Digital credit and the gender gap in financial inclusion: Empirical evidence from Kenya

Johnen, Constantin; Mußhoff, Oliver Article — Published Version Digital credit and the gender gap in financial inclusion: Empirical evidence from Kenya Journal of International Development Provided in Cooperation with: John Wiley & Sons Suggested Citation: Johnen, Constantin; Mußhoff, Oliver (2022) : Digital credit and the gender gap in financial inclusion: Empirical evidence from Kenya, Journal of International Development, ISSN 1099-1328, Wiley, Hoboken, NJ, Vol. 35, Iss. 2, pp. 272-295, https://doi.org/10.1002/jid.3687 This Version is available at: https://hdl.handle.net/10419/287809 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by-nc-nd/4.0/ RESEARCH ARTICLE Digital credit and the gender gap in financial inclusion: Empirical evidence from Kenya Constantin Johnen | Oliver Mußhoff Department of Agricultural Economics and Rural Development, University of Goettingen, Goettingen, Germany Correspondence Constantin Johnen, Department of Agricultural Economics and Rural Development, University of Goettingen, Goettingen, Germany. Email: [email protected] Abstract Formal digital credit raises hopes to decrease the gender gap in financial inclusion. However, up until now, it remains unknown whether these hopes are justified. Using nationally representative household surveys from Kenya, the present study aims to fill this gap. We find strong indication that formal digital credit, contrasting to expectations, has led to an increase in the gender gap in financial inclusion. We further find indication that the pervasive gender gap in the formal digital credit market is largely attributable to gender differences in socio-economic variables in combination with a lack of contract term heterogeneity in that market. The paper suggests that policies to strengthen women's position in society and/or to encourage contract term heterogeneity in the formal digital credit market could decrease the gender gap in financial inclusion. KEYWORDS digital financial services, formal credit, gender equality, regulation 1|INTRODUCTION Increasing financial inclusion, typically defined as access to or usage of formal financial products, is largely believed to be an important driver of economic growth (Demirgüç-Kunt & Klapper, 2013; Hannig & Jansen, 2010). A main focus within the stream of the financial inclusion literature is understanding the pervasive gender gap in financial inclusion through formal credit markets (Adegbite & Machethe, 2020; Ghosh & Vinod, 2017; Guerin, 2006; Swamy, 2014), causing thriving discussions on how to increase financial inclusion in general and for women in Received: 3 February 2022 Revised: 10 June 2022 Accepted: 24 June 2022 DOI: 10.1002/jid.3687 This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made. © 2022 The Authors. Journal of International Development published by John Wiley & Sons Ltd. 272 J. Int. Dev. 2023;35:272–295. wileyonlinelibrary.com/journal/jid particular (Banerjee et al., 2015; Bharadwaj & Suri, 2020; Ghosh & Vinod, 2017). A new innovation, formal digital credit, raises hopes to increase financial inclusion, especially for women, by leveraging the widely accessible mobile money network, which already counts 1.2 billion subscriptions worldwide (Andersson-Manjang & Naghavi, 2021). Whether these hopes are justified remains largely unknown. In a recent study, Bharadwaj and Suri (2020) conclude that digital banking (which includes digital credit) holds its promise as an important source of financial inclusion. This finding is corroborated by recent studies which show that digital credit increases the total share of (formal) borrowers (Bharadwaj & Suri, 2020; Brailovskaya et al., 2021; Johnen et al., 2021). However, whether men and women profit equally from increases in financial inclusion through formal digital credit usage is not yet determined. In fact, early reports about digital credit in Kenya show that borrowers are more likely to be male (Gubbins & Totolo, 2018). However, these reports do not distinguish between formal digital credit: that is, at least one lending party is a regulated bank and semi-formal digital credit; that is, no lending party is a regulated bank. This distinction is relevant as only users of the former are considered financially included (Demirgüç-Kunt & Klapper, 2013). In addition, it appears conceivable that average formal digital credit user characteristics could change over time and might hence have changed from the reports about early digital credit users. In fact, literature about other digital financial services shows that initial gender differences in the usage of these services decrease over time (Jack & Suri, 2011). In contrast, recent studies indicate that such a development might not take place in the formal digital credit market. Formal digital credit is inherently different from other digital financial services, as users need to qualify for usage. There is a growing body of literature which shows that algorithms assign systematically lower digital credit scores to women (Brailovskaya et al., 2021). These gender differences in credit scores might result from socio-economic differences, which for instance might cause different phone using patterns (Björkegren & Grissen, 2018). The mere existence of differing credit scores in itself does not necessarily imply a gender gap in formal digital credit users. Low credit scores can, in theory, be compensated through adjustments in contract terms, such as higher interest rates (Hoff & Stiglitz, 1990). However, in many low-and middle income countries, available formal digital credit contract terms are rather homogenous and inflexible. The reason for that is that the formal digital credit market is often dominated by very few mobile network operator-bank partnerships who offer a formal digital credit product with the same interest rate and repayment period for everyone (Blechman, 2016; Francis et al., 2017). The contract terms lenders are able to offer can additionally be delimited by market regulations, such as interest rate caps. This might lead to a further restriction in available contract terms. In Kenya for instance a single loan called M-Shwari dominates the formal digital credit market (Suri et al., 2021). Every M-Shwari loan has a fixed interest rate of 7.5% per month and a fixed repayment period of 1 month; contract terms are merely heterogeneous in credit size (Suri et al., 2021). This lack of heterogeneity in available contract terms might cause pervasive credit rationing, if none of the few available contract terms can be offered below a certain credit score (Hoff & Stiglitz, 1990). Systematically lower credit scores of women could therefore translate into a pervasive gender gap in the formal digital credit market. It remains unknown whether initial gender differences in formal digital credit usage are indeed pervasive or, similar to other digital financial services, decrease over time. Consequently, there is also still very limited knowledge about why gender differences might persist over time. The present study addresses these issues. Therefore, our two central research questions are as follows: First, does formal digital credit decrease the gender gap in financial inclusion over time? Financial inclusion is defined as total formal credit usage, consisting of both formal digital and conventional credit. It is important to analyse both forms of formal credit usage to detect potential substitution effects. Second, what explains gender differences in formal digital credit usage? To answer the research questions, we use the nationally representative Kenya FinAccess household surveys from 2009, 2012, 2015 and 2018, which entail a wide range of socioeconomic characteristics and formal digital credit usage. More precisely, to answer the first research question, we analyse the changes in the gender gap concerning financial inclusion and distinguish the relative contribution of formal digital and formal conventional credit. To answer the second research question, we conduct logistic regressions based on the latest survey round, in order to understand whether socio-economic gender differences explain the gender gap in formal digital credit usage. We compare the results to the association between socio-economic characteristics and credit usage in the semi-formal digital credit market, where a substantially larger number of lenders compete for JOHNEN AND MUßHOFF 273 borrowers with different credit products. This comparison yields some further insight whether differences in market specific characteristics, such as contract term heterogeneity, are associated with the relationship between socioeconomic variables and digital credit usage. Kenya is a particularly telling country for this investigation as formal digital credit was already introduced in 2012, which enables the authors to observe changes over time. Our results show considerable, statistically significant and persistently lower user shares of women in the formal digital credit market. The differences in the formal digital credit market have further translated into an overall statistically significantly larger gender gap in the total formal credit market since the introduction of formal digital credit, from 4.1 percentage points in 2012 (just before digital credit was offered) to 7.8 percentage points in 2018. We find that gender differences in socio-economic variables such as income and education explain a large share of the gender gap in the formal digital credit market. In contrast, we do not find any gender differences in the unregulated semi-formal digital credit market (where the available contract terms are much more heterogeneous). We interpret this finding, in accordance with economic literature, as suggestive evidence that socio-economic differences explain the gender gap in the formal digital credit market partially due to the lack of suitable contract terms for women in that market. By answering our research questions, this study contributes to several strands of literature. First, this study largely contributes to literature investigating the pervasive gender gap in financial inclusion. This stream of literature identifies several obstacles for women in the formal conventional credit market, foremost long-travel times to banks and collateral requirements (Baydas et al., 1994; Fletschner & Kenney, 2014). In the present study, we are the first to investigate the gender gap in formal digital credit, a type of credit that renders the aforementioned obstacles obsolete. We show that despite these promising characteristics, formal digital credit has largely contributed to a statistically significant increase in the gender gap in financial inclusion. Second, we further contribute to the extensive body of literature that investigates adoption and usage of digital financial services. This stream of literature finds that initial gender differences in the usage of digital financial services decrease over time (Demombynes & Thegeya, 2012; Jack & Suri, 2011). However, this stream of literature has not yet considered one of the most prominent digital financial services, that is, digital credit. We show that the gender gap in formal digital credit in contrast to other digital financial services is substantial and persistent over time. The present study is also the first to investigate which socio-economic variables explain the persistent gender gap and thereby carves out explicit policy recommendations. Third, by comparing credit usage in the regulated formal and unregulated semi-formal digital credit markets, we also add to the literature investigating regulation and competition in credit markets (Cuesta & Sepulveda, 2018; Jaffee & Stiglitz, 1990). The present study is the first to compare the association of socio-economic variables between the formal and semi-formal digital credit markets. We show that socio-economic gender differences, in line with economic theory, do not translate into gender differences in digital credit usage in the unregulated semi-formal credit market. We thereby add to thriving policy discussions, which have recently contributed to the loosening of regulations in both formal credit markets, including the retraction of interest rate caps in all credit markets in Kenya. For policymakers, the present study underscores the long-standing recommendation to decrease women's unfavourable position in Kenya, mainly with regard to income, education, and mobile phone ownership. It also highlights that the recent measure of the Kenyan government to retract interest rate caps might prove particularly fruitful for the financial inclusion of women in the short run. We proceed by reviewing relevant literature about digital credit. We then describe the data and methods used in the present study. Next, we analyse gender differences in the semi-formal and formal digital credit market in Kenya. Finally, we derive policy implications and conclude. 2|DIGITAL CREDIT BACKGROUND Three attributes are typically used to distinguish digital from conventional credit (Chen & Mazer, 2016). First, digital credit is instant; this means that the time-span between application and approval is shorter than 24 h. Second, digital 274 JOHNEN AND MUßHOFF credit is automated; this refers to the automated credit approval process, which is undertaken by algorithms, based on non-traditional data (such as mobile money or phone usage data). Due to the second attribute of digital credit, borrowers typically do not need to provide collateral. This could be particularly beneficial for women, as women often have lesser control over resources accepted as collateral (Fletschner & Kenney, 2014). Third, digital credit is remote; this refers to the possibility of conducting the entire process from credit application to credit approval from anywhere as long as the borrower has access to mobile network. This attribute may also be particularly useful for women, who often have reproductive roles in the household, not allowing them to leave the house for a long time period. This circumstance is shown to render long travel times to banks a relatively larger obstacle for women (Baydas et al., 1994; Lycette & White, 1989; Moser, 1993). Due to the remote nature of digital credit, this constraint becomes largely obsolete. Other than the universal approach distinguishing digital credit from conventional credit, the status quo of digital credit is defined by rather short-term, expensive, low-sized loans (Francis et al., 2017). From an economic perspective, these features can help hedge the risk of default (Hoff & Stiglitz, 1990). This enables digital lenders to offer credit to a larger share of the population compared to formal conventional credit lenders. It follows that through formal digital credit a larger share of the population could become able to use a form of formal credit and start to build a financial history (Blechman, 2016; Brailovskaya et al., 2021). However, not all digital lenders can (or are willing to) adjust contract terms according to the risk assessment of the applicant. Formal digital lenders in Kenya (i.e., where at least one of the lending parties is a bank) are bound to the regulatory framework of the Central Bank Kenya (Blechman, 2016). While this regulatory framework is meant to protect consumers from excessively high fees, it can also cause credit rationing when the set of available contract terms does not allow lenders to make credit offers to individuals whose risk assessment is above a certain threshold (Hoff & Stiglitz, 1990). Especially poorer individuals who are less active users of mobile phones might not generate enough data to be scored by algorithms or may be assessed as too risky given the available set of contract terms (Blechman, 2016). In fact, recent studies show that women use mobile phones less often than men and that women are assigned with lower digital credit scores on average (Brailovskaya et al., 2021). In such cases, regulation might impede financial inclusion through formal digital credit. This may be one reason why all interest rate caps in Kenya have been retracted recently (this policy measure took place after the latest survey round in the present study). However, it should be noted that the formal digital credit market in Kenya is largely dominated by a single credit product, M-Shwari, which is offered at a fixed interest rate only (Suri et al., 2021). Whether the retraction of interest rate caps therefore indeed leads to a large heterogeneity in available contract terms remains dubitable. In contrast to formal lenders, semi-formal digital lenders in Kenya are not bound to the same set of regulations and can set interest rates arbitrarily high (Blechman, 2016); borrowers of semi-formal digital credit are therefore not considered financially included 1 (Demirgüç-Kunt & Klapper, 2013). Being able to set interest rates arbitrarily high might enable semi-formal lenders to grant credit to individuals which are considered too risky for formal digital lenders; such as women whose lower average credit scores might lead to being disproportionately often considered too risky in the formal digital credit market (Blechman, 2016; Brailovskaya et al., 2021). The downside of accepting riskier borrowers is that default rates are expected to increase. Indeed, due to comparatively high default rates and questionable loan enforcement practices, unregulated semi-formal digital lenders in Kenya are by now prohibited from reporting repayment behaviour to credit reference bureaus (CBK, 2020). While this restriction for semi-formal lenders might protect consumers, it also underscores that those borrowing in the semi-formal digital sector exclusively cannot build up credit scores, and might remain financially excluded. At the same time, it should be noted that the number of semi-formal digital lenders is much higher compared to their formal counterpart. This may be attributable to the lack of regulation in this market, which causes low entry barriers. In 2017, more than 110 semi-formal digital loan products were available in Kenya (Gwer et al., 2019). This competition has thereby led to a much greater variety in available contract terms in that market; including those with excessively high interest rates. From an economic perspective the large variety of contract terms increases the chances of people to find a credit product which they qualify for. JOHNEN AND MUßHOFF 275 3|DATA AND METHODS 3.1 |Data The present study is based on the 2009, 2012, 2015 and 2018 FinAccess household surveys. The surveys represent the second to fifth in a series of surveys, which aim to gather information about the usage of financial services in Kenya. The FinAccess household surveys emanate from a collaboration between the Kenya National Bureau of Statistics (KNBS), Financial Sector Deepening Trust (FSD) Kenya and the Central Bank of Kenya (CBK). The surveys employ a two-stage stratified cluster sampling design to collect data. The first level encompasses the selection of clusters based on the National Sample Surveys and Evaluation Program, with no substitute households being allowed. In the second level, households are selected in each cluster. From each chosen household, an individual aged 16 or above is then randomly selected. According to the FSD, this procedure ensures statistically valid and nationally representative samples for individuals aged 16 and above. We excluded all individuals who are less than 18 years of age from the respective survey rounds as formal credit lenders require borrowers to be 18 years of age or older. The 2009 FinAccess household survey was conducted in February 2009 and resulted in 6598 respondents; of which three minors were excluded for the present analysis (FSD Kenya, CBK, & KNBS, 2009). The 2012 FinAccess household survey was conducted during the time period of October 2012 to February 2013 and resulted in 6449 respondents, of which 263 minors were excluded for the present analysis (FSD Kenya, CBK, & KNBS, 2013). The 2015 FinAccess household survey was conducted during the time period August to October 2015 and resulted in 8865 respondents, of which 457 minors were excluded for the present analysis (CBK, FSD Kenya, & KNBS, 2016). Finally, the 2018 FinAccess household survey was conducted during the time period of October to December 2018 and resulted in 8669 respondents, of which 402 minors were excluded for the present analysis (FSD Kenya, CBK, & KNBS, 2019). The latest survey contains a unique set of questions regarding borrowing behaviour. Our analysis of why potential differences between women and men with regard to digital credit usage exist therefore focuses on this particular survey. For this in depths analysis, we also excluded all individuals with missing values from the present analysis, leading to a total sample of 7583 individuals. We further conducted appropriate weighting adjustments for all survey rounds in order to provide results which are nationally representative. 3.2 |Methods The first objective of the present study is to investigate whether formal digital credit decreases the gender gap in financial inclusion in Kenya through total formal credit usage over time. In order to achieve that aim, we compare how total formal credit usage, consisting of either formal digital credit and/or formal conventional credit (i.e., the lending institution is a bank or the government) usage, has changed for men and women between 2009 and 2018. We use different definitions of credit usage to answer the first and second research objectives. To answer the first research objective, we define credit usage (conventional and digital) is as having ever used the respective credit form. This definition is chosen for two reasons: First, this definition is suitable for comparisons over time, and second, all survey rounds used the exact same question to elicit whether a responded has ever used the respective credit forms. The latter ensures that the presented changes in credit usage are accurately reflecting real changes in the population and are not due to changes in the framing of the question. To answer our research question, we first calculate whether gender differences in the respective years are statistically significant (first differences). Based on the first differences we then analyse changes in gender differences over time (second differences). The second objective of the present study is to understand which factors explain differences in formal digital credit usage between men and women in Kenya; specifically, with respect to gender differences in socio-economics as well as the role of market characteristics in the formal digital credit market. In order to achieve these aims we 276 JOHNEN AND MUßHOFF estimate four different models. Models 1 and 2 include formal digital credit usage as an outcome variable; these models indicate how socio-economic characteristics are associated with formal digital credit usage. Model 3 and 4 include semi-formal digital credit usage as an outcome variable and are estimated to understand whether and why gender differences might not persist in the unregulated semi-formal digital credit market. As socio-economics are not stable over time (e.g., income and age), asking whether an individual has ever used the respective form of credit might be inappropriate. The 2018 FinAccess survey entails an additional question, which elicits whether individuals have used the respective credit forms in the past 12 months. As this more appropriate definition is only available in the 2018 FinAccess Survey, we answer the second research question based on this specific survey only. The binary nature of the outcome variables implies that a logit model can be used to investigate the relationship between digital credit usage and the set of explanatory variables (Hoetker, 2007). Previous studies which investigate gender gaps have also relied on propensity score matchings and Blinder-Oaxaca decomposition (Bauer et al., 2007; Fonseca et al., 2012; Robinson-Cimpian et al., 2014). However, the functional form assumptions of the parametric Blinder Oaxaca can lead to misleading results (Mora, 2008). The propensity score matching may hence be more appropriate. However, the PSM is particularly suitable when seven or fewer events per confounder exist; with more confounders, logistic regression should be the technique of choice (Cepeda et al., 2003). Given that there are roughly 630 observations per confounder on average in the present study, logistic regressions are the most appropriate estimation technique. In the first step we estimate the Models 1 and 3, which can be expressed as yi¼αþβFemaleiþμð1Þ where yiis a binary variable that equals one if an individual has used a formal digital credit (Model 1) or semi-formal digital credit (Model 3) within the past 12 months prior to the survey and zero otherwise. Femaleiis a binary variable that takes value one if the respondent is female and zero otherwise. The random error term μis expected to have a standard logistic distribution; αand βare parameters to be estimated in Equation (1). In order to investigate which confounders explain gender differences in digital credit usage, we estimate Models 2 and 4, which can be expressed as yi¼αþβFemaleiþγXiþμð2Þ where yiis a binary variable that equals one if an individual has used a formal digital credit (Model 2) or semi-formal digital credit (Model 4) within the past 12 months prior to the survey and zero otherwise. Femaleidescribes the same variable as in Equation (1). The random error term μis expected to have a standard logistic distribution; α,β, and γ are parameters to be estimated in Equation (2). Xiis a vector of control variables, including socio-demographic and socio-economic variables. Even though some previous studies suggest to include all clinically and intuitively relevant variables in the model regardless of their statistical significance, such a procedure can introduce numerically unstable estimates in the model and difficulties for other researchers to reproduce the results (Hosmer et al., 2013). Due to these shortcomings, we follow seven steps to purposefully select the model (cf. Hosmer et al., 2013, for an in depths description). Based on previous studies in the field of credit access and MFS adoption, we initially select 12 control variables as potentially relevant, including age of the respondent and age squared (age 2 ) to capture non-linear effects (Demombynes & Thegeya, 2012; Munyegera & Matsumoto, 2016); distance to nearest financial institution (Munyegera & Matsumoto, 2016); education level of the respondent (Munyegera & Matsumoto, 2016); whether the respondent is the household head (Tran et al., 2016); the respondent's income, transformed into log form (Demombynes & Thegeya, 2012); a dummy for whether the respondent is married (Demombynes & Thegeya, 2012); a variable that captures whether the respondent owns a mobile phone (Munyegera & Matsumoto, 2016) and/or a national ID (Gwer et al., 2019); the number of household members (Munyegera & Matsumoto, 2016); a variable JOHNEN AND MUßHOFF 277 which captures province fixed effects (Sekabira & Qaim, 2017); and a dummy that captures whether a respondent lives in an urban or a rural environment (Weil et al., 2012). During the purposeful selection, four of the 12 control variables that have been initially selected are excluded due to their statistical insignificance and absence of adjustment effects; that is, marital status, household head, urban and number of household members are excluded. We also conduct robustness checks including all 12 control variables. In addition, of all clinically or intuitively relevant interaction terms only gender by education is statistically significant albeit only across a fraction of the possible combinations. The interaction term is not included in the main model but discussed separately. 3.3 |Model specification We first examine whether the analysis is subject to multicollinearity, a common identification problem with crosssectional data, by generating a pairwise correlation matrix (Table A1 in Appendix A). The matrix shows the degree of correlation amongst the variables. The highest absolute correlation coefficient is 0.34, between “education”and “age.”We further estimate the variance inflation factor (VIF) between the variables. For both models in which we add control variables (i.e., Models 2 and 4) VIFs are between 1.06 and 2.88 with a mean VIF of 1.92. We conclude therefore that multicollinearity does not seem to cause adverse consequences to the model. The goodness of fit of the model is then tested with the Archer and Lemeshow test. This test was developed by Archer and Lemeshow (2006); it is similar to the more known Hosmer and Lemeshow test, but contrarily, accounts for the complex design of the survey sampling (Archer & Lemeshow, 2006). We find high pvalues of 58.9 and 14.1 for both models 2 and 4 respectively, strongly indicating no poor fit of the respective models. To assess whether all the relevant explanatory variables are included in the model, we run a link test. The link test uses the predicted value and its square as the predictors to rebuild the model. The predictor variable should be statistically significant, but not its square. If the latter is statistically significant, this could mean the functional form of the model is inadequate or that relevant variables are omitted (Dudek & Lisicka, 2013). In both models the predictor variable (hat) is statistically significant but not its square (hat 2 ), strongly indicating that the functional form of the respective models is not inadequate. 4|RESULTS AND DISCUSSION 4.1 |Descriptive statistics of borrower characteristics Overall, 51.5% of individuals in the 2018 FinAccess Survey are women, and 48.5% are men (Table 1). This share of women is close to the national average in Kenya, of 50.5% (KNBS, 2019). As a pre-condition to be able to use any form of digital credit, individuals need a valid (national) ID. The vast majority of respondents do fulfil this pre-condition, with a mean of 94.4% owning a national ID. Differences between men and women in national ID ownership are neglectable, contrasting prior research that attributed women's lower access to financial services to lower levels of ID ownership (Demirgüç-Kunt et al., 2013). Another pre-condition to be able to use any form digital credit is having a mobile phone; mobile phone ownership among respondents is also high at 83.9%. Here, we find statistically significant differences between genders, with men being 5.4 percentage points more likely than women to own a mobile phone. We further find that the majority of individuals live less than 10 min walking distance from the nearest financial institution (60.0%). A circumstance that can be largely attributed to the vast number of mobile money agents, more than 200 000 in 2018 (CBK, 2021). Table 1further corroborates previous findings on women's unfavourable position in the Kenyan society, showing that men's average income is statistically significantly higher, with a difference of more than 50%. In addition, women are statistically significantly more likely to not have any form of formal education, with a considerable 278 JOHNEN AND MUßHOFF difference of 8.6 percentage points. Men in contrast are statistically significantly more likely to have completed secondary education or even continued education thereafter, with considerable differences of 6.8 percentage points and 3.1 percentage points, respectively. TABLE 1 Socio-economic descriptive statistics of respondents (in the year 2018) Borrower characteristics Explanation Total Sample a Mean (SD) Men a Mean (SD) Women a Mean (SD) Difference women/men b Observations 7583 3172 4411 Gender (%) Dummy (takes value 1 if respondent is female and zero otherwise) 51.54 0 100 Infrastructure access Financial institution Distance Walking Distance to nearest financial institution in minutes <10 min 60.02 61.33 58.81 2.52* 10–30 min 28.20 27.91 28.48 0.58 >30 min 11.77 10.76 12.72 1.96** Mobile phone ownership (%) Dummy (takes value 1 if respondent owns a mobile phone and zero otherwise) 83.93 86.91 81.54 5.37*** National ID ownership (%) Dummy (takes value 1 if respondent owns a national ID and zero otherwise) 94.35 94.04 94.63 0.59 Socio-demographics Age (years) Age in years 39.97 (16.39) 40.52 (16.03) 39.45 (16.70) 1,07** Income (KSh) Income in Kenyan Shilling 11 082.66 (20 973.58) 13 591.18 (23 977.63) 8723.80 (17 365.36) 4867.38*** Education (%) No education 12.21 7.80 16.35 8.55*** Some primary 22.94 20.86 24.90 4.04*** Primary completed 21.80 22.34 21.29 1.05 Some secondary 11.01 11.84 10.22 1.64* Secondary completed 18.68 22.20 15.37 6.83*** Technical/ university 13.37 14.96 11.87 3.09*** Note: Authors' own calculations based on FinAccess survey 2018. a All means are calculated under consideration of the complex survey design using Stata's svy function. b Statistical difference is calculated with the adjusted Wald test. *p< 0.1. **p< 0.05. ***p< 0.01. JOHNEN AND MUßHOFF 279 deviations in coefficient sizes between logit models and LPMs could for instance be due to unobserved heterogeneity which is uncorrelated with the confounders, which can alter the magnitude of the confounders in the logit model, but not in the LPM. Table A3 shows that the association between gender, income and semi-formal digital credit usage in Model 4 (Table 3), is robust in sign, magnitude and statistical significance throughout the different logit model specifications. The LPM further indicates that changes in education level do not statistically significantly increase the likelihood that people use semi-formal digital credit, unless they have completed secondary education or higher. As the present study only investigates average effect estimates, the LPM analysis delivers unbiased and consistent estimators (Wooldridge, 2010). We can therefore estimate whether omitted variable bias could disprove the main findings, by using the LPMs. We estimate robustness values (RVs) for the estimates of “female”(4.61%), “mobile phone ownership”(3.21%), “education”(15.10%) and “income”(9.26%) in Model 2. The RVs indicate a threshold of how much of the residual variance, both of the treatment and the outcome, an unobserved confounder would need to explain to bring the lower bound of the confidence interval (of the estimated effect) to zero (significance level 5%) (Cinelli & Hazlett, 2020). To facilitate the interpretation of what that means, we use income as a benchmark for the former three RVs (i.e., for female, mobile phone ownership, and education), as income is correlated with all confounders and a strong predictor of digital credit usage. We find in each case that unobserved confounders as strong as income would not be sufficient to explain the observed estimate; with the higher bound estimates for income being below the respective RV in each case, at 1.51% (cf. RV female), 1.44% (cf. RV mobile phone ownership) and 3.71% (cf. RV education). To interpret the RV of income, we use mobile phone ownership as benchmark. We find that an unobserved confounder as strong as mobile phone ownership (higher bound at 1.44%) would not be sufficient to explain the observed estimate on income (cf. RV income). A main finding of the second research question is that socio-economic differences explain the major share of gender differences in formal digital credit usage. We shortly discussed in the section data and methods, why logit models are an appropriate choice for the present analysis. Nevertheless, we compare the main finding of the second research question to two other statistical methods, which are often used to analyse gender gaps (i.e., Propensity Score Matching and Blinder-Oaxaca decomposition). First, we estimate a propensity score model (PSM). This method enables us to compare men and women with similar covariate values in the observed characteristics, so that comparisons between genders are not confounded by differences in covariate distributions (Rubin, 1997). We choose nearest neighbour matching as a matching technique as this form of matching has been shown to perform consistently well across a wide range of settings (Austin & Stuart, 2017). In order to obtain excellent balance for all covariates we group income in ten groups; the other covariates are the same as in Model 2. The PSM estimates an average treatment effect on the treated of 1.4 percentage points, which is marginally statistically significant at the 10% level (Table A4). In other words, being female decreases the probability of formal digital credit take up by 1.4 percentage points, when compared to men with similar propensity scores. Second, we conduct a Blinder-Oaxaca decomposition, which is often used to decompose average gender gaps into an explained and unexplained part (Bauer et al., 2007; Blinder, 1973). The results indicate that 3.1 percentage points of the 5.1 percentage point gender gap can be explained by differences in the observed covariates (Table A5). This can be interpreted as 61.8% of the gender gap in formal digital credit usage being explained by gender differences in socio-economic variables. We view the findings from the last two robustness checks as strong support that the bulk of the gender gap in formal digital credit usage can be explained by differences in the observable socio-economic variables. 5|CONCLUSION AND RECOMMENDATIONS The present study reveals that formal digital credit has largely contributed to a statistically significant increase in the gender gap in financial inclusion. The present study therefore clearly shows that despite formal digital credit characteristics, which seem particularly beneficial for women (such as the lack of collateral requirements), as of 286 JOHNEN AND MUßHOFF now formal digital credit is no means to decrease the gender gap in financial inclusion; but rather a pathway towards increasing it. The study yields further insights as to why the gender gap in formal digital credit usage in Kenya persists. We find strong indication that gender differences in socio-economic variables (foremost in education, income, and mobile phone ownership) largely explain differences in formal digital credit usage. We further find that socio-economic gender differences do not lead to a gender gap in the unregulated semi-formal digital credit market. We suggest, in accordance with economic literature, that the absence of a gender gap in the semi-formal digital credit market could be partially due to the larger heterogeneity in available contract terms compared to its formal counterpart. The analysis therefore delivers valuable indications for policymakers, who aim to decrease the gender gap in financial inclusion. First, we find strong indication that strengthening women's position in society, foremost with respect to education, income and mobile phone ownership, is likely to translate into narrowing the gender gap in the formal digital credit market. Second, we find some indication that fostering a greater variety of formal digital contract terms (for instance by removing interest rate caps or stimulating competition) could increase the user share of women in the formal digital credit market. In fact, based on a similar reasoning, the Kenyan government has recently retracted interest rate caps for both formal credit markets. We hence strongly encourage future research to investigate the effects of the retraction of interest rate caps on financial inclusion of women through formal digital credit. In addition, the present study uncovers further potential to investigate why gender differences exist in the regulated formal credit market and not in the unregulated semiformal digital credit market; especially with regard to contract term availability. Lastly, even though Kenya is a vanguard for digital credit usage, socio-cultural, political and other relevant differences should be considered when results are transferred to other contexts. We therefore encourage future research to investigate the role of formal digital credit for financial inclusion of women in other countries. ACKNOWLEDGEMENTS Open Access funding enabled and organized by Projekt DEAL. DATA AVAILABILITY STATEMENT The data of FinAccess survey 2009, 2012, 2015 and 2018 are available in “Harvard Dataverse”with the digital object identifiers https://doi.org/10.7910/DVN/VWLGY4,https://doi.org/10.7910/DVN/JSPE9W and https://doi. org/10.7910/DVN/QUTLO2. The data of the FinAccess survey 2018 are available on the website of the Central Bank of Kenya, https://www.centralbank.go.ke/finaccess/. ORCID Constantin Johnen https://orcid.org/0000-0001-5455-8008 ENDNOTE 1 Unless, of course, the individual is considered financially included through the usage of other formal financial instruments REFERENCES Adegbite, O. O., & Machethe, C. L. (2020). Bridging the financial inclusion gender gap in smallholder agriculture in Nigeria: An untapped potential for sustainable development. World Development,127, 104755. https://doi.org/10.1016/j. worlddev.2019.104755 Agarwal, S., Chomsisengphet, S., & Liu, C. (2011). Consumer bankruptcy and default: The role of individual social capital. Journal of Economic Psychology,32(4), 632–650. https://doi.org/10.1016/j.joep.2010.11.007 Andersson-Manjang, S. K., & Naghavi, N. (2021). State of the industry report on mobile money. Archer, K. J., & Lemeshow, S. (2006). Goodness-of-fit test for a logistic regression model fitted using survey sample data. The Stata Journal,6(1), 97–105. https://doi.org/10.1177/1536867X0600600106 JOHNEN AND MUßHOFF 287 Atieno, R. (2001). Formal and informal institutions' lending policies and access to credit by small-scale enterprises in Kenya: An empirical assessment. AERC research paper: Vol. 111. African Economic Research Consortium. Austin, P. C., & Stuart, E. A. (2017). Estimating the effect of treatment on binary outcomes using full matching on the propensity score. Statistical Methods in Medical Research,26(6), 2505–2525. https://doi.org/10.1177/0962280215601134 Banerjee, A., Karlan, D., & Zinman, J. (2015). Six randomized evaluations of microcredit: Introduction and further steps. American Economic Journal: Applied Economics,7(1), 1–21. https://doi.org/10.1257/app.20140287 Bauer, T., Göhlmann, S., & Sinning, M. (2007). Gender differences in smoking behavior. Health Economics,16(9), 895–909. https://doi.org/10.1002/hec.1259 Baydas, M. M., Meyer, R. L., & Aguilera-Alfred, N. (1994). Discrimination against women in formal credit markets: Reality or rhetoric? World Development,22(7), 1073–1082. https://doi.org/10.1016/0305-750x(94)90149-x Bharadwaj, P., & Suri, T. (2020). Improving financial inclusion through digital savings and credit. AEA Papers and Proceedings, 110, 584–588. https://doi.org/10.1257/pandp.20201084 Björkegren, D., & Grissen, D. (2018). The potential of digital credit to bank the poor. AEA Papers and Proceedings,108,68–71. https://doi.org/10.1257/pandp.20181032 Blechman, J. G. (2016). Mobile credit in Kenya and Tanzania: Emerging regulatory challenges in consumer protection, credit reporting and use of customer transactional data. The African Journal of Information and Communication.,2016(17), 61–88. https://doi.org/10.23962/10539/21628 Blinder, A. S. (1973). Wage discrimination: Reduced form and structural estimates. Journal of Human Resources,8(4), 436–455. https://doi.org/10.2307/144855 Brailovskaya, V., Dupas, P., & Robinson, J. (2021). Digital credit: Filling a hole, or digging a hole? Evidence from Malawi. CBK. (2020). Publication of the credit reference bureau regulations, 2020 and additional measures on credit information sharing [Press release]. Nairobi. CBK. (2021). Mobile payments. https://www.centralbank.go.ke/national-payments-system/mobile-payments/ CBK, FSD Kenya, & KNBS. (2016). FinAccess household survey 2015. Harvard Dataverse. https://doi.org/10.7910/DVN/ QUTLO2 Cepeda, M. S., Boston, R., Farrar, J. T., & Strom, B. L. (2003). Comparison of logistic regression versus propensity score when the number of events is low and there are multiple confounders. American Journal of Epidemiology,158(3), 280–287. https://doi.org/10.1093/aje/kwg115 Chen, G., & Mazer, R. (2016). Instant, automated, remote: The key attributes of digital credit. https://www.cgap.org/blog/ instant-automated-remote-key-attributes-digital-credit Cinelli, C., & Hazlett, C. (2020). Making sense of sensitivity: Extending omitted variable bias. Journal of the Royal Statistical Society, Series B: Statistical Methodology,82(1), 39–67. https://doi.org/10.1111/rssb.12348 Cuesta, J. I., & Sepulveda, A. (2018). Price regulation in credit markets: A trade-off between consumer protection and credit access. https://doi.org/10.2139/ssrn.3282910 Demirgüç-Kunt, A., & Klapper, L. (2013). Measuring financial inclusion: Explaining variation in use of financial services across and within countries. Brookings Papers on Economic Activity,2013(1), 279–340. https://doi.org/10.1353/eca.2013.0002 Demirgüç-Kunt, A., Klapper, L. F., & Singer, D. (2013). Financial inclusion and legal discrimination against women: Evidence from developing countries. World Bank Policy Research Working Paper (6416). DOI: https://doi.org/10.1596/18139450-6416 Demombynes, G., & Thegeya, A. (2012). Kenya's mobile revolution and the promise of mobile savings. Diagne, A., Zeller, M., & Sharma, M. (2000). Empirical measurements of households' access to credit and credit constraints in developing countries: Methodological issues and evidence. https://ageconsearch.umn.edu/record/16465/files/ fc000090.pdf Dudek, H., & Lisicka, I. (2013). Determinants of poverty—Binary logit model with interaction terms approach. Ekonometria., 3(41), 65–77. https://www.ceeol.com/search/article-detail?id=110743 Feder, G., Lau, L. J., Lin, J. Y., & Luo, X. (1990). The relationship between credit and productivity in Chinese agriculture: A microeconomic model of disequilibrium. American Journal of Agricultural Economics,72(5), 1151–1157. https://doi.org/ 10.2307/1242524 Fletschner, D. (2009). Rural women's access to credit: Market imperfections and intrahousehold dynamics. World Development,37(3), 618–631. https://doi.org/10.1016/j.worlddev.2008.08.005 Fletschner, D., & Kenney, L. (2014). Rural women's access to financial services: Credit, savings, and insurance. In Gender in agriculture (pp. 187–208). Springer. https://doi.org/10.1007/978-94-017-8616-4_8 Fonseca, R., Mullen, K. J., Zamarro, G., & Zissimopoulos, J. (2012). What explains the gender gap in financial literacy? The role of household decision making. The Journal of Consumer Affairs,46(1), 90–106. https://doi.org/10.1111/j.17456606.2011.01221.x Francis, E., Blumenstock, J., & Robinson, J. (2017). Digital credit: A snapshot of the current landscape and open research questions. CEGA White Paper, 1739–1776. https://people.ucsc.edu/jmrtwo/digital_credit_landscape.pdf 288 JOHNEN AND MUßHOFF Freeman, H. A., Ehui, S. K., & Jabbar, M. A. (1998). Credit constraints and smallholder dairy production in the East African highlands: Application of a switching regression model. Agricultural Economics,19(1–2), 33–44. https://doi.org/10. 1111/j.1574-0862.1998.tb00512.x FSD Kenya, CBK, & KNBS. (2009). FinAccess Household Survey 2009. FSD Kenya, CBK, & KNBS. (2013). FinAccess household survey 2013. Harvard Dataverse. https://doi.org/10.7910/DVN/ JSPE9W FSD Kenya, CBK, & KNBS. (2019). 2019 FinAccess household survey. Ghosh, S., & Vinod, D. (2017). What constrains financial inclusion for women? Evidence from Indian micro data. World Development, 92,60–81. https://doi.org/10.1016/j.worlddev.2016.11.011 Gubbins, P., & Totolo, E. [E.] (2018). Digital credit in Kenya: Evidence from demand side-survey. Kenya. Guerin, I. (2006). Women and money: Lessons from Senegal. Development and Change,37(3), 549–570. https://doi.org/10. 1111/j.0012-155X.2006.00490.x Gwer, F., Odero, J., & Totolo, E. [. E.]. (2019). Digital credit audit report: Evaluating the conduct and practice of digital lending in Kenya. FSD Kenya. Hannig, A., & Jansen, S. (2010). Financial inclusion and financial stability: Current policy issues. https://doi.org/10.2139/ ssrn.1729122 Hoetker, G. (2007). The use of logit and probit models in strategic management research: Critical issues. Strategic Management Journal,28(4), 331–343. https://doi.org/10.1002/smj.582 Hoff, K., & Stiglitz, J. E. (1990). Introduction: Imperfect Information and rural credit markets—Puzzles and policy perspectives. The World Bank Economic Review,4(3), 235–250. https://doi.org/10.1093/wber/4.3.235 Hosmer, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied logistic regression. John Wiley & Sons. https://doi.org/10. 1002/9781118548387 Jack, W., & Suri, T. (2011). Mobile money: The economics of M-PESA. National Bureau of Economic Research Working Paper No. 16721. Advance Online Publication. https://doi.org/10.3386/w16721 Jack, W., & Suri, T. (2014). Risk sharing and transactions costs: Evidence from Kenya's mobile money revolution. American Economic Review,104(1), 183–223. https://doi.org/10.1257/aer.104.1.183 Jaffee, D., & Stiglitz, J. E. (1990). Credit rationing. In Handbook of monetary economics (Vol. 2, pp. 837–888). Elsevier. https://www.sciencedirect.com/science/article/pii/s1573449805800232,https://doi.org/10.1016/S1573-4498(05) 80023-2 Jia, X., Heidhues, F., & Zeller, M. (2010). Credit rationing of rural households in China. Agricultural Finance Review,70(1), 37–54. https://doi.org/10.1108/00021461011042620 Johnen, C., Parlasca, M., & Mußhoff, O. (2021). Promises and pitfalls of digital credit: Empirical evidence from Kenya. PLoS ONE,16(7), e0255215. https://doi.org/10.1371/journal.pone.0255215 Karaca-Mandic, P., Norton, E. C., & Dowd, B. (2012). Interaction terms in nonlinear models. Health Services Research,47, 255–274. https://doi.org/10.1111/j.1475-6773.2011.01314.x KNBS. (2019). 2019 Kenya population and housing census results. https://www.knbs.or.ke/?p=5621 Loges, W. E., & Jung, J.-Y. (2001). Exploring the digital divide. Communication Research,28(4), 536–562. https://doi.org/10. 1177/009365001028004007 Lycette, M., & White, K. (Eds.) (1989). Improving women's access to credit in Latin America and the Caribbean: Policy and project recommendations.KumarianPress. Mora, R. (2008). A nonparametric decomposition of the Mexican American average wage gap. Journal of Applied Econometrics, 23(4), 463–485. https://doi.org/10.1002/jae.1006 Moser, C. O. N. (1993). Gender planning and development: Theory, practice, and training. Routledge. https://www. taylorfrancis.com/books/mono/10.4324/9780203411940/gender-planning-development-caroline-moser Munyegera, G. K., & Matsumoto, T. (2016). Mobile money, remittances, and household welfare: Panel evidence from rural Uganda. World Development,79, 127–137. https://doi.org/10.1016/j.worlddev.2015.11.006 Robinson-Cimpian, J. P., Lubienski, S. T., Ganley, C. M., & Copur-Gencturk, Y. (2014). Teachers' perceptions of students' mathematics proficiency may exacerbate early gender gaps in achievement. Developmental Psychology,50(4), 1262– 1281. https://doi.org/10.1037/a0035073 Rubin, D. (1997). Estimating causal effects from large data sets using propensity scores. Annals of Internal Medicine, 127(127), 757–763. https://doi.org/10.7326/0003-4819-127-8_Part_2-199710151-00064 Sekabira, H., & Qaim, M. (2017). Mobile money, agricultural marketing, and off-farm income in Uganda. Agricultural Economics,48(5), 597–611. https://doi.org/10.1111/agec.12360 Suri, T., Bharadwaj, P., & Jack, W. (2021). Fintech and household resilience to shocks: Evidence from digital loans in Kenya. Journal of Development Economics,153(153), 102697. https://doi.org/10.1016/j.jdeveco.2021.102697 Swamy, V. (2014). Financial inclusion, gender dimension, and economic impact on poor households. World Development,56, 1–15. https://doi.org/10.1016/j.worlddev.2013.10.019 JOHNEN AND MUßHOFF 289 Tran, M. C., Gan, C. E., & Hu, B. (2016). Credit constraints and their impact on farm household welfare. International Journal of Social Economics,43(8), 782–803. https://doi.org/10.1108/IJSE-11-2014-0243 Weil, D., Mbiti, I., & Mwega, F. (2012). The implications of innovations in the financial sector on the conduct of monetary policy in East Africa. Report Submitted to the International Growth Centre Tanzania Country Program. https:// citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.259.7015&rep=rep1&type=pdf Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data (2nd ed.). MIT Press. How to cite this article: Johnen, C., & Mußhoff, O. (2023). Digital credit and the gender gap in financial inclusion: Empirical evidence from Kenya. Journal of International Development,35(2), 272–295. https://doi. org/10.1002/jid.3687 290 JOHNEN AND MUßHOFF APPENDIX A. TABLE A1 Correlation matrix (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) Age 1.000 Education 0.344 1.000 Female 0.032 0.141 1.000 Financial Literacy 0.196 0.529 0.102 1.000 Financial Provider Distance 0.138 0.262 0.032 0.149 1.000 Household Head 0.305 0.064 0.144 0.037 0.007 1.000 Household Size 0.092 0.151 0.031 0.100 0.144 0.180 1.000 Income 0.021 0.276 0.163 0.249 0.161 0.176 0.115 1.000 National ID 0.207 0.018 0.013 0.021 0.005 0.208 0.080 0.198 1.000 Married 0.064 0.024 0.059 0.007 0.003 0.102 0.258 0.150 0.137 1.000 Mobile Phone Ownership 0.129 0.291 0.067 0.206 0.161 0.115 0.076 0.257 0.219 0.113 1.000 Province 0.082 0.075 0.000 0.110 0.068 0.069 0.107 0.186 0.049 0.032 0.092 1.000 Urban 0.218 0.318 0.029 0.222 0.355 0.031 0.215 0.232 0.017 0.032 0.174 0.353 1.000 Note:N=7583. Authors' own calculations based on 2018 FinAccess Household Survey. JOHNEN AND MUßHOFF 291 TABLE A2 Robustness checks of Model 2 (formal digital credit) Variables Main model (1) LPM (2) Added variables (3) Interaction (4) All ages + Income imputed Independent variable Female (1 =yes) 0.270*** (0.099) 0.024*** (0.008) 0.203* (0.106) 0.114 (1.042) 0.286*** (0.093) Infrastructure access Mobile phone ownership (1 =yes) 2.552*** (0.717) 0.027*** (0.005) 2.524*** (0.718) 2.547*** (0.717) 2.562*** (0.641) National ID ownership (1 =yes) 1.520** (0.600) 0.049*** (0.013) 1.529** (0.606) 1.517** (0.600) 1.896*** (0.536) Financial institution distance (Base =Under 10 min) 10–30 min 0.338** (0.143) 0.027*** (0.009) 0.293** (0.146) 0.339** (0.143) 0.373*** (0.138) >30 min 0.205 (0.212) 0.015 (0.010) 0.147 (0.218) 0.205 (0.211) 0.250*** (0.212) Socio-demographics Age (years) 0.066*** (0.023) 0.002** (0.001) 0.059** (0.024) 0.067*** (0.023) 0.0755*** (0.020) Age 2 0.001*** (0.000) 0.000*** (0.000) 0.001*** (0.000) 0.001*** (0.000) .001*** (0.000) Income (ln of KSh) 0.329*** (0.059) 0.021*** (0.003) 0.310*** (0.061) 0.328*** (0.059) 0.325*** (0.058) Education (Base =No education) Some primary 1.505** (0.587) 0.005 (0.007) 1.529*** (0.590) 1.717** (0.776) 1.513*** (0.552) Primary completed 2.249*** (0.575) 0.024** (0.011) 2.262*** (0.578) 2.360*** (0.761) 2.257*** (0.535) Some secondary 2.804*** (0.584) 0.078*** (0.016) 2.824*** (0.585) 2.956*** (0.771) 2.663*** (0.545) Secondary completed 2.726*** (0.569) 0.083*** (0.013) 2.734*** (0.572) 2.753*** (0.750) 2.626*** (0.530) Technical/university 3.151*** (0.575) 0.171*** (0.018) 3.162*** (0.579) 3.254*** (0.751) 3.121*** (0.537) Province North Eastern/Coast 0.693*** (0.231) 0.100*** (0.025) 0.660*** (0.235) 0.698*** (0.231) 0.757*** (0.214) Eastern 1.278*** (0.211) 0.120*** (0.022) 1.212*** (0.218) 1.283*** (0.211) 1.406*** (0.205) Central 0.005 (0.187) 0.020 (0.029) 0.088 (0.199) 0.012 (0.186) 0.021 (0.174) Rift Valley 0.497*** (0.160) 0.083*** (0.022) 0.425*** (0.164) 0.501*** (0.160) 0.566*** (1.150) Western/Nyanza 0.437*** (0.168) 0.076*** (0.023) 0.369** (0.176) 0.439*** (0.168) 0.497*** (0.157) Added variables Married Not included Not included 0.054 (0.122) Not included Not included 292 JOHNEN AND MUßHOFF TABLE A2 (Continued) Variables Main model (1) LPM (2) Added variables (3) Interaction (4) All ages + Income imputed Household size Not included Not included 0.008 (0.030) Not included Not included Urban Not included Not included 0.159 (0.125) Not included Not included Household head Not included Not included 0.179 (0.124) Not included Not included Interaction Female * Education (2nd difference) Not included Not included Not included Not included No education vs. Primary 0.036* (0.020) No education vs. Some sec. education 0.064** (0.032) No education vs. Technical/ university 0.065*** (0.025) Secondary compl. vs. Technical/university 0.078*** (0.030) Constant 11.918*** (1.142) 0.141*** (0.037) 11.947*** (1.170) 12.010*** (1.230) 12.316*** (1.025) Observations 7583 7583 7583 7583 8561 Note:N=7583. Robust standard errors are shown in parentheses. Authors' own calculations based on 2018 FinAccess Household Survey. *p< 0.1. **p< 0.05. ***p< 0.01. TABLE A3 Robustness checks of Model 4 (semi-formal digital credit) Variables Main model (1) LPM (2) Added variables (3) Interaction (4) All ages + Income imputed Independent variable Female (1 =yes) 0.009 (0.106) 0.000 (0.007) 0.028 (0.118) 1.176* (0.613) 0.098 (0.0966) Infrastructure access Mobile phone ownership (1 =yes) 1.158*** (0.269) 0.031*** (0.007) 1.140*** (0.271) 1.182*** (0.268) 1.521*** (0.279) National ID ownership (1 =yes) 0.724** (0.325) 0.046*** (0.012) 0.689** (0.322) 0.726** (0.326) 0.425 (0.291) Financial institution distance (Base =Under 10 min) 10–30 min 0.434*** (0.139) 0.029*** (0.007) 0.353** (0.143) 0.423*** (0.139) 0.473*** (0.132) >30 min 0.182 (0.206) 0.013 (0.011) 0.065 (0.209) 0.196 (0.206) 0.249 (0.207) Socio-demographics Age (years) 0.002 (0.024) 0.006 (0.021) (Continues) JOHNEN AND MUßHOFF 293 TABLE A3 (Continued) Variables Main model (1) LPM (2) Added variables (3) Interaction (4) All ages + Income imputed 0.012 (0.023) 0.003** (0.001) 0.015 (0.023) Age 2 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) Income (ln of KSh) 0.027 (0.042) 0.002 (0.003) 0.015 (0.041) 0.032 (0.042) 0.032 (0.042) Education (Base =No education) Some primary 0.859** (0.336) 0.005 (0.008) 0.797** (0.334) 0.382 (0.512) 0.809** (0.332) Primary completed 1.107*** (0.336) 0.015 (0.010) 1.036*** (0.334) 0.328 (0.471) 1.070*** (0.331) Some secondary 1.013*** (0.365) 0.009 (0.013) 0.947*** (0.362) 0.269 (0.504) 0.928*** (0.355) Secondary completed 1.425*** (0.343) 0.046*** (0.012) 1.336*** (0.342) 0.806* (0.478) 1.295*** (0.335) Technical/university 1.763*** (0.346) 0.087*** (0.016) 1.627*** (0.346) 1.185** (0.483) 1.694*** (0.336) Province North Eastern/Coast 0.476** (0.213) 0.030* (0.017) 0.562*** (0.216) 0.488** (0.214) 0.527*** (0.199) Eastern 0.864*** (0.199) 0.064*** (0.017) 0.928*** (0.211) 0.878*** (0.200) 0.821*** (0.193) Central 1.353*** (0.216) 0.134*** (0.028) 1.385*** (0.227) 1.374*** (0.216) 1.541*** (0.216) Rift Valley 0.686*** (0.222) 0.020 (0.014) 0.582** (0.228) 0.687*** (0.223) 0.744*** (0.216) Western/Nyanza 0.209 (0.205) 0.018 (0.015) 0.368* (0.214) 0.214 (0.206) 0.172 (0.197) Added variables Married Not included Not included 0.133 (0.128) Not included Not included Household Size Not included Not included 0.061** (0.029) Not included Not included Urban Not included Not included 0.173 (0.137) Not included Not included Household Head Not included Not included 0.065 (0.144) Not included Not included Interaction Female * Education (2nd Difference) Not included Not included Not included No education vs. Primary 0.042* (0.025) No education vs. Some sec. education 0.037 (0.027) 294 JOHNEN AND MUßHOFF TABLE A3 (Continued) Variables Main model (1) LPM (2) Added variables (3) Interaction (4) All ages + Income imputed No education vs. Technical/ university 0.005 (0.029) Secondary compl. vs Technical/university 0.044* (0.025) Constant 5.112*** (0.667) 0.043 (0.034) 5.111*** (0.676) 4.419*** (0.705) 5.206*** (0.581) Observations 7583 7583 7583 7583 8561 Note:N=7583. Robust standard errors are shown in parentheses. Authors' own calculations based on 2018 FinAccess Household Survey. *p< 0.1. **p< 0.05. ***p< 0.01. TABLE A4 Gender and formal digital credit usage (propensity score matching) ATT Std. Err. tvalue Sample size treatment Sample size control 0.014 0.008 1.795 4411 2023 Note: Authors' own calculations based on the FinAccess household survey 2018 using nearest neighbour matching. ATT =Average treatment effect of treated with female being the treatment. Dependent variable =formal digital credit usage. Number of Blocks =6; balancing property is satisfied; region of common support is [0.221; 0.849]; standard errors are simulated with 1000 bootstrap replications. TABLE A5 Gender and formal digital credit usage (Blinder-Oaxaca decomposition) Coefficient Men 0.132*** (0.008) Women 0.081*** (0.006) Difference 0.051*** (0.008) Explained 0.031*** (0.003) Unexplained 0.020** (0.008) Note:N=7583. Robust standard errors are shown in parentheses. Included covariates are equal to Model 2. Authors' own calculations based on 2018 FinAccess Household Survey. *p< 0.1. **p< 0.05. ***p< 0.01. JOHNEN AND MUßHOFF 295