Empty digital wallets: new technologies and old inequalities in digital financial services among women

ABSTRACT

Digital financial services (DFSs) may lower certain costs of accessing finance, but they bring new costs, including difficulties accessing mobile networks. Using the Demographic and Health Surveys and several geocoded databases in Nepal, the Philippines, Senegal and Tanzania, this paper studies the distribution of digital finance use among women and its enabling infrastructure, including mobile phone towers, compared to traditional finance. The potential of digital technologies to lessen inequalities depends on availability and access, particularly for women who may already face gaps in financial inclusion. Mobile phone towers are more unequally distributed than traditional banks, though mobile phone use is near universal. However, digital finance use is still low and almost as unequal as traditional finance, driven by the same inequalities. Wealth, education and location appear to be strongly associated with access to DFSs. The results suggest that old inequalities may constrain the promise of new digital technologies.

INTRODUCTION: THE PROMISE OF DIGITAL FINANCE

Previous research has suggested that digital financial services (DFSs) have the potential to reduce inequalities in access to finance and, thereby, subsequently reduce income inequalities, to the extent that access to finance affects income. However, DFS depends on infrastructure of its own: namely, mobile phone towers and strong mobile networks, which determine both the availability of services and the types of services which can be used (Klapper, 2017; Perlman and Wechsler, 2019). Given this dependency on infrastructure, will mobile DFS be able to reduce the inequalities in access to traditional finance, or will it instead worsen existing ones?

During the COVID-19 pandemic, digital technologies have come to play an even more central role, with many governments increasingly turning to mobile money or digital payments. For example, the government of Rwanda has scaled up the use of digital cash transfers, despite inequalities in who uses mobile money services (World Bank, 2021). Countries like Senegal have expanded the use of mobile money and lowered or removed fees for these services, while some (like Ghana) offer cash transfers through mobile money platforms (Gentilini et al., 2020). Who benefit these programs and whether they have the potential to reach the most vulnerable depend on the distribution of mobile DFS.

The literature has highlighted the benefits of DFS and the lower costs associated with it in terms of transportation costs and time, but has not yet systematically explored the new costs that DFS introduce. In particular, DFS carries additional costs of accessing mobile networks and requires technological literacy. These two barriers, unique to digital technologies, may be correlated with the costs of traditional finance that DFS hoped to overcome. A critical part of assessing the implications of these barriers is assessing inequalities in the infrastructure needed to access DFS. This paper addresses this topic by studying how physical infrastructure and mobile phone network quality, as well as individual characteristics like education, affect the ability to access and use DFS.

This paper contributes to the literature in several ways. First, it offers a cross-country, country-wide overview of a little-studied topic. Although many have studied the potential of mobile DFS to reduce inequalities, few have studied the inequalities within the distribution of access and use of mobile DFS. This is the first systematic attempt to document the differing salience of physical infrastructure for DFS and traditional finance. Further, it is the first to quantitatively explore the distinction between infrastructure barriers to accessing to these technologies (such as mobile phone signal quality) and individual characteristics that may affect access (such as education). This paper merges several geospatial datasets that have not been previously studied in this context along with a survey of women in four countries (Nepal, the Philippines, Senegal and Tanzania).

This paper will assess the distribution of access to the enabling infrastructure of DFS use (mobile phone towers and mobile phone ownership) and the distribution of DFS use itself. It compares the spatial clustering of these with the clustering of enabling infrastructure of traditional finance. It also considers the strength of the relationship between access to infrastructure and adaption of these services. It then explores how individual characteristics like wealth and education predict mobile phone ownership, DFS adoption and use of traditional finance.

MOTIVATION AND CONCEPTUAL FRAMEWORK

The discussion about access to mobile technologies and infrastructure has been embedded in the capabilities approach to development, which proposes that well-being is directly a function of an individual’s opportunities (Robeyns, 2006). Access to digital technologies may boost development by allowing access to institutions and infrastructures, such as healthcare systems (Haenssgen and Ariana, 2018). In the context of financial inclusion, this view underpins the concept of DFS as a gateway to inclusion in formal financial systems and an opening for the reduction of inequality in access to finance.

The success of mobile technologies for development depends on the individual ability to engage with new technologies, as well as on the infrastructures and systems these technologies rely on (Haenssgen and Ariana, 2018). This is why some have proposed new technologies as ‘amplifiers’ of inequalities – driven in part both by unequal access (such as lack of ability to afford the new technology) and unequal ability to use them (such as education or infrastructure) (Toyama, 2011). In conceptualizing the ‘digital divide’, van Dijk (2012) proposes that digital inequalities often follow age, gender, race, education, and household size, among others. This conceptualization of who can access and use digital technologies is even more relevant during and after the COVID-19 pandemic, as access to mobile phones and digital technologies and the ability to use them even more strongly govern access to formal institutions like finance.

Access to finance—perhaps supported by DFS—may help promote growth and development through two main channels: by providing credit for productive investment and by providing households insurance against shocks. In both channels, both the level of access to finance as well the distribution of access to finance may matter. The poor, in particular, may benefit more from access to finance, according to both theory and empirical evidence. For reviews of empirical evidence in both channels, see Karlan & Morduch (2010) and Demirguc-Kunt et al. (2017).

In the first channel, theories of economic growth have suggested that access to finance should promote aggregate economic growth and ultimately lower poverty rates by promoting investment in productive activities as follows (Gurley and Shaw, 1955; McKinnon, 1973; Galor and Zeira, 1993): At the household level, these theories predict that giving individuals access to finance should encourage them to invest in income-generating activities that could drive growth and poverty reduction. Theories suggest that households may find themselves in a ‘poverty trap,’ wherein those who are poor are credit-constrained and are not able to escape poverty through productive investment. In turn, these inequalities in access to finance may reduce growth (Banerjee et al., 2019). Providing access to finance therefore can help ease this credit constraint and help promote both economic growth and poverty reduction.

In the second channel, access to finance allows individuals to save or purchase insurance products that can protect them from economic shocks. In other words, increasing access to finance is a step toward completing financial markets, which should result in higher growth. In this context, inequalities in access to finance again may matter to the extent that the poor and vulnerable are those who most need insurance.

The theories of access to finance propose that individuals will only open financial accounts if the benefits outweigh the costs of doing so. In general, the benefits of access to finance are thought to be large but the costs are also often large. Further, if these costs are large and the household does not have access to credit, they may not be able to invest in a financial account or product even when the benefits outweigh the cost. This means that a lack of access to finance can be inefficient but self-sustaining. Further, these effects will be stronger for two kinds of households: (i) the poor and (ii) those with high costs of accessing finance, such as those in rural areas.

Because of this, many have proposed digital technologies, especially in the form of DFSs, as a method to close this gap of the ‘unbanked’. Especially when these take the form of mobile DFS, that is, financial services provided through mobile phones, these technologies are seen as a promising opportunity for those with little access to transportation, credit records or collateral that prevent them from participating in traditional finance (Radcliffe and Voorhies, 2012; Demirguc-Kunt et al., 2015). This approach hinges on the idea that DFS are less costly, particularly in terms of transportation costs and travel time, and those costs are more equally distributed. In East Africa especially, mobile DFS has been found to have similar benefits as traditional finance: reducing poverty, increasing entrepreneurship and agricultural earnings and increasing saving and resilience to shocks (Jack and Suri, 2011; Kirui and Njiraini, 2013; Kikulwe et al., 2014; Munyegera and Matsumoto, 2016; Suri and Jack, 2016; Lee et al., 2017; Islam et al., 2018). For a review of the empirical evidence on the effectiveness of DFS, see Karlan et al. (2016). More strongly, some authors have suggested that DFS can themselves reduce income inequality by increasing financial inclusion (Demir et al., 2020).

However, these previous models have not considered that the costs of participating in DFS are still nonzero and are still unequal. Although transportation costs are reduced, DFS still carry with them two important types of costs: costs of accessing mobile networks and costs of technological literacy. Those who wish to use DFS must purchase mobile network plans, which can be expensive. In addition, in places where mobile phone network quality is bad or non-existent, households who wish to use DFS will have to spend more time and effort doing so. Both of these costs may be unequally distributed, with poor quality particularly likely to affect households in rural or remote areas. This leads us to the first hypotheses:

• H1: Mobile phone infrastructure, specifically mobile phone towers and quality of signal, is highly spatially concentrated in the same patterns as traditional banks.

• H2: Mobile phone infrastructure is strongly positively associated with DFS use.

• H3: Patterns of spatial inequality in DFS use are similar to patterns of inequality in traditional finance use.

• H4: Wealth positively predicts DFS use.

The second type of cost is the cost the technological access needed to use DFS. Just as when accessing traditional finance, users of DFS need to be literate and able to understand the financial products they are being offered, which may take some level of education. DFS may replicate or exacerbate these problems, by requiring that users also be technology savvy and able to navigate the apps or interfaces that DFS providers offer. This leads us to the next hypotheses:

• H5: Education positively predicts DFS use.

• H6: The individual-level predictors of DFS use are similar to those of traditional finance.

The analysis that follows draws a distinction between availability, on one hand, and adaption or use of mobile DFS, on the other hand. In this context, availability refers to whether the infrastructure exists to allow for the use of mobile DFS and the individual has access to this infrastructure and technology. In their study of the Philippines, Roberts & Hernandez (2019) define access to technology in terms of five components: availability, affordability, awareness, abilities and agency. In this study, availability is analyzed in terms of ownership of a mobile phone, which is in turn affected by local infrastructure. Meanwhile, adaption or use of DFS depends on an individual’s preferences and need for DFS, the costs of the program and awareness of the services. Although many of these factors are unobservable, we can study how demographic and socioeconomic characteristics of individuals make them more or less likely to adopt DFS. Both elements—availability and adoption—reflect dimensions of inequality.

BACKGROUND AND PREVIOUS LITERATURE

Previous work has highlighted the large remaining gaps in access to finance (Chaia et al., 2009; Johnston and Morduch, 2008; de Haan and Sturm, 2017; Fungáčová and Weill, 2015). Barriers to financial inclusion include distance to financial institutions, costs of accounts, onerous documentation requirements, lack of trust and unaffordable fees (Izquierdo and Tuesta, 2015; Allen et al., 2016; Nandru and Rentala, 2019; Esquivias et al., 2020). These costs are not equally distributed, meaning that access to finance is significantly associated with education and proximity to banks (Bhanot et al., 2012; Ghosh, 2020) and is concentrated in urban areas (Martínez et al., 2013). In addition, several studies have found significant spatial spillovers of financial inclusion on the country level (Wang and Guan, 2017; Bozkurt et al., 2018).

However, few studies have looked at inequalities in the costs associated with accessing DFS and barriers to accessing these new technologies, including physical barriers such as the location of mobile phone towers. One study in West Africa found that use of mobile money was significantly more likely among those who were already banked, more educated, male and wealthier (Senou et al., 2019). In qualitative interviews, only a small portion of mobile phone users in Tanzania are able to benefit financially from this (often expensive) technology (Malm and Toyama, 2021). Others have found that, in Kenya, mobile phones appeared to amplify existing inequalities, in particular among women in rural areas (Wyche et al., 2016). A paper focusing on Nigeria highlighted that inequalities in digital technology did not appear to be growing or shrinking between 2009 and 2015: they found that all education and wealth groups had similar growth rates of mobile phone ownership over this period (Billari et al., 2020). Similar conclusions that mobile technologies appear to be replicating existing inequalities—particularly when it comes to affordability and technological literacy—have been drawn in the Philippines and Tanzania (Roessler, 2018; Roberts and Hernandez, 2019). Others have argued that access to government services and websites in Nepal has been severely limited by inequalities in technological access and literacy (Acharya, 2020).

This study focuses on women who may lag behind in financial inclusion and digital technologies. Several studies have indicated that women have significantly lower likelihood of having a financial account even in DFS, although this gap may be smaller than for other forms of finance (Senou et al., 2019; Mndolwa and Alhassan, 2020). In developing economies around the world, women are ~9 percentage points less likely to have a financial account than men, although they may be more likely in some countries like the Philippines and the gap may be smaller when it comes to digital finance like mobile money accounts (Demirguc-Kunt et al., 2018). However, it is not clear how serious this limitation is, as other studies have found that the gender gap in access to finance in Sub-Saharan Africa can be explained entirely by observable characteristics of women and women-owned enterprises (Aterido et al., 2013). That is, observable characteristics like education and wealth may matter more for women than men when accessing finance (Ghosh and Vinod, 2017). Studying the inequalities that women, who may already be disadvantaged in access to finance, face in mobile DFS is critical to understanding how to reduce gender gaps in financial inclusion.

This paper focuses on four lower-middle income countries, Nepal, the Philippines, Senegal and Tanzania. These four countries differ geographically in terms of their mobile network infrastructures and in their experience with mobile money and other DFS. Comparing and contrasting them provides valuable insights into how a country’s context affects its experience. Senegal and Tanzania are alike in that they have both had significant experience with DFS, particularly in the form of mobile money systems, while the Philippines and Nepal have much lower uptake of these technologies. Commentaries have suggested that uptake in the Philippines has been low due to the weak mobile network and a lack of trust of mobile apps, despite the central bank’s efforts to encourage and enable mobile money providers (Lal and Sachdev, 2015; THENERVE, 2019). In Nepal, commentaries suggest that a major barrier has been lack of financial literacy (Khalti Digital Wallet, 2018). Still, in all of these countries, DFS are becoming increasingly popular and growing rapidly.

DATA AND METHODS

This analysis makes use of data from several sources, which are summarized in Table 1.

Table 1

Data sources

Data source	Purpose	Level
Demographic and Heath Surveys Round VII 2016 (DHS)	Contains questions on use of financial services as well as individual characteristics	Individual women and cluster-level geocodes
OpenStreetMap	Locations of banks	Individual establishments
OpenCelliD	Locations of mobile phone towers	Individual towers
Gridded Population of the World (GPW), v4	Population density (persons/square km)	5 km × 5 km grids
NTL	Used to approximate income on a geospatial level	0.1 degree × 0.1 degree grids
Speedtest by Ookla Global Mobile Network Performance Maps	Average mobile network download speeds	Zoom level 16 web Mercator tiles (approx. 611 m × 611 m)

Demographic and Health Surveys (DHS) are a set of cross-sectional surveys designed to be representative or nearly representative of country populations. These surveys are comparable across countries around the world. Although it is primarily a survey on women’s health, the 2016 version of the DHS contains useful questions on financial and digital technology. The measure of use of traditional finance is covered by the question: ‘Do you have an account in a bank or other financial institution that you yourself use?’ The question relating to use of mobile DFS reads: ‘Do you use your mobile phone for any financial transactions?’. This excludes DFSs that may be performed in other ways, such as by an agent or on a computer. The responses to this question appear to give a reasonable approximation of the underlying financial inclusion variables, despite the imprecision of the question (Using this question, the incidences of DFS use in the DHS are similar to the figures reported in the Global Findex 2017 for the percentage of the population with a mobile money account in each country (Demirguc-Kunt et al., 2018). For Nepal, the Philippines, Senegal and Tanzania, the Findex reports 0%, 4.5%, 31.8%, and 38.5%, respectively. The corresponding figures from DHS (see Fig. 1 and discussion hereafter) are 6%, 9%, 19%, and 32%, respectively.). The countries studied are those that had the 2016 round of the DHS available when research began.

Figure 1

Spatial patterns of banks and financial institution accounts Note: The size of points is for ease of visualization only and does not reflect the size of DHS clusters or the range of mobile phone towers or banks.

It should also be noted that the question used to approximate traditional finance does not specifically reference traditional finance or exclude DFSs but is interpreted as more closely related to traditional finance due to the use of the word ‘bank’ in the question. Because of the chance that this question also contains those who only have an account with a DFS provider and not a traditional finance provider, we can also consider the variable for those who use traditional finance only and do not use mobile DFS as defined in that question. Although this interpretation of the question is likely to underestimate the true prevalence of traditional finance, it should give us a bound on the results that we expect (These results are presented alongside the results using the original variable but are generally found to be similar and as such as not discussed in detail. In Nepal, the Philippines and Senegal, we find that only a small fraction of the population used both DFS and traditional finance by these measures, so the bound is quite close. In Tanzania, there is a large discrepancy, with only 1% of the population using only a traditional finance account and not DFS. The data cannot speak to whether this finding is due to ambiguity in the phrasing of the question or if it is due to another factor, such as the fact that only those who are already financially included are aware of, able, or willing to use DFS.).

The DHS data are geocoded at the cluster level, which gives the unique opportunity to analyze elements of financial inclusion in conjunction with other geo-coded datasets. This expands the analysis significantly by allowing for a much richer understanding of the context in which the survey respondents live and the spatial patterns of financial inclusion. However, to protect respondent privacy, these geocodes have been displaced. Urban and rural clusters may be displaced up to 2 and 5 km, respectively, with 1% of rural clusters displacement up to 10 km. This lends itself to a natural way to construct measures of proximity with the data. Mobile phone towers and other clusters are considered ‘nearby’ if they are within 10 km of the cluster. Notice that this means that several isolated clusters may not have any neighbors by this measure. The average number of neighbors and the number of clusters with no neighbors are reported in Table 2. A limitation of the geospatial analyses is that DHS clusters are not spread uniformly throughout the countries—they are sampled weighted on population. This means that the data may not cover the most remote areas in each country and estimates of inequality may be underestimates.

Table 2

Cluster neighbors

	Nepal	Philippines	Senegal	Tanzania
Number of clusters	383	1212	214	608
Percent of nonzero weights	0.59	0.6	0.77	0.43
Average number of neighbors per cluster	2.3	7.3	1.6	2.6

Notes: this table gives a summary of the weight matrix used to calculate the effects of neighbors on cluster-level characteristics.

The OpenStreetMap and OpenCelliD databases offer a unique opportunity to study the locations of banks and mobile phone towers in each country, respectively. Their global scale means that they contain in-depth information on all the countries studied. However, they are crowd-sourced and may contain errors. More concerning is that, if crowd-sourced volunteers focus on the most populated areas, then this will bias estimates of the spatial inequalities in the locations of banks and mobile phone towers. The OpenCelliD dataset has been cleaned by removing the 1% of mobile phone towers with the highest ranges to remove implausible values.

Nighttime light (NTL) data have been used to estimate income inequality in previous papers and found to correlate closely with GDP per capita (Zhou et al., 2015), GDP growth (Chaiwat, 2016) and income inequality measured derived from other sources (Mveyange, 2015). NTL adds to this analysis because, unlike the DHS wealth index, it is not limited by the location of DHS clusters or sampling, allowing us to better understand spatial inequality.

Finally, the analysis also makes use of the Speedtest by Ookla Global Mobile Network Performance Maps, which provide data on mobile network performance based on speed tests conducted between quarter 1 of 2019 and quarter 3 of 2021 (Provided by Ookla and accessed December 2, 2021. Ookla trademarks used under license and reprinted with permission.). Although this date range is somewhat later than the survey dates, this is the earliest it is available. The data are distributed on a quarterly basis with a resolution of ~611 m × 611 m. For this analysis, all quarters are aggregated in order to increase sample size and the geographic spread of observations. Each cluster’s average download speed is taken to be the average speed in all the tiles within 10 km of the cluster. Some clusters have missing values if there were no speed tests conducted nearby them.

The first part of this analysis documents inequalities in both availability and adaption of traditional finance and DFS by location and household income. Spatial inequalities are quantified using the spatial Gini index and inequalities by income are measured using concentration curves.

The spatial Gini index is a measure of spatial clustering. This measure of spatial inequality was originally applied to approximating income inequality using nightlight data but can be extended to consider other geospatial data sources (Dawkins, 2004, 2006, 2007; Elvidge et al., 2012). Rey & Smith (2013) decompose this coefficient into two parts: (i) a measure of spatial autocorrelation, that is, a measure of the inequality between neighboring areas; and (ii) a measure of inequality between non-neighboring areas. In other words, the spatial Gini for each country can be written:

$$ \mathrm{G}=\frac{\sum_{\mathrm{i}=1}^{\mathrm{n}}{\sum}_{\mathrm{j}=1}^{\mathrm{n}}{w}_{i,j}\left|{x}_i-{x}_j\right|}{2{n}^2\overline{x}}+\frac{\sum_{\mathrm{i}=1}^{\mathrm{n}}{\sum}_{\mathrm{j}=1}^{\mathrm{n}}\left(1-{w}_{i,j}\right)\left|{x}_i-{x}_j\right|}{2{n}^2\overline{x}}, $$

where x_i and x_j represent the values of the variable studied in the region in grid i and j, and w_i,j is a binary variable equal to 1 if the regions are neighbors and 0 if they are non-neighbors. The first term represents the portion of inequality due to differences between neighbors and the second represents the portion due to differences between non-neighbors. Although the spatial Gini index has several limitations, including the fact that the spatial Gini can be equal for multiple different distributions, it is used in this paper because it satisfies the principles of anonymity, transfer and income-scale independence and offers an easily interpretable output (Rey and Smith, 2013). An additional benefit of using the spatial Gini in this analysis is that it is readily compared with concentration coefficients for other variables with respect to income. For the purpose of this analysis, neighbors within 10 km have w_i,j = 1. However, because of the fact that few of the ‘nearby’ neighbors are close enough to be within useful range for sharing of infrastructure or services, only the overall Gini coefficients are presented.

The next part of the analysis quantifies the relationship between infrastructure, availability and use of these services. The paper uses spatial regression techniques to study how the distribution of mobile phone towers and banking institutions predicts the use of mobile phones and traditional financial services in each cluster. This regression on the cluster level is done using OLS in the following form (OLS relies on the assumption that observations are independently drawn, which may not be the case when there are spatial dependencies between the dependent or independent variables in different clusters. However, further analysis considering spatial autocorrelation between clusters is beyond the scope of this paper and left for future research):

$$ \begin{align*} {Y}_c&={\alpha}_0+{\alpha}_1\mathrm{electricit}{\mathrm{y}}_c+{\alpha}_2\ln \left(\mathrm{mobiletower}{\mathrm{s}}_c+1\right)\\&\quad+{\alpha}_3\ln \left(\mathrm{bank}{\mathrm{s}}_c+1\right)+{\alpha}_4\ln \left(\mathrm{popde}{\mathrm{n}}_c+1\right)\\&\quad+{\alpha}_5\ln \left(\mathrm{NT}{\mathrm{L}}_c+1\right)+{\varepsilon}_c \end{align*}$$

where Y_c is the outcome variable of interest (proportion owning mobile phone, proportion using mobile DFS, or proportion using traditional finance), electricity_c represents the proportion in the cluster who have electricity, mobiletowers_c represents the number of nearby mobile phone towers (within 10 km), banks_c represents the number of nearby banks (within 10 km), popden_c represents the population density in persons per square kilometer and NTL_c represents NTLs. Additional specifications replace ln(mobiletowers_c + 1) with its equivalents for specific types of mobile towers (GSM towers, UMTS towers and LTE towers, described in detail hereafter) or include a control for the average download speed for the cluster.

In addition to exploring the cluster-level dimensions of inequality, this paper also considers individual-level inequalities, that is, differences in DFS use based on individual socioeconomic characteristics. This paper presents the results of the following logit models for each country:

$$ \begin{align*} {Y}_i&={\beta}_0+{\beta}_1\mathrm{ag}{\mathrm{e}}_i+{\beta}_2\mathrm{ag}{\mathrm{e}}_i^2+{\beta}_3\mathrm{urba}{\mathrm{n}}_i+{\beta}_4\mathrm{childre}{\mathrm{n}}_i\\&\quad+{\beta}_5\mathrm{ownlan}{\mathrm{d}}_{\boldsymbol{i}}+{\beta}_5\mathrm{marrie}{\mathrm{d}}_i+{\boldsymbol{\beta}}_{\mathbf{6}}\mathbf{wealt}{\mathbf{h}}_{\boldsymbol{i}}\\&\quad+{\beta}_7\ln \left(\mathrm{mobiletower}{\mathrm{s}}_i\right)+{\beta}_8\ln \left(\mathrm{bank}{\mathrm{s}}_i\right)\\&\quad+{\boldsymbol{\beta}}_{\mathbf{9}}\mathbf{regio}{\mathbf{n}}_{\boldsymbol{i}}+{\omega}_i \end{align*}$$

where Y_i represents the logit of the binary outcome variable of interest (whether the person owns a mobile phone, whether they use DFS, or whether they use traditional finance), age_i represents the person’s age, urban_i represents a binary variable representing whether the individual lives in an urban area, children_i gives the total number of children ever born, ownland_i represents a binary variable for whether the individual owns land, married_i is a binary variable for whether the individual is married, wealth_i represents a vector of binary variables for whether the individual belongs to a given wealth index group and region_i represents a vector regional fixed effects. For this analysis, the DHS wealth index, a measure of wealth based on ownership of certain assets and quality of housing, is used. Although it would be interesting to add distance to the nearest bank or mobile phone tower to the individual regressions, this cannot be done accurately due to the random displacement of clusters in the DHS. Instead, the models control for the log of the count of nearby mobile phone towers and the log of the count of nearby banks. The average marginal effects are reported with standard errors clustered at the DHS cluster level.

RESULTS

Inequalities in infrastructure, access, and usage

A major barrier to access to finance is the distance, or more precisely, the travel time to financial institutions. The locations of financial institutions are highly clustered and unequally spread, mostly concentrated in the capital region and other major urban centers in each country (Fig. 1). Even more significantly, the spatial Gini index for the number of banks near a cluster ranges from 0.89 in the Philippines to 0.92 in Senegal and Tanzania (Table 3). Even the fact of whether the cluster has a single nearby bank is quite unequal, especially in Senegal and Tanzania where the spatial Gini indexes are 0.64 and 0.69, respectively. In line with these, the proportion of those using traditional finance in the sense of this survey is also highly unequal, although somewhat less unequal than the distribution of banks in most cases, ranging from 0.26 in Nepal to 0.60 in Senegal.

Table 3

Spatial Gini coefficients

	Nepal	Philippines	Senegal	Tanzania
Number of banks	0.90	0.89	0.92	0.92
Have a bank	0.39	0.31	0.64	0.69
Number of mobile phone towers	0.90	0.90	0.87	0.92
Have a mobile phone tower	0.14	0.05	0.13	0.35
Number of GSM towers	0.88	0.87	0.86	0.91
Number of UMTS towers	0.91	0.90	0.88	0.93
Number of LTE towers	0.95	0.91	0.94	0.95
Download speed	0.19	0.21	0.37	0.30
Proportion with a mobile phone	0.13	0.11	0.19	0.27
NTL	0.55	0.63	0.52	0.52
Population density	0.79	0.78	0.88	0.84
Proportion using DFS	0.57	0.63	0.49	0.42
Proportion using traditional finance	0.26	0.45	0.60	0.46
Proportion using only traditional finance	0.25	0.46	0.67	0.86

Along with this strong spatial clustering, the use of traditional finance is also regressive in most cases, with the concentration curve lying below the Lorenz curve for wealth, except in Nepal, where uptake is the lowest of all the countries considered (only 6%) (Fig. 2) (The concentration curves in this analysis are on the individual level using the DHS data. Following the literature, a concentration curve is considered regressive, relatively progressive or absolutely progressive depending on whether it lies below the Lorenz curve, above the Lorenz curve but below the diagonal or above the diagonal, respectively. See for example Lustig et al., 2011; Davoodi et al., 2010.). The low uptake may not give enough variation to fully understand the relationship between wealth and DFS use. The significance of differences between the concentration curve and Lorenz curve can be tested using a dominance test following the methods summarized in O’Donnell et al. (2007). The visual patterns are statistically significant in all except the case of Tanzania, where the curves cross.

Figure 2

Concentration curves

As noted earlier, ignoring costs of accessing mobile networks and levels of technological literacy, high access to mobile phones should imply better access to DFS. Mobile phone ownership is nearly universal: across all the countries studied, the majority of the sample owns a mobile phone (Fig. 3). In fact, promisingly, mobile phone technology is relatively equally distributed geographically, with the proportion of those owning a mobile phone having a spatial Gini ranging from 0.11 to 0.27 (Table 3). And mobile phone ownership is nearly proportional to population (Fig. 2). This highlights the great potential for DFS to be inequality reducing: if everyone who has a mobile phone had access to DFS technologies, then inequalities in financial inclusion would be nearly eliminated, and (if access to finance has positive effects on income), income inequality would decrease.

Figure 3

Traditional finance account ownership, mobile DFS use and mobile phone ownership

However, although mobile phone ownership is widespread, the costs associated with accessing DFS may not be. The infrastructure that supports effective mobile DFS is more clustered. The spread of mobile phone towers is just as unequal or even more unequal than traditional financial infrastructure, with a spatial Gini ranging from 0.87 to 0.92 (Fig. 4 and Table 3). Interestingly, the patterns do not appear to be any less unequal in Nepal, where the government-owned Nepal Telecom operates many of the towers. Whether there is a single nearby mobile phone tower is much less unequally distributed. However, access to and the growth of DFS around the world is limited by poor connection and types of network connections available, which may constrain the types of DFS services available in areas with fewer mobile phone towers (Perlman and Wechsler, 2019). In fact, several authors have pointed out that network quality is an important component of access, and one in which inequalities appear to persist. In the Philippines, while many have some connection, quality digital connectivity is still unequal and replicates existing urban/rural and educational divides (Quimba et al., 2020; World Bank, 2020).

Figure 4

Spatial patterns of mobile phone towers and mobile DFS use Note: The size of points is for ease of visualization only and does not reflect the size of DHS clusters or the range of mobile phone towers or banks.

To this extent, we should aim to measure the quality of mobile network signal. Although high-quality geospatial data on signal strength is not available for research, the analysis will use two proxies: the type of mobile phone tower and the download speed available in the area. We can consider three different types of mobile phone towers: GSM towers (which provide only 2G networks), UMTS (which can provide up to 3G) and LTE towers (which can provide 4G).

The quality of service is also unequal (Fig. 6). We find that GSM towers (the lowest quality) are the most equally spread, while LTE towers are the most unequal, more unequal than any other variable studied. On the other hand, average download speeds are much more equally distributed than tower type. Even so, it is still more unequally distributed than mobile phone ownership in every case. The two indicators should be interpreted together, since neither fully captures signal quality.

Together, these findings help document the ways that—although mobile phone technology would seem to enable DFS use—inequalities in uptake persist. Despite its strong potential given the high ownership of mobile phones, DFS adaption rates are only a small fraction of phone ownership (albeit more common than traditional finance in Africa) (Fig. 3) ( The low incidence of DFS use in the Philippines is consistent with user complaints in virtual discussions about DFS services and technical support. For example, among the ‘most helpful’ reviews, 65% of users of the Philippine DFS app PayMaya report a negative experience and complain that the app ‘continuously crashes’, accounts become stuck or frozen and transactions may not complete correctly. See https://appgrooves.com/app/paymaya-by-paymaya-philippines-inc/negative. Meanwhile, apps in Tanzania may get more positive reviews, with Tigo Pesa Tanzania having 85% positive reviews among the most helpful. See https://appgrooves.com/app/tigo-pesa-tanzania-by-millicom-services-uk-ltd/negative). And, further, the use of DFS is more spatially unequal than the distribution of mobile phone ownership (Table 3). At the same time, DFS use is in general regressive, with the concentration curve for DFS lying below that for income (Fig. 2). In most cases, its patterns are similar to those of traditional finance; however, in Nepal in particular, mobile DFS is more regressive than traditional finance. Thus, neither mobile phone ownership nor wealth alone explains the inequalities in DFS.

In general, physical infrastructure tends to be clustered in cities, which also tend to be wealthier than rural areas (Fig. 5). Thus, the concentration of DFS and traditional finance in cities may reconcile the results above. In fact, the spatial Gini for population density, a measure of how tightly clustered or dispersed the population of a country is, follows the same pattern (Table 3).

Figure 5

Spatial patterns of population density and physical infrastructure Note: The size of points is for ease of visualization only and does not reflect the size of DHS clusters or the range of mobile phone towers or banks.

Figure 6

Spatial patterns of download speeds and mobile phone towers Note: The size of points is for ease of visualization only and does not reflect the size of DHS clusters or the range of mobile phone towers.

Figure 7A

Cluster-level residuals for mobile phone ownership against download speed

Figure 7B

Cluster-level residuals for DFS use against download speed

Figure 7C

Cluster-level residuals for traditional finance use against download speed

Although it has the potential to be inequality reducing because of the widespread and relatively spatially equal and progressive distribution of mobile phones, the impact of DFS on inequality may ultimately be hindered by old inequalities in physical infrastructure such as mobile phone towers, as well as inequalities by wealth, education and urban areas. Although mobile phone ownership is widespread, inequalities in the use of DFS echo the inequalities in access to traditional finance.

Inequalities in infrastructure predict inequalities in access to finance

The geolocations of the clusters allow us to study spatial inequalities in infrastructure and how they relate to access to and use of DFS and traditional finance. In particular, this analysis considers how physical infrastructure such as mobile phone towers and banks is associated with mobile phone ownership, DFS use and traditional finance use. These results cannot be said to be causal—for example, the debate remains on whether any positive relationship between infrastructure and technology is the result of supply or demand for infrastructure. For example, DFS adaption may be enabled by strong mobile phone networks, or high adoption of DFS in a certain area may cause higher demand for mobile phone networks and result in the construction of more infrastructure to support higher network usage. But instead, these results highlight an important pattern: DFS and digital technologies may not be any more equally distributed than traditional finance if they also are strongly associated with infrastructure. Another limitation of this analysis is strong multicollinearity in the covariates studied—perhaps itself an indicator of the clustering.

Descriptive statistics on the cluster level are available in Table 4. The models predict the proportion of individuals in each cluster who own a cellphone, use mobile DFS, or use traditional finance (Tables 5A–5C). The proportion of those in the cluster who report access to electricity is always positively and significantly associated with mobile phone ownership and access to finance. This indicates a basic level at which infrastructure may affect access to mobile DFS.

Table 4

Descriptive statistics, cluster level

	Nepal	Philippines	Senegal	Tanzania
Proportion owning mobile phones
Mean	0.73	0.82	0.62	0.51
Min	0.11	0.00	0.08	0.00
Max	1.00	1.00	1.00	1.00
Proportion using mobile DFS
Mean	0.06	0.09	0.19	0.32
Min	0.00	0.00	0.00	0.00
Max	0.41	0.80	0.73	0.93
Proportion using traditional finance
Mean	0.41	0.19	0.08	0.24
Min	0.00	0.00	0.00	0.00
Max	0.89	1.00	0.64	0.86
Proportion using only traditional finance
Mean	0.38	0.15	0.03	0.01
Min	0.00	0.00	0.00	0.00
Max	0.81	0.80	0.27	0.22
Proportion with electricity
Mean	0.90	0.92	0.58	0.24
Min	0.00	0.00	0.00	0.00
Max	1.00	1.00	1.00	1.00
Number of mobile phone towers
Mean	565.0	6317.0	1004.0	582.0
Min	0.00	0.00	0.00	0.00
Max	9450	93 946	14 311	11 882
Number of banks
Mean	46.0	132.0	22.0	9.7
Min	0.0	0.0	0.0	0.0
Max	668.0	2073.0	427.0	208.0
Log of population density
Mean	5.89	6.60	4.83	5.44
Min	2.22	1.60	1.59	0.15
Max	10.32	10.90	10.00	10.25
Log of NTL
Mean	3.03	3.30	3.43	3.33
Min	0.00	0.00	0.00	0.00
Max	5.43	5.50	5.39	5.55
Average download speed
Mean	13.80	14.00	28.91	12.98
Min	2.40	2.00	0.06	0.05
Max	47.10	34.00	104.08	55.97
N	383	1212	214	608
N (Download speed)	375	1208	160	431

Table 5A

Cluster-level relationships between infrastructure and access to finance

	Nepal					Philippines
	(1a)	(1b)	(1c)	(1d)	(1e)	(2a)	(2b)	(2c)	(2d)	(2e)
	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Download speed	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Download speed
Proportion with electricity	0.133***	0.043***	0.272***	0.239***	−3.909**	0.495***	0.104***	0.262***	0.198***	−0.949
	(0.050)	(0.014)	(0.038)	(0.035)	(1.573)	(0.035)	(0.021)	(0.024)	(0.019)	(0.822)
Log of cell phone towers	0.008	0.003	0.028***	0.026***	0.946***	0.015***	0.007***	0.002	0.0003	1.213***
	(0.006)	(0.003)	(0.007)	(0.007)	(0.233)	(0.004)	(0.003)	(0.003)	(0.003)	(0.108)
Log of banks	0.027***	0.012***	0.019**	0.009	−0.048	0.001	−0.002	0.028***	0.023***	0.792***
	(0.008)	(0.003)	(0.008)	(0.007)	(0.217)	(0.005)	(0.004)	(0.005)	(0.004)	(0.125)
Log population density	0.008	0.002	−0.009	−0.009	−0.248	0.013***	0.002	0.0001	−0.004	−0.091
	(0.011)	(0.005)	(0.010)	(0.009)	(0.296)	(0.005)	(0.003)	(0.004)	(0.004)	(0.127)
Log NTLs	−0.047***	−0.011**	−0.038***	−0.032**	0.509	−0.008	−0.002	−0.010	−0.007	−0.663***
	(0.014)	(0.006)	(0.014)	(0.014)	(0.363)	(0.005)	(0.005)	(0.006)	(0.005)	(0.158)
Constant	0.637***	0.020	0.205***	0.204***	13.997***	0.230***	−0.041*	−0.088***	−0.038	9.596***
	(0.062)	(0.022)	(0.048)	(0.046)	(2.091)	(0.035)	(0.024)	(0.029)	(0.024)	(0.971)
Observations	383	383	383	383	375	1212	1212	1212	1212	1208
R2	0.1618	0.1514	0.2998	0.25	0.1862	0.4238	0.05282	0.2468	0.1905	0.5866

Heteroskedasticity-robust standard errors in parentheses. ^***  P < 0.01, ^**  P < 0.05, ^*  P < 0.1

Table 5A

Cluster-level relationships between infrastructure and access to finance (continued)

	Senegal					Tanzania
	(3a)	(3b)	(3c)	(3d)	(3e)	(4a)	(4b)	(4c)	(4d)	(4e)
	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Download speed	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Download speed
Proportion with electricity	0.412***	0.304***	0.145***	0.052***	−2.875	0.370***	0.341***	0.289***	0.026***	2.105
	(0.025)	(0.020)	(0.013)	(0.007)	(4.014)	(0.030)	(0.036)	(0.036)	(0.007)	(1.482)
Log of cell phone towers	−0.013**	−0.013**	−0.010***	−0.003	−2.244	0.036***	0.016***	0.000	0.001	−0.324
	(0.006)	(0.006)	(0.004)	(0.002)	(1.529)	(0.005)	(0.006)	(0.005)	(0.001)	(0.251)
Log of banks	0.012	0.029***	0.014*	0.004	3.049	−0.025**	−0.012	0.014	0.001	−0.129
	(0.012)	(0.011)	(0.008)	(0.003)	(1.912)	(0.010)	(0.012)	(0.011)	(0.003)	(0.372)
Log population density	0.008	−0.006	−0.002	−0.002	0.607	−0.001	−0.008	−0.011*	0.000	−0.919**
	(0.009)	(0.008)	(0.005)	(0.003)	(1.309)	(0.007)	(0.007)	(0.007)	(0.001)	(0.424)
Log NTLs	−0.007	−0.008	0.0002	−0.001	−4.983*	−0.002	0.068***	0.054***	−0.003	1.517***
	(0.018)	(0.015)	(0.008)	(0.004)	(2.599)	(0.012)	(0.016)	(0.013)	(0.003)	(0.452)
Constant	0.409***	0.094**	0.025	0.023	52.346***	0.361***	0.027	0.044	0.014	13.832***
	(0.058)	(0.046)	(0.026)	(0.017)	(9.706)	(0.048)	(0.058)	(0.051)	(0.010)	(2.248)
Observations	214	214	214	214	160	608	608	608	608	431
R2	0.6076	0.4975	0.392	0.2137	0.0721	0.5484	0.4832	0.404	0.09921	0.03785

Heteroskedasticity-robust standard errors in parentheses. ^***  P < 0.01, ^**  P < 0.05, ^*  P < 0.1

Table 5B

Cluster-level relationships between infrastructure and access to finance, controlling for tower type

	Nepal					Philippines
	(1a)	(1b)	(1c)	(1d)	(1e)	(2a)	(2b)	(2c)	(2d)	(2e)
	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Download speed	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Download speed
Proportion with electricity	0.145***	0.049***	0.272***	0.234***	−4.075***	0.487***	0.112***	0.259***	0.184***	0.211
	(0.051)	(0.014)	(0.039)	(0.036)	(1.553)	(0.036)	(0.024)	(0.025)	(0.019)	(0.795)
Log of GMS towers	0.013	0.004	0.034***	0.032***	0.410	0.003	0.014*	0.014**	0.006	−0.118
	(0.011)	(0.005)	(0.011)	(0.010)	(0.368)	(0.008)	(0.008)	(0.007)	(0.005)	(0.198)
Log of UMTS towers	−0.004	0.000	−0.006	−0.006	0.706**	0.018**	−0.008	−0.010	0.0005	0.986***
	(0.010)	(0.005)	(0.011)	(0.010)	(0.310)	(0.008)	(0.010)	(0.008)	(0.005)	(0.196)
Log of LTE towers	0.026***	0.014**	0.001	−0.010	−0.574***	−0.010**	0.004	−0.002	−0.010***	0.874***
	(0.008)	(0.005)	(0.008)	(0.007)	(0.154)	(0.005)	(0.004)	(0.005)	(0.004)	(0.120)
Log of banks	0.020**	0.007**	0.021**	0.014*	−0.005	0.005	−0.002	0.030***	0.027***	0.279**
	(0.008)	(0.003)	(0.008)	(0.008)	(0.205)	(0.005)	(0.004)	(0.005)	(0.004)	(0.120)
Log population density	0.0001	−0.003	−0.007	−0.004	−0.091	0.014***	0.001	−0.0002	−0.003	−0.170
	(0.012)	(0.004)	(0.010)	(0.010)	(0.336)	(0.005)	(0.003)	(0.004)	(0.004)	(0.124)
Log NTLs	−0.050***	−0.013**	−0.038***	−0.032**	0.498	−0.004	−0.004	−0.009	−0.003	−1.074***
	(0.014)	(0.005)	(0.014)	(0.014)	(0.363)	(0.005)	(0.006)	(0.007)	(0.005)	(0.178)
Constant	0.681***	0.045**	0.201***	0.180***	13.595***	0.220***	−0.036	−0.092***	−0.053**	11.471***
	(0.067)	(0.020)	(0.053)	(0.049)	(2.300)	(0.036)	(0.025)	(0.030)	(0.024)	(1.034)
Observations	383	383	383	383	375	1212	1212	1212	1212	1208
R2	0.1808	0.1812	0.3002	0.2537	0.1968	0.4257	0.05494	0.2493	0.1958	0.6138

Heteroskedasticity-robust standard errors in parentheses. ^***  P < 0.01, ^**  P < 0.05, ^*  P < 0.1

Table 5B

Cluster-level relationships between infrastructure and access to finance, controlling for tower type (continued)

	Senegal					Tanzania
	(3a)	(3b)	(3c)	(3d)	(3e)	(4a)	(4b)	(4c)	(4d)	(4e)
	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Download speed	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Download speed
Proportion with electricity	0.414***	0.303***	0.145***	0.052***	−2.895	0.369***	0.343***	0.288***	0.026***	2.227
	(0.025)	(0.020)	(0.013)	(0.007)	(3.978)	(0.031)	(0.036)	(0.036)	(0.007)	(1.502)
Log of GMS towers	−0.008	−0.018	−0.008	0.000	−2.605	0.030***	0.019*	0.004	0.002	0.033
	(0.013)	(0.013)	(0.007)	(0.004)	(3.642)	(0.010)	(0.011)	(0.008)	(0.001)	(0.595)
Log of UMTS towers	−0.011	0.004	−0.003	−0.004	−0.024	0.012	−0.003	−0.004	−0.001	−0.475
	(0.011)	(0.012)	(0.007)	(0.004)	(3.617)	(0.011)	(0.012)	(0.010)	(0.002)	(0.526)
Log of LTE towers	−0.019	0.004	0.001	−0.002	−0.539	−0.014	−0.010	−0.012	0.0001	0.313
	(0.015)	(0.012)	(0.006)	(0.003)	(1.480)	(0.009)	(0.012)	(0.012)	(0.002)	(0.345)
Log of banks	0.027	0.024*	0.014*	0.007*	3.246	−0.017	−0.002	0.025*;	0.001	−0.284
	(0.017)	(0.013)	(0.008)	(0.004)	(2.517)	(0.014)	(0.017)	(0.014)	(0.003)	(0.447)
Log population density	0.014	−0.006	−0.002	−0.002	1.046	−0.001	−0.007	−0.011	0.0003	−0.862**
	(0.010)	(0.009)	(0.005)	(0.004)	(1.862)	(0.007)	(0.007)	(0.007)	(0.001)	(0.424)
Log NTLs	0.0002	−0.010	0.0005	0.0004	−4.741*	0.0002	0.070***	0.057***	−0.003	1.499***
	(0.018)	(0.016)	(0.009)	(0.005)	(2.652)	(0.012)	(0.016)	(0.013)	(0.003)	(0.453)
Constant	0.358***	0.101*	0.023	0.017	50.150***	0.355***	0.014	0.033	0.013	13.534***
	(0.059)	(0.052)	(0.028)	(0.019)	(11.533)	(0.049)	(0.061)	(0.053)	(0.010)	(2.375)
Observations	214	214	214	214	160	608	608	608	608	431
R2	0.613	0.4983	0.3926	0.2169	0.07368	0.5463	0.4828	0.4056	0.1009	0.03955

Heteroskedasticity-robust standard errors in parentheses. ^***  P < 0.01, ^**  P < 0.05, ^*  P < 0.1

Table 5C

Cluster-level relationships between infrastructure and access to finance, controlling for download speed

	Nepal				Philippines
	(1a)	(1b)	(1c)	(1d)	(2a)	(2b)	(2c)	(2d)
	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance
Proportion with electricity	0.112***	0.050***	0.265***	0.228***	0.491***	0.106***	0.264***	0.197***
	(0.043)	(0.016)	(0.043)	(0.038)	(0.035)	(0.022)	(0.024)	(0.019)
Log of cell phone towers	0.005	0.003	0.028***	0.026***	0.018***	0.006**	0.004	0.003
	(0.006)	(0.003)	(0.008)	(0.007)	(0.005)	(0.003)	(0.004)	(0.003)
Download speed	−0.001	0.0002	−0.002	−0.002	−0.003**	0.001	−0.002*	−0.003***
	(0.002)	(0.001)	(0.002)	(0.002)	(0.001)	(0.001)	(0.001)	(0.001)
Log of banks	0.033***	0.012***	0.022***	0.012*	0.004	−0.003	0.030***	0.025***
	(0.008)	(0.003)	(0.008)	(0.007)	(0.005)	(0.004)	(0.005)	(0.004)
Log population density	−0.006	0.001	−0.015	−0.015	0.013**	0.002	−0.001	−0.004
	(0.009)	(0.005)	(0.010)	(0.010)	(0.005)	(0.003)	(0.004)	(0.004)
Log NTLs	−0.035***	−0.010*	−0.030**	−0.026*	−0.010*	−0.002	−0.011*	−0.009*
	(0.013)	(0.006)	(0.014)	(0.014)	(0.005)	(0.005)	(0.006)	(0.005)
Constant	0.720***	0.019	0.259***	0.256***	0.259***	−0.050**	−0.071**	−0.010
	(0.056)	(0.025)	(0.056)	(0.052)	(0.038)	(0.026)	(0.030)	(0.025)
Observations	375	375	375	375	1208	1208	1208	1208
R2	0.1503	0.1493	0.2816	0.2287	0.4193	0.05265	0.246	0.1949

Table 5C

Cluster-level relationships between infrastructure and access to finance, controlling for download speed (continued)

	Senegal				Tanzania
	(3a)	(3b)	(3c)	(3d)	(4a)	(4b)	(4c)	(4d)
	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance	Proportion owning mobile phone	Proportion using mobile DFS	Proportion using traditional finance	Proportion using only traditional finance
Proportion with electricity	0.425***	0.297***	0.135***	0.047***	0.345***	0.312***	0.278***	0.027***
	(0.030)	(0.023)	(0.013)	(0.007)	(0.029)	(0.035)	(0.036)	(0.007)
Log of cell phone towers	−0.011	−0.012	−0.010**	−0.005*	0.028***	0.011*	−0.003	0.0005
	(0.008)	(0.009)	(0.005)	(0.003)	(0.006)	(0.007)	(0.006)	(0.001)
Download speed	0.0003	0.0002	−0.0001	−0.0001	0.003***	0.005***	0.004***	−0.0001
	(0.0005)	(0.0004)	(0.0002)	(0.0002)	(0.001)	(0.001)	(0.001)	(0.0001)
Log of banks	0.015	0.029**	0.014*	0.005	−0.020*	−0.011	0.018	0.001
	(0.014)	(0.012)	(0.008)	(0.004)	(0.011)	(0.013)	(0.013)	(0.003)
Log population density	0.000	−0.008	−0.001	−0.002	0.008	−0.009	−0.018**	0.0002
	(0.009)	(0.009)	(0.005)	(0.003)	(0.008)	(0.010)	(0.009)	(0.002)
Log NTLs	−0.001	−0.007	−0.0004	0.0003	−0.004	0.084***	0.069***	−0.004
	(0.019)	(0.016)	(0.009)	(0.005)	(0.012)	(0.018)	(0.015)	(0.003)
Constant	0.399***	0.091	0.030	0.032*	0.316***	−0.060	0.004	0.016
	(0.074)	(0.057)	(0.032)	(0.019)	(0.055)	(0.070)	(0.064)	(0.013)
Observations	160	160	160	160	431	431	431	431
R2	0.6228	0.5018	0.4031	0.2222	0.5625	0.5086	0.4348	0.08919

The results on the number of banks and mobile phone towers are more difficult to interpret, in part because these two measures are strongly related to each other, and it is difficult to identify separate effects for them. The number of nearby banks is positively associated with traditional finance use in all countries and all specifications. Interestingly, the log of the number of nearby banks is a positive and significant predictor for DFS use in both Nepal and Senegal, indicating that mobile DFS may not effectively combat locational barriers to access to finance. The log of the number of mobile phone towers is positively associated with the proportion owning a mobile phone and using DFS in all cases except for Senegal where it is marginally negatively significant.

Given the costs of accessing finance, we might expect NTLs (a measure of wealth) to be positively associated with use of traditional finance. The results are mixed: in Nepal, the coefficient is negative and significant; in the Philippines and Senegal, it is insignificant; and in Tanzania, it is positive and significant. In all countries, this is the same for both mobile DFS and traditional finance. The negative relationship in Nepal may represent areas of high inequality: holding the proportion that have electricity constant, more wealth may mean more inequality and thus lower access to finance and DFS.

However, the number of mobile phone towers does not tell the complete story. For example, many in the Philippines struggle with fragile and unreliable network connections, indicating that signal quality matters as well (Roberts and Hernandez, 2019). The data we have find an average mobile network download speed of 14.0 Mbps in the Philippines (Table 4). However, this a cluster-level average, so it does not necessarily reflect the average speed experienced by the population. There is a wide variance in the estimated mobile network speeds; Opensignal estimates that users in the Philippines get an average of 10.4–20.7 Mbps download speeds depending on carrier but can only reliably expect to experience speeds closer to 1.2 Mbps (Fenwick, 2021; Fogg, 2021). These speeds are on par with those measured in the other countries, although the observed average speeds are faster in Senegal. However, this higher average from Senegal should be interpreted with caution, again because these represent an average over the clusters and may have missing values. In reality, Speedtest.net reports that mobile network download speeds are similar across all four countries (Ookla, 2021).

The results in Nepal and the Philippines show that the number of mobile phone towers is strongly positively predictive of the average download speed (Table 5A). In Senegal and Tanzania, download speeds appear to be much more difficult to predict, with low regression R² and few significant predictors. This association means that inequality in the number of towers may in part reflect differences in signal quality, but this is not the full extent of how the number of towers matters. For example, capacity, reliability and continuity of signal range may also matter.

Adding type of tower as a control to the models from above, we find qualitatively similar results (Table 5B). The relationships between tower quality and technology adoption differs between the countries. In Nepal, LTE towers seem to drive DFS use, while in the Philippines and Tanzania, GSM towers are more significant. Once we control for tower type, the proportion using traditional finance is positively and significantly associated with the number of nearby banks in every country. Adding download speed as a control (Table 5C), again we find similar results. Although download speed is only a positive and significant predictor of DFS use in Tanzania, controlling for it does not eliminate the effects we found above.

Finally, for each of these outcomes, we can plot the regression residuals against download speeds to uncover any potential nonlinear associations between download speed and mobile phone ownership, DFS use and traditional finance use (Figs 7A, 7B and 7C). In general, the association between download speed and any of these outcomes is weak and rarely different than zero. As expected, controlling for other factors, there appears to be no association between traditional finance use and download speeds. This reassures us that download speeds are not simply proxying other unobserved factors associated with costs of accessing finance. In all cases, mobile phone ownership seems to increase with download speeds at very low speeds, but then the relationship becomes flat for higher speeds. This is consistent with mobile phones being less useful when quality is quite bad. However, there is no strong indication that DFS use is strongly affected by download speeds after controlling for all these other factors.

This indicates that the results are not only due to quality of service (which may be remedied by a single, high-quality tower) but also related to network capacity and other factors associated with the number of towers, issues which require much more investment to remedy.

Individual-level inequalities in access and use

However, inequalities in infrastructure only make up a portion of inequality in access: for example, infrastructure may affect availability of these technologies but individual or household-level factors like household wealth and education may determine affordability, awareness, abilities, and agency. In particular, several demographic and socioeconomic characteristics available in this survey mirror aspects of van Dijk’s digital divide, including age, education and household size, among others (van Dijk 2012).

Descriptive statistics on the individual level are available in Table 6. These models point out how the same inequalities behind traditional finance may also have strong implications for inequalities in DFS use (Tables 7A and 7B). Being in urban centers is positively associated with both traditional finance and DFS use, and is significant in the Philippines, Senegal and Tanzania. These results suggest that people living in urban areas are about 3.4–15.7 percentage points more likely to use traditional finance and 0.4–13 percentage points more likely to use DFS. Although urban location is a less strong predictor of DFS use than for traditional finance (perhaps again due to the small rates of DFS uptake in Nepal and the Philippines, making each coefficient have a smaller association with DFS use); in Senegal, belonging to an urban area is even more strongly associated with DFS use than traditional finance. These regressions hold constant the log of the number of nearby banks and mobile phone towers, so this is the association holding infrastructure constant. However, urban location may be multicollinear with these infrastructure variables. Even further, urban or rural location may proxy not only infrastructure quantity but also quality and reliability of the connection. In fact, Perlman & Wechsler (2019) propose that those living in rural areas, even when they are able to use DFS, can only access limited features and services compared with those in urban areas.

Table 6

Descriptive statistics, individual level

	Nepal			Philippines			Senegal			Tanzania
	Mean	Min	Max	Mean	Min	Max	Mean	Min	Max	Mean	Min	Max
Own mobile phone	0.72	0	1	0.82	0	1	0.61	0	1	0.51	0	1
Use mobile DFS	0.06	0	1	0.09	0	1	0.18	0	1	0.32	0	1
Use traditional finance	0.40	0	1	0.19	0	1	0.07	0	1	0.24	0	1
Log of mobile phone towers	3.44	0	9.2	5.17	0	11.5	3.84	0	9.57	2.73	0	9.38
Log of GSM towers	3.11	0	8.4	4.24	0	9.6	3.49	0	8.73	2.35	0	8.25
Log of UMTS towers	2.46	0	8.5	4.56	0	11.0	2.89	0	8.91	2.14	0	8.95
Log of LTE towers	0.25	0	5.5	2.67	0	9.7	0.67	0	6.60	0.57	0	5.91
Download speed	13.85	2.4	47.1	13.89	2.03	34.1	29.16	0.06	104.08	12.98	0.05	55.97
Log of banks	1.43	0	6.5	2.17	0	7.6	0.87	0	6.06	0.79	0	5.34
Age	29.51	15	49	30.17	15	49	28.11	15	49	28.89	15	49
Urban	0.64	0	1	0.37	0	1	0.47	0	1	0.31	0	1
Children ever born	2.07	0	15	1.88	0	18	2.58	0	13	2.87	0	17
Education in years	5.07	0	11	10.63	0	19	3.67	0	20	6.42	0	20
Male household head	0.68	0	1	0.82	0	1	0.72	0	1	0.77	0	1
Married	0.77	0	1	0.46	0	1	0.67	0	1	0.47	0	1
Wealth index = 1, poorest	0.22	0	1	0.21	0	1	0.29	0	1	0.17	0	1
Wealth index = 2, poorer	0.22	0	1	0.21	0	1	0.22	0	1	0.17	0	1
Wealth index = 3, middle	0.21	0	1	0.21	0	1	0.18	0	1	0.18	0	1
Wealth index = 4, richer	0.18	0	1	0.19	0	1	0.16	0	1	0.20	0	1
Wealth index = 5, richest	0.17	0	1	0.18	0	1	0.14	0	1	0.28	0	1
	N: 12268	N (Download speed): 12 024		N: 24063	N (Download speed): 23 982		N: 8772	N (Download speed): 6559		N: 12537	N (Download speed): 9061

Table 7A

Profiles of DFS and traditional finance users, AMEs

	Nepal				Philippines
	(1a)	(1b)	(1c)	(1d)	(2a)	(2b)	(2c)	(2d)
	Phone ownership	DFS among mobile phone owners	Traditional finance	Traditional finance only	Phone ownership	DFS among mobile phone owners	Traditional finance	Traditional finance only
Age	0.0669***	−0.005**	0.0761***	0.0724***	0.0244***	0.0138***	0.0296***	0.0221***
	(0.0031)	(0.0027)	(0.0035)	(0.0035)	(0.0017)	(0.0018)	(0.002)	(0.0019)
Age2	−0.001***	0.0001*	−0.001***	−0.0009***	−0.0004***	−0.0002***	−0.0004***	−0.0003***
	(0.00004)	(0.00004)	(0.0001)	(0.0001)	(0.00002)	(0.00003)	(0.00003)	(0.00003)
Urban	0.0412**	0.0045	0.0896***	0.0839***	0.0561***	0.0182	0.0341**	0.0191
	(0.0195)	(0.0149)	(0.0277)	(0.0278)	(0.0228)	(0.0231)	(0.016)	(0.0174)
Children ever born	−0.0274**	−0.0113	0.0004	0.0015	−0.0134**	0.0027	−0.0006	−0.0015
	(0.013)	(0.0096)	(0.0136)	(0.0139)	(0.0079)	(0.0088)	(0.0075)	(0.0071)
Education in years	0.0311***	0.0124***	0.0193***	0.0124***	0.0209***	0.012***	0.031***	0.0221***
	(0.0034)	(0.0043)	(0.0035)	(0.0034)	(0.0014)	(0.0018)	(0.0018)	(0.0017)
Male household head	−0.1346***	−0.0073***	−0.0786***	−0.0685***	−0.0199***	−0.0089***	−0.0131***	−0.009***
	(0.0014)	(0.0013)	(0.0014)	(0.0014)	(0.0007)	(0.0012)	(0.0012)	(0.0011)
Owns land	0.0613***	0.0091***	0.1208***	0.1021***	−0.0011	0.0178***	0.0495***	0.035***
	(0.005)	(0.0037)	(0.0054)	(0.0054)	(0.0027)	(0.0036)	(0.0033)	(0.003)
Married	0.1573***	0.0022	0.0837***	0.0782***	−0.012**	−0.0032	0.0145***	0.0148***
	(0.0115)	(0.0084)	(0.0141)	(0.0145)	(0.006)	(0.0057)	(0.0055)	(0.0053)
Log of mobile phone towers	−0.0127	0.0003	0.0008	0.0022	0.0056	0.0007	0.0013	0.0031
	(0.0141)	(0.0098)	(0.0139)	(0.0136)	(0.0074)	(0.0072)	(0.0066)	(0.0061)
Log of banks	0.0046	−0.0001	−0.0083	−0.0108	−0.002	−0.0014	0.0079*	0.0063
	(0.0087)	(0.0064)	(0.0088)	(0.0089)	(0.0071)	(0.0063)	(0.0051)	(0.0049)
wealth index = 2, poorer	0.0554***	0.0272**	0.1253***	0.1218***	0.0669***	0.0161	0.0094	0.0129
	(0.021)	(0.0137)	(0.0217)	(0.0228)	(0.0236)	(0.0215)	(0.0168)	(0.0184)
wealth index = 3, middle	0.0613***	0.0214**	0.1641***	0.1583***	0.1055***	0.0232	0.0456***	0.0449***
	(0.0197)	(0.0126)	(0.0201)	(0.0211)	(0.0207)	(0.0266)	(0.0179)	(0.0174)
wealth index = 4, richer	0.0749***	0.0244*	0.1661***	0.1581***	0.1628***	0.0479**	0.1054***	0.0879***
	(0.0196)	(0.0152)	(0.0271)	(0.027)	(0.0192)	(0.0235)	(0.0153)	(0.0175)
wealth index = 5, richest	0.094***	0.0295**	0.2215***	0.1962***	0.2497***	0.0865***	0.18***	0.1312***
	(0.0199)	(0.0134)	(0.0206)	(0.0219)	(0.019)	(0.021)	(0.0206)	(0.0219)
Observations	12 268	8806	12 268	12 268	24 063	19 783	24 063	24 063
McFadden Pseudo R2	0.1892	0.1005	0.1904	0.1623	0.2486	0.1059	0.2642	0.1896

Standard errors clustered at the DHS cluster level. All models estimated by OLS with region fixed effects. *** P < 0.01, ** P < 0.05, * P < 0.1

Table 7A

Profiles of DFS and traditional finance users, AMEs (continued)

	Senegal				Tanzania
	(3a)	(3b)	(3c)	(3d)	(4a)	(4b)	(4c)	(4d)
	Phone ownership	DFS among mobile phone owners	Traditional finance	Traditional finance only	Phone ownership	DFS among mobile phone owners	Traditional finance	Traditional finance only
Age	0.071***	0.0397***	0.021***	0.0071***	0.077***	0.0149***	0.0445***	0.0029***
	(0.0045)	(0.0053)	(0.0021)	(0.0016)	(0.0027)	(0.0048)	(0.0026)	(0.0009)
Age2	−0.0009***	−0.0005***	−0.0002***	−0.0001***	−0.001***	−0.0002***	−0.0006***	−0.00003***
	(0.0001)	(0.0001)	(0.00003)	(0.00002)	(0.00004)	(0.0001)	(0.00004)	(0.00001)
Urban	0.1785***	0.1304***	0.0726***	0.0325***	0.2059***	0.1238***	0.1571***	0.0038*
	(0.0339)	(0.0473)	(0.0096)	(0.0093)	(0.0392)	(0.0474)	(0.0215)	(0.0025)
Children ever born	−0.0399***	−0.0106	−0.0043	−0.0011	−0.0137	−0.0091	−0.0094	0.0002
	(0.0137)	(0.022)	(0.0076)	(0.0061)	(0.0136)	(0.0183)	(0.0135)	(0.0029)
Education in years	0.0216***	0.02***	0.0087***	0.0021**	0.0257***	0.0197***	0.0254***	0.0022***
	(0.0032)	(0.0039)	(0.0014)	(0.001)	(0.0024)	(0.0034)	(0.0022)	(0.0005)
Male household head	−0.0728***	−0.0544***	−0.0076***	0.0021***	−0.0646***	−0.0618***	−0.0486***	0.0028***
	(0.0016)	(0.0016)	(0.0006)	(0.0005)	(0.0012)	(0.0022)	(0.0013)	(0.0004)
Owns land	−0.0011	0.0161**	0.0215***	−0.0039***	0.0005	−0.0064	0.0112***	−0.005***
	(0.0043)	(0.0074)	(0.0019)	(0.0014)	(0.004)	(0.0051)	(0.0037)	(0.001)
Married	0.089***	0.023*	0.0061	0.0003	0.0229**	−0.0163	0.003	−0.003
	(0.0165)	(0.0173)	(0.0071)	(0.0051)	(0.0098)	(0.0132)	(0.0084)	(0.0024)
Log of mobile phone towers	−0.0036	−0.0111	−0.0007	0.0011	0.0073	0.0171	0.0049	0.0008
	(0.0184)	(0.0256)	(0.009)	(0.0066)	(0.0099)	(0.0143)	(0.0089)	(0.0029)
Log of banks	0.0059	0.0185	−0.0009	−0.0028	0.0092	−0.0082	−0.0065	−0.0008
	(0.0132)	(0.0151)	(0.0057)	(0.004)	(0.0099)	(0.0142)	(0.0089)	(0.0026)
wealth index = 2, poorer	0.0691*	0.0304	0.0243**	0.0077	0.0523	0.052	0.037*	−0.005
	(0.0419)	(0.0479)	(0.0142)	(0.0077)	(0.0422)	(0.0525)	(0.0234)	(0.0061)
wealth index = 3, middle	0.137***	0.0728	0.0578***	0.0237**	0.116***	0.0645	0.0849***	0.004
	(0.0359)	(0.069)	(0.0185)	(0.012)	(0.0399)	(0.0542)	(0.0251)	(0.0087)
wealth index = 4, richer	0.1961***	0.0844	0.0638***	0.0262***	0.1921***	0.0955**	0.1211***	0.0008
	(0.0286)	(0.0658)	(0.0114)	(0.0076)	(0.0386)	(0.0487)	(0.0221)	(0.0076)
wealth index = 5, richest	0.2198***	0.0911	0.0925***	0.0328*	0.2718***	0.1571***	0.2083***	0.0057
	(0.0251)	(0.0716)	(0.0157)	(0.0207)	(0.0342)	(0.0562)	(0.0229)	(0.0069)
Observations	8772	5390	8772	8772	12 537	6396	12 537	12 537
McFadden Pseudo R2	0.2037	0.1333	0.3261	0.191	0.2644	0.2002	0.2427	0.1492

Standard errors clustered at the DHS cluster level. All models estimated by OLS with region fixed effects. ^***  P < 0.01, ^**  P < 0.05, ^*  P < 0.1

Table 7B

Profiles of DFS and traditional finance users, coefficients

	Nepal				Philippines
	(1a)	(1b)	(1c)	(1d)	(2a)	(2b)	(2c)	(2d)
	Phone ownership	DFS among mobile phone owners	Traditional finance	Traditional finance only	Phone ownership	DFS among mobile phone owners	Traditional finance	Traditional finance only
Age	0.420***	−0.069*	0.413***	0.387***	0.223***	0.152***	0.267***	0.217***
	(0.0200)	(0.0380)	(0.0200)	(0.0200)	(0.0160)	(0.0200)	(0.0180)	(0.0180)
Age2	−0.006***	0.001	−0.005***	−0.005***	−0.004***	−0.002***	−0.003***	−0.003***
	(0.0003)	(0.0010)	(0.0003)	(0.0003)	(0.0002)	(0.0003)	(0.0003)	(0.0003)
Urban	0.258***	0.063	0.486***	0.448***	0.514***	0.200**	0.308***	0.188***
	(0.0820)	(0.1320)	(0.0750)	(0.0750)	(0.0720)	(0.0980)	(0.0680)	(0.0700)
Children ever born	−0.172***	−0.155***	0.002	0.008	−0.123***	0.029	−0.005	−0.015
	(0.0210)	(0.0580)	(0.0190)	(0.0180)	(0.0130)	(0.0190)	(0.0160)	(0.0170)
Education in years	0.195***	0.171***	0.105***	0.066***	0.192***	0.132***	0.280***	0.217***
	(0.0090)	(0.0180)	(0.0080)	(0.0080)	(0.0070)	(0.0130)	(0.0120)	(0.0120)
Male household head	−0.844***	−0.100	−0.426***	−0.366***	−0.182***	−0.098	−0.119**	−0.088*
	(0.0560)	(0.0880)	(0.0480)	(0.0480)	(0.0650)	(0.0690)	(0.0460)	(0.0480)
Owns land	0.385***	0.126	0.655***	0.545***	−0.01	0.196**	0.447***	0.344***
	(0.0890)	(0.1350)	(0.0770)	(0.0740)	(0.0680)	(0.0790)	(0.0600)	(0.0610)
Married	0.987***	0.030	0.454***	0.418***	−0.110**	−0.035	0.131***	0.145***
	(0.0730)	(0.1160)	(0.0770)	(0.0780)	(0.0550)	(0.0630)	(0.0490)	(0.0520)
Log of mobile phone towers	−0.080**	0.004	0.004	0.012	0.052**	0.008	0.012	0.03
	(0.0310)	(0.0510)	(0.0300)	(0.0290)	(0.0250)	(0.0390)	(0.0300)	(0.0290)
Log of banks	0.029	−0.001	−0.045	−0.058*	−0.018	−0.015	0.071**	0.062*
	(0.0400)	(0.0550)	(0.0340)	(0.0330)	(0.0320)	(0.0460)	(0.0350)	(0.0340)
wealth index = 2, poorer	0.347***	0.375**	0.679***	0.650***	0.613***	0.177	0.085	0.127
	(0.0720)	(0.1670)	(0.0820)	(0.0810)	(0.0560)	(0.1190)	(0.0900)	(0.0900)
wealth index = 3, middle	0.384***	0.295	0.890***	0.845***	0.967***	0.256**	0.411***	0.440***
	(0.0830)	(0.1860)	(0.0870)	(0.0880)	(0.0670)	(0.1200)	(0.0920)	(0.0930)
wealth index = 4, richer	0.470***	0.337*	0.901***	0.844***	1.493***	0.527***	0.951***	0.862***
	(0.0920)	(0.1930)	(0.1000)	(0.0980)	(0.0830)	(0.1210)	(0.0930)	(0.0920)
wealth index = 5, richest	0.590***	0.407*	1.201***	1.047***	2.289***	0.952***	1.651***	1.286***
	(0.1080)	(0.2160)	(0.1170)	(0.1120)	(0.1260)	(0.1270)	(0.0970)	(0.0970)
Constant	−6.385***	−2.667***	−9.195***	−8.630***	−3.996***	−6.890***	−10.570***	−9.130***
	(0.3130)	(0.5750)	(0.3200)	(0.3200)	(0.3070)	(0.4470)	(0.3280)	(0.3360)
Observations	12 268	8806	12 268	12 268	24 063	19 783	24 063	24 063
McFadden Pseudo R2	0.1892	0.1005	0.1904	0.1623	0.2486	0.1059	0.2642	0.1896

Standard errors clustered at the DHS cluster level. All models estimated by OLS with region fixed effects. *** P < 0.01, ** P < 0.05, * P < 0.1

Table 7B

Profiles of DFS and traditional finance users, coefficients (continued)

	Senegal				Tanzania
	(3a)	(3b)	(3c)	(3d)	(4a)	(4b)	(4c)	(4d)
	Phone ownership	DFS among mobile phone owners	Traditional finance	Traditional finance only	Phone ownership	DFS among mobile phone owners	Traditional finance	Traditional finance only
Age	0.397***	0.225***	0.429***	0.254***	0.454***	0.084***	0.331***	0.230***
	(0.0280)	(0.0310)	(0.0440)	(0.0550)	(0.0180)	(0.0270)	(0.0200)	(0.0690)
Age2	−0.005***	−0.003***	−0.005***	−0.002***	−0.006***	−0.001***	−0.004***	−0.003***
	(0.0004)	(0.0005)	(0.0010)	(0.0010)	(0.0003)	(0.0004)	(0.0003)	(0.0010)
Urban	0.999***	0.738***	1.479***	1.167***	1.214***	0.698***	1.169***	0.308
	(0.0820)	(0.1270)	(0.1490)	(0.1980)	(0.0840)	(0.1050)	(0.1040)	(0.2270)
Children ever born	−0.223***	−0.060***	−0.089***	−0.038	−0.081***	−0.051***	−0.070***	0.014
	(0.0190)	(0.0220)	(0.0290)	(0.0360)	(0.0140)	(0.0190)	(0.0160)	(0.0430)
Education in years	0.121***	0.113***	0.177***	0.077***	0.151***	0.111***	0.189***	0.175***
	(0.0090)	(0.0100)	(0.0120)	(0.0180)	(0.0080)	(0.0130)	(0.0100)	(0.0340)
Male household head	−0.407***	−0.308***	−0.154	0.074	−0.381***	−0.349***	−0.362***	0.225
	(0.0750)	(0.0850)	(0.1150)	(0.1440)	(0.0590)	(0.0800)	(0.0670)	(0.2060)
Owns land	−0.006	0.091	0.439**	−0.142	0.003	−0.036	0.084	−0.404*
	(0.1030)	(0.1450)	(0.1830)	(0.2380)	(0.0580)	(0.0810)	(0.0670)	(0.2290)
Married	0.498***	0.13	0.123	0.01	0.135**	−0.092	0.022	−0.243
	(0.0930)	(0.0980)	(0.1460)	(0.1830)	(0.0580)	(0.0740)	(0.0620)	(0.1970)
Log of mobile phone towers	−0.02	−0.063	−0.014	0.041	0.043*	0.096***	0.037	0.067
	(0.0240)	(0.0420)	(0.0380)	(0.0490)	(0.0230)	(0.0280)	(0.0280)	(0.0840)
Log of banks	0.033	0.105	−0.018	−0.101	0.054	−0.047	−0.048	−0.066
	(0.0400)	(0.0650)	(0.0670)	(0.0870)	(0.0480)	(0.0660)	(0.0540)	(0.1200)
wealth index = 2, poorer	0.386***	0.172	0.495**	0.276	0.308***	0.294**	0.275***	−0.405
	(0.0800)	(0.1530)	(0.2130)	(0.2630)	(0.0850)	(0.1260)	(0.1060)	(0.4440)
wealth index = 3, middle	0.767***	0.412***	1.176***	0.853***	0.683***	0.364***	0.632***	0.321
	(0.1000)	(0.1460)	(0.2140)	(0.2420)	(0.0850)	(0.1200)	(0.1030)	(0.3620)
wealth index = 4, richer	1.097***	0.478***	1.299***	0.940***	1.132***	0.539***	0.901***	0.066
	(0.1030)	(0.1470)	(0.2010)	(0.2370)	(0.0880)	(0.1200)	(0.1050)	(0.3730)
wealth index = 5, richest	1.230***	0.516***	1.883***	1.178***	1.602***	0.886***	1.550***	0.461
	(0.1230)	(0.1670)	(0.2150)	(0.2840)	(0.0960)	(0.1270)	(0.1070)	(0.3550)
Constant	−6.918***	−5.466***	−14.160***	−11.030***	−9.757***	−2.210***	−9.023***	−24.080***
	(0.4660)	(0.5140)	(0.8420)	(1.0840)	(0.3310)	(0.4760)	(0.3410)	(1.1690)
Observations	8772	5390	8772	8772	12 537	6396	12 537	12 537
McFadden Pseudo R2	0.2037	0.1333	0.3261	0.191	0.2644	0.2002	0.2427	0.1492

Standard errors clustered at the DHS cluster level. All models estimated by OLS with region fixed effects. ^***  P < 0.01, ^**  P < 0.05, ^*  P < 0.1

In all of the countries studied, inequalities in wealth and education are also carried through in DFS uptake even more strongly than in traditional finance (Fig. 5). First, in all countries, higher wealth is very strongly associated with a higher likelihood of owning a mobile phone. Those in the richest wealth quintile are 9.3–27.2 percentage points more likely to own a mobile phone, already suggesting that a lack of affordability of mobile networks may perpetuate inequalities in access. For example, in Tanzania, even the cheapest available mobile phones will cost 1/20 of annual income for those at the bottom of the wealth distribution, even before paying for a mobile phone plan, and a smartphone would cost about 1/6 of annual income (Karlsson et al., 2018). Costs for mobile broadband access are even higher: in the Nepal, the Philippines, Senegal and Tanzania, 1 gigabyte of mobile broadband costs 9.1%, 3.8%, 10.2%, and 8.7%, respectively, of average monthly income (A4AI, 2017).

Moreover, the average marginal effect associated with 1 year of education is also positive and large, with each year of education being associated with 1.2–2 percentage points higher likelihood of using DFS, compared with 0.9–3.1 percentage points higher likelihood of using traditional finance. Education in this context may proxy general literacy, as well as both technological and financial literacy. In fact, higher education also predicts mobile phone ownership, indicating the ways that digital technologies are not always accessible to those with less formal education.

In all the countries studied, the presence of a male household head is strongly negatively associated with mobile phone ownership, DFS use and traditional finance use. Previous studies have found that male-headed households are more likely to have access to finance, but women’s access to finance increases when they head their households (Demirgüç-Kunt et al., 2013; Ghosh and Vinod, 2017). In other words, this suggests that male household heads may handle financial matters for the household, leaving women without access to technology or accounts of their own.

Finally, notice that, after holding all of these factors constant, the coefficients on the log of banks and mobile phone towers are mostly insignificant. This may be due to the inclusion of region fixed effects (which may capture differences in infrastructure well if infrastructure is highly clustered by region) as well as urban/rural location.

In the Philippines, Nepal and Senegal, these models of personal characteristics explain less well the use of DFS than traditional finance (the McFadden Pseudo R² on the DFS regressions are about one half of those for the traditional finance regressions). (In interpreting the McFadden Pseudo R², it should be noted that this measure is known to be often lower than R² and should be interpreted in this light. McFadden (1977) writes that a value of 0.2 to 0.4 represents ‘an excellent fit’). This may be because these three countries have relatively low uptake of DFS compared to Tanzania (9%, 6%, and 18%, respectively). Thus, it makes sense that the regressions would not explain as powerfully the use of DFS in these countries. In contrast, the models predict use of DFS almost just as well as traditional finance in Tanzania, where uptake is higher.

Limitations

There are several limitations to this analysis. While it would be enlightening to consider on the individual- or cluster-level the importance of distance to financial institutions or mobile phone towers, the DHS data do not allow for this because of the random displacement of clusters in the data, as noted earlier. Future research should attempt to make a stronger connection between the location of physical infrastructure such as banks and mobile phone towers and individual likelihood of using traditional or digital finance.

Perhaps most importantly, we should note that the use of DFS may depend on the physical location of cash in or cash out points, where individuals can change between physical and digital money. To the extent that these are concentrated in the same areas as traditional financial institutions or other infrastructure, that may confound the associations between infrastructure and DFS use. Comprehensive data on the locations of these points are not yet available for research purposes but should be considered by future analyses hoping to isolate the effects of mobile phone networks in contrast to physical cash in/cash out locations.

DHS is not a survey on financial behavior or attitudes, so the analysis is limited by the questions available, which limits any examination of the mechanisms driving financial behavior. There are no questions, for example, on trust in financial institutions or awareness of DFS. Future research should attempt to examine whether attitudes toward finance can explain any difference between traditional finance and DFS use. In addition, the DHS survey analyzed only includes responses from women; interactions between dimensions of inequality like wealth and education may change how they are experienced by men.

Another weakness of the data is that the DHS findings come from 2016, which may be outdated given the pace of technological progress. There is also mismatch between the dates of some of the data sources, as the internet download speeds dataset only begins in 2019. This may be problematic due to the rapid spread and development of technology in recent years.

In addition, this analysis should not be interpreted causally – although the models suggest that higher wealth and education are associated with greater probability of DFS use, this does not indicate the direction of causation between them. It may instead be due to variations in other correlated variables which are unobservable in this analysis.

CONCLUSIONS

With the rise of mobile phone technology, solutions such as mobile digital finance have emerged as ways to address inequalities in physical infrastructure and access to finance. However, these technologies may simply replicate existing inequalities: while mobile phones are nearly universal, accessing DFS may come with its own costs. Specifically, DFS may depend on reliable and strong mobile networks with enough capacity, which may not exist everywhere. They may also depend on having education and technological literacy that is not needed with traditional finance.

This analysis finds that mobile DFS use is still highly unequal and that its supporting physical infrastructure such as mobile phone towers follows the same distributional patterns as traditional financial institutions. This is not entirely explained by download speeds or the quality of mobile phone towers. DFS use is associated with this highly unequal physical infrastructure.

On the individual level, characteristics like wealth, education and living in an urban area are still important determinants of use of mobile DFS. These may indicate lack of technological literacy or awareness to access DFS. Alternative explanations could be that those with lower education or wealth have less need for financial products, but this is generally not the case.

This analysis holds several insights for policy. Promoting DFS may not be an effective method to foster access to finance in remote areas, among the poor or among the uneducated. Without additional measures, these technologies may simply replicate the same inequalities as traditional finance. Policymakers hoping to use DFS to reduce inequalities should focus in parallel on reducing inequalities in digital infrastructure. Especially during and after the COVID-19 pandemic, it will become even more crucial to eliminate these persisting inequalities in access to mobile networks and digital technologies. Programs for investment in infrastructure in rural areas, as well as programs supporting affordability and digital literacy, may all help to reduce these inequalities. The most crucial thing is reducing the costs of access for these new technologies.

In 1987, Robert Solow famously noted, ‘You can see the computer age everywhere but in the productivity statistics’ (Solow, 1987). Similarly, perhaps we can say digital finance seems to be everywhere but not yet in everyone’s digital wallets.

ACKNOWLEDGEMENTS

The author wishes to thank Erwin Tiongson and Lidia Ceriani at Georgetown University, as well as two anonymous referees and the editor, for many helpful comments and insights.

FUNDING

This research did not receive any specific grant from funding agencies in the public, commercial or not-for-profit sectors.

CONFLICT OF INTEREST

None declared.

DATA AVAILABILITY

The data underlying this article are available in the Demographic and Health Surveys 2016 (available upon request at https://dhsprogram.com/), OpenStreetMap (queried from http://overpass-turbo.eu/s/Ose), OpenCelliD (available at https://www.opencellid.org/), the Gridded Population of the World Database (available at https://sedac.ciesin.columbia.edu/data/set/gpw-v4-population-density-adjusted-to-2015-unwpp-country-totals-rev11/data-download), NASA Earth Observatory (available at https://earthobservatory.nasa.gov/features/NightLights/page3.php) and Speedtest by Ookla Global Mobile Network Performance Maps (available at https://github.com/teamookla/ookla-open-data).

Footnotes

1

The latest version of this dataset is available here: http://overpass-turbo.eu/s/Ose 

2

The latest version of this dataset is available here: https://www.opencellid.org/ 

3

The dataset is available here: https://sedac.ciesin.columbia.edu/data/set/gpw-v4-population-density-adjusted-to-2015-unwpp-country-totals-rev11/data-download 

4

The dataset is available here: https://earthobservatory.nasa.gov/features/NightLights/page3.php 

5

The dataset is available here: https://github.com/teamookla/ookla-open-data