA deposits and M-PESA withdrawals. M-PESA agents purchase e-money balances from Safaricom or from customers and hold them on their own mobile phones. They also maintain cash on their premises to fulfill user withdrawal orders. Since customers need to be physically present with an agent to convert their e-money into cash or cash into e-money, the agent network maintained and operated by Safaricom is crucial in facilitating access to mobile money services. The closer agents are to households, the easier it is for customers to purchase e-balances or redeem cash from e-money sent to them. The number of M-PESA agents grew from 4,000 in early 2008 to 33,000 in 2010 and to over 88,000 in 2013 (CCK 2014). The tremendous expansion of the agent network brought cash-in and cash-out services within walking distance of a large fraction of households, providing them with the technology to save, pay, transfer, and receive money. Much smaller changes in standard cash-in and cash-out service points (bank branches, ATM network) occurred over the same period. By 2013, M-PESA users had made a total of $6.82 billion worth of deposits and transferred between each other a cumulative amount of $6.36 billion. Following the early success of M-PESA, other mobile money services were established in Kenya. However, M-PESA enjoyed a near-monopoly of mobile money transfers in Kenya over the study period, 2007–2013, be it in cash transfers, active users, or the number of agents.7 7 Furthermore, M-PESA dominates its money transfer predecessors in virtually all dimensions. Before the advent of mobile money services, the majority of households sent money through friends or bus drivers. Users found M-PESA to be faster, cheaper, safer, and more reliable than the alternatives (Jack and Suri 2011). 738 Fabregas and Yokossi 3. Data This paper combines two main sources of data to explore the effects of mobile money on economic growth: light density at night, as a proxy for economic activity, and the expansion of the mobile money agent network, to measure access to M-PESA services. Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 3.1. Night Lights To explore changes in local economic activity, one would like to use time-varying subnational measures of GDP, income, or wealth. Unfortunately, such measures do not exist in Kenya, even at the district or provincial level. Instead, light density at night, as measured by satellites of the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS), is used.8 The light data are available at a very high resolution every year from 1992 to 2013. Each global digital image recorded by DMSP satellites is made up of pixels, each of which represents a cell of 30*30 arcsecond2 , an area of about 0.9 km2 in Kenya. Attached to each pixel is a value that represents the annual average stable intensity of light emitted at night by the corresponding area. The data are already processed to remove interference from clouds, forest fires, aurorae, and other factors that are unrelated to normal human economic activity. In some years, two satellites recorded global nighttime light densi- ties. As a result, 34 satellite-years are available for the 22-year period. In years for which more than one digital image is available, the pixel values are averaged to produce an annual data set. The study period runs from 2000 to 2013 and includes roughly the same number of years before and after the launch of M-PESA.9 The analysis in this article is carried out at the pixel level with a grid that replicates that of light density images. Kenya is covered by more than 600 thousand pixels. Each pixel cell is matched to Kenya’s official administrative subdivisions. On average a district is made up of 90 thousand pixels, and a sublocation comprises 101 pixels. A map of lit and unlit pixels shows that most of Kenya has no visible light at night (fig. S1.1 in the supplementary online appendix). However, most of the country is also sparsely populated. Unsurprisingly, night light is visible in areas where most of the population and economic activity are located. Strong correlations between light density and economic performance have been documented at the country level (Chen and Nordhaus 2011; Henderson, Storeygard, and Weil 2012), including Kenya (Bundervoet, Maiyo, and Sanghi 2015). More recent work has started to document their validity at the subnational level (Michalopoulos and Papaioannou 2013; Storeygard 2016; Bruederle and Hodler 2018).10 Using an approach similar to that of Michalopoulos and Papaioannou (2013), this study first analyses how light density correlates with household wealth and electrification in Kenya using microlevel data from the Demographic and Health Survey (DHS). The analysis employs the composite wealth index, a measure of households’ cumulative living standard based on observables such as asset ownership (radio, 8 The US Air Force Weather Agency collects the DMSP data. The image and data processing comes from the NOAA’s National Geophysical Data Center. 9 Gibson (2021) describes issues with measurement errors in the DMSP data that are mean reverting. Top coding and image blurring (as described in Abrahams, Oram, and Lozano-Gracia (2018)) are particular concerns. These issues suggest that regression coefficients will be attenuated and the estimates might understate true treatment effects (Gibson, Olivia, and Boe-Gibson 2020). 10 Michalopoulos and Papaioannou (2013) provide a cross-validation of light density and subnational economic perfor- mance in African countries. Using microlevel data from the Demographic and Health Surveys, the authors show that nighttime light density has a strong positive correlation with household wealth, education, and electrification in four large African countries representative of different parts of the continent: Nigeria (West Africa), Tanzania (East Africa), Zimbabwe (South Africa), and the Democratic Republic of Congo (Central Africa). Bundervoet, Maiyo, and Sanghi (2015) estimate an annual GDP growth of 4 percent between 2000 and 2012 using night-light data for Kenya, whereas the national-accounts-based growth rate was 4.2 percent. The World Bank Economic Review 739 TV, bicycles, etc.), materials used in housing construction, water access, and sanitation.11 Averaging the wealth index over households in each DHS enumeration area and comparing it with the average lumi- nosity of pixels 7 or 10 km around the centroid of the enumeration area results in a strongly positive correlation (respectively 75 percent and 76 percent). Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 The linear fit between the average household wealth index of DHS clusters in Kenya and the average light density in a given radius around DHS clusters, along with the corresponding scatterplots, is graphed in fig. 1. The radius around DHS cluster centroids is 7 km in (a) and 10 km in (b). Both graphs are produced by combining data from the two rounds of DHS surveys in Kenya for which corresponding nighttime light data and GIS coordinates of DHS cluster centroids are available (2003 and 2008–09). Both graphs depict a strong relationship between the light density and the household wealth index. Although this cross- validation is mostly cross-sectional, while the results in this article are based on variations of pixel values over time, the strong correlation between light density and the index of household wealth in Kenya’s DHS clusters is reassuring. The correlations between a household electricity dummy variable with the average light density of pixels in a 7 or 10 km radius around DHS clusters’ centroids are respectively 72 percent and 73 percent.12 3.2. The M-PESA Agent Network Expansion Data on financial service providers’ locations in Kenya was collected by the Financial Sector Deepening (FSD) Kenya and its partners (FSD 2016). The analysis uses this data to map the M-PESA agent network in the country each year during the 2007–2013 period. The number of mobile money agents in Kenya has been undergoing rapid expansion (fig. 2). In 2013 there were more than 88,000 mobile money agents spread out across the country. This is close to 3 times the agent count in 2010 and 20 times the agent count in early 2008. However, most districts covered in 2013 were already covered in 2007. The change over time was not new districts gaining access to mobile money services but rather, within each district, more areas coming within walking distance of an agent. Since consumers convert cash into M-PESA e-money and vice versa thanks to agents who provide cash-in and cash-out services, agent proximity to a given household is a good measure of the ease of access to M-PESA services.13 The empirical analysis in this study uses the variation in the timing of cell access to an M-PESA agent. The distance from each pixel cell to the closest agent is computed for each year. These distances fall over time as the agent density increases. For each pixel cell, year 0 of access is defined, if applicable, as the year in which the distance to the closest agent falls below 5 km. The spatial distribution of mobile money agents depended on both the agent demand for licenses and Safaricom’s rollout decisions. Jack and Suri (2014) report that Safaricom rolled out to the most populated areas of the country first, but that they did not know the exact geographic locations of their agents, making it unlikely that they could have targeted particular areas within districts or cities.14 Agent applications, on the other hand, had to be screened at Safaricom central offices, creating queues in the approval process (Jack and Suri 2014) and likely introducing idiosyncratic variation in the precise timing when a local area would gain access. However, more agents from better-off areas might have sought licenses first. This 11 See http://www.dhsprogram.com/topics/wealth-index/Wealth-Index-Construction.cfm for more details on the construc- tion of this index by DHS country teams. The DHS data are downloaded from http://www.dhsprogram.com/Data/. 12 These correlations are strong despite the displacement of DHS enumeration areas’ centroids of up to 2 km for urban areas and up to 5 km for rural areas, with a further randomly selected 1 percent of rural clusters (every hundredth cluster) displaced a distance up to 10 km. This noise is willingly added to the precise geo-coordinates of DHS clusters’ centroids for privacy reasons. 13 See Jack and Suri (2014) for household-level evidence of the positive relationship between the agent density and the use of M-PESA. 14 One exception may have been Nairobi. Results are robust to its exclusion from the sample. 740 Fabregas and Yokossi Figure 1. Nighttime Light Density and Household Wealth—Kenya DHS Cluster 1.5 (a) Household Wealth Index of a DHS Cluster Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 −.5 0 .5 −1 1 0 20 40 60 Light Density in a 7km Radius of a DHS Cluster (b) 1.5 Household Wealth Index of a DHS Cluster −.5 0 .5 −1 1 0 10 20 30 40 50 Light Density in a 10km Radius of a DHS Cluster Source: Authors’ analysis using the Demographic and Health Survey (DHS) data for Kenya. Nighttime light data come from the Defense Meteorological Satellite Program’s Operational Linescan System (DMS). Note: A linear fit between the average household wealth index of DHS clusters in Kenya and the average light density in a given radius around DHS clusters along with the corresponding scatterplots are plotted. The radius around DHS clusters centroids is 7 km on the left and 10 km on the right. Both graphs are produced using the two rounds of DHS surveys in Kenya for which relevant nighttime light data and GIS coordinates of DHS cluster centroids are available (2003 and 2008–09). paper refers to cells that received agents in each rollout year as waves.15 Supplementary online appendix table S1.1 shows summary statistics by wave for the first five waves in the analysis sample (described in the next section). The data indicate that cell access to an M-PESA agent is not random: cells in earlier waves had higher pre-expansion night lights and better access to infrastructure. 15 In other words, wave 1 correspond to those cells classified as having M-PESA access in 2008, wave 2 for those in 2009, wave 3 for those in 2010, and so forth. The World Bank Economic Review 741 Figure 2. Mobile Money Agent Network Expansion Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Source: M-PESA agent data for 2007–2013 come from the FinAccess geospatial mapping collected by Financial Sector Deepening (FSD) Kenya and partners. Note: Agent locations are shown in years 2007 (left), 2010 (center), and 2013 (right). M-PESA was quickly rolled out to cover most populated areas. The expansion of the network came from a gradual increase in the density of agents in covered areas rather than from the coverage of new broad areas. 3.3. Additional Variables Administrative divisions’ data come from the 2009 master shapefile of the Kenya National Bureau of Statistics. These files contain sublocation-level population counts and densities. Baseline data on bank branches’ locations at the launch of M-PESA in 2007 is extracted from the financial service provider lo- cations data (FSD 2016). The paper also uses publicly available GIS information on the baseline locations of roads, cell towers, major electricity transmission lines, and city locations.16 16 Road data are from Hijmans et al. (2016) and city locations from Henderson, Storeygard, and Deichmann (2017). Cell tower data are accessible from OpenCell ID: http://opencellid.org. Location of electricity transmission 742 Fabregas and Yokossi Since the infrastructure variables are all proxies of urbanization and highly correlated, in the analysis they are combined into two alternative summary measures: the average distance (in km) to these locations and an index derived from a principal component analysis (pca). Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 4. Sample and Empirical Strategy This section describes the sample, empirical approach, and identification assumptions. 4.1. Analysis Sample A number of restrictions are made to the total number of pixel cells in Kenya. First, areas with densities below 4 ppl/km2 in 1999 are excluded (49 percent of the total). In these areas, light is unlikely to be a suitable proxy for economic activity because the data are too noisy (Min 2008) and including these areas might lead to spurious correlations (Cogneau and Dupraz 2014). Sublocations that have no pixel ever lit during the study period (2000–2013) are also removed. These areas are sparsely populated, with little economic activity, and are unlikely to be suitable for the analysis in this article because changes in their economic activity are unlikely to be captured by light. This leaves the sample with 31 percent of the remaining cells. Finally, to increase the likelihood that the study focuses on a valid comparison, the analysis is restricted to cells that gained access to a mobile money agent during the first five waves of the M-PESA rollout.17 This allows us to identify effects only from the variation created by the timing of access. Columns 1 and 2 in table 1 show summary statistics before imposing any restrictions (All cells). Columns 3 and 4 show descriptive statistics for the sample of cells that gained M-PESA access at any point during the study period, had population densities above 4 ppl/km2 , and were in sublocations with pixels lit (Gained M-PESA). This corresponds to 51,137 cells. Finally, columns 5 and 6 show summary statistics for the 44,105 cells employed in the main analysis, which corresponds to cells that gained an agent during one of the first five rollout waves (Analysis sample). As expected, relative to the universe of pixel cells in Kenya, the analysis focuses on cells that are richer, more urban, and closer to infrastructure. In the analysis sample, the average night-light density was 1.67 in 2007 and 3.62 in 2013. These num- bers are 0.12 and 0.26 for all cells in Kenya. The average distance to mobile money agents in the sample drops from 12.1 km in 2007 to 2.2 km in 2013. Average distances to banks, roads, and major cities are respectively 16, 3, and 23 km, and about half of the sample is under the electricity grid at baseline.18 However, as the standard deviations show, there is substantial variation in the baseline access to these various infrastructures. The pixel count weighted average of sublocation density is close to 393 ppl/km2 . An important caveat of the analysis, therefore, is that it is not nationally representative. It only applies to the experience of early access cells in selected areas. To the extent that these cells are more affluent and there might be other advantages from being an early mover, this might limit the generalizability of the study’s results. Figure 3 shows a plot of night-light averages over time and by wave, and table S1.1 summary statistics at baseline by wave. In the main analysis sample, 37 percent of the cells correspond to wave 1, 21 percent to wave 2, 19 percent to wave 3, 14 percent to wave 4, and 9 percent to wave 5. Earlier waves, and in particular wave 1, had higher levels of night lights before the M-PESA rollout. Earlier waves were also, on average, closer to banks, cities, and roads. The association in levels between earlier waves and affluence persists even when cells are compared within sublocations (table 2, columns 1–6). This will be problematic lines comes from the African Infrastructure Country Diagnostic database (AICD 2009): http://capacity4dev.ec. europa.eu/afretep/minisite/processed-gis-data. 17 Cells in later waves appear to have been growing at a faster rate relative to those in earlier waves during the pre-rollout periods. 18 A pixel cell is defined to be under-grid if it is within 10 km of major transmission lines. The World Bank Economic Review 743 Table 1. Summary Statistics All cells Gained M-PESA Analysis sample Mean Std. dev. Mean Std. dev. Mean Std. dev. (1) (2) (3) (4) (5) (6) Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Lights 2007 0.12 1.45 1.47 5.03 1.67 5.37 Lights 2013 0.26 2.27 3.22 7.54 3.62 8.00 Lit 2007 0.02 0.13 0.20 0.40 0.22 0.41 Lit 2013 0.03 0.16 0.30 0.46 0.34 0.47 Distance to agent 2007 66.36 48.48 13.32 20.91 12.12 20.22 Distance to agent 2013 30.37 24.98 2.38 1.31 2.23 1.27 Access to mobile money 2007 0.00 0.00 0.00 0.00 0.00 0.00 Access to mobile money 2013 0.14 0.35 0.94 0.24 1.00 0.00 Avg. distance to infrastructure 75.62 53.20 14.31 18.98 13.31 18.10 Distance to bank 79.91 58.39 17.46 22.13 16.09 21.73 Distance to road 9.90 9.79 3.34 2.83 3.23 2.78 Distance to cell tower 50.75 36.69 8.36 11.50 7.48 10.65 Distance to electricity tower 161.93 140.99 28.06 52.92 26.44 50.09 Under power grid 0.08 0.28 0.47 0.50 0.49 0.50 Distance to major city 77.58 46.40 24.01 24.28 23.00 23.94 Distance to Nairobi 365.87 170.57 207.76 132.20 200.04 129.09 Distance to top 5 city 273.66 168.68 88.28 81.99 84.43 78.62 Pixel sublocation density 49.80 335.73 354.11 1,149.93 393.45 1,232.73 Source: Authors’ calculations based on night-light data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMS), M-PESA agent and bank data from the FinAccess geospatial mapping, road data adapted from Hijmans et al. (2016), cell tower data from OpenCell ID, major electricity transmission lines from the African Infrastructure Country Diagnostic database, and city locations from Henderson, Storeygard, and Deichmann (2017). Note: Lights show the average night-light index in 2007. Lit is a dummy for non-zero pixel values. The access to mobile money dummy variable takes the value 1 for all years after year 0. Time-varying variables are measured before the advent of M-PESA: banks in 2007, roads in 2004, electricity transmission lines in 2007, and population density in 1999 (latest census before the launch of M-PESA). All distances are in km. Pixel sublocation density refers to the pixel-count-weighted average of sublocation population density (in ppl/km2 ). Under power grid is the proportion of the corresponding sample cells that are within 10 km of major transmission lines at the launch of M-PESA in 2007. All cells N = 674, 791, Gained M-PESA cells N = 51, 137, Analysis sample cells N = 44, 105. Table 2. Selection on Level vs. Trend Night lights ln (distance) to (Pre-rollout) Bank Major road Cell tower Power line City Std. mean Std. growth rate (1) (2) (3) (4) (5) (6) (7) Wave 1 −0.368*** −0.351*** −0.393*** −0.118*** −0.237*** 0.250*** 0.004 (0.048) (0.038) (0.034) (0.034) (0.027) (0.032) (0.066) Wave 2 −0.153*** −0.231*** −0.206*** −0.052 −0.052* 0.042** 0.048 (0.031) (0.037) (0.025) (0.037) (0.026) (0.021) (0.047) Wave 3 −0.036 −0.104** −0.049 −0.054 0.001 −0.003 −0.009 (0.037) (0.048) (0.036) (0.038) (0.032) (0.023) (0.056) Wave 4 −0.019 −0.083** −0.072** 0.003 0.011 0.018 0.074 (0.026) (0.037) (0.029) (0.035) (0.025) (0.025) (0.047) Subloc. FE Yes Yes Yes Yes Yes Yes Yes Observations 44,105 44,105 44,105 44,105 44,105 44,105 44,105 Source: Authors’ calculations based on night-light data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMS), M-PESA agent and bank data from the FinAccess geospatial mapping, road data adapted from Hijmans et al. (2016), cell tower data from OpenCell ID, major electricity transmission lines from the African Infrastructure Country Diagnostic database, and city locations from Henderson, Storeygard, and Deichmann (2017). Note: This table estimates the association between waves and the pre-rollout period log distance to banks, roads, cell towers, cities, mean nighttime lights, and the pre-period nighttime lights growth rate. Nighttime lights have been standardized. The omitted group is wave 5. Standard errors are clustered at the district level. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. 744 Fabregas and Yokossi Figure 3. Average Nighttime Lights by Wave over Time Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Source: Nighttime light data come from the Defense Meteorological Satellite Program’s Operational Linescan System (DMS). M-PESA rollout data come from the FinAccess geospatial mapping collected by Financial Sector Deepening (FSD) Kenya and partners. Note: The figure plots the yearly average of the (raw) nighttime lights for each M-PESA wave for the sample period. for the identification strategy only if these cells were trending differently, and if the empirical strategy does not correctly account for this. These issues are discussed in the following sections. 4.2. Basic Specification The unit of analysis is the pixel cell i; the time frequency t is annual. For ease of interpretation, the dependent variable is standardized to have a mean of 0 and standard deviation of 1 in the Gained M- PESA sample. The key independent variable is access to mobile money services, which is defined by the year in which a given pixel cell has its closest mobile money agent within walking distance (5 km). Different pixels gain access to mobile money services at different times, but the initial year t = 0 is uniformly defined by the year in which a cell gains access to the service for the first time. The main specification is a dynamic difference-in-differences model that takes the form std.Lit = αi + γt + βt Accessit + θ Xit + δi t + it , (1) t where std.Lit is the standardized light density at night of pixel i in year t; Accessit indicates a set of dummy variables that take the value 1 if a cell had received M-PESA access in year t, for years following year 0; α i and γ t represent respectively pixel fixed effects and year fixed effects in the 2000–2013 period; δ i t represents a linear time trend specific to pixel cell i; and Xit indicates time-varying factors that can be added as controls, such as sublocation-year fixed effects. The coefficients of interest are given by β t . They indicate the average residual difference in light density between pixels that had access to M-PESA and those that had not yet gained access, after accounting for pixel fixed effects, year fixed effects, sublocation-year fixed effects, and cell-specific linear trends. There are two reasons to prefer this dynamic specification relative to a single binary treatment model. First, this setup allows us to explore the dynamics of how night lights were affected over time, which might The World Bank Economic Review 745 be important as more people started to use M-PESA. Second, estimates with unit-specific trends can be significantly biased when the model does not fully take into consideration time-varying effects (Wolfers 2006).19 This issue is avoided by allowing the estimated effects to vary by year following the rollout. Throughout the analysis standard errors are conservatively clustered at the district level (three admin- Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 istrative levels above sublocations) to account for spatial autocorrelation. However, the results are robust to different clustering assumptions. 4.3. Identification Assumptions Equation (1) identifies the causal impact of mobile money on light density if the Accessit variables are independent of the error term, it , conditional on the pixel fixed effects, time fixed effects, linear trends, and time-varying controls. This identification assumption would be automatically satisfied if the spatial distribution of mobile money agents was exogenous from any other time-varying factor that systematically affects nighttime light density. How the empirical specification presented by this study compares to this strong exogeneity assumption is discussed next, along with a number of tests to support the underlying assumptions. As illustrated in table S1.1, agents opened first in richer cells. Including pixel fixed effects in the analysis would address the concern that the study merely captures fixed differences between different types of cells. The inclusion of sublocation-year fixed effects addresses biases that could stem from agents opening services based on the knowledge of the future trajectory of a sublocation.20 Absent pixel cell-specific trends, the identification assumption is that the underlying trends in the out- come variable would have been the same across the cells in different rollout waves. This could be violated, for instance, if agents filed more applications when they sensed that a particular area was growing within a sublocation. This concern is explored by looking for pre-trends in an event study. Consider year t = −1, the year preceding the year within which a given pixel cell came within walking distance of its closest mobile money agent for the first time. If indeed agent applicants could target growing areas within sublo- cations, one would expect to see, in the years leading to M-PESA access, a pre-trend indicating growth before year 0 of access, especially given that agent applicants had to wait for the completion of Safaricom’s approval process. This is estimated by running std.Lit = αi + γt + βt Relative Accessit + θ Xit + it . (2) t Here RelativeAccessit represents leads and lag years since period 0. Pre-trends are detected if the esti- mates of the coefficients β t are statistically significant in the years leading to access to the service and of comparable magnitude to the years following access. It would mean that the “treatment group,” which is 19 In essence, the pixel specific trends are likely to capture differences in pre-existing trends, but also differences in impact intensity following the treatment. This would lead to “overcontrolling” and result in likely biased estimates. The dynamic specification is also less likely to be affected by the biases introduced by staggered two-way fixed effects models as discussed by Goodman-Bacon (2021). 20 In this case, the results should be immune to a scenario in which a sublocation suddenly follows a different trajectory or receives a shock at the same time as some of its pixel cells gain access to M-PESA. Including sublocation-year fixed effects also helps address the concern that the results are driven by population (density) effects. The treatment period is short, so it is hard to imagine big population changes that hinge on the treatment. The inclusion of sublocation-year fixed effects also reduce concerns around businesses or households moving from treatment areas to comparison areas. Once they are included, moves from one sublocation to another or from one district to another leave the results unchanged. While short moves within a sublocation could affect the interpretation of the results, it is unlikely that a business would react fast enough and move within these small areas to be closer to agents, especially since new agents could have been closer in subsequent years. The fact that different pixels are treated at different times even within a sublocation also assuages the concerns about another intervention happening coincidentally with access to mobile money. 746 Fabregas and Yokossi Figure 4. Event Study .6 Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 .4 Lights .2 0 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 Year Source: Authors’ analysis based on night-light data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP) and M-PESA agent data from the FinAccess geospatial mapping. Note: For each pixel cell, year 0 indicates the year within which the cell’s distance to the closest M-PESA agent came below 5 km. The figure is based on the set of pixels for which a full year of access can be defined for the first five rollout waves. The dependent variable is the standardized light density at night. Pixel and year fixed effects are included, as well as sublocation-year fixed effects. Standard errors are conservatively clustered at the district level. not a fixed group here, is already experiencing a differential increase in light density before the advent of M-PESA. Estimating specification (2) yields the results presented in fig. 4. In the figure, the 95 percent confidence intervals are plotted for each year’s coefficient. Most coefficients on the years leading to access are not significantly different from zero, and they are smaller than the coefficients corresponding to the years following access to M-PESA, which are all positive and statistically significant. This is reassuring and provides support for the assumption that areas that had just gained access to mobile money services did not have different dynamics from areas that would gain access to the same services in the future. In addition, table 2 shows results from comparing early vs. later rollout cells using the pre-rollout mean level of night lights and the pre-rollout growth rate of night lights. While column 6 shows that wave 1, and to a lesser degree wave 2, cells have significantly higher levels of night lights at baseline relative to the last wave (a level difference), there are no significant differences in pre-rollout light growth rates (column 7). Finally, to also test for differential trends across cells, fig. 5 shows plots of coefficients from a specifi- cation showing differences in night lights by calendar year for waves 1 to 4 relative to wave 5.21 While there are some small deviations in early periods, there are no clear systematic pre-trends prior to the 21 Formally, an equation of the following form is estimated: std.Lit = αi + βnt (waven · cyeart ) + θ Xit + it , n t where waven is a dummy variable indicating the rollout wave n and cyeart is a vector of indicators for each calendar year between 2000 and 2013. The base year is 2007, the year prior to rollout. This approach follows that of Cook and Shah (2020). The World Bank Economic Review 747 Figure 5. Effects from M-PESA on Nighttime Lights by Year and Wave .4 Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Coefficient on Wave*Year −.2 0−.4 .2 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 Pre−Rollout Rollout Wave1 Wave2 Wave3 Wave4 Source: Authors’ analysis based on night-light data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP and M-PESA agent data from the FinAccess geospatial mapping. Note: The figure plots the coefficients from a regression equation of each wave dummy interacted with calendar year indicators for the period 2000–2013. The dependent variable is a standardized measure of night lights. Each coefficient picks up the difference between waves 1, 2, 3, and 4 relative to wave 5 on each year. Wave 1 corresponds to a rollout in 2008, wave 2 in 2009, wave 3 in 2010, wave 4 in 2011, and wave 5 in 2012. The regression controls for pixel and sublocation-year fixed effects. Standard errors are clustered at the district level. M-PESA rollout. Afterwards there is a difference in night lights, particularly for waves 1 and 2 relative to subsequent waves. Overall, this evidence does not suggest a violation of the parallel trend assumption. However, several pre-treatment periods further allow the analysis to relax the assumption of parallel trends by introducing cell-specific linear time trends. In that case, the main identification assumption is that any deviation on pixel outcomes from their trend after the rollout is due to the M-PESA agents.22 Therefore, even in the presence of different baseline levels and trends in night-light outcomes, causal effects can be identified as long as there are no unobserved cell-specific shocks affecting night lights that coincided with the agent expansion. 5. Results 5.1. Main Results Table 3 presents the estimates of the main specification, equation (1), for the analysis sample. Column 1 shows effects up to six years following the agent expansion and controls for year and pixel fixed effects. When controls for sublocation-year fixed effects are included (column 2), estimates are halved, but they remain statistically significant at the 1 percent level. Column 3 controls for the average cell-level growth rate in the nighttime lights for the pre-period 2000–2006. The results are practically identical to only 22 This is known as a “parallel growth” paths assumption rather than a “parallel trend” assumption (Mora and Reggio 2012). 748 Fabregas and Yokossi Table 3. Main Results Analysis sample Excluding wave 1 (1) (2) (3) (4) (5) Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Year 1 0.068*** 0.038*** 0.039*** 0.035*** 0.020** (0.017) (0.007) (0.007) (0.008) (0.009) Year 2 0.129*** 0.072*** 0.072*** 0.065*** 0.038*** (0.035) (0.013) (0.013) (0.014) (0.013) Year 3 0.294*** 0.160*** 0.160*** 0.152*** 0.064*** (0.065) (0.022) (0.022) (0.024) (0.018) Year 4 0.353*** 0.196*** 0.196*** 0.185*** 0.074*** (0.077) (0.026) (0.026) (0.029) (0.025) Year 5 0.517*** 0.303*** 0.303*** 0.275*** 0.097*** (0.111) (0.046) (0.046) (0.045) (0.031) Year 6 0.769*** 0.460*** 0.460*** 0.391*** (0.168) (0.071) (0.071) (0.070) Year FE Yes Yes Yes Yes Yes Pix FE Yes Yes Yes Yes Yes Subloc*time FE No Yes Yes Yes Yes Pre-period growth*time FE No No Yes No No Pix linear trend No No No Yes Yes Mean 0.04 0.04 0.04 0.04 −0.20 Pixels 44,105 44,105 44,105 44,105 27,842 Observations 617,470 617,470 617,470 617,470 389,788 Source: Authors’ analysis based on night-light data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP) and M-PESA agent data from the FinAccess geospatial mapping. Note: Dynamic difference-in-differences estimates are reported. The dependent variable is the standardized light density at night observed at the pixel-year level. For each pixel cell, year 0 indicates the year within which the cell’s distance to the closest M-PESA agent came below 5 km. Standard errors clustered at the district level are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. controlling for sublocation-year fixed effects. Column 4 introduces pixel-specific linear trends. Including linear trends reduces the point estimates slightly, but the magnitude and statistical significance remain very similar. This result was already expected, since there was not much evidence of pre-existing trends. Taking column 4 as the most conservative specification, we see that access to an M-PESA agent is associated with a night-lights increase of 0.04 standard deviations in the first year, and up to 0.39 standard deviations by year 6.23 Figure 5 suggested that wave 1 cells experienced much larger increases in lights in the post-period relative to other waves. These cells also had higher levels of night lights at baseline. To rule out concerns that there was something different about wave 1 cells that might violate the identification assumption, a specification excluding wave 1 cells and using the variation in agent access from waves 2 to 5 is also shown. The magnitudes are smaller but the effects remain statistically significant. They range from 0.02 standard deviations in year 1 to 0.10 standard deviations in year 5 (note that excluding wave 1 cells leaves the sample with only five years of post-rollout data).24 23 To benchmark the size of these effects, one could consider Henderson, Storeygard, and Weil (2012) who estimated the elasticity of GDP growth with respect to night-light growth to be 0.28 for low- and medium-income countries. However this relationship might not be representative of changes in economic activity within smaller regions. 24 Excluding wave 1 and wave 2 cells, and only using the variation from years 3 to 5 leads to very similar magnitudes to this specification (up to year 4), but these are less precisely estimated. The World Bank Economic Review 749 5.2. Robustness Checks This section discusses additional threats to the identification strategy and interpretation, and how are they tested. First, one might be concerned that the results are picking up the effects of an electrification drive in the country. Sublocation-year fixed effects are likely to address this concern already because sublocations Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 are small administrative units: it is likely that once a cell within a sublocation gained access to electricity, the entire sublocation would gain access too. However, a more direct way to check the robustness of the results to electrification changes is to look at areas that were close to major power transmission lines, and therefore already under-grid, before the launch of M-PESA. Restricting the sample to these locations leads to the results presented in columns 1 and 2 in supplementary online appendix table S1.2. The point estimates in column 2, which corresponds to the inclusion of linear trends, are statistically significant at the 1 percent level and if anything the magnitudes are larger than those of the main analysis sample. This supports the interpretation that the effects are not driven by electrification. To confirm that the effects do not simply reflect agents’ own light use, pixel cells whose centroid is closest to the agent (within 1 km) are removed. The magnitude of the estimates (table S1.2, columns 3 and 4) is qualitatively similar to the main results, indicating that the effect on night lights is not being driven by the light use of the agents themselves, as opposed to a broader increase in local activity. Jack and Suri (2014) report that Nairobi is the only area where Safaricom management might have actively granted agent licenses according to precise location. While this is unlikely to threaten the validity of the estimates, the results are robust to excluding Nairobi (table S1.2, columns 5 and 6). Differences-in-differences models with unit-specific time trends can be unreliable if the number of pe- riods before rollout is small (Wolfers 2006). Since several pre-expansion periods are available, this is unlikely to be an issue. However, extending the sample so far out might raise concerns. As an additional robustness check, the estimation sample is limited to two years before the rollout. These results are pre- sented in supplementary online appendix table S1.3. Column 2, which includes sublocation fixed effects, shows that the effects are very similar to those of table 3. Column 3 shows smaller magnitudes once linear trends are included, but this is likely misspecified since there are only a couple of periods from which to identify pre-existing trends. Column 4 shows robust effects for the sample that excludes wave 1 cells. Recent literature has pointed to issues with dynamic difference-in-differences designs. In particular, if there are heterogeneous treatment effects, some units might receive negative weights when their outcomes are aggregated (De Chaisemartin and d’Haultfoeuille 2020; Callaway and Sant’Anna 2021; Goodman- Bacon 2021; Borusyak, Jaravel, and Spiess 2021). To address this issue, results are replicated using the estimation method proposed by Sun and Abraham (2021), which addresses this potential bias in event studies in the presence of heterogeneous treatment effects. The pattern of effects is similar to that of the OLS specification, though the magnitudes of the simple specification without trends are larger and less precise, especially in later years (supplementary online appendix fig. S1.2). Robustness to different cutoffs to determine access to mobile money is also examined. The main es- timates use 5 km because it is a natural walking distance beyond which it is arguably inconvenient for potential customers to visit mobile money agents.25 However, effects are robust to variations in this 5 km cutoff (supplementary online appendix table S1.4). Nighttime lights can be defined in several ways, so alternate definitions of the dependent variable are explored. Supplementary online appendix table S1.5 shows results using the raw measure of night lights, the inverse hyperbolic sine (asinh), and the natural logarithm of lights+0.01 as the dependent variables.26 All the point estimates are positive, and for the most part, remain statistically significant. 25 It would take about 45 to 60 minutes for an average person to walk 5 km. This also sits in the middle range of the possible cutoff distances, since sublocation-year fixed effects are included and a sublocation covers on average a 9 km × 9 km area. 26 The inverse hyperbolic sine (asinh), asinh(x) = ln(x + 1 + x2 ), has a similar interpretation to that of the natural log- arithm function but it is defined at zero. 750 Fabregas and Yokossi The main specifications conservatively cluster at higher levels of geographical aggregation to address spatial correlation (Bester, Conley, and Hansen 2011; Cameron and Miller 2015). Supplementary online appendix table S1.6 shows main estimates with standard errors clustered at the district, sublocation, and pixel levels. These different cluster choices have minimal impact on inference. Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Finally, the analysis presents robustness tests to outliers. The distribution of nighttime lights is skewed. A substantial number of cells (about 49 percent in the sample) have value zero throughout the study period (see supplementary online appendix fig. S1.3 for a histogram).27 Moreover, the values of nighttime lights data are censored at 63. Supplementary online appendix table S1.7 shows how results vary depending on how outlier cells are treated. In columns 1 and 2, all cells in which the average pixel intensity for any year is 63 are removed. Results trimming the 5 percent tails are also shown (columns 3 and 4). Finally, columns 5 and 6 show results removing all cells that had zero lights during the entire study period. In all cases the estimates are positive and statistically significant. 6. Unpacking the Effects of Mobile Banking This study has documented that access to mobile money agents led to an increase in economic activity. A number of channels might drive this effect. For example, by reducing transaction costs, M-PESA has been shown to have a positive impact on the number of internal remittances (Morawczynski and Pickens 2009; Jack and Suri 2014). Remittances might stimulate the local demand and provide other members of the community with a source of credit. Households could use those increased remittances to start or grow businesses (Suri and Jack 2016). As a result, access to mobile money can improve the allocation of physical and human capital investments. Through easier payment arrangements, access to mobile money services also facilitates the trade of goods and services. For instance, firms can arrange to pay their suppliers by pushing a few buttons instead of traveling long distances. They can also be paid more easily by their customers. Households can use M- PESA to pay for a host of services from electricity bills to taxi fares.28 This can help households and firms save time, energy, and money and makes the business environment safer, as it reduces the need to carry large amounts of cash. By providing an easier and safer storage technology, M-PESA can also increase savings. The study’s approach does not allow us to pin down the precise mechanisms or the most significant driving force underlying the positive relationship between the use of mobile money and night lights. However, a natural way to gain additional insights is to understand how mobile money complements existing economic activity. Some insights are provided in this section. 6.1. Extensive and Intensive Margins To understand whether the results are driven mostly by pixels that were already lit or pixels that were never lit, results are decomposed into an extensive and an intensive margin. For the extensive margin, the focus is on the subset of the sample for which the average pixel value over the years prior to year t = 0 of access is zero. Results are shown with and without controlling for pixel linear trends (columns 1 and 2, table 4). The coefficients are smaller than the equivalent estimates for the whole sample but remain statistically significant. Columns 3 and 4 correspond to the subset of cells in the sample for 27 Note that in the sample restriction, sublocations that were never lit are dropped, but cells that were never lit within a sublocation that had positive night lights are not. 28 As a referee pointed out, a potential channel through which M-PESA might affect night lights is through the emergence of the pay-as-you-go (PAYG) solar energy model. M-PESA allows customers to pay for solar energy through small daily payments. This channel, however, is unlikely to drive the bulk of results, especially since the PAYG solar services only started to scale up after 2012 (Yadav, Heynen, and Palit 2019). In 2013, the biggest solar service provider in Kenya, M-KOPA, only had 15,000 clients (https://m-kopa.com/2013/05/22/m-kopa-surpasses-15000-customers/). The World Bank Economic Review 751 Table 4. Effects for Intensive and Extensive Margin Baseline unlit Baseline lit (1) (2) (3) (4) Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Year 1 0.010*** 0.008** 0.080*** 0.069*** (0.003) (0.003) (0.018) (0.017) Year 2 0.016*** 0.011** 0.147*** 0.126*** (0.004) (0.005) (0.037) (0.030) Year 3 0.040*** 0.034*** 0.279*** 0.249*** (0.008) (0.008) (0.055) (0.052) Year 4 0.057*** 0.046*** 0.313*** 0.271*** (0.008) (0.008) (0.070) (0.070) Year 5 0.083*** 0.068*** 0.475*** 0.415*** (0.015) (0.014) (0.105) (0.102) Year 6 0.119*** 0.093*** 0.705*** 0.585*** (0.018) (0.017) (0.155) (0.161) Year FE Yes Yes Yes Yes Pix FE Yes Yes Yes Yes Subloc*time FE Yes Yes Yes Yes Pix trend No Yes No Yes Mean −0.32 −0.32 0.60 0.60 Pixels 26,604 26,604 17,501 17,501 Observations 372,456 372,456 245,014 245,014 Source: Authors’ analysis based on night-light data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP) and M-PESA agent data from the FinAccess geospatial mapping. Note: Dynamic difference-in-differences estimates are reported. The dependent variable is standardized light density at night observed at the pixel-year level. For each pixel cell, year 0 indicates the year within which the cell’s distance to the closest M-PESA agent came below 5 km. Baseline unlit refers to the sample where the average pixel value over the years prior to year 0 of access is zero. Baseline lit refers to the sample that had a strictly positive average in the years prior to year 0. The study period runs from 2000 to 2013. Standard errors clustered at the district level are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. which the average pixel value over the years leading to mobile money access is strictly positive. The mag- nitude of these latter estimates is much larger, suggesting that mobile money’s effects are mostly driven by the intensive margin rather than the extensive margin. 6.2. Heterogeneity: Which Areas Benefit the Most? Wave 1 cells appear to benefit more from M-PESA access relative to subsequent waves. These cells also had higher night lights at baseline, and are closer to banks, roads, and cities. While there might be a number of other unobserved factors that might bolster the impact of M-PESA in these areas, this might suggest that the positive effects of M-PESA are strongest in areas that are initially richer, urban, and better connected to infrastructure. Panel A in table 5 formalizes this relationship by estimating specifications where a pooled dummy for agent access in the post period is interacted with summary measures of distance to infrastructure. Columns 1 and 2 show results with a measure of average log distance to the closest infrastructure (road, bank branch, cell tower, and electricity line). Columns 3 and 4 show results using an alternative measure based on a principal component analysis to derive an infrastructure distance index. Columns 5 and 6 show the interaction of the access variable with the log of distances to the closest town or city. The coefficients on the interaction terms are negative and statistically significant at the 1 percent level in all cases. This result suggests that the mobile money effect is more substantial the closer the area is to better 752 Fabregas and Yokossi Table 5. Interactions with Distances to Infrastructure and Cities Dependent variable: std. night lights [X ] Avg. distance PCA index Dist. city Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 (1) (2) (3) (4) (5) (6) Panel A. Full analysis sample Access 5 km 0.456*** 0.099*** −0.004 −0.025*** 0.460*** 0.115*** (0.084) (0.030) (0.019) (0.009) (0.078) (0.024) Access*[X] −0.206*** −0.056*** −0.105*** −0.020* −0.163*** −0.050*** (0.044) (0.016) (0.033) (0.010) (0.031) (0.010) Mean 0.04 0.04 0.04 0.04 0.04 0.04 Pixels 44,105 44,105 44,105 44,105 44,105 44,105 Observations 617,470 617,470 617,470 617,470 617,470 617,470 Panel B. Excluding wave 1 Access 5 km 0.255*** 0.101*** 0.019 0.007 0.259*** 0.093*** (0.061) (0.033) (0.013) (0.009) (0.074) (0.031) Access*[X] −0.103*** −0.041*** −0.056*** −0.020** −0.081*** −0.029*** (0.029) (0.014) (0.018) (0.009) (0.027) (0.010) Year FE Yes Yes Yes Yes Yes Yes Pix FE Yes Yes Yes Yes Yes Yes Subloc*time FE Yes Yes Yes Yes Yes Yes Pix trends No Yes No Yes No Yes Mean −0.20 −0.20 −0.20 −0.20 −0.20 −0.20 Pixels 27,842 27,842 27,842 27,842 27,842 27,842 Observations 389,788 389,788 389,788 389,788 389,788 389,788 Source: Authors’ calculations based on night-light data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMS), and M-PESA agent and bank data from the FinAccess geospatial mapping, road data adapted from Hijmans et al. (2016), cell tower data from OpenCell ID, major electricity transmission lines from the African Infrastructure Country Diagnostic database, and city locations from Henderson, Storeygard, and Deichmann (2017). Note: Dynamic difference-in-differences estimates are reported. The dependent variable is standardized light density at night observed at the pixel-year level. For each pixel cell, year 0 indicates the year within which the cell’s distance to the closest M-PESA agent came below 5 km. Interactions of the access variable with the log of the average distance to infrastructure (closest bank branch, road, cell towers, and electricity transmission lines) are shown in columns 1 and 2 and to a principal component index in columns 3 and 4. Columns 5 and 6 show the interactions of with the log distance to the closest city. The study period runs from 2000 to 2013. Standard errors clustered at the district level are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. infrastructure and cities. These results do not seem to be only driven by wave 1. Panel B in table 5 shows similar results when wave 1 pixels are excluded.29 There are countervailing forces in the interplay between access to infrastructure and mobile money services. On the one hand, one might predict that mobile money and other types of infrastructure like banks, would be clear substitutes. As a result, the effects of mobile money should be stronger in more remote areas. Conversely, there might be complementarities between mobile money and other types of services. For instance, mobile money agents might operate better when banks are present because it could allow them to manage their cash more easily. Similarly, households or firms might use M-PESA to take better advantage of other existing services. The results speak in favor of a complementarity channel. These results also suggest that the effect of M-PESA is more consistent with a growth effect than a mere redistribution effect. In a simple redistribution story, one would expect areas that are poorer and that have less access to potential alternatives to mobile money services (banks, road transport, etc.) to benefit more from access to mobile money by receiving remittances from wealthier areas. However, it 29 Columns 1 to 8 in supplementary online appendix table S1.8 presents similar interactions but with each component of the infrastructure summary measure. Column 7 also confirms that population density plays no key role in the mobile money effect. The World Bank Economic Review 753 is instead the initially more connected areas that benefit the most from mobile money services: with the increased capacity and ease of sending remittances, making payments, and trading, they might be able to take better advantage of their access to roads, traditional banking services, and urban amenities. Long-term benefits to M-PESA are likely to be greater as the network of customers grows and it opens Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 new opportunities for exchange (Aron and Muellbauer 2019). This prediction is tested in supplementary online appendix table S1.9, which shows the results of interacting the access indicator with the sublocation percentage of access to M-PESA by year (columns 1 to 3) and with the national percentage of access (columns 4 to 6). An increase in sublocation access is associated with greater impacts (column 1, prior to including sublocation-time fixed effects). The results are positive for increased aggregate national access. 7. Conclusion A growing body of rigorous research has started to show the positive impacts of financial products in improving economic outcomes. Yet because of market failures or mistargeting, some of these financial innovations have failed to live up to the expectations (Karlan et al. 2016). A considerable challenge going forward is understanding what products work and how to operationalize them. This article contributes to this effort by investigating the impact of access to mobile money on aggregate local economic activity. The identification strategy exploits the variation in the timing of access to M-PESA to establish causal effects. Data from the expansion of the mobile money agent network in Kenya and light density data derived from nighttime satellite images are combined to show that areas that enjoy easier access to mobile money services grow faster than less well served areas. While there are differences in area-level characteristics at baseline, there is not much evidence to suggest that different waves of cells had different underlying growth trajectories. Moreover, the results are robust to the inclusion of area-specific linear time trends, further suggesting that the impacts are not driven by fundamental differences between areas that gained access to mobile money services at different times. The effect of mobile money is stronger in areas that are initially richer, urban, and better connected to infrastructure. While the study cannot provide evidence on the mechanisms of these results, this suggests that at least in Kenya, mobile money is likely to complement other alternatives that enable people to connect, trade, and allocate investments within their network. This could also suggest that the benefits of mobile money might be higher in the longer term as countries gain increased access to complementary infrastructure, and more individuals and firms take advantage of this technology. Several caveats are worth noting. First, while this article investigates growth effects at the local level, the results do not necessarily speak to how these effects would aggregate up. Second, Kenya is one of the leading countries in the adoption of mobile money, and this might be attributed to a range of contextual factors that might not be representative of other settings. Even within Kenya, the sample of cells is not nationally representative. The focus is on the groups of cells that gained access to an M-PESA agent during the first years of rollout. This sample of cells is more urban than other regions in the country, and there might have been other advantages from being an early mover. Despite the rapid increase in the number of mobile money users worldwide, we are just starting to understand the channels through which these types of innovations affect growth. As other countries push for similar scale-ups, a more granular and precise understanding of the strength of the different channels underlying the positive effects of mobile money on economic activity would help to understand whether one should expect similar impacts in other settings. References Abrahams, A., C. Oram, and N. Lozano-Gracia. 2018. “Deblurring DMSP Nighttime Lights: A New Method Using Gaussian Filters and Frequencies of Illumination.” Remote Sensing of Environment 210: 242–58. 754 Fabregas and Yokossi Adam, C., and S. E. Walker. 2015. “Mobile Money and Monetary Policy in East African Countries.” Said Business School Discussion Paper. Oxford: University of Oxford. Available at: www.sbs.ox.ac.uk/sites/default/files/research- projects/mobile-money/Monetary-policypaper.pdf. Aggarwal, S., V. Brailovskaya, and J. Robinson. 2020. “Cashing In (and Out): Experimental Evidence on the Effects of Mobile Money in Malawi.” In AEA Papers and Proceedings, Vol. 110, pp. 599–604. Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Aker, J. C., R. Boumnijel, A. McClelland, and N. Tierney. 2016. “Payment Mechanisms and Antipoverty Programs: Evidence from a Mobile Money Cash Transfer Experiment in Niger.” Economic Development and Cultural Change 65(1): 1–37. Aron, J., and J. Muellbauer. 2019. “The Economics of Mobile Money: Harnessing the Transformative Power of Technology to Benefit the Global Poor.” Centre for the Study of African Economies. University of Oxford. Available at: https://www.oxfordmartin.ox.ac.uk/downloads/May-19-OMS-Policy-Paper-Mobile-Money- Aron-Muellbauer.pdf. Bateman, M., M. Duvendack, and N. Loubere. 2019. “Is Fin-Tech the New Panacea for Poverty Alleviation and Local Development? Contesting Suri and Jack’s M-PESA Findings Published in Science.” Review of African Political Economy 46(161): 480–95. Bester, C. A., T. G. Conley, and C. B. Hansen. 2011. “Inference with Dependent Data Using Cluster Covariance Estimators.” Journal of Econometrics 165(2): 137–51. Blumenstock, J. E., N. Eagle, and M. Fafchamps. 2016. “Airtime Transfers and Mobile Communications: Evidence in the Aftermath of Natural Disasters.” Journal of Development Economics 120: 157–81. Borusyak, K., X. Jaravel, and J. Spiess. 2021. “Revisiting Event Study Designs: Robust and Efficient Estimation.” Working Paper. Available at: https://arxiv.org/abs/2108.12419. Bruederle, A., and R. Hodler. 2018. “Nighttime Lights as a Proxy for Human Development at the Local Level.” PloS One 13 (9): e0202231. Bruhn, M., and I. Love. 2009. “The Economic Impact of Banking the Unbanked: Evidence from Mexico.” World bank policy research working paper, No. 4981. Available at: https://doi.org/10.1596/1813-9450-4981. Bundervoet, T., L. Maiyo, and A. Sanghi. 2015. “Bright Lights, Big Cities: Measuring National and Subnational Eco- nomic Growth in Africa from Outer Space, with an Application to Kenya and Rwanda.” World Bank Policy Re- search Working Paper, No.7461. Available at: https://ssrn.com/abstract=2682850. Burgess, R., and R. Pande. 2005. “Do Rural Banks Matter? Evidence from the Indian Social Banking Experiment.” American Economic Review 95(3): 780–95. Callaway, B., and P. H. Sant’Anna. 2021. “Difference-in-Differences with Multiple Time Periods.” Journal of Econo- metrics 225(2): 200–30. Cameron, A. C., and D. L. Miller. 2015. “A Practitioner’s Guide to Cluster-Robust Inference.” Journal of Human Resources 50(2): 317–72. CCK. 2014. “Annual Report for the Financial Year 2013–2014.” Technical Report, Communication Commission of Kenya. Available at: https://www.ca.go.ke/document/annual-report-for-the-financial-year-2013-2014/. Chen, X., and W. Nordhaus. 2011. “Using Luminosity Data as a Proxy for Economic Statistics.” Proceedings of the National Academy of Sciences of the United States of America 108(21): 8589–94. Cogneau, D., and Y. Dupraz. 2014. “Questionable Inference on the Power of Pre-colonial Institutions in Africa.” Working Paper. Paris School of Economics. Available at: https://halshs.archives-ouvertes.fr/halshs-01018548/. Cook, C. J., and M. Shah. 2020. “Aggregate Effects from Public Works: Evidence from India.” The Review of Eco- nomics and Statistics. Available at: https://doi.org/10.1162/rest_a_00993. De Chaisemartin, C., and X. d’Haultfoeuille. 2020. “Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects.” American Economic Review 110(9): 2964–96. Demombynes, G., and A. Thegeya. 2012. “Kenya’s Mobile Revolution and the Promise of Mobile Savings.” World Bank Policy Research Working Paper, No. 5988. Available at: https://papers.ssrn.com/sol3/papers.cfm? abstract_id=2017401#. Donaldson, D., and A. Storeygard. 2016. “The View from Above: Applications of Satellite Data in Economics.” Journal of Economic Perspectives 30(4): 171–98. Dupas, P., and J. Robinson. 2013. “Savings Constraints and Microenterprise Development: Evidence from a Field Experiment in Kenya.” American Economic Journal: Applied Economics 5(1): 163–92. FSD. 2016. “FinAccess Geospatial Mapping 2015.” Technical Report, Bill and Melinda Gates Foundation and Central Bank of Kenya and FSD Kenya. Available at: https://doi.org/10.7910/DVN/SG589T. The World Bank Economic Review 755 Gibson, J. 2021. “Better Night Lights Data, for Longer.” Oxford Bulletin of Economics and Statistics 83(3): 770–91. Gibson, J., S. Olivia, and G. Boe-Gibson. 2020. “Night Lights in Economics: Sources and Uses 1.” Journal of Economic Surveys 34(5): 955–80. Goodman-Bacon, A. 2021. “Difference-in-Differences with Variation in Treatment Timing.” Journal of Econometrics 225(2): 254–77. Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 GSMA. 2019. “Harnessing the Power of Mobile Money to Achieve the Sustainable Development Goals.” https://www.gsma.com/mobilefordevelopment/resources/harnessing-the-power-of-mobile-money-to-achieve- the-sustainable-development-goals/. Henderson, J. V., A. Storeygard, and D. N. Weil. 2012. “Measuring Economic Growth from Outer Space.” American Economic Review 102(2): 994–1028. Henderson, J. V., A. Storeygard, and U. Deichmann. 2017. “Has Climate Change Driven Urbanization in Africa?” Journal of Development Economics 124: 60–82. Higgins, D., J. Kendall, and B. Lyon. 2012. “Mobile Money Usage Patterns of Kenyan Small and Medium Enterprises.” Innovations 7(2): 67–81. Hijmans, R., M. Cruz, E. Rojas, R. O’Brien, and I. Barrantes. 2016. “DIVA-GIS, version 1.4. A Geographic Information System for the Management and Analysis of Genetic Resources Data.” http://www.diva-gis.org/datadown. Online; accessed November 2016. Jack, W., A. Ray, and T. Suri. 2013. “Transaction Networks: Evidence from Mobile Money in Kenya.” American Economic Review 103(3): 356–61. Jack, W., and T. Suri. 2011. “Mobile Money: The Economics of M-PESA.” Working Paper. National Bureau of Eco- nomic Research. Available at: https://www.nber.org/papers/w16721. ———. 2014. “Risk Sharing and Transactions Costs: Evidence from Kenya’s Mobile Money Revolution.” American Economic Review 104(1): 183–223. Karlan, D., J. Kendall, R. Mann, R. Pande, T. Suri, and J. Zinman. 2016. “Research and Impacts of Digital Financial Ser- vices.” Working Paper. National Bureau of Economic Research. Available at: https://www.nber.org/papers/w22633. Levine, R. 2005. “Finance and Growth: Theory and Evidence.” In Handbook of Economic Growth, edited by Philippe Aghion and Steven Durlauf, 865–934. Lucas, R. E. 1988. “On the Mechanics of Economic Development.” Journal of Monetary Economics 22(1): 3–42. Mas, I. 2009. “The Economics of Branchless Banking.” Innovations 4(2): 57–75. Mas, I., and O. Morawczynski. 2009. “Designing Mobile Money Services: Lessons from M-PESA.” Innovations 4(2): 77–91. Mawejje, J., and C. Lakuma. 2017. Macroeconomic Effects of Mobile Money in Uganda (No. 260017). Economic Policy Research Centre (EPRC). Report. Available at: 10.22004/ag.econ.260017. Mbiti, I., and D. N. Weil. 2015. “Mobile Banking: The Impact of M-PESA in Kenya.” In African Suc- cesses, Volume III: Modernization and Development, 247–93. University of Chicago Press. Available at: https://www.degruyter.com/document/doi/10.7208/9780226315867/pdf#page=252. Michalopoulos, S., and E. Papaioannou. 2013. “Pre-Colonial Ethnic Institutions and Contemporary African Devel- opment.” Econometrica 81(1): 113–52. Miller, M. H. 1998. “Financial Markets and Economic Growth.” Journal of Applied Corporate Finance 11(3): 8–15. Min, B. 2008. “Democracy and Light: Electoral Accountability and the Provision of Public Goods.” Mora, R., and I. Reggio. 2012. “Treatment Effect Identification Using Alternative Parallel Assumptions.” UC3M Working Papers. Universidad Carlos III de Madrid. Departamento de Economía. Available at: http://hdl.handle.net/10016/16065. Morawczynski, O., and M. Pickens. 2009. “Poor People Using Mobile Financial Services: Observations on Customer Usage and Impact from M-PESA.” CGAP Brief, Washington DC: World Bank. Available at: https://openknowledge.worldbank.org/handle/10986/9492. Munyegera, G. K., and T. Matsumoto. 2016. “Mobile Money, Remittances, and Household Welfare: Panel Evidence from Rural Uganda.” World Development 79: 127–37. Murendo, C., and M. Wollni. 2016. “Mobile Money and Household Food Security in Uganda.” GlobalFood Discussion Paper 76, University of Goettingen, Germany. Available at: https://ageconsearch.umn.edu/record/229805/. Ndirangu, L., and E. M. Nyamongo. 2015. “Financial Innovations and their Implications for Monetary Policy in Kenya.” Journal of African Economies 24 (suppl_1): i46–i71. 756 Fabregas and Yokossi Pinkovskiy, M. L. 2013. “Economic Discontinuities at Borders: Evidence from Satellite Data on Lights at Night.” Working Paper. Massachusetts Institute of Technology. Riley, E. 2018. “Mobile Money and Risk Sharing against Village Shocks.” Journal of Development Economics 135: 43–58. Robinson, J. 1979. “The Generalisation of the General Theory.” In The Generalisation of the General Theory and Downloaded from https://academic.oup.com/wber/article/36/3/734/6587137 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 other Essays, 1–76. Palgrave Macmillan, London. Storeygard, A. 2016. “Farther on Down the Road: Transport Costs, Trade and Urban Growth in Sub-Saharan Africa.” The Review of Economic Studies 83(3): 1263–95. Sun, L., and S. Abraham. 2021. “Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treat- ment Effects.” Journal of Econometrics 225(2): 175–99. Suri, T., and W. Jack. 2016. “The Long-Run Poverty and Gender Impacts of Mobile Money.” Science 354(6317): 1288–92. Wieser, C., M. Bruhn, J. P. Kinzinger, C. S. Ruckteschler, and S. Heitmann. 2019. “The Impact of Mobile Money on Poor Rural Households: Experimental Evidence from Uganda.” World Bank Policy Research Working Paper, No. 8913. Available at: https://ssrn.com/abstract=3430525. Wolfers, J. 2006. “Did Unilateral Divorce Laws Raise Divorce Rates? A Reconciliation and New Results.” American Economic Review 96(5): 1802–20. Yadav, P., A. P. Heynen, and D. Palit. 2019. “Pay-As-You-Go Financing: A Model for Viable and Widespread Deploy- ment of Solar Home Systems in Rural India.” Energy for Sustainable Development 48: 4–53.