This paper aims to assess the existence of gender discrimination in small business lending using a unique dataset from a Brazilian microfinance institution (MFI). The paper proposes a new method for detecting discrimination by comparing the gender coefficient in loan size regressions to the gender coefficient in regressions of loan default rates. Applying this method, the empirical results point to gender discrimination, as women receive significantly worse credit conditions despite being found to be creditworthy. Additionally, reducing information asymmetry through ongoing client relationships does not remedy the disadvantage faced by women borrowers.

1. Credit to Women Entrepreneurs:
The Curse of the Trustworthier Sex
Isabelle Agier and Ariane Szafarz
Women entrepreneurs are known not only to reimburse loans swifter than men, but also to
receive smaller loans. However, on average women have smaller-scope business projects
and are poorer than men. A deeper investigation is thus required in order to assess the
existence of gender discrimination in small-business lending. This is precisely the aim of this
paper. Its contribution is twofold. Firstly, it proposes a new estimation method for assessing
discrimination in loan allocation. This method operationalizes the theoretical “double
standard” approach developed by Ferguson and Peters (1995, Journal of Finance).
Secondly, this paper applies the new methodology to an exceptionally rich database from a
Brazilian microfinance institution. The empirical results point to gender discrimination.
Additionally, it is shown that reducing the information asymmetry through relationship brings
no remedy to the curse of the trustworthier sex.
Keywords: Small Business, Microcredit, Gender, Loan Size, Denial
Rate, Default
JEL Classifications: G24, L26, O16, M13
CEB Working Paper N° 11/005
February 18, 2011
Université Libre de Bruxelles - Solvay Brussels School of Economics and Management
Centre Emile Bernheim
ULB CP114/03 50, avenue F.D. Roosevelt 1050 Brussels BELGIUM
e-mail: ceb@admin.ulb.ac.be Tel. : +32 (0)2/650.48.64 Fax : +32 (0)2/650.41.88

2. Credit to Women Entrepreneurs:
The Curse of the Trustworthier Sex∗
Isabelle Agier†
Ariane Szafarz‡
This version: February 18, 2011
Abstract
Women entrepreneurs are known not only to reimburse loans swifter than men,
but also to receive smaller loans. However, on average women have smaller-scope
business projects and are poorer than men. A deeper investigation is thus required
in order to assess the existence of gender discrimination in small-business lending.
This is precisely the aim of this paper.
Its contribution is twofold.
Firstly, it
proposes a new estimation method for assessing discrimination in loan allocation.
This method operationalizes the theoretical double standard approach developed
by Ferguson and Peters (1995,
Journal of Finance ).
Secondly, this paper applies the
new methodology to an exceptionally rich database from a Brazilian micronance
institution. The empirical results point to gender discrimination. Additionally, it
is shown that reducing the information asymmetry through relationship brings no
remedy to the curse of the trustworthier sex.
Keywords: Small Business, Microcredit, Gender, Loan Size, Denial
Rate, Default
JEL codes: G24, L26, O16, M13
∗
The authors thank Cécile Abramowicz, Marie Brière, Valentina Hartaska, Marc Labie,
Bruce Wydick, and the participants to the CERMi Research Day (Mons, October 2010)
for helpful discussions and suggestions.
†
UMR 201 - Développement et Sociétés (Paris I Sorbonne / IRD) and CERMi, Email:
isabelleagier@gmail.com
‡
Université Libre de Bruxelles (ULB), SBS-EM, Centre Emile Bernheim, and CERMi,
Email: aszafarz@ulb.ac.be
1

3. It is extremely important (...) to conduct research into the social processes of
discrimination and the politics of access, control, agency, and empowerment.
Little can be assumed about gendered relations of disadvantage. They require
empirical specication which in turn requires micro-level research Saith and
Harriss-White (1999, p. 492).
1
Introduction
Women-owned businesses are taking an increasing importance in the economy. According to Jalbert (2000), the percentage of female business owners
in the world passed from 13% in 1970 to 20% in 1990.
1
Despite this favor-
able evolution, access to credit for female entrepreneurs remains a concern
for policymakers and researchers (Greene et al., 2003; Gatewood et al., 2004;
Jamali, 2009). Although women tend to create smaller rms, lack of capital
is still a major obstacle to them. Indeed, several studies show that, on average, female entrepreneurs are less nanced than male ones (see, e.g., Riding
and Swift (1990) for Canada, Verheul and Thurik (2001) for the Netherlands,
Alsos, Isaksen and Ljunggren (2006) for Norway, Alesina, Lotti and Mistrulli
2
(2008) for Italy).
By focusing on poor female entrepreneurs in developing countries, microcredit has brought to light the underestimated potential of female self- employment.
Notably, the microcredit industry has proved on a large scale that
women are more trustworthy than men in terms of repayment conduct (Armendáriz and Morduch, 2000). Still, Buvinic and Berger (1990); Fletschner
(2009) and Agier and Szafarz (2010) show that women keep being more
3
credit-rationed than men
by micronance institutions (MFIs).
At rst sight, one might be puzzled by the combination of women being
more reliable and receiving smaller loans. However, this combination does
not per se imply the presence of gender discrimination.
Indeed, gendered
repayment rates are established irrespectively of the personal and business
1 More precisely, the percentage of women business owners rose between 1970 and 1990
from 17.5% to 25% in Africa, from 8% to 11% in Asia and the Pacic, from 33.5% to 28%
in Eastern Europe, from 11% to 24% in Latin America and the Caribbean, from 11% to
19% in Western Europe and other. For the US, Gatewood et al. (2004) state that: From
1997 to 2004, the number of women-owned rms grew at a rate of 17 percent (...)
in
comparison to a 9 percent growth in the number of rms overall.
2 Some papers do not share this conclusion (Haines, Orser and Riding, 1999).
3 Credit rationing is to be understood here as lower loans granted to women, and not
higher loan denial like in Stiglitz and Weiss (1981). This point is further discussed in Agier
and Szafarz (2010).
2

4. characteristics of the borrowers.
Moreover, men and women entrepreneurs
dier in at least two respects: 1) women are poorer than men on average,
4
and 2) women have smaller-scope business projects. Besides, smaller loans
typically generate higher operational and monitoring costs for the lender
(Morduch, 1999; Armendáriz and Szafarz, 2011). Therefore, unconditional
statistics might be misleading. A deeper approach is required to reach robust
conclusions. This is precisely the aim of this paper.
In a companion paper (Agier and Szafarz, 2010) based on an exceptional
database including 34,000 loan applications from a Brazilian MFI, we have
shown that fair access to credit is compatible with the presence of a glass ceiling in loan size (larger female projects are more credit-rationed than comparable male projects). Complementing this analysis thanks to reimbursement
records from the same institution, the present paper investigates whether the
glass-ceiling eect is economically justied.
Existing evidence pertaining to access to credit in developing countries is
mostly based on household surveys. This approach provides valuable information on the demand side of the market but is unable to reect the supplyside perspective. In their literature review, Morrison, Raju and Sinha (2007)
state that: The existing research on credit markets in developing countries
admittedly scarce suggests that by and large women receive unfavorable
treatment not because of discriminatory treatment per se, but rather because
of gender dierences in individual characteristics that are relevant for loan
qualication (p. 39). However, because the body of evidence is demandsided, we argue that this conclusion is premature.
Indeed, as emphasized
by Diagne, Zeller and Sharma (2000), credit limits typically emanate from
the lenders. Unfortunately, due to data unavailability, the way MFIs assess
creditworthiness and grant loans has hardly been investigated yet, let alone
the gender issue.
5
Beneting from exhaustive information gathered by an
MFI on its loan applicants and borrowers over an eleven-year period, our
contribution aims at lling this gap.
Discrimination in the lending industry has been scrutinized in various countries, notably in the US where it is a legal oense.
6
Unfortunately, no con-
sensus has emerged so far regarding the methodology to be used (Dymski,
4 According to ILO (2009), 75% of worldwide poverty aects women.
5 Exceptions include Buvinic and Berger (1990) who obtained data from the Urban
Small Enterprise Development Fund in Peru, and Marrez and Schmit (2009) who analyze
the credit risk of a leading Maghrebian MFI.
6 The US legal framework against discrimination in lending includes the 1968 Fair hous-
ing Act, the 1974 Equal Credit Opportunity Act, and the 1975 Home Mortgage Disclosure
Act. Since 1989, the lenders must report the race and ethnicity of their loan applicants.
Race and gender discrimination has been scrutinized by, e.g., Munnell et al. (1996); Schafer
3

5. 2006). This is likely due to data-driven limitations. Indeed, authors tend to
adapt methodology to data, rather than the reverse.
Empirical tests for credit discrimination may be split into two approaches according to their underlying assumptions on gendered creditworthiness (Blanchard, Zhao and Yinger, 2008). The rst approach postulates that men and
women with identical personal and business characteristics are equally creditworthy. Hence, gender discrimination is assessed by testing whether gender
inuences the probability of loan denial (or the credit conditions). The second approach, used in this paper, avoids any prior assumption on gendered
creditworthiness.
It is more general but it necessitates data on individual
7
reimbursement records.
Discrimination is detected if a lower credit risk is
associated to a higher probability of denial (or worse credit conditions).
The contribution of this paper is twofold.
Firstly, it proposes a new esti-
mation method for assessing discrimination in loan allocation, which is well
adapted to microcredit. This method is based on the comparison of gender
coecient in loan size regression, on the one hand, and on the regression
of loss-over-loan-size ratio (hereinafter relative loss), on the other hand.
Secondly, this paper provides an original application. Exhaustive data help
robustifying the estimation with respect to the missing-variable problem
8
9
that often plagues studies on discrimination.
The empirical results point to gender discrimination. Indeed, all other things
equal, women face signicantly worse credit conditions, while being creditworthier. Additionally, it is shown that reducing the information asymmetry
through relationship brings no remedy to the handicap of being female.
The rest of the paper is organized as follows. Section 2 describes the database.
Section 3 discusses methodological issues. Section 4 provides evidence of discrimination and section 5 shows that it is not tempered by existing relationship. Section 6 concludes.
and Ladd (1982); Cavalluzzo and Cavalluzzo (1998); Ross and Yinger (1999, 2002); Blanchower, Levine and Zimmerman (2003); Han (2004); Cavalluzzo and Wolken (2005); Blanchard, Zhao and Yinger (2008). This empirical literature in surveyed in Agier and Szafarz
(2010).
7 Detailed characteristics of small-business borrowers are scarcely disclosed by the
lenders (GAO, 2008).
8 Admittedly, our results may still suer from the self-selection bias put forward by
Cavalluzzo (2002).
9 In that way, we follow Ross and Yinger (2002)'s recommendation; (...) well known
methodological problems, such as selection and endogeneity bias, could lead to disparateimpact discrimination even when the designers (...) are trying hard to avoid it. Scholarly
access to loan performance data and careful research are needed to shed further light on
these issues ( p. 298).
4

6. 2
Data
Our unique database comes from Vivacred, a Brazilian MFI. Vivacred provides credit to micro-entrepreneurs located in the Rio de Janeiro low income
communities and neighborhoods. It focuses on urban (formal and informal)
micro-businesses such as storekeepers, craftspersons, and service providers.
Vivacred started its activity in 1996 in Rocinha, the largest favela in Rio.
Five other branches were created since then: Rio das Pedras in 1998, Copacabana (now in Gloria) in 1999, Maré in 2000, Santa Cruz in 2002 and
in the city of Macaé (Rio state) in 2004. Until 2009, Vivacred was mostly
funded by the Brazilian Development Bank (BNDES). Then, Vivacred integrated the national CrediAmigo program nanced by Banco do Nordeste, a
Brazilian public bank.
Vivacred's loans are accessible to businesses with at least six months of activity.
For each application, the credit ocer in charge collects detailed
information
10
on the applicant and the guarantor, if any, and on the charac-
teristics of the business.
11
The credit ocer then provides a recommendation
to the credit committee that makes the nal decision (acceptance or denial,
and loan size). Actually, the term credit committee, used by Vivacred itself,
is misleading since it refers to a single person.
12
13
Vivacred charges the same interest rate to all its clients (3.9% per month).
Its lending methodology is based on credit rationing, rather than on adjusting
the interest rate to perceived credit risk.
Although this way of doing is
standard for MFIs, it raises ethical concerns (Hudon, 2009).
The data have been collected by the six branches of Vivacred. For the period under consideration (1997-2007), about 41,000 loans were solicited by
15,400 applicants, and about 32,000 loans were granted to 11,400 borrowers.
However, we removed the applications canceled by the clients, the contracts
with incomplete specications, the loans to Vivacred's employees, and the
few group loans. Therefore, the study is based on exhaustive data of 34,000
applications and 32,000 actual loans.
10 Private and professional addresses, birth date, birth state, marital status, gender,
dependent(s), profession, bank references, spouse's ID, current account, family consumption, family external income, full credit history (as a borrower, a borrower's spouse, or a
guarantor).
11 Location, sector, legal status, number of employees.
12 Depending on the requested amount, this person is either the branch manager, or a
senior credit ocer.
13 Banco da Mulher, a comparable non-prot institution, provides loans with rates be-
tween 3% and 5% a month, while
Fininvest, a for-prot institution, proposes consumption
credit with a monthly 12% rate.
5

7. Our dataset contains the full credit history (number of former loans, delays,
defaults, and losses) of all borrowers. A repayment is considered delayed
after 30 days, and defaulted after 180 days.
The penalty for default is
14
the client's name inclusion within the SPC register,
which is available for
consultation by any institution supplying credit, including shops.
Beyond
losing access to credit, those who are registered in SPC face serious trouble
getting a cell phone contract or buying household appliances, for example.
Table 1 gives the descriptive statistics, globally and then split by applicant's
gender, with t-tests for equality of means.
Vivacred claims no special commitment to serve women. Its clientele is balanced, with 49.6% of women over the period 1997-2007.
About the same
share (47.41%) is observed for female credit ocers, and these are more often in charge of dealing with female entrepreneurs.
Female applicants request smaller loans, to be paid back in less installments,
than men (BRL 1,237 against BRL 1,518),
amounts (BRL 891 against BRL 1,136).
15
and logically receive smaller
Additionally women are slightly
more credit-rationed than men as they get, on average, 21.4% less than they
request, against 20.7% for men. Nevertheless, men and women face similar
approval rates around 95%.
Women entrepreneurs are two years older than males (43 versus 41), less
likely to be married (43% versus 52%), and less likely to have dependents
(51% versus 53%).
Male and female applicants also dier in business characteristics.
Female-
owned businesses are smaller, in terms of both prots and sta size.
16
The
external income (i.e., income earned by any household member and unrelated
to business activity) is similar for men and women (around BRL 213 per
month).
Regarding the credit characteristics, the purpose of capital investment (as
opposed to liquidity) is present in 34% and 29% of applications from men and
women, respectively. Women need loans for both liquidity issues and loan
repayment more often than men.
Finally, the guarantor's and the client's
genders are unrelated.
Women exhibit a lower probability of delay than men (7.8% against 9.4%),
but a similar probability of default (2.9%). Most importantly, women lead
14 SPC is a national database recording bad payers.
15 The average requested amount for all applications, including the denied ones, is BRL
1,250 for women and BRL 1,524 for men.
16 Kevane and Wydick (2001) observe similar characteristics for micro-entreprises in
Guatemala.
6

8. Table 1: Global and Gender-specic Descriptive Statistics: All Applicants
All applicants M app. F app.
Mean S.D. Mean Mean
t-testb
Applicant's, ocer's and guarantor's genders
c
Female applicant
0.496
Female credit ocer
c
Female guarantor
0.500
0.474
0.499
0.459
0.490
0.430
c
0.501
0.429
0.430
−0.0311∗∗∗
−0.00106
Request, loan size, and repayment record
a
Requested Amount (BRL)
a
Loan size (BRL)
RA−LS
Rationing factor (
) (%)
RA
c
Delay (30 days)
c
Default (180 days)
a
Loss (BRL)
1,242
1,518
1,015
996
1,136
891
21.38
24.16
20.73
21.98
0.086
0.281
0.094
0.078
0.029
0.167
0.030
0.027
18.6
156.0
21.4
15.5
2.52
Relative loss (%)
1,380
1,237
15.27
2.75
2.29
280.7∗∗∗
245.3∗∗∗
−1.263∗∗∗
0.0165∗∗∗
0.003
5.888∗∗∗
0.465∗∗
Applicant's characteristics
Age (years)
c
42.2
12.0
41.2
43.2
Married
0.47
0.50
0.52
0.43
c
At least one dependent
0.52
0.50
0.53
0.51
a
Mth. ext. income (X100 BRL)
2.13
3.76
2.11
2.16
# former loans
2.25
3.27
2.35
2.15
# former loans with delay
0.04
0.21
0.04
0.04
# times as a guarantor
0.74
2.11
0.89
0.60
−1.925∗∗∗
0.0962∗∗∗
0.0169∗∗
−0.04
0.202∗∗∗
0.0077∗∗∗
0.282∗∗∗
Business characteristics
a
Business prot (X100 BRL)
9.19
13.44
10.26
8.09
Sector (trade = 1, other = 0)
c
Ocial business
0.53
0.50
0.49
0.56
0.06
0.23
0.07
0.05
# employees
0.63
2.20
0.72
0.54
2.177∗∗∗
−0.0760∗∗∗
0.0165∗∗∗
0.175∗∗∗
Credit characteristics
# installments
9.03
Capital investment purpose
c
Loan repayment purpose
c
Guarantor's involvement
Observations
a All
c
4.39
9.10
8.97
0.32
0.47
0.34
0.29
0.09
0.29
0.08
0.10
0.92
0.27
0.93
0.92
16,899
0.128∗∗
0.0518∗∗∗
−0.0171∗∗∗
0.00756∗∗
16,631
33,530
nancial values are in deated BRL (Real), the Brazilian currency. Over the period,
the Real uctuated between 0.270 and 0.588 USD.
b T-test for equality
c Dummy variables
of means between male and female applicants; *** p0.01, ** p0.05
7

9. to signicantly smaller losses for the MFI, in absolute and relative terms.
Vivacred's average relative loss is 2.8% for male borrowers and 2.3% for
female ones.
MFIs.
17
These numbers are consistent with those reported by other
In sum, irrespectively of their characteristics women receive smaller loans
and reimburse better than men. Section 4 will examine whether this evidence
resists multivariate analysis.
3
Methodology
Assessing discrimination in lending is complex for reasons pertaining to both
the underlying economic theory and intrinsic econometric issues.
As sum-
marized by Dymski (2006), the inconclusiveness of the academic literature
[can be attributed] to several factors: the ambiguity of legal and theoretical
denitions of discrimination; the inescapability of the point of view of the
observer and observed in empirical studies of racial discrimination; and the
way in which empirical methodologies require research questions to be framed
(p.215).
In this paper, we adopt a narrow denition. Namely, we dene gender discrimination in lending as the economically unjustied awarding of inferior
credit conditions to female borrowers.
This denition corresponds to the
intuition of a double-standard lending practice. It therefore excludes the socalled rational discrimination where unequal credit conditions result from
business needs.
18
Following our denition, disparate treatment (i.e., a harsher application
process for women) is a necessary but not sucient condition for gender discrimination. Indeed, disparate treatment could sometimes be economically rationalized by objective credit risk characteristics. For instance, let us
imagine just for the sake of the argument that women exhibit higher credit
17 For instance, reported default rates are: below 2.2% for CrediAmigo in Northeast
Brazil (CrediAmigo, 2009), below 5% for the Grameen Bank in Bangladesh (Morduch,
1999), and between 1 and 5.5% for rural MFIs in Indonesia, with a single exception of
12% (Robinson, 2002).
18 Rational discrimination can arise because of information costs (Lang and Nakamura,
1993). Also, when some variables aecting creditworthiness are not observable (e.g., business abilities, social connections, etc.), lenders could use gender as a proxy for credit risk.
Such a practice leads to statistical discrimination (Arrow, 1971, 1998). For instance, some
empirical papers in micronance use gender as a proxy for poverty. If gender were used in
the same way by lenders, this could lead to statistical discrimination.
8

10. risk than men, all other things equal. In such a situation, disparate treatment
could be a rational reaction from the lender. Under such circumstances, we
would not characterize the lender's attitude as gender discrimination.
Obviously, the denition of discrimination used here is based on economics,
and not on ethics. Actually, whether rationalizable or not, disparate treatment is highly questionable on ethical, and even legal, grounds. Pragmatically, our motivation for choosing such a narrow denition for gender discrimination in lending is linked to its empirical testability and the level of
conclusiveness it allows to reach. Indeed, detecting narrowly dened discrimination brings stronger conclusions on the lender's practice.
Besides, this narrow denition is close in spirit to Becker's denition of tastebased discrimination (Becker, 1971).
However, the qualication of taste-
based might look too restrictively connected to intentional prejudice.
In-
stead, we tend to view gender discrimination in lending as resulting from
mostly unintentional stereotyping shown by social psychologists (Fein and
Spencer, 1997; Kunda and Sinclair, 1999) to be a common human feature.
Indeed, Buttner and Rosen (1988) emphasize that women entrepreneurs still
suer from gender stereotypes related to their ability to eciently run a rm
(in terms of leadership, autonomy, lack of emotionalism, etc.).
On top of denitional complexity, empirical studies on discrimination in lending are often plagued by technical problems. First and foremost, because data
made available to researchers are generally insucient to trustfully reproduce
the lender's scoring process, the sources of gender gap in credit conditions, if
any, are hard to identify empirically. This paper will circumvent this serious
identication problem by using an exhaustive database.
Other challenging issues go beyond data availability (Ross, 2000).
discrimination in lending may take dierent forms.
Firstly,
Indeed, it may be ob-
served with regard to access to credit (higher denial probability), and/or
to credit conditions (higher interest rates, smaller loans, more collateral required, etc.).
19
Secondly, the lender's decision making is sequential: in the selection phase,
loans are approved or denied, and then credit conditions are set for approved
loans solely. As a consequence, loan allocation and credit conditions do not
concern the same pool of applicants.
Thirdly, the lender's assessment of creditworthiness is generally unknown.
Therefore, researchers commonly use a surrogate for creditworthiness built
19 Some authors, like Blanchower, Levine and Zimmerman (2003) and Weller (2009),
combine the two perspectives.
9

11. from a set of relevant variables (ideally, the ones used as screening devices
by the lender), referred to as controls, which aim at capturing all genderunrelated relevant variables.
This approach may suer from several draw-
backs, notably omitted variables. Inevitably, researchers are confronted to
some degree of uncertainty regarding the lender's screening process.
Fourthly, ex ante creditworthiness is, by nature, unobservable. It is typically
proxied by ex post variables like delay, default, and loss.
However, these
variables are to some extent endogenous because they are aected by the
credit conditions. For instance, default might be more frequent for larger
and therefore presumably riskier loans (Stiglitz and Weiss, 1981). Alternatively, more rationed borrowers could nd it harder to reimburse.
20
In any
case, endogeneity prevents ex post outcomes from being straightforward explanatory variables for the probability of approval and the credit conditions.
Given all these methodological limitations, how should we test for gender
discrimination in lending? We address this question by referring to the theoretical approach proposed by Ferguson and Peters (1995), who dene discrimination as the use of dierent credit standards across the two components of
the population and state that discrimination happens when a lower or equal
default rate is associated to a higher or equal denial rate, provided that at
least one inequality is strict.
The remaining of this section is devoted to
making this rule econometrically operational, and applicable to microcredit.
The lending methodology of the microcredit industry is based on standardized contracts, with typically the same interest rate for all borrowers. In that
framework, loan size is the sole credit condition that is tailored to the client's
needs by the MFI. Hence, the lender's problem may be represented as:
M ax {(1 + r) LS − E [Loss (LS)]}
LS 0
where
r
is the xed interest rate,
LS
(1)
is the loan size (denial corresponding
to a zero loan size) that is the lender's decision variable, and
is the expected loss that depends on loan size.
E [Loss (LS)]
Equivalently, this problem
writes:
E [Loss (LS)]
0
LS
M in
LS
(2)
On the empirical side, two variables are going to be explained: the loan size
(i.e., the decision variable), and the expected relative loss (i.e., the objective
20 The results in Appendix A reveal that, in Vivacred, loan size negatively impacts the
probabilities of delay and default.
10

12. function). Expectations being unobservable, we will take realized loss as a
proxy for expected loss.
21
In order to test for gender discrimination, we introduce the following notations. The loan applications are indexed by i.
22
Each application involves
several variables. First, the applicant's ex ante characteristics are:
•
Applicant's gender represented by a dummy variable:
Fi =
•
Vector
(z1i , ..., zni )
1
0
if the applicant is female
if the applicant is male
summarizing all other characteristics, including the
applicant's requested amount
RAi .
Second, the lender's decision variables are:
•
Loan approval represented by a dummy variable:
Ai =
•
Loan size:
1
0
if the loan is approved
if the loan is denied
LSi
We have explicitly split the decision variable in two parts (approval and
loan size) in order to make the impact of the selection process visible, and
subsequently apply the Heckman procedure.
Third, the ex post outcome variable is:
•
Relative loss:
Lossi /LSi
A companion paper (Agier and Szafarz, 2010) shows that, all things equal,
Vivacred's denial probability is not signicantly dierent between men and
women, but female borrowers receive signicantly smaller loans than men.
21 As loss is endogenous, the expectation error will simply be absorbed in the error term
of the regression without introducing any bias in the estimated coecients.
22 A person who introduces several loan applications will thus appear once for each
application.
11

13. Given these results, the current econometric model concentrates on loan size,
and not on denial probability:
n
βk zki +
1i
∀i
s.t.
Ai = 1
(3)
ϕk zki +
LSi = βF Fi +
2i
∀i
s.t.
Ai = 1
(4)
k=1
n
Lossi /LSi = ϕF Fi +
k=1
Equations 3 and 4 make it possible to operationalize the Ferguson and Peters
(1995) rule. Indeed, gender discrimination corresponds to the situation where
βF ≤ 0
and
ϕF ≤ 0,
with at least one strict inequality.
Moreover, the
selection issue will be addressed by using the Heckman estimation method
(Heckman, 1976, 1979).
Lastly, it is worth mentioning that Vivacred is a socially-oriented MFI, and
not a prot-oriented lender.
Does it make a dierence when it comes to
testing for discrimination? We argue that it does not, so that discrimination
in social lending may be addressed like in prot-based lending.
Our argument is the following. For the sake of self-sustainability, sociallyoriented lenders are bound to assess their applicants' creditworthiness.
In
practice, MFIs select their clients in two steps. Firstly, they dene their target pool of borrowers according to their social mission (typically, the poor
and/or unbanked entrepreneurs in a given area). Secondly, they assess creditworthiness of the applicants from this target pool basically in the same way
23
as prot-oriented institutions do.
Therefore, gender discrimination may
show o in the same way too, provided that the target pool is dened in-
24
dependently from gender considerations,
which is indeed the case for the
MFI under study.
23 This way of doing partly explains why MFIs do not reach the very poor (Rhyne, 2001).
Nevertheless, Hartarska (2005) nds evidence that in Eastern Europe and Central Asia,
MFIs with higher proportion of women on their board reach poorer borrowers. See also
Karlan and Zinman (2008) on the credit elasticities of the poor.
24 This restriction is important since some MFIs, like the Grameen Bank, serve women
solely or majoritarily. Our approach would not make sense for such MFIs.
12

14. 4
Estimation Results
4.1
Testing for Gender Discrimination
In this section, we compare the gender dummy coecients in loan size and
relative loss regressions, along the lines of the econometric methodology exposed in section 3.
In other words, we check whether the harsher credit
rationing imposed by Vivacred to female entrepreneurs is, at least partially,
attributable to repayment conduct.
We address this issue by estimating
equations (3), and (4).
In the rst regression (equation (3)) loan size is explained by the borrower's
gender, the amount requested by the borrower, and control variables. The
second regression (equation (4)) explains relative loss with the same variables. Additionally, two alternative specications for equation (4) open the
possibility of capturing the impact of credit rationing.
The borrower's gender is our explanatory variable of interest. In both equations, we control for all variables collected by Vivacred's credit ocers as well
as for this credit ocer's gender.
More precisely, the control variables in-
clude the borrower's characteristics (marital status, existence of dependents,
age, and household's extra income), the business characteristics (prots, sector, ocial status, number of employees), the loan characteristics (requested
amount, installments, loan renegotiation), and the guarantor's existence and
gender, if any. The relationship with Vivacred is accounted for by three variables: the number of former loans as a client and as a guarantor, and the
number of former loans repaid with delay (as a client, solely). Year dummies
are added to capture time heterogeneity.
Table 4.1 presents the regression results including one specication for loan
size and three specications for relative loss.
25
result for the basic formulation of equation (4).
Column (2) displays the
In columns (3) and (4),
26
the requested amount (RA) is replaced respectively by the loan size
RA−LS
).
and the rationing factor (
RA
(LS ),
25 Because, the Ferguson and Peters (1995) framework is purely theoretical (and only
considers, in its original form, denial and default), it does not state which variables are
relevant in the estimation.
26 Loan size inclusion is tricky because it is the dependent variable of the rst equa-
tion. Making it appear in the second equation could distort the impact of the borrower's
characteristics, among which gender.
However, ignoring this variable could create an
omitted-variable problem. Alternatively, we could use simultaneous-equation estimation
to account for endogeneity. However, combining such estimation with Heckman's procedure is tedious.
13

15. Table 2:
Loan Size and Relative Loss: Heckman's Regressions
(1)
(4)
Loss/LS
Loss/LS
-32.48***
-0.832***
-0.881***
-0.870***
(0.174)
(0.175)
(0.173)
0.623***
-0.000112
(0.00265)
Requested amount (RA)
(3)
Loss/LS
(5.116)
Female borrower (F)
(2)
LS
(9.03e-05)
Loan size (LS)
-0.000683***
(0.000116)
Rationing factor (
RA−LS
)
RA
0.0460***
(0.00379)
Female guarantor
-26.58***
-0.0981
-0.165
-0.0745
(5.280)
(0.180)
(0.180)
(0.178)
-50.45***
0.364**
0.312*
0.253
(5.146)
(0.175)
(0.176)
(0.174)
16.38***
-1.128***
-1.113***
-1.087***
(5.262)
(0.179)
(0.180)
(0.178)
13.29**
-0.146
-0.107
-0.148
(5.334)
(0.181)
(0.182)
(0.180)
0.613***
-0.0446***
-0.0447***
-0.0399***
(0.219)
(0.00748)
(0.00749)
(0.00744)
13.98
0.0280
-0.258
0.785**
(11.13)
(0.379)
(0.376)
(0.372)
# installments
24.65***
0.110***
0.144***
0.129***
(0.619)
(0.0211)
(0.0216)
(0.0205)
Capital investment
31.00***
-0.401**
-0.283
-0.502**
(5.811)
(0.198)
(0.198)
(0.195)
Loan repayment
81.71***
3.053***
3.142***
3.263***
(9.181)
(0.313)
(0.314)
(0.312)
External income
0.0943***
-0.000753***
-0.000528**
-0.000799***
(0.00701)
(0.000239)
(0.000240)
(0.000233)
Business prot
0.0576***
9.23e-05
0.000226***
6.98e-05
(0.00219)
(7.46e-05)
(7.52e-05)
(6.93e-05)
-24.10***
0.448**
0.445**
0.314*
(5.347)
(0.182)
(0.183)
(0.182)
Female credit ocer
Married client
Client with dependent(s)
Client's age
No Guarantor
Trade (sector)
Ocial business
0.270
0.653
0.209
(11.70)
# employees
179.9***
(0.399)
(0.401)
(0.390)
9.331***
0.00708
0.0307
0.00311
(1.213)
(0.0414)
(0.0415)
(0.0408)
37.68***
-0.292***
-0.243***
-0.238***
(1.003)
(0.0342)
(0.0350)
(0.0337)
# former loans with delay
-39.53***
0.363**
0.276*
0.376***
(4.236)
(0.144)
(0.145)
(0.143)
# times as a guarantor
10.69***
-0.141***
-0.130***
-0.117***
(1.246)
(0.0425)
(0.0426)
(0.0423)
Constant
-218.5***
2.032***
2.081***
0.583
(24.57)
(0.469)
(0.468)
(0.478)
-118.5***
3.840***
4.948***
1.671*
(30.91)
(1.053)
(1.030)
(1.004)
33,530
33,530
33,530
33,530
1,860
1,860
1,860
1,860
128,127
822.1
852.1
977.1
29
26
26
26
# former loans
Mills
Observations
Censored obs.
Wald Chi2
Degrees of freedom
Year dummies as controls (not reported). Heckman's selection: Approval by the committee.
Selection instruments: Kind of premises, source of funds, credit ocer's family status.
favela resident, and seniority, credit ocer turnover faced by the client, Rocinha branch.
Standard errors in parentheses; *** p0.01, ** p0.05, * p0.1
14

16. The requested amount acts as a proxy for the entrepreneur's project size. In
particular, it allows to take into account the fact that women typically ask
for smaller loans. By controlling for this rare piece of information, we intend
to clean the regression from the eect of gender-specic request.
including loan size,
28
27
When
we control for the level of indebtedness irrespectively
of the source of the gender gap.
The correlation between the requested amount and the loan size is high (equal
to 0.667). For this reason, we avoid putting both variables simultaneously
in the second regression. Instead, we opt for a third specication using the
rationing factor that measures the fraction of the requested amount that has
actually been granted to the applicant.
Column (1) of table 4.1 conrms that women suer from harsher credit rationing than men. Indeed, even when accounting for the selection bias and
the dierences in requested amounts, women receive signicantly smaller
loans than men.
As detailed in section 3, this result can be due to either
(economically unjustied) discrimination, or economically justied lending
practice.
The remaining columns of table 4.1 allow to disentangle these two possibilities unambiguously. Indeed, in all specications the gender dummy has a
signicant negative impact on relative loss, meaning that, all things equal,
women are creditworthier than men.The requested amount in itself has no
signicant impact (column (2)).
On the other hand, the loan size aects
relative loss negatively (but aects absolute loss positively, see Appendix A)
while the rationing factor has a positive impact.
More rationed loans are
harder to repay.
Remarkably, despite the handicap of being more credit-rationed, women manage to reimburse their loans better than men. In other words, if men and
women were equally rationed, the female repayment conduct would be even
better than it actually is.
Globally, the results are robust. The coecient of the gender dummy is about
the same in the three specications of the relative loss equation. Appendix
A proves that the same result applies to the absolute loss, the probability of
27 Still, we cannot exclude that women try to maximize their chances of getting a loan
by intentionally introducing smaller requests. If this is the case, then the request eect is
partly driven by the borrower's strategy. More generally, the identication of demand and
supply eects in credit markets is discussed by, e.g., Kanoh and Pumpaisanchai (2006);
de Janvry, McIntosh and Sadoulet (2010).
28 The inclusion of loan size in the second equation is tricky because loan size is the
dependent variable of the rst equation. However, as a matter of fact, the coecient of
the gender dummy is not much aected by such inclusion.
15

17. delay, and the probability of default. In addition, although the legitimacy of
using Heckman's estimation method is exhibited by a signicant Mills ratio,
OLS estimation brings similar features (results not reported here).
At this point, the rst conclusion of our empirical study emerges: Women
entrepreneurs are trustworthier borrowers than men, but do not benet from
this quality.
On the contrary, they face harsher credit conditions.
Conse-
quently, the Ferguson and Peters (1995) rule leads to asserting the presence
of discrimination.
Does the same conclusion apply to women involved as guarantors? In Vivacred, each contract involves at most three people from the borrowing side:
the client, the guarantor, and the client's spouse. Each of them is at risk in
case of default. Indeed, they all bear the risk of being registered in SPC (the
Brazilian insolvency register) and, consequently, experiencing serious trouble
29
in future nancial transactions.
From the lender's viewpoint, having more
people involved in a credit contract is always better.
30
This is likely the
reason why married borrowers repay better (and receive larger loans).
The coecient of the guarantor's gender reveals that female guarantors as
opposed to male guarantors have a negative impact on loan size, but no
signicant impact on relative loss.
According to the Ferguson and Peters
(1995) rule, this again should be viewed as a stigma of gender discrimination,
in a milder form though.
Incidentally, table 4.1 shows that the credit ocer's gender is signicant for
loan size, but not for relative loss. Female ocers typically oer smaller loans,
but obtain similar relative losses. Thus, viewed from the MFI's perspective,
male and female credit ocers are equally protable, although using dierent
screening processes.
Moreover, in Agier and Szafarz (2010), we show that
loan allocation by both male and female ocers leads to disparate treatment.
If, as conjectured, discrimination is attributable to gender stereotyping, then
the stereotypes are shared by male and female credit ocers.
The signs of the coecients associated to gender-neutral characteristics of the
borrowers match well with the intuition that lower loan size is associated to
higher relative loss, and vice versa. This means that the credit ocers grant
loans rationally in all respects except the applicant's gender. For instance,
married clients and older clients receive larger loans and repay better. The
same is true for borrowers with larger extra income. Applications from the
29 All guarantors provide their scal identity number (CPF), which is the code required
for registering them in SPC.
30 However, Alesina, Lotti and Mistrulli (2008) mention that the presence of a guarantor
might signal a borrower's higher credit risk.
16

18. trade sector (as opposed to services) bring smaller loans and generate higher
relative losses.
Mechanically, the loan size increases with the number of installments. For a
given loan size, the higher the number of installments, the worse the repayment conduct. As expected, all indicators of the borrower's credit history are
signicant: existing relationship (as a borrower and/or a guarantor) leads to
larger and better repaid loans. Former delays act in the opposite way.
To summarize, we have shown that gender discrimination is present in the
data. In line with the glass-ceiling theory (Agier and Szafarz, 2010), the
next section will examine in greater details the interaction between the applicant's gender and the scope of his/her project.
4.2
Interaction Between Gender and Project Scope
Up to now, we have assumed that the estimated model is fully linear. Nevertheless, loan size is likely a non-linear function of the requested amount.
Indeed, for tiny requests, it would be cost-inecient for credit ocers to
devote much attention to the specicities of the request le.
Instead, the
credit ocers may roughly examine some basic creditworthiness characteristics, and make a yes-or-no decision. They would either approve the loan as
such and oer the requested amount, or simply deny the loan.
Given that no gender gap would be observed on the loan allocation decision,
one can conjecture that gender discrimination is absent when the requested
amount is very small.
Moreover, if the observed gender discrimination is
associated with stereotyping (women are less able to run large projects),
then the gender gap in loan size should be increasing with the requested
amount. In order to investigate whether these conjectures hold in the data,
we now add a gendered interaction term in each estimated equation.
The empirical results in table 4.2 conrm the basic intuitions. The loan size
equation in column (1) features a positive coecient for the gender dummy
and a negative coecient for the interaction term.
In theory, this should
mean that women are favored for tiny loans (below BRL 100), but in practice
no actual loan lies below this limit. Thus, except for tiny loans (around BRL
100), there always exists a gender gap, and this gap is increasing with the
scope of the project.
This result is consistent with the glass-ceiling eect
unveiled by Agier and Szafarz (2010).
In contrast, column (2) shows that the relative loss is not related to the
requested amount, even when men and women are considered separately.
17

19. Table 3:
Gender Gap and Project Scope: Heckman's Regressions
(1)
RA * F
(4)
Loss/LS
Loss/LS
91.49***
-0.995***
-0.938***
-0.494**
(0.260)
(0.250)
(0.230)
0.656***
-0.000156
(0.00303)
Requested Amount (RA)
(3)
Loss/LS
(7.561)
Female borrower (F)
(2)
LS
(0.000104)
-0.0921***
0.000121
(0.00416)
(0.000143)
Loan size (LS)
-0.000701***
(0.000128)
LS * F
5.86e-05
(0.000183)
Rationing factor (
RA−LS
)
RA
0.0548***
(0.00520)
Rationing factor * F
-0.0174**
(0.00702)
Mills
-129.0***
Wald Chi2
DF
1.677*
(1.053)
(1.030)
(1.004)
33530
33530
33530
33530
1860
1860
1860
1860
130358
822.8
852.1
983.5
30
Censored obs.
4.954***
(30.69)
Observations
3.854***
27
27
27
Same controls as in table 4.1; Heckman selection: Approval by the committee.
Standard errors in parentheses; *** p0.01, ** p0.05, * p0.1
Despite being more heavily penalized, women with larger projects do not
incur higher losses.
The negative impact of loan size on relative loss does
not interact with gender (column (3)).
On the contrary, the worsening of
relative loss incurred by rationing is less pronounced for female borrowers.
This is perhaps attributable to higher female adaptability to bad circumstances.
Women cope with restricted loans better than men under similar
circumstances. In conclusion, gender discrimination is stronger for more ambitious projects. The next section examines whether relationship mitigates
discrimination.
5
Impact of Relationship
We now address the resilience of the gender gap in loan size by examining
the dynamics of the gender-specic treatment along the borrower's credit
history. Relationship between the lender and the borrower typically reduces
information asymmetry in lending(Tra and Lensink, 2007). Indeed, timely
repayments demonstrate the borrower's creditworthiness. As a consequence,
18

20. a borrower who has successfully reimbursed a rst loan will more easily obtain
a second and often larger loan, and so on. This is the basic principle
driving progressive lending (Egli, 2004).
Figure 1:
31
Evolutions of the gender-specic requested amounts and loan sizes
with respect to the length of relationship
Chakravarty and Scott (1999) show that the duration of relationship lowers
the probability of credit rationing in consumer loans. Our previous estimations (table 4.1) conrm that former loans have a positive impact on loan
size, and a negative impact on relative loss. However, as gure 1 exhibits,
after the second loan, the spread between the requested amount and the loan
size seems to stabilize.
32
Actually, successful second-time applicants request
smaller amounts than in their rst applications, but then benet from second loans higher than their rst loans. Later on, regular borrowers do not
31 Copestake (2002) emphasizes that progressive lending may also induce an increasing
inequality eect.
32 This constant spread may be seen as a steady-state equilibrium of the lending game
under credit rationing.
Indeed, the borrower knows that the lender is going to exert
credit rationing and rationally inates his/her request accordingly. Therefore, the optimal
response of the lender is to keep applying credit rationing, but in a constant - and thus
predictable - way to allow their regular borrowers accurately size their requests. If this
scenario holds, no player has any advantage of moving to an equilibrium without credit
rationing.
19

21. 33
downscale their requests with respect to the previous one anymore.
Table 4 provides additional descriptive statistics. The left side of the table
concerns the new applicants, whereas the right side concerns the known applicants. In each case, the overall and gender-specic means are displayed
for the following variables: requested amount (RA), loan approval rate, loan
RA−LS
), with the corresponding t-tests for
size (LS ), and rationing factor (
RA
equal means between genders. While requesting more on average (BRL 1,440
versus BRL 1,366), new applicants face more denial (9% versus 5%), receive
smaller loans (BRL 772 versus BRL 1,059), and are more rationed (38.9%
versus 18.2%). The global statistics are thus consistent with the asymmetric
information story.
Table 4: Descriptive Statistics for New and Known Applicants
Requested amount
a
All
(RA)
New applicants
M
F t-test
1,440
1,545
1,334
Loan approval (%)
a
Loan size (LS)
90.9
91.2
91.3
772
849
694
Rationing factor (%)
38.9
37.8
39.9
12,190
6,115
10.8∗∗∗
−0.18
11.9∗∗∗
−3.8∗∗∗
6,075
Observations
a
in BRL, *** p0.01
All
Known applicants
M
F t-test
1,366
1,519
1,209
95.0
95.1
95.3
1,059
1,190
925
18.2
18.1
18.4
21,367
10,815
17.5∗∗∗
−0.61
17.8∗∗∗
−0.81
10,552
Let us now examine whether discrimination tends to scale down with relationship. Table 4 shows that, regardless of their credit history, women are
left with identical opportunity to obtain a loan.
However, conrming the observations from gure 1, the gender gaps in both
requested amount and loan size widen with relationship. The female-overmale mean value ratios for new applicants are 86.3% for requested amount
and 81.7% for loan size, while the corresponding ratios for known applicants
are 79.6% and 77.7%, respectively. Contrastingly, only for rst loans is the
rationing factor signicantly higher for women. Perhaps with time, women
learn about the endured gender gap, and revise the scope of their projects
accordingly.
If gender discrimination were due to prejudice and/or stereotyping, relationship could reveal insucient to mitigate it. In such a case, repayment history
would not matter, and disparate treatment would subsist despite the revelation of women's creditworthiness.
33 Interestingly though, after two loans men start to progressively increase their requests
while, under similar circumstances, women seem to keep more or less the same requested
amount.
20

22. Alternatively, if gender discrimination were due to informational deciencies
(attributable to cultural reasons, for instance), credit ocers would learn
from experience and adapt their practice to the facts.
In such a case, the
intensity of discrimination would be decreasing with the number of previous
loans.
34
Relationship would then exhibit a stronger (positive) impact on
loan size for female borrowers than for male ones, allowing the former to be
treated in a progressively fairer way with time.
In order to disentangle these two possible scenarios, we add a gender-specic
relationship factor into the regressions by means of an interaction term between the number of former loans and the gender dummy. The coecient
associated to that new variable will indicate whether the impact of relationship diers across genders.
Table 5:
Gender Gap and Relationship: Heckman's Regressions
(1)
(2)
(3)
(4)
LS
Loss/LS
Loss/LS
Loss/LS
-21.38***
-0.861***
-0.896***
-0.935***
(6.182)
(0.211)
(0.211)
(0.210)
# Former loans
39.57***
-0.297***
-0.245***
-0.249***
(1.166)
(0.0397)
(0.0405)
(0.0391)
# Former loans * F
-4.946***
0.0128
0.00666
0.0289
(1.547)
(0.0527)
(0.0529)
(0.0524)
10.51***
-0.141***
-0.130***
-0.116***
(1.247)
(0.0425)
(0.0426)
(0.0423)
-38.58***
0.360**
0.275*
0.370***
(0.143)
Female borrower (F)
# Former loans
as a guarantor
# Former loans
with delay
(0.145)
(0.145)
3.850***
4.954***
1.689*
(30.92)
Requested Amount
(4.246)
-121.8***
Mills
(1.054)
(1.030)
(1.004)
X
X
Loan Size
X
Rationing factor
Observations
Censored obs.
Wald Chi2
DF
X
33,530
33,530
33,530
33,530
1,860
1,860
1,860
1,860
128,113
822.1
852.0
977.4
30
27
27
27
2nd loan
p0.1
Same controls as in table 4.1; Heckman selection: got at least a
Standard errors in parentheses; *** p0.01, ** p0.05, *
Our database includes 11,422 dierent borrowers, among which 63.31% beneted from a second loan. About one third of the newcomers dropped out
after their rst loan.
This second selection issue leads to a second use of
35
Heckman's estimation procedure.
34 For instance, Beaman et al. (2009) show on Indian data that female political leadership
weakens stereotypes about gender roles.
35 As Heckman's procedure allows one selection solely, we perform two separate exercises.
21

23. The results are presented in table 5. Expectedly, the number of former loans
has a positive impact on loan size and a negative impact on relative loss. More
troubling is the negative eect of the interaction term on loan size. While
men benet of an average extra BRL 39.57 for each former loan, women see
this bonus in loan size reduced by BRL 4.95, thus amounting BRL 34.62
only. Credit restrictions are progressively relaxed with relationship, but at a
slower pace for women. Relationship is thus less valued for females, digging
the gender gap instead of reducing it.
Loan size is also increasing with relationship as a guarantor, but the eect is
milder. A former loan as a borrower brings a bonus almost four time larger
than as a guarantor.
36
Moreover, former loans with delays have a negative
impact on loan size and, consistently, lead to higher relative losses.
6
Conclusion
The empirical approach to discrimination in the lending industry is less clearcut than in the labor market. Indeed, the literature exhibits large methodological variations, mainly data-driven, which plague the comparability of
results.
For this reason, we have adopted a restrictive denition of dis-
crimination in lending, embodying double standards solely.
Using such a
restrictive denition strengthens our conclusion that discrimination is indeed
present in the MFI under scrutiny.
In a nutshell, we have shown that, all things equal, women entrepreneurs
receive smaller loans and induce smaller losses for the lender.
This result
is consistent with the stylized facts reported by Armendáriz and Morduch
(2010).
Nonetheless, our ndings are more reliable than rough descriptive
statistics since the regressions take into account all variables actually reported
by the credit ocers, including the required amount.
Furthermore, the gender gap in loan size increases with relationship and
subsequent asymmetric information dwindling. Although being trustworthier
than men, women entrepreneurs thus seem to undergo a never-ending curse.
Starting with smaller rst loans than men, they never recover from their
First, we apply Heckman's regressions to the pool of all applicants. Second, we apply the
same regressions to the pool of applicants who beneted from one former loan at least. In
both cases, the selected clients are the ones who obtained a second loan. As both exercises
brought similar gures, we only present the results concerning the pool of all applicants.
36 Nonetheless, the impact of relationship as a guarantor is gender-insensitive (regression
not reported).
22

24. initial handicap. This nding strongly advocates for external intervention to
combat gender discrimination.
It is worth stressing that Vivacred, the MFI under scrutiny in this paper, is a
very well-run organization. The release of its remarkable database also indi-
37
cates that good governance practices are in place.
We therefore conjecture
that our ndings underestimate the global level of gender discrimination in
small-business lending.
As a matter of fact, the microcredit industry is highly subsidized internationally, notably by donors having a women empowerment agenda.
38
Given the
lack of anti-discrimination regulations in many developing countries, donors'
concern could appear as an alternative disciplining device. The main obstacle
thereto is data unavailability. The rst step should therefore be to request
more transparency in the screening processes put in place in MFIs, and more
generally in lending institutions.
The need for critical assessments of the
fairness of micronance practices is advocated by many authors, including,
e.g., Servet (2005); Rossel-Cambier (2008); Labie et al. (2010).
Data limitations are still present in our study. Although the regressions have
exploited all screening variables collected by the MFI itself, we cannot exclude
that face-to-face interviews bring unobservable but relevant gender-related
pieces of information, linked for instance to education, nancial literacy, and
attitude toward risk. Moreover, despite the exceptional exhaustivity of our
database, we do not possess information on the steps that predate the formal
loan application. For instance, an informal contact with a credit ocer might
discourage some entrepreneurs to pursue the application process. However,
it seems highly unlikely that such unobservable elements could challenge the
conclusion pointing to discrimination. More plausibly, they would reinforce
it.
The origins and consequences of prejudice and stereotyping, and the means to
deter them, are widely discussed in the socio-economic literature. We do not
elaborate further on these issues. Nevertheless, our ndings raise additional
unanswered questions.
Firstly, why are women entrepreneurs trustworthier than their male counterparts?
Do they fear penalties more, a hypothesis compatible with the
evidence that women are more risk-averse than men (Borghans et al., 2009)?
Or are they acting strategically in order to obtain better credit conditions
37 One of the authors has had the opportunity to observe Vivacred's day-to-day business
practices in details, and her ndings conrm that statement.
38 Isserles (2003) and Berkovitch and Kemp (2010) discuss the underlying ideology of
this agenda.
23

25. in the future? Anyhow, the facts documented in this paper contradict the
incentive-based argument stating that borrowers who accept less favorable
credit conditions (all other things equal) are more likely to default.
39
Secondly, why do women ask for smaller loans? Do they expect to be discriminated against
40
and refrain from applying for riskier projects thereby
creating a self-selection eect? The results in section 5 are consistent with
this scenario.
Moreover, if this explanation holds, it means that a consid-
erable amount of hidden entrepreneurial talent is wasted through rationally
expected discrimination.
Thirdly, how do women manage to reimburse better than men while being
more credit-rationed (which is detrimental to repayment conduct)? How do
household constraints interfere with female business projects? Recent studies
on intra-household relations in India (Garikipati, 2008; Guérin et al., 2009)
have shown that access to credit may increase female nancial vulnerability.
Seriously addressing these questions is necessary, at the very least for economic reasons.
Indeed, the potential for female-driven economic develop-
ment is far from being exhausted. Better knowing the needs and aspirations
of women entrepreneurs will help designing gender-conscious nancial products, as emphasized for the micronance industry by Johnson (2004); Corsi
et al. (2006); Guérin (2010) and many others.
By demonstrating that even well-run socially-oriented MFIs are not immune
to gender discrimination, this paper has stressed the importance of nding
creative solutions to the lack of capital endured by women entrepreneurs.
References
Agier, Isabelle, and Ariane Szafarz. 2010. Micronance and Gender:
Is
There a Glass Ceiling in Loan Size? Université Libre de Bruxelles CEB
Working Papers 10-047.
Alesina, Alberto F., Francesca Lotti, and Paolo Emilio Mistrulli.
2008. Do Women Pay More for Credit? Evidence from Italy. National
Bureau of Economic Research Working Paper Series 14202.
39 Actually, the original argument, as summarized by Banerjee and Duo (2010), is
phrased in terms of interest rates, and not loan size. However, the same logics should be
applicable to any credit condition, including loan size.
40 Blumberg and Letterie (2007) argue that applicants foresee pretty well the decision
procedure of the lender.
24

26. Alsos, Gry Agnete, Espen John Isaksen, and Elisabet Ljunggren. 2006. New Venture Financing and Subsequent Business Growth in
Men- and Women-Led Businesses. Entrepreneurship Theory and Practice,
30: 667686.
Armendáriz, Beatriz, and Ariane Szafarz.
2011. On Mission Drift
In Micronance Institutions. In The Handbook of Micronance, B. Armendariz and M. Labie ed. London-Singapore:World Scientic Publishing
(forthcoming).
Armendáriz, Beatriz, and Jonathan Morduch. 2000. Micronance Beyond Group Lending. Economics of Transition, 8(2): 401420.
Armendáriz, Beatriz, and Jonathan Morduch. 2010. The
Economics
of Micronance. 2 ed. Cambridge, MA:The MIT Press.
Arrow, Kenneth.
1971. The Theory of Discrimination. Princeton Uni-
versity, Department of Economics, Industrial Relations Section.
Arrow, Kenneth J. 1998. What Has Economics to Say about Racial Discrimination? Journal of Economic Perspectives, 12(2): 91100.
Banerjee, Abhijit V, and Esther Duo. 2010. Giving Credit Where it
is Due. Journal of Economic Perspectives, 24(3): 6180.
Beaman, Lori, Raghabendra Chattopadhyay, Esther Duo, Rohini
Pande, and Petia Topalova. 2009. Powerful Women: Does Exposure
Reduce Bias? Quarterly Journal of Economics, 124(4): 14971540.
Becker, Gary S. 1971. The Economics of Discrimination. 2nd ed. Chicago,
IL:University of Chicago Press.
Berkovitch, Nitza, and Adriana Kemp. 2010. Economic Empowerment
of Women' as a Global Project: The Limits of Social Change in the Neo
Liberal Era. In Confronting Global Gender Justice: Women's Lives, Hu-
man Rights, D. Bergoen, P. R. Gilbert, T. Harvey, and C. L. McNeely
ed. UK: Routledge (forthcoming).
Blanchard, Lloyd, Bo Zhao, and John Yinger. 2008. Do Lenders Discriminate Against Minority and Woman Entrepreneurs? Journal of Urban
Economics, 63(2): 467497.
Blanchower, David G., Phillip B. Levine, and David J. Zimmerman. 2003. Discrimination in the Small-Business Credit Market. Review
of Economics and Statistics, 85(4): 930943.
25

27. Blumberg, Boris F., and Wilko A. Letterie. 2007. Business Starters
and Credit Rationing. Small Business Economics, 30(2): 187200.
Borghans, Lex, Bart Golsteyn, James J. Heckman, and Huub Meijers. 2009. Gender Dierences in Risk Aversion and Ambiguity Aversion.
IZA Discussion Paper 3985.
Buttner, E. Holly, and Benson Rosen.
1988. Bank Loan Ocers'
Perceptions of the Characteristics of Men, Women, and Successful Entrepreneurs. Journal of Business Venturing, 3(3): 249258.
Buvinic, Mayra, and Marguerite Berger. 1990. Sex Dierences in Access to a Small Enterprise Development Fund in Peru. World Develop-
ment, 18(5): 695705.
Cavalluzzo, Ken.
Discrimination:
2002. Competition, Small Business Financing, and
Evidence from a New Survey.
Journal of Business,
75(4): 641680.
Cavalluzzo, Ken, and John Wolken.
2005. Small Business Loan
Turndowns, Personal Wealth, and Discrimination. Journal of Business,
78(6): 21532178.
Cavalluzzo, Ken, and Linda Cavalluzzo. 1998. Market Structure and
Discrimination: The Case of Small Businesses. Journal of Money, Credit
and Banking, 30(4): 77192.
Chakravarty, Sugato, and James S. Scott.
1999. Relationships and
Rationing in Consumer Loans. Journal of Business, 72(4): 523544.
Copestake, James.
2002. Inequality and the Polarizing Impact of Mi-
crocredit: Evidence from Zambia's Copperbelt. Journal of International
Development, 14(6): 743755.
Corsi, Marcella, Fabrizio Botti, Tommaso Rondinella, and Giulia
Zacchia. 2006. Women and Micronance in Mediterranean Countries.
Development, 49(2): 6774.
CrediAmigo. 2009. CrediAmigo Annual report - 2009.
www.bnb.gov.br.
de Janvry, Alain, Craig McIntosh, and Elisabeth Sadoulet.
2010.
The Supply- and Demand-Side Impacts of Credit Market Information.
Journal of Development Economics, 93(2): 173188.
26

28. Diagne, Aliou, Manfred Zeller, and Manohar Sharma.
2000. Em-
pirical Measurements of Households' Access to Credit and Credit Constraints in Developing Countries:
Methodological Issues and Evidence.
IFPRI Food Consumption and Nutrition Division Discussion Paper 90,
Washington, D.C.
Dymski, Gary A. 2006. Discrimination in the Credit and Housing Markets: Findings and Challenges. In Handbook on the Economics of Dis-
crimination, W.M. Rodgers III ed., 215259. Cheltenham:Edward Elgar
Publishing.
Egli, Dominik. 2004. Progressive Lending as an Enforcement Mechanism
in Micronance Programs. Review of Development Economics, 8(4): 505
520.
Fein, S, and S.J. Spencer. 1997. Prejudice as Self-Image Maintenance:
Arming the Self Through Derogating Others. Journal of Personality
and Social Psychology, 73: 44, 31.
Ferguson, Michael F, and Stephen R Peters. 1995. What Constitutes
Evidence of Discrimination in Lending? Journal of Finance, 50(2): 739
748.
Fletschner, Diana. 2009. Rural Women's Access to Credit:
Market Imper-
fections and Intrahousehold Dynamics. World Development, 37(3): 618
631.
GAO, (U.S. Government Accountability Oce). 2008. Fair Lending:
Race and Gender Data Are Limited for Nonmortgage Lending. Report to
Congressional Requesters 08-1023 T, Washington, D.C.
Garikipati, Supriya. 2008. The Impact of Lending to Women on Household Vulnerability and Women's Empowerment:
Evidence from India.
World Development, 36(12): 26202642.
Gatewood, Elizabeth J., Candida G. Brush, Nancy M. Carter,
Patricia G. Greene, and Myra Hart. 2004. Women Entrepreneurs,
Growth, and Implications for the Classroom. Coleman Foundation White
Paper Series for the United States Association for Small Business and Entrepreneurship.
Greene, Patricia G., Myra M. Hart, Elizabeth J. Gatewood, Candida G. Brush, and Nancy M. Carter. 2003. Women Entrepreneurs:
27

29. Moving Front and Center:
An Overview of Research and Theory. US-
ASBR White Papers.
Guérin, Isabelle. 2010. The Gender of Finance and Lessons from Micronance. In The Handbook of Micronance, B. Armendariz and M. Labie
ed. London-Singapore:World Scientic Publishing.
Guérin, Isabelle, Marc Roesch, Venkatasubramanian, Mariam Sangare, and Santosh Kumar. 2009. Micronance and the Dynamics of
Financial Vulnerability. Lessons from Rural South India. RuMe Working
Paper.
Haines, G.H., B.J. Orser, and A.L. Riding. 1999. Myths and Realities:
An Empirical Study of Banks and the Gender of Small Business Clients.
Canadian Journal of Administrative Sciences / Revue Canadienne des Sciences de l'Administration, 16(4): 291307.
Han, Song. 2004. Discrimination in Lending:
Theory and Evidence. Jour-
nal of Real Estate Finance and Economics, 29(1): 546.
Hartarska, Valentina. 2005. Governance and Performance of Micronance
Institutions in Central And Eastern Europe and the Newly Independent
States. World Development, 33(10): 16271643.
Heckman, James J. 1976. The Common Structure of Statistical Models
of Truncation, Sample Selection and Limited Dependent Variables and a
Simple Estimator for Such Models. Annals of Economic and Social Mea-
surement, 5(4): 120137.
Heckman, James J. 1979. Sample Selection Bias as a Specication Error.
Econometrica, 47(1): 15361.
Hudon, Marek.
2009. Should Access to Credit be a Right?
Journal of
Business Ethics, 84(1): 1728.
ILO. 2009. Small Change, Big Changes:
Women and Micronance. Inter-
national Labour Organization, Geneva.
Isserles, Robin G.
2003. Microcredit:
The Rhetoric of Empowerment,
the Reality of Development As Usual. Women's Studies Quarterly,
31(3/4): 3857.
Jalbert, Susanne E. 2000. Women Entrepreneurs in the Global Economy.
Center for International Private Enterprise, Washington, D.C.
28

30. Jamali, Dima.
2009. Constraints and Opportunities Facing Women En-
trepreneurs in Developing Countries: A Relational Perspective. Gender
in Management: An International Journal, 24(4): 232251.
Johnson, Susan.
2004. Gender Norms in Financial Markets:
Evidence
from Kenya. World Development, 32(8): 13551374.
Kanoh, Satoru, and Chakkrit Pumpaisanchai. 2006. Listening to the
Market: Estimating Credit Demand and Supply from Survey Data. Institute of Economic Research, Hitotsubashi University Hi-Stat Discussion
Paper Series d05-137, Tokyo.
Karlan, Dean S., and Jonathan Zinman.
2008. Credit Elasticities in
Less-Developed Economies: Implications for Micronance. American Eco-
nomic Review, 98(3): 104068.
Kevane, Michael, and Bruce Wydick.
2001. Microenterprise Lending
to Female Entrepreneurs: Sacricing Economic Growth for Poverty Alleviation? World Development, 29(7): 12251236.
Kunda, Ziva, and Lisa Sinclair. 1999. Motivated Reasoning with Stereotypes:
Activation, Application, and Inhibition. Psychological Inquiry,
10(1): 1222.
Labie, Marc, Pierre-Guillaume Méon, Roy Mersland, and Ariane
Szafarz. 2010. Discrimination by Microcredit Ocers: Theory and Evidence on Disability in Uganda. Université Libre de Bruxelles CEB Working Papers 10-007.
Lang, William W., and Leonard I. Nakamura.
1993. A Model of
Redlining. Journal of Urban Economics, 33(2): 223234.
Marrez, Helena, and Mathias Schmit.
microcredit:
How does gender matter?
2009. Credit risk analysis in
Université Libre de Bruxelles
CEB Working Papers 09-053.
Morduch, Jonathan. 1999. The Micronance Promise.
Journal of Eco-
nomic Literature, 37(4): 15691614.
Morrison, Andrew, Dhushyanth Raju, and Nistha Sinha. 2007. Gender Equality, Poverty and Economic Growth. The World Bank 4349.
Munnell, Alicia H., Georey M. B. Tootell, Lynn E. Browne, and
James McEneaney. 1996. Mortgage Lending in Boston: Interpreting
HMDA Data. American Economic Review, 86(1): 2553.
29

31. Rhyne, E.
2001. The Yin and Yang of Micronance: Reaching the Poor
and Sustainability. Microbanking Bulletin, 2.
Riding, Allan L., and Catherine S. Swift. 1990. Women Business Owners and Terms of Credit: Some Empirical Findings of the Canadian Experience. Journal of Business Venturing, 5(5): 327340.
Robinson, Marguerite S.
2002.
Micronance Revolution Volume 2:
Lessons from Indonesia. World Bank Publications.
Rossel-Cambier, Koen.
2008. Combined Micro-Finance:
Selected Re-
search Questions from a Stakeholder Point of View. Université Libre de
Bruxelles CEB Working Papers 08-004.
Ross, Stephen L. 2000. Mortgage Lending, Sample Selection and Default.
Real Estate Economics, 28(4): 581621.
Ross, Stephen L, and John Yinger. 1999. Other Evidence of Discrimination: Recent Studies of Redlining and of Discrimination in Loan Approval and Loan Terms. In Mortgage Lending Discrimination: A Review
of Existing Discrimination, Margery Austin Turner, Felicity Skidmore ed.,
1742. Washington, D.C.:The Urban Institute.
Ross, Stephen L, and John Yinger. 2002. The
Color of Credit: Mort-
gage Discrimination, Research Methodology, and Fair-Lending Enforcement. Cambridge, MA:MIT Press.
Saith, Ruhi, and Barbara Harriss-White. 1999. The Gender Sensitivity
of Well-being Indicators. Development and Change, 30(3): 465497.
Schafer, Robert, and Helen F Ladd. 1982. Discrimination in Mortgage
Lending. Cambridge, MA:MIT Press.
Servet, Jean Michel.
2005. Le besoin d'objectifs principaux nouveaux
pour la micronance : lutter contre les inégalités et faire face aux risques.
Techniques Financières Développement, 78: 1220.
Stiglitz, Joseph E, and Andrew Weiss. 1981. Credit Rationing in Markets with Imperfect Information. American Economic Review, 71(3): 393
410.
Tra, Pham Thi Thu, and Robert Lensink.
2007. Lending policies
of informal, formal and semiformal lenders. Economics of Transition,
15(2): 181209.
30

32. Verheul, Ingrid, and Roy Thurik. 2001. Start-Up Capital:
Does Gen-
der Matter?. Small Business Economics, 16(4): 32945.
Weller, Christian E. 2009. Credit Access, the Costs of Credit and Credit
Market Discrimination. Review of Black Political Economy, 36(1): 728.
31

33. Appendix A
Robustness check: Repayment Behavior
Table 6:
Absolute Loss, and Probabilities of Delay and Default
(1)
(4)
(5)
(6)
(7)
(8)
(9)
Loss
Loss
Delay
Delay
Delay
Default
Default
Default
-6.340***
-6.351***
-7.011***
-0.0107***
-0.0118***
-0.0123***
-0.00332***
-0.00362***
-0.00359***
(1.749)
Requested Amount (RA)
(3)
Loss
Female borrower (F)
(2)
(1.753)
(1.752)
(0.00252)
(0.00251)
(0.00263)
(0.000900)
(0.000905)
(0.000953)
0.00559***
2.94e-06**
6.97e-07
(1.22e-06)
(0.000906)
(4.92e-07)
0.00542***
-8.60e-06***
(0.00116)
Loan Size (LS)
(1.73e-06)
Rationing Factor
-3.32e-06***
(8.51e-07)
0.000861***
(0.0382)
Married client
0.191***
(6.06e-05)
0.000206***
(2.55e-05)
-8.568***
-0.0237***
-0.0249***
-0.00556***
-0.00544***
(1.803)
(1.805)
(0.00261)
(0.00259)
(0.00272)
(0.000977)
(0.000973)
(0.00104)
-0.999
-1.187
-1.826
0.00310
0.00196
0.00271
-0.000194
-0.000446
-0.000294
(1.808)
(1.803)
(0.00259)
(0.00257)
(0.00270)
(0.000891)
(0.000888)
(0.000931)
-0.154
0.00632
-0.934
0.0128***
0.0123***
0.0103***
0.00205**
0.00185**
0.00157*
(1.758)
Client with dependent(s)
-0.0244***
(1.805)
Female Credit Ocer
-8.318***
(1.800)
Female Guarantor
-8.628***
-0.00579***
(1.763)
(1.764)
(0.00257)
(0.00253)
(0.00269)
(0.000888)
(0.000884)
(0.000932)
-2.938
-2.814
-2.396
-0.00421
-0.00378
-0.00429
-0.000179
-0.000149
-0.000216
(1.821)
(1.823)
(1.823)
(0.00262)
(0.00261)
(0.00274)
(0.000903)
(0.000903)
(0.000948)
-0.229***
-0.237***
-0.221***
-0.000885***
-0.000894***
-0.000866***
-0.000191***
-0.000194***
-0.000187***
(0.0750)
(0.0751)
(0.0753)
(0.000110)
(0.000109)
(0.000116)
(4.00e-05)
(3.98e-05)
(4.25e-05)
0.458
-1.699
-2.104
-0.000676
-0.00655
0.00892
0.00181
0.000314
0.00552*
(3.807)
(3.773)
(3.755)
(0.00579)
(0.00526)
(0.00671)
(0.00218)
(0.00192)
(0.00293)
1.524***
1.513***
1.936***
0.00181***
0.00237***
0.00257***
0.000526***
0.000673***
0.000725***
(0.211)
(0.217)
(0.207)
(0.000286)
(0.000289)
(0.000297)
(0.000102)
(0.000107)
(0.000114)
Capital investment
-5.118**
-4.609**
-3.682*
-0.00722***
-0.00543**
-0.00707**
-0.00231**
-0.00178*
-0.00248**
(1.988)
(1.988)
(1.974)
(0.00270)
(0.00269)
(0.00281)
(0.000956)
(0.000964)
(0.00101)
Loan repayment
27.29***
27.07***
28.78***
0.102***
0.0991***
0.120***
0.0241***
0.0240***
0.0293***
(3.140)
(3.148)
(3.152)
(0.00831)
(0.00817)
(0.00905)
(0.00346)
(0.00344)
(0.00400)
External income
-0.00366
-0.00305
-0.000757
-1.20e-05***
-8.30e-06**
-1.09e-05***
-6.74e-06***
-5.44e-06***
-6.31e-06***
(1.77e-06)
Client's age
No Guarantor
# Installments
(0.00240)
(0.00236)
(3.35e-06)
(3.34e-06)
(3.46e-06)
(1.69e-06)
(1.69e-06)
0.00285***
0.00422***
6.47e-07
2.86e-06***
1.53e-06
2.38e-07
8.08e-07**
3.98e-07
(0.000749)
Trade (sector)
(0.00240)
0.00250***
Business prot
(0.000755)
(0.000700)
(8.83e-07)
(9.60e-07)
(9.34e-07)
(3.15e-07)
(3.78e-07)
(3.60e-07)
0.556
Ocial business
0.774
0.250
-6.67e-05
-0.000752
-0.00222
0.00193**
0.00176*
0.00150
(1.829)
(1.831)
(1.836)
(0.00259)
(0.00258)
(0.00272)
(0.000903)
(0.000904)
(0.000945)
# Employees
2.788
3.644
7.551*
0.00855
0.0195***
0.0131**
-0.000190
0.00275
0.000371
(4.002)
(4.019)
(3.940)
(0.00575)
(0.00644)
(0.00611)
(0.00222)
(0.00276)
(0.00239)
0.390
# Times as a guarantor
0.709*
0.000556
0.000979**
0.000844*
-1.96e-05
0.000182
5.50e-05
(0.416)
(0.413)
(0.000493)
(0.000462)
(0.000495)
(0.000295)
(0.000235)
(0.000302)
-1.742***
-1.780***
-1.064***
-0.0144***
-0.0131***
-0.0120***
-0.00360***
-0.00325***
-0.00308***
(0.343)
# Former loans
0.462
(0.415)
(0.351)
(0.341)
(0.000810)
(0.000786)
(0.000845)
(0.000343)
(0.000329)
(0.000367)
-1.243***
-1.272***
-1.067**
-0.00588***
-0.00543***
-0.00508***
-0.00146***
-0.00134***
-0.00128***
(0.426)
(0.427)
(0.427)
(0.00104)
(0.00103)
(0.00107)
(0.000359)
(0.000356)
(0.000369)
3.214**
3.014**
2.128
0.0228***
0.0205***
0.0217***
0.00494***
0.00441***
0.00471***
(1.448)
(1.452)
(1.441)
(0.00233)
(0.00229)
(0.00241)
(0.000862)
(0.000847)
(0.000923)
33,530
33,530
33,530
33,530
33,530
33,530
33,530
33,530
33,530
Censored obs.
1,860
1,860
1,860
1,860
1,860
1,860
1,860
1,860
1,860
Wald Chi2
582.8
564.9
567.3
1,560
1,618
1,744
659.3
679.7
727.9
26
26
26
26
26
26
26
26
26
# Former loans with delay
Observations
DF
For all regressions, Heckman selection: Approval by the credit committee. Standard errors in parentheses; *** p0.01, ** p0.05, * p0.1
The coecients of the year dummies and constants are not reported. The probabilities of delay and default are estimated by Heckman-probit (marginal eect reported).