The characteristics of microfinance participants
An analysis of the traits that influence acceptance of microfinance loans
Danny Wolf 10212418
Bachelor thesis Econometrics Supervisor: S. Stephan
Universiteit van Amsterdam June 26, 2018
Statement of Originality
This document is written by Student Danny Wolf who declares to take full respon-sibility for the contents of this document.
I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Contents
1 Introduction 1 2 Theoretical framework 3 2.1 Microfinancing . . . 3 2.2 Characteristics of participants . . . 4 2.3 Proxy variables . . . 53 Data & model 7
3.1 Data set . . . 7 3.2 Logit model for binary data . . . 10
4 Results & analysis 12
4.1 Descriptive statistics . . . 12 4.2 Logit regression results . . . 15
1
Introduction
Poverty is an issue many governments actively try to combat. One of the ways this is done is by providing small loans, which allows participating individuals to start or improve their business. Aside from government initiatives, there are also government organizations providing microfinancing. These are known as non-government organizations micro finance institutions (NGO-MFIs).
Multiple studies found the positive impact of these loans to be significant. Poor business owners less efficient due to credit constraints (Omonona et al, 2010; Feder et al, 1990). Microfinancing would be a solution to increase their efficiency by alle-viating the credit constraint. Even though microfinancing seems to reduce poverty, many individuals decline when offered a microloan to fuel a business.
This study focuses on the characteristics of the people who are offered a loan. The goal is to determine which characteristics have an influence on whether an individual accepts or declines a loan. This information can be useful for MFIs, since they can then focus their resources on informing people who are more likely to be willing to participate. They can also use this information to tailor the loan products to the needs of the people in need of credit.
The estimation is done using a logit model, which is one of the standard methods when modeling a binary variable. It is important not to omit important variables from the model. Multiple studies have found several significant variables and this
paper will combine those in a single model to test whether they significantly impact the decision to take up a loan.
It has been suggested that age, sex, family size and wealth are factors in the deci-sion to accept or decline a microloan, where younger males with a bigger family and less wealth are more likely to accept (Rasheed et al., 2016). The data being used in this paper does not contain a definitive measure for wealth. However, there are multiple variables indicating economic status available, which mostly focus on the status of housing and occupation. In this paper some of those variables are used to approximate wealth, which allows for estimation of the effect of wealth on the decision to accept or reject a microfinance loan. Three ways to approximate wealth are used and the results are compared in order to check if the regression is sensitive to which variables are chosen as proxy for wealth. Finding a good proxy for wealth can be beneficial for many future studies with the same issue.
The remainder of the paper is organized as follows. Section 2 provides more informa-tion on microfinance and discusses which variables should be included in the model. Section 3 describes the data and the econometric model. In section 4 the models are estimated, the findings are discussed and compared with each other and with previous research. Section 5 concludes.
2
Theoretical framework
This chapter gives a summary of previous related research. A description of micro-financing and the institutions are given for context. This is followed by a discussion of characteristics of participants of microfinancing in previous studies. Afterwards proxy variables are discussed.
2.1
Microfinancing
Traditional banks and credit companies require borrowers to have a certain amount of income or some other collateral to ensure the loan can be paid back. This means they do not issue loans to the poorest entrepeneurs of the population (Zhu, Ma & Shi, 2009). Microfinancing loans are provided by microfinance institutions (MFIs). They provide small loans without collateral to people who would be refused by banks and the MFIs don’t necessarily generate profit from these loans (Armendariz de Aghion & Murdoch, 2005 p.4). This system supports the poor in setting up or expanding businesses. Akanji (2006) claims that microfinancing is used as a way to reduce poverty.
2.2
Characteristics of participants
This section will discuss the traits of those who accept the loans. Previous studies conclude some socio-economic characteristics to be of importance when an individ-ual decides to reject or accept a loan.
Rasheed, Xia, Ishaq, Mukhtar and Waseem (2016) find several important charac-teristics in their study. Their study was conducted on poor farmers. They find that males are more likely to accept a loan than females, although it was barely a sig-nificant result. Their data shows family size and education to have a positive effect on the demand for credit. The farm size and income level of the farm have a signifi-cant positive effect in their study. These last two variables are statistically the most significant in their study. This indicates that already successful farms have a higher demand for credit.
Omonona et al (2010) also conducted a study on poor farmers. They find females are more inclined to accept loans, which seems to contradict the findings of Rasheed et al (2016). They also conclude that bigger families and a higher level of education leads to a higher demand for credit, which does coincide with the findings of Rasheed et al (2016). Lastly, they find that the probability of accepting a microloan increases with age.
Okurut, Schoonbee and van der Berg (2005) conclude that age, sex, education level, family size and household expenditure have an influence on an individuals decision to accept a loan. Males are more likely to accept than females. Being older, more educated, having a bigger family and having higher household expenses all positively affect the demand for credit (Okurut et al, 2005). This mostly corresponds to the results of Omonona et al (2010). They only disagree on whether males or females are more likely to accept.
Nguyen (2007) finds age to be negatively affecting the decision to accept a loan. Younger people are more likely to do so, which is the opposite of what Okurut et al (2007) and Omonona et al (2010) conclude. Nguyen (2007) also concludes gender and education do not influence the decision. He does find that the family size has a large significant positive effect on the demand for credit. Whether an individual works in agriculture or not also influences the demand, where people that do work in agriculture have a higher demand for credit (Nguyen, 2007).
Summarizing the above studies, we can conclude that family size, education level, occupation and income/expenditures affect an individuals decision to accept a loan. The studies are not clear on the effect of age and gender, where for both variables studies find negative effects, positive effects or no significant effect at all.
When a variable which affects the decision is not available, as is the case in this paper, one can use a proxy variable. These are discussed in the next section.
2.3
Proxy variables
Since income is missing from the data set used in this paper, a proxy variable will be used. This is done to remove the omitted variable bias, but using a proxy can introduce a new bias. McCallum (1972) and Wickens (1972) both show that the bias created by the proxy is always smaller than the omitted variable bias if the proxy does not contain systematic measurement errors. According to them the use of a poor proxy is still better than omitting the variable entirely. Frost (1979) continues on their findings and concludes that using poor proxies can increase bias in a model.
Morris, Carletto, Hoddinott and Christiaensen (2000) used multiple dummy vari-ables as a proxy for income. They used a weighted sum of household assets where the less common assets get a higher weight. The weights used are the reciprocal of the occurrence ratios. So a luxury asset which occurs in 15 per cent of houses would get a weight of 0.151 .
Tschirley and Mathenge (2003, pp. 8-9) provide general types of variables to use as a proxy for income. They present the possible proxies in multiple categories. In short, these categories mostly contain crop yield, amount and type of labour done and assets owned.
3
Data & model
3.1
Data set
The goal of this paper is to determine which characteristics influence an individual to take on a loan. The previous section discussed several variables which can be of importance. The effects of these variables will be examined in this paper. This section will discuss the data set and how some of the problems with the data are dealt with.
The data consists of a survey conducted in 75 villages in rural southern Karnataka, a state in India. The data used is collected by Banerjee, Chandrasekhar, Duflo and Jackson (2012). The dataset is available on the Social Networks and Microfinance project web page. It contains multiple socio-economic variables about the individ-uals in the village and their household. Important to note is that income is not included among these variables, but possible proxieas are. Six months after the sur-vey was conducted BSS, a MFI, started to offer microfinancing loans to all eligible villagers, which are women aged between 18 and 57. BSS added an additional re-striction which only allowed a maximum of one loan to be disbursed per household. Due to operational difficulties BSS only managed to start its operations in 43 out of the 75 villages, so the data used in this paper will consist of those 43 villages. Furthermore, some households didn’t provide any individual data of the inhabitants. These data points were dropped, since the individual characteristics are necessary
for the main goal of this paper. A few other households didn’t have any eligible inhabitants. These were also dropped, because these people don’t have the option to take on a loan. This leaves 4571 households in the data set.
There can be more than one eligible villager in a house, but a maximum of one loan can be accepted per house. Because of this, one cannot simply use the individual data of all the eligible villagers in the analysis. There are multiple ways to solve this, such as using averages of the household eligible inhabitants or only using the individual characteristics of the household heads. In this paper the latter option is used under the assumption that the household head influences the entire households decision to take on a loan. The household heads themselves do not necessarily have to be eligible to take on a loan. With this assumption the individual characteristics of the household heads are used in the analysis. Those characteristics are the ones discussed in section 2.2, which are family size, education level, occupation, income and age. Gender is not used as the loans are only disbursed to females in this data set. Family size, education level, income and working in agriculture are expected to positively affect the dependant variable. The effect of age is disputed among the studies and as such is expected to have no influence. However, age is included to test whether it has an effect on the dependant variable or not. Family size might get an overestimated effect, since a bigger family tends to have more eligible members and therefore a higher probability of accepting a loan.
Family size is be the number of people in the household. Occupation is represented by dummy variables for unemployed, working in agricultural sector and working elsewhere, since Nguyen (2007) concludes working in agriculture affects the decision to take on a microfinance loan. Education is an ordered variable ranging from 0 to 14 with 0 being having no education and 14 having a degree or another diploma. In the original data set it was ordered from 1 to 16 with 16 being having no education
and 15 having another diploma. Age is the age in years. Income is not available in the data set, as such using proxies will be considered.
In this paper a weighted sum is used as a proxy for income in the same way it was used by Morris et al (2000). The weights are the reciprocal of the occurrence. The assets used in this sum are some non-land assets as recommended by Tschirley and Mathenge (2003, pp. 8-9). They are whether the house is owned or rented, whether the house has private electricity or not, whether the household head has a savings account and whether the house has a private latrine.
A second analysis is done using the same sum as before, but with weights set to one. This allows for a comparison with previous regression to check whether using weights based on occurrences is an improvement.
Furthermore a third analysis is done to examine if the coefficients of the other ex-planatory variables are sensitive to the choice of proxies. The number of days worked last week is used as a proxy in the third regression. This is also a proxy for income recommended by Tschirley and Mathenge (2003, pp. 8-9). Unfortunately, some ob-servations are defined as hours per week or hours per day worked. These households are dropped and this leaves 4345 households to use in the regressions. Some indi-viduals worked 8 days last week according to the data set. These have been set to 7 days per week.
Lastly a fourth regression is done without including a proxy for income to exam-ine how much the coefficients are affected by omitted variable bias, assuming the proxies used in the other regressions are good proxies. Mani et al (2013) show that education is correlated with income. It is expected that the coefficient for education changes the most when introducing omitted variable bias.
3.2
Logit model for binary data
For the regressions, this paper uses logistic regression as introduced by Cox (1958) and it is also used by Banerjee et al (2012). The model can be written as
logit(E[Yi|Xi]) = logit(pi) = ln(
pi
1 − pi
) = Xi0β (1)
Where:
Yi is the binary dependant variable.
Xi is a vector with explanatory variables, including a constant.
logit(.) denotes the logit function. pi = P[Yi = 1|Xi]
β is a vector with coefficients need to be estimated. i is the household
The dependant variable, Yi in equation (1), is the decision made by household head
i; 0 if he rejected and 1 if he accepted the loan.
The explanatory variables are the ones discussed in section 3.1., including the dis-cussed proxy variables.
Equation (1) can be rewritten to isolate pi.
Λi(Xi0β) = E[Yi|Xi] = pi =
eXi0β
eXi0β + 1 (2)
Equation (2) shows an easier interpretation, since the probability of Yi = 1 is
ex-plicit. A positive coefficient β would mean an increase in the corresponding X would increase the probability of taking on a loan.
different for each individual. If xji denotes the jth explanatory variable for person i, then ∂Λi(Xi0β) ∂xji = βj eXi0β (1 + eX0 iβ)2 = βj∗ λi(Xi0β) (3)
Multiplying the coefficient found with logistic regression with a correction factor of λi gives the marginal effect for individual i. In the following chapter the results of
4
Results & analysis
4.1
Descriptive statistics
This section shows more information of the social-economic characteristics in the data set and these are uses to calculated the weighted sum, which is used as a proxy for income.
Figure 1 shows statistics of the household heads and their households. In this sam-ple the take-up rate of the loans is 18.6%. Around 15.3% do currently not have a job, 49.3% work in the agricultural sector and the remaining 35.4% work in an-other sector. Figures 2 and 3 are histograms of the age and education level of the household heads. All ages seem to be represented. Many of the household heads have no education at all.
To generate the weighted sum variable as used by Morris et al (2000), the frequency of the assets in figure 1 are used. The weights are the reciprocal of the occurrence.
W sum = 1 0.9 ∗ House + 1 0.631 ∗ Electricity + 1 0.386 ∗ Savings + 1 0.275 ∗ Latrine
Variable Mean Std. Dev. Min. Max. Participation 0.186 0.389 0 1 Family size 2.554 1.151 1 17 Agricultural sector 0.493 0.5 0 1 Unemployed 0.153 0.36 0 1 Age 45.73 11.535 20 95 Education level 4.751 4.517 0 14 House ownership 0.9 0.3 0 1 Private electricity 0.631 0.483 0 1 Savings account 0.386 0.487 0 1 Latrine 0.275 0.447 0 1 Weighted Asset Sum 4.001 2.662 0 8.923 Asset Sum 2.192 1.060 0 4 Work frequency 4.669 2.307 0 7
N 4345
Figure 1. Summary statistics
4.2
Logit regression results
In this section the results of the regressions are shown. Firstly the regression in-cluding the weighted sum as a proxy for income, secondly the asset sum with equal weights is used, thirdly work frequency is used as proxy and lastly the regression without any proxy. These results are compared to previous studies in section 2.2 and with each other.
Figure 4 shows the results of the regression with the weighted sum. Family size has a significant positive effect on the decision to accept a microfinance loan. This is in agreement with the results of Rasheed et al (2016), Omonona et al (2010), Okurut et al (2005) and Nguyen (2007).
Occupation seems to have an effect on the demand for credit, where being unem-ployed results in having the lowest demand. Individuals working in the agricultural sector have a lower demand than those working in different sectors. Other busi-nesses in the data might be more capital intensive than agriculture. This is the opposite result of what Nguyen (2007) found, who concluded the agricultural sec-tor has a bigger demand for credit.
Previous studies were in disagreement about the effect of age. In the data set used in this paper it appeared that households with a younger head are more likely to accept a loan.
Education is found to have a negative impact on the participation in microfinance. Rasheed et al (2016), Omonona et al (2010) and Okurut et al (2005) found positive effects. Nguyen (2007) did not find a significant effect.
The weighted sum is assumed to be positively correlated with income, which means the coefficient would have the same sign as the coefficient of income would have. As such a negative effect of income on the demand for credit is found in the data, which is the opposite of what previous studies found.
Using equation (2) in section 3.3, one can calculate probabilities to accept a loan in this model. Using the means reported in figure 1, a household with an average fam-ily size, age and education and the head working in a sector other than agriculture and with only electricity as a luxury asset used in the sum has a Λ(−1.196) = 0.232 chance of accepting a loan.
The marginal effects for this household can be calculated using equation (3). The marginal effect for working in agriculture is βagri ∗ λ(−1.196) = −0.048, which
means the probability of accepting a loan would drop from 0.232 to around 0.184 if the household head works in the agricultural sector instead of another one.
Variable Coefficient Std. Err. z-value P-value Family size .1718169 .0350397 4.90** 0.000 Agricultural sector -.2702055 .0867444 -3.11** 0.002 Unemployed -.3809391 .1334374 -2.85** 0.004 Age -.0141765 .0041237 -3.44** 0.001 Education level -.0331996 .0099631 -3.33** 0.001 Weighted asset sum -.0996155 .017063 -5.84** 0.000 Constant -.5685681 .18299986 -2.11** 0.002
N LR χ2(6) = 102.78 Pseudo R2
4345 P=0.0000 0.0246
Figure 4. Logit regression with weighted sum. Significant at ** (p<0.01), * (p<0.05).
The regression using a sum of assets with equal weights is shown in figure 5 and has similar results as the previous regression with weights based on occurrence. The coefficients are more positive although they do not significantly differ.
The coefficient corresponding to the sums does differ. Since the weighted sum in figure 4 uses weights bigger than one, the sum will be bigger than the sum used in figure 5. Figure 1 shows that the weighted sum is roughly twice as big on average
as the sum with weights equal to one. As the corresponding coefficient is around half the size, both sums have an equal effect in their model.
Using equation (2), the same imaginary family as before with an average size, age and education and working in a sector other than agriculture and with electricity as a luxury asset, has a Λ(−1.044) = 0.260 chance of accepting a loan. This is higher than the probability calculated with the previous regression as the coeffi-cients as less negative here.
Variable Coefficient Std. Err. z-value P-value Family size .1728188 .0350349 4.93** 0.000 Agricultural sector -.2539607 .0868027 -2.93** 0.003 Unemployed -.3687809 .133402 -2.76** 0.006 Age -.0141489 .0041212 -3.43** 0.001 Education level -.033015 .009936 -3.32** 0.001 Asset sum -.2524827 .0412424 -6.12** 0.000 Constant -.4285553 .1857318 -2.31* 0.021 N LR χ2(6) = 105.54 Pseudo R2 4345 P=0.0000 0.0253
Figure 5. Logit regression with equally weighted sum. Significant at ** (p<0.01), * (p<0.05).
Figure 6 contains the results of the regression with work frequency as the proxy for income. The coefficients for the proxy and the constant are not significantly different from zero in this model. The rest of the results look similar to those of the previous regressions, although all coefficients do significantly differ. Unemploy-ment and education have a bigger negative effect than in the regression with a sum of assets. It is possible that work frequency is not correlated with income enough to be used as a proxy. This would explain why a higher coefficient for education
was found. Since education is positively correlated with income (Mani et al, 2013), its coefficient would take over some of the effect the omitted income would have.
The change in the coefficient of the unemployed dummy could be explained by the nature of the proxy that is used here. Unemployed people most probably did not work in the previous week. If the difference in income between working zero and one day is bigger than the difference between working six and seven days, the coefficient should be larger at the lower values than it is here in the linear model. The unemployment coefficient could correct for this, which might have happened in this regression.
Using this model the probability of accepting a loan is Λ(−0.770) = 0.317 for the same family as used before and with an average work frequency as shown in figure 1.
Variable Coefficient Std. Err. z-value P-value Family size .1539487 .0346699 4.44** 0.000 Agricultural sector -.2811815 .0868396 -3.24** 0.001 Unemployed -.6797205 .2229284 -3.05** 0.002 Age -.0176047 .0040883 -4.31** 0.000 Education level -.0521927 .0093258 -5.60** 0.000 Work frequency -.0592087 .0324088 -1.83 0.068 Constant -.3292536 .2526249 -1.30 0.192 N LR χ2(6) = 70.67 Pseudo R2 4345 P=0.0000 0.0169
Figure 6. Logit regression with work frequency. Significant at ** (p<0.01), * (p<0.05).
The results of the regression without a proxy for income are shown in figure 7. The goal of this regression was to examine the impact of the previously used proxies on the coefficients of the other explanatory variables.
The coefficient are similar to those found with the work frequency proxy, except for the unemployment coefficient. Using days worked the previous week as a proxy for income seems unsuccessful. Other ways to quantify the amount of labour done might prove more successful.
Both sums of assets seem to perform well as a proxy for income. It was expected that omitted variable bias would effect education the most. Since the coefficient of education gets smaller when the sum is added to the regression, it seems the sum performs well as a proxy to remove the bias.
This model predicts the probability of the same family as before accepting a loan as Λ(−0.809) = 0.308. The predicted probabilities by all 4 models are shown in figure 8. The take up rate of the loans is 18.6 per cent. The fictional family used in the calculations are average except for the employment sector. As working in a sector other than agriculture leads to a higher probability of demanding credit according to all 4 regressions, it is expected that the probability they accept a loan is higher than the take up rate, which is the case with all 4 regressions.
Figure 8 also shows the results of a likelihood-ratio test. The last model without a proxy is compared with the others and the test shows that using either sum of assets significantly fits the data better than not including a proxy at all. This is not the case when using days worked last week as a proxy.
Variable Coefficient Std. Err. z-value P-value Family size .1516783 .0346397 4.38** 0.000 Agricultural sector -.2650377 .0863364 -3.07** 0.002 Unemployed -.3510628 .1326914 -2.65** 0.008 Age -.0175478 .0040867 -4.29** 0.000 Education level -.0540483 .0092676 -5.83** 0.000 Constant -.650624 .1817852 -3.58** 0.000 N LR χ2(6) = 67.37 Pseudo R2 4345 P=0.0000 0.0161
Figure 7. Logit regression without income proxy. Significant at ** (p<0.01), * (p<0.05).
Regression Λ(Xi0β) LR χ2(1) p-value Weighted sum of assets 0.232 35.41** 0.0000 Equal weight sum of assets 0.260 38.16** 0.0000 Days worked last week 0.317 3.30 0.0693
No proxy 0.308 -
-Figure 8. Lambda-values and LR-test. Significant at ** (p<0.01), * (p<0.05).
5
Conclusion
The main goal of this study was to determine which characteristics of individuals cause them to accept microfinance loans. To achieve this goal a proxy for income was to be found.
Using days worked the previous week as a proxy appeared to be a poor choice. A good choice was to use a sum of assets, which seemed to remove some of the bias and fitted the data significantly better. Using weights based on the occurrence of the assets did not have an improvement compared to using equal weights.
In the regressions with a sum of assets, family size was found to have a positive effect on the demand for credit. Education level and income were found to nega-tively impact ones decision to accept a loan. Younger people in this data set were found to be more likely to take on a loan. People working in agriculture were less likely to accept a loan than in other sectors and unemployed were less likely to do so than employed were.
The findings in this study do not correspond to those of previous studies. Most notable is the disagreement on education and income for which other studies find positive effects.
in-stead of the people who are offered a loan, since the MFI offered at most one loan per household. This also affects the estimate of the effect of family size on the demand for credit. Bigger families tend to have more eligible members. This study can be improved by using data where every individual is offered a loan without restrictions.
Bibliography
Akanji, O. (2006). Microfinance as a Strategy for Poverty Reduction: A Review of Countrys Experiences of Finance Program for the Poor. Cbn economic & financial review, 39 (4).
Armendariz de Aghion, B. & Murdoch, J. (2005). The Economics of Microfinance. Cambridge, Masschusetts: MIT Press
Banerjee, A., Chandrasekhar, A., Duflo, E., & Jackson, M. (2012). The diffusion of microfinance (NBER Working Paper, 17743).
Cox, D. (1958). The Regression Analysis of Binary Sequences. Journal of the Royal Statistical Society. Series B (Methodological), 20 (2), 215-242.
Feder, G., Lau, L., Lin, J. and Luo, X. (1990). The Relationship between Credit and Productivity in Chinese Agriculture: A Microeconomic Model of Disequilib-rium. American Journal of Agricultural Economics, 72 (5), 1151-1157.
Frost, P. (1979). Proxy Variables and Specification Bias. The Review of Economics and Statistics, 61 (2), 323-325.
Mani, S., Hoddinott, J., & Strauss, J. (2013). Determinants of Schooling: Empiri-cal Evidence from Rural Ethiopia. Journal of African Economies, 22 (5), 693-731.
McCallum, B. (1972). Relative Asymptotic Bias from Errors of Omission and Measurement. Econometrica, 40 (4), 757-758.
Morris, S., Carletto, C., Hoddinoot, J., & Christiaensen, L. (2000). Validity of rapid estimates of household wealth and income for health surveys in rural Africa. Journal of Epidemiology and Community Health, 54 (5), 381.
Nguyen, C. H. (2007). Determinants of credit participation and its impact on household consumption: Evidence from rural Vietnam. Centre for Economic Re-form and TransRe-formation, School of Management and Languages. Heriot-Watt University, Edinburgh.
Okurut, N., Schoombee, A., & Berg, S. van der (2005). Demand and Credit Rationing in the Informal Financial Sector in Uganda. South African Journal of Economics, 73 (3), 482-497.
Omonona, B., Lawal, J. and Oyinlana, A. (2010). Determinants of Credit Constraint Conditions and Production Efficiency Among Farming Households in Southwestern Nigeria. 2010 AAAE Third Conference/AEASA 48th Confer-ence. Cape Town, South Africa: African Association of Agricultural Economists (AAAE);Agricultural Economics Association of South Africa (AEASA).
Pitt, M. and Khandker, S. (1998). The Impact of GroupBased Credit Programs on Poor Households in Bangladesh: Does the Gender of Participants Matter?. Journal of Political Economy, 106 (5), 958-996. doi: 10.1086/250037
Rasheed, R., Xia, L., Ishaq, M., Mukhtar, M. and Waseem, M. (2016). Deter-minants influencing the demand of microfinance in agriculture production and estimation of constraint factors. International Journal of Agricultural Extension and Rural Development Studies, 3 (4), 45-58.
Strauss, J., Lei, X., Park, A., Shen, Y., Smith, J., Yang, Z., & Zhao, Y. (2010). Health Outcomes and Socio-economic Status Among the Elderly in China: Evi-dence from the CHARLS Pilot. Journal of Population Ageing, 3 (3), 111-142.
Tschirley, D., & Mathenge, M. (2003). Developing Income Proxy Models for Use by the USAID Mission in Kenya: A Technical Report (Tegemeo Working Paper 7).
Wickens, M. (1972). A Note on the Use of Proxy Variables. Econometrica, 40 (4), 759-761.
Zhu, X. , Ma, X., & Shi, Q. (2009). An Empirical Analysis of Reputation, Wealth and Rural Credit Rationing: Study on Supply Behavior of Different Rural Credit Organizations in Less Developed Areas. Journal of Finance and Economics, 8 (1), 23-30.
