NBER WORKING PAPER SERIES
THE IMPACT OF MINORITY REPRESENTATION AT MORTGAGE LENDERS
W. Scott Frame
Ruidi Huang
Erik J. Mayer
Adi Sunderam
Working Paper 30125
http://www.nber.org/papers/w30125
NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue
Cambridge, MA 02138
June 2022
This paper was previously circulated under the title “Minority Loan Officers and Minorities’
Access to Mortgage Credit.” For helpful comments we thank Sam Antill, Francesco D’Acunto,
Will Dobbie, Mark Egan, Kristopher Gerardi, Paige Ouimet, James Vickery, and seminar
participants at Brigham Young University, the University of Wisconsin–Madison, Southern
Methodist University, and the Mortgage Bankers Association. We also thank conference
participants at the Mid-Atlantic Research Conference in Finance and the SFS Cavalcade. The
views expressed are those of the authors and not necessarily those of the Federal Reserve Bank of
Dallas or any other entity within the Federal Reserve System. We are grateful to the Conference
of State Bank Supervisors (CSBS) for granting us access to data from NMLS Consumer
AccessSM, please see https://nmlsconsumeraccess.org/. The results and opinions are those of the
authors and do not reflect the position of the CSBS. The views expressed herein are those of the
authors and do not necessarily reflect the views of the National Bureau of Economic Research.
NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.
© 2022 by W. Scott Frame, Ruidi Huang, Erik J. Mayer, and Adi Sunderam. All rights reserved.
Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission
provided that full credit, including © notice, is given to the source.
The Impact of Minority Representation at Mortgage Lenders
W. Scott Frame, Ruidi Huang, Erik J. Mayer, and Adi Sunderam
NBER Working Paper No. 30125
June 2022
JEL No. G21,G51,J15
ABSTRACT
We study links between the labor market for loan officers and access to mortgage credit. Using
novel data matching the (near) universe of mortgage applications to loan officers, we find that
minorities are significantly underrepresented among loan officers. Minority borrowers are less
likely to complete mortgage applications, have completed applications approved, and to
ultimately take-up a loan. These disparities are significantly reduced when minority borrowers
work with minority loan officers. Minority borrowers working with minority loan officers also
have lower default rates. Our results suggest that minority underrepresentation among loan
officers has adverse effects on minority borrowers’ access to credit.
W. Scott Frame
Federal Reserve Bank of Dallas
Ruidi Huang
Southern Methodist University
6212 Bishop Blvd
Dallas, TX 75275
Erik J. Mayer
Southern Methodist University
6212 Bishop Blvd
Dallas, TX 75275
Adi Sunderam
Harvard Business School
Baker Library 359
Soldiers Field
Boston, MA 02163
and NBER
1
Better understanding the sources of racial disparities in economic outcomes is a
focus of research and policy (Lang and Spitzer 2020; Small and Pager 2020). An emerging
literature finds that racial disparities in the labor market may beget disparities in the
provision of critical services. For instance, minority students tend to achieve better
outcomes when taught by minority teachers (Dee 2004, 2005; Fairlie, Hoffmann, and
Oreopoulos 2014) and minority patients achieve better outcomes when treated by minority
doctors (Alsan, Garrick, and Graziani 2019), yet minorities are underrepresented among
teachers and doctors. Much of the work in this literature studies contexts with significant
regulatory and competitive frictions, such as education and health care. A key question is
whether the underrepresentation of minorities in the labor force and its adverse effect on
minority consumers persist even in settings with fewer frictions and greater competition.
In this paper, we study the relationship between racial disparities in the labor market
for loan officers and economic outcomes in the highly competitive market for originating
U.S. home mortgages. We find that working with minority loan officers improves credit
access for minority mortgage applicants along several dimensions. These results suggest
that underrepresentation of minorities among loan officers may have an adverse effect on
minority access to credit in the aggregate. Home equity is an important source of household
wealth and exhibits large racial disparities (e.g., Bhutta, Chang, Dettling, Hsu, and Hewitt
2020; Gupta, Hansman, and Mabille 2021; Kermani and Wong 2021). Given the
importance of access to credit for home purchases, frictions between borrowers and lenders
related to race and ethnicity may contribute to these wealth disparities.
Our analysis leverages a novel dataset linking the confidential version of the
recently expanded Home Mortgage Disclosure Act (HMDA) data to information on loan
officers from the Nationwide Mortgage Licensing System (NMLS). We start our analysis
by providing a description of the racial/ethnic composition of mortgage loan officers. As
with other white-collar professions, we find that minorities are underrepresented. In 2019,
minorities accounted for 39% of the U.S. labor force, but only an estimated 15% of
mortgage loan officers.
2
We next explore whether the race/ethnicity of loan officers and applicants matters
for mortgage application completion, approval, and origination. We find that minority
applications are about 2 percentage points less likely to be completed than white
applications handled by the same white loan officer. For minority loan officers, however,
the gap in completion rates between minority and white applicants is 1-2 percentage points
smaller. We find similar effects for loan approval conditional on application completion.
We focus on “high-discretion” loan applications, those for which automated underwriting
systems do not make definitive recommendations. For these applications, we find that
minority applicants are about 3 percentage points less likely to be approved than white
applicants working with the same white loan officer. For minority officers however, the
difference in approval rates between minority and white applicants is 1-2 percentage points
smaller.
The effects we find on overall loan origination rates are economically meaningful.
Loan applications from minority borrowers are about 5 percentage points less likely to
result in an origination than applications from white borrowers handled by the same white
loan officer, but this difference is 2-4 percentage points smaller for minority officers. In
other words, minority loan officers close much of the gap in credit access between white
and minority borrowers.
We then examine whether the expanded credit access that results when minority
officers handle minority applications leads to higher default rates using loan-level data
from the Federal Housing Administration (FHA). Conditional on standard underwriting
risk factors, minority borrowers are 1.7 percentage points more likely to default than white
borrowers working with the same white loan officer. Despite the fact that minority
applications handled by minority officers are more likely to result in loan originations, the
same loans are less likely to default. Indeed, the full difference in default rates between
minority and white borrowers (i.e., the full 1.7 percentage points) disappears when the loan
officer is a minority. Our evidence on defaults cuts against simple taste-based
discrimination explanations. Under taste-based discrimination, we would expect loan
3
applications from minority borrowers handled by white loan officers to have both lower
approval rates (as we find) and lower default rates (contrary to what we find).
To further explore the mechanism behind our results, we examine the cross section
of borrowers, financial institutions, and loan officers. Along each dimension, we find
evidence consistent with the idea that minority loan officers are better able to produce soft
information about minority borrowers. In the cross section of borrowers, our baseline
results—that minority applicants are more likely to be approved and less likely to default
when working with minority loan officers—are amplified when we look at matches
between minority borrowers and loan officers of the same race/ethnicity, for which there
is most plausibly a soft information advantage. Similarly, we find that our baseline results
are strongest for low-income borrowers, for whom traditional hard information is less
predictive of defaults and soft information is likely important (Blattner and Nelson 2021;
Di Maggio, Ratnadiwakara, and Carmichael 2021; Mayer 2021). In the cross section of
financial institutions, our baseline results are strongest for small banks, which may be more
likely to exploit soft information in lending decisions (e.g., Stein 2002; Berger, Miller,
Petersen, Rajan, and Stein 2005; Liberti and Mian 2009). Finally, in the cross section of
loan officers, we find that our baseline results are stronger for minority loan officers who
handle many minority loan applications and weaker for experienced white officers. In
addition, when minority officers work with minority borrowers, application outcomes are
less correlated with hard information on loan applications like credit scores and debt-to-
income ratios.
Taken together, these results suggest that certain loan officers may receive and act
upon more precise soft information signals about minority borrowers. A soft information
mechanism is surprising because mortgage lending in the U.S. is a setting where hard
information dominates. Most mortgage applications are evaluated with the aid of
automated underwriting systems that use hard information that has been verified with
documentation (e.g., credit reports, property appraisals, and tax returns). In light of these
institutional features, a natural interpretation of our results is that in some cases loan
officers’ soft information is a key determinant of the hard information that ends up on loan
4
applications. For instance, minority loan officers may be better informed when deciding
whether to exert effort to document alternative sources of borrower income or assemble
more compelling documentation for minority applications. Thus, by influencing the quality
of hard information on loan applications, loan officers’ soft information may influence the
loans that get originated and how they ultimately perform.
Throughout our analysis, we take two separate approaches to isolate supply-side
effects running from loan officers to mortgage applicants. First, we saturate our regressions
with fixed effects. In our most stringent specifications, we include branch-loan officer-year
fixed effects, as well as fixed effects controlling for narrow ranges of credit scores (FICO),
loan-to-value ratios (LTV), and debt-to-income ratios (DTI). These specifications
essentially compare two borrowers, one a minority and one white, with observationally
equivalent mortgage applications facing the same loan officer. Our regressions then ask
whether differences in outcomes between these borrowers vary with the race/ethnicity of
the loan officer.
Despite the tight controls in these OLS specifications, one might still worry that our
results are driven in part by endogenous matching on unobservable characteristics between
loan officers and borrowers. To address this concern, we instrument for whether an
application is handled by a minority loan officer with the share of applications at the same
branch handled by minority officers on the same day of the week in previous weeks. So
long as application quality is unrelated to loan officer work schedules, this instrument
captures exogenous variation in the probability that an application is handled by a minority
officer. We find very similar results using this identification approach. If anything, our
instrumental variables results are somewhat stronger than our OLS results, consistent with
the idea that unobservably riskier minority borrowers endogenously match with minority
loan officers.
We contribute to several strands of the vast literature on racial disparities in
economics.
1
First, we contribute to recent work showing that racial disparities in labor
1
Early foundational work includes Becker (1957), Phelps (1972), and Arrow (1972, 1973). A large literature
documents racial disparities in a variety of economic outcomes such as wealth, wages, and housing returns
5
markets can create inequalities in downstream customers’ outcomes by using
comprehensive and novel data to show that minorities’ underrepresentation in skilled labor
positions (loan officers) gives rise to racial disparities in the U.S. residential mortgage
market. Our findings highlight the importance of minority representation at firms even in
a competitive market setting with low search costs, where relationships play a relatively
small role and the logic of Becker (1957) suggests disparities should be competed away.
Second, racial disparities in the mortgage market have received considerable
attention since the seminal work of Munnell, Tootell, Browne, and McEneaney (1996).
Many studies find that minorities face lower approval rates, and if approved, pay higher
interest rates. This strand of the literature includes recent work such as Bhutta, Hizmo, and
Ringo (2021) and Bartlett, Morse, Stanton, and Wallace (2022), which use detailed data on
borrower credit scores, loan-to-value ratios, and debt-to-income ratios that have
historically not been available.
2
We contribute to this literature by showing that minority
loan officers appear to have an advantage in terms of soft information on minority loan
applicants, which allows them to achieve higher approval rates and lower default rates. In
this way, our paper relates to studies examining the effects of shared group status between
lenders and borrowers. Much of this work, including Fisman, Paravisini, and Vig (2017),
Fisman, Sarkar, Skrastins, and Vig (2020), and D’Acunto, Ghosh, Jain, and Rossi (2021),
studies the effects of shared religion or caste between loan officers and borrowers in India.
3
In the U.S., Ambrose, Conklin, and Lopez (2021) study brokered loans originated by New
Century Financial Corporation from 2003-2007 and show that brokers charge borrowers
higher fees when they are from different races/ethnicities. We contribute to this literature
(e.g., Charles and Guryan 2008; Bayer and Charles 2018; Kuhn, Schularick, and Steins 2020; Kermani and
Wong 2021).
2
Additional evidence from correspondence experiments (Hanson, Hawley, Martin, and Liu 2016) and audit
studies (Ross, Turner, Godfrey, and Smith 2008) often supports a discrimination explanation for racial
disparities in mortgage lending. However, some empirical studies also find that minority borrowers default
more (e.g., Berkovec, Canner, Gabriel, and Hannan 1998), which may be inconsistent with taste-based
discrimination.
3
A much larger literature shows that loan officers and brokers affect loan outcomes. See, e.g., Tzioumis and
Gee (2013), Drexler and Schoar (2014), Berg (2015), Cole, Kanz, and Klapper (2015), Agarwal and Ben-
David (2018), Bushman, Gao, Margin, and Pacelli (2021), and Dobbie, Liberman, Paravisini, and Pathania
(2021), among others.
6
by documenting in-group advantages based on race and ethnicity in the U.S. mortgage
market and showing that soft information remains important even in a setting with large
amounts of hard information and relatively little scope for relationship building.
Perhaps the most closely related paper to ours is Jiang, Lee, and Liu (2022), which
was developed independently and contemporaneously to our paper. They also link
confidential HMDA data to the NMLS loan officer data and find that minority loan officers
are more likely to approve minority mortgage applications. They focus on the ability of
FinTech lenders and machine learning to reduce differences in outcomes across
races/ethnicities, while we use differences within the cross sections of white and minority
workers to help pin down the economic mechanisms at work.
We argue that the joint patterns of application completion rates, approval rates, and
loan default rates, as well as the ways in which loan officers react to hard information on
applications, are all consistent with the idea that minority loan officers have a soft
information advantage in handling minority applications. Finally, our back-of-the-
envelope calculations suggest that improving minority representation among loan officers
could close nearly half of the gap in access to mortgage credit between white and minority
borowers.
2. Data and Methodology
We leverage three novel datasets to conduct our empirical analysis. First, we build
the first nationwide panel of mortgage loan officers based on licensing and registration
information from the Nationwide Mortgage Licensing System from 2012 to 2019. Second,
we connect loan officers to their mortgage applications handled in 2018 and 2019 using
the newly expanded confidential version of the Home Mortgage Disclosure Act data
maintained by the Federal Reserve System. Third, we connect loan officers to their FHA-
insured mortgage originations from 2012 to 2018 using comprehensive data provided by
the Federal Housing Administration. We supplement these datasets with information on
ZIP code-level demographic and economic characteristics from the U.S. Census Bureau.
Appendix A provides detailed definitions of all the variables used in the analysis.
7
2.1 Nationwide Mortgage Licensing System Data
The Secure and Fair Enforcement for Mortgage Licensing Act of 2008 (SAFE Act)
was designed to enhance consumer protection and reduce fraud in the mortgage market.
4
The SAFE Act requires all residential mortgage loan originators (i.e., loan officers) to be
either state licensed or federally registered. Loan officers employed by federally insured
depository institutions or their subsidiaries must be federally registered. All other loan
officers working at nonbank mortgage companies must be state licensed. Importantly for
our study, the SAFE Act requires that all loan officer licenses and registrations must be
recorded in the Nationwide Mortgage Licensing System (NMLS).
5
By 2012, all state and
federal regulators had integrated their licensing/registration with the NMLS, making it a
comprehensive registry of mortgage lenders and their loan officers.
We obtain access to the data from NMLS Consumer Access
SM
through an
agreement with the State Regulatory Registry, a wholly-owned subsidiary of the
Conference of State Bank Supervisors.
6
The data contain historical snapshots of files with
information on licenses, registrations, and other filings for loan officers, as of the end of
each calendar year from 2012 to 2019. Importantly, the NMLS assigns each loan officer a
unique NMLS ID that stays with them over time and across employment spells, allowing
us to accurately track officers throughout their career in the mortgage industry. Thus, the
data allow us to construct a nationwide loan officer-year panel which contains their name,
NMLS ID, current employer, work address, and employment history in the industry.
2.2 Identifying Loan Officer Race/Ethnicity
We do not directly observe the race and ethnicity of loan officers in the NMLS data.
Instead, we infer this information using the Bayesian Improved First Name Surname
Geocoding (BIFSG) method developed by Voicu (2018) and adopted by Ambrose,
4
See https://www.hud.gov/sites/documents/DOC_19673.PDF.
5
The NMLS was created in 2008 by the Conference of State Bank Supervisors (CSBS) and the American
Association of Residential Mortgage Regulators (AARMR), see https://nationwidelicensingsystem.org.
6
For additional information on NMLS Consumer Access
SM
, see https://nmlsconsumeraccess.org/.
8
Conklin, and Lopez (2021).
7
The BIFSG method utilizes each individual’s first name, last
name, and physical location to calculate the probability that they belong to a particular
racial/ethnic category (i.e., non-Hispanic white, non-Hispanic black, non-Hispanic
American Indian and Alaskan Native, non-Hispanic Asian and Pacific Islander, Hispanic,
and non-Hispanic other/multiracial).
The BIFSG method calculates the probability of an individual with surname s, first
name f, and ZIP code z belonging to each race group r as:
(1)
where
is the posterior probability of belonging to race group r;
is the
probability of belonging to race group r conditional on surname s;
is the probability
of having first name f conditional on race r; and is the probability of locating in ZIP
code z conditional on race r. In the denominator, we sum over the six race and ethnicity
groups.
8
The NMLS data provide us with the surname, first name, and branch office ZIP
code for each loan officer. We first standardize the loan officer first and last names and
limit ZIP codes to the first five digits. Then, we match loan officers’ last names to the 2010
U.S. Census surname list, which includes surnames found more than 100 times, covers
about 90% of the U.S. population, and lists the probability of a given surname belonging
to one of the six racial/ethnic groups. We match loan officers’ first names to a list created
by Tzioumis (2018), which is derived from mortgage application data and covers over
4,000 first names and provides the probability of each name belonging to one of the six
groups.
9
Lastly, we match branch ZIP codes to the share of each racial/ethnic group in the
ZIP code based on the 2010 U.S. Census.
10
7
The BIFSG methodology is an improvement from the Bayesian Improved Surname Geocoding (BISG)
method developed by Elliott et al. (2009), which is used by government agencies such as the Consumer
Financial Protection Bureau. For additional details, see U.S. Consumer Financial Protection Bureau (2014).
8
The BIFSG methodology assumes p(f|r) = p(f|r,s) and p(z|r) = p(z|r,f,s).
9
See https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/TYJKEZ
10
Throughout the paper, we refer to loan officers’ office addresses as “branches”, although in some cases
these could be single office lending institutions.
9
For each loan officer, we calculate the probability that they belong to each of the
six race and ethnicity groups according to equation (1). We rank the six probabilities from
highest to lowest and assign each loan officer to the specific racial/ethnic group with the
highest probability, following Ambrose, Conklin, and Lopez (2021). Voicu (2018)
conducts validation exercises in voter registration and mortgage application data to show
that this method provides classifications with high accuracy (i.e., low probability of false
positives or false negatives) and low bias (i.e., the distribution of imputed classifications is
similar to the population distribution). We describe loan officer demographics and
characteristics in Section 3 below.
2.3 Confidential HMDA Data
The Home Mortgage Disclosure Act (HMDA) requires nearly all mortgage lenders
to report detailed information on the applications they receive, and whether they originate
the loan. Only very small or exclusively rural lenders are exempt from HMDA reporting,
making it the most comprehensive source of information on mortgage applications with
over 95% coverage (Avery et al. 2017).
11
The public version of the HMDA data includes
borrower income, race, and ethnicity, as well as loan size, loan purpose (purchase,
refinance, home improvement), loan type (conventional, government-insured), loan
priority (first or second lien), the existence of any co-applicants, and property location
(census tract).
Starting in 2018, the confidential version of the HMDA data maintained by the
Federal Reserve System includes a host of new variables, which are not publicly available.
Important additions include key underwriting variables like borrower credit score, loan-to-
value ratio, and debt-to-income ratio, as well as the automated underwriting system (AUS)
used and its recommendation code. Critical to our study, the new confidential HMDA data
11
As of 2019, any depository institution must report to the HMDA database if it has: (i) at least one branch
or office in a metropolitan statistical area (MSA), (ii) at least $46 million in assets, and (iii) originated at least
25 mortgages in each of the previous two years. Non-depository institutions must report data if they have a
branch/office in an MSA (or receive at least five applications from MSAs) and originated at least 25
mortgages in each of the previous two years.
10
include the loan officer identifier (NMLS ID), which allows us to link loan officers in our
NMLS panel directly to the applications they handle.
2.4 FHA Data
Since the HMDA dataset does not include information on loan performance, we
use data from the Federal Housing Administration (FHA). The FHA operates the largest
government mortgage insurance program, which facilitates home financing opportunities
for first-time and low- and moderate-income homebuyers by guaranteeing loans with small
down payments (high LTVs) to borrowers with relatively low credit scores. In 2019, FHA
mortgages constituted 18% of all first-lien home purchase mortgage originations.
The FHA has provided the Federal Reserve Banks of Atlanta and Dallas with
comprehensive data on the population of their insured single-family mortgage originations
between 2000 and 2018. These data include mortgage terms and standard underwriting
variables such as credit score, loan-to-value ratio, debt-to-income ratio, etc. The data also
contain information on loan performance through September 2019, including an indicator
for whether the loan ever became 90 or more days past due, which is a standard definition
of mortgage default. Importantly, the data also include the loan officer’s NMLS ID, which
allows us to link loan officers to their FHA originations from 2012 to 2018.
2.5 Merges and Screens
We first merge all HMDA mortgage applications in 2018 and 2019 to information
on the loan officer handling the application from NMLS using their NMLS ID—we find
matches for 99.3% of applications. We then drop the 2.9% of applications where the loan
officer’s branch (i.e., work office) address is missing or is located outside the U.S.
Additionally, we drop the 9.4% of applications where the loan officer’s race/ethnicity
cannot be assigned using the BIFSG algorithm. Therefore, our analysis focuses on the
remaining 87% of applications where we have sufficient data on loan officers’ location and
race/ethnicity.
11
We use HMDA data on all mortgage applications to compute loan officer-year
statistics for 2018-2019, which we tabulate and discuss in the next section. However, for
our main application-level tests, we pare down the sample along a few dimensions to make
it more homogeneous, consistent with prior studies (e.g., Bhutta, Hizmo, and Ringo 2021;
Bartlett, Morse, Stanton, and Wallace 2022). First, we drop any observations where
applicant/borrower race is not reported. Second, we drop loans originated through
mortgage brokers to ensure that the borrower was dealing directly with the loan officer.
Third, we focus only on first-lien, 30-year, fixed-rate, home purchase mortgages financing
owner-occupied single-family properties. Focusing on a single mortgage contract helps
address selection concerns that would arise if we considered multiple contracts, and 30-
year fixed-rate mortgages are the most popular contract in the U.S. Focusing on purchase
mortgages also helps address selection concerns, as refinancing borrowers are more likely
to use the same loan officer as they had for their first mortgage if they had a good
experience with the first mortgage.
We follow a similar matching and filtering process for the FHA data. We find
sufficient information on loan officer location and race/ethnicity for 89.9% of FHA home
purchase mortgages from 2012 to 2018. We report summary statistics for these main
samples in subsequent sections, immediately preceding the related analyses.
3. Minority Loan Officers and Access to Mortgage Credit
This section studies the influence of minority loan officers on borrower access to
mortgage credit. We document two main results. First, minorities are underrepresented
among the ranks of loan officers. Second, applications by minority borrowers that are
handled by minority loan officers are more likely to be completed, approved, and result in
an originated mortgage. To deal with the fact that there may be endogenous matching
between loan officers and borrowers, we show that similar results obtain using an
instrumental variables (IV) approach that aims to approximate random assignment of loan
officers to borrowers.
12
3.1 Minority Representation among Mortgage Loan Officers
Table 1 presents summary statistics for the loan officers in our HMDA/NMLS
matched panel. Panel A shows a 2019 snapshot of minority representation among mortgage
loan officers. Our sample in 2019 is comprised of around 255,000 active loan officers.
12
Among these individuals, we identify 84.6% as white, 8.9% as Hispanic, 1.8% as Black,
and 4.7% as Asian. By contrast, for the U.S. population that year, the shares were white
60.7%, Hispanic 18.0%, Black 13.2%, and Asian 6.2%. If we compare the minority share
of U.S. mortgage loan officers to jobs that require similar skills, we find a persistent
minority representation gap. For example, according to the Current Population Survey
maintained by the Bureau of Labor Statistics (BLS), in 2019 non-whites accounted for
29.3% of financial managers, 28.0% of credit counselors, and 21.8% of personal financial
advisors, whereas they represent only 15.4% of the loan officers in our sample.
13
Panel B of Table 1 provides descriptive statistics on loan officer, lender, and branch
ZIP code characteristics overall and by race/ethnicity using the loan officer data for 2018
and 2019. The average level of experience for mortgage loan officers is 9.1 years. For white
officers, average experience is 9.3 years, while the figure is 7.7 years for Hispanic officers,
8.0 years for Black officers, and 8.3 years for Asian officers. We see similar patterns for
turnover and tenure at current employer. White loan officers average 41 applications per
year, while Hispanic, Black, and Asian loan officers all handle around 32 applications per
year on average. In dollar terms, the application patterns are somewhat different. Asian
loan officers handle loan applications totaling $11 million on average each year, while
Black loan officers handle $6 million. The average dollar volume of mortgages handled
by a loan officer across all races is around $9 million.
We construct Lender Minority LO %
i,f,c,t
as the fraction of minority loan officers at
the mortgage originator f that officer i works at in year t, excluding the county c where
officer i works, to measure minority representation at the institution level. The average
12
Some lenders also register other (non-loan officer) employees in NMLS. Therefore, we focus on the
matched 2018-2019 data, because we can directly confirm that the person is an active loan officer by requiring
them to handle at least one mortgage application at a HMDA-reporting institution that year.
13
For a list of professions, see https://www.bls.gov/cps/aa2019/cpsaat11.pdf.
13
value of the variable is 15%, matching the overall composition of our loan officer sample.
When we focus on minority loan officers, the variable takes on values above 15%,
indicating that there is heterogeneity in minority representation across lenders: When a
lender has more minority loan officers outside county c, it has more minority loan officers
in county c as well. In other words, Hispanic, Black, and Asian loan officers all tend to
work at lenders with more minority loan officers in other locations.
[Insert Table 1 Here]
Figure 1 illustrates our first main result. We plot the share of minority loan officers
in a ZIP code against the minority population share in that ZIP code for deciles of the
minority population share distribution. If the share of minority loan officers simply reflects
local population demographics, we would expect a 45-degree line with a y-intercept at zero
(depicted in the plot for reference). However, the figure shows that the intercept of the
best-fit line is negative, and the slope is less than one. In other words, mortgage lenders
employ fewer minorities than the ZIP code level minority share, and minorities are most
underrepresented among loan officers in the ZIP codes with the highest minority share.
Taking the far-right bin as an example, for ZIP codes in the top decile of minority share
(around 80% minority residents), only around 55% of loan officers are minorities. Internet
Appendix Table IA1 provides regression evidence corresponding to Figure 1 and shows
that similar patterns persist when we control for region, population density, average income
in the ZIP code, and average income among minorities in the ZIP code.
[Insert Figure 1 Here]
3.2 Minority Loan Officers and Mortgage Application Outcomes
This subsection shows our second main result—that mortgage loan applications
opened by minorities are more likely to be completed, approved, and ultimately originated
if they are handled by a minority loan officer. These tests use the sample of home purchase
14
mortgage applications in HMDA from 2018-2019, as described in Section 2.5. Table 2
reports summary statistics for this sample of approximately 5.65 million applications as
well as statistics splitting by both the applicant and loan officer race/ethnicity. About 10%
of opened loan applications are handled by minority loan officers, while minority
borrowers account for about 30% of all opened applications. The average applicant is 41
years old with an income of $94,000, a credit score of 725, and is requesting a loan for just
under $260,000. Minority applicants tend to have slightly lower incomes and credit scores
and slightly higher LTVs. The raw statistics also show large minority gaps in mortgage
application completions, approvals, and originations under white officers, and smaller gaps
under minority officers, which we will now examine more carefully.
[Insert Table 2 Here]
3.2.1 Completion Rates
Table 3 presents application-level linear probability models of the form:
(2)
The main independent variables of interest are indicators for the applicant filing application
i being a minority (Minority
i
), and for the loan officer being a minority (Minority Officer
i
),
as well as their interaction. We two-way cluster standard errors by lender and county.
In the first two columns of Table 3, the dependent variable is a dummy variable
indicating that an opened application i was completed (Completed
i
). The control variables
(X
i
)
are loan type indicators, ten-year bins of applicant age, centile bins of the applicant
income-to-MSA median income ratio,
14
log(loan amount), an indicator for jumbo loans,
and an indicator for joint applications. We choose these controls because they are available
14
We use the income-to-MSA median income ratio because there are cost of living differences across MSAs
and certain government program incentives (e.g., Community Reinvestment Act requirements) are a function
of MSA median income.
15
for all applications, including those that were not completed, and we refer to them in the
tables as Basic App Controls. Appendix A provides variable definitions.
Table 3 column 1 reports results conditioning on these controls as well as branch-
year and property county fixed effects. Thus, the specification exploits comparisons
between two borrowers applying at the same branch in the same year, financing properties
in the same county, with similar loan application characteristics. The coefficient β
1
on the
minority applicant dummy is negative and statistically significant, indicating that minority
applications handled by white loan officers are 2.1 percentage points less likely to be
completed than white applications handled by white loan officers. (For reference, the
baseline application completion rate is 83.7%.) The coefficient β
2
on the minority loan
officer dummy is negative but small in magnitude (-0.2 percentage points), indicating that
white applications handled by minority officers are slightly less likely to be completed than
white applications handled by white officers. The coefficient β
3
on the interaction of
minority applicant and minority loan officer is positive and statistically significant,
indicating that minority applications handled by minority loan officers are 1.5 percentage
points more likely to be completed than minority applications handled by white loan
officers.
In column 2, we replace our branch-year fixed effects with branch-year-officer
fixed effects. Thus, we can no longer identify the coefficient β
2
on the minority loan officer
dummy, but we are now exploiting variation within a loan officer. In other words, we are
taking the difference in how the same loan officer handles minority applicants relative to
white applicants and asking how that difference varies with the officer’s race. The results
are similar to those in column 1. The coefficient β
1
on the minority applicant dummy is
negative and significant, indicating that averaging across white loan officers, minority
applications are 1.9 percentage points less likely to be completed than white applications
handled by the same officer. The coefficient β
3
on the interaction of minority applicant and
minority loan officer is positive and statistically significant. When the loan officer is a
minority, this difference in completion rates between minority and white applicants is 1.1
percentage points smaller.
16
[Insert Table 3 Here]
3.2.2 Approval Rates
Next, we restrict the sample to completed mortgage applications and examine
whether minority applicants are more likely to be approved when working with a minority
loan officer. Figure 2 shows the raw data on approval rates based on whether the applicant
and loan officer are racial/ethnic minorities. The left panel plots the average approval rate
for applications handled by white loan officers by decile of borrower credit score, splitting
white and minority applicants. Minority applications handled by white loan officers have
lower approval rates across the entire credit score distribution. The right panel plots
applications handled by minority loan officers, again splitting by whether applicants are
minorities or not. In stark contrast to the left panel, the approval rates for white and minority
applicants working with minority loan officers are very similar when the credit score is
above 680. At low credit scores, minority applicants have slightly higher approval rates
than white applicants.
[Insert Figure 2 Here]
Columns 3-6 of Table 3 present formal regression evidence. Within the set of
completed mortgage applications, we run a regression in line with Eq. (2) but now the
dependent variable is a dummy indicating that application i was approved (Approved
i
). The
main independent variables of interest are again indicators that the applicant is a minority
(Minority
i
), the loan officer is a minority (Minority Officer
i
), and their interaction. The
basic application controls are again: loan type indicators, bins for applicant age and income,
log(loan amount), and indicators for jumbo loans and joint applications. Since completed
loan applications contain more information in the confidential HMDA file, we expand the
set of control variables to include indicators for narrow bins of FICO, LTV, and DTI, all
17
interacted with the loan type indicators. We refer to these as Extended App Controls. Again,
standard errors are two-way clustered at the lender and county level.
The confidential HMDA data also allow us to observe the automated underwriting
system (AUS) used to evaluate the mortgage application, and the output code it produces.
We control directly for each AUS output code with fixed effects. To further isolate the
importance of loan officers, we split our sample into low versus high-discretion cases based
on the AUS output code. We label an application as a “low discretion” case if the average
approval rate for the AUS code that the application receives is greater than 90%.
Applications receiving AUS codes with less than 90% approval rates are deemed to be
“high discretion.” We select the 90% threshold because there is a sharp break in the
distribution of approval rates (see Internet Appendix Figure IA1), but we obtain similar
results with different thresholds. By our definition, approximately 86% of applications are
low discretion.
Table 3 columns 3 and 4 examine low-discretion applications. The coefficient on
the minority dummy indicates that, even among these applications, completed minority
applications are one percentage point less likely to be approved. The small and statistically
insignificant interaction between the minority applicant and minority loan officer dummies
implies that in low-discretion cases, minority application approvals are unrelated to the
race/ethnicity of the loan officer handling the application.
In column 5 of Table 3, we re-estimate the same specification using the high-
discretion applications. The coefficient on the minority applicant dummy is now more
negative, indicating that high-discretion minority applications handled by white loan
officers are 2.9 percentage points less likely to be approved than high-discretion white
applications. (For reference, the average approval rate for high-discretion applications is
70.3%.) The coefficient on the minority loan officer dummy is small and statistically
insignificant. However, the interaction between the minority applicant and minority loan
officer dummies is now positive and significant, indicating that minority applications
handled by minority loan officers are 1.2 percentage points more likely to be approved than
minority applications handled by white loan officers. In other words, the difference in
18
approval rates between minority and white applications with high discretion is about 40%
smaller when the loan officer is a minority.
In column 6, we replace our branch-year fixed effects with branch-year-loan officer
fixed effects. Thus, we can no longer identify the coefficient on the minority loan officer
dummy, as we are exploiting variation within a loan officer. The results are similar to those
in column 5. Averaging across white loan officers, minority applicants are 2.6 percentage
points less likely to be approved than white applicants handled by the same loan officer.
When the officer in question is a minority, this difference is 1.4 percentage points smaller.
3.2.2 All-in Origination Rates
Columns 7 and 8 of Table 3 combine the two previous results, by studying all-in
probabilities that opened mortgage applications result in loan originations. As in columns
1 and 2, the sample is all loan applications, including those never completed, and hence the
tests use the basic application controls.
Column 7 shows that minority applications handled by white loan officers are 5.0
percentage points less likely to end in a loan origination than white applications. (For
reference, the average origination rate for all applications is 74.9%.) The coefficient on the
minority loan officer dummy is negative, indicating that applications handled by minority
officers are slightly less likely to result in originations. The interaction between the
minority applicant and minority loan officer dummies is positive and significant, indicating
that minority applications handled by minority loan officers are 2.5 percentage points more
likely to result in originations than minority applications handled by white loan officers. In
other words, the difference in origination rates between minority and white applications is
about 50% smaller for minority loan officers than white loan officers. Column 8 shows that
the results are similar when we replace our branch-year fixed effects with branch-year-loan
officer fixed effects.
As we show in Table 7 below, most of the effect of minority applicant-minority
loan officer pairings appears to be driven by the roughly 80% of such pairings that are
same-race/ethnicity pairings. In our main results, we continue estimate the overall effects
19
of minority applicants matching with minority loan officers because we have limited power
to distinguish between all combinations of race/ethnicity pairings between borrowers and
loan officers.
15
3.3 Instrumental Variables
In our previous tests, we rely on tight controls and fixed effects for identification.
However, the results still mix endogenous matching between loan officers and borrowers
within a branch with the causal effect of loan officer race/ethnicity on loan application
outcomes. To better identify the causal effect, we use day-of-the-week as an instrument to
capture loan officer work schedules. Specifically, for application i opened at branch b on
day of the week d in week w, we compute the share of applications opened at the same
branch b on the same day of the week d during the prior 12 weeks (w-12 to w-1) that were
handled by minority loan officers. We use this variable to instrument for 1{Minority
Officer}
i
, the dummy indicating whether application i was handled by a minority officer.
Formally, the instrument is defined as:
The first stage for this instrument relies on the idea that loan officer work schedules
are persistent. If minority loan officers have handled a larger fraction of the applications
on a specific day of the week at a particular branch in previous weeks, then they are more
likely to handle applications on that day of the week at the same branch this week. The
exclusion restriction here is that the instrument’s influence on whether a minority loan
officer handles an application is the only channel through which the instrument affects
application outcomes. This essentially boils down to the restriction that the quality of
applications arriving at a specific branch on a particular day of the week this week is
15
Internet Appendix Tables IA2, IA3, and IA4 show results disaggregating all pairings.
20
unrelated to minority loan officer work schedules at the branch over prior weeks. The first
stage of the instrumental variables regression is:
′
(3)
The dependent variable is a dummy indicating that the loan officer is a minority (Minority
Officer). The independent variable of interest is our instrument
. As in our earlier
tests, we include our basic and extended application controls, as well as property county
fixed effects. We also include branch-week and day-of-the-week fixed effects, so that the
instrument isolates within-week variation beyond consistent daily patterns that should
capture loan officer work schedules and/or rotation policies. We construct instruments for
interactions with the minority officer dummy by interacting the variable of interest with
Z
i,b,d,w
. For instance, when we are interested in the interaction of the minority applicant
dummy and the minority loan officer dummy, we construct the instrument
.
16
Table 4 Panel A reports the first stage regression results. Each column corresponds
to the sample we use to study different outcome variables. For instance, in column 1 the
sample is all opened applications, which we use to study application completion rates.
Across the columns, the instrument is strong, with first stage F-statistics in excess of 15.
The exclusion restriction would be violated if the instrument is correlated with any
unobservable determinants of application outcomes. While the exclusion restriction is
untestable, in Internet Appendix Table IA5 we report covariate balance tests that support
the notion it holds. We find little-to-no correlation between the instrument and observable
borrower characteristics.
17
16
In untabulated results, we find similar results if we simply run separate IV regressions for minority
borrowers and white borrowers. These specifications make clearer that the exclusion restriction needed is
that the instrument is uncorrelated with loan application outcomes except through its effect on loan officer
race/ethnicity, conditional on borrower race/ethnicity.
17
A particularly sharp test of covariate balance is in columns 4 and 5 of Table IA5, where we examine the
output codes from automated underwriting systems. These AUS output codes essentially aggregate
information across all observable borrower characteristics, and we use the average approval rate for a given
21
[Insert Table 4 Here]
In Table 4 Panel B, we revisit the main results from Table 3 using our instrumental
variables strategy.
18
The IV results are similar to our OLS results. Indeed, if anything, they
are slightly stronger, consistent with the idea that higher risk minority applicants tend to
endogenously match with minority loan officers (we provide further evidence of this in
Table 10 below). For instance, when we look at approval rates for high-discretion
applications, our OLS specifications in Table 3 suggest that minority applicants are 1.2
percentage points more likely to be approved when their loan officer is a minority. In Table
4, our corresponding IV estimate is 3.6 percentage points. Similarly, examining all-in
origination rates in Table 3, we find that a minority application is 2.5 percentage points
more likely to result in a loan origination when handled by a minority officer than a white
officer. In Table 4, our corresponding IV estimate is 3.8 percentage points.
Overall, the results presented in this section provide strong evidence of a causal
link between loan officer race/ethnicity and access to credit for minority borrowers. Loan
applications from minority borrowers are significantly more likely to be completed,
approved, and result in a loan origination when they are handled by minority loan officers.
4. Minority Loan Officers and Loan Performance
One possible explanation for our results on application outcomes is that minority
loan officers help riskier minority borrowers access mortgage credit. In this section, we
examine the relationship between loan officer race/ethnicity and loan performance and
show this is not the case. Minority loans handled by minority loan officers default less than
minority loan handled by white officers.
code as a continuous measure of the AUS recommendation. The balance tests show that the instrument is
uncorrelated with the AUS recommendation, suggesting that it is unrelated to any hard information about
borrower creditworthiness. We thank James Vickery for suggesting this test.
18
The sample sizes shrink for the IV tests because we drop the first three months of 2018 to construct the
instrument, and the branch-week fixed effects create some singletons that are dropped in the estimation.
22
The HMDA data from our prior tests do not include information on loan
performance. Fortunately, we have data on the population of FHA mortgages which we
can link to NMLS loan officer information back to 2012. We first confirm in the HMDA
application data that our main results from Table 3 showing minority loan officers’ impact
on credit access hold in the FHA subsample (see Table IA6). We then examine the
performance of FHA-insured mortgages made by white and minority officers from 2012
to 2018. The key result of this section is that loans to minority borrowers by white loan
officers have higher default rates than either observationally similar loans to minority
borrowers by minority officers or observationally similar loans to white borrowers.
Table 5 reports summary statistics for our sample of FHA home purchase
mortgages, as described in Section 2.5. We report statistics for the full sample of
approximately 3.39 million loans as well as splitting by both the borrower and loan officer
race/ethnicity. Compared to the average home purchase mortgage applicant (see Table 2),
FHA borrowers have lower incomes and credit scores, and request smaller loans with
higher LTVs, as the FHA program intends. As is standard in the literature, we define
mortgage default as the loan ever going 90 days delinquent, which occurs for 9.1% of the
loans in our sample. The sample splits show that white loan officers have particularly high
default rates with minority borrowers (11.7%) compared to their white borrowers (8.0%),
or to minority loan officers’ white or minority borrowers (8.4% and 9.2%, respectively).
[Insert Table 5 Here]
Before turning to formal regression evidence, Figure 3 presents the raw data. The
left panel plots average default rates for FHA mortgages handled by white loan officers by
decile of borrower credit score, splitting white and minority borrowers. White officers’
loans to minority borrowers have higher default rates, particularly for borrowers with low
credit scores. In other words, despite our results in Section 3—that minority applications
handled by white loan officers have lower completion and approval rates—here we see that
the minority loans that white officers do approve have higher default rates. At the lowest
23
credit scores, the difference is economically large, with approximately 19% of minority
borrowers defaulting, compared to 15% of white borrowers.
[Insert Figure 3 Here]
The right panel of Figure 3 shows the analogous plot for mortgages handled by
minority loan officers. Here there is little discrepancy between default rates for minority
and white borrowers. Furthermore, the default rates for both minority and white borrowers
handled by minority loan officers are similar to those for white borrowers handled by white
loan officers. Of the four sets of loans, loans to minorities handled by white loan officers
stand out as having higher default rates.
Table 6 provides regression evidence. We run linear probability models similar to
our prior tests, except the dependent variable is a dummy indicating that the loan defaulted.
The FHA Controls include log(loan amount), centile bins of the borrower income-to-MSA
median income ratio, indicators for narrow bins of FICO, LTV, and DTI, the interest rate
on the loan, and an indicator for whether the borrower is a first-time home buyer. In
addition, we include fixed effects for branch-year, property county, and month-of-
origination.
19
Columns 1 and 2 report OLS results. Column 1 shows that minority borrowers
whose applications are handled by white loan officers have a default rate that is 1.8
percentage points higher than white borrowers whose applications are handled by white
loan officers. However, the coefficient on the interaction of minority borrower and
minority loan officer is -2.2 percentage points, indicating that minority borrowers handled
by minority officers do not default more than white borrowers. In column 2, we add loan
officer-year fixed effects and find similar results.
We next turn to IV results, using the day-of-the-week instrument defined in Section
3.3 above. The instrument is the same as in Table 4, except it is calculated within our FHA
19
These month-of-origination fixed effects run from January 2012 to December 2018 and absorb any
variation in default rates due to the length of time over which we measure default. We also find similar results
if we measure default over certain time horizons, e.g., within one or two years after origination.
24
sample. Because FHA loans comprise only about 20% of the mortgage market, this
instrument is somewhat noisier in this analysis than in the full sample of loans in Table 4.
Therefore, we run the IV regression including branch-month fixed effects rather than
branch-week fixed effects. In other words, we exploit variation in the loan officer assigned
to a borrower driven by work schedules within the same month rather than the same week.
Column 3 of Table 6 replicates our OLS results from column 1 using the IV sample,
and column 4 shows the corresponding IV results.
20
If anything, the IV results are stronger
than the OLS results. Minority borrowers whose applications are handled by white loan
officers have a default rate that is 1.7 percentage points higher than other borrowers whose
applications are handled by white loan officers. However, the coefficient on the interaction
of minority borrower and minority loan officer is -5.2 percentage points, indicating that
minority borrowers handled by minority officers are not more likely to default than white
borrowers handled by white officers.
[Insert Table 6 Here]
These loan performance results are not consistent with simple explanations
centered on taste-based discrimination, which posit that white loan officers would apply
stricter standards to minority loan applications. Taste-based discrimination would predict
that loan applications from minority borrowers handled by white loan officers should have
lower completion, approval, and origination rates, as we see in the data. However, it also
predicts that loans to minority borrowers handled by white loan officers should have lower
default rates. That is not what we see in Table 6.
5. Exploring the Mechanism
20
The IV sample is smaller because we drop the first three months of 2012 and require FHA lending to occur
at the branch on the same day of the week during the prior 12 weeks in order to construct the instrument. The
branch-month fixed effects also create some singletons which are dropped in the estimation. Internet
Appendix Table IA7 reports the IV first stage regression and balance tests.
25
In this section, we further explore the economic mechanism driving our results. We
present three main results. First, the impact of working with a minority loan officer on
minority application and loan outcomes is larger for same race/ethnicity pairings, for low-
income borrowers, and at small banks, whereas it is weaker at FinTech lenders. Second,
we find particularly strong effects for two subgroups of loan officers: minority loan officers
who appear to specialize in loans to minority borrowers and white loan officers with high
industry experience. Third, these two subgroups of loan officers respond similarly to hard
information: applications with low credit scores and high DTIs handled by these officers
are more likely to be approved and less likely to default, even though low credit scores and
high DTIs are associated with higher default risk in the data. Taken together, these results
suggest that certain loan officers are better able to produce soft information about minority
applicants, allowing them to simultaneously achieve higher approval rates and lower
default rates.
In interpreting our results, some institutional context is helpful, as mortgage lending
is somewhat different from other types of lending like small business lending. At many
financial institutions, particularly the largest ones, the role of the loan officer is to help the
borrower complete the application and provide all relevant documentation, such as bank
statements and pay stubs. At these institutions, loan officers do not make credit approval
decisions, which are made by a separate group of underwriters using application
information and property appraisals. In other words, in most cases application decisions
are largely driven by the hard information captured by a completed application.
This institutional background suggests the following interpretation of our results.
Loan officers can expend effort to help borrowers complete their applications with more
compelling documentation, and minority loan officers are particularly helpful for minority
borrowers. In deciding which applications to exert effort on, loan officers use their own
personal assessments of the borrowers’ creditworthiness. These assessments then show up
in default rates if the loan is originated. In other words, even when loan officers are not
directly making mortgage approval decisions, their soft information is still important in
26
driving the hard information that ends up on loan applications and in turn the loans that get
originated and how they ultimately perform.
21
5.1 Heterogeneity Across Borrowers and Lenders
We start by examining how the impact of loan officer and borrower race/ethnicity
on mortgage approval and default rates varies in the cross section.
22
Table 7 studies the cross section of borrowers and lenders. In columns 1-3, we study
the probability that completed FHA home purchase mortgage applications in our 2018-
2019 HMDA sample are approved. In column 1, we split the interaction of minority loan
officers and minority applicants by whether the pair share the same race/ethnicity.
23
We
find that such matching is associated with a positive effect on loan approval, while the
effect is statistically insignificant when a minority borrower matches with a minority loan
officer of a different race/ethnicity. The importance of same race/ethnicity matches is
consistent with the soft information channel described above. Loan officers may be
particularly adept at helping mortgage borrowers of the same race/ethnicity complete
applications in a compelling manner. These results also suggest that the effect of loan
officer-borrower matches is driven more by homophily (i.e., preferences for shared
backgrounds or demographics) than a taste for diversity (i.e., preference for any
underrepresented group), an important distinction highlighted by D’Acunto, Fuster, and
Weber (2021).
In column 2, we focus on low-income borrowers, defined as those in the bottom
third of the ratio of applicant income to MSA median income. (Note that the controls
21
Bartos et al. (2016) use correspondence studies to examine the idea that disparities can arise when decision
makers must exert effort to acquire information. Our results can be interpreted as showing that minority
officers either have a pre-existing informational advantage when working with minority borrowers or face
lower costs of acquiring information.
22
Ideally, we would study approval decisions and subsequent defaults for the exact same set of loans. This
is not possible due to the data limitations that (i) HMDA only began tracking loan officer identifiers and
detailed borrower information (e.g., credit scores) starting in 2018, and (ii) it takes several years after loan
origination to accurately measure default rates. Therefore, we focus on FHA loans from 2012 to 2018, and
FHA applications in HMDA in 2018-2019. In Internet Appendix Table IA8, we verify that the patterns we
document for FHA loan approval hold in our broader HMDA sample as well (i.e., across all loan types).
23
Structuring the regression this way allows us to separate same race/ethnicity effects within minorities from
such effects between white borrowers and white loan officers.
27
include centile bins of the applicant income to MSA median income ratio, so this
coefficient captures the comparison between how white loan officers treat low-income
minorities relative to other low-income borrowers.) The interaction of the indicators for
minority applicant and low income shows that low-income minority applicants are 0.5
percentage points less likely to be approved than other minority applicants when handled
by white loan officers. The triple interaction of the indicators for minority applicant,
minority loan officer, and low income, however, shows that low-income minority
applicants are 1.7 percentage points more likely to be approved for an FHA mortgage when
handled by minority loan officers.
In column 3, we compare small commercial banks (i.e., less than $10 billion in total
assets) and FinTech lenders (as defined by Buchak, Matvos, Piskorski, and Seru, 2018) to
all other mortgage originators. The triple interaction of the indicators for minority
applicant, minority loan officer, and small bank implies that minority applications handled
by minority loan officers are more likely to be approved at small banks than at other
financial institutions. In contrast, the triple interaction with the FinTech lender indicator
implies that minority loan applications handled by minority loan officers are less likely to
be approved at FinTech lenders than at other financial institutions, consistent with the
contemporaneous findings of Jiang, Lee, and Liu (2022). Note that the double interactions
of either minority borrower or minority loan officer with finance institution type are
insignificant, suggesting that the effect of institution type is operating through the loan
officer-borrower match.
Columns 4-6 provide the corresponding results for defaults in our sample of FHA
loans originated between 2012 and 2018. While in column 1 we saw that minority
applications handled by minority loan officers are more likely to be approved when the
borrower and loan officers are of the same race/ethnicity, in column 4 we see that the lower
default rate for minority borrowers when working with minority loan officers (Table 6) is
driven by same-race pairings (2.7 percentage points). Similarly, while in column 2 we saw
that low-income minority applications handled by minority loan officers are more likely to
be approved, in column 5 we see that they are also less likely to default. Finally, while
28
column 3 showed larger effects of minority loan officers on increasing minority approvals
at small banks (and smaller effects at FinTech lenders), column 6 shows similar patterns in
the institutions where minority officers reduce minority default rates.
In other words, our baseline results from Tables 3 and 6—that minority applicants
are more likely to be approved and less likely to default when working with minority loan
officers—are amplified when we look at matches between borrowers and loan officers of
the same race/ethnicity, are amplified for low-income borrowers and at small banks, and
are dampened at FinTech lenders. These patterns are consistent with minority officers
having more precise soft information about minority applicants, and in particular loan
officers having more precise soft information about applicants of the same race/ethnicity.
This soft information is particularly important for low-income borrowers for whom hard
information may be noisier and at small banks where loan officers may be afforded more
discretion. As discussed above, any soft information possessed by minority loan officers
may not directly enter application approval decisions. It may instead be that loan officers
use their soft information in deciding which applicants they should exert effort to help in
the application completion process.
24
[Insert Table 7 Here]
5.2 Heterogeneity Across Loan Officers
In Table 8, we examine how our baseline results vary with loan officer
characteristics. The structure is similar to Table 7, with the dependent variables being
mortgage approvals (columns 1-2) and defaults (columns 3-4).
In column 1, we examine variation across loan officers in the share of minority
applications they handle. The triple interaction of the officer’s share of minority
24
Differences in soft information most naturally explain the differences in approval and default rates we see
between white and minority loan officers. The fact that the level of defaults is higher for white loan officers
handling minority loans suggests that the marginal costs and benefits of loan origination may differ across
loan officers. If the marginal costs and benefits were the same, in many soft information based explanations,
white officers application approval rates would fall to equalize default rates.
29
applications with the minority officer and minority borrower indicators is positive and
statistically significant. In other words, minority loan officers who handle a lot of
applications from minority borrowers are better able to facilitate minority application
approvals. The interaction of the minority borrower indicator with the officer’s share of
minority applications is statistically insignificant, suggesting that increased exposure to
minority applications has no effect on white loan officers.
In column 2, we study the effects of loan officer years of experience. Experienced
loan officers have higher approval rates, particularly for minority applications. The triple
interaction of loan officer experience, minority loan officer, and minority borrower is
statistically insignificant, suggesting that experience affects white and minority loan
officers similarly. There are two possible explanations for experience effects: learning on
the part of loan officers or survivorship. Internet Appendix Table IA9 provides evidence
of survivorship. Loan officers who approve more minority loans in a given year are more
likely to be in our sample the following year. However, the effect exists for both white and
minority loan officers, suggesting that survivorship is not driving the importance of
experience for white loan officers handling minority applications we see in column 2.
Columns 3-4 show the corresponding results for defaults. Column 3 shows that, in
general, loan officers handling larger fractions of minority loan applications experience
higher default rates. However, minority loan officers who handle many minority
applications have significantly lower default rates among their minority borrowers. The
results in column 4 show that loan officer experience is associated with lower minority
default rates, especially for white loan officers.
Overall, the tests in Table 8 show that our baseline results from Tables 3 and 6—
that minority applicants are more likely to be approved and less likely to default when
working with minority loan officers—are amplified when we look at minority officers who
specialize in handling minority applications. In addition, experienced loan officers also
achieve higher approvals and lower defaults with minority borrowers compared to their
less experienced peers.
30
[Insert Table 8 Here]
Figures 4 and 5 summarize these results. Figure 4 plots the unexplained minority
gaps in approvals and defaults against the handling officer’s minority application share, for
white loan officers (left panel) and minority loan officers (right panel). The unexplained
minority gap estimates come from OLS regressions with the full set of controls from Table
8, where the minority borrower indicator is interacted with indicators for each combination
of white/minority officer and the group the officer falls into in terms of minority application
share (<25%, 25-50%, 50-75%, >75%). The left panel does not show a strong relationship
between minority gaps in approvals/defaults and minority application shares for white
officers. In contrast, the right panel shows that minority gaps shrink—approvals rise and
defaults fall—when minority borrowers work with minority officers who appear to
specialize in minority applications.
Figure 5 similarly plots unexplained minority gaps against the handling officer’s
industry experience, split into four groups (1-4 years, 5-8 years, 9-12 years, >12 years).
The left panel shows a clear pattern: mortgage approvals rise, and defaults fall when
minority borrowers work with more experienced white loan officers. In the right panel, we
see a similar, although quantitatively weaker, pattern for minority borrowers working with
minority loan officers.
[Insert Figures 4 and 5 Here]
5.3 Differential Reactions to Hard Information
If not driven by taste-based discrimination, the differences between white and
minority loan officers we document in approval rates and default rates for minority
borrowers could stem either from different reactions to hard information or different soft
information. In Table 9, we examine the possibility of differential reactions to hard
information. For FHA loan applications by minority borrowers, we study the probability
of approval as a function of all available hard information characteristics, including
31
income, credit scores, and DTIs.
25
For compactness and interpretability, the table imposes
a linear functional form on the relationship between approval and these variables, as
opposed to the fine-grained indicator variables we used as nonparametric controls in earlier
tables. However, in Internet Appendix Table IA10 we show that similar conclusions obtain
if we take a nonparametric approach.
Columns 1 and 2 of Table 9 report the results for applications handled by white
officers and minority officers, respectively. Column 3 reports that there are statistically
significant differences between the way white and minority loan officers treat credit scores
and DTIs for minority loan applications. Minority officers penalize low credit scores and
high DTIs less than white officers. The magnitudes are economically meaningful. A 100-
point reduction in credit score reduces the likelihood that an application handled by a white
officer is approved by 4.6 percentage points, while the reduction for minority officers is
only 3.7 percentage points.
26
The remaining columns show that we see similar differences
in reactions to hard information when we compare inexperienced white loan officers to
experienced white officers, and when we compare minority officers who handle relatively
few minority applications to specialists that handle many.
27
Internet Appendix Table IA11
shows that similar results obtain when we directly analyze the reasons for rejecting a loan
application reported in HMDA. Holding fixed a borrower’s credit score and DTI, minority
borrowers with applications handled by white loan officers are more likely to be rejected
with the reported reason being “credit history” or “debt-to-income.” When minority
borrowers’ applications are handled by minority loan officers, this difference shrinks.
These different responses to hard information can explain our results on approval
rates. The fact that minority officers penalize low credit scores less means that they are
more likely to approve low credit score applicants. However, the different responses to
25
We exclude the jumbo loan indicator and loan-to-value ratios, both of which exhibit very little variation in
the FHA sample.
26
These results are consistent with Blattner and Nelson (2021), who find that credit scores are less
informative for minority borrowers. Our results suggest that minority loan officers have soft information that
can help improve access to credit for minority borrowers when hard information is noisy.
27
We classify loan officers as “experienced” if they have worked in the industry for at least 10 years. We
classify minority loan officers as “specialists” if the share of minority applications in their portfolio is over
80% (which is the median share within minority officers).
32
hard information we see cannot explain our results on default rates. If minority officers
were only relying on hard information and penalized low credit scores less, we would
expect the loans they approve to have higher default rates. This is not what we see in the
data. Instead, the data suggest that minority loan officers are using soft information and
therefore relying less on hard information.
[Insert Table 9 Here]
5.4 Other Results
In the Internet Appendix, we provide several additional analyses. First, in Table
IA12 we show that our results are not stronger at financial institutions where a larger
fraction of loan officers are minorities. This suggests that our results are driven by
information available to the loan officers, rather than policy decisions made at the
institution level. Second, Table IA12 also shows that our results do not vary depending on
how competitive the local mortgage market is, as measured by the county-level HHI. In
other words, we find no evidence that competition disciplines the way white loan officers
handle applications from minority borrowers. Third, in Table IA13, we show that
interactions between minority loan officers and minority borrowers do not affect mortgage
interest rates.
6. Matching Minority Borrowers and Loan Officers
The previous sections document that matching minority loan officers with minority
mortgage applicants can significantly increase loan originations and decrease default rates.
In this section, we first assess the extent to which these gains are captured by endogenous
matching in the existing market equilibrium with the existing pool of loan officers. We
show that there is already a significant amount of matching. We then conduct a back-of-
the-envelope calculation to provide a rough estimate of the potential benefits of increasing
minority representation among loan officers.
33
To assess the degree of matching between minority applicants and loan officers, we
estimate application-level linear probability models of the form:
. (4)
Table 10 presents the results. Column 1 reports the raw regression without any
controls. The constant of 4.8% indicates that white applicants have a 4.8% probability of
working with a minority loan officer and thus a 95.2% probability of working with a white
loan officer. The coefficient on the minority applicant indicator is 16.7%, indicating that
minority applicants have a 4.8%+16.7% = 21.5% probability of working with a minority
loan officer. In other words, there is significant matching between minority loan officers
and minority borrowers. Column 2 adds the loan type, branch-year, and county-of-property
fixed effects and shows that much of this matching is driven by the geographic colocation
of minority borrowers and loan officers. Yet, even within a given branch office, minority
applications are 4.6 percentage points more likely to be handled by minority officers. This
correlation is large relative to the baseline probability of working with a minority loan
officer of 9.7%.
The remaining columns of Table 10 examine heterogeneity in matching. Column 3
shows that low-income minority applications are particularly likely to be handled by
minority loan officers, suggesting that lenders may recognize minority officers’
information advantage with certain applicants.
28
Column 4 shows less matching at FinTech
lenders where loan officers are less likely to have personalized interactions with applicants,
reducing the likelihood of collecting soft information.
[Insert Table 10 Here]
28
By themselves, our matching results could also be consistent with taste-based discrimination. It is possible
that certain loan officers choose to avoid certain borrowers. However, given that many of our other results
are inconsistent with taste-based discrimination, an information-based explanation seems more likely. We
also note that the tendency of minority loan officers to work with the lowest income minority applicants
should work against our main OLS results in Table 3 showing better application outcomes, making them
conservative. Our slightly larger IV estimates in Table 4 also support this notion.
34
While Table 10 shows a meaningful amount of matching from the perspective of
mortgage applicants, there is a much larger amount of matching from the perspective of
loan officers because minorities are underrepresented among their ranks. In Table 10, the
probability a minority application is handled by a minority officer is 21.5%. The summary
statistics in Table 2 show that 65.3% of applications handled by minority officers are from
minority borrowers. By Bayes Rule, the difference between these two numbers stems from
the fact that the fraction of applications handled by minority loan officers (9.7%) is
significantly smaller than the fraction of minority loan applications (29.5%).
29
This implies
that increasing the proportion of minority applications handled by minority loan officers
would likely require increasing the representation of minorities among loan officers, rather
than simply reshuffling applications within the existing pool of loan officers.
How much might increasing minority representation among loan officers increase
minority access to credit? To give a rough sense of the magnitudes, suppose minority
representation among loan officers increased enough that minority applicants were as likely
to work with a minority loan officer as white applicants are to work with a white officer.
30
That is, suppose the percentage of minority applications handled by minority loan officers
increased 73.7 percentage points, from 21.5% to 95.2%. For this additional 73.7% of
minority applications, our IV results in Table 4—which (approximately) randomly assign
minority loan officers to applications—suggest a meaningful increase in the probability
that each application ultimately results in a loan origination. Table 4 column 4 shows that
the gap in origination rates for minority applications handled by white loan officers is 5.0
percentage points conditional on our basic controls, while minority loan officers increase
origination rates for minority applicants by 3.1 percentage points (0.038-0.007=0.031 in
Table 4 column 4). In other words, minority loan officers close about 62% of the gap
relative to white applicants under white officers. Applying this 62% minority loan officer
29
Formally, Pr(MinorityOfficer|Minority Applicant) = 21.5% = Pr(Minority Applicant| MinorityOfficer) x
Pr(MinorityOfficer)/Pr(Minority Applicant) = 65.3% x 9.73%/29.5%.
30
The exact increase in the fraction of minority loan officers required to attain this benchmark depends on
the patterns in geographic colocation and matching within lenders.
35
“treatment effect” to 73.7% of minority applicants would then reduce the overall minority
gap in mortgage origination rates by roughly 46%.
31
Equilibrium levels of minority representation among mortgage loan officers are
determined by many labor market factors outside the scope of our study. Yet, two points
are clear from our analysis. First, minority representation matters. It is difficult to substitute
for an adequate supply of informed minority loan officers with either skilled white officers
or through matching minority applicants to minority officers. And second, the economic
importance of minority representation is significant.
7. Conclusion
This paper studies the effect of minority loan officers on minority access to
mortgage credit using novel data linking loan applications to the loan officers who handle
them. We first show that minorities are significantly underrepresented among the ranks of
loan officers. We then document the consequences of this underrepresentation for minority
access to mortgage credit. Minority applications handled by minority loan officers are more
likely to be completed, approved, and result in loan originations than those handled by
white loan officers. Moreover, minority loans handled by minority loan officers are less
likely to default than those handled by white loan officers.
Our results suggest that minority loan officers have an informational advantage in
handling loan applications from minority borrowers. An implication of these findings is
that improving the representation of minorities among loan officers could improve credit
access and credit outcomes for minority borrowers. It is worth noting that the setting we
study, home mortgage lending, is one where hard information predominates. The
underrepresentation of minorities among loan officers could have even larger
consequences for minority borrowers in other contexts like small business lending, where
soft information is more crucial.
31
This calculation (0.62 x .737 = 0.46) is conservative for two reasons. First, it assumes that the full 5% gap
is available to be reduced to begin with (i.e., no minority applicants are already working with minority
officers). Second, it assumes matching of minority borrowers with minority loan officers, rather than
matching with an officer of the same race/ethnicity.
36
References
Agarwal, Sumit, and Itzhak Ben-David, 2018, Loan prospecting and the loss of soft
information, Journal of Financial Economics 129, 608-628.
Alsan, Marcella, Owen Garrick, and Grant Graziani, 2019, Does diversity matter for
health? Experimental evidence from Oakland, American Economic Review 109, 4071-
4111.
Ambrose, Brent W., James N. Conklin, and Luis A. Lopez, 2021, Does borrower and
broker race affect the cost of mortgage credit? Review of Financial Studies 34, 790-
826.
Arrow, Kenneth J., 1972, Models of job discrimination, Racial Discrimination in
Economic Life 83.
Arrow, Kenneth J., 1973, The Theory of Discrimination, in Ashenfelter and Rees, eds.,
Discrimination in Labor Markets. Princeton: Princeton University Press.
Avery, Robert B., Mary F. Bilinski, Brian K. Bucks, Christine Chai, Tim Critchfield, Ian
H. Keith, Ismail E. Mohamed, Forrest W. Pafenberg, Saty Patrabansh, Jay D. Schultz,
and Claudia E. Wood, 2017, A profile of 2013 mortgage borrowers: Statistics from
the national survey of mortgage originations.
Bartlett, Robert, Adair Morse, Richard Stanton, and Nancy Wallace, 2022, Consumer-
lending discrimination in the FinTech era, Journal of Financial Economics 143, 30-
56.
Bartos, Vojtech, Michal Bauer, Julie Chytilova, and Filip Attention Discrimination: Theory
and Field Experiments with Monitoring Information Acquisition, American Economic
Review 106, 1437-75.
Bayer, Patrick, and Kerwin K. Charles, 2018, Divergent paths: A new perspective on
earnings differences between black and white men since 1940, Quarterly Journal of
Economics 133, 1459-1501.
Bayer, Patrick, Fernando Ferreira, and Stephen L. Ross, 2018, What drives racial and
ethnic differences in high-cost mortgages? The role of high-risk lenders, Review of
Financial Studies 31, 175-205.
Becker, Gary S., 1957, The Economics of Discrimination (University of Chicago press).
Berg, Tobias, 2015, Playing the devil's advocate: The causal effect of risk management on
loan quality, Review of Financial Studies 28, 3367-3406.
Berger, Allen N., Nathan H. Miller, Mitchell A. Petersen, Raghuram G. Rajan, and Jeremy
C. Stein, 2005, Does function follow organizational form? Evidence from the lending
practices of large and small banks, Journal of Financial Economics 76, 237-269.
37
Berkovec, James A., Glenn B. Canner, Stuart A. Gabriel, and Timothy H. Hannan, 1998,
Discrimination, competition, and loan performance in FHA mortgage lending, Review
of Economics and Statistics 80, 241-250.
Bhutta, Neil, Andrew C. Chang, Lisa J. Dettling, Joanne W. Hsu, and Julia Hewitt, 2020,
Disparities in wealth by race and ethnicity in the 2019 survey of consumer
finances, Feds Notes, 28-2.
Bhutta, Neil, Aurel Hizmo, and Daniel Ringo, 2021, How much does racial bias affect
mortgage lending? Evidence from human and algorithmic credit decisions. Available
at SSRN 3887663.
Blattner, Laura and Scott Nelson, 2021, How costly is noise? Data and Disparities in
Consumer Credit. Working paper.
Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru, 2018, FinTech,
regulatory arbitrage, and the rise of shadow banks, Journal of Financial Economics,
453-483.
Bushman, Robert, Janet Gao, Xiumin Martin, and Joseph Pacelli, 2021, The influence of
loan officers on loan contract design and performance, Journal of Accounting and
Economics 71, 101384.
Charles, Kerwin K., and Jonathan Guryan, 2008, Prejudice and wages: An empirical
assessment of Becker’s The Economics of Discrimination, Journal of Political
Economy 116, 773-809.
Cole, Shawn, Martin Kanz, and Leora Klapper, 2015, Incentivizing calculated risk‐taking:
Evidence from an experiment with commercial bank loan officers, Journal of
Finance 70, 537-575.
CFPB, 2014, Using publicly available information to proxy for unidentified race and
ethnicity.
D'Acunto, Francesco, Andreas Fuster, and Michael Weber, 2021, Diverse policy
committees can reach underrepresented groups. National Bureau of Economic
Research, working paper No. w29275.
D'Acunto, Francesco, Pulak Ghosh, Rajiv Jain, and Alberto G. Rossi, 2020, How costly
are cultural biases? Available at SSRN 3736117.
Dee., Thomas S., 2004, Teachers, race, and student achievement in a randomized
experiment, Review of Economics and Statistics 86, 195-210.
---, 2005, A teacher like me: Does race, ethnicity, or gender matter? American Economic
Review 95, 158-165.
Di Maggio, Marco, Dimuthu Ratnadiwakara, and Don Carmichael, 2021, Invisible primes:
FinTech lending with alternative data. Available at SSRN 3937438.
Dobbie, Will, Andres Liberman, Daniel Paravisini, and Vikram Pathania, 2021, Measuring
bias in consumer lending, Review of Economic Studies, forthcoming.
38
Drexler, Alejandro, and Antoinette Schoar, 2014, Do relationships matter? Evidence from
loan officer turnover, Management Science 60, 2722-2736.
Elliott, Marc N., Peter A. Morrison, Allen Fremont, Daniel F. McCaffrey, Philip Pantoja,
and Nicole Lurie, 2009, Using the Census Bureau’s surname list to improve estimates
of race/ethnicity and associated disparities, Health Services and Outcomes Research
Methodology 9, 69-83.
Fairlie, Robert W., Florian Hoffmann, and Philip Oreopoulos, 2014, A community college
instructor like me: Race and ethnicity interactions in the classroom, American
Economic Review 104, 2567-91.
Fisman, Raymond, Daniel Paravisini, and Vikrant Vig, 2017, Cultural proximity and loan
outcomes, American Economic Review 107, 457-92.
Fisman, Raymond, Arkodipta Sarkar, Janis Skrastins, and Vikrant Vig, 2020, Experience
of communal conflicts and intergroup lending, Journal of Political Economy 128,
3346-3375.
Gupta, Arpit, Christopher Hansman, and Pierre Mabille, 2021, Financial constraints and
the racial housing gap, Available at SSRN 3969433.
Hanson, Andrew, Zackary Hawley, Hal Martin, and Bo Liu, 2016, Discrimination in
mortgage lending: Evidence from a correspondence experiment, Journal of Urban
Economics 92, 48-65.
Jiang, Erica Xuewei, Yeonjoon Lee, and Will Shuo Liu, 2022, Disparities in consumer
credit: The role of loan officers in the FinTech era, Available at SSRN 4035764.
Kermani, Amir, and Francis Wong, 2021, The determinants of racial disparities in housing
returns, Available at SSRN 3846569.
Kuhn, Moritz, Moritz Schularick, and Ulrike I. Steins, 2020, Income and wealth inequality
in America, 1949–2016, Journal of Political Economy 128, 3469-3519.
Lang, Kevin, and Ariella Kahn-Lang Spitzer, 2020, Race discrimination: An economic
perspective, Journal of Economic Perspectives 34, 68-89.
Liberti, Jose M., and Atif R. Mian, 2008, Estimating the effect of hierarchies on
information use, Review of Financial Studies 22, 4057-4090.
Mayer, Erik J., 2021, Big Banks, Household Credit Access, and Intergenerational
Economic Mobility. Available at SSRN 3816308.
Munnell, Alicia H., Geoffrey M. B. Tootell, Lynn E. Browne, and James McEneaney,
1996, Mortgage lending in Boston: Interpreting HMDA data, American Economic
Review 86, 25-53.
Phelps, Edmund, 1972, The Statistical Theory of Racism and Sexism, American Economic
Review, 62(4), 533-539.
39
Ross, Stephen L., Margery A. Turner, Erin Godfrey, and Robin R. Smith, 2008, Mortgage
lending in Chicago and Los Angeles: A paired testing study of the pre-application
process, Journal of Urban Economics 63, 902-919.
Small, Mario L., and Devah Pager, 2020, Sociological perspectives on racial
discrimination, Journal of Economic Perspectives 34, 49-67.
Stein, Jeremy C, 2002, Information production and capital allocation: Decentralized versus
hierarchical firms, Journal of Finance 57, 1891-1921.
Tzioumis, Konstantinos, 2018, Demographic aspects of first names, Scientific Data 5, 1-9.
Tzioumis, Konstantinos, and Matthew Gee, 2013, Nonlinear incentives and mortgage
officers’ decisions, Journal of Financial Economics 107, 436-453.
Voicu, Ioan, 2018, Using first name information to improve race and ethnicity
classification, Statistics and Public Policy 5, 1-13.
40
Appendix A: Variable Definitions
This table provides the definitions and data sources for all variables. Panels A, B, and C contain the variables used in the analyses of minority
representation levels, mortgage lending, and loan performance, respectively.
Panel A: Variables used to analyze minority representation among mortgage loan officers
Variable
Source
Definition
Officer
I(Minority Officer)
NMLS
Indicator equal to one if the loan officer is a racial/ethnic minority
Officer Experience
NMLS
Number of years the loan officer has worked in the industry
Tenure
NMLS
Number of years the loan officer has worked for the lender
Turnover (2018)
NMLS
Indicator equal to one if the loan officer leaves the lender next year - only defined for 2018
Apps Handled (N)
HMDA
Number of mortgage applications handled by the officer this year - inclusive of all loan types
Apps Handled ($M)
HMDA
Total dollar value (in millions) of the mortgage applications handled by the officer this year
Originations (N)
HMDA
Number of mortgages originated by the officer this year – inclusive of all loan types
Originations ($M)
HMDA
Total dollar value (in millions) of the mortgages originated by the officer this year
HP App Share
HMDA
Percentage of the applications handled by the officer this year that are for home purchases
Lender
Lender Minority LO %
NMLS
Percentage of the lender’s loan officers that are racial/ethnic minorities, excluding all officers
working at branches in the same county as the focal officer
Lender Mortgage Orig (#)
HMDA
Number of mortgages originated by the lender this year
Credit Union
HMDA
Indicator equal to one if the lender is a credit union
Mortgage Company
HMDA
Indicator equal to one if the lender is a not a depository institution
Branch ZIP Code
Minority Population Share
Census
Percentage of the population that are racial/ethnic minorities in the branch’s ZIP code
Minority to White PIPC
Census
Ratio of minorities’ personal income per capita to that of whites in the branch’s ZIP code
PIPC
Census
Personal income per capita in the branch’s ZIP code
Population Density
Census
Population per square mile in the branch’s ZIP code
Census Regions
Northeastern U.S.
Census
Indicator equal to one if the officer’s branch is in the Northeastern U.S. Census Region
Southern U.S.
Census
Indicator equal to one if the officer’s branch is in the Southern U.S. Census Region
Midwestern U.S.
Census
Indicator equal to one if the officer’s branch is in the Midwestern U.S. Census Region
41
Panel B: Variables used in the mortgage lending analyses
Variable
Source
Definition
Application Outcomes
I(Completed)
HMDA
Indicator equal to one if the application is completed and submitted for a decision
I(Approved)
HMDA
Indicator equal to one if the application is approved
I(Origination)
HMDA
Indicator equal to one if the application leads to a loan origination
Loan Pricing (for originated loans)
Interest Rate
HMDA
The interest rate on the loan (in percentage point units)
Net Discount Points
HMDA
The dollar value paid for any rate discount minus the dollar value of any lender credits,
expressed as a percentage of the loan amount (in percentage point units)
Key Independent Vars
Minority
HMDA
Indicator equal to one if the applicant is a racial/ethnic minority
Minority Officer
NMLS
Indicator equal to one if the loan officer is a racial/ethnic minority
Low Income
HMDA
Indicator equal to one if the ratio of applicant income to MSA median income is in the
bottom third of all applicants
Small Bank
Call Report
Indicator equal to one if the lender is a bank with less than $10 Billion in total assets
FinTech
See Def:
Indicator equal to one if the lender is a FinTech firm according to Buchak et al. (2018)
Officer Minority Share
HMDA
Fraction of applications handled by the officer that are from minorities (ranges 0 to 1)
Officer Experience
NMLS
Number of years the loan officer has worked in the industry
Basic App Controls
Age
HMDA
Applicant age, controlled for with 10-year bins from age 20/younger to 90/older
Income
HMDA
Applicant income, controlled for with centile bins for the ratio of income to MSA
median income (or the median income in the state’s non-MSA areas if not in MSA)
Loan Amount
HMDA
Requested loan amount in dollars, controlled for with Log(Loan Amount)
Jumbo
HMDA
Indicator equal to one if the loan amount is over the conforming loan limit in the county
Joint Application
HMDA
Indicator equal to one if there are multiple people on the application
Loan Type - Conventional
HMDA
Indicator equal to one if the application is not associated with any government program
Loan Type - FHA
HMDA
Indicator equal to one if the application is for a FHA loan
Loan Type - VA
HMDA
Indicator equal to one if the application is for a VA loan
Loan Type - FSA/RHS
HMDA
Indicator equal to one if the application is for a FSA or RHS loan
Extended App Controls (for completed apps)
Credit Score
HMDA
FICO Score, controlled for with 10-point bins ranging from 500 to 850
Loan-to-Value
HMDA
LTV ratio, controlled for with bins for each percentage point from 10% to 110%
Debt-to-Income
HMDA
DTI ratio, controlled for with bins for each percentage point from 10% to 80%
Underwriting Sys. Rec.
HMDA
Each combination of underwriting system X output code is given a number, used for FE
42
Panel C: Variables used in the FHA loan performance analysis
Variable
Source
Definition
Loan Performance
I(Default)
FHA
Indicator equal to one if the borrower ever becomes 90 or more days delinquent
Key Independent Vars
Minority
FHA
Indicator equal to one if the borrower is a racial/ethnic minority
Minority Officer
NMLS
Indicator equal to one if the loan officer is a racial/ethnic minority
Low Income
FHA
Indicator equal to one if the ratio of borrower income to MSA median income is in the
bottom third of all borrowers
Small Bank
Call Report
Indicator equal to one if the lender is a bank with less than $10 Billion in total assets
FinTech
See Def:
Indicator equal to one if the lender is a FinTech firm according to Buchak et al. (2018)
Officer Minority Share
FHA
Fraction of loans made by the officer that are to minorities (ranges 0 to 1)
Officer Experience
NMLS
Number of years the loan officer has worked in the industry
FHA Controls
Interest Rate
FHA
The interest rate on the loan (in percentage point units)
Loan Amount
FHA
The loan amount in dollars, controlled for with Log(Loan Amount)
Income
FHA
Borrower income, controlled for with centile bins for the ratio of income to MSA median
income (or the median income in the state’s non-MSA areas if not in MSA)
Credit Score
FHA
FICO Score, controlled for with 10-point bins ranging from 500 to 850
Loan-to-Value
FHA
LTV ratio, controlled for with bins for each percentage point from 60% to 97%
Debt-to-Income
FHA
DTI ratio, controlled for with bins for each percentage point from 10% to 57%
FT Buyer
FHA
Indicator equal to one if the borrower is a first-time home buyer
43
Figure 1: Minority Loan Officers Relative to Population Share
This figure shows a binned scatter plot of the percentage of loan officers that are minorities against
the ZIP code minority population share. The sample includes all loan officer-years in the
HMDA/NMLS matched panel, which covers 2018 to 2019. The bin averages are computed for
decile bins formed based on the ZIP code minority population share. The plot also includes the
best fit line and the 45-degree line for reference.
44
Figure 2: Loan Officer Race and Mortgage Approval
This figure plots mortgage approval rates for white and minority applicants against credit scores
(in decile bins). The left panel shows applications handled by white loan officers, and the right
panel shows those handled by minority officers. The sample includes all completed home purchase
mortgage applications in the HMDA database in 2018 and 2019, subject to the standard data filters
described in Section 2.5.
45
Figure 3: Loan Officer Race and Mortgage Default Rates
This figure plots mortgage default rates for white and minority borrowers against credit scores (in
decile bins). Default is defined as the loan ever becoming 90 or more days delinquent. The left
panel shows loans handled by white loan officers, while the right panel shows loans handled by
minority loan officers. The sample includes FHA home purchase mortgages originated between
2012 and 2018, subject to the standard data filters described in Section 2.5.
46
Figure 4: Loan Officer Race and Minority Gaps – Specialization
This figure plots unexplained minority gaps in mortgage approvals and defaults against the
handling loan officer’s minority application share. The unexplained minority gap estimates come
from OLS regressions with the full set of controls in Table 8, where the minority borrower
indicator is interacted with indicators for each combination of white/minority officer and the group
the officer falls into in terms of minority application share. The left panel shows the estimated
minority gaps for white officers (with 95% confidence intervals), and the right panel shows the
results for minority officers. The samples match those in Table 8. For approvals, the sample is all
completed FHA home purchase mortgage applications in the HMDA database in 2018 and 2019.
For defaults, the sample is all FHA home purchase mortgages originated from 2012 to 2018. Both
samples implement the standard data filters described in Section 2.5.
47
Figure 5: Loan Officer Race and Minority Gaps – Experience
This figure plots unexplained minority gaps in mortgage approvals and defaults against the
handling loan officer’s years of experience. The unexplained minority gap estimates come from
OLS regressions with the full set of controls in Table 8, where the minority borrower indicator is
interacted with indicators for each combination of white/minority officer and the group the officer
falls into in terms of experience. The left panel shows the estimated minority gaps for white
officers (with 95% confidence intervals), and the right panel shows the results for minority
officers. The samples match those in Table 8. For approvals, the sample is all completed FHA
home purchase mortgage applications in the HMDA database in 2018 and 2019. For defaults, the
sample is all FHA home purchase mortgages originated from 2012 to 2018. Both samples
implement the standard data filters described in Section 2.5.
48
Table 1: Summary Statistics for Mortgage Loan Officers
This table presents summary statistics for the mortgage loan officers in our HMDA/NMLS
matched panel. Panel A provides a 2019 snapshot of racial/ethnic groups’ share of the U.S.
population (column 1), and of the loan officers in our data (column 2). Columns 3-5 show similar
statistics for loan officers working at banks, credit unions, and mortgage companies, respectively.
Panel B provides summary statistics for loan officer, lender, and ZIP code characteristics for the
full 2018-2019 loan officer-year panel (columns 1-3), as well as sample means by race/ethnicity
(columns 4-7). Appendix A provides variable definitions.
Panel A: Loan Officer Race (2019 Snapshot)
U.S.
Population
Loan Officers
(N=255,277)
Bank LO
(N=134,257)
Credit Union LO
(N=23,073)
Mort. Co. LO
(N=97,947)
White
60.70%
84.59%
84.16%
86.65%
84.70%
Hispanic
18.00%
8.88%
8.80%
9.31%
8.89%
Black
13.24%
1.76%
2.21%
1.64%
1.17%
Asian
6.22%
4.74%
4.80%
2.36%
5.22%
Other
1.84%
0.03%
0.04%
0.03%
0.02%
Panel B: Summary Statistics (All Loan Officer-Years 2018-2019)
Full Sample
White
LO
Hispanic
LO
Black
LO
Asian
LO
Mean
Median
Std Dev
Mean
Mean
Mean
Mean
Officer
Officer Experience
9.11
9.00
6.38
9.32
7.73
8.02
8.26
Tenure
5.71
3.00
6.21
5.83
4.72
5.58
5.40
Turnover (2018)
0.21
0.00
0.41
0.21
0.23
0.23
0.19
Apps Handled (N)
39.91
14.00
63.07
41.32
31.53
31.62
32.94
Apps Handled ($M)
9.39
2.18
16.28
9.60
7.06
5.97
11.12
Originations (N)
25.13
8.00
38.64
26.29
18.04
16.56
20.48
Originations ($M)
6.06
1.32
10.52
6.24
4.22
3.30
7.17
HP App Share
0.37
0.27
0.37
0.38
0.33
0.24
0.34
Lender
Lender Minority LO %
0.15
0.14
0.11
0.14
0.22
0.17
0.26
Lender Mortgage Orig (#)
61,317
15,900
101,321
58,089
72,955
80,938
91,139
Branch ZIP Code
Minority Population Share
0.34
0.29
0.22
0.31
0.54
0.57
0.55
Minority to White PIPC
0.77
0.78
0.15
0.76
0.85
0.74
0.79
PIPC
40,758
37,106
16,912
40,807
37,785
35,080
47,553
Population Density
3,217
2,066
4,415
2,808
5,112
4,529
6,644
49
Table 2: Summary Statistics for Home Purchase Mortgage Applications
This table presents summary statistics for our sample of mortgage applications, which includes all
home purchase applications in the HMDA database in 2018-2019 (including those never
completed), subject to the standard data filters described in Section 2.5. Columns 1-3 present
summary statistics for application outcomes, key independent variables, as well as the basic and
extended application controls. Columns 4-7 present sample means for each combination of
white/minority loan officer and white/minority applicant. Appendix A provides variable
definitions.
White Officers
Minority Officers
Full Sample
Whites
Minorities
Whites
Minorities
N=5.65M
N=3.79M
N=1.31M
N=191K
N=359K
Mean
Median
Std Dev
Mean
Mean
Mean
Mean
Application Outcomes
Completed
83.7%
1
0.369
84.8%
81.3%
80.6%
82.8%
Approved, given completed
92.1%
1
0.270
93.7%
88.3%
90.1%
88.6%
Taken-up, given approved
97.2%
1
0.164
97.4%
96.8%
97.1%
96.8%
All-in Origination %
74.9%
1
0.433
77.4%
69.5%
70.5%
71.0%
Key Independent Vars
Minority Officer
0.097
0
0.296
0
0
1
1
Minority
0.295
0
0.456
0
1
0
1
Low Income
0.330
0
0.470
0.314
0.375
0.250
0.379
Small Bank
0.104
0
0.305
0.121
0.076
0.045
0.056
FinTech
0.127
0
0.332
0.123
0.141
0.142
0.105
Basic App Controls
Age
40.8
38.0
13.2
41.0
40.2
42.1
40.0
Income
94,209
77,000
63,796
96,483
88,151
106,500
85,658
Loan Amount
255,897
223,250
147,453
251,226
258,932
288,973
276,594
Jumbo
0.037
0
0.188
0.037
0.031
0.062
0.038
Joint Application
0.426
0
0.495
0.449
0.363
0.464
0.397
Loan Type - Conventional
0.614
1
0.487
0.653
0.508
0.642
0.569
Loan Type - FHA
0.228
0
0.420
0.185
0.331
0.155
0.351
Loan Type - VA
0.126
0
0.332
0.123
0.142
0.191
0.071
Loan Type - FSA/RHS
0.031
0
0.174
0.039
0.018
0.011
0.009
Extended App Controls (for completed apps)
Credit Score
725
731
58
731
707
732
711
Loan-to-Value
0.892
0.950
0.129
0.882
0.920
0.877
0.905
Debt-to-Income
0.389
0.397
0.103
0.378
0.415
0.387
0.425
50
Table 3: Do Minority Loan Officers Improve Minorities’ Access to Credit?
This table presents regressions examining the effect of loan officer and applicant race/ethnicity on whether mortgage applications are
completed, approved, and ultimately originated. The dependent variable in columns 1 and 2 is an indicator for the application being
completed. In columns 3-6, the dependent variable is an indicator for the application being approved. We split mortgage approval decisions
into two cases: applications where the Automated Underwriting System (AUS) gives a clear decision (columns 3-4), and applications where
the AUS does not give a definitive decision, allowing for discretion (columns 5-6). In columns 7-8, the dependent variable is an indicator
for a started application ultimately resulting in an origination. The key independent variables across all specifications are indicators for the
applicant being a minority, the loan officer being a minority, and their interaction. The sample in columns 1-2 and 7-8 includes all home
purchase mortgage applications in the HMDA database in 2018-2019 (including those never completed), subject to the standard data filters
described in Section 2.5. The sample in columns 3-6 is restricted to completed applications. Appendix A lists the controls and provides
variable definitions. The standard errors are two-way clustered at the lender and county level, and the symbols *, **, and *** indicate
significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Completed)
I(Approved)
I(Origination)
Low Discretion
High Discretion
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Minority
-0.021
***
(0.001)
-0.019
***
(0.001)
-0.010
***
(0.001)
-0.010
***
(0.001)
-0.029
***
(0.002)
-0.026
***
(0.001)
-0.050
***
(0.002)
-0.046
***
(0.002)
Minority X Minority Officer
0.015
***
(0.002)
0.011
***
(0.002)
0.002
(0.002)
-0.000
(0.002)
0.012
***
(0.004)
0.014
***
(0.004)
0.025
***
(0.003)
0.017
***
(0.003)
Minority Officer
-0.002
*
(0.001)
-0.003
(0.002)
-0.002
(0.003)
-0.005
**
(0.002)
Basic App Controls
Y
Y
Y
Y
Y
Y
Y
Y
Extended App Controls
-
-
Y
Y
Y
Y
-
-
Underwriting Sys. Rec. FE
-
-
Y
Y
Y
Y
-
-
Branch-Year-Officer FE
-
Y
-
Y
-
Y
-
Y
Branch-Year FE
Y
-
Y
-
Y
-
Y
-
Property County FE
Y
Y
Y
Y
Y
Y
Y
Y
R-Squared
0.057
0.090
0.167
0.202
0.562
0.618
0.080
0.114
Observations
5,643,662
5,625,635
3,628,307
3,609,947
544,380
514,953
5,643,662
5,625,635
51
Table 4: IV Estimates of Minority Officers’ Effect on Completions, Approvals, and Originations
This table presents instrumental variables tests examining the effect of minority loan officers on
application completions, approvals, and originations. We instrument for the application being
handled by a minority officer with the share of applications at the branch that were handled by
minority officers on the same day of the week as the current application during the prior 12 weeks
(P12 DOW Min. Off. Share). Given the branch-week fixed effects, the intra-week variation in this
instrument captures loan officer work schedules and/or rotation policies. We also instrument for
Minority X Minority Officer with the interaction of Minority and our instrument. Panel A reports
the first-stage regression results and Panel B reports the second-stage results. In columns 1 and 4,
we study application completion and all-in origination effects using all home purchase mortgage
applications in the HMDA database in 2018-2019 (including those never completed), subject to
the standard data filters described in Section 2.5, and the requirement that the application is opened
after the 12
th
week of 2018. In columns 2 and 3, we restrict the sample to completed applications
in order to study credit approval. The reported controls and fixed effects apply to both panels—
see Appendix A for variable definitions. The standard errors are two-way clustered at the lender
and county level, and the symbols *, **, and *** indicate significance at the 10%, 5%, and 1%
levels, respectively.
Sample:
All Apps
Completed Apps
All Apps
Low
Discretion
High
Discretion
(1)
(2)
(3)
(4)
Panel A: First-stage
Dependent Variable:
I(Minority Officer)
P12 DOW Min. Off. Share
0.067
***
(0.007)
0.080
***
(0.009)
0.101
***
(0.022)
0.067
***
(0.007)
R-Squared
0.550
0.557
0.597
0.550
First-stage F-stat
67.3
55.2
18.3
67.3
Panel B: Second-stage
Dependent Variable:
I(Completed)
I(Approved)
I(Origination)
Minority
-0.021
***
(0.002)
-0.010
***
(0.001)
-0.031
***
(0.003)
-0.050
***
(0.002)
Minority X Minority Officer
0.018
**
(0.008)
0.008
(0.006)
0.036
**
(0.016)
0.038
***
(0.011)
Minority Officer
-0.001
(0.033)
-0.005
(0.021)
-0.009
(0.065)
-0.007
(0.043)
Basic App Controls
Y
Y
Y
Y
Extended App Controls
-
Y
Y
-
Underwriting Sys. Rec. FE
-
Y
Y
-
Branch-Week FE
Y
Y
Y
Y
Day-of-Week FE
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
Observations
3,958,927
2,408,755
225,785
3,958,927
52
Table 5: Summary Statistics for FHA Loans
This table presents summary statistics for the sample of FHA home purchase mortgages originated
between 2012 and 2018, subject to the standard data filters described in Section 2.5. Columns 1-3
present summary statistics for default (the loan ever becoming 90 or more days delinquent), key
independent variables, as well as the control variables. Columns 4-7 present the sample means for
each combination of white/minority loan officer and white/minority borrower. Appendix A
provides variable definitions.
White Officers
Minority Officers
Full Sample
Whites
Minorities
Whites
Minorities
N=3.39M
N=2.12M
N=910K
N=80K
N=282K
Mean
Median
Std Dev
Mean
Mean
Mean
Mean
Loan Outcome
I(Default)
0.091
0
0.287
0.080
0.117
0.084
0.092
Key Independent Vars
Minority Officer
0.107
0
0.308
0
0
1
1
Minority
0.351
0
0.477
0
1
0
1
Low Income
0.330
0
0.470
0.316
0.354
0.264
0.380
Small Bank
0.125
0
0.331
0.143
0.104
0.077
0.069
FinTech
0.122
0
0.327
0.119
0.127
0.181
0.111
FHA Controls
Interest Rate
4.117
4.000
0.559
4.075
4.177
4.143
4.237
Loan Amount
185,871
166,920
90,895
179,706
192,286
206,168
205,844
Income
66,772
59,256
33,688
68,253
64,398
72,950
61,541
Credit Score
681
673
46
684
674
682
678
Loan-to-Value
0.954
0.965
0.040
0.954
0.955
0.951
0.954
Debt-to-Income
0.412
0.420
0.090
0.402
0.425
0.414
0.437
FT Buyer
0.813
1
0.390
0.779
0.867
0.810
0.895
53
Table 6: Loan Officer Race and Default Rates
This table presents regressions examining default rates based on borrower and loan officer
race/ethnicity. The dependent variable is an indicator for the loan ever becoming 90 or more days
delinquent. The key independent variables are indicators for the borrower being a minority, the
loan officer being a minority, and their interaction. Columns 1 and 2 present the OLS results for
the full sample of FHA home purchase mortgages originated between 2012 and 2018, subject to
the standard data filters described in Section 2.5. The remaining columns implement an
instrumental variables test that uses the share of FHA loans at the branch that were handled by
minority officers on the same day of the week during the prior 12 weeks to instrument for a
minority officer handling the loan. The sample is similar to the first two columns, with the
additional requirement that the application date is after the 12
th
week of 2012 and the instrument
can be computed. Columns 3 and 4 present the OLS and IV results in this sample, respectively.
Appendix A lists the controls and provides variable definitions. The standard errors are two-way
clustered at the lender and county level, and the symbols *, **, and *** indicate significance at
the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Default)
Full Sample
IV Sample
OLS
OLS
OLS
IV
(1)
(2)
(3)
(4)
Minority
0.018
***
(0.001)
0.017
***
(0.001)
0.018
***
(0.001)
0.017
***
(0.002)
Minority X Minority Officer
-0.022
***
(0.002)
-0.017
***
(0.002)
-0.023
***
(0.002)
-0.052
**
(0.025)
Minority Officer
0.002
**
(0.001)
0.002
(0.002)
0.028
(0.075)
FHA Controls
Y
Y
Y
Y
Origination Month FE
Y
Y
-
-
Branch-Year-Officer FE
-
Y
-
-
Branch-Year FE
Y
-
-
-
Branch-Month FE
-
-
Y
Y
Day-of-Week FE
-
-
Y
Y
Property County FE
Y
Y
Y
Y
R-Squared
0.103
0.158
0.229
-
Observations
3,370,855
3,297,801
2,239,026
2,239,026
First-stage F-stat
-
-
-
59.1
54
Table 7: Heterogeneity Across Borrowers and Lenders
This table presents regressions that explore the heterogeneity across borrowers and lenders in the impact of loan
officer and borrower race on mortgage approval and default. The analysis of mortgage approval (columns 1-3) is
based on completed FHA home purchase mortgage applications in the HMDA database in 2018-2019. The analysis
of mortgage default (columns 4-6) uses FHA home purchase mortgages originated from 2012 to 2018. Both samples
implement the standard data filters described in Section 2.5. The key independent variables are an indicator for the
borrower being a minority, an indicator for the loan officer being a minority, their interaction, and further triple
interactions with indicators for the borrower and officer being of the same/different race, or for the borrower having
a low income, or for the application occurring at a small bank or FinTech lender, respectively. All base terms for
the interactions are included unless they are subsumed by the binned controls or fixed effects. Appendix A lists the
controls and provides variable definitions. The standard errors are two-way clustered at the lender and county level,
and the symbols *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Approved)
I(Default)
(1)
(2)
(3)
(4)
(5)
(6)
Minority
-0.017
***
(0.001)
-0.015
***
(0.001)
-0.016
***
(0.001)
0.018
***
(0.001)
0.021
***
(0.001)
0.018
***
(0.001)
Minority Officer
-0.002
(0.002)
-0.001
(0.003)
-0.003
(0.003)
0.003
**
(0.001)
0.004
***
(0.001)
0.002
*
(0.001)
Minority X Minority Officer
0.004
(0.004)
0.007
***
(0.003)
-0.017
***
(0.002)
-0.023
***
(0.002)
Minority X Min. Off. X Other Race
0.004
(0.003)
-0.003
(0.003)
Minority X Min. Off. X Same Race
0.008
***
(0.003)
-0.027
***
(0.002)
Minority X Min. Off. X Low Inc.
0.017
**
(0.007)
-0.012
***
(0.003)
Minority X Low Inc.
-0.005
***
(0.001)
-0.010
***
(0.001)
Min. Off. X Low Inc.
-0.003
(0.006)
-0.004
(0.003)
Minority X Min. Off. X Small Bank
0.013
*
(0.007)
-0.002
(0.006)
Minority X Small Bank
-0.002
(0.003)
-0.003
(0.002)
Min. Off. X Small Bank
0.001
(0.007)
0.004
(0.006)
Minority X Min. Off. X FinTech
-0.012
**
(0.005)
0.008
**
(0.003)
Minority X FinTech
-0.005
(0.005)
0.002
(0.002)
Min. Off. X FinTech
-0.000
(0.007)
-0.001
(0.003)
Basic App Controls
Y
Y
Y
-
-
-
Extended App Controls
Y
Y
Y
-
-
-
Underwriting Sys. Rec. FE
Y
Y
Y
-
-
-
Branch-Year FE
Y
Y
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
Y
Y
FHA Controls
-
-
-
Y
Y
Y
Origination Month FE
-
-
-
Y
Y
Y
R-Squared
0.403
0.403
0.403
0.103
0.103
0.103
Observations
956,543
956,543
956,543
3,370,855
3,370,855
3,370,855
55
Table 8: Heterogeneity Across Loan Officers
This table presents regressions that explore the heterogeneity across loan officers in the impact of
race/ethnicity on mortgage approval and default. The analysis of mortgage approval (columns 1 and 2)
is based on completed FHA home purchase mortgage applications in the HMDA database in 2018-
2019. The analysis of mortgage default (columns 3 and 4) uses FHA home purchase mortgages
originated from 2012 to 2018. Both samples implement the standard data filters described in Section
2.5. The key independent variables are an indicator for the borrower being a minority, an indicator for
the loan officer being a minority, their interaction, and further triple interactions with the officer’s
minority share in their application/loan portfolio, and the officer’s years of industry experience,
respectively. Appendix A lists the controls and provides variable definitions. The standard errors are
two-way clustered at the lender and county level, and the symbols *, **, and *** indicate significance
at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Approved)
I(Default)
(1)
(2)
(3)
(4)
Minority
-0.0153
***
(0.0016)
-0.0213
***
(0.0022)
0.0165
***
(0.0014)
0.0260
***
(0.0019)
Minority Officer
-0.0034
(0.0037)
-0.0030
(0.0048)
0.0039
*
(0.0021)
0.0020
(0.0025)
Minority X Minority Officer
-0.0032
(0.0078)
0.0104
**
(0.0051)
0.0051
(0.0049)
-0.0278
***
(0.0028)
Off. Minority Share
-0.0074
*
(0.0039)
0.0078
***
(0.0018)
Minority X Off. Minority Share
-0.0031
(0.0043)
0.0002
(0.0036)
Min. Off. X Off. Minority Share
0.0057
(0.0070)
-0.0031
(0.0037)
Minority X Min. Off. X Off. Minority Share
0.0191
**
(0.0090)
-0.0340
***
(0.0073)
Off. Experience
0.0008
***
(0.0002)
0.0001
(0.0001)
Minority X Off. Experience
0.0005
**
(0.0002)
-0.0012
***
(0.0002)
Min. Off. X Off. Experience
0.0002
(0.0005)
0.0001
(0.0003)
Minority X Min. Off. X Off. Experience
-0.0006
(0.0005)
0.0008
**
(0.0004)
Basic App Controls
Y
Y
-
-
Extended App Controls
Y
Y
-
-
Underwriting Sys. Rec. FE
Y
Y
-
-
Branch-Year FE
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
FHA Controls
-
-
Y
Y
Origination Month FE
-
-
Y
Y
R-Squared
0.403
0.403
0.103
0.103
Observations
956,543
956,543
3,370,855
3,370,855
56
Table 9: Differential Reaction to Hard Information
This table presents regressions that explore loan officers’ differential reaction to hard information in mortgage approval. The dependent
variable is an indicator for the application being approved. The sample is all completed FHA home purchase mortgage applications by
minority borrowers in the HMDA database in 2018-2019, subject to the standard data filters described in Section 2.5. Columns 1 and 2
utilize the subsample of applications handled by white and minority officers, respectively. Columns 4 and 5 use applications handled by
inexperienced and experienced white officers, respectively. Columns 7 and 8 use applications handled by non-specialist and specialist
minority loan officers, respectively. Columns 3, 6, and 9 test the differences in the coefficients for the preceding two columns. The key
independent variables are an indicator for joint applications, log(loan amount), log(income), credit score (scaled by 100), debt-to-income
ratio, and age (scaled by 10). Appendix A provides variable definitions. The standard errors are two-way clustered at the lender and county
level, and the symbols *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Approved)
White
Officers
Minority
Officers
Diff.
Inexp.
White
Officers
Exp.
White
Officers
Diff.
Non-Spec.
Minority
Officers
Specialist
Minority
Officers
Diff.
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Joint Application
-0.007
***
(0.002)
-0.005
**
(0.002)
-0.006
***
(0.002)
-0.007
***
(0.002)
-0.008
**
(0.004)
-0.005
(0.003)
Log(Loan Amount)
0.031
***
(0.006)
0.036
***
(0.007)
0.035
***
(0.009)
0.027
***
(0.005)
0.029
*
(0.017)
0.036
***
(0.006)
Log(Income)
0.017
***
(0.004)
0.012
**
(0.005)
0.020
***
(0.005)
0.012
**
(0.005)
*
0.027
***
(0.009)
0.007
(0.005)
**
Credit Score / 100
0.046
***
(0.003)
0.037
***
(0.005)
**
0.049
***
(0.004)
0.042
***
(0.003)
**
0.045
***
(0.006)
0.033
***
(0.005)
***
Debt-to-Income
-0.552
***
(0.019)
-0.506
***
(0.022)
**
-0.581
***
(0.020)
-0.509
***
(0.020)
***
-0.522
***
(0.035)
-0.496
***
(0.022)
Age / 10
-0.007
***
(0.001)
-0.006
***
(0.001)
-0.008
***
(0.001)
-0.006
***
(0.001)
-0.007
***
(0.001)
-0.006
***
(0.001)
Underwriting Sys. Rec. FE
Y
Y
Y
Y
Y
Y
Branch-Year FE
Y
Y
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
Y
Y
R-Squared
0.395
0.379
0.406
0.413
0.462
0.349
Observations
310,304
89,720
178,448
128,724
25,082
64,060
57
Table 10: Does Race Influence Applicant-to-Loan Officer Matching?
This table presents regressions examining which mortgage applicants are most likely to be matched
with a minority loan officer. The dependent variable is an indicator for the application being
handled by a minority officer. The key independent variables are an indicator for the applicant
being a minority, and its interaction with indicators for the applicant having a low income, or for
the application occurring at a small bank or FinTech lender, respectively. All base terms for the
interactions are included unless they are subsumed by the fixed effects. The sample includes all
home purchase mortgage applications in the HMDA database in 2018-2019 (including those never
completed), subject to the standard data filters described in Section 2.5. Appendix A provides
variable definitions. The standard errors are two-way clustered at the lender and county level, and
the symbols *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Minority Officer)
(1)
(2)
(3)
(4)
Constant
0.048
***
(0.006)
Minority
0.167
***
(0.014)
0.046
***
(0.003)
0.038
***
(0.003)
0.048
***
(0.004)
Minority X Low Inc.
0.022
***
(0.002)
Low Income
0.003
***
(0.001)
Minority X Small Bank
-0.004
(0.006)
Minority X FinTech
-0.012
**
(0.005)
Officer Experience
-0.001
***
(0.000)
-0.001
***
(0.000)
-0.001
***
(0.000)
Loan Type FE
-
Y
Y
Y
Branch-Year FE
-
Y
Y
Y
Property County FE
-
Y
Y
Y
R-Squared
0.066
0.540
0.541
0.540
Observations
5,649,234
5,643,662
5,643,662
5,643,662
58
Internet Appendix
Figure IA1: Classifying Applications as High versus Low Discretion Cases
This figure shows a histogram of the mortgage applications in our sample based on the average
approval rate given the output code of the automated underwriting system (AUS). Our sample
includes all completed home purchase mortgage applications in the HMDA database in 2018-2019,
subject to the standard data filters described in Section 2.5. We classify an application as a “low
discretion” case if the average approval rate for the AUS code that the application receives is
greater than 90% (see the red line in the plot). Applications receiving AUS codes with less than
90% approval rates are classified as “high discretion.”
59
Table IA1: Minority Representation among Loan Officers—Stylized Facts
This table presents regressions that examine the role of economic, geographic, and institutional
factors in determining minority representation levels among mortgage loan officers. The
dependent variable is an indicator for the loan officer being a racial/ethnic minority. The key
independent variables are ZIP code demographic and economic characteristics, indicators for
census regions of the U.S., and lender characteristics. In columns 1-4, the sample includes all loan
officer-years in our HMDA/NMLS matched panel. The tests in columns 5-7 focus on loan officers
working for banks, credit unions, and mortgage companies, respectively. Appendix A provides
variable definitions. The standard errors are two-way clustered at the lender and county level, and
the symbols *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Minority Officer)
Full Sample
Banks
C.U.
Mort. Co.
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Minority Population Share
0.662
***
(0.052)
0.655
***
(0.049)
0.592
***
(0.042)
Minority to White PIPC
0.183
***
(0.031)
0.162
***
(0.026)
Log(PIPC)
0.049
***
(0.013)
0.051
***
(0.008)
Log(Population Density)
0.004
(0.003)
0.004
*
(0.002)
Northeastern U.S.
0.003
(0.017)
-0.022
*
(0.012)
Southern U.S.
-0.055
***
(0.019)
-0.041
***
(0.015)
Midwestern U.S.
-0.037
**
(0.019)
-0.060
***
(0.016)
Credit Union
-0.014
**
(0.006)
Mortgage Company
-0.057
***
(0.007)
Lender Minority LO %
0.277
***
(0.037)
0.118
***
(0.031)
0.352
***
(0.031)
Log(Lender Mortgage
Orig #)
0.002
(0.002)
0.006
**
(0.003)
0.000
(0.001)
Constant
-0.075
***
(0.011)
Year FE
-
Y
Y
Y
Y
Y
Y
Lender FE
-
-
Y
-
-
-
-
Branch ZIP Code FE
-
-
-
Y
Y
Y
Y
R-Squared
0.161
0.178
0.293
0.294
0.338
0.326
0.315
Observations
514,892
514,892
511,983
513,809
273,823
41,094
190,274
60
Table IA2: The Effect of Own-Race Officers on Application Completion Rates
This table presents regressions that examine the role of loan officers in whether or not mortgage
applicants complete an application they started. The dependent variable is an indicator for the
application being completed. The key independent variables are an indicator for the officer and
applicant being of the same race, and indicators for each applicant race (white is the omitted
group). Column 1 presents the results for the full sample. Columns 2-5 present results for the
applications handled by white, Hispanic, Black, and Asian loan officers, respectively. The sample
includes all home purchase mortgage applications in the HMDA database in 2018-2019 (including
those never completed), subject to the standard data filters described in Section 2.5. Appendix A
lists the controls and provides variable definitions. The standard errors are two-way clustered at
the lender and county level, and the symbols *, **, and *** indicate significance at the 10%, 5%,
and 1% levels, respectively.
Dependent Variable:
I(Application Completed)
Full
Sample
White
Officers
Hispanic
Officers
Black
Officers
Asian
Officers
(1)
(2)
(3)
(4)
(5)
Own-Race Officer
0.006
***
(0.001)
Hispanic
-0.004
***
(0.001)
-0.010
***
(0.001)
0.000
(0.002)
0.002
(0.007)
-0.006
(0.005)
Black
-0.012
***
(0.001)
-0.018
***
(0.001)
-0.014
***
(0.003)
-0.001
(0.005)
-0.017
***
(0.005)
Asian
-0.028
***
(0.002)
-0.035
***
(0.002)
-0.022
***
(0.004)
-0.029
***
(0.007)
-0.019
***
(0.005)
Basic App Controls
Y
Y
Y
Y
Y
Branch-Year-Officer FE
Y
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
Y
R-Squared
0.090
0.089
0.103
0.123
0.133
Observations
5,625,635
5,079,008
380,178
51,045
113,950
61
Table IA3: The Effect of Own-Race Officers on Credit Approval
This table presents regressions that examine the role of loan officers in mortgage approval. The
dependent variable is an indicator for the application being approved. The key independent
variables are an indicator for the officer and applicant being of the same race, and indicators for
each applicant race (white is the omitted group). Column 1 presents the results for the full sample.
Columns 2-5 present results for the applications handled by white, Hispanic, Black, and Asian loan
officers, respectively. The sample includes completed home purchase mortgage applications in the
HMDA database in 2018-2019 that we classify as “high discretion” and that satisfy the standard
data filters described in Section 2.5. Appendix A lists the controls and provides variable
definitions. The standard errors are two-way clustered at the lender and county level, and the
symbols *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Approved)
Full
Sample
White
Officers
Hispanic
Officers
Black
Officers
Asian
Officers
(1)
(2)
(3)
(4)
(5)
Own-Race Officer
0.008
***
(0.002)
Hispanic
-0.013
***
(0.003)
-0.021
***
(0.002)
-0.005
(0.006)
-0.010
(0.026)
-0.018
(0.012)
Black
-0.017
***
(0.003)
-0.024
***
(0.002)
-0.023
***
(0.008)
-0.010
(0.011)
-0.012
(0.011)
Asian
-0.024
***
(0.004)
-0.032
***
(0.003)
-0.015
*
(0.009)
-0.023
(0.023)
-0.011
**
(0.005)
Basic App Controls
Y
Y
Y
Y
Y
Extended App Controls
Y
Y
Y
Y
Y
Underwriting Sys. Rec. FE
Y
Y
Y
Y
Y
Branch-Year-Officer FE
Y
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
Y
R-Squared
0.619
0.618
0.649
0.670
0.653
Observations
514,953
454,693
35,964
6,493
16,031
62
Table IA4: All-in Estimate of Own-Race Officers’ Effect on Lending to Minorities
This table presents regressions that examine the all-in effect of loan officers on whether a started
application ultimately ends in an origination. The dependent variable is an indicator for the
application resulting in an origination. The key independent variables are an indicator for the
officer and applicant being of the same race, and indicators for each applicant race (white is the
omitted group). Column 1 presents the results for the full sample. Columns 2-5 present results for
the applications handled by white, Hispanic, Black, and Asian loan officers, respectively. The
sample includes all home purchase mortgage applications in the HMDA database in 2018-2019
(including those never completed), subject to the standard data filters described in Section 2.5.
Appendix A lists the controls and provides variable definitions. The standard errors are two-way
clustered at the lender and county level, and the symbols *, **, and *** indicate significance at
the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Origination)
Full
Sample
White
Officers
Hispanic
Officers
Black
Officers
Asian
Officers
(1)
(2)
(3)
(4)
(5)
Own-Race Officer
0.009
***
(0.001)
Hispanic
-0.021
***
(0.002)
-0.029
***
(0.001)
-0.012
***
(0.003)
-0.026
***
(0.008)
-0.027
***
(0.005)
Black
-0.052
***
(0.002)
-0.060
***
(0.001)
-0.057
***
(0.004)
-0.042
***
(0.006)
-0.066
***
(0.007)
Asian
-0.044
***
(0.003)
-0.054
***
(0.003)
-0.038
***
(0.005)
-0.053
***
(0.009)
-0.029
***
(0.005)
Basic App Controls
Y
Y
Y
Y
Y
Branch-Year-Officer FE
Y
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
Y
R-Squared
0.114
0.113
0.119
0.147
0.147
Observations
5,625,635
5,079,008
380,178
51,045
113,950
63
Table IA5: IV Balance Tests for the Analyses of Completions, Approvals, and Originations
This table presents the balance tests for our instrumental variables analyses. The dependent
variable is our instrument—the share of applications at the branch that were handled by minority
officers on the same day of the week as the current application during the prior 12 weeks (in
percentage point units). The key independent variables are the controls from our IV analyses,
including continuous versions of our binned controls. We also include the average approval rate
for the AUS code the application received in columns 4-5 (a continuous measure of the AUS
recommendation). In column 1, the sample includes all home purchase mortgage applications in
the HMDA database in 2018-2019 (including those never completed), subject to the standard data
filters described in Section 2.5, and the requirement that the application is opened after the 12
th
week of 2018. This sample corresponds to the IV analyses of application completions and
originations. Columns 2-5 restrict the sample to completed applications, and split the data into
cases that are deemed low versus high discretion. These samples correspond to the IV analysis of
credit approval. The p-value reported at the bottom of the columns is for an F-test of the joint
significance of the control variables. Appendix A provides variable definitions. The standard errors
are two-way clustered at the lender and county level, and the symbols *, **, and *** indicate
significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
P12 DOW Min. Off. Share
All Apps
Completed Apps
Low
Discretion
High
Discretion
Low
Discretion
High
Discretion
(1)
(2)
(3)
(4)
(5)
Age
-0.001
**
(0.000)
-0.000
(0.000)
-0.001
(0.001)
Log(Income)
-0.009
(0.009)
-0.014
(0.012)
-0.018
(0.039)
Log(Loan Amount)
0.005
(0.010)
0.006
(0.015)
-0.016
(0.037)
Jumbo
0.001
(0.026)
0.065
(0.045)
-0.004
(0.047)
Joint Application
-0.001
(0.007)
0.007
(0.009)
0.026
(0.035)
Credit Score
0.000
(0.000)
0.000
(0.000)
Loan-to-Value
-0.003
(0.035)
-0.108
(0.132)
Debt-to-Income
-0.038
(0.043)
-0.197
*
(0.107)
AUS Code Approval Rate
-0.009
(0.016)
-0.000
(0.001)
Loan Type FE
Y
Y
Y
Y
Y
Underwriting Sys. Rec. FE
-
Y
Y
-
-
Branch-Week FE
Y
Y
Y
Y
Y
Day-of-Week FE
Y
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
Y
Observations
3,958,927
2,408,755
225,785
2,408,755
225,785
p-value on Joint F-test
0.305
0.641
0.350
0.575
0.819
64
Table IA6: Replication of our Main Results in the FHA Subsample
This table presents regressions that replicate our main results using the FHA subsample of the
home purchase mortgage applications in the HMDA database in 2018-2019, subject to the standard
data filters described in Section 2.5. The tests in columns 1, 2, and 3 correspond to our main results
on application completion, approval, and origination in Table 3. Appendix A lists the controls and
provides variable definitions. The standard errors are two-way clustered at the lender and county
level, and the symbols *, **, and *** indicate significance at the 10%, 5%, and 1% levels,
respectively.
Dependent Variable:
I(Completed)
I(Approved)
I(Origination)
(1)
(2)
(3)
Minority
-0.016
***
(0.001)
-0.017
***
(0.001)
-0.045
***
(0.001)
Minority X Minority Officer
0.015
***
(0.003)
0.008
***
(0.002)
0.031
***
(0.004)
Minority Officer
-0.005
*
(0.003)
-0.002
(0.002)
-0.008
**
(0.004)
Basic App Controls
Y
Y
Y
Extended App Controls
-
Y
-
Underwriting Sys. Rec. FE
-
Y
-
Branch-Year FE
Y
Y
Y
Property County FE
Y
Y
Y
R-Squared
0.077
0.403
0.113
Observations
1,281,735
956,543
1,281,735
65
Table IA7: First Stage and IV Balance Test for the Default Analysis
This table presents the first stage regression and instrument balance test for our IV analysis of
defaults. Column 1 presents the first stage, where the dependent variable is an indicator for the
loan being handled by a minority loan officer, and the key independent variable is our instrument—
the share of FHA loans at the branch that were handled by minority officers on the same day of
the week during the prior 12 weeks. This first stage regression also includes the FHA Controls and
fixed effects used in the IV analysis. Column 2 presents the balance test, where the dependent
variable is our instrument, and the key independent variables are the controls from our IV analysis,
including continuous versions of our binned controls. The p-value reported at the bottom of column
2 is for an F-test of the joint significance of the control variables. The sample includes all FHA
home purchase mortgages originated between 2012 and 2018, subject to the standard data filters
described in Section 2.5, and the requirement that the application date is after the 12
th
week of
2012 and the instrument can be computed. Appendix A lists the controls and provides variable
definitions. The standard errors are two-way clustered at the lender and county level, and the
symbols *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
Minority
Officer
P12 DOW Min.
Off. Share
(1)
(2)
P12 DOW Min. Off. Share
0.035
***
(0.003)
Interest Rate
0.035
(0.028)
Log(Loan Amount)
0.003
(0.035)
FT Buyer
-0.006
(0.023)
Log(Income)
0.021
(0.029)
Credit Score
0.000
(0.000)
Loan-to-Value
-0.043
(0.220)
Debt-to-Income
-0.083
(0.097)
FHA Controls
Y
-
Branch-Month FE
Y
Y
Day-of-Week FE
Y
Y
Property County FE
Y
Y
Observations
2,239,026
2,239,026
First-stage F-stat
59.1
-
p-value on Joint F-test
-
0.783
66
Table IA8: Heterogeneity Across Borrowers, Lenders, and Loan Officers – All Loan Types
This table repeats the regressions from Tables 7 and 8 that examine the cross-sectional variation in the effect
of race/ethnicity on mortgage approval, except here we use the broader HMDA sample, rather than only
FHA applications. Here the sample includes all completed home purchase mortgage applications in the
2018-2019 HMDA database (subject to the standard data filters described in Section 2.5) that are deemed
“high discretion” cases based on the automated underwriting system output. The dependent variable is an
indicator for the application being approved. The regression specifications are the same as those in Tables
7 and 8. However, to improve readability, we report here the coefficients for only the key independent
variables (i.e., the triple interactions, and the Minority X Off. Experience term in column 5). We note that
the remaining interaction terms are included in the specifications, and look similar to those reported in
Tables 7 and 8. Appendix A lists the controls and provides variable definitions. The standard errors are two-
way clustered at the lender and county level, and the symbols *, **, and *** indicate significance at the
10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Approved)
(1)
(2)
(3)
(4)
(5)
Minority
-0.031
***
(0.002)
-0.029
***
(0.002)
-0.029
***
(0.002)
-0.025
***
(0.003)
-0.034
***
(0.003)
Minority X Minority Officer
0.008
**
(0.004)
0.013
***
(0.004)
-0.003
(0.008)
0.018
**
(0.009)
Minority X Min. Off. X Low Inc.
0.014
*
(0.008)
Minority X Min. Off. X Small Bank
0.018
*
(0.010)
Minority X Min. Off. X FinTech
-0.026
**
(0.010)
Minority X Min. Off. X Other Race
0.005
(0.005)
Minority X Min. Off. X Same Race
0.015
***
(0.004)
Minority X Min. Off. X Off. Minority Share
0.024
*
(0.015)
Minority X Min. Off. X Off. Experience
-0.001
(0.001)
Minority X Off. Experience
0.001
**
(0.000)
Remaining Interaction Terms
Y
Y
Y
Y
Y
Basic App Controls
Y
Y
Y
Y
Y
Extended App Controls
Y
Y
Y
Y
Y
Underwriting Sys. Rec. FE
Y
Y
Y
Y
Y
Branch-Year FE
Y
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
Y
R-Squared
0.562
0.562
0.562
0.562
0.562
Observations
544,380
544,380
544,380
544,380
544,380
67
Table IA9: Loan Officers’ Minority Lending and Survival in the Sample
This table presents regressions that examine the effect of loan officers’ minority lending on their
likelihood of remaining in the FHA sample. The sample includes all loan officer-years in the FHA
data from 2012 to 2017 (we collapse the FHA loan-level data to the officer-year level and compute
statistics on minority lending outcomes). The dependent variable is an indicator for the loan officer
remaining in the sample the following year. The key independent variables are an indicator for the
officer being a minority, and its interaction with the officer’s share of loans made to minority
borrowers (Off. Minority Share) and the difference between the officer’s default rates on loans to
minority versus white borrowers (Off. Minority Default Gap). Appendix A provides variable
definitions. The standard errors are clustered by lender, and the symbols *, **, and *** indicate
significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Remain in Sample Next Year)
(1)
(2)
(3)
(4)
Minority Officer
-0.007
(0.005)
-0.052
***
(0.008)
-0.002
(0.004)
-0.002
(0.004)
Off. Minority Share
0.018
***
(0.003)
0.011
***
(0.004)
Min Off. X Off. Minority Share
0.068
***
(0.009)
Off. Minority Default Gap
0.002
(0.003)
0.001
(0.003)
Min Off. X Off. Minority Default Gap
0.009
(0.014)
Officer Experience
0.005
***
(0.000)
0.005
***
(0.000)
0.005
***
(0.000)
0.005
***
(0.000)
Lender-Year FE
Y
Y
Y
Y
R-Squared
0.153
0.154
0.145
0.145
Observations
357,660
357,660
177,745
177,745
68
Table IA10: Differential Reaction to Hard Information – Binned Controls
This table presents results from regressions similar to those in Table 9 exploring loan officers’
differential reaction to hard information in mortgage approval. As in Table 9, the dependent
variable is an indicator for the application being approved, and we examine differences in approval
patterns between minority versus white officers (column 1), experienced versus inexperienced
white officers (column 2), and specialist versus non-specialist minority officers (column 3). The
difference from Table 9 is that here, the control variables are binned following the descriptions in
Appendix A. Therefore, instead of coefficients, this table reports the p-values from Wald tests of
the joint significance of the interaction terms between each set of control bin indicators and the
indicator for the relevant loan officer group. The sample in column 1 is completed FHA home
purchase mortgage applications by minority borrowers in the HMDA database in 2018-2019,
subject to the standard data filters described in Section 2.5. Columns 2 and 3 limit the sample to
applications handled by white and minority officers, respectively.
Dependent Variable:
I(Approved)
Minority Officers
vs.
White Officers
Exp. White Offs.
vs.
Inexp. White Offs.
Specialist Min. Offs.
vs.
Non-Spec. Min. Offs.
(1)
(2)
(3)
Joint Application X INT
0.892
0.350
0.901
Log(Loan Amount) X INT
0.507
0.276
0.482
Income Ratio Bins X INT
0.000
0.000
0.000
Credit Score Bins X INT
0.000
0.004
0.034
Debt-to-Income Bins X INT
0.000
0.000
0.000
Age Bins X INT
0.207
0.485
0.448
Binned Controls
Y
Y
Y
Underwriting Sys. Rec. FE
Y
Y
Y
Branch-Year FE
Y
Y
Y
Property County FE
Y
Y
Y
R-Squared
0.445
0.457
0.439
Observations
401,582
310,304
89,720
69
Table IA11: Reasons for Denial
This table examines the reported reasons for denying a mortgage application. The sample includes
all completed home purchase mortgage applications in the HMDA database in 2018-2019, subject
to the standard data filters described in Section 2.5. Panel A tabulates the frequency of the nine
potential denial reasons available to HMDA-reporting lenders. The columns report frequencies
within the set of all denied applicants, and for denied white and minority applicants, separately.
The regressions in Panel B examine racial differences in denial reason usage. The dependent
variable is an indicator for the application being denied due to “debt-to-income” or “credit history”
in columns 1 and 2, respectively. The key independent variables are an indicator for the borrower
being a minority, an indicator for the loan officer being a minority, and their interaction. The
specifications include fixed effects for each combination of lender, year, loan type, and narrow
bins of debt-to-income ratio (column 1) and credit score (column 2). Appendix A provides variable
and bin definitions. The standard errors are two-way clustered at the lender and county level, and
the symbols *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.
Panel A: Denial Reason Frequencies
Denial Reason
All Denials
(%)
White Denials
(%)
Minority Denials
(%)
Debt-to-Income
31.78
29.45
34.92
Credit History
17.99
17.73
18.34
Collateral
16.09
18.48
12.88
Incomplete Application
9.70
10.61
8.46
Other
8.16
8.12
8.22
Insufficient Cash
6.55
6.59
6.50
Unverifiable Info
5.84
5.10
6.83
Employment History
3.74
3.74
3.73
Mortgage Insurance Denied
0.15
0.17
0.12
Panel B: Racial Differences in Denial Reason Usage
Dependent Variable:
I(Denied for DTI)
I(Denied for Credit History)
(1)
(2)
Minority
0.0057
***
(0.0004)
0.0035
***
(0.0003)
Minority X Minority Officer
-0.0018
***
(0.0006)
-0.0026
***
(0.0005)
Minority Officer
0.0011
***
(0.0004)
0.0011
***
(0.0004)
Lender-Year-Loan Type-DTI Bin FE
Y
-
Lender-Year-Loan Type-FICO Bin FE
-
Y
R-Squared
0.496
0.235
Observations
4,054,501
4,114,910
70
Table IA12: Additional Heterogeneity in the Minority Officer Effect?
This table presents regressions that examine potential heterogeneity in the impact of loan officer
and borrower race on lending outcomes across different types of lenders and competitive
environments. The analysis of mortgage approval (columns 1 and 2) is based on completed FHA
home purchase mortgage applications in the HMDA database in 2018-2019. The analysis of
mortgage default (columns 3 and 4) uses FHA home purchase mortgages originated from 2012 to
2018. Both samples implement the standard data filters described in Section 2.5. The key
independent variables are an indicator for the borrower being a minority, an indicator for the loan
officer being a minority, their interaction, and further triple interactions with indicators for the
lender ranking in the top third in terms of their minority loan officer share (Min. Employer), or for
the application occurring in a county in the top tercile of mortgage market HHI (Low Comp.),
respectively. Appendix A lists the controls and provides variable definitions. The standard errors
are two-way clustered at the lender and county level, and the symbols *, **, and *** indicate
significance at the 10%, 5%, and 1% levels, respectively.
Dependent Variable:
I(Approved)
I(Default)
(1)
(2)
(3)
(4)
Minority
-0.016
***
(0.002)
-0.018
***
(0.001)
0.020
***
(0.001)
0.019
***
(0.002)
Minority Officer
-0.002
(0.004)
-0.003
(0.003)
0.003
(0.002)
0.003
**
(0.002)
Minority X Minority Officer
0.007
*
(0.004)
0.007
**
(0.003)
-0.021
***
(0.002)
-0.023
***
(0.002)
Minority X Min. Off. X Min. Employer
0.002
(0.004)
0.001
(0.003)
Minority X Min. Employer
-0.002
(0.002)
-0.004
*
(0.002)
Min. Off. X Min. Employer
-0.001
(0.004)
-0.002
(0.003)
Minority X Min. Off. X Low Comp.
-0.000
(0.005)
0.008
*
(0.004)
Minority X Low Comp.
0.002
(0.002)
-0.004
**
(0.002)
Min. Off. X Low Comp.
-0.000
(0.004)
-0.003
(0.003)
Basic App Controls
Y
Y
-
-
Extended App Controls
Y
Y
-
-
Underwriting Sys. Rec. FE
Y
Y
-
-
Branch-Year FE
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
FHA Controls
-
-
Y
Y
Origination Month FE
-
-
Y
Y
R-Squared
0.403
0.403
0.103
0.103
Observations
956,543
956,543
3,370,855
3,370,855
71
Table IA13: Loan Officer Race and Loan Pricing
This table presents regressions that examine the effect of loan officers on mortgage loan pricing.
The dependent variable is the interest rate (in percentage point units), and we control directly for
non-rate components of pricing with Net Discount Points. The key independent variables are an
indicator for the borrower being a minority, an indicator for the loan officer being a minority, and
their interaction. Columns 1 and 2 present the OLS results for the low- and high-discretion samples
(based on the underwriting system output), respectively. Columns 3 and 4 present the instrumental
variables tests that use the share of applications at the branch that were handled by minority officers
on the same day of the week during the prior 12 weeks to instrument for a minority officer handling
the loan. The sample includes all originated home purchase mortgages in the HMDA database in
2018-2019, subject to the standard data filters described in Section 2.5, and for the IV tests the
requirement that the application is opened after the 12
th
week of 2018. Appendix A lists the
controls and provides variable definitions. The standard errors are two-way clustered at the lender
and county level, and the symbols *, **, and *** indicate significance at the 10%, 5%, and 1%
levels, respectively.
Dependent Variable:
Interest Rate
OLS
IV
Low
Discretion
High
Discretion
Low
Discretion
High
Discretion
(1)
(2)
(3)
(4)
Minority
-0.000
(0.001)
-0.001
(0.004)
-0.002
(0.002)
-0.003
(0.006)
Minority X Minority Officer
0.007
**
(0.003)
-0.010
(0.008)
0.011
(0.021)
0.010
(0.060)
Minority Officer
0.008
***
(0.003)
0.017
***
(0.005)
0.008
(0.078)
-0.044
(0.277)
Net Discount Points
-0.156
***
(0.005)
-0.120
***
(0.008)
-0.157
***
(0.006)
-0.120
***
(0.009)
Basic App Controls
Y
Y
Y
Y
Extended App Controls
Y
Y
Y
Y
Underwriting Sys. Rec. FE
Y
Y
Y
Y
Branch-Week FE
Y
Y
Y
Y
Day-of-Week FE
Y
Y
Y
Y
Property County FE
Y
Y
Y
Y
R-Squared
0.805
0.871
-
-
Observations
2,614,823
162,732
2,233,434
138,123
First-stage F-stat
-
-
35.7
5.3
