Are Women Better Loan Officers?

This version: February 2009

Thorsten Beck* CentER, Dept. of Economics, Tilburg University and CEPR

Patrick Behr†

Goethe University Frankfurt

André Güttler‡ European Business School

Abstract

What if any is the impact of the gender of a loan officer on loan default risk? Using a unique data

set for a microbank in Albania over the period 1996 to 2006, we find that loans handled by

female loan officers show significantly lower default rates than loans handled by male loan

officers, controlling for a variety of borrower, loan, and loan officer characteristics. This effect

comes in addition to a lower default rate of female borrowers and cannot be explained by

experience differences between female and male loan officer. Our result seems to be driven by

differences in monitoring intensity, as we do not see significant differences in the acceptance

rates of loan officers of different genders.

JEL Classification: G21; J16

Keywords: Loan officers; gender; loan default; monitoring; Albania; microcredit

* Department of Economics and European Banking Center, Tilburg University, P.O. Box 90153, 5000 LE Tilburg, The Netherlands, and CEPR, Email: t.beck@uvt.nl. † Department of Finance, House of Finance, Goethe University Frankfurt, Grüneburgplatz 1, 60323 Frankfurt, Germany, Email: behr@finance.uni-frankfurt.de (corresponding author). ‡ HCI Endowed Chair of Financial Services, Department of Finance, Accounting and Real Estate, European Business School, Rheingaustr. 1, 65375 Oestrich-Winkel, Germany, E-mail: andre.guettler@ebs.edu. We thank Andreas Madestam and Harry Schmidt for very helpful suggestions and comments, and Annekathrin Entzian for assistance with the data preparation.

1

I. Introduction

What is the impact of loan officers’ gender and experience on loan default risk? While the

role of gender has been explored in a variety of fields in finance, such as investment decisions,

mutual fund management or equity analyst performance, and the behavior and importance of loan

officers in financial institutions has been studied in several recent papers, the impact of loan

officers’ gender on loan default risk has not been analyzed, yet. This paper uses a unique loan-

level data set for an Albanian microbank over the period 1996 to 2006 to assess the relationship

between borrowers’ and loan officers’ gender and the probability of loan default, controlling for a

vast array of borrower, loan and loan officer characteristics. Specifically, controlling for the

borrowers’ gender, we test whether male or female loan officers experience a lower default

probability on their loans and whether this relationship varies with the experience of loan

officers.

Understanding the relationship between loan officers’ gender and loan default risk is

interesting and important for practitioners and researchers alike. Designing incentives for loan

officers to minimize loan losses might have to take into account loan officers’ gender if empirical

findings point to differences between male and female loan officers in their screening and

monitoring quality and ability. Exploring the relationship between loan officers’ gender and

experience and loan default risk also adds to the literature on borrower-loan officer relationships.

Theory provides ambiguous predictions of why the gender of the loan officer might

matter for the default probability of “their” borrowers. Consider first the effort exerted by loan

officers in screening and monitoring borrowers. Modeling the relationship between loan officer

and bank as principal-agent relationship can help understand the incentives of loan officers to

exert effort (Agarwal and Wang, 2008). Female loan officers have typically fewer outside options

2

in the labor market and have therefore stronger incentives to excel in form of low default rates in

their loan portfolio.1 Women are typically less mobile, especially if married, and thus more

dependent on the existing job, again increasing their incentives to excel. Especially in developing

countries, women are more conservative and more afraid of social sanctions, which increases

pressure on female loan officers to perform better than their male colleagues. These arguments

are similar to arguments of why female borrowers in developing countries are typically better

clients than their male peers (Armendariz de Aghion and Morduch, 2005). On the other hand,

consider the relationship between loan officer and borrower. In patriarchic societies, male loan

officers might have a stronger standing vis-à-vis borrowers, be they male or female, in terms of

monitoring and disciplining them, thus ensuring loan repayment. In this case, we would observe

lower default probability of loans approved and monitored by male loan officers. Finally, loan

officers might have an easier time monitoring and disciplining borrowers of their own gender,

hence, we would expect to find a lower default probability of female borrowers if the loan is

approved and monitored by a female rather than by a male loan officer, with the reverse holding

for male loan officers.

We test several alternative, though not necessarily competing, hypotheses on the

relationship between loan officer’s gender and experience and loan default probability. On the

one hand, experience might be negatively related to loan default risk, if loan officers gain

expertise on screening and monitoring borrowers over time (Anderson, 2004). On the other hand,

career concerns might induce younger and less experienced loan officers to undertake a greater

effort to avoid loan losses in order to maximize their career progress and thus future income

perspectives (Agarwal and Wang, 2008).

1 Darity and Mason (1998) provide a comprehensive overview of gender discrimination in the labor market.

3

We explore these hypotheses by analyzing a unique data set on more than 43,000 loans

over the years 1996 to 2006 provided by a microcredit institution in Albania. For each loan, we

can identify the loan officer who screened the borrower and subsequently monitored her over the

lifetime of the loan. A possible default of the loan, i.e. arrears beyond a certain number of days,

can thus be directly linked to a specific loan officer. The data set includes extensive information

about borrower characteristics such as the gender or the marital status of the borrower, loan

characteristics such as size, maturity and interest rate of the loan, and loan officer characteristics

such as gender and experience within the institution. As Albania is a transition economy and

given that the lender is a typical microcredit institution, we include several variables to capture

the different lending technology and different borrower population of such a lender. Specifically,

we control for borrower characteristics like, for instance, the number of persons in the household

of the borrower or whether a phone is available in the household of the borrower, information

that is normally not used/available when using data provided by banks in developed countries.

Critically, we have information on both successful and rejected loan applicants, which allows us

to test whether differences in default risk across loan officers of different gender are driven by

selection bias to the extent that female or male loan officers select better performing borrowers

ex-ante.

Our results indicate that loans handled by female loan officers have a significantly lower

default probability than those of their male counterparts. This result is robust to controlling for

borrower’s gender and for the correlation between borrower’s and loan officer’s gender. We also

find only very little variation of women’s superior performance vis-à-vis men with their

experience as loan officer, suggesting that our results are not driven by women having harder

access to loan officer positions. This result holds over different samples. Specifically, we confirm

our finding both for first loans as well as for repeat loans of the same borrowers, with a stronger

4

effect for first loans. We interpret this as supporting for our hypothesis that female loan officers

in this microbank face stronger incentives or have better skills in dealing with borrowers, as the

agency problems between bank and borrower should be stronger for first time borrowers. We also

test for differences in the approval rates of loan applicants between female and male loan

officers. Controlling for a vast array of borrower characteristics, we cannot find any difference

between female and male loan officers in their acceptance of applicants, suggesting that the

performance advantage of female loan officers is in their monitoring of borrowers rather than

their screening. This also confirms that our findings are not driven by selection bias of female

loan officers dealing with borrowers that have ex-ante a different risk profile.

By investigating gender differences in the context of loan officers, this paper is related to

a growing body of studies on the role of loan officers in financial institutions. For instance,

Andersson (2004) finds that senior loan officers come to more consistent decisions than

inexperienced loan officers. Berger and Udell (2004) argue that the loan officers’ experience with

severe business environments decays in boom periods and, as a result, also substandard

borrowers get loans. Hertzberg et al. (2008) show that loan officers are more likely to reveal

negative information in the case of job rotation because it seems to be better if the loan officer

reveals this kind of negative information herself instead of having bad information being revealed

by a successor loan officer. Liberti and Mian (Forthcoming) find that the higher the decision

maker is in the bank’s hierarchy, the lower the importance of soft information gets because the

unverifiable soft information looses reliability over hierarchy levels. Finally, in a recent paper,

Agarwal and Wang (2008) argue that loan officer’s choice of effort depends on the incentive

scheme implemented by the bank, the information asymmetry between the loan officer and the

bank, and the loan officer’s career concerns. Our results add a new facet to this literature. They

5

suggest that not only the institutional design of financial institutions matter (Berger et al., 2005;

Mian, 2006), but also the gender of the people operating in it.

Our study is also related to the literature studying gender disparities in risk taking and

performance. Several papers have shown that female decision makers are more risk averse than

male decision makers (Barsky et al., 1997; Agnew et al., 2003) and that this higher risk aversion

affects financial decision (Charness and Gneezy, 2007; Christiansen et al., 2006; Barber and

Odean, 2001). Other authors have explored the behavior of women in different competitive

environments and their treatment within financial institutions (Gneezy et al., 2003, Forthcoming;

Niederle and Vesterlund, Forthcoming; Black and Strahan, 2001, Goldin and Rouse, 2000).

Green et al. (2008) analyze the performance of male versus female Wall Street equity analysts

and document that the male analysts seem to have better forecasting abilities, i.e. women seem to

perform worse at hard, quantifiable tasks. On the other hand, they also report that female analysts

seem to perform better at non-quantifiable aspects of the job such as client service.

Our work contributes to this literature by documenting that women may perform better

than men at quantifiable job aspects such as the management of default risk. The results further

suggest that this is not driven by a higher degree of risk aversion as we do not find significant

differences between female and male loan officers in the loan approval decision.

The remainder of the paper is organized as follows. Section II discusses the data, and

section III the methodology. Section IV presents our main results and section V contains

robustness checks and further analyses. Section VI concludes.

II. Data

We use a unique data set of both rejected and accepted loan applicants from a microcredit

lender in Albania. Specifically, we have information on over 43,000 loan applications and 31,000

6

loans given by the lender over the period January 1996 to December 2006, i.e. the first eleven

years of operation of this microcredit institution. While the microcredit lender is part of an

international network, it works with local management and loan officers. Specifically, our data

set contains information on 203 loan officers and covers five branches of the lender in the

Albanian capital, Tirana. Unlike other microcredit institutions, the lender grants only individual

(not joint liability or group) loans, for business, real estate, and consumption purposes. While the

lender clearly focuses on the low-income and small-enterprise segment, and has thus a double-

bottom line approach of both profitability and increasing access to credit, financial sustainability

and therefore profitability is the primary goal.

Table 1 provides some basic data about the lender. Specifically, it shows for the 11 years

of our sample period the number of loan applications, number of approved loans, loan

characteristics, basic borrower characteristics, loan usage and the share of female loan officers.

The lender grew substantially over the past 11 years, from originally 350 loans and 300

borrowers in 1996 to over 7,000 loans and over 6,000 borrowers in 2006. Over this period, the

approval ratio, defined as the number of approved applications divided by the number of all

applications, increased substantially. It rose from 44 percent in 1996 over 60 percent in 2000 to

71 percent in 2006, which can be partly explained by the increasing share of repeat borrowers.

The average loan size was 4,372 US dollars, illustrating that the loan portfolio of the lender

consists mainly of microloans and loans to small and medium sized enterprises (SME). While the

lender initially gave only loans for business purposes, in 2006 almost 30 percent were for

consumption purposes.2

Loan defaults in the table and our empirical analyses are defined as the occurrence of

payments being in arrears for more than 30 days, that is, if at least one of a borrower’s payments

2 Given the fungibility of resources, the share of consumption loans might actually be underreported.

7

was in arrears for more than 30 days at any point over the whole lifetime of the loan, we count

this as a loan default. As robustness tests, we also use time periods of 15, 60, and 90 days in the

empirical analyses. The default rate varied significantly over the sample period, from a high of

24.5 percent in the first year to a low of 1.3 percent in 2006, most likely reflecting an increase in

experience of the lender.3 The share of female borrowers is surprisingly low for a lender

operating in a developing economy, with, on average, only 20 percent, though increasing over the

last years of the sample period, to 25 percent in 2006. The share of female loan officers, on the

other hand, is very high with an average of 66 percent of loan officers being female. This share,

however, has been decreasing over time, dropping to below 50 percent in 2005 and 2006.

For our following empirical analysis, we restrict and cut the data in several ways. For the

main analysis, we restrict our attention to actual borrowers and their default behavior and thus

drop unsuccessful loan applicants. Second, we focus on a set of borrowers that have had only one

loan with the lender, for several reasons. The first reason is that the database we use is

constructed in a way that all socio-demographic borrower data are overwritten whenever a new

loan application is forwarded by a customer that had already applied for a loan before. Hence,

some of the socio-demographic data we use as control variables might not be up to date if we use

also further loan applications by the same borrower.4 The second reason is that the comparison of

first (and at the same time last) loan applications allows for a consistent comparison as all loan

officers have the same limited information about the respective borrower at the time of the

3 Note that the default frequency is not the yearly default frequency, but rather the default frequency of all loans being granted in 1996, 1997, and so on. Therefore, the low default frequency in the last year is partly due to the effect of loans still outstanding at the end of our sample period. 4 For instance, a certain customer might have applied for a loan in 1996 when she was not married and again in 2000 when she was married. As the data we use were provided by the lender in January 2007, the database would classify that particular customer as being married also in 1996, although in 1996 this was not the case.

8

application.5 In the case of repeat borrowers, loan officers already have historic information,

which they can take into consideration when granting and monitoring the loan. Focusing on the

first loan by each successful loan applicant thus allows us to study in a clean way gender-specific

loan officer performance effects. Third, we drop loans with missing gender information on the

borrower or the loan officer level. For that purpose, we exclude loans by borrowers classified as

corporate clients in the database because in these cases we cannot observe the borrowers’ gender

information. Fourth, we drop certain outliers from the sample. Specifically, we drop loans with

amounts of less than 100 US dollars and more than US 100,000 dollars. While very low values

might result from false entries in the database we want to exclude very large loans that do not fit

the definition of micro and SME loans. Additionally, we exclude loans with an unreasonable

borrower age (smaller than 18 or larger than 75 years). Finally, we exclude loans approved in

December 2006 as we cannot observe these loans’ performance.6 This reduces our sample from

31,000 to 6,775 loans granted by 141 loan officers for the main regression analysis. In robustness

tests we use a different cut of the data and obtain samples containing more than 14,000 loans.

We include a vast array of borrower, loan officer and loan characteristics in the regression

of loan defaults. Table 2 presents descriptive statistics and correlations for these variables.

Specifically, in addition to controlling for the borrower’s gender, we control for her marital

status, employment status (self-employed or salaried employee) and age. We expect female,

married and employed borrowers to be less likely to default, because of higher opportunity costs

of defaulting and more stable incomes. We also include the number of persons in the borrower’s

household and whether there is a phone available. While the availability of a phone might

5 This rests on the reasonable assumption that loan applicants and loan officers did not know each other before the loan application was forwarded. 6 The overwhelming majority of the loans have an installment frequency of one month. As the data covers the period until December 31, 2006, it is impossible for a borrower who was given a loan in December 2006 to default on her loan.

9

increase the ease of monitoring by the loan officers, there is no clear a-priori relationship between

household size and default probability. The descriptive statistics in Table 2A indicate that on

average 23.1 percent of the loans are given to female borrowers, while 57.7 percent are approved

by female loan officers. These numbers are similar to the ones presented in Table 1 and indicate

that the data selection process did not induce a strong sample bias. On average, borrowers are 39

years old, while loan officers are 25 years old. 75.4 percent of borrowers are married, while 12.4

percent are self-employed. On average, there are almost five persons in a borrower’s household

and there is a phone available in 93.2 percent of borrowers’ households.

The correlations in Panel B of Table 2 show that female, older, and married borrowers

and borrowers with a phone face a lower default probability, while household size and

employment status are not correlated with default probability. There are also many significant

correlations among borrower characteristics. For example, female borrowers are less likely to be

married or self-employed and live in smaller households.

We also control for several loan characteristics that might affect a loan’s default

probability. Specifically, we control for the annualized interest rate, the log of the approved

amount and the log of the adjusted maturity of the loan.7 Further, we include the ratio of

approved to applied loan amount and the type of collateral (personal, mortgage, or chattel

guarantee) provided.8 Higher interest rates can result in adverse selection of borrowers with

riskier projects and in riskier behavior of borrowers (Stiglitz and Weiss, 1981). Similarly, a lower

approved share might signal higher default risk, while longer-term loans tend to be riskier. On the

other hand, there is a-priori no clear relationship between collateral or loan purpose and default

7 Some loans in the database mature after 2006. These loans’ maturity was adjusted to December 31, 2006 in order to be able to compare the outstanding loans with already matured loans. 8 The use of chattel guarantees is quite common in developing economies as objects from the household of a borrower (such as a fridge or a television) often have very high (not necessarily monetary) values for the borrowers.

10

risk. The descriptive statistics in Table 2A show that annualized interest rates varied between 4.3

and 24 percent, with an average of 13.8 percent. The average loan size is 3,700 US dollars, while

the loan maturity varies between 1 month and 6 years, with an average of 16 months. On average,

borrowers received 88.8 percent of the amount they applied for.9 96.2 percent of all loans were

secured with chattel collateral, while 12.4 percent provided mortgages and 15.0 percent personal

guarantees.

The correlations in Panel B of Table 2 show that longer-term loans, loans with higher

interest rates and loans that are smaller relative to the amount originally applied for are more

likely to default, while loan size is not significantly correlated with default probability. Loans

with a personal guarantee are more likely to default, while other guarantees are not significantly

correlated with default probability. Larger and longer-term loans, loans with personal and

mortgage guarantees carry lower interest rates. Some of the loan characteristics are also

correlated with borrower characteristics. Female borrowers, for example, pay lower interest rates

and are less likely to default.

Finally, we control for several loan officer characteristics. Specifically, in addition to the

gender of loan officers, we include their age and the number of loan applications they have

processed, counted from the first loan they ever processed since they started working for the

lender. The correlation of age and experience with default probability is ex-ante not clear. While

age and experience might improve loan officers’ performance (Anderson, 2004), the career

concern view discussed in Agarwal and Wang (2008) would predict the opposite relationship.

The age of loan officers in our sample ranges from 19 to 32 years, with an average of 25 years.

On average, loan officers have processed already 223 loan applications. Additionally, we find

huge differences in their experience because the number of already processed loans ranges from 1

9 We winsorize the approved share at the first and 99th percentile to account for outliers.

11

to over 1,000 loans. The correlations in Table 2B indicate that female loan officers are, on

average, younger, while they do not have more experience in terms of loan applications

processed. Older analysts have processed more loan applications. Female loan officers are more

likely to process loan applications of female, younger, non-married, and not self-employed

borrowers. Female loan officers provide larger loans, for longer maturities and at lower interest

rates. They are more likely to process loans with personal or mortgage guarantees, but less likely

to process loans collateralized with chattel guarantees.

III. Methodology

We use several regression specifications to disentangle the relationship between loan

default probability and the gender of borrowers and loan officers. We pay particular attention to

loan officer experience to investigate whether different experience levels can explain our results.

Specifically, it may be that loan officer gender related loan performance differences are driven by

higher experience levels of a specific loan officer group. We explicitly control for this in our

regressions.

The significant correlations between the different borrower, loan officer and loan

characteristics in the previous section stress the importance of multivariate regressions.

Specifically, for the first set of results we utilize a binary probit model of the following form:

ijijii XDofficerloanFemaleFemaleDefault εδγββα +++++= **** 21 (1)

where Defaulti is a binary variable taking the value 1 if customer i defaulted on her loan (i.e. had

arrears for at least 30 days once during the lifetime of the loan), Femalei is a dummy variable

taking the value 1 for female borrowers, Female loan officerj is a dummy variable taking the

value 1 if the loan officer j serving borrower i is female, Di is a vector of control variables

12

referring to borrower and loan i, Xj is a vector of control variables referring to loan officer j and �

is an error term. In addition, we include dummies for the five branches of the lender to control for

potential clustering of loan officers of a certain gender or ability in a specific branch, year

dummies to control for macroeconomic factors that might affect default risk of borrowers, and

five business sector dummies (construction, production, other services, trade, transport) to control

for risk differences associated with the business sector the borrower operates in. Results for these

additional controls will be omitted from the tables. Standard errors are clustered at the loan

officer level, thus allowing for unobserved correlation between loans processed and monitored by

the same loan officer (Froot, 1989).10

Given that loan officers may be more likely to deal with borrowers of the same sex, for

our second set of results we will utilize several interaction terms to disentangle the relationship

between default probability and gender of borrower and loan officer

ijiji

jijii

XDofficerloanMaleMale

officerloanFemaleMaleofficerloanFemaleFemaleDefault

εδγβββα

+++

+++=

***

***

3

21 (2)

where the combination female borrower-male loan officer is the omitted category. The coefficient

β1 thus indicates whether female borrowers are more or less likely to default with a female than

with a male loan officer, while the difference between β2 and β3 indicates whether male

borrowers are more or less likely to default with a female than with a male loan officer. This

specification therefore allows us to not only control for the correlation between borrower and

loan officer gender, but also to distinguish between the performance difference of female and

male loan officers among borrowers of different genders. Similarly, we can assess the

10 As suggested by Petersen (Forthcoming) we also reproduced all our results using heteroskedasticity robust standard errors without accounting for cluster correlation (White, 1980), the results are significant at similar or even higher statistical levels. They are available upon request.

13

performance of female vs. male borrowers by considering the difference between β1 and β2 (for

female loan officers) and the coefficient on β3 (for male loan officers).

Finally, with our third set of results we assess whether the relationship between gender

and default probability varies with the experience of loan officers. For that purpose we add to

specification (2) interaction terms between the borrower-loan officer dummy of interest and a

variable proxying for the loan officer’s experience. Specifically, we control for loan officer

experience by interaction the borrower-loan officer gender with four experience quartiles

ijijj

ikjkjik

jkjiki

XDquartileExperienceofficerloanMale

MalequartileExperienceofficerloanFemaleMale

quartileExperienceofficerloanFemaleFemaleDefault

εδγ

ββ

βα

+++−

+−

+−+=

***

***

***

,3,,2

,,1

(3)

where k denotes the experience quartile (1: 0-25 percent, 2: 25-50 percent, 3: 50-75 percent, 4:

75-100 percent). This regression specification yields twelve borrower-gender-experience

interaction terms, the omitted category being the combination female borrower-male loan officer.

The experience proxies we use are the number of loan applications already handled by the loan

officer, the number of years the loan officer has worked for the microlender, and the loan

officer’s age. The sign and significance of the coefficient β1,1 (β1,4) indicate whether female loan

officers with very low (very high) experience have lower default rates for female borrowers than

male loan officers, independent of the experience of the male loan officers. If experience

differences drive the superior performance of female loan officers, then we would expect to find

a significant effect only for high experience levels (i.e. β1,3 and β1,4, that is, the third and fourth

experience quartile).

While specification (3) tests the influence of experience on loan officer performance for

the case of female borrowers, we also run a regression where the combination male borrower-

female loan officer is the omitted category. This specification allows us to test if and how

14

potential performance differences between female and male loan officers for male borrowers

depend on loan officer experience.

While the signs of the estimated coefficients of the explanatory variables indicate whether

an increase of that explanatory variable increases or decreases the probability of loan default, the

estimated coefficients of probit models do not allow us to assess the economic size of a change in

the explanatory variable. In the results section, we therefore only present marginal coefficient

estimates that are computed at the sample mean in order to be also able to derive the economic

significance of our results.

IV. Main results

The results in Column 1 of Table 3 suggest that female borrowers and borrowers served

by female loan officers are less risky. The default probability of female borrowers is 4.2 percent

lower than that of male borrowers across our sample of first (and last) loans. We also find that the

default probability of borrowers served by female loan officers is 4.7 percent lower than the

default probability of borrowers served by male loan officers. Both effects are economically

significant, as the average default rate in our sample is 13.5 percent. On the other hand, the

default probability does not vary with the experience of the loan officer. The number of loan

applications the loan officer has already processed, one of our proxies for a loan officer’s

experience, does not enter significantly.

Several other loan officer, borrower and loan characteristics enter significantly in the

column 1 regression of Table 3. First, older borrowers and borrowers served by older loan

officers are less likely to default. The latter result contradicts the career concern hypothesis by

Aggarwal and Wang (2008). Second, consistent with Stiglitz and Weiss (1981), the interest rate

15

is positively, significantly, and economically very substantially associated with a higher default

probability. Third, married borrowers and borrowers from households where a phone is available

are less likely to default, suggesting higher opportunity costs for these borrowers. Fourth, larger

loans and loans with longer maturities are more likely to turn non-performing. Fifth, the higher

the ratio of approved to applied loan amount, the lower is the default probability. Finally, loans

with personal guarantees are more likely to turn bad, while loans guaranteed with mortgages are

less likely to default. An explanation for this finding may be that personal guarantees, which are

third-party guarantees, induce a moral hazard, while the potential loss of the own house sets

strong repayment incentives. Overall, the fit of our model is satisfactory, with 75% of the

defaulted loans predicted correctly and 61% of the non-defaulted loans and a Pseudo R-square of

13%.11

Since the finding that female loan officers experience lower default rates might be driven

by the fact that female borrowers are less risky than male borrowers and might be more often

served by female loan officers, we next construct borrower gender-loan officer gender

combinations as dummy variables and run a regression using regression specification (2).

Specifically, we interact borrower and loan officer gender, with the combination female

borrower-male loan officer being the omitted category. In our baseline sample, 68% (55%) of

female (male) borrowers are screened and monitored by female loan officers.

Column 2 of Table 3 shows the robustness of our previous findings to controlling for the

correlation between borrowers’ and loan officers’ genders. Compared to female borrowers

monitored by male loan officers, female borrowers monitored by female loan officers have a

default probability that is 4.3 percent lower. Similarly, we find that the default probability of

11 In classifying observations, predicted probabilities significantly higher than 13.5% (average default probability) are classified as default observations and those below 13.5% are classified as no default. We adjust this benchmark depending on the sample and default definition.

16

male borrowers monitored by female loan officers is 4.8 percent lower than the default

probability of male borrowers monitored by male loan officers. This suggests that, independent

of the gender of the borrower, female loan officers are better in managing default risk.

Comparing the marginal effects of the different borrower-loan officer dummies, we also confirm

that male borrowers are more likely to default than female borrowers. In the case of female loan

officers, male borrowers default 4.4 percent more often and in the case of male loan officers they

default 3.8 percent more often. Our previous findings on the different loan officer, loan and

borrower characteristics are confirmed by this regression.

Columns 3 to 5 of Table 3 show the robustness of our results to using alternative

definitions of default. Specifically, we redefine default as having a payment in arrears for more

than 15 days (column 3), 60 days (column 4) and 90 days (column 5). Our findings are all

confirmed for the stricter default definition of 15 days. Here we also find that the advantage of

female loan officers vis-à-vis their male peers appears to be stronger for female borrowers (5.8

percent) than for male borrowers (4.9 percent). In the case of less strict definitions (columns 4

and 5), the size of the marginal effect of loan officer’s gender for female borrower declines but

stays significant, while the effect of loan officer’s gender turns insignificant for male borrowers.

Finally, in column 6 of Table 3 we confirm our findings for a larger sample of first loans

for which we have also subsequent loan information. Here, we do not restrict our attention to the

first loans that were at the same time the last (and thus only) loans by the borrowers, but we use

all first loans available in the database. As in this case we cannot be sure that the socio-

demographic information has not changed after the first loan, we exclude all socio-demographic

variables from the regression. This less strict cut of the data leaves us with a sample containing

14,020 first loans. The column 6 results of Table 3 show that even in this larger sample, we

confirm our finding that female loan officers are more efficient in preventing a loan default than

17

their male peers.12 While the marginal effects are somewhat smaller in size, we still find that

female loan officers are better in preventing loan defaults than their male peers, both for female

and for male borrowers. The results for the other controls are very similar to our previous

regressions. The overall fit of the model decreases, as can be seen from the Pseudo R-squares and

the percentages of correctly predicted observations, underlining the importance of the socio-

demographic borrower characteristics in predicting default.

Our results so far suggest that female loan officers are more efficient in preventing loan

defaults than male loan officers. However, these results might be driven by different levels of job

experience. For instance, if female loan officers were more experienced in monitoring borrowers

than their male peers, we might expect them to perform better, that is, have lower default rates.

To control for this possible driver of the results, we interact the borrower gender-loan officer

gender dummy variable with different levels of loan officer experience. Specifically, we utilize

four experience quartiles and build an interaction term for each quartile. This yields four

interaction terms for each borrower-loan officer-experience quartile combination, and twelve

interaction terms overall. As before, the combination female borrower-male loan officer is the

reference category. Loan officer experience is proxied by the number of previous loan

applications the loan officer has processed13, the number of years the loan officer has worked for

the microlender, and the age of the loan officer. The Table 4 regressions are thus based on

regression specification (3).

The Table 4, Panel A, column 1 regression shows that the advantage of female loan

officers in managing the default risk of female borrowers holds for all, but the lowest quartile of

experience. Specifically, for the second, third and fourth quartiles of job experience, we can

12 For this regression we use the 30 days in arrears default definition. In unreported regressions we confirm our earlier findings using this bigger sample without socio-demographic data for the alternative default definitions. 13 We divide the number of loan applications per loan officer by 1,000 for scaling reasons.

18

confirm that female loan officers perform better than male loan officers. The sizes of the

marginal effects are similar to our previous finding from Table 3, column 2, and first increase up

to the third quartile of experience before decreasing again. In Panel B, we use the combination

male borrower-female loan officer as reference category. We find that female loan officers are

better than male loan officers in managing default risk of male borrowers for the third and fourth

quartiles of experience. The economic significance of the effect is considerably higher than

before, in particular for very high levels of experience.

Column 2 presents the results when using the time since the loan officer works for the

microlender as experience proxy. The use of this alternative experience proxy shows that the

performance difference for female borrowers exists at all but the highest levels of experience.

Interestingly, though, the effect vanishes at the fourth quartile of experience. The magnitude of

the effect is again similar to before. We further find that the performance gap with regard to male

borrowers exists already for medium experience and remains significant up to very high

experience, being statistically significant for the second, third and fourth quartiles.

Finally, column 3 presents the results when using loan officer age as experience proxy.

While being only a crude experience proxy, the use of this third alternative does not alter our

findings. As before, the performance of female loan officers with regard to female borrowers is

only indistinguishable from male loan officers for low levels of experience, but significantly

better for the second, third and fourth quartiles of their age. For male borrowers, we again find

performance differences for high and very high experience levels.

All in all, we conclude from these tests that the superior performance of female loan

officers for female borrowers is not driven by their higher experience. Only for male borrowers

we find slight evidence that the performance advantage of female loan officers relative to their

male counterparts is significant only at experience levels above the median.

19

V. Robustness and Additional Tests

We subject our findings to several sensitivity analyses in order to test their robustness but

also to explore the channels through which the relationship between loan officer gender and

default probability works. We fist loosen the strict sample selection that we had chosen for our

baseline regression. Specifically, we expand the sample from first loans to borrowers’

subsequent, that is, repeat, loans. This allows us a robustness test in two aspects: first, we have a

different sample, but, second, we expect a less significant relationship between the gender of the

loan officer and default probability as the information asymmetries and thus agency problems

between bank and borrower should be lower. We thus would interpret a somewhat weaker

finding of a female performance advantage on the sample of subsequent borrowers as

confirmation of women’s advantage in monitoring borrowers. We also test for this directly by

including an interaction term with a variable indicating the duration of the borrower’s

relationship with the microbank.

For this robustness check we include several control variables that capture a borrower’s

loan history with the bank. While for first loans there is no loan history available, here we can

make use of historic information. Specifically, we control for the duration of the lending

relationship in years, whether any previous loan application of the borrower has been rejected and

whether the borrower has ever defaulted on any loan granted by the lender before applying for a

new loan. We thus use specification (2) and add the three control variables for the borrowers’

loan history with the bank. As in the baseline regression, we first focus on a sample of last loans

to be able to control for socio-demographic borrower characteristics. Cutting the data in this way

leaves us with 6,448 repeat loans. We then drop the socio-demographic variables and focus on a

broader sample of repeat loans. This yields a sample size of 12,940 loans. Focusing on further

20

loans and including loan history variables increases the fit of the model significantly, as can be

seen from the higher Pseudo R-square and percentages of correctly predicted defaults.

The results in column 1 of Table 5 confirm the findings and their interpretation with a

regression using repeat instead of first loans. We continue to find that female borrowers screened

and monitored by female loan officers have a lower default probability than if screened and

monitored by male loan officers, while there is no significant difference for male borrowers.

However, even in the case of female borrowers, the economic size is substantially smaller than

before, with only 1.8 percent, compared to the 4.3 percent we found in Table 3, column 2. Large

proportions of the explanatory power seem to shift to the loan history data. This observation is

consistent with Mester et al. (2007) who show that previous customer information help financial

institutions to monitor their borrowers. Specifically, we find that defaults are on average 37.1

(3.7) percent more likely if the same borrower defaulted on a previous loan (had a rejected loan

application before). In spite of this, however, we continue to find a loan officer effect. This is a

very interesting finding because it illustrates that even if historic, loan default relevant borrower

information is used, there are still differences between female and male loan officers.14 The

column 2 regression of Table 5 shows that this performance gap is not a function of how long the

borrower has been borrowing from the institution because the interaction terms between the

borrower-loan officer gender pairs and the duration of the lending relationship do not enter

significantly.

The results in columns 3 and 4 of Table 5 largely confirm these findings for a larger

sample of 12,940 subsequent loans that is not limited to last loans. As before, we do not use the

socio-demographic borrower characteristics for these regressions, which again reduces the fit of

14 Note, however, that this is only the case for female borrowers because results of unreported regressions show that male borrowers served by female loan officers do not have different default probabilities compared with male loan officers.

21

the model. The column 3 results without the interaction term show that the performance

advantage of female vis-à-vis male loan officers is now only 1.5 percent for female borrowers.

For male borrowers the advantage is 1.0 percent, but only weakly significant. The size of the

performance gap for female borrowers in column 4 remains, but looses significance, and the

unreported marginal effect for male borrowers does also not enter significantly. Again, we do not

find that the performance gap is a function of the duration of borrowers’ lending relationship with

the bank. Taken together, the results in Table 5 suggest that the performance advantage of female

vis-à-vis male loan officers continues to hold for repeat loans. However, this is true only in the

case of female borrowers. We also find that this effect is smaller for repeat loans compared with

first loans, while it is not a function of the duration of borrowers’ relationship with the bank. It

thus seems that the learning effect that reduces the performance advantage of female loan officers

vis-à-vis their male counterparts kicks in with the second loan.

Finally, we test whether the advantage of female loan officers vis-à-vis their male

counterparts arises from their better screening capacities of loan applicants. For this test, we use a

sample containing both successful and unsuccessful loan applications and run the following

regression

ijiji

jijii

XDofficerloanMaleMale

officerloanFemaleMaleofficerloanFemaleFemaleApproval

εδγβββα

+++

+++=

***

***

3

21 (4)

Specification (4) differs from (2) since the dependent variable is now a dummy variable

indicating whether a loan application was approved (Approvali = 1) or not. This enables us to test

if female loan officers are less likely than their male counterparts to accept loan applicants of a

specific gender. Performing this test allows us to exclude ex ante borrower selection as the driver

of the performance differences between female and male loan officers. Specifically, if we do not

find any significant difference between female and male loan officers, then ex ante selection does

22

not drive our previous findings. In contrast to specification (2) we are not able to use some loan-

related control variables, such as the interest rate, because these are not available at the time of

the loan application. Note also that rather than using the approved loan amount as a loan size

proxy, we use the applied loan amount, and rather than using the approved maturity we use the

applied maturity.

We test for ex ante sample selection using four different samples. At first, we use a

sample of first loan applications, which at the same time were the last applications, thus

corresponding to the specification of Table 3 (columns I to V), with 8,297 loan applicants, around

92% of which were accepted.15 Second, we drop the socio-demographic variables and include all

first loan applications, yielding a sample of 15,986 loan applications. Third, we use a sample of

repeat borrowers. Again, we run a specification with loan applications that were at the same time

last loan applications (sample size of 7,240 loan applications) and a specification without this

restriction and thus without socio-demographic borrower characteristics (14,502 loan

applications).

The results in Table 6 illustrate that our finding of a superior performance of female vis-à-

vis male loan officers is not driven by selection bias of the borrowers or better screening capacity

of female loan officers. We do not find any significant difference in the likelihood of borrowers

to be accepted by female or male loan officers, independent of whether the borrower is male or

female. Further, we do find that male loan officers are more likely to accept loan applications of

male clients. Overall, this test suggests that screening differences between female and male loan

officers do not drive the performance gap between them. It rather indicates that the results are

driven by better monitoring of the female loan officers.

15 Here we also include loans approved in December 2006, unlike for the arrears regressions.

23

VI. Conclusions

This study is, to the best of our knowledge, the first to consider gender differences in loan

officer performance. While some papers have reported gender differences with regard to

investment decisions (e.g. Barber and Odean, 2001; Charness and Geezy, 2007) or the general

behavior of women in competitive environments (e.g. Gneezy et al., Forthcoming; Niederle and

Vesterlund, Forthcoming), we provide novel results about the role of gender in financial

institutions. Contrary to Green et al. (2008) who document that women seem to perform worse

than men in quantifiable aspects of the job, we find convincing evidence that women may also

perform better than men in quantifiable job aspects such as the management of loan default risk.

Although the job environment in financial institutions is usually highly competitive, we further

find counter-evidence to several papers (e.g., Gneezy et al., 2003) which show that females

underperform their male peers in highly competitive environments. Additionally, we find that

borrowers served by older loan officers are less likely to default which is in line with results from

Anderson (2004) but on the other hand contradicts the career concern hypothesis by Aggarwal

and Wang (2008).

Our estimations also shed light on the mechanisms. We find that female loan officers are

not more or less likely to accept borrowers with the same characteristics. Further, ex-ante risk

differences captured by interest rates do not influence our findings as we explicitly control for the

interest rate in our regressions. It thus seems to be the better monitoring of borrowers that

explains the lower default risk in the case of female loan officers. Finally, there is no convincing

evidence that better experience explains the advantage of female loan officers vis-à-vis their male

colleagues.

24

Our results do not only contribute to the literature on gender differences in finance and

economics, but also to the growing body of literature on the role of loan officers in financial

institutions. They suggest that the performance of loan officers in financial institutions is not only

driven by setting the right incentives, for instance by implementing a routine job rotation

mechanism as in Hertzberg et al. (2008), or by the degree of asymmetric information in the

institution as in Agarwal and Fang (2008), but also by gender-specific differences between

female and male loan officers.

25

References

Agarwal, S., Wang, F.H., 2008, Motivating Loan Officers: An Analysis of Salaries and

Piece Rates Compensation, Working Paper, Federal Reserve Bank of Chicago.

Agnew, J., Balduzzi, P., Sundén, A., 2003, Portfolio Choice and Trading in a Large

401(k) Plan, American Economic Review 93, 193-215.

Andersson, P., 2004, Does Experience Matter in Lending? A Process-tracing Study on

Experienced Loan Officers' and Novices' Decision Behavior, Journal of Economic Psychology

25, 471-492.

Armendariz de Aghion, B., Morduch, J., 2005, The Economics of Microfinance, MIT

Press.

Barber, B.M, Odean, T., 2001, Boys will be Boys: Gender, Overconfidence, and Common

Stock Investment, Quarterly Journal of Economics 116, 261-292.

Barsky, R.B., Juster, F.T., Kimball, M.S., Shapiro, M.D., 1997, Preference Parameters

and Behavioral Heterogeneity: An Experimental Approach in the Health and Retirement Study,

Quarterly Journal of Economics 112, 537-579.

Berger, A.N., Miller, M., Petersen, M., Rajan, R., Stein, J., 2005, Does Function Follow

Organizational Form? Evidence from the Lending Practices of Large and Small Banks, Journal of

Financial Economics 76, 237-269.

Berger, A.N., Udell, G.F., 2004, The Institutional Memory Hypothesis and the

Procyclicality of Bank Lending Behavior, Journal of Financial Intermediation 13, 458-495.

Black, S., Strahan, P. E., 2001, The Division of Spoils: Rent-Sharing and Discrimination

in a Regulated Industry, American Economic Review 91, 814-831.

26

Charness, G., Gneezy, U., 2007, Strong Evidence for Gender Differences in Investment,

Working Paper.

Christiansen, C., Schröter Joensen, J., Rangvid, J., 2006, Gender, Marriage, and the

Decision to Invest in Stocks and Bonds: Do Single Women Invest More in Less Risky Assets?,

Working Paper.

Darity Jr., W.A., Mason, P.L., 1998, Evidence on Discrimination in Employment: Codes

of Color, Codes of Gender, Journal of Economic Perspectives 12, 63-90.

Froot, K.A., 1989, Consistent Covariance Matrix Estimation with Cross-Sectional

Dependence and Heteroskedasticity in Financial Data, Journal of Financial and Quantitative

Analysis 24, 333-355.

Gneezy, U., Leonard, K.L., List, J.A., Forthcoming, Gender Differences in Competition:

Evidence from a Matrilineal and a Patriarchal Society, Econometrica.

Gneezy, U., Niederle, M., Rustichini, A., 2003, Performance in Competitive

Environments: Gender Differences, Quarterly Journal of Economics 118, 1049-1074.

Goldin, C., Rouse, C., 2000, Orchestrating Impartiality: The Impact of “Blind” Auditions

on Female Musicians, American Economic Review 90, 715-741.

Green, C. T., Jegadeesh, N., Tang, Y., 2008, Gender and Job Performance: Evidence from

Wall Street, NBER Working Paper No. 12897.

Hertzberg, A., Liberti, J.M., Paravisini, D., 2008, Information and Incentives Inside the

Firm: Evidence from Loan Officer Rotation, Working Paper.

Liberti, J.M., Mian, A., Forthcoming, Estimating the Impact of Hierarchies on

Information Use, Review of Financial Studies.

Mester, L.J., Nakamura, L.I., Renault, M., 2007, Transactions Accounts and Loan

Monitoring, Review of Financial Studies 20, 529-556.

27

Mian, A., 2006, Distance Constraints: The Limits of Foreign Lending in Poor Economies,

Journal of Finance 61, 1465-1505.

Niederle, M., Vesterlund, L., Forthcoming, Do Women Shy Away From Competition? Do

Men Compete Too Much?, Quarterly Journal of Economics.

Petersen, M., Forthcoming, Estimating Standard Errors in Finance Panel Data Sets:

Comparing Approaches, Review of Financial Studies.

Stiglitz, J.E., Weiss, A.M., 1981, Credit Rationing in Markets with Imperfect Information,

American Economic Review 71, 393-410.

White, H., 1980, A Heteroskedasticity-Consistent Covariance Matrix and a Direct Test

for Heteroskedasticity, Econometrica 48, 817-838.

Table 1: Some statistics on the lending institution

This table contains a broad overview for the 5 Tirana branches of the Albanian microlender. The loan size is given in US dollars and the interest rate is per annum. The default frequency is measured as the occurrence of a borrower being in arrears for more than 30 days during the lifetime of her loan. It is not the yearly default frequency, but rather the default frequency of all loans being granted in 1996, 1997, and so on. Business loans incorporate investments into fixed assets and working capital. Real estate loan usages include the purchase, construction, improvement and extension of houses. New loan Loan usage Share Share Year of Approved Loan volume Default Business Real of female of female application Applications loans size (1,000) Borrowers frequency Loans estate Consuming borrowers loan officers 1996 794 351 3,646 2,895 297 0.245 1.000 0.000 0.000 0.195 0.644 1997 454 251 3,348 1,520 227 0.080 1.000 0.000 0.000 0.183 0.709 1998 932 481 4,616 4,302 413 0.085 1.000 0.000 0.000 0.172 0.885 1999 1,057 590 5,287 5,588 545 0.034 0.863 0.137 0.000 0.213 0.848 2000 2,390 1,438 4,062 9,709 1,277 0.102 0.695 0.305 0.000 0.193 0.616 2001 2,230 1,456 3,674 8,193 1,308 0.063 0.656 0.342 0.002 0.198 0.651 2002 2,495 1,907 5,762 14,400 1,746 0.059 0.543 0.385 0.072 0.185 0.757 2003 3,737 2,941 6,455 24,100 2,697 0.049 0.426 0.328 0.246 0.226 0.656 2004 9,656 7,836 4,068 39,300 7,010 0.108 0.398 0.303 0.299 0.242 0.522 2005 9,437 7,339 3,996 37,700 6,582 0.098 0.619 0.167 0.214 0.212 0.463 2006 9,944 7,024 3,176 31,600 6,152 0.013 0.625 0.076 0.283 0.250 0.499 Sum 43,126 31,614 179,307 Average 4,372 0.085 0.711 0.186 0.101 0.206 0.659

Table 2A: Descriptive statistics

This table contains borrower, loan, and loan officer characteristics for the 5 Tirana branches of the Albanian microlender. All variables are provided for a sub sample of 6,775 approved loans to individual, private borrowers. The table concentrates on the first and last loans for each borrower. We further drop loans with unreasonable entries for the borrower’s age (smaller than 18 or larger than 75 years), missing gender information for borrower and loan officer, and applied loan size (smaller than 100 or larger than 100,000 US dollars). Female is a dummy variable indicating the gender of the borrower (female = 1), Female loan officer is a dummy variable indicating the gender of the loan officer (female = 1), Age of borrower is the age of the borrower at the time of the loan application, Civil status is a dummy variable indicating whether the borrower is married (married = 1), Self employed is a dummy variable indicating whether the borrower is self-employed or a wage earner, Number persons household indicates how many persons other than the borrower are in the household of the borrower, Phone availability is a dummy variable indicating whether the borrower has a phone or not (phone available = 1), Approved amount is the loan size granted in US dollars, Adjusted maturity is the of the loan maturity adjusted such that no loan has a maturity greater than December 31, 2006, Interest rate is the annual interest rate charged on the loan, Approved share is the ratio of applied amount to approved amount in percent, Personal guarantee, Mortgage guarantee, and Chattel guarantee are all dummy variables indicating whether any of the three respective types of collateral are pledged by the borrower, Applications per loan officer is a loan officer experience proxy, which indicates the number of loan applications handled by the loan officer until the respective loan was granted, Age of loan officer is the age of the loan officer at the time the loan was granted measured in years. Variable Mean Minimum 25%-Quartile Median 75%-Quartile Maximum Default 0.135 Female 0.231 Female loan officer 0.577 Age of borrower 39 18 30 38 47 74 Civil status 0.754 Self employed 0.124 Number persons household 4,825 1 4 5 6 21 Phone availability 0.932 Approved amount 3,729 140 1,433 2,322 3,455 100,000 Adjusted maturity 488 31 328 450 610 2060 Interest rate 0.138 0.043 0.127 0.148 0.160 0.241 Approved share 0.888 0.300 0.789 1.000 1.000 1.333 Personal guarantee 0.150 Mortgage guarantee 0.124 Chattel guarantee 0.962 Applications per loan officer 223 1 66 163 325 1090 Age of loan officer 25 19 23 24 26 32

Table 2B Correlation matrix

This table contains the pair-wise correlations for borrower, loan, and loan officer characteristics for the 5 Tirana branches of the Albanian microlender. All variables are provided for a sub sample of for 6,775 approved loans to individual, private borrowers. Refer to Table 2A for a description of the variables and the sample selection. * indicates a significance level of at least 0.05.

Female loan Age of Civil Self Number persons Phone Approved Adjusted Interest Approved Personal Mortgage Chattel Applications

Default Female officer borrower status employed household availability amount maturity rate share guarantee guarantee guarantee per loan officer

Default

Female -0.0733*

Female loan officer -0.0533* 0.1101*

Age of borrower -0.0760* -0.0073 -0.0575*

Civil status -0.0503* -0.1221* -0.1183* 0.4867*

Self employed -0.0165 -0.0591* -0.2812* 0.0278* 0.0734*

Number persons household -0.014 -0.1548* -0.1137* 0.1417* 0.3202* 0.0822*

Phone availability -0.1097* 0.0098 -0.0027 -0.0424* -0.0546* 0.0558* 0.007

Approved amount -0.011 0.0082 0.0713* 0.0480* 0.0579* -0.0202 -0.0096 0.0323*

Adjusted maturity 0.0549* -0.0092 0.1260* 0.0476* 0.0439* -0.1799* -0.0546* 0.008 0.4295*

Interest rate 0.1024* -0.0600* -0.1255* -0.0425* -0.02 0.0831* 0.0628* -0.0684* -0.4484* -0.2889*

Approved share -0.0586* 0.0537* 0.0758* -0.0252* -0.0812* 0.0064 -0.1009* 0.0867* 0.1099* 0.1631* -0.1224*

Personal guarantee 0.0489* -0.0249* 0.0433* 0.0141 0.0481* -0.1491* 0.0302* -0.1218* 0.1957* 0.3437* -0.2573* 0.0115

Mortgage guarantee -0.0102 -0.0181 0.0886* 0.0753* 0.0811* -0.0953* -0.0304* -0.1177* 0.5675* 0.5029* -0.4196* -0.0018 0.2982*

Chattel guarantee 0.0193 -0.004 -0.0673* -0.0401* -0.0266* 0.0649* 0.0751* 0.1122* -0.2184* -0.2929* 0.1769* 0.0073 -0.1602* -0.4853*

Applications per loan officer -0.0671* 0.0496* 0.0169 -0.0306* -0.0684* 0.0860* -0.0107 0.0931* -0.1233* -0.4110* 0.0414* -0.0148 -0.2353* -0.2069* 0.1190*

Age of loan officer -0.0553* -0.0131 -0.3478* 0.0172 0.0077 0.2260* 0.0254* 0.0657* -0.0411* -0.1971* 0.0003 -0.0069 -0.0947* -0.0780* 0.0634* 0.2770*

31

Table 3: Default probability and loan officers’ gender – first loans This table contains the marginal effects of the outcome test with the gender of borrowers and loan officers. The first five regression models are based on sub samples of approved loans to individual, private borrowers. They are at the same time first and last loans per borrower. Regression model VI comprises all 14,020 first loans and does not contain socio-demographic variables. For regression models I, II, and VI, the dependent variable is the occurrence of a borrower being in arrears for more than 30 days during the lifetime of her loan. Regression models III, IV, and V use arrear definitions of 15, 60, and 90 days. The independent variables are as described in Table 2A except the number of loan applications per borrower which is divided by 1,000. Instead of the raw numbers we employ the natural logarithm for the approved amount (ln(approved amount)) and the adjusted maturity (ln(adjusted maturity)). In regression models II to VI, we interact the borrower and loan officer gender: Female & Female loan officer is a dummy variable indicating the combination of a female borrower and a female loan officer, Male & Female loan officer indicates the combination of a male borrower and a female loan officer, Male & Male loan officer indicates the combination of a male borrower and a male loan officer. The combination Female & Male loan officer serves as the reference group. We also control for five loan destinations (working capital, fixed assets, mixed purpose, real estate, and consuming), five business sectors (construction, production, trade, transport, other services), five branches, and the years from 1996-2006. Results for these control variables are omitted. Standard errors are clustered at the loan officer level. *,**,*** indicate significance at the 10%, 5% and 1% level, respectively. Marginal effects for regression model Independent variable I II III IV V VI Female -0.042*** Female loan officer -0.047*** Female & Female loan officer -0.043*** -0.058*** -0.033*** -0.035*** -0.023*** Male & Female loan officer -0.005 -0.009 -0.005 -0.012 -0.004 Male & Male loan officer 0.044*** 0.042** 0.013 -0.002 0.027*** Loan applications per loan officer 0.018 0.018 0.030 -0.024 -0.030 0.013 Age of loan officer -0.009*** -0.009*** -0.010*** -0.004** -0.003* -0.005*** Interest rate 0.882*** 0.883*** 1.078*** 0.384*** 0.296*** 0.559*** Age of borrower -0.002*** -0.002*** -0.002*** -0.001*** -0.001*** -0.002*** Civil status -0.031*** -0.031*** -0.046*** -0.007 -0.002 Self employed 0.014 0.014 0.032 0.012 0.015 Number persons household -0.001 -0.001 -0.003 -0.002 -0.002 Phone availability -0.092*** -0.093*** -0.130*** -0.038** -0.035** ln(approved amount) 0.015* 0.015* 0.012 0.022*** 0.017*** 0.015** ln(adjusted maturity) 0.029** 0.029** 0.057*** 0.000 0.000 0.050*** Approved share -0.107*** -0.106*** -0.124*** -0.075*** -0.060*** -0.043*** Personal guarantee 0.026** 0.026** 0.036** 0.005 0.009 0.014* Mortgage guarantee -0.028** -0.028** -0.037** -0.031*** -0.024*** -0.014 Chattel guarantee 0.017 0.016 0.016 0.009 0.009 0.018 Observations 6,775 6,775 7,107 6,770 6,571 14,020 Pseudo R square 0.127 0.127 0.142 0.139 0.148 0.090 Share of default correctly predicted 75.4 75.8 75.2 79.9 78,2 71.3 Share of non-default correctly predicted 61.4 61.1 62.2 60.7 65.6 62.0

Table 4: Default probability and loan officers’ gender – interaction with experience

This table contains the marginal effects of the outcome test with the gender of borrowers and loan officers together with interactions with loan officer experience. All four regression models are based on the sub sample of 6,775 approved loans to individual, private borrowers, corresponding to regression model II in Table 3. They are at the same time first and last loans per borrower. The dependent variable is the occurrence of a borrower being in arrears for more than 30 days during the lifetime of her loan. In addition to the already used independent variables described in Table 3 we interact Female & Female loan officer with the loan officer’s experience that is proxied by the number of loan applications handled by the respective loan officer until a certain loan was granted (I); the time since the loan officer works for the microlender (II), and the age of the loan officer at the time of the loan approval. To test whether the loan officer effects depends on loan officer experience we use interactions with the four experience quartiles for each experience proxy. Control variables are the same as in Table 3, results for most of these controls are omitted. In Panel A, the combination Female & Male loan officer serves as the reference group, in Panel B, the combination Male & Female loan officer serves as the reference group. Standard errors are clustered at the loan officer level. *,**,*** indicate significance at the 10%, 5% and 1% level, respectively. Marginal effects for regression model Independent variable I II III Panel A: Performance differences for female borrowers Female & Female loan officer & 0-25% Experience -0.033 -0.050* -0.026 Female & Female loan officer & 25-50% Experience -0.042* -0.044** -0.063*** Female & Female loan officer & 50-75% Experience -0.056*** -0.057*** -0.041* Female & Female loan officer & 75-100% Experience -0.052*** -0.014 -0.056*** Male & Female loan officer & 0-25% Experience 0.010 -0.020 0.022 Male & Female loan officer & 25-50% Experience 0.003 -0.007 -0.023 Male & Female loan officer & 50-75% Experience -0.008 0.006 -0.014 Male & Female loan officer & 75-100% Experience -0.037** 0.018 -0.012 Male & Male loan officer & 0-25% Experience 0.036* 0.012 0.042* Male & Male loan officer & 25-50% Experience 0.022 0.053** 0.015 Male & Male loan officer & 50-75% Experience 0.061*** 0.035* 0.045** Male & Male loan officer & 75-100% Experience 0.063** 0.098*** 0.063*** Loan applications per loan officer 0.027 0.027 Age of loan officer -0.008*** -0.007*** -0.008** Time since first loan application -0.019* Panel B: Performance differences for male borrowers Male & Male loan officer & 0-25% Experience 0.030 0.033 0.036 Male & Male loan officer & 25-50% Experience 0.022 0.062** 0.016 Male & Male loan officer & 50-75% Experience 0.073*** 0.032* 0.057** Male & Male loan officer & 75-100% Experience 0.111*** 0.083*** 0.093*** Observations 6,775 6,775 6,775 Pseudo R square 0.130 0.129 0.131 Share of default correctly predicted 74.5 75.1 75.2 Share of non-default correctly predicted 61.7 61.5 61/5

Table 5: Default probability and loan officers’ gender – further loans

This table contains the marginal effects of the outcome test with the gender of borrowers and loan officers together with interactions with the duration of the lending relationship. Regression models I and II (III and IV) are based on the sub sample of 6,448 first and last (12,940 first) loans to individual, private borrowers. The dependent variable is the occurrence of a borrower being in arrears for more than 30 days during the lifetime of her loan. The independent variables are as in previous tables except for three variables for the loan history of each borrower with the microlender: Duration relationship provides the number of years since the first loan application of the borrower, Any previous application rejected is a dummy variable indicating any previous rejection of a loan application (1 = rejection), Any previous loan defaulted is a dummy variable indicating any previous default (1 = default). We further use three interaction terms between the borrower gender-loan officer gender pairs and Duration relationship in regression models II and IV. Results for our additional control variables are omitted. Standard errors are clustered at the loan officer level. *,**,*** indicate significance at the 10%, 5% and 1% level, respectively. Marginal effects for regression model Independent variable I II III IV Female & Female loan officer -0.018*** -0.021** -0.015** -0.015 Male & Female loan officer -0.004 -0.007 -0.007 0.000 Male & Male loan officer 0.003 -0.006 0.003 0.007 Female & Female loan officer & Duration relationship 0.002 0.000 Male & Female loan officer & Duration relationship 0.001 -0.004 Male & Male loan officer & Duration relationship 0.005 -0.002 Duration relationship -0.008*** -0.011*** -0.007*** -0.005* Any previous application rejected 0.037*** 0.036*** 0.027*** 0.027*** Any previous loan defaulted 0.371*** 0.374*** 0.298*** 0.297*** Loan applications per loan officer 0.001 0.000 0.001 0.001 Age of loan officer -0.002 -0.002 -0.001 -0.001 Interest rate 0.228** 0.226** 0.270*** 0.268*** Age of borrower -0.001*** -0.001*** -0.001*** -0.001*** Civil status -0.017** -0.017** Self employed 0.011 0.010 Number persons household 0.001 0.001 Phone availability -0.033** -0.033** ln(approved amount) 0.005 0.005 0.009*** 0.009*** ln(adjusted maturity) 0.032*** 0.032*** 0.044*** 0.044*** Approved share -0.036*** -0.036*** -0.027*** -0.026*** Personal guarantee 0.000 0.000 0.008 0.009 Mortgage guarantee -0.022*** -0.022*** -0.011* -0.010* Chattel guarantee 0.015* 0.015* 0.010 0.011 Observations 6,448 6,448 12,940 12,940 Pseudo R square 0.270 0.270 0.171 0.171 Share of default correctly predicted 84.9 84.9 75.2 75.1 Share of non-default correctly predicted 70.4 70.6 67.2 67.2

34

Table 6: Loan approval and loan officers’ gender

This table contains the marginal effects for a sample selection test with the gender of borrowers and loan officers. The regression models are based on different sub samples of requested loans by individual, private borrowers: model I is based on 8,297 loan applications that are at the same time first and last applications per borrower; model II uses 15,986 first loan applications; model III employs 7,240 loan applications that are at the same time further and last applications per borrower; model IV is based on 14,502 further loan applications. The dependent variable is the approval decision (1 for an approved loan, 0 otherwise). We use a different set of control variables because we cannot use variables that are not available at the time of the loan application, such as the interest rate. Specifically, we employ the natural logarithm of the applied instead of the approved loan size, and the natural logarithm of the applied instead of the approved maturity. We use the same further control variables described in Table 3. Results for these control variables are omitted. Standard errors are clustered at the loan officer level. *,**,*** indicate significance at the 10%, 5% and 1% level, respectively. Marginal effects for regression model Independent variable I II III IV Female & Female loan officer 0.009 -0.003 -0.014 -0.014 Male & Female loan officer 0.017 0.004 -0.011 -0.015 Male & Male loan officer 0.012 0.006 0.009 0.009 Duration relationship 0.001 -0.001 Any previous application rejected -0.065*** -0.053*** Any previous loan defaulted -0.106*** -0.090*** Loan applications per loan officer 0.027 0.024 0.004 -0.013 Age of loan officer -0.006*** -0.005*** -0.004** -0.006*** Age of borrower 0.000 0.000 0.001*** 0.001*** Civil status -0.006 0.008 Self employed 0.022** 0.003 Number persons household 0.009*** 0.003* Phone availability 0.033** 0.037*** ln(applied amount) -0.009* -0.007** -0.008** -0.009*** ln(applied maturity) 0.034*** 0.022*** 0.030*** 0.040*** Personal guarantee 0.025*** 0.018*** 0.015 0.008 Mortgage guarantee 0.018 0.007 -0.014 -0.010 Chattel guarantee 0.151*** 0.096*** 0.074*** 0.036** Observations 8,297 15,986 7,240 14,502 Pseudo R square 0.094 0.095 0.130 0.171 Share of approvals correctly predicted 66.8 65.1 72.3 72.0 Share of non-approvals correctly predicted 62.5 63.4 67.9 69.5