Financial Efficiency and Social Impact of Microfinance Institutions Using Self-Organizing Maps

World Development Vol. 46, pp. 197–210, 2013 Ó 2013 Elsevier Ltd. All rights reserved. 0305-750X/$ - see front matter www.elsevier.com/locate/worlddev

http://dx.doi.org/10.1016/j.worlddev.2013.02.006

Financial Efficiency and Social Impact of Microfinance Institutions Using Self-Organizing Maps PHILIPPE LOUIS and ALEX SERET KU Leuven, Belgium

and BART BAESENS * KU Leuven, Belgium University of Southampton, UK Vlerick Business School, Belgium Summary. — This paper contributes to the literature by investigating whether the increased focus on financial self-sustainability of microfinance institutions has been disadvantageous to the target audience. We investigate the association between social efficiency and financial performance using a comprehensive data set that includes 650 microfinance institutions. A self-organizing map methodology is used to fully capture the existing heterogeneity among institutions. The results show that we cannot support the hypothesis that there exists a trade-off. On the contrary, we find evidence of a significant, positive relationship between social efficiency and financial performance. Ó 2013 Elsevier Ltd. All rights reserved. Key words — microfinance, self-organizing maps, outreach, social impact, profitability, self-sustainability

1. INTRODUCTION

private sector (Rhyne & Otero, 2006). The new focus on self-sustainability meant a schism in the management of microfinance institutions. Answering whether this new approach has resulted in less social impact is not straightforward. This is mainly caused by the wide differences between microfinance institutions. After all, “microfinance institution” is just an umbrella term that includes many different types. Firstly, MFIs operate in almost all parts of the world, making them exposed to different social and legal systems. Secondly, some MFIs operate as nongovernmental organizations (NGOs) or cooperatives whereas others operate as banks. Thirdly, some MFIs are relatively young in contrast to others which have been in business for decades. Other MFIs focus their business on supplying loans only while others offer a wide range of financial products. In sum, the diversity among MFIs makes it harder to report conclusive findings. Practitioners must always bear in mind that while an assumption might be true for one kind of MFIs, the same assumption could be violated for another kind of MFIs. Academic research on the microfinance industry must therefore try to capture as much of the aforementioned differences as possible to avoid potentially biased conclusions. For instance, Quayes (2012) divides the MFI spectrum in low- and high disclosure MFIs. Our study endeavors to go even further in the breakdown of MFIs. Firstly, we use a technique called self-organizing maps (SOM) to create graphical two-dimensional maps where similar MFIs are mapped close together

Microfinance experienced an unprecedented growth during the last decades. The United Nations proclaimed the year 2005 as the “International Year of Microcredit” (Year of Microcredit, 2005). Not surprisingly, the academic community has devoted a lot of research on the challenges microfinance faces. In recent years, it has been argued that microfinance institutions (MFIs) have abandoned their social mission due to an increased focus on financial performance. Does emphasizing self-sustainability result in mission drift (Mersland & Strøm, 2010; Hermes & Lensink, 2011)? This paper will try to add new insights and results to this debate. Initially, donor- and government-funded credit was made available to poor borrowers, typically at below-market interest rates. Their goal was to reach the “poorest-of-the-poor” (Robinson, 2001) as microcredit programs have their roots in humanitarian and developmental plans. Since funding was provided by either donors or governments, institutional selfreliance and cost control were not considered crucial. In the 1980s and 1990s, the continuing dependency on subsidies and evidence of unsatisfactory performance resulted in the development of a new microfinance premise of self-sustainability. Donors started to put forward that subsidization should only support newly established MFIs instead of keeping them constantly afloat (Morduch, 1999). They argued that cost control and efficiency would ultimately reduce the dependency of the industry on subsidies, which would allow MFIs to stay in business in the long-run. Furthermore, MFIs will bear the consequences of their own actions, eventually forcing them to act more carefully. MFIs must endeavor to generate sufficient income from the core activities and costs have to be reduced. Besides this evolution, MFIs were also confronted with greater competition and increased interest from the

* We extend our gratitude to the Flemish Research Council (Odysseus Grant B.0915.09) and the National Bank of Belgium (NBB/10/006 and NBB/11/004) for financial support. The authors would like to thank the anonymous reviewers for their valuable comments and suggestions to improve this paper. Final revision accepted: February 2, 2013. 197

198

WORLD DEVELOPMENT

and dissimilar more apart. This methodology is a superior method to graphically plot the heterogeneity among MFIs with regard to the different input variables. To our knowledge, SOM have not been used in the microfinance literature. Secondly, we have included a wide variety of explanatory input variables based on the SEEP Network (2010) standardized approach. This ensures comparability between this paper and previous/future research. Thirdly, a large number of MFIs, drawn from the Microfinance Information Exchange database (Microfinance Information Exchange, 2010c), is used. The possible existence of mission drift and its consequences has sparked off much debates in the academic community. The increase in the number of citations on this subject can be seen in panel A of Figure 1. It depicts the increase in the number of academic journal citations according to Web of Knowledge. On the y-axis, the number of annual citations using the search term “microfinance”, refined by “outreach” (and synonyms) and “sustainability” (and synonyms), is plotted. The growth in the number of citations is obvious. Panel B plots the sum of the gross loan portfolio in USD by region and confirms the unprecedented growth during the last decade. The data are gathered from the Microfinance Information Exchange database using its Cross-Market Analysis tool (Microfinance Information Exchange, 2012). It is clear that all regions experienced growth, though the growth is most pronounced in Latin America & the Caribbean and Eastern Europe & Central Asia. Section 2 provides the reader with a comprehensive literature review. The next section deals with the used self-organizing map methodology and Section 4 elaborates on the data used in this paper. In the next section, we interpret the results of the analysis and in Section 7 we validate our results. Finally, Section 8 reaches a conclusion. 2. LITERATURE The conceptual foundations of the “sustainability paradigm” stem from the failed traditional subsidized credit programs during the 1960s and 1970s (Adams, Graham, & Von Pischke, 1984; Robinson, 2001). These subsidized programs were neither successful from a social impact perspective, nor from a good corporate governance point of view since (1) subsidized interest rates combined with a relatively high cost of

making small loans ensured that loans were channeled to larger borrowers and not to poorer households, and (2) funding was provided by donors, making many subsidized credit programs susceptible to a high degree of moral hazard, resulting in widespread corruption and high default rates. General consensus among practitioners and academics on the future of microfinance focused on more sustainable and more efficient institutions. However, discussion arose between two groups of thought. On the one hand, “welfarists” tend to place relatively greater weight on outreach and its depth where depth of outreach refers to reaching the poorest clients who are very costly to serve (Brau & Woller, 2004). On the other hand, ‘institutionalists’ emphasize self-reliance and the ability to cover operating and financing costs. They claim that this approach is the only viable way to serve a large number of borrowers, resulting in a high breadth of outreach. It is important to mention that both groups ultimately want to maximize social impact, however they differ on whom to target and how to achieve this goal. The advocates of the subsidized credit delivery approach argue that the increased focus on financial efficiency eventually leads to abandoning the original social mission of serving large groups of very poor borrowers. In recent years, many discussions about microfinance focused on the hypothesis that some MFIs have in fact abandoned this social mission. It is alleged that these MFIs even engaged in aggressive marketing and strong-arm tactics to recover loans. A well documented case is the Andhra Pradesh crisis in which MFIs were accused of being responsible for suicides and community expulsions (Economist, 2010; Srivastava, Bharadwaj-Chand, & Sinha, 2010). Other authors even put forward that there exists a “huge disconnect [. . .] between the heady claims made for microfinance and the everyday reality” (Bateman, 2010, p.vi). He mentions that evidence that “mission drift” has become a serious problem is now quite overwhelming (Bateman, 2010, p.54). Many practitioners and academics agree that there could exist a trade-off between financial efficiency and social performance (Otero & Rhyne, 1994; Von Pischke, 1996; Morduch, 2000; Woller, 2002; Mersland & Strøm, 2010; Hermes, Lensink, & Meesters, 2011). They argue that reaching the poorest of the poor is more costly due to the relatively high unit cost of small loans (Von Pischke, 1996; Conning, 1999; Navajas, Schreiner, Meyer, Gonzalez-vega, & Rodriguez-

4.0e+10

Africa East Asia and the Pacific Eastern Europe and Central Asia Latin America and The Caribbean Middle East and North Africa South Asia

Gross loan portfolio in USD

Number of annual

microfinance trade-off citations

50

40

30

20

3.0e+10

2.0e+10

1.0e+10

10

0

0 2001

2003

2005

2007

2009

2011

2001

2003

2005

2007 Year

Year

Figure 1. Growth of microfinance.

2009

2011

FINANCIAL EFFICIENCY AND SOCIAL IMPACT OF MICROFINANCE INSTITUTIONS USING SELF-ORGANIZING MAPS

meza, 2000; Schreiner, 2002). Consequently, reaching more poor borrowers (i.e. increasing the MFI’s depth) comes at a price: worse financial performance. Von Pischke (1996) explains that common sense dictates a trade-off between outreach and sustainability within a medium-term horizon since some borrowers are neither creditworthy, nor committed to repaying. He states that “at some point the outreach shoe begins to pinch the sustainability foot; later it cripples it. Formal credit is therefore not a basic human right [. . . ]”. Hollis and Sweetman (1998) compared six microcredit organizations in 19th century Europe and find that institutions that relied on donor funding were less efficient. Besides, organizations that set interest rates at a realistic level were more sustainable. However, they note that culture is extremely important and that success is highly dependent on the local environment. Mosley and Hulme (1998) relate the design features of 13 MFIs to the institutions’ social impact. They argue that more sustainable institutions may have a higher impact. These MFIs charge a relatively higher interest rate, which could deter borrowers whose project(s) have relatively low rates of return. Furthermore these institutions operate deposit facilities, which offer borrowers a supplementary financial cushion and screen out prospective borrowers who lack financial discipline. Christen (2001) argues that commercialization in Latin America has led to an increasingly competitive environment and significantly larger loan balances, indicating that MFIs could target larger and less poor clients. Using a poverty outreach index, Paxton (2003) finds that sufficiently large self-reliant MFIs may hold the promise of reaching the largest number of poor. Olivares-Polanco (2005) uses data of 28 Latin American MFIs and discovers that the legal institutional structure (e.g. NGO versus other) has no effect on loan size. However, he does find that more financial efficiency may lead to larger loan sizes and less depth of outreach. Prior to 2007 most of the academic literature that dealt with the possible trade-off between self-sustainability and social impact mainly consisted of theoretical discussions and sometimes anecdotal or limited empirical evidence (Hermes et al., 2011). In 2007, Cull, Demirgu¨ßc-Kunt, and Morduch (2007) conducted a comprehensive study using data of 124 MFIs, 49 countries and four years. They discover that larger loan sizes are linked with lower average costs and that the lending type is a major determinant of financial self-sufficiency and potential outreach-profit trade-offs. Institutions that grant loans to individuals perform well on the sustainability dimension but worse on the outreach dimension. They provide, on average, larger loans and lend relatively less to women and poor borrowers. The authors suggest that lenders who try to become or are sustainable focus on granting larger loans and thus eventually target the less-poor. Their view is supported by Hermes et al. (2011), who use a sample of 435 MFIs. They find evidence that outreach and financial efficiency are negatively related. Furthermore, they point out that financial efficiency may only be enhanced by focusing less on the poorest borrowers. These results contrast with Gutie´rrez-Nieto, SerranoCinca, and Mar Molinero (2009), Mersland and Strøm (2010), Quayes (2012). MFIs that perform well on the outreach dimension are also financially efficient (Gutie´rrez-Nieto et al., 2009). With one exception, no MFIs exist that are financially non-sustainable but yet have a high outreach. Mersland and Strøm (2010) find that cost effective MFIs grant smaller loans. Over time, they cannot find evidence of mission drift. Finally, Quayes (2012) discovers a positive relationship between the depth of outreach and financial self-sufficiency except for low-disclosure MFIs. Table 1 summarizes the papers discussed in this paragraph.

199

3. METHODOLOGY (a) Self-organizing maps Kohonen maps, also called self-organizing maps (SOM), have been introduced in 1981 by Kohonen. Fields like exploratory data analysis, web usage mining (Smith & Ng, 2003), industrial and medical diagnostics (Schwartz, Smith, Churilov, Dally, & Weber, 2003), and customer profiling (Seret, Verbraken, Versailles, & Baesens, 2012) are contemporary examples of SOM analysis applications and successes. To the best of our knowledge, this is the first time this technique is used in a microfinance context and the reader interested in financial studies relatively similar to the one presented in this paper is referred to e.g. Huysmans, Martens, Baesens, Vanthienen, and Gestel (2006), Chen (2012), Carlei, Marra, and Pozzi (2012), Chen, Ribeiro, Vieira, and Chen (2013). This section is based on Kohonen (1995) and aims at giving a theoretical background to the reader, whereas an application of the technique can be found in Section 6. The two main objectives of the SOM algorithm are vector quantization and vector projection. The vector quantization aims at summarizing the data by dividing a large set of data points into groups having approximately the same number of points closest to them. The groups are then represented by their centroid points which typically are vectors obtained as the mean of the points of the respective groups (e.g. the k-means algorithm (Tan, Steinbach, & Kumar, 2006)). A typical way to assess the quality of the resulting quantization is to calculate the mean quantization error (MQE) (Kohonen, 2001; Po¨lzlbauer, 2004). It is calculated by averaging all euclidean distances between the different input vectors and their respective closest neurons. A low MQE value indicates a good representation of the input by the SOM is achieved. The interpretation of the obtained value depends on the scale of the input variables. The second objective is vector projection in which the dimensionality of the data points is reduced by projection onto lower dimensional maps (e.g. the PCA (Jolliffe, 2005)). Typically a projection to twodimensional maps is performed in order to be able to visualize and represent the different variables on classical reporting supports. The projection is performed with the neurons obtained after the quantization phase. In a case of a good projection onto the two-dimensional maps, neurons close to each other in the high dimensional space should be mapped to position close to each other in the low dimensional space. The combination of vector quantization and projection enables to explore the data and to use techniques like visual correlation analysis or clustering analysis in an intuitive manner while keeping a mapping between the input vectors and the neurons in the low dimensional space. The different steps of the algorithm are discussed in what follows. In the first step, a feedforward Neural Network (NN) is trained on the input data. The output layer is a map with a lower dimensionality and a given number of neurons. During each iteration of the algorithm, an input data vector ni, representing all the variables for the observation i, is compared with the neurons mr of the output layer using Euclidean distances. Note that the quantization objective imposes both the input vectors ni and neurons mr to have the same dimensionality. mr summarizes at the end of the algorithm a set of input vectors ni. The neuron mc with the smallest distance with regard to the input vector is identified as the Best Matching Unit (BMU): kni  mc k ¼ minr fkni  mr kg:

ð1Þ

The weights of the BMU are then modified in the direction of the input vector, leading to a self-organizing structure of the

200

WORLD DEVELOPMENT Table 1. Summary of empirical research on the potential trade-off between social and financial performance

Publication

Data source

Outreach var.

Fin. efficiency var.

Conclusion

Cull et al. (2007): Financial OLS performance and outreach: a global analysis of leading microbanks

Methodology

MicroBanking Bulletin, 124 MFIs, 49 countries, 1999–2002

Financial selfsustainability ratio

Trade-off: yes but magnitude depends on lending type

Gutie´rrez-Nieto et al. (2009): Data Social efficiency in envelopment microfinance institutions analysis

MIX Market, 89 MFIs, unknown number of countries, 2003

Avg. loan size/GNP per capita, avg. loan size/ GNP per capita of poorest 20% of population, percentage women borrowers DEA outputs: number of active women borrowers, poorest benefit dummy

DEA outputs: gross loan portfolio, financial revenue

Mersland and Strøm (2010): Microfinance mission drift?

Panel data estimation with instruments

Ratingfund, 379 MFIs, 74 countries, 1998–2008

Avg. loan size, main market, lending methodology, gender bias dummy

Avg. profit, avg. operational cost, Portfolio-at-Risk 30 days, age, assets

Hermes et al. (2011): Outreach and efficiency of microfinance institutions

Stochastic frontier analysis

MIX Market, 435 MFIs, ln(avg. loan size), unknown number of percentage of female countries, 1997–2007 borrowers

Cost function: total costs

Quayes (2012): Depth of outreach and financial sustainability of microfinance institutions

OLS, logistic MIX Market, 702 MFIs, Avg. loan size/GNI, regression, 3SLS 83 countries, 2006 percentage of women borrowers

Low positive relationship between outreach and financial efficiency; except one, no socially efficient but financially inefficient MFIs exist, NGOs more outreach No evidence of trade-off. In contrast, more financial efficiency could lead to more outreach Trade-off: more financial efficiency implies less outreach since more efficiency can only be achieved by focusing less on poorest clients No trade-off except for low-disclosure MFIs

neurons. A learning rate a(t) and a neighborhood function hcr(t) are defined as parameters of the learning function: mr ðt þ 1Þ ¼ mr ðtÞ þ aðtÞhcr ðtÞ½ni ðtÞ  mr ðtÞ:

ð2Þ

The learning-rate will influence the magnitude of the BMU’s adaptation after matching with an input vector ni, whereas the neighborhood function defines the range of influence of the adaptation. In order to guarantee the stability of the final output map, decreasing learning rates and neighborhood functions are often used at the end of the training. The obtained neurons are then projected on a two-dimensional map. Each neuron has then a representation in the high and low dimensional spaces. In Section 6, component planes are shown representing the relative values of the different neurons for the different variables on a two-dimensional map, providing the analyst with a powerful visualization facility. An exhaustive discussion of the projection approach and the influence of the parameters such as the number of neurons, the shape of the map, or the initial weights of the neurons is to be found in Kohonen (1995). (b) Two-step clustering In the application of Section 6, a two-step clustering approach is used in order to capture the structure of the data

Gross loan portfolio, total equity, debt to equity, total expense ratio, cost per borrower, Dummy = 1 if operational selfsufficiency, loan loss reserve ratio >1

and to evaluate the research question discussed in Section 1, which deals with a potential association between social impact and financial performance. The first step consists of the application of the SOM algorithm to the data set described in Section 4. The output of this step is a set a neurons summarizing the structure of the data. The neurons have the same dimensionality as the input data set and can be considered as prototypes of MFIs. Using the component planes as shown in what follows, it is possible to visualize the structure of the data and to draw interesting conclusions. As will be seen in Figure 3, some neurons are sharing the same characteristics and are grouped together on the output map, giving the analyst the possibility to identify areas having some specific properties. Although it is possible to visualize those areas and to make the analysis by only using the SOM output, it is advised to use a second clustering step in order to capture those substructures and to analyze them (Vesanto & Alhoniemi, 2000). That is why a second clustering step is applied using the well-known k-means clustering. The neurons resulting from the SOM algorithm are used as input and grouped using the k-means algorithm into a predefined number of clusters obtained using the Davies–Bouldin index as a measure of cluster quality (Davies & Bouldin, 1979). The Davies–Bouldin index is calculated as:

FINANCIAL EFFICIENCY AND SOCIAL IMPACT OF MICROFINANCE INSTITUTIONS USING SELF-ORGANIZING MAPS

index ¼

  c 1X ri þ rj ; maxi–j c i¼1 dðci ; cj Þ

ð3Þ

where c is the number of clusters, cy is the centroid of cluster y, ry is the average distance of all elements of cluster y and d(ci,cj) is the distance between the centroid of cluster i and cluster j. Since a good partitioning corresponds to a situation where the intra-cluster distances are low and the inter-cluster distances are high, the lower the Davies–Bouldin index, the better the partitioning obtained. The number of clusters leading to the best partitioning is thus chosen as a parameter of the k-means. The output of this step is a set of cluster centroids, which are averaged vectors characterizing the different clusters. Figure 2 shows the three steps of the methodology used in this paper.

201

Firstly, only MFIs that supply at least audited financial statements are included in our sample (i.e. P4 diamonds). 1 This is in line with Hartarska and Nadolnyak (2007) but contrasts with Quayes (2012), who also includes three diamonds MFIs. Secondly, although our main analysis will be performed on data of 2011, we considered it important to limit the data set to established MFIs. In order to do so, only MFIs that report without gaps during three consecutive years (i.e. 2009, 2010, and 2011) are incorporated. Thirdly, notwithstanding the strict data quality requirement (i.e. at least audited financial statements), we had to remove 126 observations after validating the data sample using a set of logical business rules. Eventually, our sample includes 650 MFIs. Data are available for six regions that cover 88 countries. (b) Outreach variables

4. DATA SET AND DESCRIPTION OF THE VARIABLES (a) Data set The data are gathered from the Microfinance Information Exchange database. The database mainly contains financial, operational, and social performance information from MFIs in all regions of the developing world and has been used in Gutie´rrez-Nieto et al. (2009), Hermes et al. (2011), Quayes (2012). At the end of 2011, the database included over 1,200 MFIs. The data have been standardized across the industry to make comparisons feasible. All data are reviewed and validated against a set of business and audit rules (Microfinance Information Exchange, 2010a). MFIs can voluntarily decide to be incorporated. Hence, the sample could be skewed toward better performing MFIs (Cull, Demirgu¨cß-Kunt, & Morduch, 2011). For our analysis, we limit the sample to established MFIs that provide externally verified data to the MIX database.

Outreach is about bringing the social benefits of microfinance to poor clients (Schreiner, 2002). The most widely used dimensions are depth and breadth of outreach. Depth of outreach refers to serving the poorest people, whereas breadth refers to serving large numbers of people, even if they are relatively less poor (Schreiner, 2002; Brau & Woller, 2004). Although the majority of microfinance clients are women, outreach to women is also often considered (Cull et al., 2007; Quayes, 2012). Given the tremendous growth of the microfinance industry, breadth of outreach has obviously increased industry wide (Quayes, 2012). Besides, since the majority of microfinance clients are women, depth of outreach has become the de facto dimension to measure social impact. In the hypothetical situation of full information availability, one would measure depth of outreach by aggregating the personal equity of each individual borrower and testing whether MFIs do provide loans to borrowers without a lot of wealth. Because this information is not easily obtainable, many authors, including Cull

Figure 2. Stepwise representation of the used methodology.

202

WORLD DEVELOPMENT Table 2. Description of the continuous variables

N Mean Std. dev. Min Q1 Q2 Q3 Max Skewness Kurtosis

Average loan size p.borr./GNI p.c. O1

Portion of women borrowers O2

Real yield on gross portfolio

Profit margin

F1

613 0.475 0.634 0.011 0.104 0.255 0.594 5.148 3.24 16.88

589 0.652 0.252 0.003 0.452 0.625 0.9 1 0.10 1.96

603 0.240 0.175 0.109 0.127 0.199 0.299 1.050 1.68 6.64

et al. (2007), Kai (2009), Mersland and Strøm (2010), Hermes et al. (2011), Quayes (2012), use the average loan size (often divided by GDP or GNI per capita) as a proxy for depth of outreach. Smaller amounts are believed to indicate greater depth. We have included three measures of outreach in this paper: depth of outreach (O1), outreach to women (O2), and breadth of outreach (O3). The average loan size per borrower/GNI per capita (depth of outreach) and the portion of women borrowers (outreach to women) are continuous variables whereas the breadth of outreach is categorical. Columns two and three of Table 2 provide the reader with information on the depth of outreach and outreach to women of our data set. Information on the frequencies of breadth of outreach can be found in Table 3. It is apparent that the sample consists of a heterogeneous group of MFIs. For instance, one MFI grants, on average, loans that are just 1.1% of the gross national income per capita whereas at the other extreme, one MFI’s average loan size is more than the fivefold of the gross national income per capita. However, the median average loan size per borrower to GNI per capita is just 0.255, indicating that most MFIs provide relatively small loans to borrowers. The discrepancy between the mean and median of depth of outreach due to outliers is also reported by Cull et al. (2007), Mersland and Strøm (2010), Quayes (2012) 2. Outreach to women is fairly high, which is in line with previous research (Cull et al., 2007; Hermes et al., 2011; Quayes, 2012). (c) Financial variables The 1980s were marked by the increasing attack on donorand government-funded credit programs as evidence mounted of poor outreach, low repayment rates, high administrative costs, widespread corruption, and a ‘grant mentality’ among clients owing to heavy subsidization (Adams et al., 1984; Robinson, 2001). The lack of efficiency and the profusion of corruption became so apparent to donors that the traditional subsidized credit delivery approach was challenged by a new premise, which believed that the only viable way to efficiently satisfy the unmet demand for microcredit would be by focusing on sustainability (Robinson, 2001). The increased focus on self-sustainability meant a schism in the management of microfinance institutions. While MFIs in the past relied greatly on subsidies just to be able to conduct their core businesses and to help cover costs, they now had to focus more on generating sufficient income from their operations as well as operating more efficiently (Woller, Dunford, & Woodworth, 1999; Morduch, 2000).

Cost per loan (in $)

F2

Gross loan portfolio to total assets F3

Debt to equity ratio

F4

Portfolio at risk >30 days F5

605 0.063 0.439 5.839 0.022 0.109 0.206 0.717 7.44 79.06

611 0.770 0.142 0.032 0.706 0.801 0.864 0.999 1.49 6.42

574 217 255 5 65 160 263 2025 3.45 18.84

581 0.064 0.103 0 0.015 0.038 0.077 1 5.64 43.88

611 4.135 5.296 24.2 1.46 3.07 5.66 62.91 5.40 53.19

F6

In order to investigate the impact of the increased focus on profitability and efficiency, we have selected six important financial variables: two profitability variables (F1–F2), two efficiency variables (F3–F4), and two related variables (F5– F6). These variables were drawn from the SEEP Network (2010) standardized approach to measure and analyze financial performance and risk management to ensure comparability between this paper and previous/future research. This standard is endorsed by over hundred organizations, such as CGAP and MIX Market (SEEP Network, 2010). Furthermore, we have tried to select variables that were used in previous research. Table 2 lists important descriptive statistics on the used variables. The profitability ratios compare MFI’s earnings, expenses, and other relevant costs to judge whether an institution will be able to continue serving its clients beyond a short-term period. The real yield on the gross portfolio (F1) indicates how much interest, fees, and commissions a MFI generates from its average gross loan portfolio (SEEP Network, 2010). Selfreliant MFIs are believed to set their interest rates high enough (Morduch, 2000). This is essential for any business that intends to continue its operations beyond the short-term. Subsidized MFIs were often requested to enforce an interest rate ceiling, which was usually set at a level less than that required to cover its costs. This in turn required often re-capitalization by donors, and discouraged private investors from supporting the industry, thus reducing the feasibility to expand operations (Fernando, 2006). We find that half of the sample collects a real yield between 12.7% and 29.9%. The mean/median real yield on the gross portfolio is equal to 24.0%/19.9%, which is lower than the number reported by Cull et al. (2007) (35.4%/30.5%). This decrease could be due to the fact that the industry has become more competitive, forcing MFIs to charge lower rates. The profit margin (F2) indicates how much of financial revenue (i.e. revenues from the loan portfolio and from other financial assets) is kept as net operating income. As such, it is a good measure of cost control and profitability. On average, MFIs realize a profit margin of 6.3%. The median profit margin is however higher (10.9%) because 111 MFIs report a negative profit margin. That means that they cannot pay their financial and operational expenses with revenues from their loan portfolio and from other financial assets, which is of course not sustainable. Efficiency ratios explain whether a MFI is serving as many customers as possible while keeping its costs under control. Variable F3 measures how much of its assets a MFI allocates to making loans. Table 2 reports that most MFIs allocate around three-quarters of their assets to their portfolio.

FINANCIAL EFFICIENCY AND SOCIAL IMPACT OF MICROFINANCE INSTITUTIONS USING SELF-ORGANIZING MAPS

203

Table 3. Description of the other variables Name

Frequency

Percentage

Breadth (O3)

618

N

Small Medium Large

Categories

257 148 213

41.59 23.95 34.47

Region (R1)

650

Africa East Asia and the Pacific Eastern Europe and Central Asia Latin America and The Caribbean South Asia

91 81 103 269 106

14.00 12.46 15.85 41.38 16.31

Type (R2)

650

Loans only Loans and deposits

295 355

45.38 54.62

Legal status (R3)

640

Bank Credit Union/Cooperative Non-bank financial institution (NBFI) Non-governmental organization (NGO) Other

65 72 220 250 33

10.16 11.25 34.38 39.06 5.16

Age (R4)

633

New and young Mature

125 508

19.75 80.25

Scale (R5)

649

Small Medium Large

157 185 307

24.19 28.51 47.30

However, 31 MFIs allocate less than half, pointing to nonessential (i.e. inefficient) use of assets. We obtain the same results as Cull et al. (2007). The cost per loan ratio (F4) shows again that there exist important differences between MFIs. While the median MFI incurs a cost of $151 per loan, 122 MFIs are able to provide loans at a cost of less than $50, in contrast to 45 MFIs that face a cost of over $500 per loan. Finally, related variables F5 and F6 are included. The portfolio at risk >30 days ratio (F5) measures the portfolio quality by calculating the portion of the portfolio whose payments are more than 30 days past due. A higher number is a potential sign of trouble. The debt to equity ratio measures a MFI’s financial leverage. A high number could indicate that the MFI has financed its growth with debt, reducing its ability to absorb unexpected losses. Two MFIs report a negative number, pointing to serious financial troubles as the book value of equity is negative. (d) Other variables Five MFI specific categorical variables are included in the analysis. Table 3 contains frequency statistics of these variables. Most MFIs in the sample are operational in Latin America and The Caribbean. With regard to the services MFIs offer, we find that around half only grants loans and half grants loans and offers deposit facilities. Voluntary savings are fundamental to poverty reduction as they allow households to smooth consumption, to reduce the holding of cash assets, and to better prepare for emergencies (Morduch, 2000). Savings mobilization programs have demonstrated that low income people will opt to start depositing their cash holdings if they have access to convenient and trustworthy saving facilities (Branch & Klaehn, 2002). Furthermore, the deposits can be used as an inexpensive source of funding for providing microcredit. As such, MFIs that offer deposit facilities may achieve a better social performance. The age variable captures how long the MFI has been in business. When age is new and young, a MFI has operated up to eight years. A MFI categorized as mature has operated

more than eight years. Given the strict data quality criteria (at least audited financial reports for three consecutive years), it is not surprising that the majority of MFIs in the sample are labeled mature. Lastly, the scale variable reports how large a MFI’s portfolio is depending on the region in which it operates. When a MFI is categorized as small, its portfolio is less than $2 million in all regions except Latin America and The Caribbean (LAC) and less than $4 million in LAC. Similarly, a medium MFI has a portfolio between $2 and $8 million in all regions except LAC and between $4 and $15 million in LAC. When the MFI’s portfolio exceeds $8 million in all regions except LAC and $15 million in LAC, it is categorized as large. 5. DATA PROCESSING All variables described in Section 4 are used as input to the SOM analysis. However, continuous variables are categorized in two groups: low and high. The median is chosen to be the cut-off point between the two categories. Values below the median are categorized as low and values at and above the median are categorized as high for all variables except depth of outreach (O1) where values below the median are categorized as high and values at and above the median were categorized as low. The rationale for discretization of the continuous variables is that the distribution of these variables confirms the existence of extreme values and outliers (see Table 2). Since other authors kept the outliers (e.g. Cull et al., 2007) and thus in the interest of comparability, we decided that removing this outliers is not advisable. 6. RESULTS (a) Concise exploration of SOM maps (i) Introduction Highly dimensional data, such as the sample used in this paper, are hard to visualize. A self-organizing map (SOM)

204

WORLD DEVELOPMENT

attempts to overcome this difficulty using a two dimensional map which plots the structure of the data by grouping similar observations together and dissimilar observations apart. Such an ordered display facilitates interpretation of the data set structures. The relative value in the different regions of the self-organizing map is visualized by different colors. In some areas the observations will have a high relative value compared to the sample for that particular variable, and subsequently this area will be black. However, in other areas, the relative value will be low and consequently, this area’s color will be almost white. In sum, one extreme is black, values near the middle are gray, and the other extreme is almost white. The SOM output of Figure 3 may look challenging to understand at first sight. To facilitate interpretation, the SOM’s area is divided into nine equally spaced parts using a grid. A1 and C3 denote the upper-left and lower-right corner, respectively. Before fully discussing the results, we will introduce the reader using an example. Banks (Legal status variable R3) represent around 10% of the sample. They are mainly located in region C3. Next, we would like to analyze the properties of these banks. In order to do so, we must remember where these banks are located in the SOM area (recall region C3). With respect to age (variable R4), we notice that banks are clearly mature because the equivalent location (again region C3) is black (i.e. high relative value) for the category mature and almost white (i.e. low relative value) for the category new and young. Next, we find that banks are mainly located in Eastern Europe and Central Asia, and Latin America and the Caribbean (region variable R1). Furthermore, scale variable R5 reveals that banks are, not surprisingly, large. Variable R2 about the type of MFI shows that banks offer both loans and deposit services. (ii) Outreach variables Three outreach variables are included in the SOM analysis: depth of outreach O1 (low-high), outreach to women O2 (low-high), and breadth of outreach O3 (small-medium-large). It appears that there exists a very clear relationship between the depth of outreach and the percentage of women borrowers. This relationship is also reported in Hermes et al. (2011), Quayes (2012). MFIs that combine a high depth of outreach and a high outreach to women are often NGOs and Non-bank financial institution (NBFIs). They tend to be located in East Asia and the Pacific, Latin America and the Caribbean, Middle East and North-Africa, and South Asia. However, in Latin America and the Caribbean some MFIs also score poor with regard to depth of outreach and female borrowers. Compared to the other geographical regions, Eastern European and Central Asian, and African MFIs do not perform well. Breadth of outreach relates orthogonally to depth of outreach and outreach to women (i.e. the regions with high/low relative value of O1 and O2 are unrelated to the regions with high/low relative value of O3). This implies that breadth is not a determinant of depth of outreach. As such, the potential hypotheses that more breadth of outreach results in less depth, and vice versa, as mentioned in Schreiner (2002), are refuted. For instance, in South Asia, MFIs are able to combine high depth and high breadth of outreach. In contrast, in Eastern Europe and Central Asia some MFIs neither achieve depth nor breadth of outreach. MFIs with a low breadth are often NBFIs and NGOs that operate mainly in Eastern Europe and Central Asia, and Latin America and the Caribbean. (iii) Financial variables The analysis of the financial variables is less straightforward compared to the interpretation of the outreach variables due

to the increased heterogeneity. As mentioned in the introduction, the used SOM technique is able to fully capture this heterogeneity. This is a major advantage over regular statistical techniques that calculate results “on average”, hence disregarding potentially interesting information when the data set is heterogeneous. Six financial variables are used as inputs to the SOM analysis. A high yield (variable F1) is charged to lenders by the majority of new and young MFIs. These MFIs are often NBFIs and NGOs, in contrast to banks and credit unions & cooperatives, which charge a lower yield. Variable F2, profit margin, is a good indicator of cost control and profitability. African MFIs score poorly with respect to profit margin. The proportion of the assets that is allocated to the loan portfolio (variable F3) captures the efficiency of a MFI. The SOM analysis shows that making a large share of the assets available to the loan portfolio is often associated with a higher profit margin. Cost of loans (F4) is also an important efficiency measure. Perhaps not surprisingly, the majority of MFIs that incur a low cost per loan achieve a high depth of outreach, which could indicate that a low cost per loan is a condition to achieve depth of outreach. With regard to the portfolio at risk >30 days variable (F5), it can be noticed that a high number is somewhat related to a lower allocation of assets to the loan portfolio. Finally, MFIs with a high debt to equity ratio (variable F6) tend to be larger. (b) Cluster description Next, we assign the different MFIs into groups which are called clusters so that the MFIs in the same cluster are more homogeneous to each other than to those in other clusters. The methodology used to construct the clusters was explained in Section (b). Eventually, 13 clusters are designed. Figure 4 plots the cluster structure. Recall that this plot can just be thought as a layer on top of the original SOM plots. As such, each position on the cluster figure is equal to the same position on the SOM plot. In order to analyze our research questions, we determine the properties of each cluster’s centroid, which is its geometric center. The clusters will be discussed based on the number of MFIs they contain. Cluster 6 includes 107 MFIs. These MFIs operate as NBFIs and NGOs in South Asia. Their outreach to women is high and they are able to achieve a high depth of outreach while their breadth of outreach is medium to large. Their real yield is not high and most of them obtain a low margin. However, these MFIs incur a low cost per loan. Sixty-eight MFIs are covered by cluster 9. These large and mature operate MFIs have the following traits: they perform bad on the outreach dimension with depth and outreach to women being poor. However, breadth of outreach is medium to large. From a financial perspective, this cluster has a high cost per loan, a low yield, a low portfolio to assets ratio, and a high debt/equity ratio. As such, these MFIs are socially as well as financially inefficient. The next cluster, cluster 8, contains 59 MFIs. These small MFIs provide loans only and are located in Latin America and The Caribbean. They do not achieve an excellent outreach since breadth of outreach is small. From a financial perspective, these MFIs do not perform very well. They achieve a low profit margin in spite of charging a high real yield. They do not allocate a lot of their assets to their portfolio and their portfolio at risk is high. Cluster 5 can be described as performing well with respect to outreach (high depth, high percentage of women borrowers, and a good breadth of outreach). The 58 MFIs in this cluster, which are mature and small to medium, are able to work financially

FINANCIAL EFFICIENCY AND SOCIAL IMPACT OF MICROFINANCE INSTITUTIONS USING SELF-ORGANIZING MAPS

205

Figure 3. SOM output.

efficient. They are characterized by a low portfolio at risk, a high real yield, and a low cost of loans. This cluster proves that it is possible to combine social performance and financial sustainability. The subsequent cluster, cluster 4, encompasses 50 MFIs that achieve a very bad outreach. Depth of outreach and outreach to women are low, and breadth of outreach is small. They charge a lower interest rate, have a high profit margin, and the cost of a loan is high. In the previous paragraphs, we have discussed five clusters that cover more than half of all the MFIs in the sample

(52.62%). The other clusters are smaller and will be summarized in Table 4. This table and its interpretation will be discussed in the next section. (c) Social and financial impact assessment As mentioned in the introduction, “microfinance institution” is just an umbrella term that covers a wide array of substantially different institutions. Using self-organizing maps, we accounted for this heterogeneity and generated

206

WORLD DEVELOPMENT

Figure 4. Cluster structure.

two-dimensional maps where similar MFIs are mapped close together and dissimilar more apart. In a next step, clusters were formed. Each cluster is determined more by some of the input variables. For instance, a cluster is determined by two social variables and four financial variables. This implies that all MFIs in this cluster have the same values for those variables. The remaining three variables are not significantly shared by the MFIs of the cluster. In a next step, we have to determine the properties of each cluster with regard to its social and financial efficiency. In order to do so, we created an ordinal four-point scale, ranging from ++: very good to : very bad. Intermediate steps are +: good and : bad. A cluster was categorized as very good if it, compared to the other clusters, excelled at all variables. For instance, a very good social impact cluster could score excellent on all outreach variables. As such, good and bad clusters are identified by the majority of variables. For example, a financially bad MFI could perform substandard on four financial variables and perform above standard on the other two financial variables. The result of each cluster with regard to its social and financial dimension is summarized in Table 4. The clusters (first column) are sorted in descending order of the number of observations (i.e. MFIs) in column two. The column labeled percent shows the percentage of MFIs in the dataset that are in a particular cluster. The two next columns calculate the cumulative number of observations and the cumulative percentage. The last two columns indicate how well the cluster performs with regard to social and financial performance.

The 13 clusters vary substantially. For instance, cluster 6 performs very well on the social dimension but the result with regard to the financial performance is less impressive. In contrast, cluster 9 scores substandard on both dimensions. Table 5 summarizes the results. The left panel ranks the clusters with respect to their social and financial performance. For instance, both clusters 12 and 7 perform badly on both dimensions. These two clusters contain a total of 90 MFIs (last column). The right panel of Table 5 provides the reader with a twoway contingency table. At first sight, it is apparent that most of the MFIs are located on or close to the diagonal. Extremes, such as ++ (i.e. very good) for one dimension and  (i.e. very bad) for the other dimension do not exist. To test this more formally, we compute several statistics that describe the association between the social and the financial dimension of the two-way table of frequencies. Based on simple probability, the expected number of MFIs for each cell can be calculated. The greater the difference between the observed and expected cell counts, the stronger the evidence that the two dimensions are actually associated. As such, it becomes less likely that the null hypothesis of independence holds. In a first step, we calculate the Pearson v2 for the hypothesis that the rows and columns in the right panel of Table 5 are independent. Since v2(9) = 248.35 obviously exceeds the critical value, we can reject the null hypothesis and conclude that social impact is indeed associated with financial performance. In a second step, we want to test the strength of the existing association. The first test we use, the Goodman and Kruskal’s gamma, calculates the number of concordant and discordant pairs of observations to measure the strength of association where 1 indicates perfect negative association and +1 perfect agreement. No association is found when gamma is zero. Concordant pairs score well/ badly on both dimensions whereas discordant pairs perform well for one dimensions but badly for the other dimension. When there is a predominance of concordant pairs, gamma will be closer to +1. We find that gamma is equal to 0.3770, with an asymptotic standard error of 0.041. The value of gamma indicates that there exists a positive relationship between social and financial efficiency. The second test is Kendall’s sb, which is similar to gamma except that it uses a correction for ties. Again, the strength of association lies between 1 and +1. Kendall’s sb amounts to 0.2679, with an asymptotic standard error of 0.030. The null hypothesis that social and financial performance are independent is again outright rejected. The positive relationship shows that a good financial performance is associated with a good outreach. Likewise, low social efficiency is associated with poor financial governance. Panel b of Table 5 can be represented graphically to facilitate interpretation. A good method is a mosaic plot, which

Table 4. Cluster properties with respect to outreach and financial performance Cluster

Obs

Percent

Cum.

6 9 8 5 4 12 7 3 2 11 1 10 13

107 68 59 58 50 48 42 40 39 37 37 35 30

16.46 10.46 9.08 8.92 7.69 7.38 6.46 6.15 6.00 5.69 5.69 5.38 4.62

107 175 234 292 342 390 432 472 511 548 585 620 650

Cum. Perc.

Social performance

Financial performance

16.46 26.92 36.00 44.92 52.62 60.00 66.46 72.62 78.62 84.31 90.00 95.38 100

++  + ++      ++  + +

   + +     ++ + + 

FINANCIAL EFFICIENCY AND SOCIAL IMPACT OF MICROFINANCE INSTITUTIONS USING SELF-ORGANIZING MAPS

207

Table 5. Tables of frequency counts by cluster for social and financial performance (a) Summary of performances by cluster Soc. perf Fin. perf. ++    +    ++ +  +     ++ ++  + + + (b) Two-way table of frequencies of MFIs Social performance  40 90 50 0 180

Obs.

6 12.7 8,13 9 5 4 3 2 11 1 10

107 90 89 68 58 50 40 39 37 37 35

 68 39 37 0 144

consists of groups of rectangles that represent the cells in the contingency table. Each cell in the right panel of the table is represented by a rectangle whose area is equal to the number of MFIs in that cell. The width of a rectangle is the same for all rectangles in a particular column and is equal to the fraction of MFIs that fall into that specific outreach category. The height of the rectangle is the proportion of MFIs that, given a certain outreach category, fall into a specific financial performance category. The color of the rectangle indicates the financial performance category. For instance, the bottom left rectangle consists of 40 MFIs that perform very badly on social and financial performance. The width of the rectangle is equal to the proportion of all MFIs that have a very bad outreach (i.e. 180/650) and the height amounts to the fraction of these MFIs that perform very bad on financial efficiency (i.e. 40/180). The mosaic plot in Figure 5 graphically confirms the previous findings. MFIs that are unable to achieve a good social impact are often performing substandard with respect to the financial dimension (bottom left corner). In contrast, MFIs that are characterized as having a good outreach usually perform well financially (top right corner). 7. VALIDATION Validation tries to demonstrate that the obtained results scientifically answer the research question. We will apply two types of validation: historical validation using data from 2010 and 2009, and statistical validation using a different categorization (three and four categories instead of two). (a) Historical validation This paper uses data from the Microfinance Information Exchange database. Recall from Section (a) that the sample was limited to established MFIs that provide externally verified data. All included MFIs had to supply at least audited financial statements. Furthermore, only MFIs that reported without gaps during three consecutive years (i.e. 2009, 2010, and 2011) were incorporated. After validating the data sample

+ 0 89 35 0 124

++ 0 107 58 37 202

Total 108 325 180 37 650

Fraction by social performance .25 .5 .75

0

1

1 Fraction by financial performance

Fin. perf.   + ++ Total

Cluster(s)

.75 ++ + --

.5

.25

0 --

-

+

++

Outreach

Figure 5. Mosaic plot.

using a set of logical business rules, the sample included 650 MFIs. Whereas the main analysis used the data from 2011, we will now test the research question on data of 2010 and 2009. Again, all variables described in Section 4 are used as input to the SOM analysis. Like the 2011 analysis, continuous variables were categorized in two groups: low and high. The median was chosen to be the cut-off point between the two categories. Values below the median were categorized as low and values at and above the median were categorized as high for all variables except depth of outreach (O1) where values below the median were categorized as high and values at and above the median were categorized as low. Similar to the main analysis, the properties of each cluster were determined using an ordinal four-point scale, ranging from ++: very good to : very bad. Intermediate steps are +: good and : bad. Table 6 contains two two-way frequency tables that count the number of MFIs in each social/financial performance

208

WORLD DEVELOPMENT Table 6. Tables of frequency counts for social and financial performance in 2010 and 2009

Fin. perf.

Social performance 



+

++

Total

(a) Two-way table of freq. of 2010  72  0 + 27 ++ 36 Total 135

45 60 85 0 190

0 72 105 0 177

0 0 89 59 148

117 132 306 95 650

(b) Two-way table of freq. of 2009  76  46 + 63 ++ 0 Total 185

68 72 17 0 157

0 45 120 0 165

0 0 108 35 143

144 163 308 35 650

Table 7. Tables of frequency counts for social and financial performance using 3 and 4 cat. in 2011 Fin. perf.

Social performance 



+

++

Total

(a) Two-way table of freq. using three categories  100  146 + 0 ++ 0 Total 246

86 0 34 0 120

0 72 77 0 149

0 0 65 70 135

186 218 176 70 650

(b) Two-way table of freq. using four categories  139  0 + 49 ++ 45 Total 233

45 0 122 0 167

0 103 0 68 171

0 19 60 0 79

184 122 231 113 650

category. For both years, the same association statistics are calculated. In 2010, v2(9) = 424.92,c = 0.5225 (ASE= 0.046), and sb = 0.3965 (ASE = 0.036) and in 2009, v2(9) = 414.67, c = 0.7221 (ASE = 0.029), and sb = 0.5443 (ASE = 0.025). The same conclusion as in the main analysis holds, social impact is indeed associated with financial performance.

Table 8. Mean quantification error summary Year

# of categories

MQE

2011 2011 2011 2010 2009

2 3 4 2 2

1.3967 1.7536 1.9446 1.4312 1.4176

(b) Validation using a different categorization In contrast to the main analysis, all continuous variables except depth of outreach (O1) are now categorized in three or four groups. Each group contains the same number of MFIs. Using three or four equal frequency grouping intervals could add more information to the analysis. Similar to the main analysis, the properties of each cluster were determined using an ordinal four-point scale, ranging from ++: very good to : very bad. Intermediate steps are again +: good and : bad. Table 7 contains two two-way frequency tables that count the number of MFIs in each social/financial performance category. Again, association statistics are produced. Using three categories, v2(9) = 704.70, c = 0.7592 (ASE = 0.022), and (ASE = 0.018). Four categories yield sb = 0.6188 v2(9) = 654.20, c = 0.2932 (ASE=0.040), and sb = 0.2392 (ASE = 0.033). We again confirm the findings of the main analysis hold, social impact is indeed related to financial performance.

(c) Mean quantification error Finally, in Table 8, we provide the reader with the mean quantification error (MQE) of each model. The idea behind the MQE is explained in Section (a) but recall that it is the average of the euclidean distances between every input vector and its respective closest neuron. The MQE depends on the scale of the input variables but a low MQE value indicates a good representation of the input by the SOM is achieved. As can be seen in the table, increasing the number of categories results in a higher MQE, which makes sense because the number of neurons stays equal. 8. CONCLUSION From the 1980s onward, microfinance institutions had to start dealing with a changing operational environment that re-

FINANCIAL EFFICIENCY AND SOCIAL IMPACT OF MICROFINANCE INSTITUTIONS USING SELF-ORGANIZING MAPS

quired more cost control and efficiency. This schism in the management of microfinance institutions was primarily caused by the mounting evidence of disappointing performance, both socially and financially. Proponents of this change argued that more financial control would ultimately allow more social impact. Besides this evolution, greater competition and more private sector interest further demanded cost control and efficiency gains. In recent years, some microfinance policymakers alleged that the increased focus on financial efficiency has pushed the quest for social impact into the background. For instance, Muhammad Yunus stirred up the debate by claiming that “[. . .] a worrying “mission drift” in the motivation of those lending to the poor. Poverty should be eradicated, not seen as a money-making opportunity.” (Yunus, 2011). Academic literature on this subject is sparse and before 2007 often consisted of just anecdotal evidence (Hermes et al., 2011). A major difficulty in analyzing MFI data is the diversity between institutions. “Microfinance institution” is merely an umbrella term. As such, any statistical technique that endeavors to provide insights in the research question must be able to deal with this heterogeneity. We have used self-organizing maps (SOM) to graphically plot the heterogeneity among MFIs with regard to the different input variables. In a next

209

step, the results from the SOM algorithm are used as input and grouped using the k-means algorithm into a predefined number of clusters obtained using the Davies–Bouldin index as a measure of cluster quality. Subsequently, the centroids of these clusters were analyzed with respect to their social and financial performance using an ordinal four-point scale ranging from ++: very good to : very bad. Several statistics that measure the degree of association between the social and the financial dimension were calculated using the differences between the observed and expected cell counts. The analysis was performed on MIX Market data of 2011 using a data sample of 650 MFIs that satisfy several data quality requirements. We find that the hypothesis of independence between financial performance and social outreach is rejected, pointing toward an association between the two dimensions. However, our results do not suggest a trade-off. On the contrary, we find evidence of a significant, positive relationship between social and financial efficiency. Consequently, the theory that social performance comes at the prize of less financial performance does not seem to be valid. This result supports the conclusion of Gutie´rrez-Nieto et al. (2009), Mersland and Strøm (2010), and Quayes (2012) but contrast with the findings of Cull et al. (2007) and Hermes et al. (2011).

NOTES 1. The level of data reliability is expressed in diamonds and ranges between one and five. The higher the number of diamonds, the higher the level of data quality (Microfinance Information Exchange, 2010b).

2. Mersland and Strøm (2010) do not divide the average loan size per borrower by the GNI per capita. Quayes (2012, Table 1) reports he uses the average loan size per borrower while he actually uses the average loan size per borrower to GNI per capita.

REFERENCES Adams, D. W., Graham, D. H., & Von Pischke, J. D. (1984). Undermining rural development with cheap credit. Westview Press. Bateman, M. (2010). Why Doesn’t Microfinance Work? : The Destructive Rise of Local Neoliberalism (p. 256). London: Zed Books. Branch, B., & Klaehn, J. (2002). The keys to striking the balance: An introduction to savings mobilization. In Striking the balance in microfinance: A practical guide to mobilizing savings (p. 382). Washington, DC: Pact Publications. Brau, J. C., & Woller, G. M. (2004). Microfinance: A comprehensive review of the existing literature. Journal of Entrepreneurial Finance and Business Ventures, 9, 1–26. Carlei, V., Marra, A., & Pozzi, C. (2012). Public governance, human capital and environmental outcomes: an analysis based on selforganizing maps. Environmental Policy and Governance, 22, 116–126. Chen, M.-Y. (2012). Visualization and dynamic evaluation model of corporate financial structure with self-organizing map and support vector regression. Applied Soft Computing, 12, 2274–2288. Chen, N., Ribeiro, B., Vieira, A., & Chen, A. (2013). Clustering and visualization of bankruptcy trajectory using self-organizing map. Expert Systems with Applications, 40, 385–393. Christen, R.P. (2001). Commercialization and mission drift: The Transformation of Microfinance in Latin America. Technical Report 5 Consultative Group to Assist the Poor. Conning, J. (1999). Outreach, sustainability and leverage in monitored and peer-monitored lending. Journal of Development Economics, 60, 51–77. Cull, R., Demirgu¨ßc-Kunt, A., & Morduch, J. (2007). Financial performance and outreach: a global analysis of leading microbanks. The Economic Journal, 117, F107–F133. Cull, R., Demirgu¨ßc-Kunt, A., & Morduch, J. (2011). Does regulatory supervision curtail microfinance profitability and outreach?. World Development, 39, 949–965.

Davies, D., & Bouldin, D. (1979). A cluster seperation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-1, 224–227. Fernando, N.A. (2006). Understanding and dealing with high interest rates on microcredit. Technical Report. Gutie´rrez-Nieto, B. n., Serrano-Cinca, C., & Mar Molinero, C. (2009). Social efficiency in microfinance institutions. Journal of the Operational Research Society, 60, 104–119. Hartarska, V., & Nadolnyak, D. (2007). Do regulated microfinance institutions achieve better sustainability and outreach? Cross-country evidence. Applied Economics, 39, 1207–1222. Hermes, N., & Lensink, R. (2011). Microfinance: Its impact, outreach, and sustainability. World Development, 39, 875–881. Hermes, N., Lensink, R., & Meesters, A. (2011). Outreach and efficiency of microfinance institutions. World Development, 39, 938–948. Hollis, A., & Sweetman, A. (1998). Microcredit: What can we learn from the past?. World Development, 26, 1875–1891. Huysmans, J., Martens, D., Baesens, B., Vanthienen, J., & Gestel, T. (2006). Country corruption analysis with self organizing maps and support vector machines. In H. Chen, F.-Y. Wang, C. Yang, D. Zeng, M. Chau, & K. Chang (Eds.), Intelligence and security informatics. Lecture notes in computer science (Vol. 3971, pp. 103–114). Berlin, Heidelberg: Springer. Jolliffe, I. (2005). Principal component analysis. In Encyclopedia of statistics in behavioral science. John Wiley & Sons Ltd. Kai, H. (2009). Competition and wide outreach of microfinance institutions. Economics Bulletin, 29, 2628–2639. Kohonen, T. (1995). Self-organizing maps. Springer. Kohonen, T. (2001). Self-organizing maps. Springer series in information sciences series. Springer-Verlag GmbH. Mersland, R., & Strøm, R. Ø. y. (2010). Microfinance mission drift?. World Development, 38, 28–36. Microfinance Information Exchange (2010a). About MIX.

210

WORLD DEVELOPMENT

Microfinance Information Exchange (2010b). Frequently Asked Questions (FAQs). Microfinance Information Exchange (2010c). Profiles & Reports j MIX Market. Microfinance Information Exchange (2012). Cross-market analysis. Morduch, J. (1999). Microfinance promise. Journal of Economic Literature, XXXVII, 1569–1614. Morduch, J. (2000). The microfinance schism. World Development, 28, 617–629. Mosley, P., & Hulme, D. (1998). Microenterprise finance: Is there a conflict between growth and poverty alleviation?. World Development, 26, 783–790. Navajas, S., Schreiner, M., Meyer, R. L., Gonzalez-vega, C., & Rodriguez-meza, J. (2000). Microcredit and the poorest of the poor : Theory and evidence from Bolivia. World Development, 28. Olivares-Polanco, F. (2005). Commercializing microfinance and deepening outreach? Empirical evidence from Latin America. Journal of Microfinance, 7, 47–69. Otero, M., & Rhyne, E. (1994). The new world of microenterprise finance: Building healthy financial institutions for the poor. Kumarian Press. Paxton, J. (2003). A poverty outreach index and its application to microfinance. Economics Bulletin, 9, 1–10. Po¨lzlbauer, G. (2004). Survey and comparison of quality measures for selforganizing maps. In J. Paralicˇ, G. Po¨lzlbauer, & A. Rauber (Eds.), Proceedings of the fifth workshop on data analysis (WDA’04). Sliezsky dom, Vysoke´ Tatry (pp. 67–82). Slovakia: Elfa Academic Press. Quayes, S. (2012). Depth of outreach and financial sustainability of microfinance institutions. Applied Economics, 44, 3421–3433. Rhyne, E., & Otero, M. (2006). Microfinance through the next decade: Visioning the who, what, where, when and how. ACCION International. A Paper Commissioned by the Global Microcredit Summit 2006. Retrieved from: http://livestats.pointloma.edu/sites/default/files/bio_ files/microfinance_through_the_next_decade_0.pdf. Robinson, M. S. (2001). The microfinance revolution: Sustainable finance for the poor. The World Bank.

Schreiner, M. (2002). Aspects of outreach: A framework for discussion of the social benefits of microfinance. Journal of International Development, 14, 591–603. Schwartz, D., Smith, K.A., Churilov, L., Dally, M., & Weber, R. 2003. Design and application of hybrid intelligent systems. Chapter Improving risk grouping rules for prostate cancer patients using self-organising maps (pp. 126–135). Amsterdam, The Netherlands. SEEP Network (2010). Pocket guide to the microfinance financial reporting standards measuring financial performance of microfinance institutions. Technical Report October. Seret, A., Verbraken, T., Versailles, S., & Baesens, B. (2012). A new SOMbased method for profile generation: Theory and an application in direct marketing. European Journal of Operational Research, 220, 199–209. Smith, K. A., & Ng, A. (2003). Web page clustering using a self-organizing map of user navigation patterns. Decision Support Systems, 35, 245–256. Srivastava, R., Bharadwaj-Chand, S., & Sinha, P. (2010). The ugly underbelly of microfinance. The Times of India. Tan, P., Steinbach, M., Kumar, V., et al. (2006). Introduction to data mining. Boston: Pearson Addison Wesley. The Economist (2010). Discredited. Vesanto, J., & Alhoniemi, E. (2000). Clustering of the self-organizing map. IEEE Transactions on Neural Networks, 11, 586–600. Von Pischke, J. D. (1996). Measuring the trade-off between outreach and sustainability of microenterprise lenders. Journal of International Development, 8, 225–239. Woller, G. M. (2002). The promise and peril of microfinance commercialization. Small Enterprise Development, 13, 12–21. Woller, G. M., Dunford, C., & Woodworth, W. (1999). Where to microfinance. International Journal of Economic Development, 1, 29–64. Year of Microcredit 2005 (2005). International Year of Microcredit 2005. Yunus, M. (2011). Sacrificing microcredit for megaprofits. The New York Times