India and Ghana.
Steffen Eriksen
∗Supervisor: prof. dr. B.W. Lensink
Department of Economics, Econometrics and Finance
Submitted 8th July 2014
Abstract
Using data from two field studies in India and Ghana, a cross country com-parison of the impact of microcredit is conducted. The impact of microcredit is evaluated through seven main outcome variables. Five of which are represented by indices, constructed using a summary indexation from a large set of individ-ual outcomes. Propensity score matching is combined with a double difference model to assess the impact before and after the provision of microcredit. In India five out of seven outcome variables show a positive impact, indicating a generally encouraging effect of microcredit. Conversely, only one outcome vari-able reveals a positive effect in Ghana, and the impact is negative for two of the outcomes, thus suggesting that there are differences between countries. We argue that the main difference between the two set of results is due to loan use.
Keywords: Microcredit, Impact evaluation, Summary indexation, Cross Coun-try Comparison
JEL classifications: C43, D14, O14
∗
1
Introduction
Ever since the first microloan was granted, the promise of microcredit has been the root of intense debate about if, when, and where it is actually an effective tool against poverty in developing countries. The access to financial services have been expanded in many developing countries as a consequence of microcredit (Humle et al. (2009)).1 However, the conclusions drawn about the impact of microcredit appears to be somewhat mixed. Snodgrass and Sebstad (2002) find no increase of the average income when comparing a treatment to a control group in Zimbabwe. Similarly, no short-term increase was found on average consumption, in a randomized control trial in India conducted by Banerjee et al. (2014). Another study discovered that an expansion of credit increases the income of the household and likewise employment for poor households in Mexico (Buhn et al. (2013)). Also McKenzie and Woodruff (2009) found that by just giving small loans, the increase in income of the household can be substantial. In addition to improving the income of poor households, microcredit is also considered to be a way in which to improve the female position within the household. Abdul et al. (2014) found a positive impact of woman empowerment in a study conducted in Malaysia. On a similar note, Angelucci et al. (2013) reveals evidence that microcredit does empower women. However it did not improve other indicators, such as income, expenditures or education.
Both theoretically and empirically a tension is present. On the one hand the ex-pansion of microcredit aims to increase access to funds. On the other hand, some express that this expansion of microcredit might lead to overborrowing and put poor households into to a debt spiral, making them worse off than before credit was pro-vided. This aspect has drawn a lot of attention in many developing countries, and specially due to the microfinance crisis in India in 2010.
Microcredit is seen as a key development tool and has continued to show growth in both Asian and sub-Saharan countries, which are the two areas of interest in this study. The impact on many different indicators are considered and as is shown by Rooyen et al. (2012), the evidence shows that microcredit does good as well as harm, to the living conditions of the poor. They evaluate the impact of microcredit in
sub-1
Saharan countries on several indicators including; income, savings, expenditures, asset accumulation, housing improvements, as well as a vast number of non-financial indicators. Chemin (2008) uses similar indicators to investigate the cost and benefits of microcredit from Bangladesh using matching. The results obtained are mainly positive, but the magnitude is less than previously thought. These and many other studies point to that the main objective of microcredit is to reduce poverty as mea-sured by income or expenditures.2
A lot of other benefits can be derived from microcredit. Prema (2009) considered that being a member of MFI not only gives rise to economics empowerment of its members, but also that members experienced a significant rise in self-esteem, self-respect as well as leadership qualities. More specifically, Chowdhury and Chowdhury (2011) analyze the effect of women empowerment through the provision of microcredit. As the vast majority of microcredit clients are women, the empowerment of women has long been associated with microcredit. They find that participating women were more empowerment than their non-participating counterparts. However, they argue that the empowerment is only a short term effect, but will affect long run indicators such as the household income and expenditures.
Financial literacy is also seen as an important tool to improve the long run effects of microcredit. By providing training to the clients, their financial behavior can be changed in such a way that long run effects are affected in a positive way. As shown by Sayinzoga et al. (2014), providing financial literacy training, increased the par-ticipants financial literacy, and changed their savings and borrowing behavior. They support the claim that microcredit programs may benefit from these training mod-ules to enhance their development impacts, but they fail to find any spillover effects from farmers, who were given the training, to their peers. In general there has been a lot of focus on building up human capital among the clients. In a paper by Karlan and Valdivia (2011) they measure the marginal effect of adding business training in Peru using a randomized control trial. They find little or no evidence in changes in key business outcomes, but business knowledge improvements are observed, along with the retention rate of the clients.
The findings shown above, show that there is still room to rethink the basic strategies when conducting research in the field of microfinance. Evidence points towards
considerable heterogeneity in treatment results, suggesting that impact will differ by area. In light of these findings, this study will compare two MFIs in two different countries and try to provide new evidence for the heterogeneity in the impact of microcredit. The two MFIs that will be studied are located in India and Ghana. The first is a typical Asian country, whereas the second is a typical sub-Saharan country.
The main problem when measuring impacts is how to deal with endogeneity issues. Often in impact evaluation studies, like this one, a Randomized Control Trial (RTC) is conducted, as this is the first best solution to deal with endogeneity issues. How-ever, when a RCT is not available, alternative measures have to be used to measure the impact of interest. Endogeneity leads to biased and inconsistent estimates of all parameters in the regression even if only one of these parameters suffer from endo-geneity, which then leads the researcher to draw incorrect conclusions.3 Endogeneity can be caused for several reasons, including: relevant omitted variables, measurement error, simultaneity, self-selection and serial correlation in the presence of a lagged dependent variable. However, as shown by Ebbes (2004) the four latter cases is a special case of the first. As an alternative solution to the endogeneity problem, this study uses Propensity Score Matching (PSM) combined with a Double Difference (DD) model to measure the impact before and after the provision of microcredit. In cooperation with the Dutch commercial bank ING and Netherlands Platform for Microfinance (NPM)4, this study analyses the impact of microcredit using two field studies conducted in India and Ghana. The impact of microcredit will be studied using seven main outcome variables. Each of these outcomes analyzes a way in which microcredit affects the lives of the clients and their families. The seven main outcome variables are: The education level of the oldest son and daughter in the household, household assets, housing, women empowerment, household expenditures and financial literacy. The last five of these outcome variables are constructed as indices, each combining multiple independent outcomes into a summary index. The results are mixed. For the India sample, an overall positive impact of microcredit
3
More formally, consider the following regression equation: y = β1x1+ β2x2+ i. If either x1 or
x2 were to be correlated with i (Cov(x1, i) 6= 0) that parameters would be endogenous and give
biased and inconsistent estimates of all parameters.
4
is found with five out of seven outcome variables, showing a significant positive im-pact. The biggest effects are found in financial literacy and household expenditures, which seem to be the main channels from which the poor households experience the effects of microcredit. The results from the Ghana sample are more moderate. Only the household expenditures shows a significant increase, which is smaller in magni-tude than the one found in India. A summary effect decrease is found in both assets and women empowerment. By looking into the individual components of the asset it is found that the negative impact for assets is driven by a decrease in livestock. It could be that some of the livestock are sold off to finance the increase of household expenditures, or that it simply takes more time in Ghana for the effect of microcredit to emerge. Put together, the results differ rather substantially from the two coun-tries, suggesting a heterogeneous impact of microcredit across boarders. The main explanation for these differences argued here, is due to difference in loan use. A list experiment reveals that larger percentage of borrowers in Ghana uses their loans for consumption purposes rather than a productive investment in the household business and/or farming activity.
In the next section, a theory of change will be provided, suggesting why microcredit may, (or may not,) have an effect on the seven outcome variables and ultimately the overall wellbeing of the poor households. Section 3 follows with a description of the methodology used in this study. Section 4 explains the method used to construct the indices representing five of the outcome variables. Section 5 presents the sampling method and provides a detailed description of the data used in this study, paying special attention to the possible problems with the data. Section 6 will provide the results of the study. Section 7 will compare the results between the two countries and try to explain the differences. Section 8 concludes.
2
Theory of Change
level of the oldest son and daughter of the household. All of these outcomes, with the exception of financial literacy, can be seen as a long term effect of the provi-sion of microcredit. Financial literacy, can be seen as a more intermediate effect of microcredit. The reason for the inclusion of these different indices/variables will be explained below.
When receiving a microcredit loan, the proceedings of the loan are often spent on education of not only the loan recipient, but more specially on the children of the household. Sending the children to school will develop their skills. This will then enhance their future possibilities of receiving a better job and hence make them more likely to escape poverty. In this study, the variables used to measure the effect of the provision of microcredit on education, are the education level of the oldest son and daughter in the household.
Loans are often used to build an enterprise’s or a household’s inventory of physical as well as productive assets. Productive assets can be used to generate income without actually selling the asset itself. Other physical assets can only generate cash when sold off. Many studies have indicated that asset accumulation increases with the cumulative value of the loans taken by the household. This is in turn generally associated with the duration of participation in the microcredit program.5 Building up assets, builds up the wealth of the household and therefore reduce poverty. This effect of asset accumulation is what is intended to be captured with the asset index. The housing index can also be linked to above discussion of household assets. When the household builds up assets and hence wealth, this wealth is often reflected in improvements in the household. In this study the Housing index incorporates im-provements of the household in the form of; Number of rooms, roof type, sanitation facilities and electricity connection. In particular, the roof type of the household gives a good indication about the poverty level of the household.6 As the size of the microcredit loans is not sufficient for covering the cost of roofing, the finances used for roofing (whether its cash or bounded in assets), will have to be accumulated over time.
5
Sebstad and Cohen (2000) review more than ten microcredit impact studies that looked for changes in household assets. They note that a certain duration of program participation is needed before impact on household/enterprise assets will occur.
6
From the very beginning of microfinance, the majority of microfinance clients have been women, and the share of women served by MFIs have only increased over time. The reasons for MFIs to specifically target women are many. Firstly, as MFIs focus on informal micro businesses, and these businesses, are mostly run by females, it is only natural that the female shares are large. Secondly, women are more risk averse and therefore choose more conservative investments, which leads to a better and more steady payment pattern. Thirdly, It is the case that women tend to be housebound in the discussed cultures. This reduced mobility makes it cheaper for the bank to monitor women. Lastly, women often do not have other alternatives due to credit constraints, so they self-select them into microfinance. This leads to a reduction of moral hazard, as women are less likely to shirk.7 For these reasons, the empowerment of women is an important indicator for the effects of microcredit. When women are more empowered and hence have more influence on the decision making in the household, it might enhance the performance of the household business. Additionally, other measures might also be positively affected by an increase in women’s empowerment. Women’s decisions tend to be biased in favor of expenditures in the home, thus more money will be spent on, for instance, education of the children.
The index for household expenditures provides the most important measure for poverty reduction considered in this study. In these types of studies, household expenditures are often used as a proxy for income, as the concept of income can be hard to grasp for many poor households (Armendáriz and Morduch (2010)). The measure of household income can therefore come with a great deal of error. It can therefore be beneficial to look at household expenditures or a combination of both. The index for household expenditures considered here, is a combination of both an expenditure measure and an income measure of the household. An increase in the expenditures of the household (and thus income) can be directly related to a reduc-tion in poverty, and is therefore an important indicator of the overall wellbeing of the household.
Financial literacy is associated with the client who takes the responsibility to inform herself of the products she purchases and to understand the contracts she signs. How-ever, the foundation is to have the knowledge and competence to make an informed
7
financial decision. For people in the developing world, this mostly refer to basic concepts of safe and secure savings, budgeting and wise borrowing. Knowing this can help the borrowers to make better decisions with regards to their finances and help them improve upon the measures mentioned above, such as assets and house-hold expenditures. Hence, financial literacy does not directly lead to a reduction of poverty, but rather serves as an intermediate effect improving the other measures from which poverty can be reduced. Many programs strive to change the financial behavior of the clients and emerging evidence from many of these programs does point to changing behavior.8
A theory of change can be modeled in various ways, for example by using a results chain. A results chain establish a pathway through which impacts are achieved. Beginning with the resources available at the start and ending with the long-term goals of the project, it lays out a causal logic from the very start of the project. Figure 1, illustrates the results chain as used in this project. Each of the elements contained in the chain are shown below.
Figure 1: Results chain
Inputs: This includes resources which are at the disposal for the project, including staff and budget. In the context of this project, it includes -Trained staff from Basix and OI
Activities: Actions taken or work performed in the process of converting the inputs into outputs. -Provide credit to poor households -Provide orientation training (covers terms and conditions of the products offered by the MFI and enhancing their knowledge on good repayment & borrowing practices, benefits of savings and management of household budget)
Outputs: This contains the goods and/or services that is produces by the activities of the project. - Clients who are able to invest in the household business/farm-ing activity -’Trained’ clients who are aware of their financial status and the conditions of the loan
8For a detailed overview of the current evidence of the changing financial behavior and its effect
Outcomes: This covers the results which are likely to be achieved when the target population makes use of the outcomes of the project. As an intermediate effect, the client should have an improved financial literacy score. In the medium to long run, the six remaining outcomes should show positive effects. The impact of the provision of microcredit will be tested through each of these outcome variables. In addition the individual components of the indices will be tested as well.
Final Outcomes: The final goals of the project, which are typically achieved in the long run. The long term goal for this project is a general reduction in poverty and improved living conditions of poor households.
For several reasons, these predictions may be overly optimistic. The following "risks" should be considered and kept in mind.
1. The orientation training provided by the two MFIs may be carried out by unqualified trainers. If this happens, then even if the material is good, the impact of the orientation training will be lower, and the clients might not have achieved the intended level of awareness.
2. The loan could be used for other purposes than otherwise stated. It might be used for household consumption rather than a productive investment in the household business and/or farming activity. A list experiment was implement to test for this risk.
3. Instead of spending the proceedings of the loan on education for the children, the expansion of the household business and/or faming activity may result in more children engaging in income generating activities for the household. It might be that the children are needed to help out with the household business and/or farming activity, rather than going to school. Separate regressions were run to check for an increase in children who engage in income generating activities for the household (See appendix C).
3
Methodology
To analyze the impact of microcredit on the education indicators, the five indices as well as its individual components, PSM combined with DD was used, which should provide unbiased estimates of the treatment on the outcome variables of interest. Additional covariates are included to control for various characteristics of the house-holds. This is necessary to obtain unbiased measures of the treatment effects. Having a randomized control trial would theoretically have made it unnecessary to include controls as the provision of microcredit would have been orthogonal to household characteristics. Including additional covariates are thus necessary to provide unbi-ased estimates.
When measuring the impact of an intervention like providing people microfinance, the most difficult task is to separate the causal role of microfinance from other causes. The observed changes in an outcome variable, cannot solely be attributed to microfinance. There can many other factors driving the results. Respondents in the treatment group could for instance be wealthier than the control group, when the program started. This and other factors can make some people more likely to self-select them into a program, thus causing self-selection bias. This and programme placement bias frequently occur in the context of microcredit studies. Programme placement bias refers to when there are difficulties in finding a location at which the control group’s physical, economic and social environment matches that of the treatment group.9 The PSM method is developed to draw causal inferences in non-experimental studies as is the case here.
The objective is to observe changes that could be attributed to participation in a microfinance program and how a respondent would have done if not participating in the program (Armendáriz and Morduch (2010)). However, answering this question becomes difficult when the microcredit loans are not assigned randomly, as was the case in this study. PSM can then be used to construct an observational equivalent to a randomized experiment, by making the treatment and control group similar in
9
terms of observable characteristics and possibly reduce selection bias.10 The propen-sity score is defined as the conditional probability that an individual is assigned to a treatment group (Rosenbaum and Rubin (1983)). By ’matching’ on observable characteristics a control which is similar to the treatment group can be constructed, thus reducing programme placement bias. The basic idea is to construct individuals from one group who are observationally similar to the individuals in the other group, in terms of characteristics which is not affect by the intervention. Often this include baseline characteristics, as they are clearly not affected by the intervention. Note that PSM assumes selection on observables and hence unobservable characteristics can still affect the result, and thus still bias the results.
The propensity is then estimated using a logit model, controlling for a set of matching variables. After estimating the propensity score and ensured that there is enough common support11, the matching itself can be done, from which the treatment effect can be obtained.12
The next step is to apply a DD model. The DD method relies on comparison of respondents from the two groups, treatment and control, before and after the inter-vention. It essentially compares the treatment and control group in terms of outcome changes over time relative to the outcomes observed at the baseline. The DD es-timator allows for unobserved heterogeneity such as differences in innate ability or personality across treatment and control units. DD assumes this unobserved hetero-geneity is time invariant, so the bias cancels out through differencing. Econometri-cally speaking the double difference estimator is given by the following expression:
Yijt= α + β1P ostt+ β2DjT + β3P osttDjT + β4X + ijt (1)
Where Yijtdenotes an outcome variable for respondent i in group j at time t, DTj is a dummy variable equal to one if the respondent belonged to the treatment group, P ostt is a binary variable that takes the value one if the observation corresponds to
10
PSM is not a miracle cure for controlling selection bias in impact assessments. The debate on the appropriateness of evaluation methods currently used to account for biases is far from over. One point is clear, that, accounting for selection bias should be one of the prerequisites for future impact studies.
11
To ensure that there was enough common support, treated individuals with a propensity score lying outside the range of propensity scores for respondents in the control group, were discarded.
12For the matching itself, a local linear regression matching using a normal kernel distribution
the post-treatment time period, X is a set of controls, and ijtdenotes the error term.
Then, β3 in (1) is the treatment estimate of the intervention’s impact on outcome Y. That is, β3 measures the difference between the treatment and control group in
the growth of outcome Y. It is an unbiased estimate of the average impact of being assigned to the treatment group on the dependent variable Y. In some of the models, where no recall information is available, the estimate for equation (1) only includes endline values.
The double difference approach described in (1) is equivalent to a fixed effect model, with time fixed effects as we have a balanced panel. Hence, the DD model used here control for household fixed effects.
Some of the outcome variables in the study are binary. In these cases, a linear probability model (LPM) is estimated and the marginal effect of DtC for the impact of microcredit on outcome Y is reported.13 Additionally, estimates of β3 without additional controls are reported in appendix C for the case of the summary index results. Results for the individual outcomes excluding control variables are omitted. For the India sample the standard errors are clustered at the Village level, while the standard errors are clustered at the branch level for the Ghana sample.14 It is important to cluster the standard errors, as the data might be subject to intraclass correlation, that is, households in the same village/branch are likely to be more similar on a wide variety of measures than households that are not part of the villages. The higher intraclass correlation, the less unique information each household provides. This has to be taken into account when running the regressions by inflating the standard errors.
By applying a combination of PSM and DD selection bias can be controlled, pro-gramme placement bias and time invariant heterogeneity. Only applying PSM would still leave a potential bias due to unobserved heterogeneity. Thus a respondents who
13In recent literature, Puhani (2012) shows that in a nonlinear difference-in-difference, such as
the one used in this study, the cross difference is not equal to the treatment effect. Instead the treatment effect comes from the cross derivative (or cross difference) of the conditional expectation of the observed outcome minus the cross derivative of the conditional expectation of the potential outcome without treatment. Although this calculation of the treatment effect is appealing, this study will follow common practice in the field and report the estimate of β3 in the case of a LPM.
14
have a higher innate ability would be more likely to get a larger impact of microcredit than respondents with a lower innate ability. In a setting where the microcredit loans have not been assigned randomly, PSM can then be used to construct an observa-tional equivalent to a randomized experiment, by making the treatment and control group similar in terms of observable characteristics. It ensures that respondents in the control and treatment group who differ to much are not taken into account at all, as they are dropped from the common support. So by matching on observable char-acteristics PSM can control for self-selection as well as programme placement bias, for observations on the common support. On the other hand, applying DD with control can also control for these biases, however, it does not dropped observations who differ to much like PSM.
An important thing to note. Although applying a combination of PSM and DD can solve many potential biases in the data, there still might be some problems left. As PSM assumes selection on observables, bias can still be caused by selection on unobservables, which is the drawback of the PSM method. Adding DD to PSM can help picking up the time invariant heterogeneity, but bias can still remain due to unobservables.
4
Summary Indices
The provision of microcredit is evaluated upon a total of 36 outcomes. 2 of these relate to education, 20 to household assets, 4 to housing, 5 to women empowerment, 2 on household finances and 3 on financial literacy. However, testing multiple out-comes using (1) independently increases the probability of committing at type I error (rejecting a true null hypothesis) for at least one of the outcomes in question (Duflo et al. (2007)). To overcome this problem, the variables can be grouped into indices. A total of five indices is constructed in this paper, each representing one of the outcome variables explained in section 2: Assets, housing, women empowerment, household expenditures and financial literacy. The two outcome variables for education are not represented by indices as they already consist of only one variable.15
15
There exist various ways of constructing these summary indices, and the method chosen for this paper is the one by Anderson (2008). The method combines a given number of variables into one summary index, where each of the standardized com-ponents are weighted using the inverse covariance matrix of the standardize out-comes for each of the groups (see below). Anderson’s method can be said to be an extension of the method proposed by Kling et al. (2007), whose summary index is an unweighted average of its standardized components. Mainly for this reason is Anderson’s method preferred. By weighting the inputs of the index by the in-verse covariance matrix, a new input which has a high covariance with some of the other components of the index, will receive a low weight as it provides little new information. Likewise, low covariance will provide a higher weight as a lot of new information is added to the index. The coefficients of the summary indices should then be regarded as effect sizes when interpreting the results later in the analysis. The procedure of Anderson (2008) can be summarized into the following 5 steps:
1. All the outcomes in question have to be coded such that the ’better’ outcome is in the positive direction.
2. Standardize each of the outcomes by subtracting the mean from the baseline control group and then divide by the standard deviation of the baseline control group. Doing this normalizes the outcomes so they are on a comparable scale. 3. Define the number of groups you wish to use and assign each outcome to one
of these groups.
4. Construct the summary index variable, which is a weighted average of each of its inputs. When constructing this summary index, weight each of its inputs by the inverse of the covariance matrix of the standardize outcomes for each of the groups.
5. Regress the new summary index using the chosen specification (DD in this case) to estimate the effect of microcredit on the group. A standard t-test assess the significant of the coefficient, and thus does not require any correction.
After covered the indexation technique, the process to obtain the estimates can then be summarized into the following steps:
1. Drop from the initial sample all respondents who had previous microfinance. This is done as the ’baseline’ characteristics of these individuals are affected by the intervention as they have already received microcredit at the recall moment. Thus, this will make matching on baseline characteristics and hence PSM problematic.
2. Match observations from the treatment and control group using PSM and drop all observation who are not on the common support.
3. Generate summary indices using the approach described by Anderson (2008) above. Skip this step if examining individual results.
4. Estimate the DD model using the indices generated in step 3. For individual results estimate the DD model using the outcomes from step 2.
5
Sampling and Data
This section explains the data collection process and outlines the criteria for sample selection. I also summarize the data and check the balance of the sample in terms of the observable characteristics chosen, to see if the treatment and control group are similar at the baseline (recall moment).
5.1 Data Sampling
Limited (Basix16) branches operating in the northern district of Bihar. Originally, the sample was to be drawn from northern and southern India, however, it was dif-ficult for Basix to provide information on active borrowers in southern India as its operations were heavily affected by the microfinance crises. For Ghana, Opportunity International Savings and Loans Limited (OI17) participated in the study and the data was collected from ten of their branches, spread over two regions (Ashanti and Brong-Ahafo). It was initially intended to only sample borrowers from rural areas, but urban clients were included as well, as it was later discovered that obtaining the sample from only rural borrowers would be hard. Basix and OI were asked to partici-pate in the study, based on their membership of the NPM and previous collaboration with ING.
In India the sample consisted of 719 respondents for the treatment group and 717 for the control group. In Ghana the sample size were 521 and 511 for the treatment and control group respectively. The criteria for sample selection was the following:
1. Treatment: Active female group borrowers, living in rural areas, who have not borrowed from MFI’s or formal financial sources prior to 2010 or later than 2011.
2. Control: Females selected form the same location as that of the treatment village, and being those who had never received credit from MFI’s or any other formal source.18
The survey in India was conducted from December 2013 to January 2014, while the sampling for Ghana was conducted from February to March 2014. In India the interviews was conducted at the villages where the respondents live. The respondents for the treatment group were picked out during the center meeting. The people in the control group were randomly identified in the survey villages for interviews. The participation of the respondents was voluntary and free of any charges.
16
Basix is the financial part of the Basix group, which provides predominantly microcredit, technical assistance and insurance to the rural poor people.
17OI is a leading savings and loans institute in Ghana and began its operations in 2004. It mainly
specializes on savings, but also provides microcredit to poor people in rural as well as urban areas.
18
For the Ghana sample, the survey was organized at the OI branches, where all the selected respondents gathered for the interviews. Unlike in India, the clients were provided travel allowance, snacks and a gift (T-shirt) for participation.19 As a strat-egy to gather a control group, the treatment clients were asked to bring along people living in their local neighborhood, who had never received any microcredit loans. The OI branch staff then validated the control group respondents for interviews. The sampling method applied was the recall method for collection of the baseline sample. The recall method asks the respondent to answer a question for a given recall moment in addition to answering the question at the present moment. Due to budget and time constraints, this was done rather than the conventional way of first collecting a baseline sample and then returning one or more times in the future to collect more data points. The recall point chosen was October 2010, giving a period of between two and three years for microfinance to have its impact. Figure 1 below graphs this time horizon.
Figure 2: Time frame
Using this recall moment and treatment selection criteria, the borrowers of Basix should be in their 3rd loan cycle by the time of the survey, as the standard loan cycle used by Basix is 12 months. However, due to the microfinance crisis in 2010, it was hard to find enough clients matching the criteria, as Basix temporarily had to stop lending in Bihar. Therefore, the criteria for selection to the treatment group was expanded to include borrowers who still were in their second loan cycle and who have not borrowed from any formal sources prior to the recall moment.
The borrowers from OI who matched the criteria would be in their 4th to 8th loan cycle as the standard loan cycle used by OI was 6 months. When collecting the sample, urban clients had to be included as well, as it was otherwise hard to find
19The difference between the compensation in India and Ghana is a cultural matter. In India
enough clients for the treatment and control group. In the final sample, around 39% of the borrowers came from Urban branches, whereas the remaining belonged to rural branches.
After the sample was collected it was discovered that 99 and 36 borrows from Basix and OI respectively, have had formal loans prior to 2010. These clients were therefore dropped from the sample.20 The sample considered for the study was thus, 620 treatment and 717 controls for India, and 485 treatment and 511 controls for Ghana.
5.2 Sample Issues
Most often panel survey data are collected by interviewing the same household at different points in time. Sometimes households can also be asked only once, where they then answer questions about their current and past situations. While the first method is seen to give more precise and reliable estimates, it is more time consuming and comes with potential attrition bias as it can be hard to survey all the same households again. As this method of collecting data is costly, using retrospective survey data has become more popular, it does not come without problems as shown by Nicola and Giné (2014). Although the costs associated with information gathering are lowered, the estimates becomes less reliable the further away the recall moment is in the past. It is a well established fact that the magnitude of the measurement error is correlated with the length of the recall period (Tourangeau (2000)). In addition to the risk of respondent not recalling the events precisely, the estimates might suffer from further attrition bias when asked about unpleasant events (Skowronski et al. (1991); Holmes (1970)). The noise in the recall data then leads to what is known as attenuation bias, and may thus bias the coefficients that show the program impact downward.21
As mentioned above, the recall method migrates possible attrition problems the data might have, as there are no households "dropping out". However, due to the
20
Some would argue, that including these borrowers in the sample would make one able to check for long term effects of microfinance. However, the ’baseline’ characteristics of these individuals are affected by the intervention as they have already received microcredit at the recall moment. Thus, this will make matching on baseline characteristics and hence PSM problematic. Estimation including these borrowers were conducted and no difference in the results were found.
21
selection criteria for the treatment group employed in this study, the data might still be subject to attrition bias. Recall that the borrowers in the treatment group all started to borrow between 2010 and 2011 and thus have been borrowing between two and three years at the time of the survey. Thus, one does not observe any households who dropped out in this period. It seems quite reasonable to assume that Basix and/or OI had some clients dropping out doing that period. Given this assumption the impact can either be upward or downward biased, depending on the reason for the households to drop out. If the failures are more likely to drop out, the impact will be overestimated as only the ’good’ borrowers are left. It is also possible that the pattern is reversed and some borrowers are doing so good, that they leave the microfinance program as they then qualify for loans from regular commercial banks. However, the first case is often the one assumed in practice, hence our results maybe upward biased.
5.3 Data Summary and Balance Check
Tables 1 and 2 present the summary statistics and balancing tests for the control variables used in this study for India and Ghana respectively. Similarly, Table 3 and 4 displays the summary statistics for the outcome variables.
In comparative studies, where one wishes to compare two groups of respondents, causal inference often necessitates effective adjustment for important covariates. In particular, in absence of random assignment, one can expect to find some differences on observed covariates between the control and treatment group. Several popular adjustments exists to ensure balance between the control and treatment group, in-cluding matching and stratification of the propensity score.22 Although debated in the literature, a balancing test can generally be described as to determine if the observable controls are distributed similarly between the two groups in question.23 If any significant difference exists between the two groups, a balancing test should pick up this difference and indicate that the current composition of the data could lead to a biased estimate of the treatment effect. The balancing test is conducted as a OLS regression in which the variable in question is regressed on a constant and
22Under successful random assignment, the two groups should naturally be balanced and we
would expect that no adjusts are needed.
23
a treatment dummy. The standard errors in the regressions are clustered at the vil-lage level and branch level for the India and Ghana sample respectively to account for intraclass correlation. The coefficient of the treatment dummy now gives us an unbiased estimator of the difference between the control and treatment group at the recall moment.
Table 1 and 2 contains all the controls used throughout the different models estimated in this study. Some of the covariates are self-explanatory, while others needs further explanation. The binary variable "Family type" identifies if the household is a joint family or a nuclear family. A family is considered nuclear is the respondent is only living with the spouse and children. Anything else is considered joint. The "Caste" variable refers to which caste the respondent belongs to. The caste system in India is very vast and complicated. To simplify matters, the caste system is categorized into five broad categories doing the data collecting for the study. The categories include: Other Backward Caste (OBC), Schedule Caste (SC) and Schedule Tribe (ST), General Caste (GC) and Muslims. For a more detailed explanation of the different caste categories, see appendix B. Villages is not used as a control in Ghana due to the large number of villages compared to the number of observations (396 villages compared to 996 observations).
For the India sample, only "Household members" are different between the con-trol and treatment group at baseline. The magnitude of the "Household member" variable shows that treatment household have six members compared to five in the control group (rounded). Similarly for the Ghana sample, "Household members" are imbalanced, suggesting that the treatment group on average have 4 household members, compared to 3 in the control group (rounded). Additionally the treatment group are approximately 7 years older than the control group. One notable difference between the two samples include that the family type in India is more likely to be nuclear in India, while in Ghana more joint (64% contra 90%).24
<< Insert table 1 and 2 here >>
Table 3 and 4 displays the seven outcome variables considered in this study. The two "SonEduc" and "DaughterEduc" measures the education level of the oldest son and daughter in the household. The second outcome variable is called "Assets" and
24
is an index that captures the amount of assets a household possess. Besides physical assets such as cellphones and sewing machines, it also contains productive assets such as livestock like pigs, goats and poultry. The second index is called "Housing" and refers to the quality of the house. The third index presented here is named "Women empowerment". This index measures to which degree the respondent have influence on the decision making in the household. The fourth index is called "Household expenditures" and is constructed as a composition of house hold expenditures and household income, measured at a monthly basis. The fifth and final index is called "Financial literacy". This index measures the level of financial literacy of the respon-dent based on three questions regarding awareness of financial matters. No balancing test for the financial literacy index was conducted, as only there are only endline val-ues available and hence cannot be a balancing test. For additional information of which variables each of the indices are constituted by, see appendix A. Note that the mean of the control group and the constant term representing the control group in the balancing test are 0. This is due to the construction of the index where each of the variables are standardized based on the control group observed at the baseline. For India there is no apparent balance between the treatment and control group with exception of the education level of the oldest daughter in the household. The outcomes suggest that the treatment group in the baseline have an oldest son in the household, who are more educated, that the households in the treatment group are richer in terms of more assets and higher expenditures, better houses and are more aware of their financial situation. For both groups it holds that the women belonging to the treatment group are more empowered at baseline than the women in the control group. Additional, for the Ghana sample the balancing tests shows that only the Household expenditure index and the education variables are unbalanced. However, in the DD model this is controlled for.
<< Insert table 3 and 4 here >>
6
Results
This section will present the main results conducted in this paper. First, the statistics underlying the estimated propensity score will be presented and discussed. Next, the results using the DD after PSM are presented. The next section of the paper will then discuss the difference in the results between the two countries.
When using PSM, it is important to choose a good set of observable characteristics to match upon. These variables may not be affected by the treatment. Using baseline characteristics can help avoiding this problem and make it easier to obtain a good set of covariates.25 For this study the choice of covariates is the same as the controls used when estimating the DD model and the list can therefore be found in table 1 and 2 for India and Ghana respectively. After estimating the propensity score using a logit model controlling for the chosen set of covariates, the common support can then be assessed. If there are regions of the set of covariates does not overlap for the two groups, then matching is not justified. The estimated treatment effect can only be defined conditionally on the region of the overlap, the common support. Table 5 provides a description of the propensity score at baseline data for India and Ghana. It describes the inferior bounds for the propensity score, along with the number of observations dropped from the data as a result of being off the common support.
<< Insert table 5 here >>
The number of inferior blocks for India is four. This number ensures that the mean propensity score is not different for treated and controls in each blocks. The corre-sponding number of inferior block for Ghana is six. The table displays the allocation of control and treatment respondents within each of these block. Panel C gives the number of observations dropped from the sample for no being on the common sup-port. These numbers are quite low and indicates that the set of covariates seems appropriate. After ensuring a proper common support, the matching can be done.
25Recall, that this is the reason why people who received previous microcredit were dropped
6.1 PSM with Double Difference results
In table 6 and 7 the DD results are presented using the process described above. It describes the effect of microcredit on the seven outcome variables. The row labeled ’Treatment’ represents the β3 coefficient from equation (1), and estimate the sum-mary effect size increase/decrease of microcredit on the treatment group from the recall period till now. The main text presents the results including additional con-trol, where appendix C shows the results without the additional controls. Appendix C furthermore presents the results for the individual variables used in each of the five indices, thus enabling us to see which components of the indices affect the results. It is also worth noting that the set of controls used throughout all the regressions is identical to the set of covariates used for the PSM estimation earlier.
6.2 Basix - India
For the India sample we observe from the estimates that providing microcredit has a significant effect on the education level of the oldest son in the household, while no effect is found on the education level for the oldest daughter. The difference between the treatment and control group in the evolution of the education for the oldest son in the household is then 0.169 (p = 0.009). Providing microcredit also significantly leads to a summary effect size increase in the Housing of 0.070 (p = 0.007). The estimated summary effect size for women empowerment (WE) is also significant and has a magnitude of 0.093 (p = 0.000). Household expenditures (HHE) shows an effect size increase of 0.415 (p = 0.000). The final column displays the effect size of financial literacy (FL) and shows a significant coefficient of 1.137 (p = 0.000). Recall that there is no recall data available for the finalcial literacy (FL) index, thus the estimate does not rule out time invariant heterogeneity, which means that this estimate could be affect by unobserved variables such as innate ability and personality.
few robust determinants of change on the outcome variables. << Insert table 6 here >>
6.3 Opportunity International - Ghana
Unlike the India sample, only one summary effect size increase is found, namely a summary effect size increase on household expenditures. In contrast to India, summary effect size decreases are found on WE and Assets. The remaining coefficient for the summary effects are small in magnitude and insignificant. Additionally, no effects on the oldest son’s and daughter’s education are found, thus showing no impact of microcredit on the majority of outcome variables considered in Ghana. Comparing to the results without additional controls in appendix C, we see a sig-nificant effect on both the education level of the oldest son and daughter in the household. Like with the sample from India, some of the controls show a significant impact in some of the models, but not across tables.
Although the summary effect are insignificant, looking into the individual compo-nents several effects of microcredit can be found. Negative impacts for compocompo-nents of the asset index are found. This includes: Sewingmachine, Mobilephone, furniture, number of cattle and pigs. These results suggest that in the Ghana sample, the provision of microcredit can be even harmful, at least when considering the assets of the household.
<< Insert table 7 here >>
7
The Difference in the Impact of Microcredit
This section will elaborate on different reasons why the results differ between the two countries. Keep in mind that this can only be explorative and will thus not provide any hard evidence on why the results may differ between the two countries. By comparing the two sets of outcomes in table 6 and 7, it is evident that the impact of microcredit is smaller on all the outcome variables in Ghana. For instance, the summary effect size of Household expenditures in India is 0.415, while the effect size is 0.311 in Ghana. Likewise the summary effect size increase on housing and financial literacy are 0.070 and 1.137 in India, while the effect sizes in Ghana are -0.018 and 0.029 and insignificant. In the cases assets and WE, the sign is even reversed, thus showing a negative impact in Ghana, while positive in India. Also for the education variables, the magnitude of the estimates found in India are higher than the corresponding ones found in Ghana.
The impact on each of the indices could also be expected to be different as the indices might have different horizons in terms of impact manifestation. As discussed above in section two, the FL index can be viewed as an intermediate effect, while indices such as assets and housing are more long term effects. It can thus be expected to take longer before any significant impact is observed. As microcredit’s main focus is on poverty reduction, the index for household expenditures can be seen to be the most important. Expenditures are used as a proxy for income and thus an increase in household expenditures reflects a poverty reduction. While both countries show a significant summary effect size increase in expenditures, the negative impact on household assets in Ghana could reflect that this increase could arise from selling of assets to finance other consumption. However, there might be several other reasons for these results as will be explored below.
7.1 The use of microcredit
In this study, a list experimented in connection with the main survey was conducted, trying to address this issue. When asking a question of interest directly in a survey, it is more than uncertain that the respondent would answer truthfully, specially when the question deals with a sensitive topic (Tourangeau and Ting (2007)). To overcome this, researchers have developed a variety of techniques to obtain truthful answers from the respondents. 26 List randomization provides a simple way for respondents to report on sensitive behavior without making the researcher able to identify the individual response. This anonymity makes the respondent more likely to answer truthfully, as they feel more safe. To implement such a list experiment, the treatment group is divided 50/50 into two subgroups. The first group receives a short list of statements and asked how many of are true. The second group are given the same list, but with one additional statement to capture sensitive behavior. The additional statement, and hence the item of interest was: "The main expense financed by my first loan refers to household items, such as food, a TV, a radio etc.". By simply subtracting the mean number of true statements of second group from the mean number of true statements of the first group, the proportion of respondents engaging in the sensitive behavior can then be calculated. For comparison, the sensitive question was also asked directly in the survey, as to see how much the respondents would lie. The outcome of the list experiment are reported in table 8.
<< Insert table 8 here >>
The results reveals a striking contrast between direct questioning and list randomiza-tion. When asked directly about the sensitive behavior, only 2.4% and 0.01% admit to use their first loan on household items in India and Ghana respectively. On the other hand, using the list randomization technique suggest that 6.4% of the India sample used the loan for household items. In Ghana, this number was a striking 41.2%. The estimates are large, but are consistent with other studies in the field.27
26
These techniques include direct as well as indirect methods. Examples of direct methods include: Matching the gender of the respondent and the surveyor, using forgiving language or collecting data in private. Indirect methods include: randomized response technique, the bogus pipeline, and list randomization. For more discussion about the different techniques see e.g. Karlan and Zinman (2012).
27
7.2 Alternative explanations
This subsection will look at alternative explanations for the differences in outcome between the two countries. The first thing to look into would be how the two micro-credit programmes are structured. While both of the MFIs provide micromicro-credit along with savings and insurance products, they differ quite a bit in their business model and hence in their loan sanctioning process. Basix provides its loan services through three channels: Direct individual lending, individuals organized as Joint Liability Groups (JLG) and bulk lending to Self Help Groups (SHGs).28 Apart from offer-ing loans, Basix also offers non-financial services coveroffer-ing: Institutional development service, livelihood financial services and agriculture & business development services. Basix has a defined policy for its credit operations. This include the selection of the area from which to operate, how the branch are to be set up and the village and client selection. The details of are laid out in Appendix D.
OI uses two credit methodologies: Group and individual loans.29 Unlike Basix, OI has invested heavily in technology through the years and continuously tries to reinvent itself. They have recently introduced Automatic Teller Machines (ATMs) and Point-of-Sale (POS) devices. This along with mobile banking has enabled clients to conduct their transactions conveniently and safely. OI also offers financial literacy training in addition to their financial products. A brief description of their credit operations can also be found in Appendix D.
Some differences between the two programmes exist. For instance, the way in which the control was selected. While in India the control group was randomly identified in the survey villages, the control group was ’selected’ by the treatment group in Ghana. A person from the treatment group was asked to bring along a person living in their local neighborhood, who have not yet received microfinance. This approach could easily cause problem, as it is very likely that these ’controls’ are already affected by microcredit through possible spillover effects.. This makes the control contaminated and thus not valid.
Some other differences include their approach to technology and the size of their operations. OI has invested large sums in ATMs, POS and mobile banking, trying
28With 73.3% JLG loans took up the largest part of Basix’s loan portfolio. 29
to expand their outreach as much as possible. They are at the same time much smaller measured by total number of active borrowers. At the recall moment (2010) a total number of 41,836 active borrowers were registered at OI, while this number was 1,114,468 by Basix. One thing worth noticing though is how these numbers evolved in the period considered for this study. At the time of the survey (2013) these numbers were 69,369 and 377,421 for OI and Basix respectively.30 It is interesting to see the development in these numbers. While OI has seen a steady increase in the number of active borrowers as well as branches, Basix has experienced a sharp decrease in the number of active borrowers, but an increase in the number of branches.31 Apart from the microcredit programmes, the borrowers in the two countries could also differ on observable characteristics. To make a comparison, the summary of some selected observable baseline characteristics have been gathered in table 9 below:
<< Insert table 9 here >>
Note that we are not able to compare the two groups based on the indices as they are standardized via a vis their own controls at baseline. Of notable differences we see that the respondents in Ghana are about 5 years older, the household size is smaller. For the monetary observables, the households in India saves on average US$7 more than the households in Ghana. However, even if we for a moment were to assume that all relevant observable characteristics were included in the model, thus estimating the best possible model, the outcomes could still differ among the two countries due to unobservable characteristics. For instance, the entrepreneurial skill of the borrowers in India could be higher, their innate ability could be higher, they are more optimistic etc. Regardless of how extensive the set of control variables are, the main outcomes could still differ due to unobservable characteristics.
8
Conclusion
Using two field studies from India and Ghana, the impact of microcredit could be assessed using seven main outcome variables.
Using recall data with a horizon of three years, positive effects of microcredit is found
30The number of active borrowers were measured in December in each of the years for OI and
in March for Basix
31
on five out of seven outcome variables in India. Five of these outcome variables are indices created using a summary indexation technique where each of the individual outcomes are combined into an index representing the main outcome variable. A summary effect size increase is found on financial literacy which can be seen as an intermediate effect through which the other measures can be improved. Similarly , significant summary effect size increases are found on; housing, women empowerment and household expenditures. A positive effect on the education level for the oldest son in the household is also found. The impacts found, indicate that the oldest son in the household is more educated, households are having a larger income (measured by expenses), better housing, financial knowledge and the women are more empowered. Only the education level for the oldest daughter and the index for household assets show no effect.
The outcomes from the sample in Ghana shows a much different impact of the provision of microcredit. We fail to find any significant impacts on two of the indices and any of the education variables. An increase in the household expenditures index is found. While this is good news, the magnitude of the impact is smaller than the one found in India. The provision of microcredit is found to have a significant summary effect size decrease on women empowerment as well as household assets, and when looking at the individual components, negative effects are found on household assets related to livestock. This raises the concern that this reduction in livestock could contribute to the increase in household expenditures, as livestock are sold off to purchase other items.
Although the estimation methods and models used to assess the impact are the same across the countries, heterogeneous impact of microcredit are found. Several factors could be able to explain these differences. The main reason presented here is the difference in the credit use. A list experiment was conducted to assess what the main use of the credit was. The majority of the loans provided by both Basix and OI are intended to serve as a productive investment in the small enterprise or farming activity the household engages in. However, the list experiment revealed that the percentage of clients who mainly spend their loan for consumption purpose were 6.4% and 41.2% in India and Ghana respectively.
criteria for the control group all differ quite a bit between the two MFIs. Secondly, if one were to assume that all observable factors of interest were included in the model, unobservable characteristics can still differ between the countries, and hence changing the impact. DD might control for time invariant heterogeneity, but not for time variant ones. Lastly, when applying PSM, one accepts the assumption that selection can be made upon the observables chosen. Hence, unobservables characteristics might still cause bias in the sample.
References
Abdul, Sazali; Al-Mamun, Adbullah; Mohammad N.H., Mazumder, and Zhan, Su. Empirical investigation on the impact of microcredit on women empowerment in urban peninsular malaysia. Journal Of Developing Areas, 2014.
Anderson, Michael L. Multiple inference and gender differences in the effects of early intervention: A reevaluation of the abecedarian, perry preschool, and early training projects. Journal of the American Statistical Association, 2008.
Angelucci, Manuela; Karlan, Dean S., and Zinman, Jonathan. Win some lose some? evidence from a randomized microcredit program placement experiment by com-partamos banco. Economics Letters, 2013.
Armendáriz, Beatriz and Morduch, Jonathan. The Economics of Microfinance. The MIT Press, second edition, 2010.
Banerjee, Abhijit; Duflo, Esther; Glennerster, Rachel, and Kinnan, Cynthia. The miracle of microfinance? evidence from a randomized evaluation. MIT Department of Economics and Abdul Latif Jameel Poverty Action Lab (J-Pal) Working Paper, 2014.
Buhn, M.; Ibarra, L., and McKenzie, D. Why is voluntary financial education so unpopurlar? experimental evidence from mexico. The World Bank, Development Research Group, 2013.
Chemin, Matthieu. The benefits and costs of microfinance: Evidence from bangladesh. Journal of Development Studies, 2008.
Chowdhury, Sarahat Salma and Chowdhury, Sifat Adiya. Microfinance and women empowerment: A panel data analysis using evidence from rural bangladesh. In-ternational Journal of Economics and Finance, 2011.
Chunrong, Ai and Norton, Edward C. Interaction terms in logit and probit models. Economics Letters, 2003.
Cohen, Monique and Nelson, Candace. Financial literacy: A step for clients towards financial inclusion. 2011 Global Microcredit Summit, 2011.
Deaton, Angus. The analysis of household surveys: A microeconometric approach to development policy. Baltimore: Johns Hopkins University Press., 1997.
Duflo, Esther; Glennerster, Rachel, and Kremer, Michael. Using Randomization in Development Economics Research: A Toolkit in T. Paul Schultz, and John Strauss (eds.) Handbook of Development Economics. Elsevier Science Ltd, first edition, 2007.
Ebbes, Peter. Latent Instrumental Variables – A New Approach to Solve for Endo-geneity –. PhD thesis, University of Groningen, 2004.
Heckman, James J. and Jeffrey, A. Smith. Assessing the case for social experiments. Journal of Economic Perspectives, 1995.
Holmes, D. Differential change in affective intensity and the forgetting of unpleasant personal experiences. J. Personal. Soc. Psychol, 1970.
Humle, D.; Moore, K., and Barrientos, A. Assessing the insurance role of microsav-ings. DESA Working Paper No. 83., 2009.
Karlan, Dean S. and Zinman, Jonathan. Lying about borrowing. Journal Of The European Economic Association, 2007.
Karlan, Dean S. and Zinman, Jonathan. List randomization for sensitive behav-ior: An application for measuring use of loan proceeds. Journal of Development Economics, 2012.
Kleyman, Yevgeniya N. Testing for covariate balance in comparative studies. Ph.d. dissertation, University of Michigan, 2009.
Kling, J.; Liebman, J., and Katz, L. Experimental analysis of neighborhood effects. Econometrica, 2007.
McKenzie, D. and Woodruff, C. Experimental evidence on the returns to capital and access to finance in mexico. The World Bank Economics Review, 2009.
Nicola, Frensesca de and Giné, Xavier. How accuratre are recall data? evidence from coastel india. Journal of Development Economics, 2014.
Prema, Basargekar. Microcredit and a macro leap: An impact analysis of annapurna mahila mandal (amm), and urban microfinance institution in india. The IUP Journal of Financial Economics, 2009.
Puhani, Patrick A. The treatment effect, the cross difference, and the interaction term in nonlinear ’difference-in-differences’ models. Economics Letters, 2012. Rooyen, C. Van; Stewart, R., and De Wet, T. The impact of microfinance in
sub-saharan africa: A systematic review of the evidence. World Development, 2012. Rosenbaum, P.R. and Rubin, D.B. The central role of the propensity score in
obser-vational studies for causal effects. Biometrika, Vol. 70, No. 1. pp. 41-55., 1983. Sayinzoga, Aussi; Bulte, Erwin, and Lensink, Robert. Financial literacy and financial
behavior: Experimental evidence from rural rwanda. Working paper, 2014. Sebstad, Jennefer and Cohen, Monique. Microfinance, Risk Management and
Poverty. Pact Publications, 2000.
Sharma, M. and Zeller, M. Placement and outreach of groupbased credit organiza-tions: The cases of asa, brac, and proshika in bangladesh. World Development, 1999.
Skowronski, J.; Betz, A.; Thompson, C., and Shannon, L. Social memory in everyday life: recall of self-events and other-events. J. Personal. Soc. Psychol., 1991. Snodgrass, Donald and Sebstad, Jennifer. Clients in context: The impacts of
mi-crofinance in three countries: Synthesis report. MSI International, AIMS Project January, Washington, DC., 2002.
Tourangeau, Roger. Remembering what happend: memory errors and survey reports. Psychol. Pres, 2000.
Table 1: Summary statistics and balancing tests for baseline controls (India).
Summary Statistics Balancing tests Dependent variable No. C Mean C No. T Mean T Treatment
(1) (2) (3) (4) (5) Son age 616 6.963 568 7.060 0.097 (0.813) Daughter age 626 4.859 577 5.565 0.706* (0.081) Material status 694 2.140 620 2.132 -0.008 (0.821) Family type 694 0.646 620 0.637 -0.008 (0.794) Caste 694 2.630 620 2.570 -0.059 (0.466) Age 694 33.228 620 33.831 0.603 (0.325) Household members 694 5.457 620 5.834 0.377*** (0.002) Branch 694 2.007 620 2.029 0.022 (0.885) Village 694 29.189 620 29.369 0.181 (0.954)
Table 2: Summary statistics and balancing tests for baseline controls (Ghana).
Summary Statistics Balancing tests Dependent variable No. C Mean C No. T Mean T Treatment
(1) (2) (3) (4) (5) Son age 616 4.858 466 6.876 2.017*** (0.008) Daughter age 466 5.419 456 7.651 2.232*** (0.003) Material status 458 1.994 485 2.231 0.237** (0.023) Family type 499 0.898 485 0.866 -0.032 (0.257) Age 499 32.303 485 38.602 6.299*** (0.000) Household members 499 3.417 485 4.115 0.699*** (0.002) Branch 499 6.116 485 6.016 -0.100 (0.852)
Table 3: Summary statistics and balancing tests for summary indices and education variables at baseline (India).
Summary Statistics Balancing tests Dependent variable No. C Mean C No. T Mean T Treatment
(1) (2) (3) (4) (5) SonEduc 533 0.223 462 0.351 0.127** (0.001) DaughterEduc 545 0.132 462 0.167 0.035 (0.377) Assets 694 0.000 620 0.332 0.335*** (0.000) Housing 694 0.000 620 0.649 0.668*** (0.000) Women Empowerment 694 0.000 620 0.739 0.712*** (0.000) Household expenditures 694 0.000 620 0.691 0.671*** (0.000) Financial Literacy 694 0.000 620 1.150 n.a.
Table 4: Summary statistics and balancing tests for summary indices and education variables at baseline (Ghana).
Summary Statistics Balancing tests Dependent variable No. C Mean C No. T Mean T Treatment
(1) (2) (3) (4) (5) SonEduc 419 0.473 397 0.768 0.296*** (0.002) DaughterEduc 415 0.405 394 0.787 0.382*** (0.001) Assets 511 0.000 485 0.039 0.036 (0.649) Housing 511 0.000 485 0.075 0.085 (0.254) Women Empowerment 511 0.000 485 0.408 0.422*** (0.000) Household expenditures 511 0.000 485 0.195 0.176** (0.015) Financial Literacy 511 0.000 485 -0.006 n.a.
Table 5: Description of propensity score at baseline line
Responded group Inferior of block of p-score Control Treatment Total Panel A: India 0.2 136 81 218 0.4 388 333 721 0.5 144 170 314 0.6 25 36 61 Total 694 620 1,314 Panel B: Ghana 0.184 15 2 17 0.2 128 29 157 0.3 98 53 151 0.4 160 205 365 0.6 88 181 269 0.8 10 15 25 Total 499 485 984
Panel C: Off support
India 0 23
Ghana 0 12
Panel D: Region of Common support
India [0.315 ; 0.719] Ghana [0.184 ; 0.919]
Table 6: Summary index results using Propensity Score combined with Double Difference (India).
(1) (2) (3) (4) (5) (6) (7)
VARIABLES SonEduc DaughterEduc Assets Housing WE HHE FL
Treatment 0.169*** 0.047 0.045 0.070*** 0.093*** 0.415*** 1.137*** (0.009) (0.181) (0.216) (0.007) (0.000) (0.000) (0.000) Constant 0.694** 0.213 0.154 -0.443* -0.402 -1.178*** 0.586 (0.023) (0.139) (0.542) (0.063) (0.119) (0.000) (0.114) Observations 1,517 1,703 2,628 2,628 2,628 2,628 1,316 R-squared 0.428 0.351 0.051 0.154 0.150 0.288 0.128
Table 7: Summary index results using Propensity Score combined with Double Difference (Ghana).
(1) (2) (3) (4) (5) (6) (7)
VARIABLES SonEduc DaughterEduc Assets Housing WE HHE FL
Treatment 0.013 0.004 -0.092*** -0.018 -0.030* 0.311*** 0.029 (0.580) (0.904) (0.003) (0.528) (0.087) (0.000) (0.679) Constant -0.271 -0.330 -0.074 -0.016 -0.609*** -1.344*** 0.330 (0.111) (0.107) (0.278) (0.908) (0.000) (0.000) (0.423) Observations 1,499 1,464 1,968 1,968 1,968 1,968 978 R-squared 0.767 0.679 0.023 0.003 0.085 0.168 0.017
Table 8: Comparison of direct report and list randomization estimates from Basix and OI.
India Ghana Loan use
Direct report
Proportion reported this use 0.024 0.001 SE 0.154 0.036
N 620 768
List randomization
Mean of "Yes" responses for short list 2.073 2.231 SE 0.057 0.081
N 385 295
Mean of "Yes" responses for long list 2.717 2.644 SE 0.044 0.037
N 385 295
Difference (Proportion reporting this use) 0.064 0.414
SE of difference 0.071 0.090
p-value from t-test 0.000 0.000
N 770 590
Comparison of direct report and list randomization
List randomization minus direct report 0.040*** 0.413***
