Financial Inclusion and Wealth Inequality:
Asset Accumulation or Predatory Lending?
By Demi Baks*
Supervisor: prof. dr. B.W. (Robert) Lensink
§8 January 2021
Abstract
In this thesis, I construct new composite indices to measure financial inclusion and financial availability, using factor analysis. Consequently, these indices are used to study the effect of broadening financial access and usage on wealth inequality, as measured by wealth Gini coefficient. The results indicate a positive relationship, which is weakened when predatory lending regulations prevent high cost of credit. Moreover, a negative relation with wealth per capita is found, as well as increasing low wealth population shares in more inclusive financial systems. This implies the current focus on boosting financial inclusion worldwide increases wealth inequality and lowers wealth levels, although follow-up research will be necessary to test whether these results remain valid over longer time periods and with more precise data.
Keywords: Financial inclusion · Wealth inequality · Development economics · Factor analysis
* Faculty of Economics and Business, University of Groningen, P.O. Box 800, Groningen, the Netherlands.
E-mail: d.baks@student.rug.nl. Phone: +31 6 23791145. Student Number: S2963485.
§ Faculty of Economics and Business, University of Groningen, P.O. Box 800, Groningen, the Netherlands.
E-mail: b.w.lensink@rug.nl. Phone: +31 50 3633712. Room: 5411-847 (Duisenberg Building).
2 1. Introduction
In 2017, almost one-third of the world population did not have an account at a financial institution or via a mobile money provider (Demirgüç-Kunt et al., 2018). This mostly concerns developing countries, where this percentage is 37 percent, although still 6 percent of high- income countries are also excluded from the financial system. These numbers led to the World Bank’s Universal Financial Access 2020 initiative, which aims at including 1 billion people into the financial system through account access (World Bank, 2015).
Although financial inclusion is not a Sustainable Development Goal (SDG), it is mentioned explicitly as a mean to achieve some goals and considered to play a key role in achieving many of the goals (Klapper et al., 2016). The SDGs have been endorsed by all member states of the United Nations (UN) in the 2030 Agenda for Sustainable Development (United Nations, 2015).
The achievement of these 17 ambitious goals is considered to create a better world and focus on economic, social, and environmental development.
Existing empirical literature shows several positive macroeconomic effects from advanced financial inclusion. For example, broader access to finance positively affects economic growth (Sahay et al., 2015; Kim et al., 2018). Also, financial stability may be gained, when supervision and regulation are present (Sahay et al., 2015; Danisman & Tarazi, 2020). Furthermore, enhanced financial inclusion is linked to reduced income inequality and lower poverty rates (Fouejieu et al., 2020; Park & Mercado, 2018).
Personal wealth creates economic security and resilience to overcome financial shocks. Also, it may serve as an additional income source and enables larger investments. Besides, most income Gini coefficients range between 0.3-0.5, whereas these coefficients for wealth are mostly between 0.6-0.8 (Davies et al., 2011). So, inequality levels are much higher for wealth than for income. Therefore, it is important to extend the empirical analysis beyond income inequality, to see if the equalizing effect of financial inclusion also holds for wealth measures.
Although global wealth levels are growing, not everyone benefits from this increase. In 2019, the richest 1% collectively owned 44% of global wealth, while the bottom 50% held less than 1% of total global wealth (Shorrocks, A. et al., 2019). Since the unbanked population are often the poorest in society, improvements in financial inclusion are likely to affect them the most.
The possibility to save at a financial institution might positively impact the ability to accumulate
assets (Ouma et al., 2017; Célerier & Matray, 2019; Stein & Yannelis, 2020). However,
expanded credit access to individuals who cannot afford this or against unreasonable conditions,
3 could lead to over-indebtedness and lower net wealth (Von Fintel & Orthofer, 2020). The relationship between financial inclusion and wealth inequality will be further explored by using worldwide unbalanced cross-county panel data over the time period 2014-2017.
Financial inclusion is measured by the construction of a new financial inclusion index using factor analysis. Variables related to outreach, usage, and access are taken from the Financial Access Survey (IMF, 2018) and the Global Findex Database (Demirgüç-Kunt et al., 2018).
They are jointly analysed for potential common factors, which results in two indices for financial inclusion and financial availability. Their effect on wealth inequality is examined using panel analysis, where wealth inequality is measured by the annual wealth Gini coefficient of the Credit Suisse Resource Institute (Shorrocks et al., 2014; 2017).
The results show a positive relationship between financial inclusion levels and wealth inequality;
countries where more people have access to and make use of financial products are associated with higher wealth Gini coefficients. This increasing impact of financial inclusion on wealth inequality is weakened when regulations prevent high cost of credit via interest rate caps.
Moreover, financial inclusion is negatively related with wealth per capita, indicating a lowering impact on countries’ average wealth. This wealth deteriorating effect of financial inclusion is also found when population shares within specific wealth ranges are examined.
However, these findings do not necessarily indicate that the current focus on expanding financial inclusion worldwide is unbeneficial for the poor. The short time period of the analysed data lowers the validity of these results for the longer term. Also, wealth data remains mostly based on combined datasets and statistical imputations, which increases the probability of measurement errors. Besides, future prospects and social security systems are not considered in the calculation of net wealth. Therefore, follow-up research will be necessary to test whether these results remain valid over longer time periods and with more precise data.
The remainder of this thesis is structured as follows. The following section reviews the available
literature on financial inclusion and wealth inequality more extensively, after which section 3
describes the multiple regression techniques and empirical methods used to study this
relationship. The fourth section contains a specific description of all variables used and their
sources, followed by descriptive statistics. The regression results are presented and discussed
in the fifth section, after which the final section will conclude this thesis and present suggestions
for future research.
4 2. Literature review
This section will analyse the existing literature on financial inclusion and wealth inequality.
First the relation between financial inclusion and the broader concept of financial development will be addressed, as well as both their multidimensional characters. This multidimensionality induced researchers to combine multiple indicators into a single financial inclusion index using distinct methods. After reviewing these methods, some recognized macroeconomic effects of financial inclusion are discussed. Lastly, a theoretical relationship between financial inclusion and wealth inequality is proposed through asset accumulation and predatory lending channels.
2.1 Financial Development
The core function of the economy’s financial sector is to harmonize savings supply and credit demand across agents and time. The financial system comprises roughly two tiers: financial institutions and financial markets, which are both influenced by the financial infrastructure present (World Bank, 2019). Financial institutions include the traditional banks, but also insurance companies or nonbank financial institutions. Financial markets focus on the trade in, for example, stocks, bonds, or financial derivatives. Additionally, the infrastructure of the financial system concerns regulation, payments and settlement systems implemented, or available information-sharing systems like credit-ratings.
In a perfectly competitive environment, the financial system would operate frictionless and prices would be set at the minimum, fair level. Although this is not the current situation in the real world, efforts are made to further approach this optimal efficiency ideal. Financial development regards this process of cost-reducing in the provision of financial services, for example regarding information or transaction costs, thus easing market imperfections.
Čihák et al. (2012) propose a multidimensional, broad approach, which defines financial
development as improvements in five key financial functions to the overall economy. These
functions assess possible investments and allocate capital accordingly, monitor and govern
agents towards which capital was allocated, facilitate agents with tools to trade, diversify and
manage their risk, mobilize and pool savings, and ameliorate exchanging goods, services and
financial instruments. As perfect measures of these functions for financial institutions and
financial markets are lacking, they are proxied by four characteristics: depth, access, efficiency,
5 and stability. This results into the structure of financial development presented in Figure 1, including each characteristic’s variable suggested for benchmarking.
For many years, empirical research mostly proxied financial development using variables related to financial depth (Beck et al., 2007). However, the forementioned multidimensional approach extends financial development measurement beyond financial depth; it presents financial access, efficiency, and stability as integral dimensions of financial development as well. Especially the financial institutions’ access dimension is closely related to the concept of financial inclusion, as it determines the broadening aspect of financial development by assessing a financial system’s capacity to serve the whole economy instead of solely large corporations and the wealthy.
2.2 Financial Inclusion
The World Bank considers financial inclusion to be the situation where “individuals and businesses have access to useful and affordable financial products and services that meet their needs – transactions, payments, savings, credit and insurance – delivered in a responsible and sustainable way”. Their Universal Financial Access 2020 initiative urged governments, central
Financial Development
Financial Institutions
Depth -
Privatesector credit to GDP
Access -
Accountsper 1,000 adults
Efficiency - Net interest margin
Stability - Z-score
Financial Markets
Depth - Stock market capitalization plus domestic private debt securities to GDP
Access - Percent of market capitalization outside of top ten largest companies
Efficiency - Turnover ratio for stock market
Stability - Volatility of stock price index
Figure 1: Financial development structure, characteristics, and main benchmarking variables from Čihák et al. (2012).
6 banks, and financial institutions to focus on promoting and broadening financial inclusion worldwide (World Bank, 2015). The ultimate goal is enabling all adults to have access to finance, where the initiative’s commitment aims at an increase of 1 billion people by 2020.
Although owning a transaction account is first prioritized, it is also emphasized that account usage is not always a direct result to having account access. Therefore, account ownership does not fully capture a country’s financial inclusion degree and more should be considered.
It is also possible to define financial inclusion using its opposite, namely financial exclusion.
Four main categories of people who are financially excluded can be identified, both voluntary and involuntary (World Bank, 2014). People without a need for financial services voluntary exclude themselves from the financial system, while those with cultural or religious beliefs against these services might also choose to not participate. Involuntary exclusion appears when people are excluded from i.e., credit due to fair economic reasons, like those without any income, or it could result from market failures, such as discrimination, information asymmetry, or other non-economic barriers. This approach considers countries fully financially inclusive when this last category is successfully eliminated.
2.3 Financial inclusion indexing methods
Most researchers create a financial inclusion index by combining sub-indices which capture the interaction between causal variables within the multiple dimensions. An early paper proposing such an index is by Sarma (2008; 2011), who considers three dimensions: availability, access, and usage. Each dimension is measured with key variables, respectively the number of bank branches and the number of ATMs per 100,000 of population, the number of bank accounts per 1,000 persons, and the volume of credit plus deposit as proportion of a country’s GDP.
Dimension indices are separately calculated, after which the weighted average of these sub- indices results in the financial inclusion index. Equal weights are adjusted for lower data adequacy of some variables, so rather arbitrary.
This indexing method is mostly based on the widely known Human Development Index
calculation (UNDP, 2018). However, its weighting approach is rather arbitrary and assumes
perfect substitutability between dimensions. Parametric weighting methods can be used to
overcome these problems and create a more data-driven weighting scheme. There is much
diversity in methods used, but also in dimensions choice and which determinants are used to
measure these dimensions.
7 Cámara and Tuesta (2014) first determine theoretically three dimensions and its indicators, after which two-step Principal Component Analysis (PCA) is used to determine the weighting schemes. Additionally, they emphasize the benefits of combining the mostly used supply-side information with demand-side information. The access dimension is still constructed with supply-side information, specifically the numbers of branches and ATMs per 100,000 of the population and per 1,000 kilometres squared, but the usage dimension uses demand-side information on the percentages of the population who owns at least one financial product, who keeps savings in a formal financial institution, and who has a formal loan. Lastly, they include a barrier dimension, which focuses on the unbanked population’s arguments for involuntary exclusion. The sub-indices are combined into one financial inclusion index, using PCA again.
The coefficient of variation (CV), which equals the standard deviation divided by the mean value, is used by Wang and Guan (2017). This method sets each indicator’s weight equal to the ratio of its CV to the total sum of CVs and is very similar to the approach of Cámara and Tuesta (2014). Two dimensions are considered: access and usage. Access is proxied by debit card ownership, account ownership, ATMs per 100,000 adults and branches of commercial banks per 100,000 adults. The usage dimension is assessed using percentage of people who used checks to make payments, who used electronic payments, and who saved or had a loan at a financial institution, as well as outstanding deposits and outstanding loans relative to GDP.
Both dimensions are thus determined using both supply- and demand-side information. After each dimension’s indicator is calculated, these are combined into one single financial inclusion index using CV again.
Although the weighting schemes are now data-driven, the separation of indicators into dimensions is still solely theoretical. Mialou et al. (2017) use factor analysis to establish the separate dimensions and their weights, whereafter the index is calculated by non-linearly aggregating the dimension indicators. This method recognizes a latent structure in correlated indicators’ variations, and determines weights endogenously based on these covariations.
The three dimensions considered are outreach, usage, and quality of financial services. The
outreach dimension is similarly measured as in Sarma (2008), using the numbers of branches
and ATMs in a country. However, these are now geographically rescaled to number per 1,000
kilometres squared, because one of the main arguments for financial exclusion is the proximity
of the financial sector . The usage dimension is measured by the total numbers of household
depositors and household borrowers per 1,000 adults. The quality dimension is disregarded in
the index computation for data scarcity reasons.
8 2.4 Macroeconomic effects
A financial inclusion index helps to investigate its effect on several macroeconomic and social- economic variables. Although research in this area is not very extensive, multiple studies have considered effects on, i.e., economic growth, financial stability, poverty, and wealth accumulation. Furthermore, the impact of financial inclusion on inequality is also part of the literature, but these studies focus mostly on income inequality instead of wealth inequality.
A positive effect of financial inclusion on economic growth is found by Sahay et al. (2015), where additional controls for financial depth also indicate that high financial depth levels are associated with declining marginal growth benefits of financial inclusion. So positive effects are larger in countries with low financial inclusion and less financial depth, and vice versa. The quasi-experimental situation following the recent development in Islamic financing, which lowered Muslims’ voluntary financial exclusion, led to a confirmation of this positive relationship between financial inclusion and economic growth in Organisation of Islamic Cooperation (OIC) countries (Kim et al., 2018).
Moreover, the multiple efforts to increase a country’s financial system’s inclusiveness affect financial stability differently. Increased access to automated teller machines (ATMs), branches, banking accounts, and improving depositors’ diversity are all measures stimulating economic growth without weakening financial stability (Sahay et al., 2015). Additionally, the stability impact of financial inclusion is stronger among young, unemployed, undereducated adults and in rural areas (Danisman & Tarazi, 2020). However, credit expansion is associated with mixed stability effects, where supervision and regulation strengths indicate either stability losses or gains from broadening access to credit (Sahay et al., 2015).
As financial inclusion focuses on more equal access to financial services, it is also interesting
to investigate who benefits from the increased economic growth. A negative relationship is
found between financial inclusion and income inequality measured by income Gini coefficient
(Park & Mercado, 2018; Fouejieu et al., 2020). This relationship is stronger in more stable
financial systems, and weakened by tightening monetary policy (Fouejieu et al., 2020). Besides,
when considering the poverty headcount ratio instead of income Gini coefficient, the negative
relationship with financial inclusion holds (Park & Mercado, 2018). This indicates increased
financial inclusion not only leads to a more equal income distribution, but also helps increase
the incomes of the least-well off.
9 2.5 Wealth inequality
The empirical focus has been on income inequality, when evaluating the effect of broadening financial inclusion on inequality. However, income is not the only or all-encompassing proxy for economic well-being; wealth is at least as important as income (Headey & Wooden, 2004).
Income depicts only the money flow into households, providing no information on their expenses and financial obligations, whereas wealth combines information on households’
(non-)financial assets and their liabilities (Davies et al., 2017).
Positive net wealth creates economic security and resilience to overcome financial shocks. Also, it may serve as an additional income source, for example through interest on savings or dividend payments on stocks. Furthermore, wealth enables larger investments, for example in personal education or by buying a house (Keeley, 2015). Especially these more long-term benefits are the main motivation behind the efforts to enhance financial inclusion (World Bank, 2015).
Wealth data availability has improved over the last decade, which has led to a few empirical studies beyond income inequality. For a ratio comparing countries’ mean wealth with their top wealth levels, Fouejieu et al. (2020) find a negative correlation with financial inclusion. This implies wealth and income might have a similar relationship with financial inclusion, as higher financial inclusion might be associated with an enhanced ability to accumulate wealth.
Three quasi-experimental studies indeed find such a positive relation between financial inclusion and wealth accumulation. Mobile money introductions in Africa led to both rising numbers of savers and increased saving amounts (Ouma et al., 2017). Also, in the developed world, there have been developments that provided a positive shock to financial inclusion. The U.S. interstate branching deregulation exogenously increased households’ financial inclusion, who then invested more in assets (Célerier & Matray, 2019). Also, especially low-income households showed a significant improvement in their ability to accumulate wealth. A similar study used the creation of the Freedman’s Savings Bank as an exogenous increase of financial inclusion among free Blacks. They also found that households owning a financial account had higher income, business ownership, and real estate wealth (Stein & Yannelis, 2020).
Nevertheless, another recent study on inequality in South-Africa found a positive relationship
between financial inclusion and wealth inequality (Von Fintel and Orthofer, 2020). Improved
inclusion indeed enabled the middle class to increase their savings and thereby improve their
wealth position. Adversely, the poor saw their wealth shares reduced, as expanded access to
10 credit led to predatory lending
1. Such a pattern of large debt increases for poorer households compared to significant saving by the wealthy is also mentioned by Saez (2017). This disparity has a stimulating effect on wealth inequality, which is likely to persist over time.
Although comparability between available studies is low due to differing definitions, measurements and coverage of financial inclusion and wealth inequality, there seem to be two channels through which financial inclusion affects wealth inequality. On the one hand, more financial inclusion is associated with asset accumulation via savings and beneficial investments, which may decrease wealth inequality. On the other hand, broadening access to credit may adversely affect the poorest’ net wealth if it leads to over-indebtedness of the poor and predatory lending. Supervision and regulation preventing excessive borrowing may mitigate such a positive relationship between credit expansion and wealth inequality.
3. Methodology
The proposed relationship is investigated using cross-country data for both financial inclusion and wealth inequality for the period 2014-2017. Since financial inclusion is a latent variable, available information on its dimensions is used to assemble a comprehensive financial inclusion index. Wealth inequality is analysed by countries’ wealth Gini coefficient. Furthermore, the analysis is extended to wealth per capita and the population distribution among wealth ranges to include absolute wealth levels. Multiple control variables are included, among which a dummy for the presence of predatory lending regulation. This section first discusses how the financial inclusion index is created, followed by the proposed econometric approaches to analyse its relationship with the wealth measures. Lastly, the potential endogeneity problem is addressed.
3.1 Financial inclusion index
Although the literature has not found consensus on a financial inclusion index construction method, there are three concepts which are featured in most methods: outreach, access, and usage. Whereas outreach demonstrates the financial system’s presence in a country, both access
1 Predatory lending refers to the situation where people are convinced, or even seduced, in taking a loan which they cannot afford. Often these loans have high interest rates and unreasonable conditions.
11 and usage indicate whether it is in fact widely attainable and utilized. Therefore, the former is closer related to financial availability, while the latter two best align with the World Bank’s financial inclusion definition previously discussed. They ensure financial inclusion indicates whether people have actual access to financial products and services, but also if these are useful, affordable and meet their needs. The hypothesized structure is graphically depicted in Figure 2.
Multivariate analysis is suited for analysing multidimensional datasets to detect patterns and relationships between several variables at once. Therefore, these statistical techniques are used to reduce complex, highly correlated data into fewer variables, which is quite useful to construct composite indices. Such an index is preferred over using similar variables separately for simplicity reasons and to avoid multicollinearity.
The financial inclusion index will be constructed using factor analysis, which explores the data to find common factors in all observable variables. A relevant common factor represents a latent variable, for example a country’s financial system’s inclusion level, which influences the observed variables. Each variable’s unique variance explains its remaining individual deviation and is also able to capture possible measurement errors.
Factor analysis creates uncorrelated combinations for the original variables. The model assumes the existence of common and unique factors in the dataset that explain all indicators’ variance.
Consequently, every input variable can be depicted as a combination of the common factors and an error term. This can be modelled as
𝑥
𝑗= 𝛼
𝑗,1𝐹
1+ 𝛼
𝑗,2𝐹
2+ ⋯ + 𝛼
𝑗,𝑛𝐹
𝑛+ 𝑒
𝑗(1)
Figure 2: The hypothesized structure for financial inclusion and financial availability in this thesis.
12 where every 𝐹
𝑖represents a common factor, or latent variable, and 𝛼
𝑗,𝑖is the weight given to the 𝑖th factor in explaining the variation of the 𝑗th standardized variable. All remaining variance is explained by the unique variance 𝑒
𝑗. There are multiple methods to construct the factors, but the method of principal components is mostly used in index creation (OECD, 2008).
The number of relevant factors can be determined following several ‘rules. The most commonly used method is the Kaiser criterion, which only retains factors with eigenvalues of at least one (Kaiser, 1974). This indicates that the factor explains more variance than an individual indicator, thus, including the factor has added value.
The rotated factor loadings are used to determine how much variables load on each common factor and, accordingly, develop a weighting scheme for the composite index per factor. The index is estimated by the following regression method
𝐼
𝑖= 𝛼
1𝑥
1+ 𝛼
2𝑥
2+ ⋯ + 𝛼
𝑛𝑥
𝑛(2)
where 𝛼
𝑖indicates the weight for the 𝑖th variable 𝑥
𝑖, with 𝑖 = 1, … , 𝑛.
3.2 Panel analysis
As the same countries are observed over time, the data qualifies as panel data. There are multiple techniques to linearly regress panel data. Besides pooled ordinary least squares, a fixed effects or random effects model can be employed. A pooled regression assumes no uniqueness among observed individuals, whereas the latter two methods control for time-invariant heterogeneity between them. Such individual characteristics are often present in cross-country data, which is confirmed by rejecting the null hypothesis of zero individual effects in both the F-test and the Breusch-Pagan Lagrange multiplier (LM) test (Breusch & Pagan, 1980). The general estimated model then becomes
𝑦
𝑖,𝑡= 𝛽
0+ 𝛽
1𝐹𝐼
𝑖,𝑡+ 𝛽𝑥
𝑖,𝑡′+ 𝑑
𝑡+ 𝑐
𝑖+ 𝑢
𝑖,𝑡(3)
where the dependent variable 𝑦
𝑖𝑡is country 𝑖’s wealth Gini coefficient at time 𝑡. Furthermore,
𝐹𝐼
𝑖,𝑡is its financial inclusion index, 𝑥
𝑖𝑡′represents all control variables, including country-
specific characteristics and macroeconomic determinants, 𝑑
𝑡is a time-dummy, 𝑐
𝑖is the
13 unobserved country-specific effect, and 𝑢
𝑖,𝑡is the error term. The estimated parameter 𝛽
1indicates the investigated effect of financial inclusion on wealth inequality.
The assumptions regarding the country-specific variation in observations differ for the fixed effects and the random effects model. The fixed effects model considers countries’ time- invariant individual characteristics as part of the explanatory variables, whereas the random effects model assumes these characteristics as part of the error term. Therefore, the country- specific effect must be uncorrelated with the independent variables in the random effects model, but this rather strong assumption is not necessary in the fixed effects model.
The Hausman test compares the coefficients from the fixed effects regression with those from the random-effects regression, after which it can be determined which model best suits the data (Hausman, 1978). The random effects model is more efficient under the null hypothesis of no systematic difference, but under the alternative hypothesis only the fixed effects model is consistent. Since the test fails to reject the null hypothesis, with a chi-squared test statistic of 0.85 and a p-value of 0.999, the random effects model is estimated for wealth inequality.
3.3 Absolute wealth levels
The Gini coefficient describes the distribution of wealth generally well and also allows negative values, yet it remains a relative summary measure. Therefore, the analysis is extended to other wealth measures as well, which better examine changes in countries’ ability to accumulate wealth.
First, absolute wealth levels are not considered in the Gini coefficient, so it says little about how wealthy countries are. A country with extremely low wealth can still have an equal wealth distribution, hence a low Gini, while unequal countries may be very wealthy. The analysis of wealth per capita provides more information on wealth levels themselves and if more financial inclusion actually leads to wealth accumulation.
Moreover, the Gini coefficient summarizes the whole wealth distribution into one ratio, which
does not provide a full comparison. For example, two completely different wealth distributions
can lead to the same Gini coefficient. Data on countries’ population spread between low wealth,
middle wealth and high wealth ranges better describes the distributional effects within countries
from increased financial inclusion.
14 Again, country-specific characteristics prove to be present, so the same model is used as presented in equation 3. However, 𝑦
𝑖,𝑡now represents either wealth per capita or the population share within a wealth range, and the Hausman test rejects the null hypothesis of no systematic difference in coefficients. Therefore, these regressions are estimated using the fixed effects model.
3.4 Predatory lending
Financial inclusion enables day-to-day payments and savings opportunities, but it also includes improved access to credit for households and businesses. Borrowing affects net wealth through its liabilities component, but may also create wealth, when used for profitable purposes.
However, when loans are granted to people who cannot afford them or against unreasonable conditions, increased financial inclusion might negatively affect wealth.
Effective regulation may reduce such predatory practices, but not all countries have this.
Therefore, the model is extended with a dummy variable to account for regulations preventing predatory lending, proxied by the use of interest rate caps, and its cross-product with financial inclusion. This interaction term captures the regulation environment’s effect on the relationship between financial inclusion and wealth inequality. The model then becomes
𝑦
𝑖𝑡= 𝛽
0+ 𝛽
1𝐹𝐼 + 𝛽
2𝐹𝐼 ∗ 𝑟
𝑖+ 𝑥
𝑖𝑡′𝛽 + 𝑑
𝑡+ 𝑟
𝑖+ 𝑐
𝑖+ 𝑢
𝑖𝑡(4)
where 𝑟
𝑖is the dummy for regulations preventing predatory lending, which is 1 if interest rate caps are used in country 𝑖 and 0 otherwise.
3.5 Endogeneity
Both financial inclusion and wealth inequality are related to individuals’ financial situation.
This may cause simultaneity, when both are affected by an outside influence, or reversed causality between wealth inequality and financial inclusion. For example, reduced wealth inequality might inspire political pressure to create a more inclusive financial system. Likewise, selection bias could be present when higher wealth leads to more demand for financial products.
Such an endogeneity problem, where an independent variable is correlated with the error term,
15 creates potential biases in the regression results. The method of instrumental variables overcomes these biases by replacing the endogenous variable with valid instruments. Validity is achieved when an instrument is both exogenous and sufficiently correlated with the endogenous variable it is replacing.
The instrumental variables used for financial inclusion are countries’ legal origin, following the law and finance theory, and the normalized absolute value of their capital city’s latitude, following the endowment theory (Beck et al., 2003; Beck & Levine, 2005; Yermack, 2018).
Since these variables are constant over time, the standard two-stage least squares regressions are estimated for the most recent year only. The instruments’ relevance is confirmed in the first stage regression with an F-test, while their exogeneity is proven by the Sargan test of overidentifying restrictions (Sargan, 1958).
Assuming valid instruments, it is also possible to test whether endogeneity is in fact present in the model. The Durbin (1954) and Wu-Hausman (Wu, 1974; Hausman, 1978) statistics reject the null hypothesis of exogeneity for the wealth Gini coefficient, but fail to do so for the models where wealth per capita and population shares in wealth ranges are the dependent variables.
Therefore, only the regression on wealth inequality is estimated using instrumental variables.
Although this method addresses the potential endogeneity problem for financial inclusion, it cannot overcome possible endogeneity of the other explanatory variables. Also, the data’s time- series dimension is not utilized anymore, while wealth inequality is likely to have some time dependence as well. Therefore, most researchers use a generalized-methods-of-moments (GMM) estimation for dynamic panel data (Beck et al., 2007; Kim et al., 2018; Islam &
McGillivray, 2020; Fouejieu et al., 2020). For example, the method by Arellano and Bond (1991) uses the explanatory variables’ lagged values as instruments instead of external instruments. However, the limited data availability over time prevents using this method in this thesis.
4. Data and descriptive statistics
This section describes the raw variables, regarding countries’ financial systems and wealth
measures, as well as the statistical processes to accommodate the analyses described in the
previous section. Lastly, the chosen or constructed variables will be characterized using
summary statistics.
16 4.1 Financial inclusion and availability
As with financial inclusion itself, its dimensions cannot be directly observed and measured, but it can be inferred from data on related variables. As discussed in section 3.1, the construction of the financial inclusion index is based on three dimensions: outreach, access, and usage. There are multiple databases regarding financial development, which can be used to collect information on variables within these three dimensions.
The outreach variables are from the supply-side data collected in the Financial Access Survey of the International Monetary Fund (2018). The data come from central banks and other financial regulators and include information on 189 countries since 2009. The variables of interest are the number of bank branches and the number of ATMs in a country. Both variables are measured per 1,000 kilometres squared to best proxy geographic availability. For the access and usage variables, the demand-side Global Findex Database (Demirgüç-Kunt et al., 2018) is used. This database is published triennial since 2011, and data is collected by means of surveying over 150,000 adults in more than 140 countries worldwide. The aggregate numbers are nationally representative, enabling comparison over time and between countries. The variables of interest indicate the percentages of total population aged over 15 who own an account, possess a debit or credit card, save, borrow, deposit money, withdraw money, make payments, and receive payments. Their full descriptions are included in appendix A.
Unfortunately, the survey questions regarding deposits, withdrawals, and payments were not yet included in the Global Findex’s 2011 edition. Additionally, its triennial update of 2020 has not yet been published. Therefore, only data concerning the years 2014 and 2017 are analysed in this research. For these years, the mentioned variables’ descriptive statistics are presented in Table 1.
The validity of combining variables into a single index can first be assessed with the correlation
matrix. High correlation between indicators hint at the possibility of shared common factors,
which supports constructing an index from these variables. The correlation matrix in Table 2
shows a moderate positive relationship between the outreach variables, which have a low to
negligible correlation with access and usage variables. Also, there exists very high correlation
among most access and usage variables, though this is more moderate for the variable
associated with borrowing.
17
Variable Obs Mean Std. Dev. Min Max
Branches per 1,000 km
2317 30.380 36.825 0.039 111.823
ATMs per 1,000 km
2316 44.523 53.686 0.027 165.651
Account ownership 267 0.598 0.289 0.064 1.000
Debit card ownership 267 0.435 0.313 0.005 0.988 Credit card ownership 267 0.194 0.208 0.000 0.826
Saved 267 0.241 0.194 0.009 0.793
Borrowed 267 0.125 0.073 0.004 0.405
Withdrawn money 257 0.487 0.299 0.055 0.990
Deposited money 257 0.479 0.294 0.048 0.987
Made payments 267 0.440 0.306 0.007 0.989
Received payments 267 0.391 0.256 0.028 0.929
Table 1: Summary statistics for the financial system outreach, access, and usage variables. Detailed variable definitions and sources can be found in appendix A.
Branches per 𝐤𝐦𝟐 ATMs per 𝐤𝐦𝟐 Account Debit card Credit card Saved Borrowed Withdrawal Deposit Made payments Received payments
Branches
per km2 1.00
ATMs per
km2 0.60 1.00
Account 0.09 0.35 1.00
Debit card 0.01 0.31 0.93 1.00
Credit card 0.16 0.39 0.77 0.80 1.00
Saved 0.17 0.39 0.82 0.83 0.83 1.00
Borrowed -0.11 0.12 0.48 0.46 0.49 0.50 1.00
Withdrawal 0.05 0.35 0.96 0.96 0.83 0.85 0.52 1.00
Deposit 0.05 0.35 0.96 0.96 0.82 0.85 0.51 0.99 1.00
Made
payments 0.12 0.33 0.93 0.92 0.84 0.86 0.49 0.92 0.92 1.00
Received
payments 0.03 0.26 0.93 0.92 0.79 0.83 0.52 0.94 0.95 0.96 1.00
Table 2: The correlation matrix for the financial system outreach, access, and usage variables.
18 Additionally, the variables’ correlation structure is tested for diagonality and sphericality through two multivariate tests. Both likelihood-ratio tests with first-order Bartlett correction strongly reject the null hypothesis of independent variables. Therefore, a factor analysis is conducted for these variables.
There are two factors which have an eigenvalue higher than one, so the other factors are not retained (Kaiser, 1974). The rotated factor loadings indicate how much a variable ‘loads’ on that factor, which are displayed in Table 3 below. Rotation maximizes the individual indicators’
loading on individual factors, which improves their interpretability. The results confirm the hypothesized separation of outreach variables from access and usage variables since they load on different factors.
Variable Factor 1 Factor 2 Uniqueness
Branches per 1,000 km
2-0.016 0.916 0.161
ATMs per 1,000 km
20.285 0.837 0.219
Account ownership 0.951 0.109 0.083
Debit card ownership 0.957 0.041 0.082
Credit card ownership 0.861 0.200 0.218
Saved 0.886 0.202 0.174
Borrowed 0.602 -0.179 0.606
Withdrawal 0.977 0.078 0.039
Deposit 0.977 0.081 0.039
Made payments 0.955 0.117 0.074
Received payments 0.965 0.018 0.069
Eigenvalue 7.678 1.558
Table 3: The rotated factor loadings and uniqueness per indicator variable resulting from the factor analysis.
The Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy is calculated to verify the variables have enough in common for factor analysis application. The value of 0.91 indicates high commonality between the variables, which is considered ‘marvellous’ (Kaiser, 1974).
Another assessment method is the Cronbach’s Alpha, which measures how closely related a
group of variables are, so their internal consistency. Respectively, the two factor alphas are 0.97
and 0.74, which suggest the observable variables define the latent variables well and factor
analysis is fit to analyse the data.
19 The composite indices for financial inclusion and financial availability are constructed via the regression method depicted in equation 2. For each observable variable, the factor rotated loading’s squared value represents how much of the total unit variance is explained by the latent variable. These values can be interpreted as each indicator’s weight and are shown in Table 4.
Variable Financial inclusion Financial availability
Branches per 1,000 km
20.02 0.58
ATMs per 1,000 km
20.05 0.50
Account ownership 0.12 -0.02
Debit card ownership 0.12 -0.06
Credit card ownership 0.11 0.05
Saved 0.12 0.04
Borrowed 0.07 -0.17
Withdrawn 0.13 -0.04
Deposited 0.13 -0.04
Made payments 0.13 -0.02
Received payments 0.12 -0.08
Table 4: Regression scoring coefficients used to construct financial inclusion and financial availability indices.
For comparison, the standardized indices are normalized to generate a financial inclusion and financial availability index which range between 0 and 1. These are presented in ranking order in appendix C. However, for use within a regression the composite indices 𝐼
𝑐are transformed to lie between −∞ and ∞ and be normally distributed (Wang & Guan, 2017). The transformation is as follows
𝑌 = ln (
𝐼𝑐1−𝐼𝑐
) (5)
4.2 Wealth measures
The Credit Suisse Resource Institute publishes the Global wealth report on a yearly basis since
2010. These reports and corresponding databooks contain information regarding global
household wealth levels and its distribution between and within countries for countries
worldwide. Reliable wealth data is low in availability, but almost all high-income countries and
the largest countries, including India and China, have available wealth data. Their ultimate data
20 coverage is high, through combining multiple datasets and statistically imputing missing data.
Although this lowers the trustworthiness of the provided data, it can still be considered relatively trustworthy results, as long as better alternatives lack.
In the reports, wealth, or net worth, is defined as “the marketable value of financial assets plus non-financial assets (principally housing and land) less debts” (Davies et al., 2017). They exclude public pensions, and measure wealth per individual adult. The data availability for financial inclusion variables restricts the time period for wealth inequality as well. Therefore, only the wealth distribution data for the years 2014 and 2017 are used in this research (Shorrocks et al., 2014; 2017).
The paper by Davies et al. (2017) explains how the wealth estimates are calculated by using a three-step approach. The first step combines household balance sheet data with household survey data to determine average wealth levels per country. This already covers 66 percent of global population, after which countries that lack this data are given estimates using econometric techniques.
The second step evaluates wealth patterns within countries more closely. For 33 countries
2, direct wealth distribution data is available, including most high-income countries and the most populous developing countries: India, Indonesia, and China. The distribution data was transformed into the form of cumulated wealth shares, which can be used to construct a Lorenz curve. Two large datasets used are the ECB’s Household Finance and Consumption Survey and the Luxembourg Wealth Study. For 135 other countries, the relationship between wealth distribution and income distribution is used to estimate their wealth distribution. For the 33 countries with both income and wealth distribution data, a distinction was made between countries in North America and Europe, and countries in the rest of the world. Their average wealth to income ratios at the percentiles 10, 20, …, 90 were computed per region. These ratios were then used to scale up the income Lorenz curves into a wealth Lorenz curve. The 33 countries with both data on both distributions served as a control group for these ratios, and the estimates appeared close to the reported values.
Thirdly, the wealth distribution pattern is adjusted in the top-tail because these individuals are likely to be under-represented and their wealth undervalued. Information regarding the
2 An overview table with all 33 countries can be found in appendix B.
21 wealthiest persons worldwide is published by, among others, Forbes Magazine, which are used to estimate wealth patterns more accurately within countries.
The estimated Lorenz curves indicate the cumulative proportion of the populations’ total wealth held by the bottom 𝑥 percentage share of the population. A 45-degree line from the origin represents full equality, where i.e., 40 percent of the population holds 40 percent of total wealth and so on. The Gini coefficient expresses the ratio of the area between this equality line and the Lorenz curve over the total area under the line of equality. The closer the Lorenz curve is to the equality line, the smaller this Gini coefficient is, with a Gini coefficient of zero for full equality and a Gini coefficient of 1 indicating full inequality. Again, since the Gini coefficient is also a ratio between 0 and 1, its value is transformed before estimating the regression model. The same transformation is used as for the financial inclusion and financial availability indices, which is presented in equation 5.
Besides information regarding wealth inequality, the Global wealth databooks also contain data on absolute wealth levels. It shows countries’ wealth levels per capita and the distribution of adults (%) by wealth range (USD). There are three household wealth ranges: under 10,000 USD, 10,000-100,000 USD, and over 100,000 USD. These numbers indicate wealth patterns within and between countries beyond inequality.
Additionally, the percentage of adults whose net worth is less than 10,000 USD provides some indication of asset-based poverty in a country. Generally, poverty is defined as income insufficiency, but it may also be viewed as a lack of sufficient wealth resources to afford basic needs over a certain period (Shapiro & Wolff, 2001; Brandolini et al. 2010). Nevertheless, setting a universal wealth threshold has been proven difficult, for example due to differences in cost of living.
4.3 Control variables
Multiple macroeconomic control variables are included in the regressions to control for
influences on wealth inequality apart from financial inclusion. Berisha and Meszaros (2020)
present three macroeconomic determinants of wealth inequality dynamics: economic growth,
inflation, and interest rates. The arguments for their positive or negative relationship are related
to the notion that household in the middle three quintiles of the wealth distribution have often
high debt-to-income ratios compared to the poorest and the richest quintiles.
22 Economic growth is proxied using the annual GDP per capita growth rate from the World Bank.
It is expected to especially increase less-wealthy households’ net wealth, as an increase in income could either help pay off debt faster or give more room to save. Therefore, positive economic growth is expected to decrease wealth inequality.
Inflation is measured as the annual change in the cost of acquiring an average basket of goods and services. A higher inflation rate lowers the real value of outstanding debt. Again, this especially helps the lower and middle classes who have relatively higher debt-to-income ratios.
At the same time, inflation increases housing prices, which are often their primary assets.
Therefore, also inflation is expected to be negatively related with wealth inequality.
Contrarily, higher interest rates increase the debt burden via higher interest payments, which could lead to higher wealth inequality. However, low interest rates also decrease the return on savings of the wealthiest, which could mitigate such an increasing effect on wealth inequality.
Interest rates are proxied by the commercial bank prime lending rate, which indicates the interest charged on new loans for the most credit-worthy customers.
Besides these macroeconomic influences, a year dummy is included to control for time-effects, while the possible negative relationship with human capital is investigated using a proxy of mean years of schooling (Hasan et al. 2020). For the analyses regarding wealth per capita and population shares, population growth is included as control variable as well.
Moreover, the presence of predatory lending restrictions is added in separate regressions to test whether regulation changes the investigated relationship. Regulation regarding predatory lending varies widely between countries, but interest rate caps are a common measure to prevent unreasonable cost of credit. Worldwide, 74 countries restrict lending rates via at least some form of ceiling, which are often imposed with the intention to protect consumers (Ferrari et al.
2018). Therefore, the presence or absence of interest rate caps is used to proxy predatory lending regulation. The dummy variable equals one when regulation is present, whereas it is equal to zero if regulation is absent.
4.4 Descriptive statistics
The descriptive statistics of the original variables are presented in Table 5, for the full dataset
of unbalanced panel data. Data availability varies among the variables and is lowest for the
information regarding a country’s financial system’s inclusion and availability.
23
Variable Obs Mean Std. Dev. Min Max
Financial inclusion index 234 0.42 0.29 0.00 1.00
Financial availability index 234 0.37 0.23 0.00 1.00
Wealth Gini coefficient 338 0.71 0.10 0.31 0.94
Wealth per capita 338 57,089.86 102,941.20 114.00 587,649.00
% with low wealth 338 0.66 0.30 0.02 1.00
% with medium wealth 338 0.25 0.20 0.00 0.83
% with high wealth 338 0.10 0.17 0.00 0.72
Control variables
GDP per capita growth 335 1.97 2.67 -7.35 7.77
Inflation rate 340 4.78 7.45 -1.54 47.78
Interest rate 323 10.95 7.82 0.75 52.10
Schooling 336 8.57 3.13 1.40 14.10
Population growth 338 1.37 1.33 -4.54 6.74
Regulations 340 0.44 0.50 0.00 1.00
Instrumental variables
English legal origin 167 0.32 0.47 0.00 1.00
French legal origin 167 0.43 0.50 0.00 1.00
German legal origin 167 0.04 0.19 0.00 1.00
Scandinavian legal origin 167 0.03 0.17 0.00 1.00
Socialist legal origin 167 0.18 0.39 0.00 1.00
Latitude 167 0.29 0.20 0.00 0.80
Table 5: Descriptive statistics for the full dataset of unbalanced panel data. The financial inclusion and financial availability indices are constructed in this research and rescaled using a min-max normalization. Low wealth indicates a wealth under 10,000 USD, medium wealth ranges between 10,000-100,000 USD, and high wealth implies individuals with wealth over 100,000 USD. All instrumental variables are dummy variables, as is the control variable regarding regulation, which indicates whether countries use interest rate caps. Detailed variable definitions and sources can be found in appendix A.
24 5. Results
The constructed financial inclusion and financial availability indices are used to investigate their influences on wealth inequality and absolute wealth levels. The regression results are reported and further described in the following section.
5.1 Wealth inequality
The results from the regression estimations on the wealth Gini coefficient are presented in Table 6. Although not all control variables are statistically significant, their coefficient signs are robust across estimations and in line with the predictions discussed above. Economic growth negatively affects wealth inequality, while increased interest rates are associated with rising inequality. Education has a significant negative effect, which implies education may help alleviate wealth inequality. The negative sign of the time dummy displays a negative trend over the time period between observations.
The results show positive and statistically significant coefficients for financial inclusion in the panel analyses, as well as in the instrumental variables regressions. This suggests a positive relationship, indicating that improved financial inclusion will lead to higher wealth inequality.
The positive effect of financial inclusion is even greater and more significant in the instrumental variable regressions, which are expected to be more robust for potential endogeneity problems.
Although the presence of regulations preventing predatory lending does not significantly affect wealth inequality, its presence weakens the increasing effect of financial inclusion.
Adversely, the coefficients associated with financial availability are not significant in any of the regressions. Therefore, it appears that wealth inequality is not influenced by whether the financial system is more present in a country, although proximity is often mentioned as a financial barrier. It could also signal that physically presence is becoming less important with the rise of mobile banking worldwide.
5.2 Absolute wealth levels
The analysis is extended from wealth inequality to absolute wealth levels. The estimated
equation remains similar, but now the dependent variable is wealth per capita instead of the
wealth Gini coefficient. Also, the Hausman test’s null hypothesis is rejected for this new
estimation so the first column of Table 7 presents the coefficients from the fixed effects model.
25 (Dependent variable: Wealth Gini coefficient)
VARIABLES (1) (2) (3) (4)
Financial inclusion (FI) 0.064* 0.105** 0.478*** 0.580***
(0.035) (0.047) (0.151) (0.151)
FI * Regulations -0.083* -0.196**
(0.045) (0.093)
Financial availability -0.016 -0.009 -0.032 -0.005
(0.030) (0.030) (0.043) (0.046)
GDP per capita growth -0.051*** -0.052*** -0.031 -0.037*
(0.012) (0.012) (0.021) (0.021)
Inflation rate 0.021 0.028 0.019 0.070
(0.031) (0.031) (0.078) (0.083)
Interest rate 0.054 0.070 0.340*** 0.359***
(0.063) (0.063) (0.126) (0.123)
Schooling -0.034 -0.026 -0.156*** -0.134**
(0.022) (0.022) (0.057) (0.057)
Time -0.065* -0.065*
(0.037) (0.037)
Regulations 0.086 0.132
(0.092) (0.112)
Constant 1.316*** 1.162*** 1.890*** 1.556***
(0.263) (0.273) (0.483) (0.494)
Observations 201 201 111 111
Number of countries 122 122
Estimation method RE RE IV IV
R-squared 0.11 0.15
Sargan test 0.95 0.99
Table 6: Effect of financial inclusion on wealth inequality. The robust standard errors are presented in parentheses.
Statistical significance is indicated with ***, ** , * for significance at the 1%, 5%, and 10% levels, respectively.
The dependent variable is the log-transformation of the wealth Gini coefficient. Financial inclusion and financial availability are the log-transformations of their respective normalized indices, following equation 5. GDP per capita growth equals the annual growth rate of real GDP per capita. Inflation rate equals the logarithm of the annual growth rate of the consumer price index. Interest rate equals the logarithm of the commercial bank prime lending rate. Schooling is the average years of schooling at time 𝑡. The time dummy equals 1 for the year 2017 and 0 for the year 2014. The regulations dummy equals 1 if interest rate caps are used and 0 otherwise. Specifications (1) and (2) are estimated using a GLS random-effects (RE) model with clustered standard errors at the country level.
The overall regression R-squared is reported. Specifications (3) and (4) are estimated for the year 2017 using a two-stage least squares instrumental variables regression (IV) with robust standard errors. Instrumental variables include legal origin dummies, the normalized absolute value of the capital city’s latitude, and their cross-products with the regulations dummy. For the IV regressions, the p-value for the Sargan test of overidentifying restrictions is reported.
Normally, a fixed effects model cannot accommodate time-invariant variables, which the regulations dummy for predatory lending is, but the Mundlak approach helps solving this problem. By including variables’ group-means, the random effects model can be used even when the explanatory variables are correlated with the country-specific effect (Mundlak, 1978).
The second column of Table 7 extends the model with controls for predatory lending regulations.
26 (Dependent variable: Wealth per capita)
VARIABLES (1) (2)
Financial inclusion (FI) -0.105* -0.116*
(0.057) (0.071)
FI * Regulations 0.017
(0.108)
Financial availability 0.008 0.006
(0.040) (0.042)
Population growth 0.066*** 0.064***
(0.023) (0.023)
GDP per capita growth 0.027* 0.027*
(0.014) (0.014)
Inflation rate -0.020 -0.020
(0.022) (0.023)
Interest rate -0.070 -0.067
(0.079) (0.085)
Schooling 0.237* 0.231
(0.137) (0.141)
Time -0.082* -0.080*
(0.042) (0.044)
Regulations 0.427**
(0.176)
Constant 7.532*** 9.908***
(1.288) (0.907)
Observations 201 201
Number of countries 122 122
Estimation method FE RE
R-squared 0.24 0.24
Table 7: Effect of financial inclusion on absolute wealth levels. The robust standard errors are presented in parentheses. Statistical significance is indicated with ***, ** , * for significance at the 1%, 5%, and 10% levels, respectively. The dependent variable is the logarithm of the wealth per capita. Financial inclusion and financial availability are the log-transformations of their respective normalized indices, following equation 5. Population growth equals the annual growth rate of mid-year population. GDP per capita growth equals the annual growth rate of real GDP per capita. Inflation rate equals the logarithm of the annual growth rate of the consumer price index. Interest rate equals the logarithm of the commercial bank prime lending rate. Schooling is the average years of schooling at time t. The time dummy equals 1 for the year 2017 and 0 for the year 2014. The regulations dummy equals 1 if interest rate caps are used and 0 otherwise. Specification (1) is estimated using a fixed-effects (FE) model with clustered standard errors at the country level. The within regression R-squared is reported.
Specification (2) is estimated using aGLS random-effects (RE) model with clustered standard errors at the country level. Following the Mundlak approach, variables’ group-means are included but not reported. The within regression R-squared is reported.
