did microfinance increase poverty in urban zimbabwe? · pdf filedid microfinance increase...

28
Did Microfinance Increase Poverty in Urban Zimbabwe? By Kristina Meier and Robert Rudolf* *Georg-August-University Göttingen, Germany Göttingen, April 2010 Working paper: Please do not cite. Abstract Microfinance has received much attention during the last two decades with positive examples from Bangladesh, Indonesia or Bolivia, among others. However, the start of rigorous impact evaluation at the end of the 1990s showed that positive impacts perceived were partly due to self-selection of more able clients into lending projects. Furthermore, positive impacts were often biased towards the initially richer and more dynamic poor. For the case of Zimbabwe, several microfinance institutions (MFI) have started their services during the 1990s. Yet, it seems that they have not hit on fertile ground since popularity of formal group lending schemes remains low. Analyzing a two-wave household and enterprise panel data set from Zimbabwe’s largest MFI Zambuko for 1997 and 1999, this paper investigates the poverty impact of microfinance in urban Zimbabwe controlling for issues of endogeneity and selection bias. We find a significantly negative impact of borrowing from Zambuko on urban households’ real per-capita income controlling for the intensity of program uptake and drop-outs. The negative impact increases in absolute terms when we additionally control for endogenous program participation. On the one hand, reasons for this negative result might lie in the low efficiency of Zimbabwe’s microfinance institutions. Furthermore, related literature suggests that informal agents do not trust MFIs. On the other hand, MFIs in Zimbabwe stand in direct competition with indigenous savings clubs, so called ma rounds. These popular rotating savings and credit associations (ROSCAs) are traditionally embedded in the Zimbabwean culture and historically approved, thus solving problems of moral hazard and adverse selection at very low costs. Formal group lending schemes should therefore find their economic niche and target those poor that are not accessed by informal solutions. (JEL C34, C81, G29, O16, O17, O18, O19) Keywords: Endogeneity; Selection bias; Microcredit; Income poverty; Impact analysis; Zimbabwe; ROSCA

Upload: phamlien

Post on 18-Mar-2018

220 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

� 頴頴� � 省省省蘟 Ю Ю Did Microfinance Increase Poverty in Urban Zimbabwe?

By Kristina Meier and Robert Rudolf*

*Georg-August-University Göttingen, Germany

Göttingen, April 2010

Working paper: Please do not cite.

Abstract

Microfinance has received much attention during the last two decades with positive examples from Bangladesh, Indonesia or Bolivia, among others. However, the start of rigorous impact evaluation at the end of the 1990s showed that positive impacts perceived were partly due to self-selection of more able clients into lending

projects. Furthermore, positive impacts were often biased towards the initially richer and more dynamic poor.

For the case of Zimbabwe, several microfinance institutions (MFI) have started their services during the 1990s. Yet, it seems that they have not hit on fertile ground since popularity of formal group lending schemes remains low. Analyzing a two-wave household and enterprise panel data set from Zimbabwe’s largest MFI Zambuko for 1997 and 1999, this paper investigates the poverty impact of microfinance in urban Zimbabwe controlling for issues of endogeneity and selection bias. We find a significantly negative impact of borrowing from Zambuko on urban households’ real per-capita income controlling for the intensity of program uptake and drop-outs. The

negative impact increases in absolute terms when we additionally control for endogenous program participation.

On the one hand, reasons for this negative result might lie in the low efficiency of Zimbabwe’s microfinance institutions. Furthermore, related literature suggests that informal agents do not trust MFIs. On the other hand, MFIs in Zimbabwe stand in direct competition with indigenous savings clubs, so called ma rounds. These popular rotating savings and credit associations (ROSCAs) are traditionally embedded in the Zimbabwean culture and historically approved, thus solving problems of moral hazard and adverse selection at very low costs. Formal group lending schemes should therefore find their economic niche and target those poor that are not

accessed by informal solutions. (JEL C34, C81, G29, O16, O17, O18, O19)

Keywords: Endogeneity; Selection bias; Microcredit; Income poverty; Impact analysis; Zimbabwe; ROSCA

Page 2: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

1. Introduction

Over the last decades, microfinance has been promoted as an important tool for

poverty reduction and the achievement of the millennium development goals. Access to

affordable credit is crucial for hundreds of millions of people in poor economies who face low

labor productivity on the one hand and high returns to capital on the other. Former United

Nations Secretary General Kofi Annan, when launching the International Year of Microcredit

in 2005, pointed out that sustainable access to microfinance helps alleviate poverty by

generating incomes, creating jobs, allowing children to go to school, enabling families to

obtain health care and empowering people to make the choices that best serve their needs.

Moreover, intangible assets like confidence, self-esteem and dignity, particularly of women,

can be developed through microlending practices. However, recent impact evaluations have

found that the picture is not always that rosy. Although microfinance is found to have rather

positive than non-existent or negative effects in most studies, consensus exists among

scholars that microcredits are no panacea for poverty reduction and its related development

challenges. Particularly since most microfinance institutions (MFI) are not self-sufficient but

depend on substantial amounts of external subsidies, a closer look has to be taken whether the

money invested really leads to significant welfare gains.

This paper aims to re-examine the impact of microfinance on household per-capita

income in Zimbabwe, using two-wave 1997-1999 panel data on clients and non-clients of the

MFI Zambuko. The data was collected by USAID’s Assessing the Impact of Microenterprise

Services (AIMS) Project in order to facilitate an impact assessment of USAID’s aid flows to

Zimbabwe’s largest MFI. In their internal impact assessment (Barnes et al. 2001) the authors

found that there have been positive impacts of the participation in Zambuko on household

durable assets, as well as on the value of funeral-related assistance to non-household members.

From these limited positive findings they derive the overall conclusion that

“The results from the impact analysis of the survey data, which controlled for specific, initial

differences, make a strong case that Zambuko’s program has a positive impact on its clients.”

(Barnes et al. 2001)

In our paper we will have a second look at the same data set, trying to assess the impact of

microcredits on household welfare by using a more extended set of tools than Barnes et al.

(2001).

Page 3: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

The most substantial methodological problem of any quantitative income evaluation is

selection bias. In the context of microfinance evaluations it can arise from two different

sources, non-random program placement and non-random participation. The former does not

play a major role in our study. Most urban Zimbabwean poor can choose from various MFIs

run by NGOs, commercial banks or the government. Branch offices of Zambuko exist in all

major towns and cities. The latter source of bias arises when individuals self-select into

treatment, that is, those deciding to take a microloan might differ from the non-clients with

regard to individual, unobservable characteristics, which might also influence the outcome

under consideration. Several ways of dealing with this problem are offered in the literature.

This paper applies both simple and more sophisticated methods to control for selection

bias, thus aiming to increase the validity of the results. Making use of the panel structure of

the data, we investigate the impact of microcredits with the help of simple difference-in-

difference estimations (supported by propensity score matching), as well as multivariate

cross-sectional and first-difference estimations. Instrumental variable (IV) techniques are then

used in order to correct for potential remaining time-variant selection bias. Besides using a

dummy variable indicating treatment, we also use a continuous credit variable. This dose-

response design, given the information about the credit history of all new and additional

repeat clients, is a big advantage of this data set. It is like having several control groups that

differ in the uptake of the treatment, allowing for a more detailed analysis than a simple

dummy treatment variable. Furthermore it allows us to control for group contamination if

treated individuals drop out or non-treated enter treatment. Last but not least we are able to

shed some light on the potential relationship between MFIs and rotating savings and credit

associations (ROSCAs). The so-called ma-rounds, traditional indigenous savings clubs, play a

crucial role in understanding the Zimbabwean savings and credit market like in most African

economies. Club membership is common among approximately half of all Zambuko clients in

our sample. This substitute relationship has been largely ignored by former studies evaluating

the impact of microcredits.

Contrary to the prior internal impact evaluation conducted by AIMS we find that

microloans did not have positive welfare effects. Results rather seem to suggest a negative

impact of Zambuko credits on household per-capita incomes. Although clients who took their

first loan with Zambuko in 1997 were, ceteris paribus, significantly richer than matched non-

clients shortly after first credits were issued out, clients’ incomes grew at much lower rates

thereafter. Considering all clients (those first borrowing in 1997 (new clients), as well as those

who by then had already been taking loans from Zambuko (repeat clients)) and all non-clients

Page 4: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

over the observed period 1997-1999, it looks as if the higher the borrowing between the two

years, the lower was income growth.

This paper is organized as follows. Section 2 gives a short overview over the related literature.

Section 3 describes the data and methodology used for the analysis. In section 4 the results

obtained are interpreted, and possible explanations are given. Section 5 concludes.

2. Related Literature

2.1 The Impact of Microfinance

Mixed evidence is found when looking at the welfare impact of microfinance on the

poor. Since the introduction of statistical methods able to capture potential selection bias in

the evaluation of microfinance during the 1990, no study has shown strong robust impacts yet.

One thread of the literature finds limited positive effects on poverty reduction and argues in

favor of microfinance while trying to isolate crucial determinants of successful lending

programs. A second thread warns that microcredits often only benefit the relatively richer and

more dynamic poor while leaving the most deprived households indebted and even more

vulnerable than before. Whereas there is considerable evidence from rural microfinance

programs, urban studies still remain scarce.

Pioneering the positive thread of the literature, Pitt and Khandker (1998) find that borrowing

from group-lending schemes of three different MFIs in Bangladesh increased consumption of

poor households. Increases were higher for female than for male borrowers. Yet, as Murdoch

(1998) argued, Pitt and Khandker’s results might in part reflect selection bias rather than real

program impact. He re-examines the same data and confirms that microfinance had in fact a

poverty reducing impact in Bangladesh. However, he also addresses the limitations of

borrowing by showing that programs only rarely generated additional employment for others

and that success had been limited in regions with highly seasonal income patterns and low

population densities. McKernan (2002) finds that besides the positive impact credits had in

Bangladesh, noncredit aspects of the three MFIs also led to strong income effects. Khandker

(2005) expands the Bangladeshi data set to a two-wave panel and finds that credit-induced

welfare gains can be seen not only on the clients’ household level but also on the village level

due to positive spillover effects.

As microcredits have spread across borders at a rapid pace over the last two decades,

there are now impact evaluation studies available from all continents. Using panel data from a

Peruvian MFI, Tedeschi (2008) finds positive effects from microborrowing on enterprise

revenues. Vogelgesang (2001) shows for Caja los Andes in Bolivia that clients with a higher

Page 5: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

number and a higher average size of prior loans presented relatively higher growth rates.

Patten et al. (2001) shows for Bank Rakyat Indonesia that well-designed microlending and –

saving schemes can work even in times of crisis. Moderate positive effects were also found in

Mosley’s study (1996) of Bancosol in Bolivia where about 25 percent of the

microentrepreneurs in any cohort showed spectacular gains to borrowing, 60 to 65 percent

stayed about the same level, and 10 to 15 percent went bankrupt.

However, the microcredit hype has not always led to positive outcomes. Several lines

of critique can be distinguished. Gulli (1998) argues that credit is not always the main binding

factor for growth of microenterprises since poor people demand a wider range of financial,

business development and social services in order to meet their individual business and

household purposes. Moreover, it has been argued by several studies that microfinance does

not reach the very poor (Hulme and Mosley 1998; Coleman 2006; Kondo et al. 2008; Niño-

Zarazúa 2007). These studies mainly argue that microfinance rather serves the “active” or

“dynamic” poor, while the poorest households are often not reached by MFIs. Results of this

group of studies suggest both positive and negative impacts of borrowing, mainly depending

on a household’s relative income position and its “entrepreneurial spirit”. Closely related are

different impacts for different lending forms. Niño-Zarazúa (2007) finds for Mexico that

group-lending schemes reduced the poverty gap, but had no impact on the incidence of

poverty. In contrast, individual lending significantly reduced moderate poverty, but had no

effect on extreme poverty.1 Coleman (2006) points out the importance of minimum efficient

loan sizes. Analyzing two microfinance programs in Thailand, he observes high drop-out rates

of women who cited that loans were too small to make any significant investment that could

have significantly improved their incomes. Copestake et al. (2000), using a pipeline designed

data set for urban Zambia, find that fifty percent of the clients dropped out after their first loan.

Only those who could secure a second experienced a positive impact from borrowing on

household income and enterprise profits.

As the discussion on the effectiveness of microfinance for poverty reduction continues,

most authors nowadays admit that microcredit can no longer be regarded as a panacea for

poverty reduction. It is crucial that programs are well embedded in varying cultural and

institutional contexts. What works in a specific cultural and economic setting does not

necessarily have to work when the settings change (see for example Aghion and Morduch,

2005).

1 It is a common finding that the poorest of the poor are more likely to take group loans. “Richer” poor are able

to offer at least some collateral and are thus more likely to secure an individual loan.

Page 6: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

3. Empirical Analysis

3.1 Data and Variables

We use urban Zimbabwean panel survey data collected by USAID’s AIMS project.

Two-wave longitudinal household and enterprise surveys designed for the evaluation of

microfinance impacts were conducted in 1997 and 1999.2 A quasi-experimental survey design

was adopted. Interviews were conducted with a randomly selected sample of

microentrepreneurs who were clients of Zambuko in 1997 in three geographic areas: Greater

Harare (Harare and Chitungwiza), Bulawayo and Mutare. These cities represent a large share

of Zimbabwe’s urban population and contained about 60 percent of all Zambuko clients in

May 1997. In order to set up a counterfactual for the evaluation, a random sample of non-

client microentrepreneurs who met Zambuko’s basic eligibility criteria was selected and

interviewed. Non-clients were matched to clients according to geographic community, gender

and enterprise sector. In the 1997 baseline survey, a total number of 691 microentrepreneurs

were interviewed: 393 clients and 298 non-clients.3 More than three-fourth were women. In

the 1999 follow-up survey, 87 percent of the 1997 sample could be re-interviewed: 344

clients and 255 non-clients.

The NGO Zambuko started in 1992 and opened branch offices in all of Zimbabwe’s

major towns and cities. It thus developed Zimbabwe’s largest MFI by the end of the 1990s.

Zambuko’s microfinance program emphasizes on microcredits, supplemented by business

management training. It offers group co-guaranteed loans as well as individual loans backed

by guarantors. To be eligible for a Zambuko loan, microentrepreneurs have to meet the

following criteria: a) no other credits received before from a formal organization for their

enterprise, b) be the sole or joint owner of their enterprise that has to be at least six months

old, and c) not be employed elsewhere on a full-time basis. Most loans have a maturity of six

to twelve months and are repaid on a monthly basis. Zambuko’s interest rates follow market

interest rates, i.e. the 1999 annual interest rate was 48 percent. Loans received in 1997 equal

on average half a client’s monthly household income. According to clients’ statements,

around 80 percent of the loans were used for enterprise purposes, mainly for purchasing

2 This period was just before the emergence of the economic crisis in Zimbabwe. Nevertheless, the beginning of

a high inflation period from 1999 and the HIV/AIDS crisis need to be controlled for in our analysis. 3 About half of the microenterprises in the study are in manufacturing. They are primarily focused on knitting

sewing and crocheting products for sale. Forty percent of the enterprises are in trade and less than ten percent are engaged in services, agriculture and food preparation.

Page 7: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

inputs or stock in bulk. While default in general is a minor problem, repayment on time was

increased by the introduction of stronger enforcement mechanisms by 1999.

Table 1 shows descriptive statistics of the most important demographic and socio-

economic variables of our sample. The sample can be split up into three major groups: new

clients, who received their first credit in the first half of 1997; repeat clients, who have been

receiving credits from Zambuko before 1997; and non-clients, who have never received a

credit from a formal MFI. The first two groups will serve as treatment groups and the latter

will be the control group in the first part of our analysis. It can be seen that new and non-

clients are very similar in their average household per-capita incomes and in most other

variables in the initial year 1997. However, new clients seem to be slightly more educated,

tend to live in larger households and have higher wage incomes. In contrast to new and non-

clients, repeat clients are the richest of the three groups. They also present the highest levels

of schooling, live in the biggest households, and earn the highest wage incomes. Compared to

the new clients, who mostly start with group loans in Zambuko (62 percent), 84 percent of the

repeat clients receive individual loans in 1997. In their main economic activities, all three

groups are very similar in that about 50 percent are active in manufacturing and another 40

percent in trade and services.

We further split up the samples of new and repeat clients into those who in 1999

continue to be clients in Zambuko (cont) and those who departed from the MFI (depart).

Interestingly, among the new clients there is no big difference between these groups. If there

was a difference at all, then we might see that program drop-outs were slightly richer in 1997.

However, when looking at repeat clients, we see a completely different picture. Those who

left Zambuko between 1997 and 1999 were on average much poorer than the continuing

clients already in 1997. They had borrowed almost the same total amount of credit from

Zambuko by then. Interestingly, there is a significant difference between continuing and

departing clients concerning ROSCA participation. While only 25 percent of the departing

repeat clients were members of a savings club in 1997, 67 percent of their continuing

counterparts participated in such a club in the same year. Then, after leaving the MFI,

departing clients strongly increased their ROSCA membership, suggesting that MFIs and

savings clubs function as substitutes for credit.

By 1999, about 60 percent of all new clients and almost 50 percent of all repeat clients

had left Zambuko. New clients in general have seen an average real per-capita income

increase of about +4 percent between 1997 and 1999. This effect was about the same among

continuing and departing clients. However, non clients, who represent our control group,

Page 8: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

experienced an average per-capita income growth of +15 percent, almost three times as much

as new clients. In contrast, repeat clients experienced high income losses of around -9 percent

over the two years. Table 1 also reveals that repeat clients experienced a strong increase in

demographic burden. Dependency ratios rose by 30 percent for continuing and by 50 percent

for departing repeat clients. Considering the older age of repeat clients and their relatively

high ratio of financial shocks due to illness and death already before 1997, this might simply

be an age and mortality effect. Besides that, it is likely that richer repeat households “adopted”

friends or relatives who were sick or in need for help. Regarding ROSCA membership, it is

interesting to see that departing and non-clients strongly increased their membership in these

savings clubs between 1997 and 1999, which was not the case for the continuing clients. This,

in turn, indicates that it was not the case that credit demand decreased in hard times. Rather it

seemed that formal credit was substituted by informal credit. It needs to be seen whether

relative returns to formal credits can explain this finding.

Table 1 about here

In the following, the most important variables used in the analysis will briefly be

introduced. The dependent variable used throughout the main analysis is the logarithm of real

per-capita household income. Household income here comprises all sources of income,

including enterprise net revenue, as well as earnings from employment. Although we are

aware of the shortcomings of the income measure, it still seems to us the best available proxy

of overall household welfare. Two variables are used to capture the effect of microlending.

One is “treatment”, a dummy determining whether a household has taken a loan with

Zambuko or not. Depending on the analysis, client status will be further differentiated with

regard to new and repeat clients. In addition to that we also use the continuous treatment

variable “creditstock”, which is an accumulative measure of all credits ever taken from

Zambuko by a household. Moreover, a number of covariates influencing household income

are used in the analysis. The variables “age” and “sex” specify age and gender of the

household head who is deemed to be decisive for the level of total household income.

Moreover, “edu” is the maximum years of schooling among head and spouse. The variable

“dep_ratio” measures the ratio of economically dependent to economically active persons in a

household. Together with “shockilldead” which is a dummy taking the value 1 if there has

been a serious illness or a death in the household causing financial repercussions during the

past 24 months, it will control for particular effects of HIV/AIDS. “lninc_wage” is the

Page 9: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

logarithm of real household monthly net wage/salary income. Being prone to endogeneity

with our dependent variable, it will only enter the anaylsis in the first lag. “sim” is the inverse

simpson index, indicating the diversity of income sources of a household. It is affected by the

number of income sources, as well as how evenly income is spread across these sources.

Values range from 1 to max. number of income sources. If all sources do not contribute

equally to total income, then the index is less than the total number of income sources.

“tradeservice1”, “foodagric1” and “manufac” are sector dummies. Trade is combined with

service, and foodindustry with agriculture due to the low share of service, food, and

agriculture. The variable “ma_rounds” is a dummy controlling for ROSCA membership.

“chi”, “kad”, “hwa”, “bul” , “mut” and “har” are city dummies.

3.2 The selection problem

The effect of microcredits on any outcome variable should be the same for all clients

and non-clients if we were able to control for different demographic and socio-economic

variables, either through matching techniques or multivariate regression. However, if those

who are taking credits from Zambuko self-selected into the microcredit program due to some

unobserved characteristics (motivation, ability) and, if these unobserved characteristics also

determine our outcomes, then selection bias arises in a simple OLS estimation. For example,

if all high able microentrepreneurs apply for a loan and all less able do not, then the effect of

microfinance is very likely to be overestimated by simple mean comparison or OLS.

To formalize this, consider a linear regression of the form

�� = � + �� + �� + �� (1)

with Pi being the treatment status of individual i (i.e. Pi = 0 if no treatment occurred, and Pi =

1 if treatment was received), Oi the outcome indicator under consideration and xi a set of

control variables. The presence of bias manifests itself in a correlation between the treatment

status Pi and the error term ��. This means that the programme influence cannot be reliably

estimated.

Bias-causing unobservable factors can be either time-invariant or time-variant. Appropriate

solutions would be differencing over time for the former and IV procedures for the latter.

Several authors have shown that a non-consideration of the selection problem leads to

substantial bias in the treatment effect (see for example McKernan 2002, Tedeschi 2008). A

number of standard tools is available for dealing with selection bias. In the following we will

briefly outline the methods used in this paper.

Page 10: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

3.3 Difference-in-difference estimation

Since we have access to longitudinal data, we are able to apply double differencing techniques.

The method of difference-in-difference (DID) is a possible, yet data intensive way of getting

rid of the unobserved heterogeneity causing selection bias, as long as this heterogeneity can

be assumed to be invariant over time. The simplest form of difference-in-difference involves

the comparison of averaged before-after outcomes for the treatment and the control group4.

The average treatment effect is estimated in equation (2):

)(1

)(1

11

before

l

after

l

m

l

before

k

after

k

n

k

OOm

OOn

I −−−= ∑∑==

(2)

Here, n and m denote the total number of individuals in the treatment group and control group,

respectively.

The precision of DID estimates largely depends on the comparability of the treatment and the

control group. Thus, matching procedures are crucial for a reliable DID estimate. As

mentioned above, within the AIMS survey design non-clients were matched to clients

according to gender, enterprise sector and geographic location. Moreover, they were required

to fulfill Zambuko’s basic eligibility criteria. The results are a first good match between

treatment and control group in the sample and allow us to estimate the DID. However, these

matches are not very likely to provide highest matching precision since matched

microentrepreneurs can still suffer from important differences in many demographic and/or

socio-economic variables. Therefore, we decided to increase matching quality, using

propensity score matching.

Propensity score matching

Propensity score techniques try to match untreated and treated individuals on a set of

indicators deemed to influence the outcome under investigation. An obvious problem here is

the multidimensionality of the matching problem, as the set of indicators grows large. A

solution to this is the so-called propensity score matching, which reduces the problem to one

4 It has to be noted that the first wave of the panel used here is not a true baseline in the sense that at the

time of the survey new clients (with their first Zambuko loan in 1997) had already received their first loan on

average 4 months ago. Any immediately appearing short-term effects will therefore not be captured by the DID

analysis. Later on this problem will be dealt with by using the continuous variable “creditstock”, rather than a

dummy indicator.

Page 11: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

dimension, namely the so-called propensity score (Rosenbaum and Rubin, 1983).

The propensity score PSi can be interpreted as an estimate of individual i’s probability of

receiving treatment. It is usually estimated using a logistic regression function of the form:

� � = �����(����) (3)

with xi denoting a vector of covariates that determine the probability of receiving treatment.

In a second step, untreated individuals are matched to the treated ones on the basis of the

estimated propensity score. There are a number of possible matching algorithms. The one

used in this paper is the so-called nearest neighbor matching. For each observation from the

treatment group we find five individuals from the control group who are closest with regard to

the propensity score. Then the average of these five observations is used as a counterfactual

for the treated individual. With these matched pairs the DID method described earlier is

repeated.

A noticeable drawback of propensity score matching is its reliance on the so-called

assumption of unconfoundedness:

O┴P|X

meaning that it is assumed that all relevant differences between treated and non-treated

individuals are captured by the covariates X, and that therefore assignment to treatment is not

influenced by further, unaccounted for covariates. Needless to say that this is a rather strict

assumption.

Difference-in-difference results in Table 2 can provide a first indication of the direction of the

overall impact of borrowing on average household welfare. Using different matching types,

the first three columns give DID-estimates for a number of selected outcomes, comparing new

with non-clients.5 We find that 36 out of 51 coefficients have a negative sign. Significant

negative effects are found for school enrollment, particularly of women, housing

investment, savings increase, household assets and enterprise net revenue. In contrast,

significant positive effects could only be found for the possession of a fan and the

consumption of meat/chicken/fish.

Table 2 about here

When comparing initially richer repeat clients with non-clients or new clients we find

5 Outcome variables were considered when an effect was suggested by economic theory. Outcomes analyzed but

not presented here showed no significant effect. They can be obtained from the authors upon request.

Page 12: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

rather mixed effects. Compared to non-clients, repeat clients had significantly larger average

gains in female enrollment, in the consumption of meat/chicken/fish and in the possession of

fans. Yet, they had significantly lower housing investment than non-clients. Compared to new

clients, repeat clients had significantly higher gains in total household assets but significantly

lower gains in TV possession. We believe that comparisons in column 4 and 5 are not as

straightforward due to different spending patterns of initially richer repeat clients. They are

more likely to purchase more expensive household assets and enterprise fixed assets than new

or non-clients.

3.4 IV-Approach to control for selectivity bias

The above procedures are rather simple to quantify but can lead to biased estimates

when either self-selection is not only due to time-invariant unobserved heterogeneity or when

matching procedures are not effective. A good matching procedure can reduce selection bias

to a certain extent but it cannot fully correct for this bias. In this case, instrumental variables

present a possible alternative. The bias-causing correlation between the regressor and the error

term shown in equation (1) is overcome by isolating and using only that part of the

endogenous regressor, which is not influenced by the investigated outcome. In order to

successfully implement the IV-approach, suitable instruments need to be available. In our

case, an appropriate instrument would be a variable, which highly influences the participation

in a microfinance programme, while at the same time being uncorrelated with household per

capita income. Assuming the existence of a vector z of suitable instruments, the unbiased

treatment effect can then be determined using a two-stage-least-squares estimation of the form

�. �� = ��� + ���� + �� (4)

��. �� = � + �� + ��� + �� (5)

Instruments that have been proposed by earlier studies were related to eligibility

criteria (Pitt/Khandker, 1998) or to the costs associated with receiving treatment (Niño-

Zarazúa, 2007). Eligibility criteria are not a feasible option in our study since all individuals

in our sample were selected such that they meet the eligibility criteria of Zambuko. Thus we

focus on transaction costs related to micro credits. A suitable candidate is the presence of a

savings account. It has to be noted here that, unlike with many other MFIs, it is not mandatory

to leave a deposit in order to receive a credit from Zambuko. Also, many non-clients in the

sample possess formal savings accounts, yet not as much as the clients. As is always the

problem with instrumental variables, the question of true exogeneity of the chosen instrument

persists and cannot be fully answered. Since an influence of savings accounts on our

Page 13: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

dependent variable, household per capita income, does not seem totally implausible, we try to

reduce this possible endogeneity by lagging the instrument. Although this obviously does not

solve the problem for the 1997 analysis, it should improve results for the 1999 data. Since we

were only able to identify one suitable instrument for the cross-section analysis, we cannot

rely on any overidentification test, such as the Sargan or Hansen-test. However, for the panel

analysis we are able to use the first lag of the endogenous regressor (in this case the

continuous treatment variable credit stock) as an additional instrument, and the results of the

Hansen-test suggest an exogenous set of instruments. There seem to be few doubts regarding

the relevance of our instrument, with the F- statistic always well above 30 (see Tables 3 and

4). It seems that individuals that are more familiar with bank institutions through formal

savings schemes are more likely to receive information about formal micro credit

opportunities.

One can increase efficiency of the estimation when exploiting the non-linearity of the

treatment equation combined with full maximum likelihood. Our main interest is still in the

outcome equation:

�� = � + �� + �� + �� (6)

Now the binary treatment variable is derived from an unobservable latent variable:

��∗ = ��� + �, (7)

where the vector ci contains the same set of explanatory variables as xi above plus our

instrument which help to identify the treatment effect in the two equation system. The

decision to obtain a credit is made according to the rule:

�� = ! 1, if ��∗ > 0 0, otherwise/ (8)

Treatment is received only if �� = 1. The errors of the two equations �� and � are assumed to

be bivariate normal with mean zero and variance-covariance matrix:

01 22 13 (9)

In contrast to the above standard IV-model, normality is not only assumed in the outcome

equation, but also in the treatment equation. This property yields an additional exclusion

restriction and makes, the correct specification of the treatment equation assumed, this

estimator more efficient than the standard IV. However, in the presence of heteroskedasticity,

using robust standard errors might reduce efficiency (Wooldridge 2002, 625). We can

estimate the model by full maximum likelihood. Therefore we derive the following joint

densities:

Page 14: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

4(�, � = 1) = 5 6(� − � + �8, ) 9 :;<= (10)

4(�, � = 0) = 5 6(� − � + �8, ) 9 =:; (11)

The likelihood function to be maximized is then:

>(?, @, ) = ∏[4(�, � = 1)]D [4(�, � = 0)]�<D (12)

Model with censored continuous treatment variable (creditstock)

The methods described above were based on the use of a binary treatment variable. In

a next step, we try to take all available credit information into account and use a continuous

treatment variable instead of the treatment dummy. We use the total amount of credit ever

received from Zambuko by a household and call it creditstock. The argument is that

households who have been taking credits over several periods should show stronger treatment

effects than those who have just started borrowing or who have never taken a credit (see also

Pitt and Khandker, 1998). This method also enables us to best account for the high drop-out

rates of both new and repeat clients between 1997 and 1999 in our sample.

3.5 Panel Regressions

Extending the above DID framework to a multivariate regression framework naturally

leads to panel regressions applying fixed effects (FE) or, in the case of two periods equivalent,

first differences (FD) estimation. Besides the possibility of controlling for unobserved, time-

invariant heterogeneity on the household and regional level, panel data further enables us to

analyze dynamic effects of borrowing on household outcomes. Yet, panel first differences

estimation alone can lead to inconsistent estimates when unobserved characteristics vary over

time. We can solve this problem by applying the instrumental variable approach discussed

earlier. Using the continuous variable “creditstock” as the indicator of treatment, the equation

to be estimated is

��E = �8�E + ?F�E + ��E (13)

Differencing this equation at two points in time yields:

∆��E = �∆8�E + ?∆F�E + ∆��E (14)

The differenced equations can then be estimated via two-stage least squares employing

appropriate instruments to control for endogenous program participation.

Page 15: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

3.6 Regression Results

Table 3 and 4 contain the main results of our regression analysis. While Table 3

estimates over new and non-clients only, Table 4 additionally includes repeat clients into the

sample. The first three columns of Table 3 present estimation results for equation (1), (5) and

(6) estimated by OLS, two-stage least squares, and full maximum likelihood respectively.

Both variants of the treatment variable are used in different models, the binary treatment

dummy and the continuous creditstock. OLS results indicate that new clients, ceteris paribus,

have significantly higher per-capita incomes in 1997 than non-clients. This might be either

due to high initial income increases right after credits were issued out or to selection bias

favoring more able clients to become part of the MFI. As soon as we control for potential self-

selection in columns 2 and 3 with the help of our instruments, the effect of credit is still

positive but no longer significant. When we take a look at the same regressions for 1999

(columns 4-6), we find that now the OLS treatment coefficients cease to be significant, too.

Moreover, coefficient magnitudes have all declined by 1999. This gives us a first hint that

new clients at least did not improve their income situation compared to non-clients over the

two years. This suggestion is confirmed when looking at the panel regression results depicted

in the last two columns of Table 3. These are the results of the estimation of the first-

difference model of equation (14), the first being a simple linear FD and the second a two-

stage instrumental variable model. We find a negative significant impact of the amount of

loan taken between 1997 and 1999 on real per-capita income growth. The result holds at the

10 percent significance level.6 Controlling for potential endogeneity of treatment (FD-IV)

even increases the magnitude of coefficient. Thus, it might be the case that, in contrast to their

poor performance, new clients might on average have more favorable unobserved

characteristics than non-clients.7

Table 3 about here

6 Low significance might be due to the small sample characteristic of our data set. Although the main data set is

not as small as the final sample, there is a significant amount of missing values in the data reducing the final sample. Comparing descriptive statistics of those who have full entries and those who had to be dropped because of missing values indicates no sample selection problem. 7 Overidentification tests (Hansen J-statistic, LR overid test) support the validity of the instruments used. The

LR-test of independent equations never rejects the null hypothesis that the correlation between the errors is zero which would indicate that there is no selection bias in our model. However, non-rejection of these tests is a common finding in the selection models literature and does not necessarily mean that the data is free of selection bias (e.g. Niño-Zarazúa 2007, Main and Reilly 1993).

Page 16: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

Table 4 basically confirms the results found in Table 3 by re-estimating the above

equations now including repeat clients in the sample, and using only “creditstock” as the

treatment measure. Again we see that a larger total amount of Zambuko credit between 1997

and 1999 is associated with lower income growth.

Table 4 about here

3.7 Distribution of Impacts among Income Quintiles

Results found above indicate that the average impact of Zambuko’s credits on client’s

per-capita income happened to be negative between 1997 and 1999. Now we want to take a

closer look at how these impacts were distributed between richer and poorer client’s

households. In order to provide some descriptive insights into the distribution, we examine

median values of income growth and borrowing from Zambuko of the five 1997 per-capita

income quintiles in Table 5. The table compares absolute per-capita income growth by

quintile of new clients, repeat clients and non-clients. Several tendencies can be observed in

the table. First, as expected, clients’ creditstocks increase with higher initial income. Richer

households demanded and were also granted larger amounts of credit between 1997 and 1999.

Second, income gains are distributed much more equally among non-clients than among

clients. Third, highest median per-capita income losses were experienced by the richest

quintile. Although this was the case for the fifth quintile within all three groups, we see that

the richest clients’ median losses were about 3.5 times (new clients) and 2 times (repeat

clients) as high as the richest non-clients’ median losses. Moreover we find that poorest

clients fared comparatively well.

Table 5 about here

Therefore, it seems that it was mainly the richest Zambuko clients who experienced

the highest income losses between 1997 and 1999. Bearing in mind that here we are not

controlling for other covariates nor for selection bias, the distributional statistics suggest that

being a Zambuko client led to worse performances for all income quintiles except the first.

However, it remains an unanswered question why particularly clients from the richest income

quintile in 1997 had such bad records.

Page 17: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

4. Explaining the poor performance of Zambuko

Two main explanations might help in answering the question of why Zambuko’s

performance was rather unsuccessful during the late 1990s in urban Zimbabwe. Explanation

one could be that the period was characterized by a worsening of the general socio-economic

situation. The HIV/AIDS pandemic had just reached its peak around the time of our study.

Moreover, Zimbabwe’s economy was just starting to slip into a severe downturn with a

stagnating economy and rising inflation. Thus, one could think that there might have been a

special situation which undermined MFI’s positive impact. Yet, in the course of our analysis

we are able to control for these effects. Particularly with the help of our control group of

matched non-clients, we are able to satisfactorily account for time trends. Explanation two

could lie in the nature of Zimbabwe’s financial system. Related literature suggests that

efficiency and trust in the formal MFI sector in Zimbabwe is rather low and that at the same

time a relatively more efficient system of informal rotating savings and credit associations

(ROSCA), so-called ma rounds exist, which is historically embedded in the Zimbabwean

culture.

4.1 Inflation and HIV/AIDS

One might argue that the time the data was recorded, 1997 and 1999, was marked by

two important confounding factors: Inflation and HIV/AIDS. Concerning inflation, it is true

that the late 1990s saw the beginning of the high inflationary period in Zimbabwe. While

annual inflation was only 14 percent in 1997, it soon jumped to 32 percent in 1998 and 70

percent in 1999. As this might have influenced the way of doing business for

microentrepreneurs and microcredit clients, a few points should be considered. First of all,

although inflation was quite high, national accounts as well as our own data show that real

per-capita incomes did not deteriorate significantly between 1997 and 1999. 8 The real

economic slump started after 1999. Second, the control group in our data set did not receive a

credit but experienced the same macroeconomic effects and thus serves as a valid control for

potential negative time trends. And third, a general increase in prices works in favor of loan

recipients as long as interest rates are not increased accordingly. In fact, in the case of

Zambuko interest rate adjustments did not keep pace with inflation due to slow decision

8 While GDP growth began to be slightly negative, out data suggests real per-capita income losses only for the

richest clients and non-clients.

Page 18: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

processes within the MFI. Thus, high drop-out rates of Zambuko cannot be explained by

inflation directly.

HIV/AIDS, the second national burden at the time, worsened living standards in

Zimbabwe seriously during the 1990s. During the times of the survey an estimated one

quarter of the adults aged 15 to 49 were HIV/AIDS infected and approximately 1.5 million

adults and children were living with AIDS. The pandemic affected our analysis in two ways:

The scenario of a chronically ill member in a credit group reflected an additional risk for the

whole group since overall default risks increased. Second, if a member of a borrower’s

household suffers from a chronic disease and eventually ceases to work this puts additional

financial strain on the borrowing household making it more difficult to pay back the loan. In

the above analysis we tried to control for these circumstances by including the dependency

ratio and the shock variable. Thus, it seems that the general socio-economic environment

cannot explain the negative performances of clients.

4.2 MFI efficiency and financial alternatives for microentrepreneurs in

urban Zimbabwe

What is the return of microcredits relative to other forms of credit available? What are

other potential sources that poor households can revert to in order to start or expand a

business? Considering different financial options for urban microentrepreneurs in Zimbabwe

we encounter four different sources of credit. First and widely used, borrowers will turn to

their nearest social network and borrow from friends and/or relatives. Second, they might

approach informal money lenders, in Zimbabwe known as chimbadzo. However, chimbadzo

do not represent a popular credit source since they charge interest rates of often more than 100

percent per month and are perceived immoral or even dangerous. Then of course, poor

entrepreneurs can apply for a group or individual credit at a MFI. Although several programs

were established during the 1990s, it seems that their outreach has been rather small. This in

turn could be due to the large popularity of the fourth option, informal savings and credit

clubs, so-called ma rounds (“the rounds”), which also serve as a source of credit. In these

clubs typically an informal group of seven members pool their savings and the fund then

rotates from one woman to the next.9 Chamlee-Wright (2002 and 2005), conducting 150 in-

9 A typical weekly rounds club consists of seven traders each paying Z$450 per week. This provides an early

recipient with a loan of Z$3,150 (well-established members usually get loans first), which is even higher then the median loan of a Zambuko client in 1999 (Z$2,500) who joined the bank two years ago. In contrast to Zambuko,

Page 19: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

depth interviews, most of them with female traders in the Mbare market area in 1999, finds

that in the urban Zimbabwean context, formal group lending programs have failed to attract

the vast majority of informal entrepreneurs. This is in line with other related literature

(Corsepius 1988, Raftopoulos and Lacoste 2001, O’Reilly 1996) which shows that informal

savings associations among the marginalized African population have a long tradition in

Zimbabwe. Similar ROSCA systems are very popular in Sub-Saharan Africa (see e.g.

Gugerty (2007) for Kenya, van den Brink and Chavas (1997) for Cameroon, O’Reilly (1996)

for Zambia). Chamlee-Wright (2005) argues that MFI group loans present a poor cultural fit

for the Zimbabwean context. In contrast, savings clubs evolved historically and are

consequently well embedded in the Zimbabwean culture. Through successful repetitions

(often without interruption) of rounds, clubs have gained wide acceptance in the sense that

reliable relationships can be expected. Informational costs typically arising from moral hazard

and adverse selection are reduced through effective forms of informal monitoring and

enforcement. Since clubs exclusively focus on the collection and rotation of savings,

transaction costs are low and once a ma round is established, there is no need for members to

meet very often. In contrast, taking a group loan from a MFI can be much more costly. It is

usually associated with application costs, membership costs, costs due to regular group

meetings and higher interest rates. In addition, ma rounds usually pay a higher tribute to age,

social status and privacy of their members than externally financed MFIs.

Besides economic incentives to join a ROSCA there seem to be also substantial social

benefits. Corsepius (1988, p. 33) shows that already in the 1980s the main reason for being

part of a savings club in Zimbabwe was not so much because of limitations in the formal

credit market but more because of social reasons. Gugerty (2007), analyzing 70 ROSCAs in

western Kenya argues that while saving requires self-discipline, ROSCAs can provide a

collective mechanism for individual self-control. Strong social ties within ROSCAs provide

high social pressure to avoid default or non-contribution. In contrast, MFI credit groups are

often a rather loose and/or anonymous combination of individuals. If a MFI client defaults on

a credit, she will not lose her face in front of her close friends. In the worst case, she will lose

the possibility to take another credit from this MFI. Thus, the relative incentive to save might

be higher in a ROSCA. This can in turn raise the motivation to lead a profitable business.

Concerning the efficiency of MFI programs in Zimbabwe, Chamlee-Wright (2005)

criticizes the lack of market orientation of formal lending programs like Zambuko. Since they

are financed by foreign donors, MFIs do not constantly orient their services at their client’s

ma round loans have to be paid back within 7 weeks (Zambuko: 6-12 months at an annual interest rate of 48 percent in 1999). The average exchange rate for Chamlee-Wright’s study was Z$38/US$1.

Page 20: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

needs. A particular concern is that trainings offered by MFIs are often too technical and one-

sided in that they focus only on skills transmission of a very small number of products or

industries. Besides that, entrepreneurial skills, which are most important for business success,

such as innovation and market discovery skills are often neglected, thus alienating clients

from market realities. Consequences are saturated markets and economically unsustainable

enterprises. Thus, it is not surprising that only 6 out of 150 women in her sample were clients

of a MFI. In contrast, 51 percent of the interviewed women started their business in part or

entirely from personal savings, including ma rounds. The second most important loan source

for starting a business was financial help from friends or relatives (38 percent).10

5. Conclusions

In contrast to the AIMS internal impact evaluation study, our analysis provides evidence that

formal microcredits from the largest Zimbabwean MFI, Zambuko, had on average negative

effects on household welfare. Using different matching techniques, double differencing

results show for several outcome variables that average impact effects in most dimensions of

welfare were rather negative than positive. More sophisticated econometrical techniques are

then applied on the variable household per-capita income which is deemed to proxy overall

household well-being best. Controlling for a set of individual and household variables,

general socio-economic trends, and non-random program participation, we find negative

significant effects of borrowing from Zambuko on household per-capita income between

1997 and 1999 in urban Zimbabwe. Estimating over new clients and non clients, the higher

the borrowing between the two years, the lower was income growth. These results are

confirmed when adding repeat clients to the sample. New clients tend to self-select into

treatment due to unobserved characteristics like motivation or ability which also affect their

outcomes positively. Even without accounting for this bias, clients performed worse than non-

clients over the two years. The negative effect increases in absolute value when controlling

for self-selection. Ordering households by income quintiles, we find that the richest 20

percent of households had the worst performance among new clients and non-clients.

Over the period of our study we observe extremely high drop-out rates among Zambuko

clients. More than half of all clients left the MFI. Drop-outs strongly raised their participation

in informal savings clubs substituting formal by informal credit. Ma rounds savings clubs,

10

A variable controlling for ROSCA participation has been included into the regressions in a further step.

However, it did not show significant effects. Thus, it seems that ROSCA participation did not lead to increased incomes. But there was also no decrease as seen in the participation in MFIs.

Page 21: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

Zimbabwe’s traditional ROSCAs, might be the key to understanding the poor performance of

microfinance in Zimbabwe. These well-established informal systems of saving and lending

seem to be culturally much more anchored and accepted within the urban Zimbabwean

informal sector. Savings clubs have a long history in Zimbabwe and have therefore achieved

to offer services at lower costs. More than half of all Zambuko clients were member of a

savings club at the time of their membership with the MFI.

Which implications can be derived from our analysis? The analysis suggests that US ODA

investments into the largest Zimbabwean MFI at the end of the 1990s were allocated

suboptimal. MFIs did not help to reduce poverty and to raise the living standards of the

marginalized. Since additional qualitative assessment of other papers suggests that building

up trust is an important factor in credit relationships, the question is whether these programs

might lack trust among microentrepreneurs and what could be done about this. Future reforms

within MFIs and the factor time could improve the performance. Several other problems

within MFIs seem to keep them from offering efficient services at low costs. As long as this is

the case, ma rounds might be the predominant alternative in the field of group loans. Given

the relative efficiency of savings clubs, MFIs might be better advised to redirect their

targeting towards those poor that are excluded from existing financial systems and those who

demand further financial services like individual microloans.

Page 22: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

Bibliography

Aghion, B. and Murdoch, J. (2005). The Economics of Microfinance. Cambridge: MIT Press, Ch. 1, 2.

Barnes, C., Keogh, E., Nemarundwe, N. (2001). Microfinance Program Clients and Impact: An Assessment of Zambuko Trust, Zimbabwe. AIMS Report. USAID, Washington, D.C.

Chamlee-Wright, E (2002). Savings and Accumulation Strategies of Urban Market Women in Harare, Zimbabwe. Economic Development and Cultural Change, 50, 979–1005.

Chamlee-Wright, E (2002). Entrepreneurial Response to “Bottom-up” Development Strategies in Zimbabwe. The Review of Austrian Economics, 18(1), 5-28.

Coleman, B. (2006). Microfinance in Northeast Thailand: Who Benefits and How Much? World Development, 34(9) 1612-1638.

Copestake, J., Bhalotra, S. and Johnson, S. (2000). Assessing the Impact of Microcredit on Poverty: A Zambian Case Study. University of Bath. Occasional Paper.

Corsepius, U. (1988). Savings Mobilisation in Developing Countries. Intereconomics, Jan/Feb 1988.

Gugerty, M. K. (2007). You Can’t Save Alone: Commitment in Rotating Savings and Credit Associations in Kenya. Economic Development and Cultural Change, 55, 251–282.

Gulli, H. (1998). Microfinance and Poverty: Questioning the conventional wisdom. Inter-American Development Bank. Washington, D.C.

Heckman, J. (1978). Dummy Endogenous Variables in a Simultaneous Equation System. Econometrica, 46(4), 931-959.

Hulme, D., Mosley, P. (1998). Microenterprise finance: Is there a conflict between growth and poverty alleviation? World Development, 26(5), 783-790.

Khandker, S. (2005). Microfinance and poverty: evidence using panel data from Bangladesh. The World Bank Economic Review, 19(2): 263.

Kondo, T., Orbeta, A., Dingcong, C., and Infantado, C. (2008). Impact of Microfinance on Rural Households in the Philippines. PIDS Discussion Paper Series, No. 2008-05.

Maddala, G. S. (1983). Limited-dependent and qualitative variables in econometrics, Econometric Society Monographs No. 4, Cambridge University Press.

Main, B. and Reilly, B. (1993). The Employer Size-Wage Gap: Evidence for Britain. Economica, 60, 125-142.

McKernan, S.M. (2002). The Impact of Microcredit Programs on Self-Employment Profits: Do Noncredit Program Aspects Matter? The Review of Economics and Statistics, Vol. 84, No. 1, Pages 93-115.

Murdoch, J. (1998). Does Microfinance really help the Poor? New Evidence from Flagship Programs in Bangladesh. Harvard University Working Papers.

Niño-Zarazúa, M. (2007). The impact of credit on income poverty in urban Mexico. An endogeneity-corrected estimation. Sheffield Economic Research Paper Series, SERP Number: 2007005.

Page 23: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

O’Reilly, C. (1996). Urban Women’s Informal Savings and Credit Systems in Zambia. Development in Practice, 6(2), 165-169.

Patten, R., Rosengard, J., and Johnston, D. (2001). Microfinance Success Amidst Macroeconomic Failure: The Experience of Bank Rakyat Indonesia During the East Asian Crisis. World Development, 29(6), 1057-1069.

Pitt, M.M. and Khandker, S.R. (1998). The impact of Group-Based Credit Programs on Poor Households in Bangladesh: Does the Gender of Participants Matter? Journal of Political Economy, 106(5), 958-996.

Raftopoulos, B. and Lacoste, J.-P. (2001). Savings Mobilisation to Micro-Finance: A Historical Perspective on the Zimbabwe Savings Development Movement. Conference Paper.

Rosenbaum, P. and D. Rubin (1983). The Central Role of Propensity Score in Observational Studies for Causal Effects. Biometrika 70, 41-55.

Tedeschi G.A. (2008). Overcoming Selection Bias in Microcredit Impact Assessments: A Case Study in Peru. The Journal of Development Studies, 44(4), 504-518.

Van den Brink, R. and Chavas, J.-P. (1997). The Microeconomics of an Indigenous African Institution: The Rotating Savings and Credit Association. Economic Development and Cultural Change, 45(4), 745-772.

Vogelsang, U. (2001). The impact of microfinance loans on the clients’ enterprises: Caja los Andes, Bolivia. Gk working paper series no. 2001-03, University of Mannheim.

Wooldridge, J.M. (2002). Econometric analysis of cross-section and panel data. MIT Press, Cambridge, MA.

Page 24: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

Table 1: Summary Statistics

Variable NEW97 NON97 REPEAT97 NEW99 NON99 REPEAT99

all cont depart all cont depart all cont depart all cont depart

HH Income p.c. 776 754 792 777 1323 1810 812 806 789 818 896 1203 1626 740

Growth HH inc p.c. +4% +5% +3% +15% -9% -10% -9%

Age 40.21 39.95 40.42 39.61 43.51 43.79 43.23 42.84 41.77 43.61 43.20 44.94 45.38 44.47

Sex .54 .49 .57 .59 .61 .50 .73 .66 .64 .67 .67 .61 .57 .65

Maxeducation 8.04 8.22 7.90 7.45 8.62 8.79 8.45 8.43 8.81 8.16 7.45 8.41 8.47 8.35

HH Size 5.49 5.49 5.49 4.77 5.72 5.88 5.55 6.02 5.73 6.22 5.24 5.62 5.38 5.88

Dependency ratio 1.93 1.86 1.99 1.68 1.53 1.51 1.56 2.07 2.08 2.07 1.73 2.13 1.95 2.32

Shock (Illness, Death) .38 .37 .39 .35 .43 .43 .43 .33 .34 .33 .32 .33 .30 .37

Wage income 852 831 867 582 1225 1786 635 1032 944 1095 835 1328 1560 1074

Diversification Index 1.73 1.86 1.63 1.6 1.71 1.79 1.63 1.82 1.77 1.87 1.81 1.96 1.93 1.99

Manufacturing Sector .51 .56 .48 .53 .56 .55 .58 .50 .53 .47 .52 .63 .64 .62

Trade/Service Sector .46 .42 .48 .44 .35 .38 .33 .46 .44 .48 .43 .30 .30 .30

Food/Agric Sector .03 .02 .04 .03 .09 .07 .10 .04 .03 .04 .04 .07 .06 .07

Total stock of credit 1777 1678 1851 0 5104 5336 4860 3744 6328 1887 0 9669 14219 4695

Savings account .97 1 .95 .76 .91 .98 .85 .90 .95 .85 .82 .92 1 .84

Travels to sell .57 .61 .54 .40 .49 .52 .45 .52 .59 .46 .43 .57 .57 .56

Group loan .62 .64 .61 0 .16 .21 .14 .61 .64 .60 0 .19 .18 .21

Individual loan .38 .36 .39 0 .84 .79 .86 .38 .36 .39 0 .80 .82 .79

Savings club member .54 .56 .53 .37 .46 .67 .25 .64 .56 .70 .56 .53 .62 .44

N 138 59 79 150 82 42 40 153 64 89 166 90 47 43

Incomes are real monthly Zimbabwean dollar (ZWD) with base year 1997 and use national CPIs

Page 25: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

� 頴頴� � 省省省蘟 Ю Ю Table 2: Difference in Difference estimates

New vs. Non Clients Repeat vs. Non Repeat vs. New

DID by matching type DID DID

Impact Indicator AIMS AIMS 1:1 PSM PSM PSM

Household income

Totalhhincome -689.25 -1082 -1128 -165 448

Totalhhincome p.c. -128.53 -77 -161 -108 -40

Education

Enrollment 6-21 male -6.21 -2.25 -10.49 -14.19 -7.76

Enrollment 6-21 female -11.05 -16.96* -3.4 13.60* 7.31

Enrollment 6-21 -6.16* -6.85 1.18 -4.39 1.58

Consumption

Meat/chicken/fish (days per week) .05 .22 .61* .74* -.24

Housing investment -2556.07* -5417 -3092 -6641** 393

HH Assets

Fridge .02 .09 .002 -.08 .01

TV .05 .09 .04 -.03 -.17**

Fan .04 .08 .12* .17** -.09

Total HH Assets -4296.87 -7014* -6327** 5093 12520**

Savings Increase -.16** -.13 -.17* -.12 .01

Enterprise

Net revenue (enterprise 1) -602** -662 -1130 -1232 -1354

Net revenue (all enterprises) -534** -766 -1068 -871 -564

Fixed Assets (enterprise 1) -281 485 -717 13000 19002

Fixed Assets (all enterprises) -316 1461 -360 14771 20888

Individual Empowerment

Ability to face the future -.05 -.04 -.038 .07 .04

AIMS matching refers to all clients versus all non-client for whom the respective variable was available; AIMS 1:1 matching

compares only those matched pairs that had no missing values in either of the periods; PSM = propensity score matching

(nearest neighbor approach, n=5); Significance levels: ***/**/* denote .01, .05 and .1 respectively (t- or Wilcoxon-test);

Incomes are real monthly Zimbabwean dollar (ZWD) with base year 1997 and use national CPIs.

Page 26: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

Table 3: Treatment Regression Results I (NEW vs. NON-Clients)

LNREALPERCAPITAINCOME

Cross-sectional models Panel models

OLS97 IV97 ML-Pr97 OLS99 IV99 ML-Pr99 FD FD-IV

Treatment (dummy) .174* .314 .203 .089 .182 .023

(.052) (.258) (.416) (.351) (.602) (.930)

Treatment (creditstock) .025** .042 .011 .022 -.173* -.349*

(.041) (.258) (.355) (.602) (.058) (.064)

N 276 276 276 278 278 278 233 233

Treatment (dummy)

LR test of ind. eq. (rho=0) (p-value) .915 .792

F-statistic first stage 38.7 35.0

Overid test, Hansen/LR (p-value)

Treatment (creditstock)

F-statistic first stage 39.8 38.2 35.94

Hansen J-statistic (p-value) .442

P-Values (robust t-statistic) in parenthesis. Significance levels: ***/**/* denote .01, .05 and .1 respectively. Spatial differences are

controlled for using town dummies. Incomes are real monthly Zimbabwean dollar (ZWD) with base year 1997 and use national CPIs.

Control variables: age, age squared, sex, maximum years of schooling of head or spouse, household size, dependency ratio, shock of

illness or death of a household member, log of wage income, Simpson diversification index, sector and town dummies. Instruments:

savings account 1997, lag of lnzstockcredit (for FD-IV only).

Page 27: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

Table 4: Treatment Regression Results II (NEW&REPEAT vs. NON-Clients)

LNREALPERCAPITAINCOME

Cross-sectional models Panel models

OLS97 IV97 OLS99 IV99 FD FD-IV

Treatment (creditstock) .038*** .077** .022** .063 -.156** -.317*

(.000) (.028) (.017) (.101) (.047) (.074)

N 352 352 353 353 297 297

F-statistic first stage 36.6 32.6 51.1

Hansen J-statistic (p-value) .322

P-Values (robust t-statistic) in parenthesis. Significance levels: ***/**/* denote 0.01, 0.05 and 0.1 respectively. Spatial differences are

controlled for using town dummies. Incomes are real monthly Zimbabwean dollar (ZWD) with base year 1997 and use national CPIs.

Control variables: age, age squared, sex, maximum years of schooling of head or spouse, household size, dependency ratio, shock of illness

or death of a household member, log of wage income 1997, Simpson diversification index, sector and town dummies. Instruments: savings

account 1997, lag of lnzstockcredit (for FD-IV only).

Page 28: Did Microfinance Increase Poverty in Urban Zimbabwe? · PDF fileDid Microfinance Increase Poverty in Urban Zimbabwe? ... last two decades with positive examples from Bangladesh,

Table 5: Income growth and borrowing between 1997 and 1999

by 1997 per-capita income quintiles

Q1 Q2 Q3 Q4 Q5

Median absolute income growth, per capita (NEW CLIENTS) 191.3 73.5 137.9 24.8 -667.5

Median creditstock 1999 (NEW CLIENTS) 2000 2000 3000 2000 4250

N 23 22 22 22 22

Median absolute income growth, p. c. (REPEAT CLIENTS) 64.1 -135.3 -63.2 94.7 -366

Median creditstock 1999 (REPEAT CLIENTS) 3000 6800 7000 11300 8500

N 13 13 13 13 13

Median absolute income growth p. c. (NON-CLIENTS) 141.2 136.8 128.1 155.4 -182.2

N 25 24 25 24 24

All monetary values are in real Zimbabwean dollars (ZWD) with base year 1997 and use national CPIs. Incomes are monthly household

incomes.