Amin 2003

Journal of Development Economics 70 (2003) 59 – 82
www.elsevier.com/locate/econbase
Does microcredit reach the poor and vulnerable?

Evidence from northern Bangladesh
Sajeda Amin a, Ashok S. Rai b,*, Giorgio Topa c
a
Population Council, One Dag Hammarskjold Plaza, New York, NY 10017, USA
b
Economics Growth Centre, Yale University, P.O. Box 208269, New Haven, CT 06520, USA
c
Department of Economics, New York University, 269 Mercer St., New York, NY 10003, USA
Abstract
This paper evaluates whether microcredit programs such as the popular Grameen Bank reach the
relatively poor and vulnerable in two Bangladeshi villages. It uses a unique panel dataset with
monthly consumption and income data for 229 households before they received loans. We find that
while microcredit is successful at reaching the poor, it is less successful at reaching the vulnerable.
Our results also suggest that microcredit is unsuccessful at reaching the group most prone to
destitution, the vulnerable poor. Our main contribution is to explicitly evaluate the targeting of an
antipoverty intervention using the efficient risk-sharing framework in Townsend [Econometrica 62
(1994) 539-591].
D 2002 Elsevier Science B.V. All rights reserved.
JEL classification: O16; I38; Q12

Keywords: Poverty; Vulnerability; Microcredit; Targeting; Risk sharing; Grameen Bank
1. Introduction
Subsidized credit has a disappointing history of being politically manipulated and

diverted from its intended beneficiaries, the poor. Instead, the rich have benefited
disproportionately from such programs. For instance, 80% of the US$56 million subsidies
provided to Costa Rica’s largest bank in 1974 went to large wealthy farmers (Vogel, 1984).
The nonpoor were just as likely as the poor to participate in an Indian government
subsidized credit scheme in the 1980s (Ravallion and Datt, 1995). Yet in recent years,
*
Corresponding author.
E-mail addresses: [email protected] (S. Amin), [email protected] (A.S. Rai),
[email protected] (G. Topa).
0304-3878/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 0 4 - 3 8 7 8 ( 0 2 ) 0 0 0 8 7 - 1
60 S. Amin et al. / Journal of Development Economics 70 (2003) 59–82
subsidized lending programs such as the Grameen Bank have become ‘‘the world’s hot
idea to reduce poverty’’ (New York Times, 1997). These ‘‘microcredit’’ schemes provide
small loans without collateral to households excluded by the formal financial sector in
many developing countries (Morduch, 1999). While there have been numerous attempts to
assess the impact of these programs on a variety of outcomes (Pitt and Khandker, 1998a;
Morduch, 1998; and others), there has been little research on how well targeted modern
microcredit programs actually are.1
This paper uses a unique data set from two villages in Northern Bangladesh to test if
members of microcredit programs are poorer and more vulnerable than nonmembers. A
household is defined as poor if it has low consumption levels, and vulnerable if it is unable to
smooth consumption in the face of idiosyncratic income fluctuations. We use consumption
and income data for 229 households for 12 months in 1991 –1992 to identify households that
are poor and vulnerable. We then check to see which of these households joined a
microcredit program by 1995. Since microcredit organizations only just began to give loans
in the two villages in 1991 –1992, we can ignore issues of endogeneity for the most part. The
three programs in our study, Grameen Bank, Bangladesh Rural Advancement Committee
(BRAC) and Association for Social Advancement (ASA) are the largest microcredit
providers in Bangladesh, and among the largest in the world.
We find that modern microcredit programs are definitely more successful at reaching
the poor than their predecessors. The probability that a microcredit member is below the
poverty line is substantially higher than that of a randomly picked household in both
villages. A 24% decrease in monthly consumption at the mean increases the probability of
joining a microcredit program by about 6– 7%. In contrast, microcredit is less successful at
reaching the vulnerable. We find that the vulnerable are more likely to join microcredit
programs only in the richer of the two villages. Most crucially, we find no evidence that
microcredit reaches the households most in need of assistance, the vulnerable poor.
Indeed, our results suggest that the vulnerable poor are excluded from microcredit in the
poorer village.
Many rural households lack insurance against risks of sickness, floods, crop damage
and fluctuations in prices. Of these, household-specific risks but not aggregate shocks can
be insured against at the village level. We think of households that are unable to perfectly
insure themselves in the event of household-specific shocks as vulnerable. The recent
World Development Report uses a similar definition: ‘‘vulnerability measures. . . the
likelihood that the shock will result in a decline in well-being’’ (World Bank, 2001, p.139).
In principle, both rich and poor households could be labeled as vulnerable according to
our measure.2 Our data show that poorer households are more vulnerable than the rich,
however. Reaching the vulnerable poor is considered to be crucial to any poverty reduction
1
Navajas et al. (1998) and Pitt and Khandker (1998b) are exceptions. Unlike this paper, they use data on
households after they have joined microcredit programs.
2
There is an issue of semantics here. The inability to perfectly insure against idiosyncratic risk has different
welfare implications for a poor than for a rich household, since the former is more likely to fall into destitution
following an adverse income shock. Therefore, some readers would no doubt have preferred that we use the term
‘‘vulnerable’’ only for uninsured poor households and not for uninsured rich households. Since we separately
analyze whether microcredit reaches the poor, the uninsured, and the uninsured poor, we do (indirectly) address
their concern.
S. Amin et al. / Journal of Development Economics 70 (2003) 59–82 61
strategy (World Bank, 2001). Further, protecting the vulnerable poor is an explicit goal of
microcredit practitioners and a means to promote repayment.3 Grameen explicitly bundles
insurance with credit provision: borrowers have access to repeat loans from a disaster fund
and a group savings fund when they are hit with an adverse shock. Impact studies
emphasize microcredit’s role in smoothing consumption (Morduch, 1998; Pitt and
Khandker, 1998a).
The main contribution of this paper is to explicitly evaluate the targeting of an
antipoverty intervention using the general equilibrium framework of risk sharing in
village economies (Townsend, 1994). Since the presence of vulnerable households
indicates a market or institutional failure, it can be welfare improving to assist the
vulnerable, particularly the vulnerable poor. We compute several alternative vulnerability
measures to control for measurement error and to allow for different specifications of the
basic risk-sharing model. Our results on the relationship between vulnerability and
microcredit membership are robust to these alternatives. We also compare our vulner-
ability measure to more commonly used measures of consumption variability, and argue
that such measures are inadequate proxies for vulnerability. In contrast to our vulnerability
measure, we find that rich households have more variable consumption than the poor.
The rest of the paper is organized as follows. Section 2 describes the two villages and
the data. Section 3 derives a measure of vulnerability and describes its estimation. Section
4 reports our findings on whether microcredit reaches the poor and vulnerable. Section 5
concludes.
2. Data
The study uses transactions data collected over 12 months in two villages (called A and
B to preserve anonymity) in the district of Rajshahi in northwest Bangladesh. Table 1
reports the distribution of occupations within each village. Village A is primarily
agricultural while village B is more diversified in its income sources: only 20% of the
households in village A but over half the households in village B do not report agriculture
or daily labor as their main occupation. Village B has several small shops, a marketplace
(haat) that meets twice a week and attracts 200 vendors, and local government offices. All
major marketing activities for village A are held in marketplaces outside the village. Both
villages grow three rice crops a year. In addition, village A grows betel leaf, a cash crop,
and village B has several jointly owned mango orchards (Amin, 1998).
In 1991, there were 395 and 398 households, respectively, in the two villages. Of these,
120 households were sampled in each village. Male-headed households had a 1/4 chance
of being surveyed, while all female-headed households were sampled. The lack of
complete data for few households brought the number of units in our sample down to
112 for village A and 117 for village B. Households were followed for 12 rounds and data
3
According to Grameen Bank’s deputy managing director: ‘‘In view of the vulnerabilities of its targeted
clientele, Grameen has carefully built into its credit delivery system innovative safety features. These act like
shock absorbers, enabling Grameen members to better cope with natural as well as man-made disasters. Without
them the credit delivery system would not have functioned at all. . .’’ (Shams, 1992, p. 13, emphasis added).
Table 1
Summary statistics *
Village A (112 households) Village B (117 households)
Mean SD CV Mean SD CV
Food consumption 324.00 171.05 0.53 351.73 227.61 0.65
All consumption 388.38 210.35 0.54 466.93 308.29 0.66
All income 429.69 347.49 0.81 537.62 564.04 1.05
All revenue 537.23 467.85 0.87 681.84 713.47 1.05
Medical expenditures 36.06 50.37 1.40 57.56 77.20 1.34
Household size 4.64 2.43 0.52 4.60 2.02 0.44
Age of HH head 41.77 14.34 0.34 43.66 13.97 0.32
Number of children 1.66 1.30 0.78 1.57 1.20 0.77
Number of old people 0.19 0.46 2.35 0.21 0.45 2.11
Agricultural land 904.99 2750.02 3.04 766.45 1980.07 2.58
* All statistics are weighed to correctly reflect the proportion of female-headed households in the population.
* The first five variables are measured in units of 1992 taka per adult equivalent per month. Agricultural land is
measured in decimals (100 decimals = 1 acre).
* Exchange rate in 1992 was US$1 = 38 taka.
on income, expenditure, asset transactions, time use, loans and gifts were collected at each
round.4 Each round corresponds roughly to a calendar month, with rounds starting in
September 1991 for village A and October 1991 for village B.
Two consumption measures were created for each of the sampled households over the 12
months: food consumption and all (nondurable) consumption. Food consumption includes
consumption from own production of wheat and rice, purchased wheat and rice, other food
purchases (e.g. vegetables and pulses), other food consumption from own produce, net
meals received as wages or gifts. All consumption adds expenditure on services and other
nondurable purchases (e.g. tobacco and medicines). Measures of household income and
revenue were also created.5 All income includes net profits from own crop production, net
wages earned, net profits from trading, self-employment and business activities, and rent.
Revenue comprises gross profits and wages earned. Neither the all income measure nor the
revenue measure includes net borrowing and saving or net gifts received. Each of these are
assumed to be smoothing devices used to augment consumption when incomes are low or to
put aside resources when incomes are high. Field observation and the detailed transaction
data indicate that zero nominal interest loans were common within both villages, and gift
exchange in the form of meals or food was widespread. Table 2 reports summary
statistics for the two villages. The unit of observation is a household. Village B is
wealthier than village A: adult equivalent monthly consumption, income, and revenues
4
Resident research teams of two male and two female interviewers who were recent university graduates
lived in each village between June 1991 and November 1992. The principal investigator spent approximately 1
week every month in the villages to supervise and participate in data collection.
5
Consumption, income and revenue are in per adult equivalent terms throughout the paper. The following
age – sex weights were used: 1.0 for adult males, 0.9 for adult females, 0.94 for males aged 13 – 18, 0.83 for
females aged 13 – 18, 0.67 for children aged 7 – 12, 0.52 for children aged 4 – 6, 0.32 for toddlers aged 1 – 3, and
0.05 for infants. These weights are the same as those used by Townsend (1994) which are based on a south Indian
dietary survey.
Table 2
Household categories and microcredit membership
Village A Village B
Total number of sampled households 112 117
Landless (arable land Q 0.5 acres) 42 47
Landed 70 70
Female-headed 24 31
Male-headed 88 86
Uneducated household head 71 71
Educated household head 41 46
Members of microcredit NGO 38a 38b
Nonmembers 74 79
Members of ASA 12 3
Members of BRAC 16 19
Members of Grameen 10 17
In Village B, one HH is a member of both BRAC and Grameen.
a
Fourteen microcredit members were landless in 1992 in Village A.
b
Seventeen microcredit members were landless in 1992 in Village B.
are all higher in B. The mean adult equivalent monthly income is 429 taka
(approximately US$11) in village A and 537 taka in village B (approximately
US$14 using the 1992 exchange rate of US$1 = 38 taka). The daily agricultural wage
in both villages in 1991– 1992 was 20 taka plus two meals, valued at about 7 taka
each. So a day’s agricultural work was worth less than US$1. Since the coefficient of
variation of consumption is lower than that of income or revenues for both villages,
there appears to be some risk sharing by households. There is considerable idiosyn-
cratic (and hence diversifiable) risk in this economy: incomes and revenue do not
comove across households.
Wodon (1997) calculates a poverty line of 425 taka monthly per capita consumption for
Rajshahi for 1991 – 1992 using the cost of basic needs method.6 According to his
estimates, 62% of rural Rajshahi and 47% of rural Bangladesh is below the poverty line.
Village A is slightly poorer than the average Rajshahi village: 68% of the sampled
households are below the poverty line. On the other hand, village B is slightly richer than
the average Rajshahi village but still poor relative to the national average: 54% of its
sampled households are below the poverty line.7
A resurvey of both these villages was carried out in 1995. In particular, we have
information on the number of households (or their splits) that had joined Grameen, BRAC
and ASA by 1995.8 This is the unique feature of the data that we exploit in our analysis. In
6
The poverty line was set by computing the (district specific) cost of a food basket that enabled households
to meet the normative nutritional requirement of 2.5 kcal, and adding to this an estimated allowance for nonfood
consumption.
7
The percentages of households below the poverty line have been weighted to reflect the oversampling of
female-headed households.
8
Households that split between 1992 and 1995 were treated as a single unit. Field observations suggest that
split households maintain very close social and economic ties, and appear to act as a single large unit.
1991– 1992, when the first survey was conducted, Grameen Bank had only begun to
establish their presence in the two villages. In village A, 5 sampled households took their
first loans from Grameen before the end of the survey in 1992, of which 2 took loans in the
last 2 months of the survey. In village B, 14 sampled households took Grameen loans
before the end of the survey in 1992, of which 4 households took loans in the last quarter
of the survey.9 By 1995, Grameen, BRAC and ASA had firmly established their loan
program in these villages. Throughout this paper, consumption and income data (and
therefore our measures of vulnerability) are based on the 1991– 1992 data. Microcredit
membership data, on the other hand, are derived from the 1995 resurvey. A household’s
average consumption level and vulnerability may of course change between 1991– 1992
and the year just before it joined a microcredit program, but we have no data for the
intervening period to control for this possibility. However, once a microcredit program
begins lending in an area it expands quickly and other microcredit programs follow soon
after. So it is likely that households that joined microcredit programs did so soon after the
1991– 1992 data was collected.
Table 2 summarizes the village composition in terms of several household categories
(landlessness, female headship, education level of the household head).10 In addition, the
number of households with at least one microcredit member by 1995 is reported. About
one third of all sampled households in each village had joined a microcredit program by
this time. Grameen membership had risen to 10 sampled households in village A and 17
sampled households in village B, and total microcredit membership was 38 sampled
households in each village.
Microcredit loan use is typically quite fungible. So even though the loans are some-
times given for production, they can be diverted for consumption smoothing. Todd (1996)
describes how paddy husking and the purchase of cows are commonly reported uses of
Grameen loans, but in practice loans are often used to lease land and repay other loans (a
consumption smoothing activity).
3. Measuring vulnerability
Under commonly made assumptions (separability of consumption and leisure, com-

mon rates of time preference, additively separable preferences over time) efficient risk
sharing within a village implies that household consumption should move only with
aggregate consumption and not with household income (Deaton, 1997, pp. 372 –383;
Townsend, 1994). We measure vulnerability based on this risk-sharing test. Instead of
looking at specific smoothing mechanisms (such as loans, gifts, savings or asset sales),
9
For those households that had already joined Grameen before the end of the survey, our estimates of poverty
and vulnerability levels may be biased by the loans taken from Grameen. We discuss the likely direction of such
bias, wherever applicable, in Section 4.
10
In regressions of household average monthly (all) consumption on household characteristics, we found that
female headed and larger households, and households with less arable land and more elderly members had lower
consumption in both villages.
vulnerability is derived from household outcomes (consumption and income). In this

section, we describe how we estimate a baseline vulnerability measure for each household
and propose alternative vulnerability measures to study its robustness. We also report on
the correlation of vulnerability with poverty and with measures of the variability of
consumption.
3.1. Risk sharing
A key feature of efficient risk sharing within the village is that changes in log
marginal utility of consumption must be equated across households at each date and
state. Suppose households have a constant absolute risk aversion (CARA) utility
function:

h 1 h cht
u ðcht Þ ¼ h nt exp r h
ft nt
where cth denotes consumption for household h at time t, nth is the (age – sex adjusted)
number of male adult equivalents in the household at time t, fth is a preference shock
and r is the coefficient of absolute risk aversion. Efficient risk sharing then implies the
following for each household h:
h
c 1 1
D th ¼ jt Dlnfht ð1Þ
nt r r
where Dxt u xt xt 1 for any variable xt, and jt is the first difference in logarithms of
the appropriately discounted multiplier associated with the aggregate resource constraint.
So consumption across agents should comove (modulo variations due to preference
shocks), and changes in a household’s consumption should not be affected by changes
in that household’s income.
Eq. (1) constitutes the basis of our estimation strategy. If full risk sharing is in place
and preference shocks can be treated as mean zero error terms that are uncorrelated with
changes in income and with time dummies, then changes in per-adult-equivalent
consumption over time should comove across households. Household consumption
should only be affected by aggregate fluctuations in the village, and not by idiosyncratic
shocks to the household’s own income or resources. Our estimation strategy focuses on
identifying individual households within each village that are vulnerable to idiosyncratic
risk.
So far we have implicitly assumed that the household utility stays the same if one
doubles both the total consumption in the household and the number of adult equivalents.
But a bigger household may be more efficient. First differencing consumption will
eliminate economies of scale that are fixed through time, however. Only three households
in village A and two in village B change composition during the 12 months of the sample,
so we can safely ignore household economies of scale.
3.2. Estimation
We estimate a linear regression model based on Eq. (1) to identify vulnerable

households. If we treat the fht preference shocks as mean zero error terms that are
uncorrelated with the other regressors, Eq. (1) suggests the following regression
equation:
Dc̃ht ¼ ah Dỹht þ /t MDt þ eht ; ð2Þ
where c̃th u cth/nth denotes per-adult-male equivalent consumption of household h in

month t, ỹth is (per-adult-equivalent) household income at time t, and MDt is a month
dummy, that equals one for observations at time t, zero otherwise. The coefficient /t
captures (1/r)jt from Eq. (1), which is proportional to a measure of the aggregate
resource constraint at date t. The error term, eth, is assumed to be uncorrelated with the
right-hand side variables and to be mean zero. We assume the following covariance
structure:
2
1. sht u Var(eth) = sh2 bh, t;
2. sh,ts u Cov(eth,esh) = 0bh, t p s;
3. shk,ts u Cov(eth,esk ) = 0bh p k, t p s;
4. shk,t u Cov(eth,etk ) = 0bt, h p k.
The first assumption on heteroskedasticity is motivated by the results of several tests

(such as the White general test and the Glesjer test based on the regression of the squared
residuals on several household variables). Intuitively, it seems reasonable to allow the
variance of the residuals to vary across households, since they have very different sizes,
landholdings, consumption and income levels. The other assumptions are quite standard.
We have tested for the presence of contemporaneous correlation across households using a
variety of methods, and we find very little evidence, if any, of such correlation. We
estimate Eq. (2) via FGLS, postulating that the individual household variance depends on
several observable characteristics (such as landholdings or household size) in the
following way:
s2h ¼ s2 expðbVzh Þ:
The regression equation (Eq. (2)) has been estimated numerous times in the literature
on efficient risk sharing, for the special case in which ah = a for all h. Under the null
hypothesis of full risk sharing in the village as a whole, Eq. (1) implies that a must be
equal to zero. A significantly positive estimated coefficient â implies that the full risk-
sharing hypothesis can be rejected for the village as a whole. Both villages in our dataset
fail the full risk-sharing test. For village A, â = 0.0366 (with a p-value of 0.005) and for
village B, â = 0.0299 (with a p-value of 0.019).
For this paper, we are interested in identifying specific households for which the
implications of full risk-sharing models are rejected. So we estimate a separate ah
parameter for each household h. The baseline regression (Eq. (2)) is estimated separately
for each village. The distribution of the estimates âh is depicted in Fig. 1. Nineteen
households in village A and 18 in village B had positive and significant âh using a 10%
statistical significance level.11 The mean of the âh estimates is 0.11 in village A and 0.19 in
village B.
The theoretical framework only offers a test of whether full risk sharing holds for a
given household. To interpret the size of a significantly positive vulnerability coefficient,
we consider autarky with limited savings possibilities as an alternative to the full risk-
sharing model. Under this alternative, ah would be positive. As the household moves from
full risk sharing to autarky, its ah parameter will increase.12 Our baseline measure of
vulnerability therefore is the estimated âh.
One may be concerned that the observed income fluctuations Dỹth may just reflect
seasonal variations that affect the village as a whole (such as planting and harvest, rainy
and dry seasons), and not household-specific shocks. But since such village-wide shocks
are controlled for by the month dummy in the baseline regression (Eq. (2)), our âh
estimates detect responses of household consumption to household-specific shocks
controlling for village-wide fluctuations. Further, there is substantial household specific
(and hence insurable) risk in these economies. The deviations of household income from
the village average are largely uncorrelated across households.
So far we have restricted all agents to have the same coefficient of absolute risk
aversion r. Our vulnerability estimates from regression (Eq. (2)) may be biased as a
consequence. In particular, our âh estimates may simply reflect individual differences in
the attitude towards risk: a less risk averse household may be associated with a higher âh,
whereas a more risk averse household may exhibit a lower âh. If some agents were more or
less risk averse than others, with rh denoting household h’s degree of risk aversion, then
Eq. (1) would become (ignoring the preference shock term for simplicity):
h
c 1
D th ¼ jt : ð3Þ
nt rh
Suppose the true model were Eq. (3) butP we estimate regression (Eq. (2)) instead. Then
/̃t is an estimate of mjt (where m ¼ N1 h mh and mh u r1h ), and we have to evaluate the
possibility that omitting the term (mh m)jt may bias our âh estimates. Suppose we assume
that mh is uncorrelated with Dỹth, since the latter reflects idiosyncratic shocks that hit
household h which should be independent of that household’s degree of risk aversion. That
implies our âh estimates from regression (Eq. (2)) will be biased only if there is a nonzero
correlation between Dỹth and jt, i.e. between changes in individual income and changes in
the aggregate resource constraint over time. We proxy for jt using the change in aggregate
consumption. In the data, the correlation between changes in household income and
changes in aggregate consumption is negligible and statistically insignificant for both
villages (0.009 in village A and 0.025 in village B). Therefore, differences in risk
aversion will not bias our vulnerability estimates.
11
These numbers fall to 14 and 10, respectively, using a 5% significance level.
12
Roughly one quarter of the estimated parameters are negative, but fewer than 6% of the total are negative
at the 10% level. A couple of households in either village have an estimated vulnerability parameter larger than
one. This is presumably because of measurement error.
Fig. 1. Estimated Vulnerability Coefficients.

3.3. Robustness
We estimate several alternatives to the baseline vulnerability measure for each household.
(1) Truncated. The first round of data collected in both villages is more susceptible to
measurement error than subsequent rounds because the enumerators did not have a starting
inventory of household grain and foodstuff to go by. Therefore, we estimate ah parameters
excluding the first round of data.
(2) Medical expenses. The baseline regression model (Eq. (2)) assumes that preference
shocks are uncorrelated with the other covariates. We also estimate a version of Eq. (1)
where household specific medical expenditures Xth in each period are used to proxy for the
preference shock fth:
Dc̃ht ¼ ah Dỹht þ /t MDt þ dXth þ eht ð4Þ

Sickness shocks can reduce a household’s desire to consume, for instance, and can be
correlated with income shocks.
(3) Female headship. The baseline regression (Eq. (2)) assumes common risk aversion.
While we do not have the degrees of freedom necessary to directly estimate a household-
specific risk aversion coefficient, it is possible to allow the coefficient of absolute risk
aversion to vary across groups. Female headship is associated with higher vulnerability in
regressions of our baseline vulnerability measure on household characteristics.13 There-
fore, it may be the case that female-headed households are more risk averse than male-
headed households. A version of risk-sharing implication (Eq. (1)) allowing for the
coefficient of absolute risk aversion to differ across female- and male-headed households
is
h
c 1 1
D th ¼ F jt F Dlnfht ð5Þ
nt r r
where F is a Household Dummy that is equal to one if household h is a female headed and
zero otherwise. Therefore assuming, as before, that preference shocks can be treated as
mean zero error terms that are uncorrelated with the other regressor, we also estimate the
following version of the model:
Dc̃ht ¼ ah Dỹht þ /t MDt þ /Ft MDt F þ eht ð6Þ
Thus, we estimate a /t parameter for male-headed households and a /t + /1t parameter for
female-headed households to reflect possible differences in risk aversion between the two.
(4) Common components. Measurement error that is common to consumption and
income would bias our vulnerability estimate in Eq. (2). There are components of
consumption (such as net meals received as wages and own produce consumed) that
are also included as income. So we exclude these common components from the measure
of income in our estimation of household vulnerability.
13
Vulnerability is more elusive than poverty. The adjusted R2 from regressions of vulnerability on household
characteristics were much lower than the adjusted R2 from regressions of average household all consumption on
household characteristics. Further, only female headship in the pooled regression and in the village B regression
was significantly negative, by itself and with controls for other household characteristics.
(5) Stronger definition. Our baseline vulnerability measure identifies a household as

vulnerable even if it achieves the ‘‘average’’ level of vulnerability for the village as a
whole. A stronger definition of vulnerability would only consider those households
vulnerable who achieve less insurance than the average in the village. So we construct
the following measure from the baseline: vulnerability is âh if â h>â where â is the village
vulnerability coefficient (Section 3.2), and vulnerability is 0 otherwise.
(6) CRRA utility. To check if our measure of vulnerability is robust to the choice of
utility function, we use a constant relative risk aversion (CRRA) utility specification,
" 1c #

h h 1 h h cht
u ct ¼ f n ;
1c t t nht
to derive the counterpart to Eq. (1):

h
c 1 1
Dln th ¼ jt Dlnfht
nt c c
We estimate the following regression:
Dlnc̃ht ¼ ah DlnR̃ht þ /t MDt þ eht

where R̃th represents per-adult-equivalent revenues for household h at month t. We use
revenues instead of income because the latter can sometimes take negative values, while the
former cannot. Revenues proxy for idiosyncratic risk. The distribution of the âh estimates is
shown in Fig. 1. Twenty-six households had significantly positive âh estimates in village A
and 29 households had significantly positive âh estimates in village B.
Though these six alternative measures of vulnerability are either derived from different
specifications of the baseline model or attempt to control for measurement error, it is
reassuring that the identity of households that have significantly positive âh estimates
remains stable across the alternatives. With the baseline measure a total of 37 households in
both villages have positive and significant âh estimates (at the 10% significance level). With
the female headship measure, 37 households have positive and significant âh estimates, and
only 2 of those households did not have positive and significant âh estimates in the baseline.
With the common components measure, 27 households had positive and significant âh
estimates, and only 3 of those households did not have positive and significant âh estimates
in the baseline. The pattern is similar for the other alternative vulnerability measures.
Consequently, we are confident that the baseline vulnerability measure is fairly robust. In
what follows, we will present results first using the baseline measure, and then using the
alternative measures as a crosscheck.
3.4. Vulnerability, variability and poverty
As one would perhaps expect, poorer households tend to be more vulnerable than richer
households in our data. Vulnerability is significantly negatively correlated with con-
sumption using the pooled data and in village B (Table 5). This finding is stable across all
the alternative vulnerability measures, except for CRRA measure where the correlation is
Table 3
Poverty and microcredit membership: Test of stochastic dominance
Pooled data Village A Village B
Members minus (t-statistics) Members minus (t-statistics) Members minus (t-statistics)
nonmembers nonmembers nonmembers
0.0256 0.7547 0.0154 0.4090 0.0803 1.5172 *
0.0825 1.3334 0.0175 0.2315 0.1768 1.9994* *
0.1911 2.7245* * 0.0665 0.6636 0.3131 3.1653** *
0.1819 2.9260** * 0.1228 1.3485 0.2004 2.1735* *
0.1081 2.0111* * 0.0892 1.2023 0.1889 2.1811* *
0.0739 1.8470* * 0.0183 0.3170 0.1005 1.4370 *
0.0535 1.7661 * 0.0417 1.2758 0.0762 1.4407 *
0.0533 2.0499* * 0.0278 1.0369 0.0897 1.9051* *
0.0200 1.2411 0.0139 0.7298 0.0385 1.2249
The columns report the differences between the CDF of consumption for members and nonmembers, at several
points. Under the null hypothesis that the two samples come from the same distribution, each term is distributed
as a Student’s t with (here) nine degrees of freedom. The t-statistics are in parentheses. The distribution for
members first order dominates that for nonmembers if no term is significantly greater than zero while at least one
is significantly negative. Likewise, dominance of nonmembers over members requires than no term be
significantly negative while at least one is significantly positive.
* Statistically significant at 10% level.
** Statistically significant at 5% level.
*** Statistically significant at 1% level.
negative but not significant. The same pattern emerges from a comparison of the average
vulnerability of households below the poverty line with those above the poverty line in
Table 6. Households below the poverty line have significantly higher average vulnerability
Table 4
Vulnerability and microcredit membership: Test of stochastic dominance
Pooled data Village A Village B
Members minus (t-statistics) Members minus (t-statistics) Members minus (t-statistics)
nonmembers nonmembers nonmembers
0.0200 1.2411 0.0139 0.7298 0.0256 0.9957
0.0204 0.7392 0.0015 0.0448 0.0250 0.6214
0.0160 0.2794 0.0154 0.4090 0.0209 0.2652
0.0021 0.0409 0.0395 0.4634 0.0965 1.2233
0.0005 0.0191 0.1045 1.3033 0.0405 0.9241
0.0004 0.0155 0.0431 0.9472 0.0270 0.7477
0.0130 0.6993 0.0417 1.2758 0.0398 1.2679
0.0132 1.4080 * 0.0278 1.0369 0.0263 1.4389 *
0.0132 1.4080 * 0.0139 0.7298 0.0263 1.4389 *
The columns report the differences between the CDF of vulnerability for members and nonmembers, at several
points. Under the null hypothesis that the two samples come from the same distribution, each term is distributed
as a Student’s t with (here) nine degrees of freedom. The t-statistics are in parentheses. The distribution for
members first order dominates that for nonmembers if no term is significantly greater than zero while at least one
is significantly negative. Likewise, dominance of nonmembers over members requires than no term be
significantly negative while at least one is significantly positive.
* * Statistically significant at 5% level.
Table 5
Correlations between vulnerability, variability, consumption, and microcredit membership
CV V All Food MC
(consumption) (log(consumption)) consumption consumption member
(1) Pooled Data
Vulnerability (baseline) 0.05 0.04 0.12* * 0.14* * 0.07
Vulnerability (truncated) 0.00 0.05 0.14* * 0.16** * 0.04
Vulnerability 0.05 0.04 0.13* * 0.15* * 0.07
(medical expenditures)
Vulnerability (female) 0.04 0.03 0.12* * 0.14* * 0.08
Vulnerability (common) 0.02 0.06 0.10 * 0.11* * 0.03
Vulnerability (strong) 0.15* * 0.11* * 0.15* * 0.16** * 0.05
Vulnerability (CRRA) 0.10 * 0.09 * 0.03 0.03 0.05
CV (consumption) 1.00 0.82** * 0.10 * 0.09 * 0.03
Var(log(consumption)) 1.00 0.02 0.03 0.02
Average all consumption 1.00 0.94** * 0.21** *
Average food consumption 1.00 0.20** *
(2) Village A
Vulnerability (baseline) 0.05 0.04 0.01 0.03 0.07
Vulnerability (truncated) 0.11 0.15 * 0.08 0.06 0.08
Vulnerability 0.05 0.03 0.01 0.03 0.06
Vulnerability (female) 0.04 0.01 0.03 0.05 0.04
Vulnerability (common) 0.09 0.14 * 0.11 0.10 0.08
Vulnerability (strong) 0.17* * 0.16* * 0.00 0.03 0.11
Vulnerability (CRRA) 0.05 0.06 0.07 0.10 0.10
CV (consumption) 1.00 0.82** * 0.11 0.13 * 0.20* *
Var(log(consumption)) 1.00 0.02 0.03 0.22* *
Average all consumption 1.00 0.95** * 0.12
Average food consumption 1.00 0.11
(3) Village B
Vulnerability (baseline) 0.05 0.04 0.21* * 0.24** * 0.16* *
Vulnerability (truncated) 0.10 0.06 0.21* * 0.24** * 0.17* *
Vulnerability 0.05 0.04 0.22** * 0.25** * 0.16* *
Vulnerability (female) 0.04 0.04 0.22** * 0.26** * 0.17* *
Vulnerability (common) 0.04 0.00 0.10 0.12 * 0.13 *
Vulnerability (strong) 0.16* * 0.09 0.25** * 0.26** * 0.16* *
Vulnerability (CRRA) 0.14 * 0.11 0.10 0.11 0.14 *
CV (consumption) 1.00 0.81** * 0.09 0.06 0.14 *
Var(log(consumption)) 1.00 0.05 0.03 0.17* *
Average all consumption 1.00 0.95** * 0.27** *
Average food consumption 1.00 0.28** *
Each cell reports the correlation coefficient between the row and the column variables.
* * Statistically significant at 5% level.
** * Statistically significant at 1% level.
Table 6
Vulnerability, variability, consumption by poverty status and microcredit membership
Poor Nonpoor p-value Member Nonmember p-value
(1) Pooled data
Vulnerability (baseline) 0.18 0.10 0.093 0.20 0.13 0.168
Vulnerability (truncated) 0.21 0.10 0.059 0.20 0.15 0.257
Vulnerability 0.18 0.10 0.082 0.20 0.13 0.158
Vulnerability (female) 0.19 0.11 0.112 0.21 0.13 0.125
Vulnerability (common) 0.09 0.08 0.443 0.12 0.07 0.304
Vulnerability (strong) 0.25 0.18 0.058 0.25 0.21 0.249
Vulnerability (CRRA) 0.19 0.14 0.153 0.19 0.16 0.256
CV (consumption) 0.43 0.52 0.004 0.46 0.47 0.336
Var(log(consumption)) 0.19 0.23 0.068 0.20 0.21 0.387
% Below poverty line 76.30 56.70 0.000
Average all consumption 361.20 440.18 0.000
Average food consumption 287.55 346.59 0.000
(2) Village A
Vulnerability 0.11 0.12 0.458 0.08 0.13 0.228
CV (consumption) 0.43 0.54 0.037 0.40 0.50 0.006
(3) Village B
Vulnerability 0.27 0.09 0.026 0.32 0.13 0.061
CV (consumption) 0.43 0.52 0.029 0.51 0.44 0.060
Each cell reports the average of the row variable, by poor (nonpoor) or by member (nonmember).
A household is poor if it is below the poverty line, and nonpoor otherwise.
The p-value refers to the test that the difference between poor and nonpoor (or member and nonmember) is equal
to zero.
than those above the poverty line for the baseline and half the alternative vulnerability
measures using the pooled data, and for the baseline and all but one of the alternative
vulnerability measures in village B (Table 6).14
Our measure of vulnerability is quite distinct from measures of consumption variability,
such as the coefficient of variation (CV) of consumption or the variance of log
consumption.15 The baseline and most alternative measures of vulnerability are not
significantly correlated with either of the two measures of consumption variability (Table
5). Though poor households tend to be more vulnerable than the rich, rich households
have more variable consumption than the poor. Household consumption is positively
correlated with the CV of consumption using the pooled data and in village A for food
consumption (Table 5). Households below the poverty line have a significantly lower CV
of consumption than those above the poverty line in each village (Table 6). Households
below the poverty line also have a significantly lower variance of log consumption than
households above the poverty line using the pooled data (Table 6). In addition, while
consumption variability may just capture differences in risk aversion across households,
our vulnerability measure is unbiased even when some households are more risk averse
than others, provided the degree of risk aversion is uncorrelated with shocks to household
income (Section 3.2). Consequently, for the rest of the paper, we will be concerned with
vulnerability and not with consumption variability.
4. Results
In this section, we test if households that joined microcredit programs by 1995 were
poorer and more vulnerable in 1991– 1992 than households that did not join microcredit
programs by 1995. We do this by comparing average consumption and average vulner-
ability of members with nonmembers, testing whether the distribution of consumption and
vulnerability of members first order stochastically dominates that of nonmembers, and by
conducting probit regressions of microcredit membership on consumption, vulnerability
and household characteristics.
4.1. Poverty
There is clear evidence that households that joined microcredit programs by 1995 were
poorer in 1991 –1992 than those that did not join. The Cumulative Distribution Functions
(CDFs) of average monthly consumption and income for microcredit members and
nonmembers are shown in Fig. 2. In both villages, and more so in village B, the
distribution for nonmembers is shifted to the right. The average monthly consumption
for members is significantly lower than that of nonmembers using the pooled data, and in
both villages separately (Table 6). In village A, average monthly all consumption is 36
taka (approximately US$1) lower for members, which is 10% of the village average
14
Jalan and Ravallion (1999) also find that the poor are significantly more vulnerable than the rich in rural
China.
15
Morduch (1998) uses the latter as a proxy for household vulnerability.
Fig. 2. Poverty, vulnerability and microcredit membership.
monthly consumption. In village B, average monthly all consumption is 88 taka

(approximately US$2) lower for members, which is 20% of the village average monthly
consumption. Correlations between average monthly consumption and membership are
also significantly negative using the pooled data and in village B and negative but barely
insignificant in village A (Table 5).
In addition to just looking at specific moments of the distribution of consumption of
members and nonmembers, we also use a nonparametric test to check if the distribution of
average monthly consumption for nonmembers first order dominates that for members.
This test is based on Anderson (1996) and is described in more detail in Appendix A. The
results in Table 3 show that the distribution of consumption for nonmembers significantly
first-order stochastically dominates the distribution for members using the pooled data and
in village B. The test of first-order stochastic dominance goes in the same direction, but is
76
S. Amin et al. / Journal of Development Economics 70 (2003) 59–82
Table 7
Probit regression of microcredit membership on vulnerability, consumption and HH characteristics
(1) * (2) * (3) * (4) * (5) * (6) *
Pooled data
Vulnerability 0.0693 (0.318) 0.0425 (0.544) 0.1059 (0.518) 0.0561 (0.403)
Average consumption 0.0007 (0.000) 0.0007 (0.000) 0.0006 (0.004)
Agricultural land 0.00004 (0.127) 0.0000 (0.819)
Poor (dummy) 0.1920 (0.007)
Vulnerability Poor 0.0620 (0.732)
Female headed 0.0489 (0.633)
Age of head 0.0087 (0.009)
Household size 0.0183 (0.550)
HH structure 0.0012 (0.989)
Uneducated head 0.0598 (0.406)
# Old in HH 0.0103 (0.905)
# Children in HH 0.0296 (0.432)
Pseudo R2 0.0032 0.0453 0.0183 0.0465 0.0345 0.0914
Village A
Vulnerability 0.0913 (0.498) 0.1023 (0.457) 0.2643 (0.250) 0.0237 (0.862)
Poor (dummy) 0.2148 (0.042)
Age of head 0.0035 (0.515)
# Old in HH 0.1114 (0.429)
# Children in HH 0.0719 (0.265)
Pseudo R2 0.0028 0.0209 0.0126 0.0241 0.0376 0.1383
S. Amin et al. / Journal of Development Economics 70 (2003) 59–82

Village B
Vulnerability 0.1378 (0.048) 0.1006 (0.144) 0.0263 (0.908) 0.0506 (0.441)
Poor (dummy) 0.1715 (0.068)
Age of head 0.0130 (0.003)
# Old in HH 0.1114 (0.278)
# Children in HH 0.0234 (0.659)
Pseudo R2 0.0193 0.0653 0.0286 0.0763 0.0579 0.2517
* The marginal effects dF/dx, evaluated at the sample mean, are reported here. p-values are reported in parentheses.
* All statistics are weighed to correctly reflect the proportion of female-headed households in the population.
77
barely insignificant, in village A. Thus, we have a strong indication that members are
poorer than nonmembers, since first-order stochastic dominance implies both second- and
third-order dominance.
Another way of evaluating the success of microcredit programs in reaching the poor is
by using the poverty line. The proportion of microcredit members below the poverty line is
79% in village A and 74% in village B, which is significantly higher than the proportion of
nonmembers below the poverty line in both villages (Table 6). The proportion of
microcredit members below the poverty line is also higher than the village averages,
the Rajshahi district average and the national average of 47% (see Section 2). Therefore,
the probability that a household was below the poverty line when it received a loan is
substantially higher than the probability of a randomly picked Bangladeshi household
being below the poverty line.
A probit regression in Table 7 shows that a 100 taka or 24% decrease in monthly
consumption at the mean gives a household a 6% or 7% higher probability of receiving a
loan on average. This is statistically significant using the pooled data and in village B
conditional on vulnerability (column 4) and controlling for other household character-
istics (column 6). This is significant for village A controlling for vulnerability (column
4) and significant at the 11% level controlling for other household characteristics
(column 6).
Since one-third of the households that eventually became microcredit members in
village B joined before the end of the 12 rounds of data collection in 1992, there is the
possibility that the loans they received may have affected the consumption of those
households, and thus could bias our results. Provided loans have a nonnegative impact on
household consumption, however, such a bias will only strengthen our finding that
microcredit reaches the poor. Our results may therefore underestimate microcredit’s
effectiveness at reaching the poor in village B.
Finally, notice that having more arable land is not significantly associated with a lower
probability of joining a microcredit program (Table 7, column 3). So that though
microcredit does reach the consumption-poor, there is no evidence that it reaches the
relatively landless.16
4.2. Vulnerability
There is weak evidence that households that joined microcredit programs by 1995 were
more vulnerable in 1991 –1992 than those that did not join. The Cumulative Distribution
Functions (CDFs) of vulnerability for microcredit members and nonmembers are shown in
Fig. 2. Notice that though members appear to be more vulnerable in village B, non-
members appear to be more vulnerable in village A. The average vulnerability of members
is significantly higher than that of nonmembers in village B but not in the pooled data
(Table 6). This is robust to all alternative vulnerability measures except for the Common
16
Only households with less than half an acre of cultivated land are typically eligible to join Grameen,
BRAC and ASA. Table 3 shows that over half the households that joined microcredit programs in 1995 did not
meet this eligibility criterion in 1991. Since some of these households may have split or sold land to meet the
landholding criterion, this may be an overestimate of mistargeting.
measure, where it is barely insignificant. Correlations between vulnerability and member-

ship are also significantly negative only in village B but not in the pooled data (Table 5),
and this finding is robust to alternative vulnerability measures. The results from the
nonparametric tests in Table 4 show that the distribution of vulnerability for members
significantly first order stochastically dominates the distribution for nonmembers using the
pooled data and in village B.
The probit regression in Table 7 yields the same pattern across the two villages (column
1) and again it is robust to most alternative vulnerability measures. Raising vulnerability
by 0.2 at the mean raises the probability of being a microcredit member by about 3% in
village B. Interestingly, vulnerability does not significantly increase a household’s
probability of joining if we control for poverty (column 4) and other household character-
istics (column 6). This gives a potential explanation for the differences across villages:
since village B is richer than village A, vulnerable households may be more likely to join
in village B only because they are not as poor.
This result is subject to the following caveat. It seems reasonable to suppose that
microcredit membership either decreases a household’s vulnerability or leaves it unaltered.
Since a third of the microcredit members in village B joined Grameen before the end of the
12 rounds of data collection in 1992, these households may appear less vulnerable than
they were before joining Grameen. Consequently, we may underestimate microcredit’s
effectiveness at reaching the vulnerable in village B.
4.3. Interaction
Finally, we wish to check if vulnerable households below the poverty line have a higher
probability of joining microcredit programs than vulnerable households above the poverty
line. We are particularly interested in this since the vulnerable poor are those that would
likely gain the most from gaining access to smoothing devices, thus avoiding possible
destitution. Therefore, we introduce an interaction term in the probit regression in Table 7
(column 5). The coefficient on the interaction term is insignificant using the pooled data
and in village B. But interestingly, the interaction has a significantly negative coefficient in
village A (and the coefficient remains significantly negative when we replace the baseline
vulnerability measure with the alternative vulnerability measures that use truncated data,
control for medical expenses and exclude common components). For households that are
below the poverty line in village A, raising vulnerability by 0.1 or 9% at the mean lowers
the probability of joining a microcredit program by about 2.2%, and this is significant at
the 10% level.
So microcredit does not reach the vulnerable poor in the relatively richer village B, and
appears to exclude the vulnerable poor in the poorer village A. In other words, poor
households that do join tend to have better access to insurance and smoothing devices than
those who do not. The vulnerable poor may either choose not to join or they may be
excluded by the microcredit program. There is some anecdotal evidence consistent with
the latter. For instance, a bank officer told Todd (1996, p.173) that he was looking for
borrowers that were ‘‘not hopeless.’’ This suggests that subsidized loans may not be the
appropriate strategy to reach the vulnerable poor, because of their potentially higher
default risk.
Microcredit’s success at reaching the poor is robust to introducing the interaction term.
Households below the poverty line have a 19% higher chance of joining a microcredit
program using the pooled data (Table 7, column 5). This coefficient remains significantly
positive when we use alternative vulnerability measures instead of the baseline.
5. Conclusions
This paper uses panel data from two Bangladeshi villages to test if microcredit reaches
the poor and vulnerable. This analysis is possible due to the convenient timing of data
collection. Households were extensively surveyed in 1991 – 1992, when microcredit
programs had only a small presence in the study villages. Households were subsequently
resurveyed in 1995 by which time microcredit programs had firmly established them-
selves.
This paper studies the interaction between microcredit selection and the vulnerability of
rural households. Vulnerability refers to the inability of households to insure against
idiosyncratic risks, and it is distinct from measures of consumption variability. It provides
an example of how an antipoverty program, which is successful at reaching the poor, may
exclude those most in need of assistance, the vulnerable poor. The forces that make some
poor households vulnerable may also make them greater risks for microcredit providers,
however. This may explain why microcredit programs are unsuccessful at reaching the
vulnerable poor in these villages, and suggests that subsidized credit may have limits as an
antipoverty strategy.17
Our vulnerability estimates for instance are based on monthly data collected over 12
months. It would be interesting to see if estimates of household vulnerability are stable
from year to year, or if some households are more vulnerable to annual income shocks
than to monthly shocks. It would also be useful to know if microcredit’s success at
reaching the vulnerable poor depends on village characteristics, such as average village
poverty or vulnerability. How does the process of forming borrower groups affect the
likelihood that the vulnerable poor join microcredit programs? Do relatively poor and
vulnerable households had higher default rates after they join microcredit programs? We
leave these questions for future research with more detailed data sets.
Acknowledgements
We thank Chris Hartwell, Anna Musatti and especially Dilip Parajuli for excellent
research assistance. Suggestions from two anonymous referees, Abhijit Banerjee, Marcel
Fafchamps, Chris Flinn, Wilbert can der Klaauw, Anna Paulson, Robert Townsend, Frank
Vella and seminar participants at Boston University, Brown, Harvard, M.I.T., Penn State,
17
Subsidized microcredit may need to be supplemented with grants that reach the vulnerable poor. Rai and
Sjöström (2001) show giving both grants and subsidized loans is the efficient government intervention in a model
where the poor are credit constrained.
Population Council, and Yale are much appreciated. Ashok Rai gratefully acknowledges
initial funding from the Social Science Research Council’s IPFP program. Giorgio Topa
gratefully acknowledges financial support from the C.V. Starr Center at NYU. All errors
are our own.
Appendix A. First-order stochastic dominance tests
Suppose the cumulative distribution function of consumption for N (for nonmembers)

and M (for members) is given by FN (c) and FM (c), respectively. Let C be the range of
consumption for both these distributions first-order stochastic dominance of distribution N
over M is equivalent to the condition:
FN ðcÞVFM ðcÞ; FN ðci ÞpFM ðci Þ for some i; bcaC: ð7Þ

Anderson’s (1996) test of first-order stochastic dominance is quite straightforward to
implement. The idea is to partition the combined sample for N and M into k equal intervals
and compute theN empirical frequencies of the N and M samples in each interval: for
x
example, pNi ¼ niN , i = 1,. . .,k, where xiN is the number of observations in the N sample that
fall in interval i, and nN is the total number of observations in the N sample. Let If be a
k k lower-triangular matrix ofPones. Since the cumulative distribution function at a point
j
j can be computed as Fðcj Þ ¼ i¼1 pi, a test of condition (7) translates into the following
hypothesis test:
H0 : If ð pN pM Þ ¼ 0 against H1 : If ðpN pM ÞV0
In particular, first-order dominance of distribution N over M requires that no element of the

vector If ( pN pM) be significantly greater than zero, while at least one element is
significantly negative. The test is symmetric, so first-order dominance of distribution M
over N requires that no element of the vector If ( pN pM) be significantly negative, while
at least one element is significantly positive. Finally, the test statistic If ( pN pM) is
asymptotically distributed as a N(0,IfVIfV) under the null hypothesis, where V can be
estimated using the empirical frequencies p of the combined sample. See Anderson (1996)
for the details.
References
Amin, S., 1998. Family structure and change in rural Bangladesh. Population Studies 52, 201 – 213.
Anderson, G., 1996. Nonparametric tests of stochastic dominance in income distributions. Econometrica 64,
1183 – 1193.
Deaton, A., 1997. The Analysis of Household Surveys: A Microeconometric Approach to Development Policy
(Johns Hopkins/World Bank).
Jalan, J., Ravallion, M., 1999. Are the poor less well insured? Evidence on vulnerability to income risk in rural
China. Journal of Development Economics 58 (1), 61 – 81.
Morduch, J., 1998. Does Microfinance Really Help the Poor? New Evidence from Flagship Programs in Ban-
gladesh, mimeo, New York University.
Morduch, J., 1999. The microfinance promise. Journal of Economic Literature 37, 4 (December).
Navajas, S., Meyer, R., Gonzales-Vega, C., Schreiner, M., Rodriguez-Meza, J., 1998. Microcredit and the Poorest
of the Poor: Theory and Evidence from Bolivia, mimeo, Ohio State University.
New York Times, 1997. Micro-Loans for the Very Poor (editorial), February 16.
Pitt, M., Khandker, S., 1998a. The impact of group based credit programs on poor households in Bangladesh:
does the gender of participants matter? Journal of Political Economy 106, 5.
Pitt, M., Khandker, S., 1998b. Credit Programs for the Poor and Seasonality in rural Bangladesh, mimeo, Brown
University and World Bank.
Rai, A., Sjöström, T., 2001. Grants Vs. Investment Subsidies. Harvard Center for International Development
Working Paper.
Ravallion, M., Datt, G., 1995. Is targeting through a work requirement efficient? Some evidence from rural India.
In: van de Walle, D., Nead, K. (Eds.), Public Spending and the Poor: Theory and Evidence. Johns Hopkins
Univ. Press, Baltimore.
Shams, K., 1992. Designing Effective Credit Delivery System for the Poor: The Grameen Bank Experience.
Grameen Bank, Dhaka.
Todd, H., 1996. Women at the Center. Westview Press, Boulder.
Townsend, R., 1994. Risk and insurance in village India. Econometrica 62, 539 – 591.
Vogel, R., 1984. The effect of subsidized agricultural credit on income distribution in Costa Rica. In: Adams, D.,
Graham, D., von Pischke, J.D. (Eds.), Undermining Rural Development with Cheap Credit. Westview Press,
Boulder, pp. 133 – 145.
Wodon, Q., 1997. Food energy intake and the cost of basic needs: measuring poverty in Bangladesh. Journal of
Development Studies 34 (2), 66 – 101.
World Bank, 2001. World Development Report on Poverty and Development. World Bank, Washington, DC.

Amin 2003

Uploaded by

Copyright:

Available Formats

Amin 2003

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Amin 2003

Uploaded by

Copyright:

Available Formats

Journal of Development Economics 70 (2003) 59 – 82

Does microcredit reach the poor and vulnerable?

JEL classification: O16; I38; Q12

Subsidized credit has a disappointing history of being politically manipulated and

Under commonly made assumptions (separability of consumption and leisure, com-

vulnerability is derived from household outcomes (consumption and income). In this

3.1. Risk sharing

We estimate a linear regression model based on Eq. (1) to identify vulnerable

Dc̃ht ¼ ah Dỹht þ /t MDt þ eht ; ð2Þ

where c̃th u cth/nth denotes per-adult-male equivalent consumption of household h in

The first assumption on heteroskedasticity is motivated by the results of several tests

Fig. 1. Estimated Vulnerability Coefficients.

Dc̃ht ¼ ah Dỹht þ /t MDt þ dXth þ eht ð4Þ

(5) Stronger definition. Our baseline vulnerability measure identifies a household as

to derive the counterpart to Eq. (1):

We estimate the following regression:

Dlnc̃ht ¼ ah DlnR̃ht þ /t MDt þ eht

3.4. Vulnerability, variability and poverty

Fig. 2. Poverty, vulnerability and microcredit membership.

monthly consumption. In village B, average monthly all consumption is 88 taka

S. Amin et al. / Journal of Development Economics 70 (2003) 59–82

measure, where it is barely insignificant. Correlations between vulnerability and member-

Appendix A. First-order stochastic dominance tests

Suppose the cumulative distribution function of consumption for N (for nonmembers)

FN ðcÞVFM ðcÞ; FN ðci ÞpFM ðci Þ for some i; bcaC: ð7Þ

H0 : If ð pN pM Þ ¼ 0 against H1 : If ðpN pM ÞV0

In particular, first-order dominance of distribution N over M requires that no element of the

You might also like