Macroeconomics of microfinance on a sub-national level

Contents

1 Introduction

2 Data
2.1 Variable Description
2.2 Preliminary data analysis

3 Estimation Strategy

4 Results
4.1 Operational self-sufficiency
4.2 Payments at Risk

5 Conclusion

A Appendix

List of Tables

1 Summary statistics of institution, state and national data

2 Random-effects estimators for self-sufficiency

3 Fixed-effects estimators for PAR-30

4 List of states and institutions

1 Introduction

Microfinance is arguably one of the most popular and best-researched ap­proaches towards poverty eradication in the developing world. Ever since Muhammad Yunus founded the Grameen Bank of Bangladesh in 1983, there have been recurring waves of enthusiasm about microfinance’s potential to lift millions out of poverty through economic empowerment. The emergence of skeptical voices proclaiming the failure of microfinance as a means of socio­economic inclusion resulted in extensive debates over the ability of microfinance to reach poor and very poor people (Johnson (2009)),

The varying levels of success between microfinance institutions (MFIs) have inspired a whole new field of research dedicated to finding the determinants of MFI success. For proponents of 'traditional’ microfinance, the most important measure of success is social performance in terms of the number of people whose financial and social situation is improved through the issued loans. Attempts to quantify these livelihood improvements are being made in the form of social audits by social investors like Oikocredit (Lapenu and Ledesma (2011)),

The other pillar of an MFI’s success is its financial performance, which is no longer seen as necessarily conflicting with its social mission (Johnson (2009)), While the development community has been debating and evaluat­ing microfinance for decades, it has only recently attracted the attention of financial markets due to its success in terms of growth, financial inclusion of the marginalized, and return on investment. Apart from achieving high profit margins in a few cases, microfinance funds have been praised for their potential for reducing portfolio variability (Krauss and Walter (2009)),

Since performance indicators vary more widely across individual MFIs than across countries, it is little surprising that most research into the determinants of social and financial performance has been focused on 'best practices’ on the institution level. In doing so, however, the influence of macroeconomic and institutional factors has largely been neglected. The first to systematically analyze this relationship are Ahlin et al. (2011), whose paper serves as my main referenee. By merging data on 373 MFIs from 74 countries with country-level economic and institutional data over a period of twelve years, they are able to show a strong complementarity between MFI performance and the economy at large. More specifically, they find that MFIs have better chances of covering costs when per capita income is growing, but face difficulties expanding their client base in a context of high workforce participation. Critical indicators of MFI health such as loan default rates also appear to be influenced by the macroeconomic environment,

In the spirit of opening MFIs to the opportunities provided by investment funds, it is pertinent to ask how performance prognoses can be improved. Since the explanatory potential of MFI-level factors has been largely exploited, it is time to take a closer look into the macroeconomic determinants of MFI perfor­mance, First of all, MFIs rarely aspire to cover the entire national market for mierofinaneial services, instead opting to serve a specific region or to specialize in products like microsavings accounts or microinsurance. Second, their client base is mostly confined to the informal part of the economy, which is subject to regional influences and has been shown to move relatively independently from the official economy (Patel and Srivastava (1996)), This implies that the nationally surveyed figures are not representative of the economic environment in which the majority of microentrepreneurs operate. In order to amend this problem, this thesis tests the hypothesis that more precise estimates of MFI performance can be obtained by using data collected on a sub-national level,

I make this argument by essentially emulating the model proposed by Ahlin et al. (2011), but concentrating on Mexican and Brazilian MFIs and plugging in macroeconomic data from the federal states in which these MFIs operate. In a second step I repeat the regressions with national economic data and compare the results with a focus on the coefficients and goodness of fit.

The choice of Mexico and Brazil was informed by several factors, the fore­most being the availability of annual data for geographical sub-divisions. Both countries span several climatic zones and consist of socio-economically very diverse federal states. In contrast to other developing countries with these characteristics, such as India and Nigeria, both Mexico and Brazil conduct statistical surveys annually (for inflation and employment data even once a month) and on the state and municipality levels. Furthermore, their micro­finance sectors are rather competitive so that no single institution yields a dominant market share. Another aspect that renders these countries ideal for the purpose of this thesis is that federal law only allows registered banks to offer savings accounts. This means that MFIs; only source of funds, apart from the profits they generate, are donors and investors, who in turn expect a high degree of transparency.

The remainder of the paper proceeds as follows. Section 2 contains a de­scription and preliminary analysis of the data. Section 3 details the model specification procedure, whereas empirical results can be found in Section 4, Section 5 briefly summarizes the main findings and gives recommendations for further research.

2 Data

The data set consists of microfinance institution (MFI) and macroeconomic data, the latter collected both on a federal and state level. Data were collected over three up to eight fiscal years and arranged by observational unit, so that the resulting panel is composed of vertically stacked time-series data sets.

All data on MFI performance were obtained from the Mix Market archive (mixmarket.org) during January and February 2013, This non-profit organi­zation was founded by the Consultative Group to Assist the Poor (CGAP) as a business information provider in the microfinance sector and is widely recognized as the most authoritative source for microfinance data. Its website contains qualitative and quantitative information on approximately 2000 MFIs and more than 100 institutional investors.

My choice of MFIs was informed by the completeness of information and the time period covered by the reported data. Completeness and reliability of information is represented by Mix Market using a five-diamond rating scheme in which a score of five guarantees best reporting standards, I included only institutions scoring at least three out of five diamonds throughout the relevant period, in order to ensure continuity and minimize the risk of measurement er­rors due to differing accounting standards across time. Furthermore, the data set only includes institutions that provide annual information for at least three consecutive years, and were still in operation in 2013, Due to the unavailabil­ity of reliable data on crucial Brazilian economic indicators for the years up to 2003, 2004 was chosen as the cutoff year. The data set thus encompasses the years 2004 to 2011, the latter being the most recent fiscal year for which MFI data are published on the website. Although Ahlin et al (2011) don’t specify which MFIs their sample includes, they also source MFI data from the Mix Market over the period 1996 to 2007, using essentially the same selection cri­teria as detailed above, I therefore assume a certain degree of overlap between their sample and mine. Since most MFIs contained in the present sample were however only established or started reporting data after 2004, the panel is un­balanced, However, this should not pose a problem, given that the year an observational unit enters the panel seems not to be related to the endogenous variables in the model.

Altogether, my data set contains 53 MFIs , each with data for 3-8 years over the period 2004-2011, This includes institutions with initially as little as 200 borrowers, as well as one with roughly 2,000,000 borrowers. The breakdown by institutional type is: 32 mon-bank financial institutions’ (XBFI), 16 ‘non­profits’ (XGO), three ‘credit unions’ and two institutions officially recognized as banks. Although Cull et al (2007) show that several MFI indicators are strongly correlated with their respective institutional type, I deliberately left it out as a dummy variable. This is due to the fact that most Mexican MFIs in the sample are XBFIs, while the vast majority of their Brazilian counterparts are XGOs, leading me to suspect an institutional selection bias that would only dilute the significance of the included country dummies.

The data set includes 34 Mexican and 19 Brazilian institutions whose ge­ographical distribution by regions (federal states in brackets) is as follows: 15 from the Federal District of Mexico, 5 from Northern Mexico (Nuevo Leon and Baja California Sur), 6 from Central Mexico (San Luis Potosi, Hidalgo, Jalisco, Guanajuato and Morelos), 8 from Southern Mexico (Chiapas); 7 from Northeast Brazil (Bahia, Maranhao, Pernambuco, Piaui, Ceara and Paraiba), 4 from Southeast Brazil (Sao Paulo and Rio do Janeiro), and 8 from South Brazil (Santa Catarina and Parana)¹, While the absence of truly rural MFIs means that this sample is far from representative for the whole microfinance universe of these two countries, the Mix Market!s central role in the micro­finance community guarantees that the sample comprises almost all MFIs of interest to institutional investors. Therefore the apparent imbalance in favor of Mexican MFIs is not owed to a sampling bias, but reflects the fact that the concept of microfinance was adopted earlier in Mexico and that its number of MFIs that reached a certain point of maturity is consequently higher,

2.1 Variable Description

The estimated variables obtained from the Mix Market are operational self­sufficiency and PAR-30, Operational self-sufficiency is a measure of financial efficiency equal to annual operating revenues divided by annual administrative and financial expenses. If the resulting figure is greater than 1, the organization under evaluation is able to cover administrative costs with client revenues and is considered to be operationally self-sufficient.

The variable PAR-30 (portfolio at risk over 30 days) expresses the percent­age of the loan portfolio behind schedule with payments for more than thirty days. It acts as an early indicator of default problems which, if left unchecked, can seriously undermine the MFFs lending mission and lead to bankruptcy,

I include further non-diseretionary MFI variables as controls. These are age, calculated from the year the MFI first submitted data to the Mix Market, the number of borrowers, and average loan size. The latter two are lagged by one year to alleviate endogeneity concerns, and reflect MFI size as well as possible economies of scale. The age variable represents operational experience and is included to capture the effects of differing managerial competence. Since financial level variables are reported on the Mix Market in current USD, the values for the variable average loan size were discounted to 2011 values by multiplying them by the respective USD deflators. The reason for choosing 2011, and not 2004, as the monetary base year is that the majority of MFIs in the sample were established after 2004, Expressing monetary variables in 2004 prices would thus have meant an extrapolation that I choose to avoid.

In line with the paper’s objective, data for all of the macroeconomic vari­ables were collected twice; on a national level and on the level of individual states. With a few exceptions, the variables described below were obtained from Mexico’s and Brazil’s respective statistical institutes, the Tnstituto Na­tional de Estadistica y Geografia’ (IXEGI) and the Tnstituto Brasileiro de Geografia e Estatistica’ (IBGE),² MFI and state-level data are summarized in Table 1,

The most pivotal economic parameter in the data set is inflation, since it not only serves as an independent variable, but is used to discount values reported in current local currency to 2011 levels. From the several inflation indices available in both countries for different purposes, I chose those based on the cost of a typical basket of goods for low-income households, because this measure best represents the economic reality of a majority of MFI clients. In Mexico this is the Tndiee National de Preeios al Consumidor’ and in Brazil the Tndiee National de Pregos ao Consumidor’, Since these inflation rates vary widely across federal states, I calculate separate currency deflators for each state and apply them to all income-related variables, except in the benchmark regressions with national-level variables,

A central control variable is Income j,t - 1, the real GDP per capita expressed in 2011 PPP dollars and lagged by one year. While these figures are readily available in Brazil for both administrative levels, in the case of Mexico I cal­culate them by dividing each state’s total GDP by its population at the time. The Mexican population census is conducted every five years, so the missing population values are interpolated linearly. The variable income growth is di­rectly derived from the real GDP per capita by division of a year’s value by the previous year’s value.

Workforce participation rate describes the percentage of working-age per­sons in an economy who are either employed or actively looking for an employ­ment, In both Mexico and Brazil, working age is defined as 15 to 64 years. This indicator is a good measure of employment opportunities in the formal labor market, which may either be complementary to micro-entrepreneurial activities or crowd them out. Since the Mexican states only survey workforce participation every five years as a part of the national population census, I interpolate the missing values linearly. The Brazilian labor participation data, updated annually, show disparities between states but little variation from year to year. This slow-moving quality supports the validity of linear interpolation of Mexican states’ missing values, resulting in a negligible loss of variance.

Income inequality is included as a variable in the form of the Gini coef­ficient., which can take values between zero and one, A Gini coefficient of zero expresses perfect equality, where everyone has an exactly equal income, whereas a Gini coefficient of one expresses maximal inequality where only one person disposes of all the income. In contrast to Ahlin e.t al. (2011) I chose the Gini variant based on disposable household income, which is calculated on income after taxes and transfers. This is a more adequate measure of eco­nomic opportunities than the alternative Gini coefficient on individual pre-tax income, because it accounts for differing tax burdens and household struc­tures. However, the Gini coefficient should only be interpreted jointly with the real income per capita level, as high inequality in a wealthy state may still imply better basic services for the poor than those available in a more egalitarian but poor state. For the Mexican states, data are available for the years 2000, 2005, 2008 and 2010, The missing values were therefore obtained by linear interpolation. For Brazil, Gini coefficients on disposable household income are only available on the level of macroregions, which are composed of soeio-eeonomieally and climatically similar states.

The variable industry describes the share of the industrial sector’s value- added in total GDP, It features in the model as an indicator of labor-intensive employment opportunities which may crowd out the establishment of small enterprises. While the Brazilian IBGE reports this ratio directly for the in­dividual states, in the Mexican ease I computed it by dividing the industrial value-added by the states’ total GDP,

Apart from these, my data set includes values for three macroeconomic variables not used by Ahlin et al. (2011), These are unemployment, literacy and minimum wage,

I chose to include unemployment as an additional measure of labor market opportunities, because it is more volatile than both workforce participation and industry’s share of GDP, Unemployment is therefore better suited to cap­ture fluctuations of the business cycle that don’t necessarily translate into GDP growth. An important aspect of unemployment figures in both Mexico and Brazil is that, due to the lack or symbolic nature of unemployment ben­efits, people don’t directly report their employment status to the responsible ministries. Instead, unemployment rates are estimated from regular household surveys conducted by the statistical institutes, which the government then publishes as the official unemployment figures.

The Mexican ‘Encuesta Xaeional do Oeupaeion y Empleo’ (EXOE) is a trimestrial household survey with a sample size of 120,260 that reports unem­ployment rates on the national and state level. The Brazilian ‘Pesquisa Mensal do Empregcy (PME) on the other hand is a monthly household survey with a sample size of about 38,000 and is conducted in the major metropolitan areas. Another more representative household survey, the annual ‘Pesquisa Xaeional Amostral de Domicilios! (PXAD) reports unemployment rates on the national and maeroregional level. Since all MFI data are year-averages, I also con­densed the trimestrial and monthly unemployment rates to year-averages by computing their arithmetic means.

In some eases Brazihs metropolitan areas (all of them state capitals) coin­cide precisely with an MFPs operational radius, while unemployment data for the other MFIs had to be approximated on the basis of the only available state- level data from the 2010 national census, I did this by matching each state either with its capital or with the macroregion in which it is located. Then, all metropolitan and regional values were divided by the respective 2010 val­ues, These ratios were then multiplied with the corresponding state values from the 2010 census. The resulting state values for the years 2004-2009 and 2011 thus emulate the employment trends of their respective macroregions or capitals. The unemployment rates for Parana and Santa Catarina, for exam­ple, both follow the labor market movements of Southern Brazil, while at the same time remaining consistently different due to their anchorage in different 2010 rates. The other states are matched as follows: Bahia emulates the labor market movements of its capital Salvador; Pernambuco emulates its capital Re­cife; Rio de Janeiro emulates its homonymous capital; Sao Paulo emulates the Southeastern region; Maranhao, Piaui, Ceara and Paraiba emulate the Xorth- eastern region. This approximation is legitimized by the fact that Brazihs metropolitan areas are the growth motors of the wider regions in which they are located, and that their economic cycles affect the labor market in their periphery through large-scale workforce migration (cf. Baeninger (2012)), Us­ing metropolitan unemployment figures therefore provides the highest possible granularity, especially for those MFIs that operate exclusively within these metropolitan areas.

Another variable not present in the Ahlin e.t al. (2011) study is literacy, expressed as the percentage of citizens over the age of 15 than can read and write simple texts. Although functional literacy might have been a better measure of economically relevant education, Mexico and Brazil use different definitions, which would have invalidated any country comparison.

The third additional economic variable is the minimum wage, which I in­clude as a measure of the opportunity costs a micro-entrepreneur faces. The underlying assumption is that rising minimum wages might crowd out en­trepreneurial activities and thereby diminish MFIs’ client base, Mexico’s min­imum wage is fixed on a per-hour basis, whereas Brazil’s is a month’s wage. To ensure comparability with the variable income per capita, I convert both to annual values by presuming 26 workdays per month of eight hours each. While Brazil’s minimum wage is the same for the entire country, Mexico dis­tinguishes three wage categories on a geographical basis. Of the states present in the sample, the Federal District and Baja California Sur fall into Area A which receives the highest minimum wage, followed by Jalisco and Xuevo Leon in Area B, All other states belong to Area C and are afforded slightly lower minimum wages. Minimum wage is the only variable in the data set that has the same values for both the state-level and national-level regressions.

The last of my additions to the original model are the geographical dummy variables, which I introduce in three consecutive steps. At first I run OLS regressions with a single country dummy that takes the value one for Brazilian MFIs and zero for Mexican ones. After specifying a model, I substitute the country dummy with regional dummies in accordance with the macroregions mentioned above. Then I compare the two models with respect to R[2] and the three information criteria to cheek whether the regional dummies have a higher explanatory power than the mere distinction between countries. In a last step, I remove the regional dummies and specify a model with dummies for each of the federal states, and repeat the comparison described in the preceding step. For both outcome variables, using state instead of country dummies significantly improves the model, which in part validates my main hypothesis.

However, I also omitted a number of variables included in Ahlin et al. (2011) for practical reasons. These are private credit bureau coverage, foreign direct investment (share of GDP), remittances, as well as several institutional indicators like corruption perception and the time to enforce contracts. While all of these indicators can be expected to vary widely from state to state, data for them is only available from the World Bank’s World Development Indicators on a national level. Especially the institutional variables are of a very slow-moving nature and would have resulted in quasi-multieollinearity. By excluding them, whatever variance they might have explained is instead captured by the geographical dummies,

2.2 Preliminary data analysis

Before regressing self-sufficiency and payments at risk on the other variables, some valuable information can be obtained by simply analyzing the data as they are.

The outcome variable operational self-sufficiency shows few unexpected features. With a mean of 115% and a median of 112%, a majority of MFIs contained in this sample manage to break even in most years and generally show the same level of efficiency as those in Ahlin et al (2011), The only truly exceptional value of 4,7% can be observed in a Mexican XGO, which however achieves self-sufficiency in later years. It is noteworthy that standard deviation is higher between than within observational units (0,26 versus 0,16), implying either the existence of fixed effects or the strong influence of different economic environments. Generally speaking, there is a trend of underperfor- manee during the initial years of operation, which is followed by a gradual improvement. Whether this observation is due to learning processes or an improving macroeconomic climate remains to be seen.

The other outcome variable, PAR-30, warrants a closer look. With a me­dian of 3,7% of the gross loan portfolio at risk against a mean of 6,6%, it exhibits a strongly positive skew. Contrary to the overall improving self­sufficiency indicator, there seems to be no singular chronological trend. While some institutions start with exceptionally high default rates that gradually fall over the years, others start with a comparatively healthy loan portfolio and only experience a spike in default rates after 2008. A possible explanation is that initially lax MFIs undergo a learning process, whereas the later increase in loan delinquency might be due to macroeconomic shocks. Extreme values include one institution reporting zero percent PAR-30 throughout the obser­vation period, and others with over 50% of their loan portfolio at risk during the first three years of operations. The fact that standard deviation is about 0.068 both between and within observational units suggests that there are no significant MFI-level determinants apart from an institution’s age.

Table 1: Summary statistics of institution, state and national data

Abbildung in dieser Leseprobe nicht enthalten

Note: All macro variables are reported on two levels. The first row summarizes state data while the second row summarizes national data. The number of active borrowers is reported in 1000s; income per capita in 1000s of constant 2011 international $; and minimum wage in constant 2011 international $.

The discrete variable age takes all integral values between one and sixteen, although only a few institutions reach an age over eight by the end of the observation period, Brazilian MFIs in the sample tend to be younger than their Mexican counterparts, reflecting the earlier adaptation of the microfinance concept in Mexico,

The average number of borrowers is by far the most diverse variable in the data set, with values ranging from 210 to 1,962,000 borrowers. It it also extremely skewed, with the median number of borrowers at 10,397 and the mean at 84,492, I therefore use the logged version of the variable to reduce the leverage of outliers. Although the determinants of borrower growth is not specifically analyzed in this thesis, it should be noted that it is itself an outcome variable that accounts for the entire within-variance in the number of borrowers. To at least alleviate the endogeneity issues resulting from this relationship, I lag the number of borrowers by one year.

The same is essentially true for the other MFI size control variable, average loan size., whose values range from 125 to 10,451 constant 2011 international US-$, The median loan size in the sample is $404,7, whereas the mean value is over twice as high with $848,84, For the same reasons as with the number of borrowers, the average loan size is first lagged by one year and then logged. Contrary to what can be observed for the number of borrowers, however, stan­dard deviation is much higher between observational units than within them ($1,096 versus $660), As a consequence, this control variable has the poten­tial to capture some of the between-unit variance that would otherwise be attributed to fixed effects. Another interesting observation is the negative cor­relation between the number of borrowers and average loan size. At —0,069 (p-value 0,25) this relationship may not appear significant, but the microfi­nance literature suggests that borrower growth can be achieved by offering smaller loans that serve the needs of lower-income customers.

The macroeconomic variable income per capita is remarkable in its disparity across states, Mexico hosts both the poorest and the richest state in the sample, with Chiapas at $2,550 and the Federal District at $28,518 per capita (convert to 8-year averages). Due to its strongly trimodal distribution with peaks at around $5,000, $13,000 and $27,000, and no observations at all around $20,000, it is not suited to make consistent predictions for the whole income range, I consequently treat income as a mere control variable to test whether the coefficients of other regressors depend on an economy’s income level.

Whereas investors can treat the relative level of income between states as fixed in the short term, the annual growth in real GDP per capita is much more volatile. Estimating its influence on MFI performance is therefore of special interest not only for investors but for the institutions themselves, A Chi-square test for normality shows that growth rates follow a roughly normal distribution (p-value = 0,0724) with a mean of 2,46% and a standard deviation of 4,32%, Since most of this is within-unit variance, the conditions in terms of economic growth are very similar for all MFIs, The main difference here is that economic cycles occur slightly time-delayed across states.

Inflation rates in both Mexico and Brazil declined sharply after the crip­pling hyperinflation that marked the 1980s and 1990s (Levy (2006)), This is illustrated by the values in the sample, which reach a maximum of 8,64% and a minimum of only 1,75%, with a concentration between 3% and 5%, Although the between-groups standard deviation of 0,6% seems conspicuously low, in­flation rates diverge widely between states in most years, and thus legitimize the disaggregation of national inflation rates.

Workforce participation is at first glance a very uniform variable with over half of all values concentrated between 60% and 63%, However, some states witness a slowly declining workforce participation, while other states within the same country see the share of economically active people increase. This may be a symptom of employment opportunities in the long run, as long­term unemployed persons leave the workforce altogether. At the same time, a positive economic outlook may encourage others, especially women, to join the workforce for the first time. Both trends are important for MFIs, because when labor opportunities are scarce more people may be forced to become

self-employed and to take out loans.

Unemployment, as explained in the previous chapter, has potentially simi­lar effects on MFI performance. While workforce participation is an indicator of long-term developments, however, the rate of unemployment captures the more short-lived economic cycles. Furthermore the highly significant correla­tion of 0,33 makes it virtually impossible to separate the two without causing endogeneity issues. In this sample the unemployment rate reaches values be­tween 1.9% and 13,9%, with about half of all observations in the interval 5% to 7,5%, Interestingly, the state with the lowest unemployment rates throughout the observation period is Chiapas in the notoriously underdeveloped Mexican south. This can probably be attributed to the fact that a large part of Chi­apas’ population are subsistence farmers, which are not part of the formal workforce. In the absence of worthwhile urban employment opportunities, the number of people actively looking for wage employment is consequently lower. The opposite is true for the Brazilian states of Bahia and Pernambuco, where land scarcity forces the rural surplus population to migrate to the cities, where they either join the ranks of the unemployed or try to start their own busi­ness, thus becoming MFI clients (Baeninger (2012)), Almost all states in the sample experienced an upsurge in unemployment in 2008 or 2009 that can be attributed to the global financial crisis,

Mexico and Brazil are two of the most unequal countries on earth in terms of income distribution. From a level of 0,529 in 2004, Mexico’s Gini coefficient declined steadily to 0,505 until 2008, when the US mortgage crisis spilled over and rising unemployment caused inequality to rise again, Brazil, on the other hand, saw its Gini coefficient plunge from 0,577 to 0,528 between 2004 and 2011, largely thanks to a large-scale conditional cash-transfer program initiated in 2003 (cf. Soares e.t al. (2010)), On the state level the divergence is even greater, especially in Brazil where both the most unequal (0,571) and the most egalitarian state (0,419) in the sample can be found. Also on the state level, inequality is highly significantly correlated with several other macroeconomic variables: income (—0.34), workforce participation (—0.39), industry (—0.4) and literacy (—0.68). Even if inequality was a mere product of these factors, it would not necessarily be redundant in my model because it captures the degree of a society’s duality. In a context of extreme inequality, the informal and formal sectors of an economy tend to move independently from each other, which according to Fields (2009) makes it difficult for small enterprises to access the broader national market.

Industry's share of GDP can be regarded as another measure of employment opportunities, but in this case specifically for the better-paying formal sector. The opportunity costs of forfeiting an employment in the industrial sector in order to start one’s own enterprise are consequently higher than for an unskilled laborer in the informal sector. The degree of industrialization in the sample varies from 11.3% to 46%, remaining fairly stable over the years but differing widely between states. This gives the variable industry characteristics of a fixed effect on the state level.

The variable literacy is included as a proxy for a wide array of compo­nents that together constitute human capital. The intuition here is that, like most macroeconomic parameters, literacy impacts MFI performance indirectly through the success or failure of the small businesses that resort to the MFIs for credit. Although there are certainly better measures for the kind of human capital needed to run a business, such as university attendance rates or the availability of vocational training, literacy is the only educational indicator available for all states. It can however be reasonably assumed that the level of literacy in a given geographical division is correlated with its number and quality of universities, as well as other education opportunities. The literature rates in the sample range from 75.7% in the Brazilian state of Piaui to 98% in the Federal District of Mexico, with half of the observations over 95.5%. There is a certain convergence of the initially very disparate literacy rates to­wards the end of the observation period, yet fundamental differences between states persist. Therefore literacy may capture some of the variance that would otherwise remain unobservable fixed effects. Another striking property of liter­acy is its highly significant correlation with both income (0.84) and workforce participation (0.576).

The variable minimum wage is remarkable in several ways. First of all, it is subject to only minor state-level differences, while diverging starkly between Mexico and Brazil. Whereas Mexico’s regional minimum wages are simply adjusted for inflation every year and are therefore stagnated at around 2,000 constant US-$, Brazil’s minimum wage surged from about $2,400 in 2004 to just below $3,600 in 2011 (see table x). In the ease of Brazil, the different values across states are mostly negligible and a result of the different inflation rates used for discounting to 2011 prices. What makes the minimum wage interesting as an explanatory variable is that the two countries’ different wage policies constitute a quasi-experimental situation in which Mexico can be regarded as the control group. So if a rise in minimum wages does indeed affect one of the MFI outcome variables by crowding out micro-entrepreneurial activities, the coefficient for min_wage should more closely represent the 'true’ parameter of this relationship than is the ease for the other regressors.

3 Estimation Strategy

In accordance with Ahlin e.t al. (2011), I estimate several variants of the fol­lowing model:

Abbildung in dieser Leseprobe nicht enthalten

where yj is a vear-t outcome of MFI i located in state j; M it is a vector of MFI-specific control variables in year t; Xj t is a vector of macroeconomic variables describing state j in year t; incj, t-l is the income level of state j in year (t-1); and GEO is a vector of geographical dummy variables, consisting of either one country dummy, six regional dummies or 18 state dummies.

The main dependent variables are operational self-sufficiency and the de­fault indicator PAR-30, Other MFI performance indicators were tentatively estimated in the same way, but were found to depend solely on institution- specific factors that are not the concern of this thesis.

The baseline specification of the model includes three MFI-specific control variables: age, log of the lagged number of borrowers, and log of the lagged average loan size. Contrary to Ahlin et al. (2011) I leave out institution- type dummies because in my sample they exhibit such a strong correlation with the country in which the MFIs are located that a selection bias is to be suspected. The baseline set of macroeconomic variables consists of income growth, workforce participation and industry, along with the lagged level of real income per capita as a control.

The general procedure is the same for both dependent variables. First I plug the state-level values of the macroeconomic variables and the country dummy into the baseline model and run an OLS regression. Then I con­duct consecutive F-tests with the remaining independent variables in the data set, first one at a time, then in groups of variables bound by mutual correla­tion, The macroeconomic baseline variables are also subjected to F-tests and omitted in ease they fail to show any significant relationship with either the dependent variable or the other regressors. The control variables age, number of borrowers, loan size and income level are exempt from this treatment.

This process is then repeated first with the regional and then the state dummies instead of the country dummy. After comparing the three resulting models with respect to adjusted R[2] and three information criteria (Akaike, Schwarz and Hannan-Quinn) I choose one of these specifications based on the balance between degrees of freedom and the explained variance. Although the increased number of dummy variables implied by higher granularity automat­ically increases the proportion of total variation of outcomes explained by the model, this may not justify the loss of degrees of freedom in the comparatively small sample. It should also be noted that the resulting three models do not necessarily contain the same regressors since the variance attributed to one variable in the first model can be absorbed by the disaggregated geographical dummies on the next deeper level, thus rendering the variable redundant.

In a next step, the chosen model is tested for heteroscedasticity with a Breuseh-Pagan test, which I follow up with a Hausman specification test to check whether a random effects estimator would yield consistent results. If the assumption of heteroscedasticity is confirmed and depending on the outcome of the Hausman test, I proceed with either a random effects or a fixed ef­fects estimator. In the latter case the geographical dummies are automatically omitted.

After a final model has been specified and estimated, I substitute the state- level variables in the model with their national equivalents and compare the two regressions with respect to R[2] and the three information criteria listed above. This is the central part of the thesis as it allows the direct comparison between the goodness of fit provided by national data versus state data. However, the question here is not so much whether increased data granularity improves the goodness of fit at all. This is virtually guaranteed if the macroeconomic context as represented in the model plays any meaningful role in determining self-sufficiency and default rates. Instead, the purpose of this thesis is to examine whether the gains in predictive precision are sufficient to justify the more time-consuming accumulation of sub-national data,

4 Results

This section discusses the models obtained through the specification process described above. The term 'baseline model[5] henceforth refers to the respective model specifications using state-level variables. Considering the comparatively small sample size, the focus of this section is not on the exact variable coeffi­cients, but rather on their signs and the comparison between the baseline and the national models.

4.1 Operational self-sufficiency

The model specification that I consider most appropriate for estimating the outcome variable operational self-sufficiency takes the form of a random effects estimator. The coefficients and significance levels for both the state-level and national variants are reported in Table 2,

Tabic 2: Random-effects estimators for self-sufficiency

Abbildung in dieser Leseprobe nicht enthalten

Note: Each column corresponds to a separate regression, with the results of the baseline model on the left and the benchmark model with national data on the right. Coefficients are reported with their respective significance symbolized by one to three asterisks for the 10%, 5%, and 1% levels respectively. Standard errors are reported in parentheses.

Apart from the MFI-specific control variables, the final model includes the variables income, a quadratic in income, workforce participation, gini coef-fieient, industry, literacy and minimum wage. The initial baseline variable income growth is missing in this model, because neither did it exhibit a sig­nificant correlation with the outcome variable, nor did its omission impact the model in any perceivable way. The geographical dummy variables, although jointly highly significant on the state level, were excluded after the transition from an OLS to a random effects estimator. This decision is supported by the considerable improvement in all three information criteria upon omission of the dummy vector, implying that the gained degrees of freedom far outweigh the loss of explained variance,

A test for non-linearity resulted in the inclusion of a quadratic in the lagged level of income. According to the baseline model, the relationship between income and self-sufficiency is (ceteris paribus) given by self_suff = 0.251184 x income — 0.00746562 x income [2].

This implies positive but decreasing marginal effects of the income level on self-sufficiency, up to a turning point of $16,823, The only states well beyond that threshold are the Federal District of Mexico and Xuevo Leon, whereas Sao Paulo crossed it as recently as 2010, It thus appears harder to break even in poorer states where inadequate infrastructure raises transaction costs, among other possible factors. On the other hand, if a state is too rich there are more commercial banks who can afford to squeeze out MFIs by offering more competitive loan conditions (cf. Harper and Singh (2005)),

The model associates one percentage point of workforce participation with a 3,3 percentage point higher self-sufficiency ratio, while one percentage point of industry is associated with a 2,4 percentage point lower ratio. These rela­tionships are significant at the 5 and 1% levels, respectively. Since both are, by extension, measures of employment opportunities in an economy, they should be analyzed together. One interpretation of the coefficient signs is that an eco­nomically active population stimulates the growth of small business, whereas an abundance of jobs in the industrial sector lowers the demand for financial services as a result of fewer self-employed people.

Inequality, expressed by the gini coefficient, is positively associated with operational self-sufficiency, though only significant at the 10 percent level. If there is indeed a causal effect, it might be that the poor are more dependent on microfinance services in an unequal environment where commercial banks apply stricter rules for creditworthiness.

According to the baseline model, literacy is a negative predictor of self­sufficiency at the 10 percent significance level. Due to literacy’s strong and highly significant negative correlation with inequality (—0,68), both are likely to affect MFI performance in part through the same channels. In terms of entrepreneurial opportunities, illiteracy is a stronger factor of economic exclu­sion than merely unequal income distribution. For this reason MFIs offering financial literacy programs in areas with low education levels have a com­petitive edge over commercial banks that binds their customers and ensures self-sufficiency.

The last economic variable, minimum wage, also seems highly significant in that every $100 of annual minimum wage is associated with a 2,2 percentage point lower self-sufficiency ratio. This relationship can be intuitively explained by the reduced incentive for self-employment in a context where minimum wages are sufficient to sustain a family, which depletes the MFIs’ potential client base,

A conspicuous aspect of the present model is the absence of the more volatile variables income growth, inflation and unemployment. In the case of unemployment, any possible effects are most likely captured by the themati­cally related but slow-moving variables workforce participation and industry. This suggests the interpretation that short spells of unemployment alone don’t drive significantly more people into self-employment, whereas a constantly high workforce participation combined with scarce employment opportunities in the labor-intensive industrial sector incites entrepreneurship and thus benefits

MFIs.

The same seems to be true for real ineome growth, whose year-to-year fluc­tuations may matter less to small enterprises and MFIs than the long-term growth, or lack thereof, reflected in per capita ineome levels. In this respect, Ahlin et al. (2011) find that growth rates exhibit a significant influence on several MFI performance indicators in a rudimentary model, but lose this sig­nificance once MFI size controls are included. They conclude that the size controls (number of borrowers, average loan size) “capture some effects of per­sistent macroeconomic growth, given that high macroeconomic growth can lead to MFI growth,"

As for inflation, the impact on self-sufficiency and the other coefficients is negligible, I deduce from this observation that decades of hyperinflation have shaped both Mexico and Brazil to such a degree that the comparatively small fluctuations in inflation rates in the observation period do not decisively influence domestic investment decisions. This interpretation is supported by recent findings of the International Monetary Fund (Aliehi et al (2011)),

Even without these high-frequency variables, the differences between the baseline model and its national equivalent remain considerable (see Table 2), Xot only do the national data fail to detect any significant relations between the macroeconomic variables and self-sufficiency, but all three model selection cri­teria favor the baseline model with state data. Contrary to the results discussed above, the national model associates literacy positively with self-sufficiency. Furthermore, it attributes a markedly lower influence to workforce participa­tion, and associates inequality with much higher levels of self-sufficiency than the baseline model. One possible explanation for these discrepancies is that the baseline model overfits the data, simulating significant relations between variables that cannot be replicated with other data. It is however far more likely that the national data simply don’t accurately represent the MFIs’ eco­nomic environment, especially considering that the national data are weighted averages across all 59 federal states in the two countries, of which only 19 are included in the sample.

The coefficient estimates for the MFI control variables, on the other hand, are fairly similar in both models. As expected, age and the two dimensions of MFI size are all positive predictors of self-sufficiency. In the baseline model, an additional year of experience is associated with a 3,7 percentage point higher self-sufficiency ratio, A one-percent increase in the number of borrowers and average loan size is related to a 4,1 and 6,4 percentage point higher self­sufficiency, respectively,

4.2 Payments at Risk

The model specification for the outcome variable PAR-30, from here on referred to as delinquency rate, takes the form of a fixed effects estimator. Coefficients and significance levels for the baseline and national model are reported below in Table 3,

Following a test for linearity, I add a quadratic in age to the other MFI controls to capture the effect of diminishing experience gains per year of age. The macroeconomic regressors included in the model are per capita income, workforce participation, unemployment, inequality and minimum wages,

A per capita income of $1,000 is associated with a two percentage point higher delinquency rate. This highly significant relation could be due to the availability of alternative credit sources in richer states, which makes debtors less dependent on one single institution and thus lowers their incentives to fully comply with established dates.

Workforce participation is a negative predictor of loan delinquency at the 1% significance level. One percentage point of workforce participation is as­sociated with a 1,6 percentage point lower delinquency rate. One percentage point of unemployment, on the other hand, is just as significantly associated with a 2,4 percentage point higher delinquency rate. Since both measures reflect the extent of labor market opportunities, they should be interpreted together, A household survey conducted by the Inter-American Development

Table 3: Fixed-effects estimators for PAR-30

Abbildung in dieser Leseprobe nicht enthalten

Note: Son Note to Table 2.

Bank (Llisterri e.t al. (2006)) suggests the economic necessity to start a busi­ness as a common link, whereby the two factors balance each other out. In a context of high workforce participation the dependency ratio in households is lower, meaning that working individuals have fewer family members to sus­tain. This reduces the incentive for self-employment in the same way that persistently high unemployment raises it. In the latter ease, businesses born out of necessity instead of economic opportunity are more likely to default on their loans, which is reflected by increasing delinquency rates.

Inequality, while not being statistically significant in the present model,

is a negative predictor of loan delinquency. Similar to my interpretation of the income coefficient, the dependency of microentrepreneurs on specific MFIs might be higher in unequal regions due to the restricted access to alternative credit sources. This dependency makes debtors keen to respect amortization dates in order not to lose their good credit standing with the MFI,

Lastly, the model associates $100 of minimum wage with a 0,9 percentage point decrease in delinquency rates, A study recently published by the Brazil­ian Finance Ministry (da Silva et al. (2012)) finds conclusive evidence that retail loans extended to customers are a major determinant of small business loan default. Since rising minimum wages continually improve the financial situation of the predominantly poor client base, retail businesses become more solvent themselves, which in turn reduces their propensity to default on loans. For Mexico this relationship may be irrelevant since Mexican minimum wages have over the last decade stagnated in real terms.

As in the self-sufficiency model, I omitted the variables growth and inflation because the null hypothesis of irrelevance to the model could not be rejected in either ease. The prevalence of the industrial sector and literacy also seem to play no significant role in MFIs; delinquency rates.

The coefficients estimated by the national model diverge even further from those of the baseline model than was the ease with self-sufficiency. This is most noticeable in the opposite coefficient signs attributed to the variables income, unemployment and minimum wage, while the effect of workforce participation is estimated to be significantly lower than in the baseline model. Again, I conclude that this discrepancy is owed to a large part to the inadequacy of using national data when the observed economic units, in this ease MFIs, don’t operate in the whole country, but instead on a regional or state level. This conclusion is supported by the goodness-of-fit measure R[2] as well as the three model selection criteria, all of whom favor the baseline model.

The only aspect on that the two models agree is the role of MFI-speeifie fac­tors, Both the baseline and the national model predict that delinquency rates decline with each year, up to a turning point of 14 and 12 years, respectively. Both the number of borrowers and average loan size are positively and sig­nificantly associated with higher delinquency rates, which literature suggests is a consequence of loan portfolios growing faster than the MFIs; capacity to effectively recover loans (ef, Gonzalez (2010)),

5 Conclusion

This thesis analyzes microfinance institutions in a sub-national context by estimating the role of state-level macroeconomic factors in the performance of 53 microfinance institutions from Mexico and Brazil, The focal outcome variables operational self-sufficiency and PAR-30 are first regressed on state- level data and then on national data as a benchmark.

From the perspective of an investment analyst, the more precisely the per­formance of portfolio items can be predicted, the better. In the ease of in­vestments in microfinance funds, systematic MFI evaluations have hitherto largely neglected the influence of the host countries[5] macroeconomic environ­ment, Based on the seminal findings of Ahlin et al. (2011), I go one step further by showing that increasing data granularity from the national to the state level yields more precise predictions of MFI success,

I also find that short-lived fluctuations in the economic performance of a state are less important for MFI success than the state’s overall development level. In the models employed in this paper, a state’s development level is manifested in its per capita income, workforce participation, level of industri­alization, unemployment rate, income inequality, literacy rate and minimum wage. While location alone is certainly not an MFI’s destiny, it plays an unde­niable role in its failure or success and should therefore be taken into account. So even when an MFI evaluation’s focus is on managerial competence and good practices, the macroeconomic environment in which the MFI is situated ought to be controlled for. Furthermore, wherever MFIs operate only in parts of a country, data on this geographical sub-division should be preferred to the much more approximate national data. At least for federative countries like Mexico and Brazil with extensive data on the state level, the somewhat tedious effort of gathering these data is far outweighed by the gains in predictive precision.

A Appendix

Table 4: List of states and institutions

References

Ahlin, C., Lin, .J. and Maio, M. (2011), Where does microfinance flourish? microfinance institution performance in macroeconomic context. Journal of Development Economies, 95, 105-120,

Alichi, A., Catao, L., Kamenik, O., Kim, H., Lanton, D., Leigh, D., Pescatori, A., Portillo, R., Simon, J. and Zanna, F. (2011). World Economic Outlook (September). Tech, rep,, International Monetary Fund,

BAENINGER, R. (2012), Migratory turnover: a new look at internal migration in Brazil, Revista Interdiseiplinar da Mobilidade Humana, 20, 77-100,

Cull, R., Demirguc-Kunt, A. and Morduch, J. (2007). Financial per­formance and outreach: A global analysis of leading mierobanks. Economic Journal, 117, F107-F133,

da Silva, A., Terra, J., Marins, M., Beatriz, M., das Neves, E. and Carlos, A. (2012), Credit default and business cycles: An empirical inves­tigation of brazilian retail loans. In Estudos Eeonomieos, vol, 70, Brasilia: Ministerio da Fazenda,

FIELDS, G. (2009), Segmented labor market models in developing countries. In The Oxford Handbook of the Philosophy of Economic Science, Oxford University Press,

Gonzalez, A. (2010), Is microfinance growing too fast?, mIX Data Brief No, 5.

Harper, M. and Singh, S. (2005), Small customers, big market: commercial banks in micro ft,nance. ITDG Publishing,

Johnson, S. (2009), Microfinance is dead! long live microfinance! critical reflections on two decades of microfinance policy and practice. Enterprise Development and Micro finance, 20, 291-303,

Krauss, N. and Walter, I. (2009), Can microfinance reduce portfolio volatil­ity? Economic Development and Cultural Change, 58, 85-110,

LAPENU, C. and LEDESMA, G. (2011), Responsible investment in microfi­nance: The value added of social audits for the fund managers. Enterprise Development and Microfinance, 22, 291-303,

Levy, M. D. (2006), Sound monetary policy, credibility, and economic per­formance, Cato Journal, 26, 231-245,

Llisterri, .J. .J., Kantis, H., Angelelli, P. and Tejerina, L. (2006), Is youth entrepreneurship a necessity or an opportunity? a first exploration of household and new enterprise surveys in latin ameriea. In Development Department Technical Papers Series, Washington, D.C.: Inter-American De­velopment Bank,

Patel, U. R. and Srivastava, P. (1996), Macroeconomic policy and output comovement: The formal and informal sectors in india. World Development, 24 (12), 1915-1923.

Soares, F. V., Ribas, R. P. and Osorio, R. G. (2010). Evaluating the impact of brazil’s bolsa familia: Cash transfer programs in comparative perspective, Latin American Research Review, 45 (2), 173-190,

[...]

---

¹ See Appendix for a comprehensive list of MFIs

² All original data can be found either as Excel files or as web links on the enclosed CD-ROM.