the impact of micro credit on income poverty (working paper)
DESCRIPTION
This paper presents an estimation of the impact of microcredit on income poverty, following aquasi-experiment specifically designed to control for endogeneity and selection bias in the contextof urban Mexico. Although we find impacts on income poverty, the magnitude of the effect ismarginal and only significant among moderate poor. We find no evidence of impacts on extremepoverty. The evidence points to a link between poverty impacts and lending technology. Rigidscreening and monitoring devices used by group lending contracts generate a utility cost ofborrowing that undermines the efforts of poverty alleviation.TRANSCRIPT
1
The impact of microcredit on income poverty. An endogeneity-corrected estimation in urban Mexico
Miguel Niño-Zarazúa1
Summary This paper presents an estimation of the impact of microcredit on income poverty, following a quasi-experiment specifically designed to control for endogeneity and selection bias in the context of urban Mexico. Although we find impacts on income poverty, the magnitude of the effect is marginal and only significant among moderate poor. We find no evidence of impacts on extreme poverty. The evidence points to a link between poverty impacts and lending technology. Rigid screening and monitoring devices used by group lending contracts generate a utility cost of borrowing that undermines the efforts of poverty alleviation. Keywords: endogeneity; selection bias; microcredit; urban poverty; Mexico
Introduction
Microcredit has become an instrument for poverty alleviation widely used in the developing
world. The Microcredit Summit campaigners have recently reported that about 3316 microcredit
programs were reaching more than 133 million borrowers worldwide (Daley-Harris 2007), despite
the fact that the existing empirical evidence on the impact of microcredit remains inconclusive, and
methodologically contested (see Morduch and Haley 2002 for a review). Following global trends,
the Mexican government launched in 2001 the National Programme for Financing the
Microentrepreneur (PRONAFIM), a state revolving loan fund established with US $100 million to
support, through capital subsidization, the expansion of an infant microcredit sector in the country.
The role of PRONAFIM was perceived as complementary to a wider financial reform that under a
new regulatory framework organises the overall non-banking system. The number of microcredit
programs supported by PRONAFIM rapidly increased, from just eleven in 2001 (receiving US $7.9
million) to 80 by the end of 2005, recipients of more than US $26 million2.
PRONAFIM was launched under the general proposition that access to credit, per se, is a significant
determinant of increasing household income and thus, alleviating poverty. However, most
empirical studies that have tested such proposition focus on the rural context and suffer, with a few
2
exceptions, from endogeneity and selection bias, and those that have successfully controlled for
these estimation constraints (e.g. Pitt and Khandker 1998a; Coleman 1999) follow methodological
approaches that are difficult to replicate in the urban context, where many microcredit programs in
the developing world actually operate.
This paper presents an alternative methodological approach to control for endogeneity and
selection bias in the impact analysis of microcredit, using data from a quasi-experiment
operationalised in the context of urban Mexico. Our methodology, which is discussed in section
one, also allows the evaluation of potential differences between different lending technology
regarding poverty impacts, particularly when the utility cost of borrowing is included in the
analysis. The econometric estimation procedure discussed in section two is formulated to test for
the underlying assumptions of no endogeneity and selection bias. Section three examines the
proposition regarding microcredit as a significant determinant of an income rise, whereas section
four analyses poverty impacts. Section five concludes with some policy recommendations.
1. Research design
We designed a type of quasi-experiment that is often referred to as a non-equivalent, post test-only
quasi-experiment (Campbell and Stanley 1966), in which two groups of households are sampled:
treatment and control. Data was collected at the household level to address, as suggested by Hulme
(2000) in this journal, the fungibility problem. A major problem that emerges with the non-
equivalent, post test only quasi-experiment, referred hereafter as quasi-experiment, is that these two
groups may differ in important ways that influence the decision of borrowing and thus, the
outcome of interest. In other words, there might be unobservable factors related to e.g. individual
efforts, abilities, preferences and attitudes towards risk that affect the selection process and thus the
income variable. We refer to this potential problem as a demand-related bias. A fundamental
assumption here is that participation in a microcredit program is always voluntary.
Another potential selection problem emerges from the implicit nature of credit markets. Even if we
3
had a control group willing to take risks and borrow from a microcredit organization, we may still
face selectivity discrimination made by the lender or group members that screen out applicants who,
for instance, are not residents of the neighbourhood where the microcredit program operates. We
refer to this potential problem as a supply-related bias. In this sense, the selection process, iI , is
defined by two components: 1) the decision of a household of whether or not to participate in a
microcredit program ( *
1I ), and 2) the decision of lenders (or group members) of whether or not to
accept the applicant ( *
2I ) (see figure 1).
INSERT FIGURE 1 ABOUT HERE
Although we cannot observe households that choose either to participate or not, and households
that are either accepted or rejected by the lender, i.e. 1 2I I I= + , we can specify the distribution of
households that have self-selected to participate in the credit program, and have been accepted by
the lender, i.e. 1 2I I I= ⋅ , with a time-variance difference that accounts for the length of
membership. Consequently, those households who had self-selected to participate in a credit
program and had been accepted by the lender, and therefore were actively participating in the
credit program were eligible to be sampled as members of the treatment group. Similarly, those
households who had self-selected to participate in a credit program and had been accepted by the
lender, but had not received a loan by the time the quasi-experiment was conducted, were eligible
to be sampled as the members of the control group.
We also followed a geographical criterion, i.e. we operationalised the quasi-experiment among
households living in the same neighborhood, in areas with a minimum level of socio-economic
homogeneity, where the comparison between treatment and control groups was reasonable. By
following this criterion, it was possible to hold constant factors such as infrastructure, costs of
inputs, and local prices that could otherwise cause an endogeneity problem. A high population
density in poor urban areas makes possible to follow this approach. As a result, selection bias and
4
endogeneity are assumed to be controlled through the process of data collection itself. We test for
such assumptions in the econometric procedure presented in section 2.
Given the homogeneity of household characteristics, a sample survey was the preferred type of
data collection (Babbie 1990; Keppel 1991). The sampling strategy was implemented using a
multistage procedure in the form of clusters (Fink and Kosecoff 1985): first, we had access to a list of
program participants (both treatment and control) from three case-study organizations (the
clusters), and who lived in the selected areas. Participants with loan in arrears were included in the
list. In the second stage, both treatment and control groups were selected at random. The survey
was administrated face-to-face employing, as instrument of data collection, a semi-structured-
interview format, which embodied a combination of methods in a dominant-less dominant design
form, and where a simultaneous triangulation was straightforwardly implemented (Morse 1991)3.
In the end, we surveyed 148 households: 55 participating at Community Financial Services
(Fincomun) and living in San Miguel Teotongo, a neighborhood located to the eastern periphery of
Mexico City; 46 participating at Centre for the Assistance of the Microentrepreneur (CAME) and
living in the Chalco Valley, one of the most densely populated municipalities in the country located
to the eastern periphery of the Metropolitan area of Mexico City; and 47 participating at Programs
for Women (Promujer) and living in Tula City and the surrounding areas, a locality about two
hours from Mexico City. Thus, we have three locations, one for each organization.
Unlike most microcredit programs operating in Mexico, Fincomun mostly relies on individual
lending, and demands, as a result, physical collateral and guarantees as enforcement mechanisms.
CAME, on the other hand, employs a credit-only village banking approach, and consequently
exploits joint liability as enforcement device. Similarly, Promujer use a credit-plus village banking
approach that combines credit with education and training as part of the services provided (see
table 1 for more details).
INSERT TABLE 1 ABOUT HERE
5
2. The econometric estimation procedure
To begin with the impact analysis, our exposition considers the case where household i decides to
participate in a credit program. The amount of credit is exogenously determined by the lender L,
who sets up a maximum threshold according to level of program participation. The lender is
expected to exploit several screening, incentive and enforcement devices to deal with the problems
of moral hazard and adverse selection that are related to borrowers’ behavior (Hoff and Stiglitz
1990). Some of these devices are, inter alia, progressive lending, compulsory savings schemes and
periodical repayment schedules in group meetings.
Given the nature of fragmented credit markets, the demand for credit is rationed by the lender
(Stiglitz and Weiss 1981), and endogenously determined by household characteristics such as the
stock of human capital, individual preferences and attitudes towards risks. Our primary interest is
to estimate the impact of microcredit on the outcome of interest, which is observed through the
income variable, iY . Thus, we consider the following model:
i i i iY X I uβ δ= + + (1)
where i
X is a vector of household characteristics and i
I is a dichotomous variable with value
= 1I if household i is a program participant, = 0I otherwise. The model measures the impact of
program participation by the coefficient of the parameter estimate, δ . Note that the variable i
I
cannot be treated as exogenous if a problem of selection bias is anticipated. Initially, we consider a
specification equation in the form:
1 1 1 1i i i iY uβ δ += X + I (for program participants) (2)
2 2 2 2i i iY uβ= X + (for non participants) (3)
*
1 1 1 1I γ ε−= Z (4)
*
2 2 2 2I γ ε−= Z (5)
6
where iI is defined, as illustrated in figure 1, by *
1I and *
2I . However, since we only observe
household i self-selecting to participate in a credit program and being accepted by the lender, then
we can specify the distribution of households that are accepted to participate in the program ( *
2I )
and then estimate the parameter 2γ , if such households have previously self-selected ( 1 1I = ).
Maddala (1999) suggests to define *
2I over a sample population, e.g. identify households living in
the same neighborhood, and then analyze the model from the truncated sample, where the
parameters 1γ and 2γ can be estimated by maximizing a likelihood function, e.g. Probit or Tobit.
Thus, the observed i
Y is defined as 1i iY Y= if 1
i=I , and 2i i
Y Y= if 0i
=I , where the participation
decision function is given by *
i i iI Z γ ε= = . Since treatment and control groups were sampled to
satisfy condition 1 2I I I= ⋅ , i.e. both groups are participants with a time-variance difference that
accounts for the length of membership, then the following specification can be used:
1 1 1i i iY uβ= X + (for treatment group) (6)
2 2 2i i iY uβ= X + (for control group) (7)
and
*
1 2 1 2
( )1 0 ( )
( )
ii i i i i
i
ZE Y I E Y I X V
Z
φ γβ β σ
γ= − = = − + +
Φ (8)
where *
2 1( )ε εσ σ σ= − ; ( )φ ⋅ and ( )Φ ⋅ are the density of the distribution function and the
cumulative distribution function of the standard normal, respectively, and ( ) 0E V = . If we
encounter a selection problem, then * 0σ > . In other words, households with comparative
advantages will benefit more from a microcredit program than disadvantaged households.
Although the selection problem is assumed to be controlled through the process of data collection,
we need to test this assumption. In order to do so, we follow a Heckit estimation procedure
(Heckman 1979) with an identifying instrumental variable (IV)4.
7
2.1 The Heckman procedure with iI as endogenous regressor
This Maximum Likelihood method follows the model:
y
i i y i iY X I uβ δ= + + (9)
I
i i I i iI X Z uβ γ= + + (10)
where i
X is a vector of household characteristics, and i
Z , the identifying instrument. i
Z is an
observable variable distinct from those in i
X that affect iI but not the outcome of interest iY
conditional on iI . In other words, the instrument must be partially correlated with iI , i.e. the
coefficient on i
Z must be nonzero, 0γ ≠ , so ( , ) 0I
i iCov Z u ≠ , while
iZ must be uncorrelated with
iY , so ( , ) 0y
i iCov Z u = , where the projected error, E( ) 0y
iu = is uncorrelated with
iZ . Selecting an
appropriate instrument becomes a crucial and complex task for the estimation.
The Heckit procedure allows testing for the assumption of no self-selectivity by estimating the
inverse Mills ratio, ( )( )( )
φλ
⋅⋅ ≡
Φ ⋅, resulting from the relationship between the density of the
distribution function, ( )φ ⋅ , and the cumulative distribution function of the standard normal, ( )Φ ⋅ .
As suggested by Heckman (1979), we can estimate consistently the parameters I
β and γ by
exploiting the properties of the first stage Probit estimation and then get the estimated inverse Mills
ratio, λ∧
. In the second stage we obtain the parameters y
β and δ from Ordinary Least Squares
(OLS) with the inverse Mills ratio added to the regressors as follows:
y
i i y i y i iY X L I M uβ θ δ λ= + + + + (11)
We have included in (11) a vector of credit market characteristics, i
L , which captures the effects of
other credit agents such as banks, moneylenders and rotating savings and credit associations
(ROSCAS) that compete with microcredit programs (see table 12 for details about independent
8
variables). The rationale behind incorporating i
L relies on the principle that if we do not control for
the effects of such agents on iY , then the parameter δ may be inconsistent, i.e. we could wrongly
attribute some outcomes to the microcredit program when in fact come from, for example,
ROSCAS.
The two-stage Least Square (2SLS) procedure yields consistent estimates in the parameter of
interest δ (Wooldridge 2002) where M and λ are the inverse Mills ratio and its parameter
estimate, respectively. A simple way of testing for self-selectivity is under the null hypothesis of no
selection bias, 0 : 0H λ = , using the usual 2SLS t statistic. If 0λ ≠ , then the selection problem is
significant.
2.2 Selecting the instrumental variable
It is a common practice among microcredit programs to demand, as a screening device, periodical
repayment schedules, often in compulsory group meetings. Group meetings can be highly time-
intensive and potentially costly for program participants. In the beginning we considered an
observable variable with computational values that varied from household to household and
which reflected the heterogeneity of the utility cost of program participation. We computed this
variable by estimating the cost of transportation per credit cycle T
iC , that captures the geographical
characteristics of the accessibility to credit, in addition to the opportunity cost of borrowing, O
iC , as
a proxy of the income forgone per credit cycle for attending periodical group meetings. We
transformed this variable into logarithmic form, in order to test for the underlined assumptions of
no correlation between the identifying instrument and the income variable as follows:
LGINCOMEPC = β1AVEDUi + β2HOWNERi + β3HESTATEi + β4TIMEBUSi + β5WWORKERi + β6DEPENDRATIOi + β7AGEi + β8WOMANi + β7MARITALi + θ1ROSCASi + θ2FORMALCREDITi + θ3MONEYLENDERi + δLGMAXCREDITi + γLGCOSTBORROWPCi
We find that, in the case of Fincomun, the coefficient γ of LGCOSTBORROWPC reported p-values
9
of the t statistic that reject the null 0 : 0H γ = at 5% level of significance (see table 2), throwing out
any possibility of using this variable as the identifying instrument in the impact estimation for the
three institutions participants in the study as a whole5.
INSERT TABLE 2 ABOUT HERE
We also tried to derive the identifying instrument from the first component of
LGCOSTBORROWPC, i.e. the logarithm of the cost of transportation per credit cycle, log T
iC . The
argument relies on the idea that there is a correlation between program participation and
accessibility that emerges from two sources:
1) Microcredit programs impose, due to transaction costs, lending restrictions to households living
faraway from the branch. Regarding this particular issue, the Managing director of Fincomun
mentioned in an interview that a fundamental principle for the organisation was to operate in a
geographical radius that did not exceed a journey of 30 minutes walking or by public transport to
house of the borrower.
2) A process involving an individual choice, where households reporting high transaction and
opportunity costs of participation would either have high incentives to borrow the largest amount
of credit available, in order to compensate these costs, or may simply decide to drop out or not to
participate in the first place.
Our survey collected information on the cost of transportation; however there were substantial
missing values in the dataset that reflected the individual choice of walking to attend periodical
group meetings. For that reason, we decided to explore the attributes of the time dimension that
capture the information about the distance from the residence (or businesses) of program
participants to the branch, as a proxy of credit accessibility.
Data about the time that participants spent since they left home (or business) until they arrived at
the branch was computed and weighted when public transport was used in order to capture the
10
time that they would have consumed if they had walked to the branch6. We coded this identifying
instrument as DISTANCE. When the reduced form equation (11) was estimated with DISTANCE
as identifying instrument for each microcredit program, the p-values of the t statistic for the
coefficient γ rejected the null of 0 : 0H γ = , i.e. it reflected the statistically significance correlation
between accessibility and participation; however, when we included iZ in equation (11), the
parameter estimate γ accepted the null of no correlation against the income variable, iY (see table
3). As a result we were able to use DISTANCE as the identifying instrument for the Heckit
procedure.
INSERT TABLE 3 ABOUT HERE
One of the reasons for choosing the Heckit procedure is due to its structural qualities. On the one
hand, it enables us to test for the assumption of no selection bias by exploiting the non-linearity
properties of the inverse Mills ratio (coded in the regression equation as MILLS). As discussed
above in section 1, we operationalised the quasi-experiment in a way to control for selection bias;
however, we still need to test for the assumption of no selectivity problem, 0 : 0H λ = , allowing i
I
to be endogenous by using the 2SLS t statistic for ^
λ .
On the other hand, the Heckit procedure allows us to test for the quality of the instrumental
variable and the robustness of the estimation. In order to do so, the identifying instrument
DISTANCE is included in (11) alongside with the other exogenous variables, including the inverse
Mills ratio. The identification is achieved by exploiting the properties of the inverse Mills ratio that
result from the non-linear relationship of the exogenous variables in the reduced form equation
(11). After running the identification equation, the coefficients of the endogenous explanatory
variable in the estimation equations as well as the Mills ratio for each organization under study
remained stable (see table 4)7. The consistency of the results confirm the robustness of DISTANCE
as the instrumental variable, and allow us to accept the null of no self-selectivity, confirming that
11
we followed an appropriate methodological procedure during the process data collection.
INSERT TABLE 4 ABOUT HERE 3. Results from the second-stage Heckit: the impact of program participation on household
income
We have estimated i
Y as the dependent variable in (11) by employing the logarithm of income per
capita and three different definitions of income per adult equivalent. The use of adult equivalence
scales is generally justified to take into account economies of scale and intra-household resource
allocation. The first equivalence factor (IAE1) follows the approach adopted by Rothbarth (1943);
the second equivalence factor (IAE2) follows Wagstaff and van Doorslaer (1998), whereas the third
equivalence factor (IAE3) follows the OECD modified scales based on Hagenarrs et al, (1998).
Results from the income estimation equation are presented in table 5. Since the coefficient of the
inverse Mills ratio reveals no evidence of selection bias, we are able to concentrate on the OLS
estimation. The parameter estimate δ of the impact variable, i
I , reports the difference in the mean
log income per adult equivalent of treatment households relative to the control group. The slope
coefficients show, as expected, a positive sign for each of the three credit programs; however, the
coefficients are only statistically significant different from zero in the case of Fincomun.
INSERT TABLE 5 ABOUT HERE
In order to calculate the percentage change in income per adult equivalent of treatment households
relative to the control group, we take the antilog of the parameter estimate i
I and compute
( 1) 100eδ − × (Halvorsen and Palmquist 1980). For example, if we estimate the antilog of δ for the
logarithm of monthly IAE1, we obtain 0.548 1.7297e = , suggesting that ceteris paribus, the median
income per adult equivalent of treatment households at Fincomun was higher than that of the
control groups by about 73%. Surprisingly, the parameter δ is positive but not significantly
different from zero in the case of CAME and Promujer. In other words, although there might be a
positive impact of program participation on the level of income, the empirical evidence does not
12
confirm this relationship.
Note that δ reports the average impact of program participation; however, it does not take into
account the effect of borrowing over time. Treatment households with say five years of
membership are expected to report greater impacts than those households with just one or two
years of membership. This is in part due to the effects of progressive lending, an incentive device
extensively used by microcredit programs. In order to address this issue we extend the Heckman
procedure to a Tobit selection equation in section 3.1.
3.1 A Tobit selection equation: the impact of microcredit on household income
We replace the treatment dichotomous variable i
I in equation (11) by a continuous variable, i
C ,
that now measures the amount of credit borrowed during the last credit cycle. We assume that i
C
is exogenously determined by the lender L, who defines this maximum threshold according to level
of participation in the program. Thus we have the following specification equation:
* c
i i c i iC X Z uβ γ= + + (12)
where
*max(0, )i i
C C= , i.e. (13) * * if 0
i i iC C C= > (for treatment group) (14)
*0 if 0i i
C C= ≤ (for control group) (15)
and
2~ (0, )i iu X Normal σ
Consequently, i
C takes a maximum value and a lower threshold zero in the form of a censored
Tobit model (Tobin 1958) with a 0i
C > for treatment groups and 0i
C = for control groups8. In this
way we believe to capture a more precise measure of the impact of microcredit. Note that the Tobit
model implies that the probability of observing 0i
C > and 0i
C = are ( )φ ⋅ and *( 0) (0)i
p C < = Φ ,
respectively, where ( )φ ⋅ and ( )Φ ⋅ denote the density function and the cumulative density
13
function of the standard normal. These assumptions are very similar to those implied in the Heckit,
but now the log-likelihood function takes the form:
0 0
ln ln ln ln 1i i
i i c i c
C C
C X XL
β βσ φ
σ σ> =
− = − + + − Φ
∑ ∑ (16)
which generates the conditional mean function of the observed dependent variable i
C that is
censored at zero for control groups and have disturbances normally distributed, which can be used
to estimate the determinants of the level of borrowing by treatment and control groups alike9. This
is actually the reason of using a Tobit specification equation. If no censoring had occurred, the
Tobit model would be inappropriate (Maddala 1999). We estimate a borrowing function for the
level of program participation, which is determined by the marginal effects of the amount of credit
borrowed during the last credit cycle, as follows:
c
i c i c i i c iC X Z L uα β γ θ= + + + + (17)
where i
X and i
L are the same vectors of household and credit market characteristics, respectively
derived earlier in equation (11) and i
Z is a vector of observable variables distinct from those in i
X
that affect i
C but not the outcome of interest iY conditional on i
C that play the role of identifying
instruments. c
α , c
β , γ and c
θ
are the intercept and the unknown parameters, respectively,
whereas c
iu
is the error term, which captures unmeasured household characteristics that determine
borrowing levels. The function for the income per adult equivalent, conditional upon the level of
program participation i
C takes the form
y
i y i y i y i iY X L C uα β θ δ= + + + + (18)
where y
α , y
β , y
θ and δ are the intercept and the unknown parameters respectively, while y
iu is
the error term reflecting unmeasured determinants of iY that vary from household to household.
14
Since i
C is included as the explanatory variable in (18), we need to identify an instrument,
additional to DISTANCE, to control for policy-specifics that affect the amount of credit and not only
the accessibility to it. This instrument must satisfy the same conditions as in the Heckit in order to
estimate the 2SLS Tobit procedure, the type of method that Amemiya (1984) refers to as Type III
Tobit model. We derive this estimation equation as follows:
i y i y i y i i i
Y X L C R eα β θ δ υ= + + + + + (19)
where i
R and υ are the predicted Tobit residuals and its parameter estimate, respectively, and
( )y y
i i i ie u E u R≡ − , where ( , )i i
e R are assumed to be independent of i
X , i.e. ( , ) 0i i iE e X R = . The
predicted residuals from the Tobit equation are estimated when 0i
C ≥ in (17) and then included as
another regressor in (19) to yield consistent and efficient estimators (Wooldridge 2003). The null of
no selection bias is tested in similar fashion as in the Heckit; however, now the 2SLS
heteroskedasticity-robust t statistic on the predicted residuals is used: when 0υ ≠ , a selection
problem is encountered.
We identify, as the additional instrumental variable, the length of membership, which is computed
as the number of years of program participation and coded as MEMBERSHIP. This variable is
related to progressive lending, an incentive device extensively used by microcredit programs to
deal with moral hazard and reduce operational costs in the long run. When equation (17) was
estimated with DISTANCE and MEMBERSHIP as identifying instruments, the p-values of the t
statistic for the coefficient γ for each microcredit program rejected the null of 0 : 0H γ = , reflecting
the statistically significance correlation between the level borrowing, i
C and the two instruments
in i
Z ; however, when i
Z was included in equation (18), the parameter estimate γ accepted the null
of no correlation against iY (see table 6)10. As a result we were able to use DISTANCE and
MEMBERSHIP as identifying instruments for the Tobit selection procedure.
15
INSERT TABLE 6 ABOUT HERE
Note that the predicted residuals from the second-stage Tobit selection equation presented in table
7 (and coded as RESID) report insignificant levels in the parameter estimates υ , confirming, as in
the Heckit procedure, the assumption of no selection bias. In this sense, the evidence suggests that
increasing levels of borrowing are a function of policy-specifics that are exogenously determined,
and linearly correlated to progressive lending (captured by the length of membership).
INSERT TABLE 7 ABOUT HERE
Now, in order to confirm the assumption of exogeneity, we exploit the qualities of the Hausman’s
procedure (Hausman 1978) by testing under the null hypothesis that the asymptotic covariance
matrix of the OLS estimator is not systematically larger than the 2S-Tobit selection equation. In
order words, we examine under the null if lim 0p =d , where 2bS Tobit OLS
B−= −d , whereas under
the alternative, lim 0p ≠d . Following Greene (2003:83) we compute the Hausman statistic in
STATA as follows:
{ } ( )1
2
2 2 2ˆ ˆ ˆˆ ˆ ˆ( ) ' . . Var . . Var ( )
d
S Tobit OLS S Tobit OLS OLS S TobitH b B Est Asy b Est Asy B B b Jχ
−
− − −= − − − → .
The Hausman statistics are 2 (13)χ = 0.24, 2 (13)χ = 0.13, and 2 (12)χ = 2.11 for the sample surveys
at Fincomun, CAME and Promujer, respectively. Consequently, we cannot reject the null that the
ˆOLS
B and 2ˆ
S Tobitb − are both consistent, and ˆ
OLSB efficient relative to 2
ˆS Tobit
b − . In this sense, by
following a geographical criterion during the process of data collection, we were able to control for
potential endogeneity problems, allowing us to concentrate on the OLS results in table 7.
The parameter estimate δ of the impact variable, i
C , reports a positive sign for each microcredit
program; however, the coefficients were only significantly different from zero in the case of
Fincomun. More precisely, the results suggest that a one percent increase in the amount of capital
borrowed from Fincomun gives rise to a 0.064% in income per adult equivalent, ceteris paribus. This
result is important for two reasons:
16
First, it confirms that our findings are in line with the results reported in table 5; however, by
substituting i
C for i
I we were able to discount the effects of older borrowers on the average
impact of program participation. Second, the result confirms the findings reported elsewhere, in
relation to small (or insignificant) impacts of microcredit on income, e.g. Morduch (1998) and
Coleman (1999). Now, based on the evidence reported by Hulme and Mosley (1996), we consider
the following proposition:
Proposition 1. Microcredit programs reporting poverty impacts have the largest effects among those closest
to, or above, the poverty line.
In order to test Proposition 1, we examine in section 4 the relationship between the severity of
income deprivation and the impacts of microcredit programs.
4. The impact of credit on poverty reduction To begin the discussion, we estimate the incidence of poverty among program participants. We
follow Sedesol (2002) criteria in order to derive three different thresholds of income deprivation for
urban areas11: 1) a food-based poverty line that identifies the lowest threshold of income required to
fulfill the minimum nutritional requirements to have a healthy living. This threshold, referred
hereafter as PL1, is computed at 784.5 pesos per month and measures the incidence of extreme
poverty. 2) A capabilities-based poverty line that includes to the food-based basket, other components
such as health care and education. This threshold, referred hereafter as PL2, is computed at 1507.5
pesos per month. 3) An asset-based poverty line (PL3) that measures the incidence of moderate poverty
and is computed at 1881 pesos per month. We also include the World Bank’s US$ 2 a day poverty
line for comparative purposes.
The use of several critical thresholds of human deprivation is justified for two reasons: firstly, there
is a widespread recognition that the conventional World Bank’s poverty lines are too low for the
existing domestic prices in Mexico. Secondly, by computing several poverty lines we can analyze
how deep microcredit programs reach the poor and measure the magnitude of poverty impacts by
17
levels of deprivation. The estimation of the incidence of poverty and poverty gap are presented in
table 8.
INSERT TABLE 8 ABOUT HERE
We observe a larger poverty incidence among treatment households at CAME and Promujer than
at Fincomun; however, only in the case of Fincomun we find, when computing PL3 and PL2, a
significant association between treatment and control groups in relation to the incidence of
poverty. The estimated poverty gap also shows a longer distance between actual incomes and the
poverty line among participants at Fincomun than at CAME and Promujer. Poor borrowers at
Fincomun had to cover, on average, an income shortfall of 545 pesos per month in order to cross the
PL2, whereas poor borrowers at CAME and Promujer had to cover only 304 and 319 pesos,
respectively.
The empirical evidence suggests that some programs (in this case Fincomun) might be more
effective at reporting poverty impacts but only by lifting those closest to the poverty line (as stated
in Proposition 1). Other programs (e.g. CAME and Promujer) might be more effective at having
impacts on the poverty gap but by doing so, they report insignificant impacts on the overall
incidence. In order to explore this issue, we compute the marginal effects of credit across the
poverty lines, using a Probit estimation equation as follows:
i i i i
PL C uα δ= + + (20)
where i
PL is a binary variable that takes the values
1 if household is below poverty line
0 otherwise
th
i
iPL
and i
C is the same continuous variable in (19) that measures the amount of credit borrowed in
logarithmic form. We have run (20) with i
PL adopting different poverty lines and using by default,
the definition of income per adult equivalent 1. We have also included, for comparative purposes,
18
iI as substitute of
iC , where
iI is the dichotomous variable used in the Heckit with value = 1I for
treatment households and = 0I for control groups. By computing the marginal effects of i
C we
are able to estimate the impact of a relative change in the amount of credit borrowed by a poor household
on the probability of staying in poverty. Alternatively, by including i
I in the Probit equation, we
are able to estimate the impact of the choice of a poor household to participate in a microcredit program
on the probability of staying in poverty. The results are presented in table 9.
INSERT TABLE 9 ABOUT HERE
The slope coefficient of i
C reports negative signs when PL2 and PL3 are included as thresholds of
income deprivation; however, they are only statistically significant In the case of Fincomun. Other
things held constant, the impact of a relative change of x% in the level of borrowing by poor
members was a decline of about –0.038x% in the probability of staying below PL2, and the
magnitude of this impact was marginally greater (-0.043x%) when PL3 was included as the poverty
measure. Although the parameter δ reports negative signs when PL1 is used as the poverty line,
the statistical significance is not different from zero. In other words, we do not find evidence of
poverty impacts when only the extreme poor are included in the analysis.
Unexpectedly, Promujer reports positive signs and statistical significance in the slope coefficient of
iC when PL1 is used in the estimation equation. This suggests that, other things held constant, a
relative change in the amount of credit borrowed by a woman in extreme poverty will increase her
probability of staying in that level of deprivation. Although the slope coefficient of i
C shows
negative signs, we find no statistical significance to confirm poverty impacts in the case of CAME.
These results must be interpreted with caution in view of the relative small number of participants
in extreme poverty, although they appear to confirm Proposition 1 regarding larger poverty
impacts among those at the upper limits of deprivation. This should not be surprising since, as
Hulme and Mosley point out: “those with higher incomes have greater range of investment opportunities,
19
more information about market conditions and can take on more risk than the poorest households without
threatening their minimum needs for survival” (1996:109).
A question naturally arises at this point: why are there impact differences between microcredit
programs even when they operate under similar market conditions? During the process of data
collection, a large percentage of borrowers (97.2% in the case of Fincomun, and 78.6% and 100% in
the cases of Fincomun, CAME and Promujer, respectively) reported having been engaged in
income generating activities that often took place at street markets or premises far away from their
place of residence. It appears that the urban poor, unlike in the rural context, often travel long
distances in search of livelihoods. Treatment groups at Fincomun, for instance, spent on average 78
minutes in a return journey from their place of work to the branch, and treatment groups at CAME
and Promujer travelled 65 and 82 minutes, respectively. Program participants repeatedly expressed
their frustration with the rigidity of periodical repayment schedules in compulsory meetings,
which are extensively used as monitoring device. CAME and Promujer, for instance, request to
attend group meetings on a weekly basis. Since Fincomun relies mostly on individual lending, it
does not require periodical meetings but just weekly instalments12. To illustrate this, consider the
following examples:
Case 1: Mrs T lives in the Chalco Valley and has been member of one of CAME´s village banks for
almost 7 years. She sells shoes at street markets that are usually held on particular days of the
week. For that reason, she travels from square to square across Mexico City. When she was asked
to give her opinion about CAME, she replayed: “I cannot make repayments every week, I don’t have
problems with the interest rate but I don’t like when they (credit officers) force us to come every week. I have
a business to attend, you know, and it is far away…”(Interview: Int2-11032004).
Case 2: Mrs C lives in Tula City in the state of Hidalgo and has been member of Promujer for more
than a year. She sells home products with relatives and friends, and at street markets in the
surrounding areas. Since she is in charge of childcare (she has three small children), she is able to
20
work only 2 hours per day, although she expressed a desire of spending more time on her business.
When she was asked to give her opinion about Promujer, she replayed: “What I don’t like from
Promujer is that we have to come every week and wait hours and hours for some comrades that come late. I
have lots of problems to be sitting here waiting for them (group-members). And now because of the
meetings I cannot pick up my children from school. To be honest with you, I will leave the group as soon as
the credit cycle ends…” (Interview: Int5-06042004).
Although periodical group meetings are important monitoring mechanisms that reduce
informational costs to the lender, they actually transfer these costs, as pointed out by Stiglitz (1990),
to the borrower. The income that the urban poor have to forego as the consequence of attending
group meetings can be significant.
Some readers may ask: if rigid monitoring devices exacerbate the opportunity cost of program
participation, why then do participants remain in the program? Credit markets in Mexico are
highly fragmented, and as long as interest rates and loan contracts offered by microcredit programs
are in better conditions than those of close competitors such as moneylenders, borrowers may
simply decide to remain in the program in order to take advantage of progressive lending, which
de facto generates micro-rationing13, and minimise this cost. A change in household income will
shift the budget constraint to the right and consequently, the opportunity of cost of program
participation. In that perspective, we consider the following proposition for empirical investigation:
Proposition 2. Microcredit performs as a normal good and after an income effect; the elasticity of demand for
microcredit in relation to the opportunity cost of program participation will be greater than one.
In order to test Proposition 2, we compute a Tobit model following the same estimation procedure
of section 3.1 that takes the form:
1 2
O c
i c i c i c i c iC X C L uα β β θ= + + + + (21)
where i
C measures the amount of credit borrowed and O
iC , the opportunity cost of program
21
participation, both expressed in logarithmic form. i
X and i
L are the same vectors of household and
credit market characteristics, respectively, derived earlier in (19). O
iC is computed by estimating
the time spent in compulsory group meeting plus the time spent on travelling to attend such
meetings. The sum was weighted by the estimated earnings forgone, based on actual values
reported from income generating activities, and the outcome multiplied by 16 given the number of
periodical repayments per credit cycle14. Estimation results, which are presented in Table 10,
confirm Proposition 2: a one percent change in the opportunity cost of program participation gives
rise to a 1.614 percentage change in the demand for microcredit at Fincomun, ceteris paribus, and
this figure was in the order of 1.716 and 1.513 at CAME Promujer, respectively.
INSERT TABLE 10 ABOUT HERE
Estimation results also reflect the impact of rigid monitoring devices on household income. For
example, the opportunity cost of program participation that participants at CAME absorbed was in
the order of 1008 pesos and represented, on average, about 17% the capability-based poverty line
(PL2), vis-à-vis 886 pesos reported from participants at Fincomun.
5. Conclusions and policy recommendations
Our results are suggestive for two reasons: 1) although expanding access to credit can be an
important instrument in the fight against poverty, design factors can constrain the magnitude of
the expected outcome. Our findings point to a linkage between lending technology and poverty
impacts. Rigidity in group lending can increase the utility cost of program participation to such
levels that undermine the efforts of poverty alleviation. In that context, experimentation should be
encouraged, and perhaps facilitated by governmental agencies such as PRONAFIM beyond the
narrowed objective of expanding credit access, to improve lending technology and practices. This
could be done through a number of possible policy actions that are summarised in table 11.
INSERT TABLE 11 ABOUT HERE
22
2) Although we find poverty impacts from individual lending technology, the magnitude of the
impacts were only significant at the upper limits of deprivation, where the poor and moderate poor
are grouped, suggesting a strong relationship between the level of household welfare and poverty
impacts. However, we find no evidence of microcredit impacts on extreme poverty. Our results
question the ability of microcredit to reduce extreme poverty in urban areas and call for more
research on alternative instruments such as nutrition, literacy and health care, which in
combination with credit, can be more effective to avoid extreme deprivation in urban areas.
INSERT TABLE 12 ABOUT HERE
References Amemiya T. (1984) Tobit models: a survey, in: Journal of Econometrics, Vol. 84 pp. 3-61 Babbie, E. (1990) Survey research methods. Belmont, CA: Wadsworth Campbell, D. and Stanley, J. (1966) Experimental and Quasi-experimental designs for research, in: N.L. Gage (Ed), Handbook of research on teaching, pp. 1-76. Chicago: Rand McNally Coleman B. (1999) The impact of group lending in Northeast Thailand, in: Journal of Development Economics, Vol. 60, pp. 105-141. Daley-Harris, S. (2007) State of the Microcredit Summit Campaign. Report 2007. Fink, A. and Kosecoff, J. (1985) How to conduct surveys: A step-by-step guide. Beverly Hills, CA: Sage Publications Greene W. (2003) Econometric Analysis. Prentice Hall, Fifth edition. Halvorsen R. and Palmquist R. (1980). The interpretation of dummy variables in semilogarithmic equations, in: American Economic Review, Vol. 70 No. 3, pp. 474-5 Hagenaars A. de Vos K. and Zaidi A. (1998). ‘Patterns of poverty in Europe’ in: The Distribution of Welfare and Household Production: International Perspectives, Cambridge University Press Cambridge Hausman J. (1978) Specification test in Econometrics, in Econometrica, 46, pp 1251-1271 Heckman J. (1979) Sample selection bias as a specification error, in Econometrica, Vol. 46, pp, 153-61 Hoff, K and Stiglitz, J (1990). Introduction: Imperfect information and rural markets –Puzzles and policy perspectives, in: World Bank Economic Review, No. 4, pp. 235-249
23
Hulme, D (2000) Impact Assessment methodologies for Microfinance: theory, experience and better practice, in: World Development, 28(1), pp. 79-98
Hulme D. and Mosley P. (1996) Finance against poverty. Vol. I, II, London: Routledge. Klein L (1961) An introduction to Econometrics, Prentice Hall, Englewood Cliffs, N.J. Keppel, G. (1991) Design and analysis: A researcher’s handbook. Englewood Cliffs, NJ: Prentice Hall. Maddala G. (1999) Limited dependent and qualitative variables in Econometrics. Econometric Society Monographs, Cambridge University Press Morduch J. (1998) Does microcredit really help the poor: new evidence from flagship programs in Bagladesh. Department of Economics and HIID, Harvard University and Hoover Institution, Stanford University. Morduch J. and Haley B. (2002) Analysis of the effects of microfinance on Poverty Reduction. New York University: NYU Wagner working paper No. 1014. Morse J. (1991) Approaches to qualitative-quantitative methodological triangulation, in: Nursing Research, 40(1), pp. 120-123 Niño-Zarazúa, MA (2007). The impacts of microcredit on income poverty, labour and well-being: A quasi-experimental study in urban Mexico. Doctoral thesis. Department of Economics, University of Sheffield. Secretaria de Desarrollo Social (Sedesol) (2002) Medición de la Pobreza. Variantes metodológicas y estimación preliminar, Comité Técnico para la Medición de la Pobreza. Serie: documentos de investigación, julio. Stiglitz, J. (1990) Peer monitoring and credit markets, in: The World Bank Economic Review, Vol. 4, No. 3, pp. 351-366 Stiglitz, J. and Weiss, A. (1981) ‘Credit Rationing in Markets with Imperfect Information’ American Economic Review, 71, pp 393-410 Pitt, M. and Khandker, S. (1998a) Household and intra-household impact of the Grameen Bank and similar targeted credit programs in Bangladesh, in: Journal of Political Economy, Vol. 106, 558-596 Pitt, M. and Khandker, S. (1998b) The impact of group-based credit programs on poor households in Bangladesh: Does the gender of participants matter? in: Journal of Political Economy, Vol. 106, 958-996 Rothbarth, E. (1943). ‘Note on a method of determining equivalent income for families of different composition’, in C. Madge (ed) War-time Pattern of Saving and Spending, Cambridge University Press, Cambridge.
24
Tobin J. (1958) Estimation of relationships for limited dependent variables, in: Econometrica, Vol. 26, pp. 24-36 Wagstaff and van Doorslaer (1998) Income-related inequalities in health: some international comparisons, in: Journal of Health Economics, Vol 16, Issue 1, pp. 93-112 Wooldridge J. (2002) Econometric analysis of cross section and panel data. London: MIT Press. Notes
1 Contact at: University of Sheffield. 9 Mappin Street, S1 4DT Sheffield, United Kingdom. Tel: +44 114 222 3343, Email: [email protected] 2 For more details about the microcredit sector in Mexico, see Niño-Zarazúa (2007). 3 For details about the instruments of data collection, contact the author at [email protected]. 4 See Wooldridge (2002), Greene (2003) and Maddala (1999) for a detailed discussion on the properties of the identifying instrument. 5 We tested the 0γ ≠ condition in (11) and (12) by computing a heteroskedasticity-robust t statistic after OLS
estimation. 6 In order to estimate these weights, we asked programme participants about the approximate distance they travel to attend group meetings. 7 We adopted Lawrence Klein’s rule of thumb (1961), to test for potential problems of collinearity. We did not find evidence of collinearity.
8 Since we have a data-censoring case demanding the variable *
iC to follow a homoskedastic normal
distribution, we use a logarithmic transformation in our estimation strategy to make this assumption more reasonable. 9 For further details on the derivation of the conditional mean functions, see Greene (2003). 10 We repeated Klein’s rule of thumb (1961) to test DISTANCE and MEMBERSHIP for potential problems of collinearity. We did not find collinearity problems. 11 The Consumer Price Index was used to deflate the levels of prices in 2004 12 Fincomun has recently opened an account at HSBC where borrowers can deposit their weekly instalments. This initiative has benefited borrowers by reducing both transaction and opportunity costs. 13 Micro-rationing reflects a situation in which even those with access to credit are still credit constrained. 14 We acknowledge the possibility of being overestimating the opportunity cost of borrowing if other
household members cooperate with the borrower and work in the business while attending group meetings. We find that a large percentage of borrowers (47%) worked alone, and about half of those were single mothers having their business as the only source of household income. If we take into account the risk associated to the income loss, and if group meetings are arranged during the prime selling hours of the business, then we could be underestimating this cost.
25
Figure 1. The selection process for program participation. Adapted from Maddala (1999)
Population
Households self-excluding to participate in a credit
program1
(Ι 0)=
Households self-selecting to participate in a credit
program 1
(Ι 1)=
Households accepted by the lender (or group members) to
participate 2
(Ι 1)=
Households rejected by the lender (or group members) to
participate 2
(Ι 0)=
Household participants that have received a credit
(TREATMENT GROUP)
Household participants that have not received a credit (CONTROL GROUP)
26
Table 1. Characteristics of the case-study microcredit programs Information corresponding to 2004
Characteristics FINCOMUN CAME PROMUJER
Type of organisation Credit Union Non-Governmental Organisation Non-Governmental
Organisation Year of establishment 1994 1991 2001
Founders Juan Diego Foundation, a
catholic group
Foundation for Community Assistance, belonging to the
Archdiocese of Mexico Pro-Mujer International
No of branches 27 5 21 Personnel 339 580 45 Lending methodology Individual lending Credit-only village-banking Credit-plus village-banking Interest rate (per annum)
72% 60% 72%
Borrowers (000) 25.8 40 11.8 Women borrowers 60% 80% 100% Gross loan portfolio (000 MEX$)
169,725 58,000 13,739
Average outstanding loan (000 MEX$)
6.6 1.5 2.1
Loan loss reserve ratio 2.7% 1.8% 2.9%
Area under study San Miguel Teotongo, in the
Metropolitan area of Mexico City Chalco Valley, in the
Metropolitan area of Mexico City Tula City and the surrounding
areas in the state of Hidalgo Screening devices
Periodical repayment schedules
16 to 24 weekly instalments at Fincomun officers or HSBC
branches
16 weekly instalments in compulsory group meetings.
12 to 24 weekly or fortnightly instalments in compulsory
group meetings Compulsory savings 10 as % of loan 10-12 as % of loan 10-12 as % of loan
Others The use of palm pilots for
screening loan applications No No
Enforcement devices Guarantees Yes, two guarantees Yes, through joint liability Yes, through joint liability
Physical collateral Yes, with a value twice the credit
size. No No
Penalties On loan defaults and interest in
arrears. Legal actions.
On absence and late attendance to group meetings. Interest in
arrears and legal actions
On absence and late attendance to group meetings. Interest in
arrears and legal actions Incentive devices
Progressive lending Additional loans for a maximum
of 50% of previous credit
Additional loans based on a fixed loan schedule and compulsory savings. Upper loan limits at
20,000 pesos
Additional loans based on a fixed loan schedule and
compulsory savings.
Others Voluntary savings products and
certificates of deposits Life Insurance to cover loan
balance. Revolving fund.
Training in financial literacy, business development and
health care
Table 2. Identifying equations on functional form Logarithm of the cost of borrowing (LGCOSTBORROWPC) as identifying instrument Dependent variable in (10): logarithm of the maximum amount of credit borrowed (LGMAXCREDIT) Dependent variable in (11): logarithm of monthly income per adult equivalent 1 in pesos of 2004 (LGINCOMEPC)
FINCOMUN CAME PROMUJER Eq. (10) Eq. (11) Eq. (10) Eq. (11) Eq. (10) Eq. (11)
LGCOSTBORROWPC 1.574 0.325 1.705 0.082 1.458 0.055 (21.18)*** (2.05)** (10.74)*** (0.62) (14.61)*** (0.50) Observations 55 55 46 46 47 47
R-squared 0.44 0.49 0.41
Absolute value of t statistics in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%
27
Table 3. DISTANCE as identifying instrument Dependent variable in (10): logarithm of maximum amount of credit borrowed (LGMAXCREDIT)† Dependent variable in (11): logarithm of monthly income per capita in pesos of 2004 (LGINCOMEPC)
FINCOMUN CAME PROMUJER Eq. (10) Eq. (11) Eq. (10) Eq. (11) Eq. (10) Eq. (11)
DISTANCE 0.028 -0.000 0.073 0.005 0.066 -0.005 (1.88)** (0.09) (2.15)** (0.94) (1.92)* (1.57)
Absolute value of t statistics in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% † The Heckman procedure transforms LGMAXCREDIT into a dummy variable for treatment group = 1 if Ii > 0.
Table 4. Robustness of DISTANCE as instrumental variable Endogenous explanatory variable in (11): Logarithm of the maximum amount of credit borrowed (LGMAXCREDIT)† Dependent variable in (11): logarithm of monthly income per capita (LGINCOMEPC)
FINCOMUN CAME PROMUJER
Eq. (11) on functional
form
Eq. (11) with
DISTANCE
Eq. (11) on functional
form
Eq. (11) with
DISTANCE
Eq. (11) on functional
form
Eq. (11) with
DISTANCE
LGMAXCREDIT 0.591 0.595 0.103 0.088 0.629 0.582 (2.48)** (3.39)*** (0.59) (0.90) (1.98)** (1.88)* MILLS 0.258 0.653 0.089 0.043 -0.053 0.261 (0.58) (1.57) (0.67) (0.15) (0.14) (1.05) DISTANCE 0.002 0.006 -0.006 (0.32) (1.13) (1.06)
Absolute value of z statistics in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1% † The Heckman procedure transforms the LGMAXCREDIT variable into a dummy for treatment group = 1 if Ii > 0
Table 6. Identifying instruments for the Tobit selection equation Dependent variable in (17): logarithm of the maximum amount of credit borrowed (LGMAXCREDIT) Dependent variable in (18): logarithm of monthly income per adult equivalent 1 in pesos of 2004 (LGINCOMEPAE1)
FINCOMUN CAME PROMUJER Eq. (17) Eq. (18) Eq. (17) Eq. (18) Eq. (17) Eq. (18)
MEMBERSHIP 2.235 (6.80)***
-0.024 (0.19)
2.074 (6.78)***
0.018 (0.29)
5.487 (10.36)***
-0.003 (1.22)
DISTANCE 0.060 -0.001 0.058 0.004 0.042 0.340 (2.60)** (0.41) (1.76)* (0.88) (2.84)*** (1.65)
Absolute value of t statistics in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%
28
Tab
le 5
. Th
e im
pac
t o
f p
rog
ram
par
tici
pat
ion
on
ho
use
ho
ld i
nco
me
En
do
gen
ou
s ex
pla
nat
ory
var
iab
le (
iI
in E
qu
atio
n 1
1): L
og
arit
hm
of
the
max
imu
m a
mo
un
t o
f cr
edit
bo
rro
wed
(L
GM
AX
CR
ED
IT)
†
FIN
CO
MU
N
CA
ME
P
RO
MU
JER
P
oo
led
sam
ple
OL
S
Hec
kit
O
LS
H
eck
it
OL
S
Hec
kit
O
LS
H
eck
it
LG
MA
XC
RE
DIT
0.
553
0.59
5 0.
126
0.08
8 0.
110
0.58
2 0.
313
0.11
5
(2.5
3)**
(3
.39)
***
(0.8
1)
(0.9
0)
(0.7
3)
(1.8
8)*
(3.5
2)**
* (1
.75)
* D
epen
den
t v
aria
ble
(i
Y i
n E
qu
atio
n 1
1):
log
arit
hm
of
mo
nth
ly i
nco
me
per
cap
ita
in p
eso
s o
f 20
04 (
LG
INC
OM
EP
C)
MIL
LS
0.65
3 (1
.57)
0.04
3 (0
.15)
0.26
1 (1
.05)
0.12
9 (0
.61)
L
GM
AX
CR
ED
IT
0.54
8 0.
588
0.14
0 0.
099
0.10
2 0.
701
0.31
5 0.
121
(2
.57)
**
(3.2
7)**
* (0
.91)
(1
.00)
(0
.67)
(2
.33)
**
(3.5
9)**
* (1
.81)
* D
epen
den
t v
aria
ble
(i
Yin
Eq
uat
ion
11)
: lo
gar
ith
m
of
mo
nth
ly i
nco
me
per
ad
ult
eq
uiv
alen
t 1
in
pes
os
of
2004
(L
GIN
CO
ME
PA
E1)
a/
MIL
LS
0.67
1 (1
.57)
-0.0
10
(0.0
3)
0.
293
(1.1
8)
0.
118
(0.0
8)
LG
MA
XC
RE
DIT
0.
605
0.55
4 0.
109
0.06
3 0.
067
0.69
1 0.
314
0.11
1
(2.9
1)**
* (3
.05)
***
(0.8
0)
(0.6
8)
(0.4
4)
(2.5
3)**
(3
.75)
***
(1.7
4)*
Dep
end
ent
var
iab
le (
iY
in E
qu
atio
n 1
1): l
og
arit
hm
of
mo
nth
ly i
nco
me
per
ad
ult
eq
uiv
alen
t 2
in
pes
os
of
2004
(L
GIN
CO
ME
PA
E2)
b/
M
ILL
S
0.
676
(1.5
7)
0.
183
(0.6
5)
0.
294
(1.2
8)
0.
226
(1.0
9)
LG
MA
XC
RE
DIT
0.
611
0.55
8 0.
095
0.06
6 0.
065
0.73
7 0.
313
0.10
9
(2.9
3)**
* (3
.14)
***
(0.7
1)
(0.7
0)
(0.4
3)
(2.7
5)**
* (3
.74)
***
(1.6
9)*
Dep
end
ent
var
iab
le (
iY
in E
qu
atio
n 1
1): l
og
arit
hm
of
mo
nth
ly i
nco
me
per
ad
ult
eq
uiv
alen
t 3
in
pes
os
of
2004
(L
GIN
CO
ME
PA
E3)
c/
M
ILL
S
0.
661
(1.5
7)
0.
180
(0.6
3)
0.
311
(1.3
5)
0.
219
(1.0
5)
Ro
bu
st t
sta
tist
ics
in p
aren
thes
es
* si
gn
ific
ant
at 1
0%; *
* si
gn
ific
ant
at 5
%; *
** s
ign
ific
ant
at 1
%
† T
he
Hec
km
an p
roce
du
re t
ran
sfo
rms
LG
MA
XC
RE
DIT
in
to a
du
mm
y v
aria
ble
fo
r tr
eatm
ent
gro
up
= 1
if
I i >
0
a/ I
AE
1 fo
llo
ws
the
app
roac
h d
evel
op
ed b
y R
oth
bar
th (
1943
).
b/
IA
E2
foll
ow
s th
e ap
pro
ach
dev
elo
ped
by
Wag
staf
f an
d v
an D
oo
rsla
er (
1998
).
c/ I
AE
3 fo
llo
ws
the
OE
CD
mo
dif
ied
sca
les
bas
ed o
n H
agen
arrs
et.
al
(199
8).
29
Tab
le 7
Th
e im
pac
t o
f cr
edit
on
ho
use
ho
ld i
nco
me
En
do
gen
ou
s ex
pla
nat
ory
var
iab
le (
iC
in E
qu
atio
n 1
7): L
og
arit
hm
of
the
max
imu
m a
mo
un
t o
f cr
edit
bo
rro
wed
(L
GM
AX
CR
ED
IT)
FIN
CO
MU
N
CA
ME
P
RO
MU
JER
P
oo
led
sam
ple
OL
S
2S-T
ob
it
OL
S
2S-T
ob
it
OL
S
2S-T
ob
it
OL
S
2S-T
ob
it
LG
MA
XC
RE
DIT
0.
065
0.07
0 0.
014
0.00
3 0.
015
-0.0
43
0.03
7 0.
044
(2
.82)
***
(1.4
1)
(0.8
0)
(0.0
9)
(0.8
3)
(0.9
4)
(3.6
7)**
* (2
.38)
**
Dep
end
ent
var
iab
le (
iY
in E
qu
atio
n 1
8): l
og
arit
hm
of
mo
nth
ly i
nco
me
per
cap
ita
in p
eso
s o
f 20
04
(LG
INC
OM
EP
C)
RE
SID
-0.0
07
(0.1
2)
0.
012
(0.4
1)
0.
048
(1.3
0)
-0
.008
(0
.42)
L
GM
AX
CR
ED
IT
0.06
4 0.
075
0.01
5 0.
003
0.01
5 -0
.049
0.
036
0.04
5
(2.8
8)**
* (1
.57)
(0
.89)
(0
.07)
(0
.79)
(1
.12)
(3
.77)
***
(2.5
1)**
D
epen
den
t v
aria
ble
(i
Yin
Eq
uat
ion
18)
: lo
gar
ith
m
of
mo
nth
ly i
nco
me
per
ad
ult
eq
uiv
alen
t 1
in
pes
os
of
2004
(L
GIN
CO
ME
PA
E1)
a/
RE
SID
-0.0
14
(0.2
5)
0.
014
(0.4
7)
0.
052
(1.4
6)
-0
.010
(0
.52)
L
GM
AX
CR
ED
IT
0.07
0 0.
085
0.01
2 0.
004
0.01
0 -0
.045
0.
037
0.05
0
(3.2
1)**
* (1
.96)
* (0
.77)
(0
.14)
(0
.56)
(1
.03)
(3
.91)
***
(2.8
7)**
* D
epen
den
t v
aria
ble
(i
Yin
Eq
uat
ion
18)
: lo
gar
ith
m
of
mo
nth
ly i
nco
me
per
ad
ult
eq
uiv
alen
t 2
in
pes
os
of
2004
(L
GIN
CO
ME
PA
E2)
b/
RE
SID
-0.0
19
(0.3
9)
0.
008
(0.2
8)
0.
045
(1.2
8)
-0
.016
(0
.89)
L
GM
AX
CR
ED
IT
0.07
0 0.
010
0.01
0 0.
037
0.05
0
(3.2
4)**
* (0
.69)
(0
.55)
(3
.89)
***
(2.8
7)**
* D
epen
den
t v
aria
ble
(i
Yin
Eq
uat
ion
18)
: lo
gar
ith
m
of
mo
nth
ly i
nco
me
per
ad
ult
eq
uiv
alen
t 3
in
pes
os
of
2004
(L
GIN
CO
ME
PA
E3)
c/
R
ES
ID
0.08
6 (1
.94)
* -0
.019
(0
.39)
0.00
4 (0
.12)
0.
007
(0.2
5)
-0.0
47
(1.1
0)
0.04
7 (1
.34)
-0.0
16
(0.8
4)
Ro
bu
st t
sta
tist
ics
in p
aren
thes
es
* si
gn
ific
ant
at 1
0%; *
* si
gn
ific
ant
at 5
%; *
** s
ign
ific
ant
at 1
%
a/ I
nco
me
per
ad
ult
eq
uiv
alen
t 1
foll
ow
s th
e ap
pro
ach
dev
elo
ped
by
Ro
thb
arth
(19
43)
b/
In
com
e p
er a
du
lt e
qu
ival
ent
2 fo
llo
ws
the
app
roac
h d
evel
op
ed b
y W
agst
aff
and
van
Do
ors
laer
(19
98).
c/
In
com
e p
er a
du
lt e
qu
ival
ent
3 fo
llo
ws
the
OE
CD
mo
dif
ied
eq
uiv
alen
ce s
cale
bas
ed o
n t
he
wo
rk o
f H
agen
arrs
et.
al,
(19
98).
30
Table 8. Poverty incidence and poverty gap among program participants Figures in percentages
FINCOMUN CAME PROMUJER Concept Control Treated Control Treated Control Treated
Overall 34.5 65.5 39.1 60.9 44.7 55.3 Asset-based poverty line (PL3) ≤ 1881 pesos per month 73.7** 36.1 77.8 67.9 61.9 53.9 Poverty gap 44.8 39.3 34.1 25.4 23.3 30.6 Depth of poverty (in pesos) 842 738 642 477 439 576 Capabilities-based poverty line (PL2) ≤ 1507.5 pesos per month 63.2*** 27.8 50.0 42.9 33.3 38.5 Poverty gap 38.1 36.2 35.0 20.2 17.5 21.2 Depth of poverty (in pesos) 574 545 527 304 263 319 Food-based poverty line (PL1) ≤ 784.5 pesos per month 15.8 11.1 11.2 0 0 3.9 Poverty gap 43.4 28.2 13.5 0 0 5.1 Depth of poverty (in pesos) 341 221 106 0 0 43 World Bank's poverty line ≤ US$ 2 a day 15.8 8.3 5.6 0 0 0 Poverty gap 33.6 23.7 1.1 0 0 0 Depth of poverty (in pesos) 225 159 7 0 0 0
The statistically significant association in the cross-tabulations are indicated by the Chi-square values for the cell as a whole at 0.001 (*); 0.01 (**); 0.05 (***); and 0.1 (****) levels of significance.
Table 9. Probit: the effect of program participation on the probability of staying in poverty
Explanatory variables: i
C is the logarithm of the maximum amount of credit borrowed (LGMAXCREDIT).
iI is a dummy variable for treatment groups = 1
FINCOMUN CAME PROMUJER Pooled sample Independent variable: Dummy variable = 1 if IAE1 ≤ poverty line a/ with
iI with
iC with
iI with
iC with
iI with
iC with
iI with
iC
-0.379 -0.051 -0.350 -0.419 Coef
(0.82) (1.12) (0.98) (1.08) -0.074 -0.009 -0.036 -0.003
World Bank poverty line ≤ US $2 a day ∂Φ
∂X
(0.82) (1.12) (0.98) (1.08)
-0.217 -0.029 0.178 -0.229 -0.027 Coef
(0.49) (0.66) (5.83)*** (0.72) (0.79) -0.046 -0.006 0.003 -0.031 -0.003
Incidence of extreme poverty PL1 ≤ 784.5 pesos per month
∂Φ
∂X
(0.49) (0.66) (5.83)*** (0.72) (0.79)
-0.925 -0.100 -0.180 -0.019 -0.137 -0.013 -0.327 -0.390 Coef
(2.49)** (2.58)*** (0.47) (0.47) (0.36) (0.29) (1.53) (1.67)* -0.353 -0.038 - 0.071 -0.007 -0.051 -0.005 -0.127 -0.015
Incidence of poverty PL2 ≤ 1507.5 pesos per month
∂Φ
∂X
(2.49)** (2.58)*** (0.47) (0.47) (0.36) (0.29) (1.53) (1.67)*
-0.989 -0.108 -0.301 -0.030 -0.206 -0.029 -0.467 -0.055 Coef
(2.61)*** (2.73)*** (0.72) (0.70) (0.55) (0.64) (2.15)** (2.31)** -0.375 -0.043 -0.099 -0.010 -0.080 -0.011 -0.178 -0.021
Incidence of moderate poverty PL3 ≤ 1881 pesos per month
∂Φ
∂X
(2.61)*** (2.73)*** (0.72) (0.70) (0.55) (0.64) (2.15)** (2.31)**
Robust z statistics in parentheses * significant at 10%; ** significant at 5%; *** significant at 1% a/ Income per adult equivalent 1 (IAE1) follows Rothbarth (1943)
31
Table 10 Tobit estimation equation with LGOPPORTCOSTPC as explanatory variable Dependent variable in (21): Logarithm of the maximum amount of credit borrowed (LGMAXCREDIT)
FINCOMUN CAME PROMUJER
LGOPPORTCOSTPC 1.614 1.716 1.513 (21.10)*** (10.96)*** (13.51)***
Observations Left-censored
55 19
46 18
47 21
Absolute value of t statistics in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%
Table 11: Summary of findings and policy recommendations
Findings Policy actions for experimentation Expected benefits
Positive and statistically significant elasticity of income per adult equivalent in relation to credit in the case of Fincomun: a one percent increase in credit borrow gives rise to a 0.064% in income per adult equivalent. Although participants at CAME and Promujer also report positive elasticities, the statistical insignificance does not confirm this effect. A significant poverty impact from participants at Fincomun. The magnitude of the impact is linked to the level of household welfare: a one percent change in credit was predicted to reduce the probability of staying in poverty below a capabilities-based poverty line (PL2) by about -0.038% vis-à-vis -0.043% reported from households below an asset-based poverty line (PL3) that measures moderate poverty. No statistical significance was reported from participants at CAME but only an increased probability for extreme poor women borrowing from Promujer of staying below the food-based poverty line (PL1) that measures extreme deprivation.
Redefinition of credit targeting that could be facilitated by relaxation of donor conditionality. Additionally, more research on alternative instruments such as nutrition, literacy and health care that in combination of credit, could be more effective to alleviate extreme poverty.
Potential trickle down and wider effects in benefit of the extreme poor.
An elastic demand for microcredit in relation to the opportunity cost of program participation: a one percent change in the opportunity cost of program participation gives rise to a 1.614, 1.1716 and 1.1513 percentage change in the demand for credit by borrowers at Fincomun, CAME and Promujer, respectively.
Re-design rigid screening and monitoring devices that in combination with lending technology (e.g. mobile banking) could cut down time in group-meetings. Donors could play a role beyond the objective of expanding credit access, and facilitate research, experimentation and dissemination of innovations and best practices.
Reduction in the utility cost of program participation, with positive impacts on household income, poverty and market efficiency.
Incentive devices such as progressive lending can cause micro-rationing, and rigid screening and monitoring devices such as periodical group meetings can exacerbate the utility cost of program participation and undermine potential impacts on poverty alleviation.
Removal of upper limits of progressive lending that could be linked to the introduction of individual lending for “graduated” borrowers.
Reduction in micro-rationing with potential positive effects on client retention and improvements in financial self-sufficiency.
32
Table 12. List of variables
Independent variables Definition Obs Mean S.D. Min Max
Contained in i
X
AVEDU Years of education 148 7.047 3.777 0 17 HOWNER If household owns residence = 1 148 0.682 0.467 0 1 HESTATE If house is still in construction = 1 148 0.791 0.408 0 1 TIMEBUS Years in business 148 5.162 5.746 0 30 WWORKER Number of household members with a
waged job 148 0.547 0.703 0 3
DEPENDRATIO Dependency ratio (number of children, students and old members / household size)
148 0.498 0.222 0.125 1
AGE Age of borrower 148 42.189 10.846 19 74 WOMAN If borrower is woman = 1 148 0.730 0.446 0 1 MARITAL If borrower is in a relationship = 1 148 0.757 0.430 0 1
Contained in i
L
ROSCAS If borrower participates in rotating savings and credit association = 1
148 0.453 0.499 0 1
FORMALCREDIT If borrower have received loans from institutional lenders = 1
148 0.054 0.227 0 1
MONEYLENDER If borrower have received loans from moneylenders
148 0.095 0.294 0 1
Instrumental variables
DISTANCE Distance from branch to place of residence or business (in minutes)
148 32.365 21.716 10 100
MEMBERSHIP Years of membership 148 1.704 1.944 0 8
Dependent variables
LGMAXCREDIT Logarithm of the maximum amount of credit borrowed in the last credit cycle
148 5.475 4.466 0 10.621
LGOPPORTCOSTPC Logarithm of the opportunity cost of borrowing per credit cycle
148 3.880 3.204 0 8.006
LGINCOMEPC Logarithm of income per capita 148 7.296 0.594 5.438 8.868 LGINCOMEPAE1 Logarithm of income per adult
equivalent 1 148 7.452 0.571 5.733 9.055
LGINCOMEPAE2 Logarithm of income per adult equivalent 2
148 7.724 0.545 6.114 9.315
LGINCOMEPAE3 Logarithm of income per adult equivalent 3
148 7.895 0.543 6.324 9.512
POORPL1 If household’s income is below poverty line 1 = 1
148 0.068 0.252 0 1
POORPL2 If household’s income is below poverty line 2 = 1
148 0.405 0.493 0 1
POORPL3 If household’s income is below poverty line 3 = 1
148 0.581 0.495 0 1
POOR2US If household’s income is below US $2 a day = 1
148 0.047 0.213 0 1