The Microfinance Business Model: Enduring Subsidy and Modest Profit

Recent evidence suggests only modest social and economic impacts of microfinance. Favorable cost-benefit ratios then depend on low costs. This paper uses proprietary data on 1,335 microfinance institutions between 2005 and 2009, jointly serving 80.1 million borrowers, to calculate the costs of microfinance and other elements of the microfinance business model. It calculates that on average, subsidies amounted to $132 per borrower, but the distribution is highly skewed. The median microfinance institution used subsidies at a rate of just $26 per borrower, and no subsidy was used by the institution at the 25th percentile. These data suggest that, for some institutions, even modest benefits could yield impressive cost-benefit ratios. At the same time, the data show that the subsidy is large for some institutions. Counter to expectations, the most heavily-subsidized group of borrowers is customers of the most commercialized institutions, with an average of $275 per borrower and a median of $93. Customers of nongovernmental organizations, which focus on the poorest customers and women, receive a far smaller subsidy: the median microfinance nongovernmental organization used subsidy at a rate of $23 per borrower, and subsidy for the nongovernmental organization at the 25th percentile was just $3 per borrower.


Introduction
Microfinance institutions aim to serve customers ill-served by traditional commercial banks. The success of microfinance in achieving wide scale reach -one count includes 211 million customers globally --has inspired social business initiatives in energy, health, education and other sectors. 1 Microfinance, though, has taken a beating in recent years. Six prominent randomized controlled trials, for example, found only a small average impact of microcredit access on marginal borrowers, though the studies found some "potentially important" (though modest and not clearly robust) impacts on "occupational choice, business scale, consumption choice, female decision power, and improved risk management" (Banerjee et al 2015, p. 14). 2 While perhaps disappointing to microfinance advocates, these modest impacts could nonetheless feed into sizable benefit-cost ratios if the costs are proportionally small too. This is indeed a fundamental premise of microfinance.
1 Data are as of December 31, 2013, reported as part of the Microcredit Summit's State of the Campaign Report 2015. Data are from https://stateofthecampaign.org/data-reported/, accessed 4-15-16. 2 As Banerjee et al (2015) describe, the six studies do not provide the final word on microfinance/microcredit impacts. Most important, the studies measure impacts only on marginal borrowers. Some borrowers were determined to be not creditworthy and would have been excluded from being served, for example, but were instead served for the purposes of the study. Other studies measured impacts in new regions for the microlenders, or new populations. Still, the studies are not far from earlier studies that credibly attend to selection biases (see, e.g., Armendàriz and Morduch 2010).
By focusing on costs, this study contributes to the missing half of the conversation about the costs and benefits of microfinance. We measure the size of subsidies using proprietary data We calculate that on average, subsidies amounted to $132 per borrower, but the distribution is highly skewed. The median microfinance institution used subsidies at a rate of just $26 per borrower, and no subsidy was used by the institution at the 25 th percentile.
These data suggest that, for some institutions, even modest benefits could yield impressive cost-benefit ratios. At the same time, the data show that the subsidy is large for some institutions. Counter to expectations, the most heavily-subsidized group of borrowers are customers of the most commercialized institutions, with an average of $275 per borrower and median of $93. Customers of NGOs, which focus on the poorest customers and on women, receive far less subsidy: the median microfinance NGO used subsidy at a rate of $23 per borrower, and subsidy for the NGO at the 25 th percentile was just $3 per borrower. 3 While most firms earn positive accounting profits, only a minority earn economic profit (which accounts fully for the opportunity costs of inputs). Accounting profit reflects an institution's ability to cover its costs with its revenues, without accounting for implicit grants and subsidies. We find 67 percent of institutions were profitable on an accounting basis (weighted by the number of borrowers per institution; just 58 percent were profitable weighted by institutional 3 As a robustness check, we estimated these figures on a subset of the sample (814 institutions) for which we had complete data on every variable. Those results are very slightly lower than those reported below (the results from the balanced panel were so similar that we do not report them). With the balanced panel, we calculate that on average, subsidies amounted to $128 per borrower, with a median of $21 per borrower and again no subsidy at the 25 th percentile. The average subsidy for commercial banks is $255 per borrower and median of $89. The median (non-profit) microfinance NGO used subsidy at a rate of $21 per borrower, and subsidy for the NGO at the 25 th percentile was just $2 per borrower. assets). Turning instead to economic profit (with the local prime rate as the alternative cost of capital), we find that only 36 percent of institutions were above the profit bar (weighted by the number of borrowers per institution). Just 18 percent of institutions were profitable when weighted by their assets.
The analysis highlights the challenge created by high fixed costs in lending. The median unit cost is $14 in operating expenses for each $100 of loans outstanding, and high fixed costs imply cost advantages when making larger loans (holding all else the same). The median commercial microfinance bank makes loans that are, on average, three times larger than the median NGO (after controlling for local conditions). That helps the median commercial microfinance bank reduce unit costs to 11 percent --versus 18 percent for the median NGO.
Institutions respond by raising interest rates. Consistent with the pattern of costs, NGOs charge more than commercial microfinance banks. After adjusting for inflation, the median microfinance lender charged borrowers 21 percent per year, as measured by the average real portfolio yield. NGOs, the institutions that tend to serve the poorest customers, lent at an average of 28 percent per year after inflation. For-profit commercial microfinance banks, in contrast, charged an average of just 22 percent per year. But these averages are deceiving. Once the data are disaggregated by target market, the analysis shows the opposite: conditional on the scale of lending, for-profits tend to charge higher interest rates and non-profits have been more successful in reducing costs and cutting interest rates and fees. This is consistent with the finding that it is not NGOs, but instead commercialized microfinance banks, that use the most subsidy per borrower.
Finally, the findings contrast with arguments that microfinance subsidies are transitional.
Subsidies should play a role in helping institutions get started, according to the argument, but 5 they should phase out within a decade, allowing the unsubsidized market to take over. (An exception is made for subsidies targeted to institutions serving the poorest and most costly customers.) Our analysis of global data shows that subsidies in fact continue to be important in microfinance, even for older institutions. Summing across the 1,335 institutions, the total subsidy -both implicit and implicit --was $4.9 billion per year. 4 Of the total subsidy, 76% went to the 932 institutions that are older than ten years. Most (99.95%) of the subsidy takes the form of equity grants and cheap capital rather than direct donations. We conclude with reflections on next steps for a more transparent policy conversation around the optimal use of subsidy in the microfinance market.

Method and data
The data are from the global database of microfinance institutions collected by the MIX Market.
Within the microfinance sector, the MIX Market is responsible for collecting and disseminating financial data on microfinance institutions, and its database is the largest industry data source on the finances of microfinance institutions. 5 The raw data reflect local reporting standards, and the MIX Market adjusts the data to help ensure comparability across institutions when measuring financial performance. We begin with the MIX Market adjustments and then make further adjustments. MIX Market adjustments are made for inflation, the cost of subsidized funding, current-year cash donations to cover 6 operating expenses, donated goods and services, loan write-offs, loan loss reserves and loan loss provisioning. In addition, the MIX reclassifies some long-term liabilities as equity, and reverses any interest income accrued on non-performing loans. We further adjust the data to reflect ideas consistent with economic definitions of profit.
The MIX Market presents a calculation of profitability: i.e., the financial self-sufficiency (FSS) ratio. This notion of financial self-sufficiency is meant to indicate whether an organization can continue operations without external donor funding, but the FSS ratio falls short of accounting for inputs at their opportunity costs. The MIX Market reports that they make a costof-funds adjustment to account for the impact of "soft loans." The MIX Market calculates "the difference between what the MFI actually paid in interest on its subsidized liabilities and what it would have paid at market terms." To do that, the MIX Market uses data for shadow interest rates from the IMF's International Financial Statistics database, using the country's deposit rate as the benchmark. 6 Yaron (1994) and Shreiner and Yaron (2001) argue that this adjustment is inadequate and that the FSS thus over-states financial self-sufficiency. The deposit rate provides a benchmark for the cost of borrowing by microfinance banks that is too low: The interest rate spread (the difference between the interest rate charged by banks to private sector customers when lending and the interest rate that the private sector offers to its depositors) is generally over 5 percentage points. (2014 World Bank data, for example, show that the interest rate spread for low income countries as a group was 11.2 percentage points and 6.4 percentage points for middle income countries as a whole.) 7 Moreover, many institutions, are not legally able to collect deposits, and even those 7 that are able to do so face transactions costs associated with deposit collection. In addition, the FSS calculation implicitly (and implausibly) assumes that an institution's equity-holders seek no real return to their investments.
By using a more appropriate measure of the cost of capital and applying it to equity as well as debt financing, we obtain a clearer view of microfinance profitability and subsidy. Our analyses assume that, if they needed to borrow on the market, microfinance institutions could obtain capital at a country's prime interest rate (the rate offered to banks' safest and most favored customers). This is a conservative correction in that it does not reflect the risks of lending to institutions whose loans are typically only partially secured with collateral, and even this adjustment has large effects.
The definition of economic profit is closely related to the subsidy dependence index (SDI) developed by Yaron (1994) and explored further by Schreiner and Yaron (2001) and Manos and Yaron (2009). But rather than calculate an index, we focus on the distribution of subsidy in the context of the microfinance business model. Key variables include: Financial Self-sufficiency ratio. In the present sample we analyze the most recent data on MFIs between 2005 and 2009. The entire database includes 3,845 institution-years, reflecting 291 million borrower-years. We focus on a cross-section with the most recent data for each institution. Most of the most recent 9 data are from 2009, a year in which the data include 930 institutions with a combined 80.1 million borrowers.
The largest sample we use contains data on 1,335 institutions: 90 for-profit banks, 235 credit unions and cooperatives, 465 NGOs, 401 non-bank financial institutions (NBFIs), and 102 rural banks. Non-bank financial institutions are a broad range of institutions that generally span the space between NGOs and banks, and we divide the sample between institutions with forprofit legal status (300 institutions) and those with not-for-profit status (101 institutions). In addition, we analyze two aggregate categories defined by the MIX Market: 826 institutions with not-for-profit legal status, and 499 institutions with for-profit legal status. 9 The key relationships are analyzed by comparing means and distributional parameters of subgroups within the sample. A series of LOWESS (non-parametric smoothed) bivariate regressions describe the distributions of the data, and multivariate regressions are used to control for relevant covariates.
A major focus is how key variables like profit, cost, interest rates, and subsidy vary with the average loan size of microfinance institutions. The average loan size variable is a proxy for the income level of customers, drawing on evidence that poorer customers tend to take smaller loans. The variable is measured at the institution-level and is an average of loan sizes that could vary broadly within the institution. To control for different levels of income and development across regions, we normalize the average loan size variable by dividing it by the country's GNI (gross national income) per capita, measured at the 20 th percentile. The step of dividing by GNI per capita is relatively standard, but it creates a potential distortion in countries in which there is substantial income inequality, making loan sizes seem relatively small compared to countries at a similar level of average GNI but with lower inequality. We thus normalize by GNI per capita at the 20 th percentile of the population to address inequality within countries.
We use the entire sample in regressions (including non-parametric regressions), but we present graphical results only for the segment of the sample containing the bulk of institutions.
The figures thus cover normalized loan sizes of 0 through 5. Half of institutions have normalized average loan sizes between 0 and 1. Only a quarter of institutions have normalized average loan sizes larger than 2.5.
Figures 1 and 2 present the data as it varies by normalized average loan size. Figure 1 shows that most South Asian microfinance institutions are concentrated in the 0-1 range. The top panel of Figure 1 shows that institutions in Latin America and the Caribbean and Sub-Saharan Africa are more widely dispersed. The bottom panel of Figure 1 shows that, as expected, nonprofits make smaller loans than for-profits, though there is considerable overlap, and some forprofits are found at the lowest ranges. NBFIs make larger loans on average (median = 1.1), and banks are still larger (median = 3.4)at the upper reaches of the sample. There is limited overlap between NGOs and commercial microfinance banks.

Average loan size and fixed costs
Much of our interest is in the pattern of financial variables across institutions in different market segments. We use (normalized) average loan size as a rough proxy for the income level of customers.
Summary statistics. Table 1 gives summary statistics on the distribution of average loan size. For the full sample, the average loan size (normalized as described above) is 2.4, but the median is substantially lower at 1.0, reflecting a long upper tail. At the 75 th percentile, the normalized average loan size is 2.5, so roughly a quarter of the sample is above the sample mean. Table 1 shows how average loan size varies across types of institutions. The row on NGOs, for example, shows a median of 0.5, a figure substantially below the median for banks (3.6). 10 As in previous analyses, NGOs and banks look and behave differently, a motivation for the disaggregation here. The mean (normalized average) loan size for banks is 6.9 and the mean for NGOs is 1.4. We asserted that NBFIs span the space between NGOs and banks, consistent with the mean average loan size for for-profit NBFIs of 2.8 and the mean for non-profit NBFIs of 2.4. Table 2 shows how different the institutions are by the gender of borrowers. The median commercial microfinance bank serves a base that is 50 percent female. The median NGO, in contrast serves a base that is 80 percent female. The NBFIs are again in the middle.

Interest rates
Figures 4, 5, and 6 show how average loan size matters to the business models of the institutions. Figure 4 gives the real (inflation-adjusted) average portfolio yield of the institution. This is a measure of average interest rates, calculated by dividing the total interest earnings and fees by the size of the loan portfolio. The figure shows that most real interest rates vary between 20% and 40%, with larger loans under 30% and smaller loans above 30%.
NGOS tend to cluster to the left and banks tend to cluster to the right, with NBFIs spanning the middle space. This is consistent with the definition of the x-axis: loan sizes are smaller on the left of the figure and larger moving to the right. Table 1 showed that the median across the sample is 1.0, so half the sample is clustered at the very left end of the figure, where average interest rates are considerably higher than to the right. The figure shows that institutions making the smallest-sized loans charge the highest average interest rates. Taking average loan size as a proxy for poverty levels, the figure shows that the poorest customers in the microfinance sector pay the highest interest rates.
Tables 3 and 4 back this up. Table 3 gives nominal interest rates charged to customers, given by the average portfolio yield (earnings from lending divided by the size of the loan portfolio). The average is 34 percent and median is 29 percent. NGOs tend to charge their customers higher rates than commercial microfinance banks (the mean is 36 percent versus 31 percent), though for-profit NBFIs charge the highest rates on average (mean = 39 percent). These rates are nominal, though, and the more telling data are in Table 4, which gives the real portfolio yield (i.e., inflation-adjusted). The general patterns persist, but the numbers are smaller. The inflation-adjusted average is now 25 percent and median is 21 percent. NGOs again tend to charge their customers higher rates than commercial microfinance banks (the mean is 28 percent versus 22 percent), though for-profit NBFIs now look similar to NGOs (mean = 28 percent).

13
The data both affirm and complicate a statement made on the CGAP website: "For-profit MFIs … don't generally charge their clients more than non-profit MFIs." 11 Tables 3 and 4 show that for-profits charge slightly more on average, but the distribution of real portfolio yields is largely overlapping. That picture, though, is given nuance in Figure 4. The figure shows that when the data are segmented by customer scale (as given by normalized average loan size), banks charge less because they cluster at larger loan sizes. NGOs charge relatively less when attention is limited to smaller loan sizes. The general picture shows that the CGAP response is based on an apples to oranges comparison. Once the scale of loans is considered, the for-profit providers are seen to charge higher rates in the markets where NGOs tend to cluster.
Regression analysis. The pattern of interest rates holds after controlling for other variables. Table   5 presents regressions that show a quadratic relationship between real portfolio yield and average loan size, controlling for other factors. We estimate the following equation describing variation in yields: (1) Yi = α + β1Avg Loan Sizei + β2Avg Loan Sizei 2 + β3Regioni + β4Agei + β5Assetsi + β6Ownershipi + β7Ownership*Loan Sizei + β8Ownership*Loan Sizei 2 + εi Where Yi is the real portfolio yield of microfinance institution i. Controls include regional dummy variables; the age and size of each microfinance institution (measured by total assets); and ownership type using the same categories as in the tables presented thus far --bank (forprofit), credit union/cooperative (not-for-profit), NGO (not-for-profit), NBFI (for-profit), NBFI (not-for-profit), and rural bank. We interact the ownership type indicator variables with average loan size (divided by the per capita income at the 20 th percentile of the population) to allow the 14 relationship between loan size and yields to vary across types of institutions. The omitted ownership category is not-for-profit NBFIs. Thus, β1 and β2 describe the relationship between loan size and yields for that group of institutions. To assess whether that relationship is significant for other ownership types, we add β1 to β7 and β2 to β8 (see t-tests at the bottom of the table). β7 and β8 also provide tests of the whether the coefficients for the average loan size variables for other ownership types are statistically distinguishable from those for institutions in the omitted category. Standard errors are clustered at the country level. Table 5 shows that portfolio yields are significantly lower in Europe and South Asia, and for older and larger institutions. In models 2-5, the coefficient for average loan size is negative while that for the square of average loan size is positive, thus confirming the quadratic relationship in Figure 4. In model 5, the lack of statistical significance of the interactions between the ownership type variables and average loan size indicates that the declining quadratic relationship for not-for-profit NBFIs (the omitted category) holds also for other ownership types.
This is also confirmed for NGOs, for-profit NBFIs, and credit unions/cooperatives by the significant t-statistics at the bottom of the table. The patterns are similar for rural banks, but the cell size is small and the coefficients are not estimated with much precision. The exception to the declining quadratic relationship between loan sizes and yields is for-profit microbanks.
Coefficients for their interactions are significant and of the opposite sign as those for not-forprofit NBFIs, and the t-tests at the bottom of the table indicate a marginally significant declining relationship between loan size and yield for banks, but no significance on the interaction with the square of loan size (and thus less evidence of a quadratic relationship). The less pronounced patterns for banks are also suggested by Figure 4. Table 6 gives the operating expense ratio, an institution's total operating expenses divided by the loan portfolio. This is roughly an institution's transactions costs per dollar lent. The patterns mirror the patterns for interest rates: NGOs have higher costs than commercial microfinance banks (a median of 18 versus 11, and a mean of 23 versus 16). Not-for-profits as a group have slightly higher costs but are essentially indistinguishable from for-profits on average.

Operating expenses
Costs are partly fixed and partly variable. With high fixed costs, larger-sized loans have lower unit costs, giving a cost advantage (all else the same) to institutions making larger loans.
Differences in unit costs emerge when disaggregating by average loan size. Figure 5 shows that unit costs are substantially higher when loans are small, reflecting the relatively large fixed costs involved in microfinance operations. The low-end institutions with higher operating expenses also charge higher interest rates. The figures thus show why institutions charging higher interest rates are not necessarily more profitable -and below we show that they are not, generally.
Averages again are misleading: the figure also shows that NGOs have brought down costs on the low end, and NGOs have lower costs in the part of the distribution that they dominate (i.e., between a normalized average loan size of 0 and 1).
The regression results in Table 7 show quadratic relationships between operating expenses (per dollar lent) and average loan size, as seen in Figure  with each other indicates that they are describing related aspects of the business models used by different institutions, and the environments in which they operate (as reflected in the significant coefficients for the control variables).
One difference between the operating costs and yields regressions is that average loan size and its square are not statistically significant for institutions in the omitted category (not-forprofit NBFIs) in model 5, though the signs and magnitudes for those variables are not far off from those in models 2-4, in which all institutions are grouped together. The coefficients on the interactions between loan size (and its square) and the ownership type variables and the t-tests at the bottom of the table indicate that the quadratic relationship between operating costs and loan size is especially pronounced for the not-for-profit NGOs. Those tests also indicate significant relationships for average loan size and its square for the for-profit NBFIs. The loan size variables are also marginally significant for commercial microfinance banks. In all, there is a strong correspondence between the portfolio yields and operating costs regressions across types of institutions.

Profit: Financial self-sufficiency
We begin with the MIX Market's measure of profitability, the financial self-sufficiency (FSS) ratio. The FSS captures the difference between revenues and expenses, with adjustments made to account for some implicit subsidies.
Summary statistics. and for-profit institutions (FSS = 102 at the median).
In the analyses that follow, we show that these patterns result from the assumptions in the MIX Market formula. Even though the institutions are deemed "financially self-sufficient" or close to it, there is still substantial subsidy running through the sector once the shadow cost of capital is defined at a realistic level and applied broadly across financial categories.

Subsidy per dollar lent and economic profit
Tables 9 and 10 give our calculation of the subsidy per dollar lent. The first important step in the calculation is to use the prime rate as the shadow cost of capital. We use the local prime rate, with the idea that the institution would have to turn to local sources for financing if soft loans were not available. The local interest rates reflect regional economic conditions, and they allow us to abstract from currency risk, political risk, and similar concerns when making cross-country financial comparisons.
The second important step is to account for returns to equity. In the MIX Market's FSS calculations, it is assumed that equity donations get zero real return (the only adjustment is for inflation). Table 9 gives the resulting data. The mean subsidy per unit lent is just 10%, and 2% at the median. The 25 th percentile is zero, showing that at least a quarter of the sample is unsubsidized by this measure.
The problem, as noted above, is that it's the wrong measure. Were the donated equity to be replaced with equity provided by commercial investors, a competitive return would be expected. In accord with that logic, donated equity should also be valued at the shadow capital cost (which we, conservatively, take to be the local prime interest rate). The resulting data are in Table 10 which shows modestly larger subsidies. The mean subsidy per unit lent is now 13%, and 5% at the median. Subsidy for the 25 th percentile is again zero, showing that at least a quarter of the sample is unsubsidized by this measure.
Turning to categories of institutions shows that the subsidy per dollar is highest for the institutions focused on poorer customers (as proxied by loan size). NGOs have an average subsidy per unit lent of 18 percent and a median of 8 percent. In contrast, commercial microfinance banks have a mean of 15 percent and a median of 8 percent. In line with these results, not-for-profits have a mean subsidy per dollar lent of 15 and a median of 6, in contrast to a mean of 10 percent and median of 3 percent for for-profit institutions.
Taken together, the results seem to suggest that subsidies are targeted toward poorer households, and that, as a fraction of loans received, the poorest gain most from subsidies. This pattern is clear in Figures 7 and 8, which show that subsidies per dollar lent are the highest for institutions with the smallest average loan sizes, falling sharply for institutions serving better-off customers (as proxied by loan size). These calculations mirror the data in Table 10 and use the same subsidy definition. Figure 9 shows how the data on subsidy per unit lent lines up with the gender-orientation of institutions. The relationship for NGOs is flat. For NBFIs, however, the largest subsidy per unit goes to institutions that tend to favor men. The curve for banks draws on a small sample (n=46) but shows a generally pro-female orientation of subsidy.
We use regressions to test whether the bi-variate relationships between subsidies and proxies for target market (average loan size and the share of lending to women borrowers) hold when we control for additional variables that could account for the level of subsidies received by microfinance institutions. The equation that we estimate is: The dependent variable, Subsidy, is measured as either subsidy per dollar lent or average subsidy per borrower for microfinance institution i. The subsidy calculations use the local prime lending rate as the shadow cost of capital, as described above in the text. In equation (2), average loan size is the proxy for an institution's target market. In some specifications, we replace average loan size with the share of lending to women as our proxy for target market. As in the regressions relating average loan size and portfolio yields/operating costs, we include dummy variables for different ownership types, and we also interact those variables with our proxies for target market. Similarly, we include regional dummy variables and the age and size of each institution as control variables.
In our fullest specifications, we include portfolio yields, the ratio of operating costs to assets, and the ratio of capital costs to assets as explanatory variables. These controls are routinely used in regression analyses describing microfinance profitability, portfolio quality, and other outcomes. 12 Given the tight links between average loan size and yields/operating costs described above we expect the inclusion of these variables to substantially reduce the explanatory power of the target market variables in the subsidy regressions. To the extent that 12 See for example Cull et al. (2007). 20 this is so, we will view subsidies, loan pricing, and operating costs as elements of a package designed to serve a particular target market. In short, these are related components of a specific business model. But model 3 indicates that the situation is more complicated than that. The negative significant coefficient for average loan size indicates that subsidies are strongly declining for institutions in the omitted category, not-for-profit NBFIs. Insignificant coefficients for the interaction between loan size and the NGO, for-profit NBFI, and microbank dummy variables mean that we cannot reject the null that the negative relationship between loan size and subsidy per dollar is the same for them as for not-for-profit NBFIs, although the t-statistics at the bottom of the table indicate that the relationship is much weaker in a statistical sense for those groups.
For rural banks and credit union/cooperatives, the interactions suggest that the subsidy per dollar increases with average loan size, but again those comprise a very small share of the institutions in our data set. Finally, when the yields and costs variables are included in model 4, there is no longer a significant negative relationship between loan sizes and subsidy per dollar for any of the institutional types, suggesting subsidy, operating costs, and loan pricing are an interrelated package designed to target specific market segments.
In contrast, there is not a significant positive relationship between the share of female borrowers and subsidy per dollar in any of the regressions in  Figure 9, in which no ownership type displays a strong positive relationship between subsidy per dollar and the share of lending to women, and NBFIs that lend more heavily to men are those that rely most on subsidies. The regressions in Table 12 merely clarify that forprofit NBFIs are the ones driving the negative relationship between subsidy per dollar lent and the share of lending to women in Figure 9.

Subsidy per borrower
The picture changes when we turn to subsidies per borrower, rather than per unit lent. Table 13 shows that the mean subsidy per borrower is $84 and $10 at the median when assuming that equity-holders only need to keep abreast of inflation. Still, most for-profits are subsidized. Figure 10 gives two views of the data. The first gives data using official exchange rates, and the second gives data with purchasing power parity (PPP) exchange rates. The top panel shows a clear upward-sloping line, such that institutions offering the largest-sized loans end up more heavily subsidized than institutions making the smallest loans. The same is true for the data with PPP exchange rates, with a dip to the right in a location with sparse data. Figure 11 shows the parallel figures but disaggregated by the type of borrower. The subsidy per borrower stretches toward $200 for institutions making the largest sized loans. In PPP terms, that is roughly $500. Table 16 which  The third pair of columns makes a modest adjustment, assuming that the appropriate opportunity cost of capital should be given by the US prime lending rate. The perspective is that the donors, most of which are based in richer countries like the US, might see that as their benchmark for lending in the market. Even with this modest adjustment, now only roughly 45 percent of the sample is seen as profitable (weighted by the number of borrowers per institution; just 30 percent were profitable by this definition when weighted by their assets). In the final pair of columns, the most realistic assumption is used: the prime rate in the institutions' local market.
This accommodates local inflation and the ability to raise money on local markets. Now, the percentage of institutions that are profitable falls to 36 percent when weighted by borrowers and just 18 percent when weighted by assets.
It is sometimes argued that larger institutions tend to be more profitable than smaller ones. Thus, while there may be many unprofitable institutions, most people are served by profitable institutions and most assets are held by profitable institutions. That possibility is not borne out in the data. The final result shows that, rather than being commercially viable, just over two-thirds of microfinance customers are served by institutions not earning economic profit, and roughly 80 percent of assets in the sector are held by institutions that are not truly profitable -once realistic (but still conservative) assumptions about capital costs are included in calculations.

Women
The second proxy for the social outreach of institutions is the fraction of active borrowers who are women as a fraction of all active borrowers. Figure 3 shows that average loan sizes and a pro-female focus are negatively correlated, in line with the assumption that smaller loans tend to be made to customers who are poorer and less connected to the broader financial system. As institutions make larger loans, their focus is also more heavily on men. The negative relationship plays out through the relationships for subsidy described below. Regression analysis. We did not see a significant positive relationship between the share of female borrowers and subsidy per dollar in any of the regressions in Table 12. In Table 18, models 1 and 2 show that, on average, subsidy per borrower is negatively linked to the share of lending to female borrowers in our sample. This provides another strong indication that subsidies are not used to target women by the institutions in our sample. However, as in Table 17, the coefficients for the simple ownership dummy variables in Table 18 play an important role. With the exceptions of the rural banks dummy and the not significant coefficient for the omitted category (not-for-profit NBFIs), these are all positive and highly significant ranging from $223 for non-profit credit unions/cooperatives to $630 for for-profit NBFIs. This suggests that 26 institutions that make no loans to women have very high subsidies per borrower. And the large negative coefficients (in absolute value) for the interaction terms for banks, credit unions/cooperatives, NGOs, and for-profit NBFIs confirm that the average level of subsidy per borrower drops precipitously as institutions devote a higher share of their loans to women. In all, the regressions provide strong evidence that subsidies are not being targeted to support lending to women.

Changes over time
Microfinance experts have argued that institutions should aim to be free from subsidies after roughly seven years from their start. For example, the Consultative Group to Assist the Poor (2006) released a widely-distributed summary document, Access for All, which argued that "Donor subsidies should be temporary start-up support designed to get an institution to the point where it can tap private funding sources, such as deposits." (Helms 2006) To explore this, in Table 19, we break the sample into institutions younger than 10 years and those that are 10 years older or more. The median age in the younger group is 5 years, the median age in the older group is 18. The difference in age is large enough that we ought to see the older group with less subsidy if they follow the expert guidelines.
We show that that the guidelines are routinely violated. The older group has somewhat larger loan sizes (a normalized average loan size value of 2.5 versus 2.2 for the younger group) and smaller subsidies per borrower. They have reduced subsidy, even if it is not eliminated. The subsidy per dollar lent is 20 percent for the younger group and 9 percent for the older group (using the local prime interest rate as the alternative cost of capital and making adjustments to both equity and debt). But when we turn to subsidy per borrower, we see an average of $106 for the older group and $172 for the younger (and $20 versus $37 in the medians). The differences are not large in an absolute sense, and they clearly counter the notion that subsidy would disappear. Figure 15 gives results on profitability that disaggregates the results in Figure  There are good reasons that subsidy does not disappear. First, subsidy may continue to be optimal (e.g., market failure may persist, as may externalities). Second, subsidized credit may continue to be available in quantity, so the institutions take advantage of it -while donors feel pressure to move large amounts of capital to places where it will be invested relatively safely.
Third, institutions that are expanding continue to be in start-up mode as new regions and new products develop. Thus the idea that they have a single start-up period does not accord with the reality of institutions that, even at 18 years of age, continue to expand into new markets.
The patterns are generally consistent with the role of subsidy entering through "soft loans" and "soft equity." In that case, the total amount of subsidy tends to increase with scale.
Since institutions tend to get larger as they get older, it follows that (all else the same) subsidy per borrower naturally grows over time, rather than diminishing as microfinance rhetoric suggests.

Conclusion
The microfinance business model is challenging by definition: If achieving success was possible with standard banking procedures and products, there would be no need for microfinance.
The finding that subsidies are relatively large and enduring for some commercial microfinance institutions does not imply that microfinance commercialization is a failure or that investors should turn from microfinance. But it reinforces the need for cost-benefit determinations. In a related way, dependence on subsidies does not disappear as institutions get older, and in fact the older institutions continue to use considerable subsidies. The evidence poses a challenge for the narrative that subsidies are helpful at first but will naturally diminish over time.
The greatest challenge is that the long-standing rhetoric on subsidies and commercialization -which generally argues against the continued use of subsidies --appears to be consistently out of alignment with realities in practice. Having a transparent conversation about the uses and patterns of subsidies is an important step to making sure that subsidies are being used optimally. The evidence suggests that subsidies are likely not being used optimally.
By tilting away from those who may be able to benefit most from subsidies (poorer customers and women), microfinance subsidies support institutions that may be worthy of support, though perhaps not the most worthy, at least from the vantage of traditional social analysis.
The findings also point to the importance of pursuing new ways to change the cost structure faced by most microfinance institutions. Digital payments and innovations like mobile money have the potential to create business models that allow for reaching the poorest customers sustainably (Gates and Gates 2015). If hopes prove real, they may provide the elusive path for microfinance to reach its promise as a "social business." 29 Finally, the finding that per-borrower subsidies are in fact relatively small for parts of the NGO sector, especially institutions more focused on women and those institutions making smaller loans, reinforces the need for cost-benefit analyses to complement impact studies. Our cost calculations place into context pessimistic conclusions based only on impact studies. In some cases, the findings on cost and subsidy may even reverse those pessimistic conclusions.             0.0015 0.00918 Notes. *, **, *** represent significance at the 10, 5, and 1 percent levels, respectively. The omitted category in models 3 and 4 is not-for-profit NBFIs. All models estimated using OLS with standard errors clustered at the country level. Original, underlying data provided by Microfinance Information eXchange, Inc. (MIX).  0.556 0.00668 Notes. *, **, *** represent significance at the 10, 5, and 1 percent levels, respectively. The omitted category in models 3 and 4 is not-for-profit NBFIs. All models estimated using OLS with standard errors clustered at the country level. Original, underlying data provided by Microfinance Information eXchange, Inc. (MIX). Note: Subsidy total= Opportunity costs for equity capital (Inflation rate) -Profit before tax + Adjusted in kind subsidy + Opportunity costs for loan capital (Prime -actual paid rate)   0.501 0.206 Notes. *, **, *** represent significance at the 10, 5, and 1 percent levels, respectively. The omitted category in models 3 and 4 is not-for-profit NBFIs. All models estimated using OLS with standard errors clustered at the country level. Original, underlying data provided by Microfinance Information eXchange, Inc. (MIX). 0.986 0.939 Notes. *, **, *** represent significance at the 10, 5, and 1 percent levels, respectively. The omitted category in models 3 and 4 is not-for-profit NBFIs. All models estimated using OLS with standard errors clustered at the country level. Original, underlying data provided by Microfinance Information eXchange, Inc. (MIX).