Master Thesis Commercialisation and outreach of Microﬁnance Institutions in Latin America

(1)

Master Thesis

Commercialisation and outreach of Microfinance

Institutions in Latin America

An efficiency frontier analysis

Oliver Henrich

∗

Supervised by Francesco Cecchi

23rd June 2015

Abstract

This thesis uses data on 395 microfinance institutions (MFIs) obtained from the MIXMarket database over a 10 year time period and applies data envelopment analyses (DEA) and stochastic frontier analyses (SFA) to investigate the question of whether there is a trade-off between fin-ancial efficiency and outreach to the poor in Latin America. Different finfin-ancial, social, cost, and profit efficiencies have been estimated and were afterwards related to variables that capture the depth of outreach, such as percentage of female borrowers and the average loan balance per borrower. I find evidence that outreach to the poorest of the poor has a negative impact on financial, cost, and profit efficiency, as well as that financially efficient MFIs are less socially effi-cient in Latin America. Serving poorer clients also seems to be positively related to the number of borrowers reached. The results have important implications for policy makers, microfinance practitioners, and donors.

JEL Code: C23, D24, G21, O54

Keywords: Microfinance, efficiency, outreach, data envelopment analysis, stochastic frontier analysis, Latin America

∗_{Faculty of Economics and Business, University of Groningen, PO BOX 800, 9700 AV Groningen, the}

(2)

1 Introduction

“Sometimes I describe poor people as the bonsai tree. If you take the seed of the tallest tree in the forest and put it in a flower pot, it grows only as big as the pot will let it. There’s nothing wrong with the seed; simply we did not give it enough space to grow. Poor people are bonsai people. There’s nothing wrong with their seed, society never allowed them the space to grow as tall as everybody else.” - Muhammad Yunus, Nobel Peace price laureate and founder of the Grameen Bank (Mosher, 2012)

Starting four decades ago microfinance has been used extensively to reduce world poverty. Microfinance institutions (MFIs) have specialised in providing credit and other finance related services (savings, insurance, education) to low income individuals and people who have no access to commercial banks. These types of borrowers usually live close to the subsistence level, where a negative income shock, such as death or illness, can have exist-ence threatening repercussions. Allowing these people to have savings during good times and to borrow in bad times, can therefore help to reduce this kind of risk and maintain a stable level of consumption. Furthermore, loans give these people the opportunity to be able to use their entrepreneurial skills and set up their own income generating busi-nesses to escape poverty. The MFIs registered in the MIXMarket1 _{database alone, report}

a global outreach to 96 million people for 2013 and the global microfinance market is expected to grow by 15%-20% in 2015 (Etzensperger, 2014). The development of mi-crofinance hence seems to be a true success story and some even count it to be one of the top innovations of the last 30 years.2 But this positive view about microfinance has recently come under scrutiny.

Targeting the poorest of the poor is generally more costly than providing equivalent services to the normal population. For this reason, a lot of MFIs are highly subsidised by the government or charity organisations. Lately however, there has been a shift towards focusing more on financial sustainability and efficiency. This is mainly due to “growing commercialization and competition coupled with withdrawal of subsidies” (Abate et al., 2014, p. 924) in the microfinance sector. These developments result in a new dual objective for MFIs, which have to now not only focus on their outreach to the poor, but also have to put more emphasis on their financial sustainability and efficiency. They need to reduce their costs and cover these with the returns of their gross loan portfolio. This leads to the question of whether MFIs are able to combine these two objectives or

1_{The MIXMarket database only contains the MFIs that voluntarily share their information with}

MIX-Market. Therefore, it only provides a rough insight into the actual global microfinance environment. The true number of active borrowers is therefore much higher. Nevertheless, it is the most reliable and comprehensive database for MFIs available.

2

(4)

whether a trade-off emerges between financial efficiency and outreach.

Theoretically, the two objectives can be combined, if the requirements for financial sus-tainability improve the efficiency of allocating resources and attract commercial funds. These can then be used to increase the number of loans to the poor and thus may sup-port the outreach objective. On the other hand, the two objectives can be conflicting, if financial efficiency crowds out small loans that are demanded by the poorest of the poor because these are costly to service. The negligence of poorer clients due to a focus on financial sustainability and efficiency is often called a “mission drift” of MFIs in the literature.

Especially for policy makers, microfinance practitioners, and socially responsible in-vestors, it is very important to know what the effects of the commercialisation of MFIs are. To investigate the trade-off between financial efficiency and outreach, this thesis is focused on Latin American, since its microfinance market is considered as one of the most developed and diverse markets in the world (Miller, 2003). Moreover, many Latin American MFIs are under pressure to switch from an unregulated, non-profit institution to a regulated shareholder form (Mersland and Strøm, 2009). I use data envelopment analyses (DEA) and stochastic frontier analyses (SFA) to estimate the financial, social, cost and profit efficiency of Latin American MFIs. I then link these efficiency measures to measures of outreach to detect a possible mission drift. For the analysis I use data on 395 different MFIs in 17 countries over the period 2004-2013.

The remainder of the thesis is organised as follows. Section 2 discusses the possible reasons for a mission drift and provides a literature review on the trade-off between financial sustainability, efficiency and outreach. Section 3 outlines the current situation of the Latin American microfinance market. In section 4 the concept of efficiency and a general description of the methodology used in this thesis is explained. Section 5 describes the specific estimation strategy. Section 6 continues with a description of the dataset and is followed by the presentation of the estimation results in section 7. Section 8 summarises the findings and concludes the thesis.

2 Literature Review

2.1 What is mission drift and what are the possible reasons?

(5)

by subsidies. The insitutionalists, on the other hand, determine the goal of microfin-ance as constructing a system that can offer financial services to a large number of poor people on a sustainable basis (Bhatt and Tang, 2001). Their approach emphasises the sustainability and efficiency of institutions and puts less weight on the single individual borrower. With regard to the potential trade-off between financial sustainability and outreach, the institutionalists argue, that financial sustainability and commercialisation of microfinance can increase the number of borrowers and is therefore in line with the outreach objective.

In contrast, the welfarists believe that financial sustainability and outreach are substi-tutes, not complements. They argue that MFIs will finally have to charge higher interest rates to become sustainable and that this will be detrimental to the poorest borrowers, since they will not be able to afford loans anymore. Others fear that MFIs will stop serving the poorest of the poor, who often live in rural areas and instead switch to serve more people in urban areas. This would be advantageous for the MFIs, since they find less risky clientele in urban areas and they can provide their services for lower costs because they do not need to travel to remote areas to meet their clients (Chao-Béroff, 1997). Due to the high fixed costs per loan, MFIs may also decide to disburse bigger loan sizes to improve their financial margins and thus serve less poor clients (Hulme and Mosley, 1996). Basharat et al. (2015) argue that regulations, which are overly focused on financial goals and sustainability, may force MFIs to neglect their social goals to comply with these rules. Consequently, according to the welfarists, these developments will lead to a “mission drift” of MFIs, which will lead to abandonment of their social mission and it will be replaced with the objective of becoming financially sustainable institutions. But why should there be a move towards more commercialisation of MFIs at all?

A study by Rhyne and Otero (2006) mentions four big changes that MFIs will face in the future, or are already facing and which will increase the importance of financial efficiency and self-sustainability. First of all, an increase in competition among MFIs in several countries can be observed. This will give the borrowers more choice in picking the best MFI. The MFIs, on the other hand, have to become more efficient and enhance their product portfolio to remain competitive which will lead to additional services such as savings accounts or insurances. The second driving factor is the entry of commercial banks into the microfinance sector that will further contribute to more competition. The big success of early entrants in this market has changed the perceived risk of it and made it an attractive field to operate in (Isern and Porteaus, 2005).

(6)

without any external distortions, and contributes to the development of efficient and sustainable MFIs (Rhyne and Otero, 2006). A fifth change, mentioned by Hermes et al. (2011), describes the increased interest of private investors in investing in MFIs. So called microfinance investment vehicles (MIVs) allow investors to invest in social responsible projects and take part in the success of MFIs.3 _{They argue, that this may further push}

MFIs to more sustainability and efficiency to attract more investors.

2.2 What is the evidence for a mission drift?

Cull et al. (2009) conduct the first comprehensive study of 346 leading MFIs worldwide, serving nearly 18 million customers with total assets of $25.3 billion (PPP adjusted), over the period 2002-2004, to address several questions related to mission drift. They mainly focus their analysis on the differences between non-governmental organisations (NGO), non-bank financial institutions (NBFI) and commercial microfinance banks. By using the average loan size, which is a common proxy for outreach, they find that for the median bank average loan size is four times that for the median NGO. The percentage of female borrowers, another proxy used for outreach because women tend to belong to the poorest of the poor, mirrors this finding. For the median NGO 85% of the borrowers are female, whereas for the median bank only every second customer is a woman. Although women are thought to be more risk averse than men, they seek smaller loans, which increases the transaction costs for the MFIs (Armendáriz and Morduch, 2010). The focus of banks on larger loans and less women also has an effect on the costs to disburse a loan. While the median bank only pays $0.12 per dollar lent, the median NGO has to pay $0.26.

A rather surprising finding by Cull et al. (2009), and contrary to the fears of the welfarists, is that the median NGO charges their customers a 12% higher interest rate per annum than the median banks. The authors explain this finding with the cost structure of the institutions. NGOs have to charge higher interest rates to offset their higher unit costs. Additionally, NGOs usually do not profit from economies of scale since they serve less people than banks do. According to Dorfleitner et al. (2013), the interest rate charged to borrowers increases by 0.41 percentage points for every 1 percentage point increase in the operating expense ratio. The studies by Rosenberg et al. (2009) and Cotler (2010) also find that smaller loan sizes are negatively correlated with interest rates.

The evidence with regard to the impact of regulatory requirements on outreach is rather mixed. Mersland and Strøm (2009) find that regulated and unregulated institutions do not differ in their financial efficiency. Also Hartarksa and Nadolnyak (2007) discover no relationship between regulation and outreach. Other authors (e.g. Makame and

3_For _a _good _overview _of _the _current _market _for _MIVs _see

(7)

Murinde, 2006; Barry and Tacneng, 2011; Basharat et al., 2015), however, come to different conclusions and argue that there is a negative relationship between regulatory status and outreach.

The study by Cull et al. (2009) confirms the findings of Gonzales (2007) who finds, for a sample of 1,003 MFIs in 83 countries, that larger loan sizes are a significant way to improve financial efficiency by lowering operating expenses. Hermes et al. (2011) who use a SFA of 435 MFIs over a period of 11 years, find that MFIs with lower loan sizes and a higher number of female borrowers are less cost efficient. Other authors find a similar trade-off between financial sustainability and outreach using cross-country panel data on MFIs (see, e.g., Annim, 2012; Louis and Baesens, 2013; Vanroose and D’Espallier, 2013; Kar, 2012). On the other hand, Serrano-Cinca et al. (2011), using a DEA of 89 MFIs, find a positive but low correlation between financial and social efficiency, but none between social efficiency and profitability. Also Bédécarrats et al. (2012), analysing data from 295 MFIs in 51 countries, argue that social and financial performance are compatible “when trade-offs and synergies are combined cleverly following a well planned social performance management strategy” (Bédécarrats et al., 2012, p.23). Recently, Quayes (2015), using 764 MFIs from 87 countries, even finds that greater depth of outreach has a positive impact on the financial performance of an MFI.

Focusing more on a regional level, Omri and Chkoundali (2011) employ a static panel data analysis of 16 Mediterranean MFIs over the period 2001-2008. They find not only a positive link between average loan size and profitability, but also that the profitability is strongly linked to the number of outstanding loans per women. Makame and Murinde (2006) investigate the question about mission drift by looking at 33 MFIs in five East African countries over the period 2000-2005. They use different measures for depth of outreach and find evidence for a trade-off between outreach, sustainability, and efficiency. Also Kablan (2012), looking at 104 MFIs in the West African Economic and Monetary Union and using a DEA, finds that MFIs that stress outreach are less financially efficient. Regarding Latin America, Olivares-Polanco (2005) uses data from 28 MFIs to conduct a multiple regression analysis. He finds that there is no impact of the type of institution on the size of loans. Competition and financial sustainability, however, appear to lead to larger loan sizes and less depth of outreach. Servin et al. (2012) apply a SFA for 315 Latin American MFIs and find that “NGOs and cooperatives have a lower technology level than banks and NBFIs, because of their stronger focus on social goals and their far more severe funding constraints (Servin et al., 2012, p. 2144). In the same vein, Gutiérrez-Nieto et al. (2007) find, by using a DEA of 30 MFIs, that especially NGOs try to give out a large number of loans and aim at supporting as many individuals as possible in Latin America.

(8)

et al. (2014) use a SFA of 107 MFIs in Ethiopia and discover that outreach and cost-efficiency are negatively related to each other. Also McIntosh and Wydick (2005), who look at Uganda, find a negative relationship between increased competition and social performance. On the other hand, Crawford et al. (2014), using a DEA of 13 Cambodian MFIs over a period of 6 years, argue that commercially focused MFIs are not less efficient at reaching the poor than non-profit ones. Similarly, Lebovics et al. (2014), looking at a sample of Vietnamese MFIs, find no relation between financial and social efficiency. To summarise the above review, there are mixed results regarding a possible mission drift of MFIs. This thesis further contributes to the discussion about a trade-off between outreach and financial efficiency. By using DEA and SFA, financial, social, cost and profit efficiency scores are determined for Latin American MFIs and are later related to outreach variables. Latin America is a particular interesting market to look at, since it is more developed and diverse than markets in other regions (Miller, 2003). Additionally, the changes, mentioned by Rhyne and Otero (2006), seem to have already reached Latin America. Christen (2001) finds high levels of profitability, increasing levels of competition, and predominance of regulated institutions relative to other regions. The focus on a single region also helps to gain a rather homogenous sample of MFIs, where regional disparities are less likely to bias the results. To my knowledge this is the first study that uses such a comprehensive efficiency analysis to investigate a possible trade-off between financial efficiency and outreach in Latin America.

3 The Latin American microfinance market

The appearance of the first MFI in Latin America can be traced back to the early 1970s, when Acciòn started its programme in 1973 in Brazil (Rhyne, 2001). The big success of early pioneers combined with the prospect of poverty reduction, contributed to the prolif-eration of MFIs in Latin America in the following decades. According to the MIXMarket (2015) database, 355 institutions have been operating in 2013 – a significant increase from the 222 in 2004.4

Figure 1 depicts the development of number of loans outstanding and the number of active borrowers from 2004 to 2013. During this period the number of outstanding loans almost quintupled to 21 million and the number of active borrowers increased to 18.5 million.

Decomposed into country observations, Peru ranks first in terms of both number of loans outstanding and number of active borrowers in 2013 (see Table 1).There is, however, an

4_{The period from 2004 until 2013 is used here, since it contains observations for all countries in the}

(9)

2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 5 10 15 20 25 Year Amoun t in m illion

Number of outstanding loans Number of active borrowers

Figure 1: Number of outstanding loans and number of active borrowers over time

Source: MIXMarket (2015)

uneven development in microfinance across regions. Marulanda and Otero (2005) find that the market penetration rate is highest in Bolivia with 55.7%, whereas the lowest rate can be found in Venezuela with 0.8%. Microfinance still lags behind in larger countries like Argentina and Brazil; and seems to work better in smaller and medium sized countries like Peru, Colombia or Ecuador. Ramirez (2004) explains this by the increased presence of state development banks and the limited success of NGOs in the larger countries. On a global scale, Latin America is now the second biggest market for microfinance behind South-East Asia regarding active borrowers (65.2 million borrowers). However, serving only a fraction of the borrowers in Asia, the gross loan portfolio of Latin American MFIs is almost 4 times higher ($72.5 billion against $18.9 billion). Miller (2003) argues that the demand for bigger loans in Latin America is due to the higher gross national product (GNP) per capita. Also the lower population density in Latin America is thought to be responsible for a smaller outreach (Olsen, 2010). Nevertheless, MFIs in Latin America are growing at a much faster rate than MFIs in other regions of the world - between 20 and 40 percent in the period of 2000-2005 (Berger et al., 2006) - which indicates that there is still unmet demand in this market.

(10)

Table 1: Number of MFIs, outstanding loans and active borrowers per country in 2013

Country Number of Number of Number of

MFIs outstanding loans active borrowers

Peru 57 6,280,128 5,002,018 Mexico 52 3,365,729 3,389,722 Colombia 31 3,121,252 2,756,421 Brazil 9 2,130,945 1,898,000 Ecuador 51 1,648,141 1,454,543 Bolivia 20 1,294,972 1,223,656 Paraguay 7 1,047,899 808,566 Dominican Republic 14 496,795 445,530 Chile 3 450,108 272,275 Nicaragua 23 369,227 333,555 Guatemala 15 211,479 356,825 Honduras 27 195,427 219,489 El Salvador 13 143,155 139,787 Argentina 9 138,791 24,536 Haiti 5 130,430 129,666 Panama 5 42,299 41,056 Costa Rica 14 19,658 17,689 Source: MIXMarket (2015)

formal banking sector started taking place in the early 1990s. The rapid expansion of credit services lead to a lack of funding resources - which were still mainly donor funds - and made this step necessary to meet the demand of the market. The most prominent example of such an “upscaling” movement is the development of BancoSol in Bolivia. Starting as PRODEM, a NGO, in 1987, it disbursed over $28 million loans and funded 45,000 micro-enterprises in only five years (Anon., 2006). Having the status of a NGO, however, restricted PRODEM to further meet the increasing demand, since it was legally restricted to raise significant amounts of funds from commercial markets and it was not allowed to offer other financial services (except credits) to its clients. To circumvent these restrictions, BancoSol, the world’s first bank fully dedicated to microfinance, was founded in 1992. Today, BancoSol is successfully serving around 250,000 clients with a gross loan portfolio of $1 billion (MIXMarket, 2015).

(11)

the microfinance market and set up their own special divisions or affiliated companies for microfinance.5 _{In addition, Latin America drew the attention of more and more individual}

and commercial investors which further contributed to the growth in this sector. Pension funds, investment banks, capital development funds, and social responsible investors are becoming increasingly important in the funding of MFIs (Miller-Sanabria and Narita, 2008).

Secondly, to remain attractive for investors, MFIs have progressively shifted their focus towards financial sustainability and profitability. Christen (2001) finds that MFIs in Latin America are not only more profitable than the ones in other regions of the world, but that they also show returns that are sometimes even higher than the returns of commercial banks in this region. Navajas and Tejerina (2006) point out that the return on assets of Latin American MFIs was equal to more than two times the world’s average in 2005. Thirdly, commercialisation and the rising participation of different players in the microfinance market have lead to increased competition and market saturation in some countries. Especially in Bolivia, but also increasingly in Nicaragua, El Salvador, Honduras and Paraguay (Anon., 2004) an increased indebtedness of clients can be observed who take credits from several different MFIs to repay their loans. This is not only detrimental to the clients, but also affects the portfolio performance of the institutions negatively. Exposed to the new competitive environment, MFIs are reacting with better client responsiveness and an increased improvement of efficiency. They ameliorated their service delivery, moved more towards individual lending and started to offer a bigger range of products (e.g. saving accounts and micro-insurance).

Lastly, Christen (2001) points out a trend towards more regulation. MFIs that want to adopt a commercial approach, need to be regulated to be able to receive funds from the formal banking system. Hence, a more regulated microfinance market is the consequence of commercialisation. Table 2 gives an insight into the different properties of regulated and unregulated institution in Latin America in 2013.

The market is almost evenly split between regulated and unregulated MFIs in terms of number of institutions. However, when it comes to the number of outstanding loans and number of active borrowers, the regulated institutions possess a market share of 72% and 68% respectively. The median average loan balance per borrower of regulated institutions is almost 4 times higher than the one of unregulated institutions. Moreover, regulated institutions have fewer female borrowers, which could be an indication for a mission drift of commercialised MFIs in Latin America.

To summarise the above discussion, the rapid growth of the Latin American microfinance sector in the last decades has lead to an unprecedented move towards commercialisation.

5_{See, for example, Chowdri and Silva (2004) for a case study about commercial banks entering the}

(12)

Table 2: Characteristics of regulated and unregulated MFIs in 2013 Regulated Unregulated

Number of MFIs 156 198

Number of outstanding loans 15,100,000 5,981,277

Number of active borrowers 12,700,000 5,853,993

Average loan balance (median) 2,480 633

Percentage of female borrowers (median) 47.125 67.67

To stay competitive and to attract funding from outside investors, MFIs are increasingly focusing on efficiency and financial sustainability. Whether this shift in focus has negative ramifications and leads to a negligence of the social mission of MFIs, will be analysed in the remainder of this thesis.

4 Methodology

4.1 The concept of efficiency

Before I describe the econometric approaches to measure efficiency, it is necessary to introduce the idea of efficiency in more detail. A first definition of what we call technical efficiency today, can be found in the work of Koopmans (1951). According to him a producer is technically efficient “whenever an increase in one of its coordinates (the net output of one good) can be achieved only at the cost of a decrease in some other coordinate (the net output of another good)” (Koopmans, 1951, p.60). Stated otherwise, a firm is technically efficient if it maximises its outputs given a certain vector of inputs. Whereas Koopmans’ definition is focusing on output maximisation, the so called output-oriented approach, Shepard (1953) introduces the input orientated view. He considers a firm as efficient, if it minimises its inputs to produce a given amount of outputs.

(13)

outputs is available and one assumes firms to be profit maximisers or cost minimisers, then this information should be incorporated into the measure of efficiency. It could therefore be possible, that firms, which are operating on the technical efficiency frontier, are still considered inefficient since they fail to maximise profits or minimises costs with this input-output mix. Hence, allocative efficiency depicts the ability to combine inputs and outputs in such a way as to maximise profits or minimise costs. The product of technical and allocative efficiency determines the economic efficiency, which is also often called profit or cost efficiency depending on the point of view.

Figure 2 gives a graphical representation of the different efficiency measures for an output-oriented profit maximiser and an input-output-oriented cost minimiser.6

(a) Profit efficiency (Output-oriented) (b) Cost efficiency (Input-oriented)

Figure 2: Profit and cost function

Source: Meesters (2009)

In Figure 2a it is assumed that one input (X) is necessary to produce two outputs (Y1

and Y2). The x-axis denotes the output of Y1 and the y-axis the output of Y2. The

isoquant curve which is given by the points S’S determines the production frontier given a fixed amount of X. Output combinations that lie on this frontier, as point Q and Q’, identify fully technically efficient producers. For technical inefficient producers, however, a point on the production frontier may not be feasible. Take, for example, a producer that produces the output combination Y₁∗and Y₂∗ which is given by point P in the graphic. This point is not on the production frontier and a technical efficient producer would be able to produce the output combination denoted by point Q with the same amount of input X. I am now able to take the ratio of OP to OQ as a measure of technical efficiency. In this sense a value of 1 denotes a technically efficient producer that operates on the production frontier, whereas lower values indicate more technical inefficiency.

So far information on prices and its influence on the determination of efficiency was disregarded. Nonetheless, prices are, as stated above, crucial to determine the profit

(14)

efficiency of a producer. Therefore, it is assumed that the prices for output Y1 and Y2 are

given by p1 and p2 respectively. The proportion of these prices is used to draw the line

A’A. Thus, every point on this line yields the same profit and the further the line is shifted to the right, the higher is the amount of profit the firm can generate. Consequently, the optimal production point is reached where the price line is tangent to the production frontier (point Q’). By comparing point Q and point Q’ it can be seen that both points are on the production frontier and thus technical efficient. The former, however, fails to maximise profits with this choice of output combination. Producing the same proportion of Y1 and Y2 point R is located on the A’A line and can be used as a reference point to

determine the allocative mismatch. The ratio of OQ to OR is a measure of allocative efficiency. It has the value of 1 if the producer is allocative efficient and produces at the tangency point. Lower values indicate an allocative mismatch between outputs and profit maximisation. Technical efficiency and allocative efficiency can be combined to determine the profit efficiency. For a producer that produces at point P, the profit efficiency is given by the ratio of OP to OR; or simply the product of technical and allocative efficiency. Figure 2b represents an input-oriented cost minimiser. In this case two inputs X1 and

X2 are necessary to produce one output Y. Similar to above, the optimal production

function is given by the points S’S. Since the producer is facing a minimisation problem now, the function is mirror-inverted to the case above. Points that are farther away from the frontier, as point P, are considered inefficient, because the same outputs could have been produced with the inputs given by point Q. The measure of technical efficiency can now be described by the ratio of OQ to OP.

The proportion of the two input prices is used to draw the line A’A. It reflects the cost line and it should be shifted as low as possible to minimise costs. In the optimum, the producer would choose the input combination given by point Q’, because the cost line is tangent to the production function at this point. Producing at point Q would be technically efficient, it fails, however, to minimise costs and is thus not allocative efficient. Allocative efficiency can be measured by OR over OQ. The product of allocative and technical efficiency yields the cost efficiency which is given by the ratio of OR to OP.

4.2 How to estimate efficiency

(15)

can be applied afterwards to give a better insight into the determinants of efficiency and inefficiency respectively.

To determine the efficiency frontier, accounting measures of costs, inputs, outputs and profits are required. With regard to the technical approach to measure efficiency, Berger and Humphrey (1997) state that “there is really no consensus on the preferred method for determining the best-practice frontier against which relative efficiencies are measured” (1997, p.177). Concerning the measurement of efficiency of financial institutions they say, that there exist at least five different approaches which are used in the literature. These approaches differ in the functional form of the best practice function, whether or not random errors are considered, and in the distributional assumptions of the inefficiency term. The following two sections elaborate on the general characteristics of two of these approaches in more detail, since these are most commonly used in frontier analyses of MFIs, as can be seen in the literature review. Section 5 then describes the explicit specifications used for this thesis.

4.2.1 Data envelopment analysis

Data envelopment analysis (DEA) is a non-parametric mathematical programming ap-proach to estimate a best-practice frontier (Charnes et al., 1978; Banker et al., 1984). The frontier is constructed as piecewise linear combinations connecting a set of best-practice observations (Berger and Humphrey, 1997) and the inefficiency of firms is measured rel-ative to this frontier. Due to the fact that it uses linear combinations, it is not necessary to specify a functional form of the production function. This can be seen as an advantage over parametric approaches, since efficiency scores are not dependent on the specification of the production function and a possible bias can therefore be avoided. On top of that, a DEA model does not require assumptions about the distribution of efficiency scores. It is less data demanding and it can handle any kind of input and output without the need to standardise the data (Lebovics et al., 2014).

Before a DEA can be used, two more assumptions about the specification need to be made. First, an assumption about the nature of return to scales is necessary. It can be distinguished between a constant return to scale (CRS) and variable return to scale (VRS) model. Under the CRS model, there is no relationship between the scale of operations and the efficiency level. A VRS model allows to calculate pure technical efficiency scores, which are not influenced by scale efficiencies. Since the correlation of the efficiency scores gained by the CRS model and the VRS model is relatively high (ranges from 0.83 to 0.96), the CRS model is used in this thesis. This model is also used in several other studies (e.g., Lebovics et al., 2014; Gutiérrez-Nieto et al., 2009).

(16)

model is used. According to Kumbhakar and Lozano Vivas (2005), output levels are mainly determined by demand factors in the banking sector and banks are more focused on cost-minimisation. The same can be argued for MFIs and therefore the input-oriented model is chosen here.

The mathematical expression for the input-oriented CRS7 _{model is as follows:}

min θ,λ θ, subject to: −yi+ Y λ ≥0, θxi− Xλ ≥0, λ ≥0. (4.1)

Assume that there are M outputs and K inputs available for N firms. xi is a Kx1 vector

of inputs and yi is a Mx1 vector of outputs for each firm i. X represents a KxN input

matrix, Y the MxN output matrix for all firms and λ represents a Nx1 vector of constants. By solving the minimisation problem above, one contains a scalar θ which represents the technical efficiency score for each firm i. It satisfies the condition θ ≤ 1, with a value of 1 indicating a fully efficient firm, which operates on the production frontier.

A further assumption is necessary if panel data is available. Tulkens and Eeckaut (1993) point out three methods that can be used if time-series are available. The first option is to pool the data and estimate a single frontier over time. This would lead to a best-practice frontier which does not change over time. The second option estimates a frontier for every time period, which allows for technological progress and regress. I am using an intermediate of these two options for this thesis. A sequential frontier is estimated by continuously adding data from successive time periods. This option allows to rule out technological regress and assumes that a certain technology remains available forever. A big drawback of DEA is the assumption of the non-existence of random errors. This means that the deviations of firms from the optimal frontier due to, for example luck or measurement errors, are considered as inefficiencies. Small changes in the random error of firms on the frontier may thus have a large cumulative effect on the aggregate ineffi-ciencies, since all the other firms are measured relatively against these firms (Matoušek and Taci, 2004). Furthermore, the lack of a distributional assumption for the efficiency scores makes it difficult to draw inferences about possible determinants of efficiency scores. Moreover, the DEA model assumes a rather homogenous sample.

Given the kind of heterogenous sample in this study and taking into account the just mentioned caveats, the results of the DEA should be interpreted with caution. An

ap-7_{For more information about DEA and the specification of output-oriented and VRS models, see Coelli}

(17)

proach that does not have these kinds of drawbacks is the stochastic frontier analysis (SFA), a parametric approach. It allows to estimate the coefficients of the production function and the inefficiency term simultaneously and is specifically suited to work with unbalanced panel data (Battese and Coelli, 1995). Therefore, it is used as another eco-nometric approach to investigate the question of a mission drift in Latin America in this thesis.

4.2.2 Stochastic frontier analysis

At first, the specification of a SFA to determine technical efficiency scores is introduced. The parametric approach differs from the non-parametric approach by making a func-tional assumption about the production function and by introducing a random error. Deviations from the production function can therefore be split into an inefficiency term and error term that accounts for random deviations. Aigner, Lovell, and Schmidt (1977) and Meeusen and van den Broeck (1977) independently proposed the most commonly used parametric approach today; the stochastic frontier analysis.

A general specification of the production frontier model without a random error is given by

yi = f (xi;β) ∗ T Ei, (4.2)

where yi denotes the observed scalar of output by producer i, xi is a vector of inputs

used by the producer and β is a vector of technology parameters which can be estimated by regressing y on the x’s. f () represents a production function which is used by the industry. The translog (Christensen et al., 1973) is an example for a widely used and accepted functional form for a production function, but also other specifications are often applied. Among these are the Cobb-Douglas, the Fuss normalised quadratic (Morrison and Berndt, 1981), the generalised translog (Caves et al., 1980), the composite cost (Pulley and Braunstein, 1992) and the Fourier flexible form (Gallant, 1982).

The choice of the functional form depends on the researcher’s opinion about which spe-cification approximates the underlying production function best. I am using a translog production function for this analysis, because it allows more flexibility than a Cobb-Douglas function and it is widely used in efficiency analyses of financial institutions. Besides, the translog function is only an extension of the Cobb-Douglas function and includes additional interaction terms of the inputs and their squared values. Hence, via a simple likelihood-ratio (LR) test for the joint significance of these additional terms one can distinguish between the right functional form.8

(18)

T Ei denotes the technical efficiency of every producer i. It can be expressed as

T Ei =

yi

f (xi;β)

, (4.3)

which is the ratio of observed output to maximal feasible output. It is therefore bounded between 0 and 1. A producer with a technical efficiency of 1 would produce on the produc-tion frontier, whereas lower values of T E indicate a higher level of technical inefficiency. So far, equation 4.2 and 4.3 disregard an error component and are completely determin-istic. It is now assumed that there is a shock variable which differs for every producer but which is completely random. These shocks can occur, for example, due to measurement errors or just luck. The stochastic component in the model is denoted by exp(vi). Due to

the introduction of a random error term, the technical efficiency also becomes stochastic. It is therefore assumed, that technical efficiency also follows a specific distribution which is expressed as T Ei = exp(−ui) in this model and where ui ≥ 0 since it is required that

T Ei ≤ 1. The new stochastic production function can thus be written as

yi = f (xi; β) ∗ exp(vi) ∗ exp(−ui). (4.4)

The random error term vi is usually assumed to be normally distributed, whereas there

is no consensus about the distribution of the efficiency term ui. Aigner, Lovell, and

Schmidt (1977) themselves propose a half normal distribution while Meeusen and van den Broeck (1977) suggest a truncated normal or exponential distribution and Berger and Humphrey (1997) favour a gamma distribution. However, the distributional assumption of the efficiency term does not appear to have a big impact on the estimated efficiencies (Cummins and Zi, 1998). Additionally, it is assumed that vi and ui are independent of

each other, so a proper distinction between random effects and efficiency can be made. The coefficients in equation 4.4 can then be estimated by using a maximum likelihood estimator.

(19)

The general specification of the production function to be estimated, can be written as

ln(yit) =f (xit;β) + vit− uit (4.5)

and has the following properties

vit∼iidN (0, σ2v) (4.6) uit∼E(σuit) (4.7) σuit = v u u texp(δ0+ n X m=1 δmzitm). (4.8)

Similar to the notion above but now taking into account that panel data is available, yit denotes the output of producer i at time t. xit represents a vector of inputs used by

producer i at time t and β denotes the vector of parameters to be estimated.

uit captures the technical inefficiencies and is exponentially distributed. σuit can be

interpreted as the expected inefficiency of MFI i at time t. z determines a vector of n variables that are used to determine the inefficiency of MFI i at time t. δ are the coefficients to be estimated and the main focus of this thesis.

As stated above, other distributions like a truncated normal distribution are often as-sumed for the inefficiency term. However, in this case these assumptions would lead to a non-convergence of the likelihood maximisation routine. Meesters (2014) shows that the cause for non-convergence may occur if the true distribution of inefficiencies is expo-nential. He shows that in this case some of the parameters in the model are diverging to minus or plus infinity which means their ratio is equal to σ of equation 4.8. This is the case for the data-set used for this thesis and hence the exponential distribution is used here.

By focusing only on technical efficiency, the problem arises that only one variable can be used as an output. Nevertheless, it can be argued that MFIs have more output objections, since they try to achieve financial and social goals. A solution to this problem is to estimate a cost efficiency (CE) or profit efficiency (PE) model which can be written as:

ln(T Cit) =c(yit,wit;β) + vit+ uit (4.9)

ln(πit) =p(yit,wit;β) + vit− uit. (4.10)

T Cit and πit represent the total costs and the net profit of a producer i at time t and

(20)

efficiency model, the input variables differ. yit is a vector of outputs produced by the

producer i at time t and, since prices are necessary now, wit denotes a vector of input

prices. Furthermore, in the CE model the inefficiency term uit is added because an

inef-ficient producer faces higher costs than an efinef-ficient one. Unlike in the TE and PE model, a producer hence faces a minimisation problem and the CE model is input-oriented. All the other properties and definitions are identical with the TE model above.

5 Empirical model specifications

5.1 Data envelopment specifications

A two-stage DEA model is used to investigate the trade-off between financial efficiency and outreach. In the first stage the financial and social efficiency scores are calculated separately. A second stage regression is then applied to explore the possible determinants of the efficiency scores. As already mentioned in the previous section, a functional form of the production function is not necessary and any kind of input and output combination can be used. This allows to include several output variables that all capture the financial and social objectives of MFIs respectively.

Before the relevant input and output variables are discussed, a decision has to be made about whether the DEA is applied using the production approach (Benston, 1965) or the intermediation approach (Sealy and Lindley, 1977). This topic is highly debated in the literature and both approaches have their strengths and weaknesses. Under the production approach, MFIs are considered as producing units similar to manufacturing companies. They use capital and labour as inputs to produce certain outputs as, for example, loans, deposits or number of borrowers. On the other hand, the intermediation approach regards MFIs as financial intermediaries that use deposits and borrowings to produce certain outputs. Hence, the role of deposits differs between the approaches; once used as an output in the former and as an input in the latter approach. The production approach is applied here, since MFIs emphasise the granting of loans rather than collecting deposits (Gutiérrez-Nieto et al., 2009) and because this approach seems to be better suited for bank branches with low autonomy in loan policy (Bikker and Bos, 2008).

To calculate the technical efficiency scores, I am using three inputs which are widely used in DEA and SFA when the production approach is applied. The first input is assets, which is the total of all net assets expressed in constant 2004 US dollars.9 _{This input}

has been used by, for example, Berger and Humphrey (1997), Seiford and Zhu (1999),

9_{A table with the official definitions used by MIXMarket for all the variables in this thesis can be found}

(21)

Luo (2003) and Servin et al. (2012). The second input represents the labour used in the production and consists of the total number of staff members per MFI. It has been proposed by Sherman and Gold (1985), Athanassopoulos (1997), Berger and Humphrey (1997), Seiford and Zhu (1999), Luo (2003) and Servin et al. (2012) among others. Lastly, operating expenses,10 _{defined as, “expenses related to operations, including all personnel}

expense, depreciation and amortization, and administrative expense” (MIXMarket, 2015) and also expressed in monetary units, is commonly used in efficiency analyses of financial institutions (e.g. Athanassopoulos, 1997; Berger and Humphrey, 1997; Pastor, 1999; Gutiérrez-Nieto et al., 2009; Servin et al., 2012).

To capture the financial objectives of MFIs, I am using the gross loan portfolio and financial revenue11 _{as output variables. The gross loan portfolio includes outstanding}

principals due for all outstanding client loans. It is a widely used measure of financial efficiency in the literature (e.g. Sherman and Gold, 1985; Athanassopoulos, 1997; Berger and Humphrey, 1997; Wheelock and Wilson, 1999; Gutiérrez-Nieto et al., 2009; Lebovics et al., 2014). Financial revenue depicts the ability of a MFI to generate revenue from the gross loan portfolio and other financial assets. Among others, Pastor (1999), Seiford and Zhu (1999), Gutiérrez-Nieto et al. (2009) and Lebovics et al. (2014) have used this variable in their analyses.

Regarding the social objectives of a MFI, the number of active women borrowers12 and the reciprocal of the average loan balance per borrower divided by gross national income (ALB ) are included as output variables. These measures capture the sole objective of lending to the poorest of the poor. Lending to female borrowers has widely been used as a measure of depth of outreach (e.g. Bassem, 2008; Gutiérrez-Nieto et al., 2009; Kablan, 2012) because lending to women is generally associated with lending to the poor. Also ALB has been used in several studies to capture the outreach to the poor (e.g., Hermes et al., 2011; Olivares-Polanco, 2005). Higher values of ALB means that loans with higher principal values are disbursed, which indicates less depth of outreach. The average loan balance per borrower is divided by the gross national income, to account for the fact that some countries are richer than others and therefore clients demand bigger loans per se in these countries. The reciprocal is taken, so higher values of both measures, indicate a bigger outreach to the poorest.

It has to be noted that the number of active women borrowers and ALB are only a rough approximation of outreach to the poorest. The measure only covers one aspect of outreach, namely depth of outreach. Schreiner (2002) argues that outreach may have

10_{Operating expenses have been computed by multiplying the operating expenses to assets ratio by}

assets.

11_{Financial revenue has been computed by multiplying the financial revenue to assets ratio by assets.} 12_{The number of female borrowers has been computed by multiplying the percentage of female borrowers}

(22)

more dimensions such as the value a microfinance loan has for the client, the cost of the loan to the client, breadth of outreach, length of outreach and the scope of outreach. Also Paxton (2003) says that the size of the loan may be related to the type of loan and the lending methodology of the MFI. However, given the fact that other dimensions of outreach are difficult to measures and that these measures are commonly used in the literature, these variables are most suitable for this thesis.

After the social and financial efficiency scores have been computed, two Tobit regressions are carried out to examine the determinants of financial and social efficiency scores in the second stage. Since the efficiency scores are bounded within the range from 0 to 1, an ordinary least square estimation would lead to biased results for the efficiency parameter because it assumes a normal and homoscedastic distribution of the disturbance and the dependent variable (Maddala, 1986). The Tobit model used in this study has the following properties: y₀∗ =β0x0+ ε0 y0 =y0∗ if 0 < y ∗ 0 < 1 otherwise, y0 =0 if y∗0 ≤ 0 and y0 = 1 if y0∗ ≥ 1, ε0 ≈N (0, σ2). (5.1)

x0 and β are the vectors of explanatory variables and its coefficients respectively; y0

and y₀∗ are the vectors of the observed DEA efficiency score and the vector of the latent variable. The value of the coefficients are found by maximising a likelihood function afterwards. The exact specification to be estimated becomes:

EF Fit=β0+ β1OF Fit+ β2AGEi+ β3EF F 2it+ β4OP EXit+ β5RISKit + β6P ROD + β7REGi+ 15 X j=8 Yt+ 31 X j=16 Ci+ εit. (5.2)

(23)

variable is expected to enter the respective equation with a negative sign. To capture the influence of cost efficiency on the inefficiency measures, the ratio of operating expenses to total assets is included as a variable (OPEX ). Becoming more cost efficient is a mean for MFIs to become more attractive for investors and to become more profitable. However, this might be detrimental to the poorest of the poor, since disbursing smaller loan sizes usually entails high transaction costs. Following the work of Kablan (2012) the value of all loans outstanding that have one or more instalments of principal past due more than 30 days divided by the gross loan portfolio (RISK ) is added. It is a proxy for the different risk-taking strategies of the MFIs. The higher the value of RISK, the more the MFIs mismanage their client base and the riskier they operate. PROD captures the influence of productivity on the efficiency scores and is defined as the number of active borrowers divided by amount of staff members. To account for regulatory status (REG ), year (Y ) and country effects (C ) several dummy variables are added further. REG takes the value of one if MFIs are regulated and zero otherwise.

5.2 Stochastic frontier specifications

5.2.1 Technical Efficiency

At first, I am looking at the technical efficiency of producing financial and social outputs. As in the DEA, the production approach is used for the SFA. The financial output is given by the gross loan portfolio and the social output is the number of active borrowers. Capital and labour are used as inputs. Since a production function is estimated here, financial revenue, for example, can not be used as an output because it describes the result of the produced units (loans) and would therefore not be in line with a production function but more with a profit function. The same argument holds for female borrowers and average loan size as outputs. MFIs produce a certain amount of borrowers and this is consequently taken as an output. Unlike in the DEA, this measure captures the width of outreach only. The production function has a translog specification and can be written as:

ln(yit) =β0+ β1ln(ASSit) + β2ln(P ERSit) + β3ln(OP EXit) + β4ln(ASSit)2

+ β5ln(P ERSit)2+ β6ln(OP EXit)2+ β7ln(ASSit) ln(P ERSit)

+ β8ln(ASSit) ln(OP EXit) + β9ln(OP EXit) ln(P ERSit)

+ β10CRISIStln(ASSit) + β11CRISIS ln(P ERSit) (5.3)

+ β12CRISIS ln(OP EXit) + β13CRISIS + β14T + β15T2

(24)

y determines the gross loan portfolio or the number of active borrowers for firm i at time t and ASS, PERS and OPEX are the total amount of assets, personnel and operating expenses for these firms. CRISIS is a dummy variable for the 2007-2008 financial crisis and T denotes a time trend to capture technological progress. To account for the different risk taking strategies of MFIs, the equity to total assets ratio (EQUITY) and loan loss reserves divided by gross loans outstanding (LLR) are added. The inclusion of these variables in a SFA was suggested, among others, by Fries and Taci (2005), Lensink et al. (2008), and Hermes et al. (2011). Servin et al. (2012) find that MFIs with different ownership types use different technologies and have different efficiencies. Therefore, three dummy variables (TYPE) for the four ownership types are added. Although the above model can handle panel data, it does not exploit the properties of panel data (Meesters, 2009). Greene (2004) therefore proposes to add country dummies or even MFI dummies. Due to the large amount of cross-sections compared to the time dimension, it is only possible to add country dummies (C) in this case here.

As mentioned before, the main aim of this thesis is not to interpret the determinants of the production function. It is rather to find out what the drivers of a deviation from this production function are and to investigate the trade-off between efficiency and outreach to the poorest. To analyse this relationship, an empirical model is specified where the expected inefficiency σ_uit2 is used as a dependent variable and measures of outreach and other control variables are regressed on it. The inefficiency equation is estimated simultaneously with the production function and is specified as:

σ_uit2 = exp(δ0+ δ1OF Fit+ δ2AGEi+ δ3RECALBit (5.4)

+ δ4F EM ALEit+ δ5REGi).

(25)

5.2.2 Cost and profit efficiency

To be able to both include financial and social objectives in a SFA, I am estimating the cost and profit functions with gross loan portfolio and number of active borrowers as outputs. Since an estimation of cost and profit efficiency requires data on input prices, I am using the intermediation approach in this case because the sample only contains data for the price of financial capital and not physical capital. Hence, the MFIs are now considered to use borrowings, deposits and labour to produce outputs. The translog specification is expressed as:

ln(yit) =β0+ β1ln(GLPit) + β2ln(BORit) + β3ln(Rit)

+ β4ln(SALARYit) + β5ln(GLPit)2+ β6ln(BORit)2

+ β7ln(Rit)2+ β8ln(SALARYit)2+ β9ln(GLPit) ln(BORit)

+ β10ln(GLPit) ln(Rit) + β11ln(GLPit) ln(SALARYit)

+ β12ln(SALARYit) ln(BORit) + β13ln(SALARYit) ln(Rit) (5.5)

+ β14ln(Rit) ln(BORit) + β15ln(EQU IT Yit) + β16ln(LLRit)

+ β17CRISIS + β18CRISIS ln(GLPit) + β19CRISIS ln(BORit)

+ β20CRISIS ln(Rit) + β21CRISIS ln(SALARYit) + β22T

+ β23T2+ 26 X j=24 βjT Y P Ei+ 42 X k=27 βkCi+ εit,

with yit = T Cit, measuring the total costs13 of firm i at time t, and εit= νit+ uit for the

cost function. Equation 5.5 can simply be transformed into a profit function by using the net profit14 _{of firm i at time t, y}

it= πit and εit = νit− uit (see, e.g. Jiang and Yao, 2010;

Aiello and Bonanno, 2013).

GLP and BOR denote the output variables gross loan portfolio and number of active bor-rowers respectively. SALARY is measured as the total operating expenses per employee of an MFI and represents the price of labour. R is the MFI’s total financial expenses15_per

dollar of liability16 and denotes the price of financial capital. All the other variables have the same definition as in section 5.2.1 and equation 5.4 is again used to investigate the determinants of cost and profit inefficiency. If there is a trade-off between outreach and efficiency, I expect the signs of both ALB and FEMALE to be positive in the inefficiency equations, meaning that MFIs which reach out to poorer people are less profitable and

13_{The total costs have been computed by multiplying the total expenses to total assets ratio by total}

assets.

14_{The net profits have been computed by multiplying the profit margin by financial revenues.}

15_{The financial expenses have been computed by multiplying the financial expenses to total assets ratio}

by total assets.

16_{The liabilities have been computed by multiplying the debt to equity ratio by total equity. Equity has}

(26)

less cost efficient. Table 3 summarises all the specifications used in this thesis.

Table 3: Empirical specifications

DEA SFA

Financial Social Financial Social Cost/Profit

Efficiency Efficiency Efficiency Efficiency efficiency

Assets Assets Price of

Inputs Operating expenses Operating expenses financial capital

Personnel Personnel Salary

Gross loan Female Gross loan Active Gross loan

portfolio borrowers portfolio borrowers portfolio Outputs

Financial Reciprocal Active

Revenue ALB borrowers

ALB denotes the average loan balance per borrower divided by gross national income.

6 Data description

Data has been drawn from the Microfinance Information Exchange (MIX) database, which is a well renowned publicly available database where MFIs and supporting organ-isations share data. It is MIX’s goal to increase transparency in the microfinance finance industry through data collection and analysis. MIX uses data from audits, internal finan-cial statements, management reports and complements this data with questions directly to the MFI. Furthermore, 135 quality checks are used to ensure the accuracy of the data.17

It should be noted, however, that the MFIs reporting to MIX are not representative of the full population of microfinance institutions. The data over-represents institutions that both have a commitment to financial sustainability and that are willing to comply with the MIX’s relatively rigorous reporting standards. The reporting MFIs are therefore more biased towards a good financial performance (Bauchet and Morduch, 2010). The sample consists of 2,065 observations over a period of 10 years from 2004 until 2013, after adjustments for missing data and outliers have been made. Table 4 describes the number of observations per year for which data is available. For the year 2004 there are only 94 observations available, whereas this number increases up to 260 observations in 2009.

In total, data on 395 different MFIs in 17 Latin American countries was collected for this time period. Only in 2004 there are no observations for Panama and in 2012 and 2013

(27)

Table 4: Observations per year Year Number of observations

2004 94 2005 139 2006 206 2007 229 2008 254 2009 260 2010 259 2011 259 2012 194 2013 171 Total 2065 Source: MIXMarket (2015)

Table 5: MFI-year observations Number of years Number of MFIs

1 54 2 36 3 42 4 37 5 39 6 40 7 33 8 49 9 43 10 22 Total 395 Source: MIXMarket (2015)

none for Chile available, which leads to 16 country observations for these years. The observations per country are rather unbalanced with 250 and more observations for Peru and Ecuador; and less than 30 for Chile, Haiti and Panama. However, Mexico represents the country with the most MFIs in the sample regarding the whole time period (see Appendix A).

Considering the number of years for which data is available for the different MFIs, it can be seen in Table 5 that there are 22 MFIs for which data over the whole period is available. For almost 60% of the MFIs in the dataset at least 5 or more year observations are available and only 54 MFIs provide data for just one year.

The sample used here represents relatively old MFIs. 80% of all MFIs are at least 9 years or older. Regarding the regulatory status, only 148 out of the 395 MFIs are regulated and they count for 38% of all the observations. Split up into different ownership types, the sample contains 33 banks, 64 credit unions and cooperatives, 123 NBFIs and 175 NGOs.18 NBFIs offer similar services to those offered by banks but they are licensed under a separate category which may be due to lower capital requirements, to limitations on financial service offerings, or to supervision under a different state agency. However, they are shareholder firms which distribute excess profits to their shareholders. Cooperatives and credit unions are non-profit organisations, which are owned and controlled by its members. They are still allowed to distribute profits to its members, whereas NGOs have a genuine non-profit objective which prohibits the distribution of any profits.

Table 6 contains the descriptive statistics for all the variables used in this thesis divided by regulatory status. All the monetary variables used in this thesis were adjusted for

(28)

inflation and are expressed in 2004 US constant dollars. The consumer price indices of the respective country were taken from the International Monetary Fund database.19

Table 6: Descriptive statistics

Variable Regulated Unregulated Total

Total assets (in millions of USD) 109 10.3 46.8

Operating expenses (in millions of USD) 12.9 1.93 6.00

Personnel 547 123 281

Price of financial capital (in %) 7.59 10.46 9.39

Salary per year (in thousands of USD) 11.33 8.429 9.507

Gross loan portfolio (in millions of USD) 87.7 8.47 37.9

Financial revenue (in millions of USD) 27.3 3.41 12.3

Number of active borrowers (in thousands) 73.4 17.3 38.2

Number of female borrowers (in thousands) 46.0 12.6 25.0

Female borrowers (in %) 54.5 66.9 62.3

ALB 0.59 0.33 0.43

Reciprocal ALB 3.9 9.1 7.18

Capital-Asset ratio (in %) 21 41 33

Loan loss rate (in %) 1.5 2.1 1.9

Number of Offices 34 12 20

Borrowers per staff member 120 133 128

Portfolio at Risk 30 days (in %) 5.51 6.96 6.42

Operating expense ratio (in %) 14.98 21.69 6.42

Total costs (in millions of USD) 22.2 2.81 10.0

Net profit (in millions of USD) 4.86 1.04 2.46

Note: All values are mean values over the time period 2004-2013.

ALB denotes the average loan balance per borrower divided by the gross national income of the respective country.

7 Empirical Results

Before looking at the determinants of the efficiency and inefficiency scores, I start by reporting the average efficiency scores per regulatory status for the six models. As can be seen in the last column of Table 7, Latin American MFIs are on average more efficient with respect to financial outputs than social outputs. Since a fully efficient producer in the DEA model has a value of 1, a financial efficiency score of 0.72 indicates that, on average, Latin American MFIs could reduce their inputs by 28% keeping outputs at the same level. With respect to reaching out to the poorest of the poor, they could even reduce their inputs by almost 75%. Clearly, only focusing on the depth of outreach does

(29)

Table 7: Average efficiency scores per regulatory status

DEA Regulated Unregulated Total

Financial efficiency 0.743 0.717 0.723 Social efficiency 0.172 0.310 0.259 SFA Financial efficiency 0.936 0.886 0.905 Social efficiency 0.763 0.788 0.779 Cost efficiency 0.921 0.899 0.908 Profit efficiency 0.608 0.564 0.581

not seem to be the only objective of Latin American MFIs. The higher social efficiency findings in the SFA, which only look at the width of outreach, confirm this finding. Hence it seems that the MFIs in our sample are more concerned about the width of outreach. Distinguishing between regulated and unregulated MFIs, it can be noted that regulated MFIs have, on average higher financial and lower social efficiency scores. This is true for both the DEA and SFA model specifications. On top of that, they are also more cost and profit efficient than their unregulated counterparts. This may be seen as a first evidence that there is a trade-off between commercialisation and social outreach, considering that “it can be assumed that licensed, regulated microfinance institutions have already adopted a commercial approach” (Christen, 2001, p. 2).

Next, I investigate the relationship between financial and social efficiency by looking at the Tobit regression results for the DEA specifications. The first column of Table 8 provides the outcomes of the analysis using the financial efficiency scores as the de-pendent variable. The table shows that the main variable of interest, social efficiency, seems to be associated with financial efficiency. It enters the equation with the expected negative sign and is highly significant, pointing towards a trade-off between outreach and financial efficiency. The variables for number of offices and productivity are both positive and significant, meaning that bigger and more productive MFIs are showing a higher financial performance, confirming the assumption of economies of scale. However, rather unexpected, they seem to be less cost-efficient because the coefficient for the operational expense ratio is positive. In line with Table 7, regulation has a positive effect on finan-cial efficiency, whereas risk-taking negatively influences the finanfinan-cial efficiency and the coefficient for age is not significant.

(30)

Table 8: Determinants of financial and social efficiency using DEA

Variable Financial Efficiency Social Efficiency

[1] [2] Offices 0.0002∗∗∗ -0.0002∗∗ (0.010) (0.017) Age 0.0033 -0.0018 (0.646) (0.828) Financial efficiency -0.1575∗∗∗ (0.000) Social efficiency -0.1129∗∗∗ (0.000)

Operational expense ratio 0.1523∗∗∗ 0.7012∗∗∗

(0.000) (0.000)

Portfolio at Risk 30 days -0.0982∗∗∗ 0.0196

(0.002) (0.586) Staff productivity 0.0707∗∗∗ 0.2444∗∗∗ (0.000) (0.000) Regulated 0.0197∗∗∗ -0.0591∗∗∗ (0.002) (0.000) Constant 0.2168∗∗∗ -1.0110∗∗∗ (0.000) (0.000) Number of observations 2065 2065 Log Likelihood 1572.26 1335.94 LR χ2₍₃₂₎ _1580.99∗∗∗ _2115.78∗∗∗

Note: Although dummy variables for time and country effects were included in the estimation, they were omitted here for brevity and only the most relevant coefficients are reported. A table including time and country dummies can be found in Appendix D.

p-values are given in parentheses.

∗∗∗_,∗∗_{, and}∗ _{denote statistical significance at the 1%, 5%, and 10% levels, respectively.}

negatively, whereas a higher productivity and lower cost efficiency increase the social efficiency of MFIs. I also find that the size of MFIs has a negative effect on the social efficiency while the coefficients for age and risk are not significant.

Summarising the results of the DEA models, there is reason to believe that a trade-off between social outreach and financial efficiency does exist in Latin America. The financial and social efficiency variables show the expected negative signs in the respective equations and are both significant on the 1% significance level. Nevertheless, as already mentioned in section 4.2.1, these results have to be interpreted with caution due to the disregard of random errors in the DEA models and the characteristics of the underlying sample. The remainder of this paragraph therefore uses SFA models to give more insight into a possible trade-off between outreach and financial efficiency.

(31)

cost and profit efficiency respectively. I do not report the estimated coefficients of the production, cost and profit functions here due to brevity and because the determinants of inefficiency are the main focus of this thesis.20 _{As can be seen by the results of the LR}

tests, a translog function is preferred to a Cobb-Douglas function in all the specifications.

Table 9: Determinants of financial and social inefficiency using SFA

Variable Financial Inefficiency Social Inefficiency

[1] [2] Offices -1.5879∗∗∗ -0.1042 (0.000) (0.696) Age 0.1864 0.4940∗∗ (0.306) (0.042) Reciprocal of ALB 0.0306∗∗∗ -0.0877∗∗∗ (0.000) (0.000)

Percentage of female borrowers -0.1218 -0.0311

(0.762) (0.948) Regulated -0.4149∗∗∗ 0.0358 (0.008) (0.853) Constant -4.2671∗∗∗ -2.47221∗∗∗ (0.000) (0.000) Number of observations 2065 2065 Log Likelihood 897.38 -1250.68 Wald χ2(36) 223141.42∗∗∗ 15495.60∗∗∗ Cobb-Douglas LR χ2₍₆₎ _168.70∗∗∗ _89.76∗∗∗

Note: The coefficient estimates of the production function can be found in Appendix E. ALB denotes the average loan balance per borrower divided by the gross national income of the respective country.

In the first column of Table 9 the SFA estimation results of the financial inefficiency, with gross loans as outputs, are shown. The coefficients for size and regulatory status have negative and significant signs. Again, a negative sign in the inefficiency equation means that these variables have a negative impact on inefficiency. Size and being regulated thus influence the financial efficiency positively. With respect to the variables that capture the outreach measure, only the reciprocal of ALB is significant and positive. Hence disbursing smaller loans to the poorest clients has a negative impact on the financial efficiency and points towards a possible mission drift.

Column 2 in Table 9 shows the outcomes for the SFA when the active number of borrowers is used as an output. Note, that this is a measure of width of outreach not depth as

20_{The estimated coefficients of the production, cost and profit function can be found in the Appendix E}

(32)

in the DEA model. For this reason, it does not contain any information about the wealth of the borrower. Size, percentage of female borrowers and regulatory status are all not significant different from zero and do not seem to influence this social efficiency measure. The reciprocal of ALB enters the inefficiency equation with a negative sign and is significant. MFIs, focusing on poorer clients, are apparently more efficient in reaching a bigger number of borrowers. This would confirm the assumption, that commercialised MFIs are more selective in disbursing loans since they only see richer clients as their customers. Additionally, I find that younger MFIs are more socially efficient.

In the last two SFA models, I include both the gross loan portfolio and the number of active borrowers as output variables and investigate the determinants of cost and profit inefficiency. The results of the estimation are given in Table 10. With respect to the cost

Table 10: Determinants of cost and profit inefficiency using SFA

Variable Cost Inefficiency Profit Inefficiency

[1] [2] Offices 0.1978 -0.2383 (0.505) (0.227) Age 0.3949 -0.0818 (0.398) (0.615) Reciprocal of ALB -0.0077 -0.0063 (0.698) (0.475)

Percentage of female borrowers 1.9064∗∗ 1.3386∗∗∗

(0.027) (0.000) Regulated -0.6214 -0.1833 (0.132) (0.187) Constant -5.8894∗∗∗ -1.2053∗∗∗ (0.000) (0.000) Number of observations 2065 1759 Log Likelihood -223.54 -2069.73 Wald χ2(42) 59278.94∗∗∗ 9716.13∗∗∗ Cobb-Douglas LR χ2₍₁₀₎ _160.36∗∗∗ _98.03∗∗∗

Note: The coefficient estimates of the cost and profit function can be found in Appendix F. 306 observations for the profit function were dropped due to negative values for profit. ALB denotes the average loan balance per borrower divided by the gross national income of the respective country.