The Price Elasticity of Microcredit
An Empirical Study of Microcredit Clients’ Elasticity to
Interest Rate Changes
Master Thesis
Thesis submitted in partial fulfillment of the requirements to obtain the degree of Master at the University of Groningen
by Ying Wang
Student number 1702734 Study program MSc Finance Supervised by Francesco Cecchi
1. Introduction
Microfinance, including microcredits, micro-savings, micro-insurance and money transfers collectively, intents to provide financial services to low-income clients who are precluded from the traditional banking services (van Rooyen et al., 2012). Distinct from traditional and commercial banking, microfinance, which provides services that cover a wider range of market geographically, is one of the most dynamic industries worldwide in terms of development operation and poverty alleviation (García-Pérez et al., 2017). On the one hand, microfinance institutions (MFIs) devote to improve the lot of the poor individual so as to enhance international development and buffer poverty. MFIs provide the poor households microcredit of loans needed without collateral requirements, so that they do not have to lend from informal domestic money lending institutions that are charging for extremely high interest rates with a purpose to exploit profitability from the poor (Ibtissem and Bouri, 2013). On the other hand, MFIs have a wider range of population to diversify risks comparing to
traditional financial institutions in order to achieve objectivities and sustainability. Therefore, microfinance offers a win-win position to both the demand (borrowers) and the supply (lenders) sides of funds (Cheng, 2007).
The sensitivity of credit demand to the interest rate changes could be captured by price elasticity, which is the proportionate change in quantity demand to a given change in price (Anderson et al. 1997). Findings regarding to the sensitivity of credit demand in microcredit market are rather mixed. Among all, Karlan and Zinman (2008) even found kinked demand curves. Their salient finding indicates that there is a certain point of interest rate on the demand curve beyond which the credit demand decreases sharply with increased interest rates, while demand differs only slightly with any interest rate changes at and below that certain point. This result is rather interesting because it reveals the different impacts on microcredits demand of interest rate decreases and increases. The core interest of this paper lies on microloan
borrowers’ sensitivity towards loan price changes. Main focuses lie on three
such as borrowers’ age or education levels are not taken into account in this study. A limitation of the paper is that the first assumption is rather strong considering a possible existence of reverse causality. However, the assumption is not necessary to be unrealistic especially for nonprofit MFIs, since their clients have very closed to zero price elasticity. Another limitation addresses to the selection bias of the data due to missing observations, because the possibility that missing observations might contribute to total different results is unable to be eliminated. Nevertheless this study provides two possible policy implications. One is that in order to correct
unnecessarily high lending interest and prevent profit exploitation to the poor
individual, it seems feasible to leave for-profit MFIs to the market competition while keep nonprofit MFIs closely monitored. Moreover, purposes of outreach and female empowerment require more approaches rather than merely lowering microcredit interest rates.
The structure of this paper is as follow. Section 2 includes reviews on related literatures, Section 3 illustrates hypotheses, Section 4 presents in detail the
methodology used to test for hypotheses, Section 5 provides summary discusses for the data, Section 6 summarizes main empirical results and discusses limitations, and Section 7 presents conclusion.
2. Literature Review
There have been many attempts to investigate the price sensitivity for microcredit. Influential advocates promote that the poor are insensitive to prices for microcredit because poor borrowers who are eager for capital should have high marginal returns to their investments due to diminishing marginal return to capital with scale, and moreover, the existence of informal moneylenders who charge extremely high interest rates is a good support (Dehejia et al., 2012). This argument is buttressed by several literatures. Bell et al. (1997) ran a survey on selected rural Punjabi households and estimated models based on the survey data to analyze the credit market in the area. The results of their study suggested that credit demand of the poor households is inelastic to interest rates, and exceed credit demand flushed from regulated credit market to informal credit market due to credit rationings. Furthermore, Kochar (1997) analyzed Indian cultivator households’ credit demand in both informal or private sector and the regulated government sector based on the All-India Debt and
Investment Survey (AIDIS). The outcomes of his study showed that in general, the
opposite. Dehejia et al. (2012) examined the impacts of increased interest rate on microfinance in the slums of Dhaka, Bangladesh by investigating the records
collected by a specific microfinance program that offers a flexible lending policy. By analyzing the interest rate variations between branches under an unexpected price change, they found that borrowers are sensitive to interest rate increase and the decrease in demand for loans is correlated with decreased in interest income earned from smaller-scale and newer customers. Moreover, Karlan and Zinman (2013) partnered with a Mexican MFI to run a field experiment by using randomized controlled trail (RCT) in Mexico. In this experiment, they randomly assigned
participants into a treatment group that paid a lower interest rate and a control group that paid a higher interest rate. The results from the experiment showed that lower interest rates generate significantly more borrowings and the amount of borrowing becomes more sensitive to price changes over time. Likewise, Briones (2009) conducted an econometric test with survey data of small rice farmers in Philippines. Equal amounts of borrowers were selected from both representative lending
cooperatives and control groups across regions to form the dataset to run regressions. The findings of this study reported that the response of formal borrowing to lending rate is unitary elastic1. In addition, Bogan et al. (2015) implemented a unique survey methodology on clients of a large non-profit MFI in Latin America to analyze the intensive margin2 of microcredit demand. In the survey, participants were asked whether and which amounts they would borrow if interest rate increases by
percentages. Their results demonstrated that the microcredit demand is indeed elastic to interest rate changes. Apart from these, Karlan and Zinman (2008) implemented an experiment partnered with a large and mature local consumer microcredit lender who operates in the cash loan market in order to derive clients’ demand curves in South Africa. They sent out letters with randomized loan maturities and interest rates that were either below, above or at the lender’s standard rates to their former clients who earned good repayment histories, and provided approved borrowers contract rates that were lower than the ones they offered in the letters so that the borrowers could revise the sizes and maturities of their loans. Based on this experiment, they found that demand curves are kinked and slightly downwards sloping but relatively flat with respect to changes of interest rates below the lender’s standard rates, while price sensitivity increases sharply with a steeper slope when interest rates offered to the
clients are above the standard rates. Their findings implied that microcredit borrowers are relatively insensitive to credit prices changes at interest rates that are equal and below the lender’s standard rates but they become sensitive to rate changes once the rates surpass the standard points.
The existing literatures have certainly addressed important implications to the sensitivity of credit borrowers to interest rate changes. However, these existing results are fairly mixed and lack of a consensus. Besides supports for both elastic and
inelastic price elasticity for microcredit, the result proposing kinked demand by Karlan and Zinman (2008) should not be neglected. The theory of kinked demand dates back to the 1930’s when Hall and Hitch (1939) brought out the concept (Spengler, 1965). In the theory, the demand curve is kinked with the average cost at the peak to each seller of the same good, price cut is very likely to be matched by competitors while a price increase is not in a competitive market (Hall and Hitch, 1939). Thus, an individual firm will face a demand curve with a kink at the existing price level (Sen, 2004). Even if a discussion of the theory is not an attempt of this paper, a rather clear indication can be addressed based on the theory. Microcredit borrowers could have different elasticities responding to credit prices deviating to different directions from the existing price level, which Karlan and Zinman (2008) defined as the standard rate. Considering the existence of informal moneylenders, it is not unreasonable to expect the poor individuals to shift their borrowings from MFIs to moneylenders if MFIs increase their interest rates, while this shift is unlikely if MFIs lower their rate.
Current studies investigate the price elasticity of microcredit demand through a single big and mature MFI or a unique microfinance product the authors selected. Dorfleitner et al. (2013) claimed that MFIs are distinguished with specific
these pattern is consistent with liquidity constrains that decreases with income. Therefore, it is sensible to assume the poorer borrowers to be less elastic to the same amount of interest rate changes comparing to the less poor ones. In additional to a finding of Kusum Mukherjee (2013) stating that MFIs and moneylenders are equally exploitative to the ultra poor borrowers without market competition, the strength of competition between different markets where nonprofit and profit MFIs anticipate is also an important factor. Consequently, not discriminating the types of MFIs in analyses might induce distortion and misinterpretation in the findings.
The responds of borrowers to microcredit interest rate changes can as well be affected by gender. Female borrowers are assumed to be belonging to the poorer individuals and they are also found to be facing relatively more critical lending criteria’s, such as paying comparatively higher interest rates (Dorfleitner et al., 2013), receiving relatively smaller loan amounts and facing a glass ceiling for loans (Agier and Szafarz, 2012). As a result, female microcredit borrowers are to be facing relatively server credit rationing and having less alternative accesses to microloans. Therefore, female borrowers are supposed be less elastic regarding to interest rate changes comparing to male borrowers. In additional to gender issue, previous
research found a trade-off between microfinance outreach and sustainability, implying that MFIs with a bigger outreach are less efficient (Hermes et al., 2011). Female borrowers are treated as an indication to MFIs outreach (Hermes et al., 2011) while profit margin and risk coverage can represent MFIs sustainability. Considering the trade-off between outreach and sustainability, these variables from the two categories are supposed to affect the microcredit lenders loan amounts in the opposite ways. Finally, another issue addresses to which rate is more appropriate to be regarded as the interest rate for microloans also deserves attention. A popular proxy for microcredit interest rate is yield on gross portfolio provided by the Microfinance Information Exchange (MIX Market). The variable is expressed as the total of all incomes from loans as a percentage of the lender’s average annual gross loan portfolio, and it is calculated as financial revenue from loans divided by average
gross loan Portfolio adjusted by inflation. However, Rosenberg et al. (2013) claimed
sensitivity. For such, Rosenberg (2002) proposed a pricing formula that includes five elements, which are administrative expenses, loan loses, costs of funds, desired
capitalization rate and investment income respectively. He stated that the interest
rates calculated with this formula are the rates that institutions need to realize if they want to fund their growth primary on commercial funds and these rates yield useful approximations. In other words, the formula provides the substantial interest rates for MFIs. Nevertheless, Dorfleitner et al. (2013) found that nearly any MFIs are using this formula to set their interest rates. They further observed that actual rates were below what Rosenberg (2002)’s pricing formula proposed because the donations amounts received by some MFIs were not included in the formula. Instead,
Dorfleitner et al. (2013) argued the most accurate formula to proxy effective interest rate should include the average gross loan portfolio, the write-off ratio, the total fees
and commissions and the total interest incomes, because an effective lending rate
should include not only the existing and efficient loans but also the loss loans that were not re-paid. For this reason, this lending rate measurement is more appropriate to be applied when analyzing the demands side of microcredits.
In summary, existing studies testing microloan borrowers elasticity responding to loan interest rate changes are country/regional specific and are small-scale
experiments implementing different methodologies, which do not have macro-level focus across countries. These studies show heterogeneity across country/regional bases but somehow failed to capture the common features. An obvious shortcoming will be that the results from these studies might not be applicable to a different geographic location. Furthermore, experiments studies paired with a single specific MFI could suffer from selection bias and thus failed to conclude the behaviors of borrowers who did not select the studied MFI. Besides, MFI heterogeneity such as profit status is not discussed. Moreover, considering the fact that interest rate charge to borrowers is individual specific, what rate to be used to proximate the aggregate interest rate is also lack of consensus. This paper fills these gaps by analyzing microcredit borrowers’ elasticity responding to loan price changes in an aggregate macro-level and discriminating types of MFIs by their profit status, geographic locations and years of experience. In addition, an advance feature of this paper is to use the interest rate proxy proposed by the pioneer work of Dorfleitner et al. (2013).
3. Hypotheses
3.1 Price elasticity of microcredit
when the price of that good decreases (Anderson et al., 1997). The theory describes an inversed relationship between the dynamic of price and the quantity demand of a good, which indicates a downward sloping demand curve in the diagram with price and quantity demand. The sensitivity of quantity demand to price changes is measured by the price elasticity of demand, which is the proportionate change in quantity
demand to a given change in price (Anderson et al. 1997). Hildenbrand (1983) has proven with a mathematics model that within a group of individuals who have different total expenditures but share the same preference relation, the partial market demand curve for every good is strictly decreasing with higher prices. More recently, Maldonado and Oliveira (2011) provided a micro model to reveal the consumption habits of loans in traditional credit market. Based on the model, they showed that loan demand curves are downward sloping and kinked with respond to interest rate
increases. In microfinance market, microcredit or microloan is served as a commodity supplied by MFIs to a group of poor individuals and the price of this commodity is seen as the costs for the poor households to obtain the credit. Even though the poor households differ in total expenditures, they have the same preference relation, which in other words is that they are indifferent to prefer more wealth to less wealth. The costs incurred to the poor’s borrowings are composed with nominal interest rate payments, borrower loan transaction costs and changes in purchasing power of money (Adams and Nehman, 1979). Although the changes in purchasing power of money is beyond the focus of this paper, the effective lending interest rate that consists with nominal interest rate and transaction costs is arguably representative to the price of the microcredits to the borrowers.
Comprehensively, it is convincing to assume that the demand of the poor individuals for microcredit also follows the law of demand theory. This implication goes with an inference that microcredit demand is inversely related to the changes of effective lending interest rates. Nevertheless, an essential point regarding to quantify of demand needed to be specially stated. The real quantity of demand for credits is bigger than the magnitude MIX Market data capture. Because some people who want microcredits are not creditworthy, which indicates that MFIs cannot lend to them without suffering from unsustainable default levels (Anand and Rosenberg, 2008). The dataset fails to capture the demand of un-credit-worthy borrowers. Therefore, the data represents quantify of credit demand mixed with result from credit screening. Consequently, the demand defined in this paper is not the actual quantity demand representing the whole population, but rather the quantity demand of credit worthy borrowers. In summary, the first hypothesis is developed as below:
Hypothesis 1 (H1):
3.2 Asymmetric price elasticities of microcredit
Karlan and Zinman (2008) found kinked demand curves for microcredit demands corresponding to price changes, which indicates asymmetric elasticities of borrowers to interest rate changes. Their results showed inelastic demand for microcredits at and below a standard rate, while demand becomes elastic once the standard rate is
surpassed. The theory of kinked demand by Hall and Hitch (1939) illustrates that in a competitive market the effect of a price cut on increased demands is less than the effect of a price increase on reduced demands. Microfinance market differences from the traditional finance market with a goal to help the poor and reduce poverty. Based on this goal, it is reasonable to expect MFIs to match up with each other for interest rate reduction whenever it is possible, such as receiving donations or subsidies, and experiencing other costs reductions. In a similar logic, MFIs are different from local moneylenders in the nature of whether to explore profits from the poor individuals. Thus, it is rather unlikely for all MFIs to follow the same action to increase their interest rates when a few MFIs in the market higher their interest rates, considering a higher interest rate asked by a MFI could be an individual case due to the cause from a higher marginal cost the MFI faces. If this is not the case, any profit-exploitation purpose will contradict the microfinance institutions’ operational goal. Therefore, impact of an interest rate reduction is assumed to be smaller than the one of an interest rate increase.
From another aspect, taking a different insight from behavioral finance,
Kahneman and Tversky (1991) suggested that loses cause greater emotion impact on an individual than does an equal amount of gains. An increased interest rate to the poor borrowers is equivalent to a lost while a lowered interest rate is regarded as a gain. Their responses to a higher interest rate therefore are to be bigger than to a lower interest rate. Consequently, they are supposed to be more sensitive to interest rate increase than to interest rate decrease. In summary, the second hypothesis is illustrated as:
Hypothesis 2 (H2):
Credit-worthy microcredit borrowers have higher price elasticity corresponding to interest rate increase than interest rate decrease.
3.3 Heterogeneous impacts of MFIs profit status on microcredit demand
Serrano-Cinca and Gutiérrez- Nieto (2014) suggested that non-government organizations (NGOs), which are similar to nonprofit organizations (NPOs) (Mersland, 2008), are attracted to serve the poorer of the poor. The Law of
Diminishing Return states when all other inputs are held fixed, increasing one input
of the poor start with less wealth and have a relatively wider range of investment possibilities that yield higher returns to choose from comparing to the less poor. Because the less poor individuals have already explored comparatively more
investment opportunities to generate more wealth than the poorer ones. Additionally, the poorer borrowers are bind with liquidity constrains and therefore have smaller ranges of choice when selecting MFIs to borrow from (Serrano-Cinca and Gutiérrez- Nieto, 2014). As a result, they are assumed to be more eager for microcredits and to be less elastic corresponding to the loan price changes. Interpreting the inference differently, nonprofit MFIs that are serving a poorer client base are supposed to face less elastic demand corresponding to interest rate comparing to for-profit MFIs and consequently, the third hypothesis follows as:
Hypothesis 3 (H3):
Nonprofit microfinance institutions have less elastic clients to interest rate changes than for-profit microfinance institutions.
3.4 Other control variables
Three variables profit margin, risk coverage and percentage of female borrowers respectively, which represent MFIs sustainability and outreach are as well included into the regressions. Profit margin represents sustainability and therefore is assumed to have a positive effect on the loan demand MFIs are able to undertake. Risk coverage is the ability of MFIs to cover their loss loans, it is indirectly indicating MFIs sustainability thus is also supposed to have a positive effect. Finally, female empowerment has been an important target for microfinance (D’espallier et al., 2011). Previous researches proved female borrowers received smaller amounts of
microcredit, face more restrictions (Agier and Szafarz, 2012) and female borrowers are paying relatively higher interest rates (Dorfleitner et al. (2013). Besides
Kahneman and Tversky (1991) claimed that females are more risk averse. It is therefore reasonable to expect MFIs serving a higher percentage of female clients undertake lower amounts of loan demand, and female borrowers have relatively lower elasticity to interest rate changes. In conclusion, the forth and fifth hypotheses follow as:
Hypothesis 4 (H4):
Microfinance institutions that have higher profit margin and risk coverage, but lower percentages of female borrowers have bigger amounts of loan issued to credit-worthy borrowers.
Hypothesis 5 (H5):
4. Methodology
4.1 Model selection
The dataset is an unbalanced panel obtained from an online platform Microfinance Information Exchange (MIX Market). Therefore, selecting a best-fitted model in prior is essential. All outcome tables in the model selection process are presented in
Appendix 1. In the model selection process, the first step is to run two pooled ordinary least squares regressions (pooled OLS) with and without control variables
respectively, which are stated as:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖𝑡+ 𝑒𝑖𝑡 (1)
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖𝑡 + 𝛽3∗ 𝑙𝑛𝑃𝑀3,𝑖𝑡+ 𝛽4∗ 𝑙𝑛𝑅𝐶4,𝑖𝑡+ 𝛽5 ∗ 𝑙𝑛𝐹𝐵5,𝑖𝑡+ 𝛽6∗ 𝑅𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝐴𝐺𝐷7,𝑖𝑡+ 𝛽8∗ 𝑃𝑆𝐷8,𝑖𝑡+ 𝑒𝑖𝑡 (2)
where 𝑙𝑛𝐷𝑝 represents the natural logarithm of the aggregate demand of credit-worthy borrowers, 𝑙𝑛𝐿𝑅 the natural logarithm of the effective lending rate, 𝑙𝑛𝑃𝑀 the natural logarithm of the profit margin, 𝑙𝑛𝑅𝐶 the natural logarithm of the risk coverage, 𝑙𝑛𝐹𝐵 the natural logarithm of the percentage of female borrowers, 𝑅𝐺𝐷 the region dummy variable, 𝐴𝐺𝐷 the dummy variable for the age of MFIs, 𝑃𝑆𝐷 the profit status dummy variable and 𝑒𝑖𝑡 the error terms that capture all other features not explained by the control variables. Additionally, 𝑖 donated for microfinance institution where 𝑖 = 1,2,3 … and 𝑡 for time period where 𝑡 = 1,2 … 19. However, the pooled OLS regressions ignore the feature structure of a panel dataset to assume homogeneity across entities in all time periods. Therefore the pooled OLS assume identical demand respond from clients of all MFIs in the entire sample dataset across time. This
assumption is obviously too strong for a big unbalanced panel dataset. As a result, a
Lagrange multiplier test (LM test) is applied sequentially to test for the random
effects. The outcomes of the LM test points at a rejection to the null hypothesis of no random effects at 5% significance level, which indicates that there is heterogeneity in the dataset and thus the pooled OLS regressions are not fitted. Consequently, the second step is to run regressions with equation (2) based on a fixed effects (FE) and a
random-effects (RE) models.
A fixed-effects model eliminates any variables that are time invariant. A
random-effects model can solve the problem to test time invariant variables but the model
variables in the regression models. This assumption is highly impractical and can be simply violated with this unbalanced dataset. For instance, issues such as religious believes could also have an impact on credit demand and religious believes are likely to be related to region locations. These two models are distinguished thus a Hausman
test is further applied to test for the consistencies of coefficients in the FE and the RE
models for model selection. The purpose to test the consistencies of the coefficients is to check which model provides estimator that converges to the true parameter values. The outcomes of the Hausman test indicates that coefficients in the RE model are not consistent and therefore, the fixed effects (FE) model is more appropriate.
The final step comes to select between an entity fixed-effects (EFE) and a time
fixed-effects (TFE) models. An entity fixed effects model control for omitted variables
that are time invariant but vary across units. While a time fixed effects model control for omitted variables that are unit invariant but vary over time. At this step, an F-test to test an null hypothesis of all coefficients of year (time) dummy variables are jointly equal to zero is run. The outcomes of the F-test reject the null hypothesis at 5% level and consequently a time fixed effects model is finally selected for all regressions. The
fixed-effects (FE) regression model with time dummy is therefore:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖𝑡 + 𝛽3∗ 𝑙𝑛𝑃𝑀3,𝑖𝑡+ 𝛽4∗ 𝑙𝑛𝑅𝐶4,𝑖𝑡+ 𝛽5 ∗ 𝑙𝑛𝐹𝐵5,𝑖𝑡+ 𝛽6∗ 𝑅𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝐴𝐺𝐷7,𝑖𝑡+ 𝛽8∗ 𝑃𝑆𝐷8,𝑖𝑡+ 𝛽9∗ 𝑌9,𝑡+ 𝑒𝑖𝑡 (3) where 𝑌 is the year dummy variable.
4.2 Data selection bias
The practice in analysis of the dataset is only feasible to analyze the observations on units for which a complete time series is available, for a consideration that the dataset is an unbalanced panel with missing observations. The main concern for an
unbalanced panel dataset is the reasons why data are missing. The reasons for data attrition could be at random3 or selection bias. The data attrition is not a problem once the cause is at random and can be solved using statistics technique such as multiple
imputations4. Conversely, the possibility of selection bias arises with missing
observations that are not at random. In case of selection bias, sample randomization is not achieved. In the presence of selection bias, how sampling takes place from the underlying population is determined by a rule instead of random sampling. For this reason, a dataset selected under this selection rule is not representative of the true
3Missing data completely at random means the missing data are a complete random subset of
the data. Missing data at random means the missing data has the possibility to be a non-random subset but this possible non-non-randomness can be totally explained by variables in the data.
population. Therefore inferences and conclusions based on this observed dataset can be distorted using standard methods (Verbeek and Nijman, 1992). Consequently, a procedure to test for data selection bias is applied in order to ensure the reliability of the results.
Verbeek and Nijman (1992) proposed a few simple tests, among which the variable adding test is applied in this paper to for detection of selection bias in the dataset. The procedure of the variable adding test is to include three variables in the regression model on the unbalanced panel and test for the significances of the
coefficients of the three included variables. The significances of the three coefficients are the proof for a presence of selection bias of the dataset. The three variables to be included in the regression model recommended by Verbeek and Nijman (1992) are respectively: 𝑇𝑖 represents the number of waves individual 𝑖 participates in the dataset, 𝐶𝑖 a dummy variable equals to 1 if individual 𝑖 is observe in all time periods and 0 otherwise, and 𝑅𝑖,𝑡−1 another dummy variable equals to 1 indicating individual 𝑖 is observed in the previous section and 0 otherwise, for which 𝑅𝑖,0 is 0 by
assumption. A regression model includes these three variables are stated as:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖𝑡 + 𝛽3∗ 𝑙𝑛𝑃𝑀3,𝑖𝑡+ 𝛽4∗ 𝑙𝑛𝑅𝐶4,𝑖𝑡+ 𝛽5 ∗ 𝑙𝑛𝐹𝐵5,𝑖𝑡+ 𝛽6∗ 𝑅𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝐴𝐺𝐷7,𝑖𝑡 + 𝛽8∗ 𝑃𝑆𝐷8,𝑖𝑡+ 𝛽9∗ 𝑌9,𝑡+ 𝛽10∗ 𝑇10,𝑖+ 𝛽11∗ 𝐶11,𝑖+ 𝛽12∗
𝑅12,𝑖,𝑡−1+ +𝑒𝑖𝑡 (4)
The outcomes suggested that 𝑇𝑖 is omitted under a fixed effects (FE) model
because the variable is time invariant and the effects have already been captured in the model. However, the significance of 𝐶𝑖 and 𝑅𝑖,𝑡−1 at 1% level indicate that the dataset is indeed suffering from selection bias. One way to deal with this selection bias is to select a subsample during a time period in which both 𝐶𝑖 and 𝑅𝑖,𝑡−1 are not
significantly different from zero, by manipulating data periods and 𝑇𝑖. The subsample will therefore not suffer from selection bias and the results should not be distorted. The regression model with selected unbiased subsample can therefore be stated as:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖𝑡 + 𝛽3∗ 𝑙𝑛𝑃𝑀3,𝑖𝑡+ 𝛽4∗ 𝑙𝑛𝑅𝐶4,𝑖𝑡+ 𝛽5 ∗ 𝑙𝑛𝐹𝐵5,𝑖𝑡+ 𝛽6∗ 𝑅𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝐴𝐺𝐷7,𝑖𝑡 + 𝛽8∗ 𝑃𝑆𝐷8,𝑖𝑡+ 𝛽9∗ 𝑌9,𝑡+ 𝛽10∗ 𝑇10,𝑖+ 𝛽11∗ 𝐶11,𝑖+ 𝛽12∗
𝑅12,𝑖,𝑡−1+ 𝑒𝑖𝑡 , 𝑤𝑖𝑡ℎ 𝑇𝑖 ≥ 𝑛 (5)
Nevertheless, there are several disadvantages of analyzing the subsample instead of the whole sample. First of all, selecting a subsample reduces the sample size as well as the information carried by the observations not included in the subsample.
subsample because it is now less possible for the subsample to represent the real population. Moreover, the results based on the subsample could severely deviate from the real situation. Taking a hypothetical example, in the original unbalanced panel dataset 50% variables of MFI A and B are observed equally, but only 10% variables are observed of A while 90% of B in the subsample due to the selection of sample periods. The results based on the subsample therefore are clearly only applicable for MFI B. All together, a serious consequence of using a subsample will be that the interpretation of the results is merely applicable for MFIs including in the subsample. All outcomes of the subsample selection show that these disadvantages indeed can cause serious problem in the case with this unbalanced panel. Over 90% of
observations are eliminated in the subsample even in the best scenario. If includes the whole sample period, a subsample contains only 69 out of 18,448 observations is selected. Consequently, the results based on the subsample can only be applied to the few MFIs included while not be generalized. As such, a representative unbiased subsample is unable to be selected and the original unbalanced panel will be applied in analyses.
4.3 The reverse causality problem
Concerning the possibility of endogeneity in the regressions model, reversed causal relationships could be existed between the effective lending rate (𝐿𝑅) and the
aggregate demand (𝐷𝑝), as well as between the profit margins (𝑃𝑀) and the aggregate demand (𝐷𝑝). In case of the reverse causality, the inferences and interpretations of the results can be pseudo. Because in additional to the effects of the independent
variables on the dependent variable intended to be tested, the dependent variable as well has a reversed causal effect on one or more of the independent variables. In this case, the causal link from independent variables to the dependent variable is pseudo even if the outcomes are significant. An attempt to mitigate the reverse causality problem is to substitute the two independent variables mentioned above with the one-year lagged values of their owns’. As a result, the regression model becomes:
in the previous period. An essential note addresses to an important assumption under which the supply of microcredit is constant and the only consideration MFIs make is the creditworthiness of the borrowers. In reality, the assumption seems to be not practical but the issues is not to be discussed in this paper.
Another important note to be mentioned is that whether to use one-year lag variables or no lag variables in equations running from (1) to (4), selections outcomes are identical. Namely, both lag and no lag variables indicate a time fixed
effects (FE) model on the unbalanced panel sample to be the most appropriate.
However the one-year lag variables revision is prioritized because it can mitigate the reverse causality problem. In conclusion, equation (6) using a time fixed effects (FE) model is ultimately selected for analyses.
4.4 Asymmetric price elasticity and groups comparisons
In order to test H2, H3 and H5, a model for group comparisons is established. The purpose of H2 is to test whether borrowers are equally sensitive regarding to interest rate increase and interest rate reduce. A dummy variable is created and included into the regression model to separate the sample into two groups. An interaction variable interacting the dummy variable and the one-year lagged lending rate is also included into the regression. The time fixed effects model is established as below:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1+ 𝛽3∗ 𝐿𝑅𝐷3,𝑖,𝑡−1+ 𝛽4
∗ (𝐿𝑅𝑖,𝑡−1∗ 𝐿𝑅𝐷𝑖,𝑡−1) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6∗ 𝐴𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡 + 𝛽8∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (7)
where the 𝐿𝑅𝐷 is the dummy variable for changes of the lending rates. 𝐿𝑅𝐷 equals to one if the change of lending rate from the previous period is positive, and equals to zero if the change is negative. The value is set equal to “missing” if the changes are zero. Consequently, for group where interest rate changes are positive, equation (7) becomes:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1+ 𝛽3∗ 1 + 𝛽4
∗ (𝐿𝑅𝑖,𝑡−1∗ 1) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6∗ 𝐴𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡+ 𝛽8 ∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (8)
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1+ 𝛽3∗ 0 + 𝛽4
∗ (𝐿𝑅𝑖,𝑡−1∗ 0) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6∗ 𝐴𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡+ 𝛽8 ∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (9)
To test whether the absolute value of 𝛽2+ 𝛽3+ 𝛽4 is significantly bigger than absolute value of 𝛽2 alone can detect whether borrowers are more elastic to interest rates increase than decrease.
MFIs’ profit status is time invariant and it is omitted in a time fixed effects model. For this reason, a same group comparison procedure will be applied to test H3. A time fixed effects model is as:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1+ 𝛽3∗ 𝑃𝑆𝐷3,𝑖+ 𝛽4
∗ (𝐿𝑅𝑖,𝑡−1∗ 𝑃𝑆𝐷3,𝑖) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6 ∗ 𝐴𝐺𝐷6,𝑖𝑡 + 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡 + 𝛽8∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (10)
where 𝑃𝑆𝐷 is the profit status dummy variable equals to one if the MFI is for-profit and to zero if is nonprofit. For the same reason, the data can be grouped by the dummy variable. The equation (10) for nonprofit MFIs is therefore:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2
∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1 +𝛽3∗ 0 + 𝛽4∗ (𝑙𝑛𝐿𝑅𝑖,𝑡−1∗ 0) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6 ∗ 𝐴𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡+ 𝛽8∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (11)
and for for-profit MFIs equation becomes:
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1+ 𝛽3∗ 1 + 𝛽4
∗ (𝐿𝑅𝑖,𝑡−1∗ 1) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6∗ 𝐴𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡+ 𝛽8 ∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (12)
Based on the same logic, in the case which absolute value of 𝛽2+ 𝛽3+ 𝛽4 is significantly bigger than absolute value of 𝛽2 alone, there is evidence to show
borrowers of nonprofit MFIs are less elastic to credit price changes than clients of for-profit MFIs.
The test for H5 follows the same procedure. The percentage of female borrowers is divided by a dummy variable into high and low percentages groups. The time fixed
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1+ 𝛽3∗ 𝐹𝐵𝐷3,𝑖,𝑡+ 𝛽4
∗ (𝐿𝑅𝑖,𝑡−1∗ 𝐹𝐵𝐷3,𝑖,𝑡) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6∗ 𝐴𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡 + 𝛽8∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (13)
where 𝐹𝐵𝐷 is the dummy variable that equals to one for group with high female borrowers percentage and to zero for group with low female borrowers percentage. Thus, for high percentage group equation (12) will become
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1+ 𝛽3∗ 1 + 𝛽4
∗ (𝐿𝑅𝑖,𝑡−1∗ 1) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6∗ 𝐴𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡+ 𝛽8 ∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (14)
and for low percentage group becomes
𝑙𝑛𝐷𝑝,𝑖𝑡 = 𝛽1+ 𝛽2
∗ 𝑙𝑛𝐿𝑅2,𝑖,𝑡−1+𝛽3∗ 0 + 𝛽4∗ (𝐿𝑅𝑖,𝑡−1∗ 0) + 𝛽5∗ 𝑅𝐺𝐷5,𝑖𝑡+ 𝛽6 ∗ 𝐴𝐺𝐷6,𝑖𝑡+ 𝛽7∗ 𝑃𝑆𝐷7,𝑖𝑡+ 𝛽8∗ 𝑌8,𝑡+ 𝑒𝑖𝑡 (15)
Likewise, if absolute value of 𝛽2+ 𝛽3+ 𝛽4 is significantly bigger than the absolute value of 𝛽2alone, the group with higher percentage of female borrowers is more sensitive to interest rate changes than the low percentage group. It will then further indicate female borrowers are relatively more elastic to credit price changes.
5. Data
5.1 Dataset
All data are obtained from the web-based platform Microfinance Information Exchange, which is also known as the MIX Market. The MIX Market is a nonprofit organization founded in 2002 that works directly with microfinance providers to generate reported data. The MIX’s local teams of analysts collect and review the data reported by their co-working microfinance providers to ensure the accuracy,
reliability and validity of the information in the datasets. They also double check against source documents such as audits and ratings, and standardize the data according to internationally accepted accounting standards aiming to provide for a more useful intra-regional comparison.5
The original dataset includes self-reported data for 22 variables, excluding time and MFI identities variables, from 2,958 microfinance institutions distinguished into profit and nonprofit status across seven regions. The seven sample regions defined by the MIX Market are Sub-Saharan Africa, South Asia, East Asia and the Pacific,
Eastern Europe and Central Asia, Latin America and the Caribbean, North America
and Middle East and North Africa, respectively. The MFIs are also sorted into three categories by their establishment ages, where New includes MFIs exist between one and four years, Young includes MFIs exist between four and eight years and Mature includes MFIs exist longer than eight years. All currency variables are measured in USD and data period is between year 2000 and 2018 but with gaps as well as missing variables. The complete original dataset is therefore an unbalanced panel containing 18,506 observations.
Due to the duplication problem in the dataset, 58 observations have to be dropped. The cause of the duplication is that some MIFs reported their data twice a year instead of once a year comparing to the rest of the majority of MFIs in the dataset. These 58 observations came from 24 both for-profit and nonprofit MFIs at all age levels (new, young and mature), which are locating across all six regions except for North
America. However, not all of the semi-yearly reported data are complete or consistent,
some MFIs reported data for only a few variable during a half-year period in June but disclosed more in December. For instance, a mature nonprofit MFI named HFSKS from South Asia reported semi-yearly in 2004 in June and December for a few variables in the dataset but this frequency only appeared once. This MFI also disclosed data yearly in 2003 and 2005. Therefore, the observations from this MFI have to be dropped out of the dataset. After the elimination of duplication procedure, only one MFI is eliminated from the dataset even though 58 observations are dropped. The final sample is formed with an unbalanced panel dataset composed with 18,448 observations from 2,957 MFIs for a 19 years’ period between 2000 and 2018. For region Sub-Saharan Africa, 2,058 observations are reported by nonprofit and 1,593 are by for-profit MFIs, while 452 observations are by MFIs who did not state their profit status. For South Asia, 1,767 observations are reported by nonprofit and 1,184 are by for-profit MFIs, while 160 observations are by MFIs who did not state their profit status. As to East Asia and the Pacific, 898 observations are reported by nonprofit and 1,046 are by for-profit MFIs, while 352 observations are by MFIs who did not state their profit status. When it comes to Eastern Europe and Central Asia, 1,449 observations are reported by nonprofit and 1,630 are by for-profit MFIs, while 106 observations are by MFIs who did not state their profit status. For Latin America
and the Caribbean, 2,969 observations are reported by nonprofit and 2,041 are by
are by for-profit MFIs, while 18 observations are by MFIs who did not state their profit status. Finally, there are merely two observations reported by one MFI who did not state its profit status from region North America. Statistically, observations proportion of nonprofit and for-profit MFIs are 53% and 41% respectively, and 6% came from MFIs without clarified profit status. Across the six regions excluding
North America, the most observations are reported from Latin America and the Caribbean, which takes up to 27%, the second is from Sub-Saharan Africa as 22%.
The least observations came from Middle East and North Africa, which takes only 4%. Besides, 25 observations for variable percentage of female borrowers have to be replaced as missing values because the values are unrealistically higher than 100%, among which nine MFIs are from region South Asia, five from Sub-Saharan Africa, nine from East Asia and the Pacific, and two from Eastern Europe and Central Asia. The descriptive statistics are generated in Table 1. The dataset is analyzed with statistics software STATA/SE15.0 version.
Table 1. Dataset descriptive statistics
Variable Observation Mean Std. Dev. Min Max
MFI ID number 18,448 108072.8 17069.28 100000 176824
Fiscal year 18,448 2008.92 4.072.933 2000 2018
Gross loan portfolio1 17,910 393.000 52 600.000 0 7 050 000.000
Write-off Ratio 12,234 .024 .241 -.127 25.711
Administrative expense/assets 11,540 .083 .115 -1.083 8.475
Loan loss rate 13,292 .071 3.944 -1.420 445.253
Total expense/assets 13,851 .261 .247 -1.083 12.752
Average assets1 14,044 66.8 473 0.008 37.100
Financial revenu1 16,373 85.200 10 900.000 -35.9 1 390 000.000
Financial revenue/assets 13,838 .269 .198 -1.083 12.644
Financial revenue from loans1 13,178 106.000 12 100.000 -35.9 1 390 000.000
Other financial revenue1 11,316 0.069 623 -0.275 66.300
Average gross loan portfolio1 14,356 50.8 300 0 16.300
Gross loan portfolio to assets 17,287 .787 1.933 0 126.816
Assets1 17,444 432.000 57 100.000 0 7 540 000.000
Percentage of female borrowers 13,605 .649 .279 0 1.00
Profit margin 16,057 1.509 638.397 -35495.62 72098.13
Risk coverage 12,597 1445.646 80290.8 -6.331 4728662
Fee and commission income on loan1 10,675 12.400 1 280.000 -0.001 1 32 000.000
Yield on gross portfolio (nominal) 11,747 .332 .221 -13.329 11.478
Yield on gross portfolio (real) 11,715 .247 .212 -13.052 10.623
1 Value is measured in million USD
The proxy variable representing borrowers’ aggregate demand for microcredit, is generated as 𝐷𝑝 and is consisted with two indexes provided by the MIX Market: gross
loan portfolio (𝐺𝑝) and loan loss rate (LL), respectively. The formula to calculate variable 𝐷𝑝 is stated as below:
𝐷𝑝 = 𝐺𝑝∗ (1 + 𝐿𝐿) (16)
Equation (16) generates the demand aggregately including the loans that were not repaid. It is of importance to mention that there are also other proxies for microcredit demands yet different proxies capture different aspects of demand. For instance, the intensive margin, which can be represented by the variable average loan balance per
borrower provided by the MIX Market, measures the loan amounts changes per
borrower, can be a proxy for demand. However, this margin variable measures demand for microcredit at an individual level and credit-worthy individuals who desire for loans are not necessary to borrow at all time (Anand and Rosenberg, 2008). Using this proxy could underestimate borrowers price elasticities aggregately because the insensitive respond of some individuals to price changes might simply due to the reason of unnecessary. Furthermore, the extensive margin that measures the changes of the numbers of loan clients can also be a proxy for demand. Since the
increased/decreased number of new borrowers is also largely affected by MFIs screening criteria’s, demand measured by this proxy could carry even more effects from the credit supply side. It is as a result not an adequate proxy to measure
borrowers responds to price changes. Consequently, considering the main attempt is to analyze microcredit borrowers’ elasticity towards credit price changes at a cross-country level, aggregate demand is more suitable and using other proxies could add noises to the results.
5.3 The effective lending interest rate (LR)
In order to study borrowers’ real price elasticity, it is also crucial to define an effective interest rate that is representative to the actual amount the borrowers are charged to acquire the credits. Several existing studies used the proxy yield on gross
portfolio provided by MIX-Market, while this proxy has been criticized to
underestimate the real amount of interest rates that borrowers are actually paying (Dorfleitner et al. 2013). Intentionally, Dorfleitner et al. (2013) introduced a different proxy for the interest rate when only MFI-level data are available, which they defined as the effective lending rate (LR). They included a variable write-off ratio to contain the percentage of the highly unlikely repaid debts on basis of the yield on gross
calculation of the effective lending rate (LR) to include the following variables: 𝐿𝑅 represents the effective lending rate, 𝐼 the interest rate income, 𝐺𝐿𝑃 the average gross
loan portfolio, 𝑊𝑂𝑅 the write-off ratio and 𝐹𝐶 the total fees and commissions income on loan portfolio. Variables in italics are the variables provided by the MIX Market while 𝐼 the interest rate income is to be calculated based on the formula as below:
𝐼 = 𝑌𝐼𝐸𝐿𝐷 ∗ 𝐺𝐿𝑃 − 𝐹𝐶 (17)
where 𝑌𝐼𝐸𝐿𝐷 is the variable yield on gross portfolio (nominal) provided by the MIX Market. Dorfleitner et al. (2013) consequently finalized a formula to calculate the effective lending rate as:
𝐿𝑅 = 𝐼
𝐺𝐿𝑃 ∗ (1 − 𝑊𝑂𝑅)+ 𝐹𝐶
𝐺𝐿𝑃 (18)
The lending rate calculated with equation (18) is claimed to be a more appropriate measurement for the actual interest rates the borrowers paid to receive microloans (Dorfleitner et al. 2013). The reason is that this lending rate proxy includes those issued but not repaid microloans so as to scales up the underestimated interest rate proxy yield on gross portfolio. Therefore, the lending rate is more appropriate to use to analyze the demand side.
5.4 Summary of regression variables
The ultimate model has been selected based on the procedures discussed in the methodology section. As a result of these selection procedures, the observations reduce dramatically and the summary statistics of the regression variables is presented in Table 2.The model selected is a time fixed effects (FE) model using one-year lag variables, the equation is as below:
lending rate is calculated with interest rate income subtracting total fees and
commissions, it implies that some microcredit borrowers paid interest rates that are under the costs of their loans and the lenders bear these costs. MFI’s profit margin ranges from minus 354962% to 7209813% with a mean of 150.9%. MIX Market describes profit margin as MFIs net operating income divided by financial revenue. It represents MFIs profitability from earnings on their core business operations and determines what percentage of revenues generated from those activities remains after all costs and expenses. Risk coverage ratio that measures how prepared a MFI is to cover credit loans loses ranges from minus 633.1% to 472866166% with a mean of 144564.6%. Finally, female borrowers composed from zero to 100% with a mean of 64.9%. Therefore, majority of microcredit borrowers were female.
Table 2. Regression variables descriptive statistics
Variable Observation Mean Std. Dev. Min Max
Aggregate demand1 13,285 52.3 274 -0.015 13.100
Effective lending rate 8,522 .342 .198 -1.352 2.985
Profit margin 16,057 1.509 638.397 -35495.62 72098.13
Risk coverage 12,597 1445.646 80290.8 -6.331 4728662
Percentage of female borrowers 13,605 .649 .279 0 1.00
1 Variable is measured in million USD
6. Results interpretations
6.1 Summary regressions results
Comparing to matured MFIs, young MFIs have lower loan demand while the new MFIs have the lowest demand from borrowers. Besides, comparing to the MFIs locating in Sub-Saharan Africa, those in two regions, Eastern Europe and Center
Asia, and East Asia and the Pacific, face a lower demand while MFIs in the other
three regions all have higher loan demand. The region North America where data are only available from one single MFI is eliminated completely due to missing
observations by the model. For-profit MDIs in general also have higher loan demand comparing to nonprofit peers. However, result for Eastern Europe and Center Asia is not significant at all three significance levels, and results for East Asia and the
Pacific, Middle East and North Africa, and South Asia are only significant at 10%
level. All other results are significant at 1% level. Observations in the pooled OLS regression with control variables reduce from 6,770 to 4,183 but the R-square has increased from 4.6% to 20.9%. In a short conclusion, the pooled OLS seems to support the argument that microcredit borrowers are rather elastic to interest rate changes.
The outcomes of the fixed effects model suggested that every 0.01% interest rate increase will reduce loan demand by 0.7%, this is significant at 1% level, likewise every 0.01% increase in MFIs profit margin will increase demand by 0.04% and it is significant at 10% level. Also every 0.01% reduction in female borrowers over time increase the aggregate demands by 0.1% and the result is significant at 5% level, while the result for MFIs risk coverage is not significant at all three levels. Similar to the results from the pooled OLS, comparing to matured MFIs the young ones have lower and the new MFIs have the lowest demand from borrowers. All these results under the fixed effects model are based on the variance among MFIs that changes across time periods because the model eliminates all time invariant variables such as profit status and region. When the random effects model is applied, every 0.01% interest rate increase reduce loan demand by 0.72%, every 0.01% increase in MFIs profit margin increase demand by 0.05% and they are both significant at 1% level. Besides, exactly the same as the FE model, every 0.01% reduction in female borrowers increase the aggregate loans demands by 0.1% and it is significant at 5% level. The result regarding to MFIs risk coverage stays insignificant at all three levels. As for region, profit status and age effects, the results from the random effects model are rather similar to the ones from the pooled OLS except for Middle East and North
Africa whose result is not significant under the random effects model. The results
from the fixed effects and random effects models both support an elastic demand to price, but considering the coefficients of the RE model are not consistent under the
Hausman test, results from the FE model is seen to be more plausible even though the
Table 3. Correlation coefficient explanatory variables
01 02 03 04 05 06 07 08 09 10 11 12
01.Effective lending rate1 1.00
02.Profit margin1 -0.168 1.00
03.Risk coverage2 -0.057 0.100 1.00
04.Percentage of female
borrowers2 -0.118 -0.021 0.027 1.00
05.East Asia and the Pacific -0.031 0.062 -0.064 -0.029 1.00 06.Eastern Europe and Central
Asia 0.081 -0.026 -0.036 0.027 0.451 1.00
07.Latin America and the
Caribbean -0.024 0.021 -0.024 -0.066
0.554* 0.557* 1.00
08.Middle East and North Africa -0.003 -0.064 -0.075 -0.009 0.326 0.352 0.422 1.00
09.South Asia 0.301 0.031 -0.143 -0.175 0.470 0.495 0.577* 0.348 1.00
10.Age new -0.152 -0.009 0.015 -0.010 -0.025 -0.110 0.005 -0.042 -0.101 1.00
11.Age young -0.177 -0.016 -0.012 0.124 -0.063 -0.114 -0.018 -0.056 -0.146 0.438 1.00
12.Profit -0.213 0.076 -0.025 0.064 -0.075 0.023 0.044 0.122 -0.114 -0.054 -0.032 1.00 *Value higher than 0.50
The result of the time fixed effects model as well suggests that borrowers are elastic responding to interest rate. Nevertheless, result showing that every 0.01% Table 4. Regressions results summary statistics
OLS Pooled OLS FE model RE model TFE model Effective lending rate1 -0.803*** -0.878*** -0.695*** -0.715*** -0.279***
(-18.14) (-15.31) (-13.38) (-15.17) (-7.94)
Profit margin1 0.128*** 0.042** 0.048*** 0.082***
(5.12) (2.97) (3.50) (8.82)
Risk coverage2 0.157*** -0.005 0.012 -0.007
(7.95) (-0.37) (0.97) (-0.84) Percentage of female borrowers2 -0.181*** -0.098* -0.098* 0.023
(-3.39) (-2.12) (-2.36) (0.75)
Sub-Saharan Africa 0 0 0 0
(.) (.) (.) (.)
East Asia and the Pacific -0.300** 0 -0.492** 0
(-2.63) (.) (-2.60) (.)
Eastern Europe and Central Asia -0.101 0 -0.073 0
(-0.97) (.) (-0.42) (.)
Latin America and the Caribbean 0.380*** 0 0.514*** 0
(4.20) (.) (3.33) (.)
Middle East and North Africa 0.409** 0 0.435 0
(2.97) (.) (1.68) (.) South Asia 0.285** 0 0.271 0 (2.64) (.) (1.56) (.) Age mature 0 0 0 0 (.) (.) (.) (.) Age new -1.330*** -1.546*** -1.465*** -0.387*** (-9.32) (-18.85) (-18.67) (-6.83) Age young -0.890*** -0.879*** -0.854*** -0.095** (-12.46) (-21.37) (-21.42) (-3.22) Nonprofit 0 0 0 0 (.) (.) (.) (.) Profit 1.116*** 0 1.112*** 0 (21.69) (.) (11.32) (.) Constant 15.133*** 14.986*** 15.775*** 14.693*** 17.149*** (262.08) (125.09) (219.84) (95.75) (120.90) R-Square3 Observations 0.046 6770 0.209 4183 0.218 4183 0.188 4183 0.668 4183 T-statistics in parentheses. Symbol * P < 0.05, ** P < 0.01, *** P < 0.001. 1Value lag of one year in natural logarithm. 2Value in natural logarithm.
increase of interest rate only reduce loan demand by 0.28%, which is more than a half smaller comparing to the one of pooled OLS model. Besides, a 0.01% increase in profit margin increase the demand for 0.08%. Under the time fixed effects model, new MFIs existing for less than four years also have lower loan demand comparing to matured MFIs that are older than eight years, while the magnitude is much smaller than under the pooled OLS model. However, young MFIs under the time FE model are reported to have rather similar demand as their mature peers. All these results are significant at 1% level except for young MFIs, which is only significant at 10% level. Conversely, the results regarding to risk coverage and percentage of female
borrowers are not significant. Since the result under a Lagrange multiplier test
suggested that there is variance across time periods and the majority of time dummies in the fixed effects model are significantly different from zero, the results under the time fixed effects model are superior to the ones under the entity fixed effects model. Besides the Lagrange multiplier test, a much higher within R-square value in the TFE model is a further support.
In conclusion, results show that microcredit borrowers are indeed elastic to interest rate changes and therefore, H1 is supported. As for H4, which expects both
profit margin and risk coverage to have a positive effect while percentage of female borrowers to have a negative effect on aggregate credit demand, the results are mixed.
Profit margin of a MFI seems to have a positive effect on their loan demand while risk coverage not. Interestingly, even though all the results for risk coverage are not significant, it carries different signs under different models. A possible explanation will be that this variable has no explanatory power in these regression models, but the actual causes of is result require further researches. Likewise, when it comes to
percentage of female borrowers, the result under the TFE model is not significant at
all three levels and the coefficient also carries a different sign comparing to the previous three models. A time fixed effects model shelters the results from picking up influence of aggregate time-series trends that are irrelevant to causal relationships. For instance, there could be an aggregate trend with both higher microcredit demand and less female borrowing, but this trend is not a proof to a causal relationship. The change of coefficient signs indicates that the fixed effects model falling to control for year effects suffers from bias by omitting time invariant variables. The result for
percentage of female borrowers is insignificant nevertheless it is unwise to conclude
for an irrelevance. Rather, it could imply possible omitted variables on MFI-level if only considering the insignificant coefficient itself, and due to the fact that number of female borrowers is closed to being a time-invariant variable, the insignificant result is not unexpected under a fixed effects model (Dorfleitner et.al, 2013). Seeing from the data, it is indeed the case, since the average mean of the changes of percentage of
reveal the true relationships, but this result seems to have no support at aggregate level to the finding of Agier and Szafarz (2013) stating that female borrowers face a glass ceiling and receive smaller amounts of loan.
6.2 Summary group comparisons results
In this section, results for equation (7) to equation (15) to test H2, H3 and H5 are reported. All outcomes are included in Table 5. The results in Table 5 suggest that microcredit borrowers have different price elasticities responding to interest rate increase and decrease. Borrowers’ elasticity to interest rate increase is 0.18% higher than to interest rate decrease, for every 0.01% interest rate change. However, a concern is that these results are not significant at all three significance levels and therefore, H2 is not supported strictly speaking. Furthermore, clients of for-profit MFIs are more sensitive to interest changes than those of nonprofit MFIs. Results show a 0.21% higher price elasticity of for-profit MFIs clients comparatively and they are significant at 5% level thus H3 is strongly supported. Finally, H5 is also strongly supported at 5% level. Outcomes show that MFIs who have higher percentage of female borrowers have nearly to zero price elasticity while their peers who are serving comparatively lower percentage of female clients face a price elasticity as 0.24%. In summary, microcredit borrowers respond differently towards interest rate price increase and price decrease, it is an implication for a kinked demand curve. A deeper insight into the results suggests an insensitive reaction towards interest rate decrease based on two considerations. Firstly, the elasticity responding to interest rate increase is very closed to the one estimated under the time fixed effects model. Moreover, the result regarding to responds to price reduction is not significant. It is therefore not unreasonable to infer an insensitivity of borrowers to interest rate decrease. Seeing the results for H3 and H5, female borrowers and clients of nonprofit MFIs have nearly zero or very closed to zero elastic to price changes. This finding buttresses the view promoted by influenced advocates saying that the poor individuals are insensitive to price (Dehejia et al., 2012), and they will borrower at any rate. From the opposite aspect, male borrowers and clients who are credit worthy enough to be accepted to lend by profit MFIs care more about the prices they have to pay for loans and will borrower less if interest rate increases. As a result, creditworthiness and whether there are available alternatives affect borrowers elasticity to loans price largely. If
translating to the supply side of story, risk and costs for lending and market competition have big impact on MFIs’ decisions to the rates they charge.
Several robustness checks are applied by substituting the one-year lag lending rate variable with two other proxies as yield on gross portfolio nominal and yield on gross
portfolio adjusted with inflation respectively. Moreover, tests by removing one
Table 5. Group comparisons results summary statistics
Positive/negative effective lending rate changes
Nonprofit/profit MFIs
High/low percentage of female borrowers
Effective lending rate1 -0.166*** -0.0787* -0.236***
(-4.50) (-2.29) (-6.13)
Negative effective lending rate change
dummy=0 0
(.) Positive effective lending rate change
dummy=1 -0.0664
(-1.64) Effective lending rate1 * Effective lending
rate change dummy -0.0536
(-1.67)
Nonprofit dummy=03 0
(.)
Profit dummy=13 0
(.)
Effective lending rate1 * nonprofit dummy 0
(.)
Effective lending rate1 * profit dummy -0.210***
(-3.99)
Low percentage of female borrowers=0 0
(.)
High percentage of female borrowers=1 0.145*
(2.25) Effective lending rate1 * Low percentage of
female borrowers 0
(.) Effective lending rate1 * High percentage of
female borrowers 0.100* (2.11) Constant 16.99*** 16.84*** 16.62*** (127.53) (130.49) (118.94) R-Square2 0.503 0.552 0.556 Observations 4551 6640 5844 T-statistics in parentheses. Symbol * P < 0.05, ** P < 0.01, *** P < 0.001. 1Value lag of one year in natural logarithm. 2Value is within R2.
variable the risk coverage are also performed with those two other proxies
representing the interest rate. Summary of these robustness checks are presented in Appendix 1. Outcomes of the robustness checks confirm the findings. Another issues addresses to the possible endogeneity in the model. Even though the model has
applied the lagged values to mitigate the problem, it is not completely eliminated. One way to solve the problem is to apply instrumental variables techniques and it involves selected valid instrument variables (Wooldridge, 2012), which is far from the attempt of this paper.
6.4 Limitations and further discussion
One serious concern is to the fact that the unbalanced dataset clearly suffers from selection bias due to missing variables not at random. Even if the fixed effects model is thought to be limiting the selection bias (Mummolo and Peterson, 2018), an attempt following Verbeek and Nijman (1992) to select an unbiased subsample from the unbalanced panel failed. This results in a main limitation of this paper. Since the reasons for missing observations are hard to identify, it is uncertain whether those missing observations might change the results. Furthermore, a second limitation will be on the possibility of endogeneity considering the methodology design in this research. The interpretation of findings is based on an assumption that MFIs determine the interest rates considering merely the supply side, which means the demand side has no power to affect the decisions made from the supply side. This assumption is rather strong especially when the existence of reverse causality cannot be certainly denied. Even though alternative solutions are feasible, due to amounts of technical difficulties it is out of the attempt in this paper. Additionally, another fact to be made cleared, is the difficulty to separate supply effect and demand effect in the model. It is quite appealing to conclude all results to be demand driven, but in reality the possibility is rather vague due to the nature of the studied data. The data describes the so call “realized demand” but not the real demand of how much the poor
individual actually want to borrow, which is to be captured by survey data. The term “realized demand” mentioned is stating for the loan demand approved and issued by MFIs and this amount excludes the demand from un-credit-worthy borrowers, based on another assumption that MFIs only consider borrowers creditworthiness.
