Full text of Economic Insights (Federal Reserve Bank of Philadelphia) : First Quarter 2022, Volume 7, Issue 1

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

First Quarter 2022
Volume 7, Issue 1

Banking Trends
Meet the New Borrowers
The Costs and Benefits of
Fixing Downtown Freeways
Q&A
Research Update
Data in Focus

Contents
First Quarter 2022

A publication of the Research
Department of the Federal
Reserve Bank of Philadelphia
Economic Insights features
nontechnical articles on monetary
policy, banking, and national,
regional, and international
economics, all written for a wide
audience.
The views expressed by the authors are not
necessarily those of the Federal Reserve.
The Federal Reserve Bank of Philadelphia
helps formulate and implement monetary
policy, supervises banks and bank and
savings and loan holding companies, and
provides financial services to depository
institutions and the federal government. It
is one of 12 regional Reserve Banks that,
together with the U.S. Federal Reserve
Board of Governors, make up the Federal
Reserve System. The Philadelphia Fed
serves eastern and central Pennsylvania,
southern New Jersey, and Delaware.

1

Q&A…

2

Banking Trends

9

Meet the New Borrowers

17

The Costs and Benefits of Fixing Downtown Freeways

23

Patrick T. Harker
President and
Chief Executive Officer

Volume 7, Issue 1

29

with Jeffrey Brinkman.

Has automated underwriting banished racial discrimination from the mortgage
markets? Edison Yu provides some new evidence to help answer this question.

Igor Livshits discovers that lenders may be aggregating dispersed information
about new borrowers, giving those borrowers access to credit even though they
lack a history of repayment.

Urban freeways played a key role in the growth of suburbia—and in the decline of
central cities. Jeffrey Brinkman and Jeffrey Lin calculate the costs and benefits of
burying those urban freeways.

Research Update
Abstracts of the latest working papers produced by the Philadelphia Fed.

Data in Focus
ATSIX

Michael Dotsey
Executive Vice President and
Director of Research
Adam Steinberg
Managing Editor, Research Publications
Brendan Barry
Data Visualization Manager

ISSN 0007–7011

Connect with Us
We welcome your comments at:
PHIL.EI.Comments@phil.frb.org
E-mail notifications:
www.philadelphiafed.org/notifications
Previous articles:
https://ideas.repec.org/s/fip/fedpei.html
Twitter:
@PhilFedResearch
Facebook:
www.facebook.com/philadelphiafed/
LinkedIn:
https://www.linkedin.com/company/
philadelphiafed/

About the Cover
Philadelphia from the Art Museum
The cover photo of Philadelphia encompasses almost every stage of the city’s
development. In the midground, partially obscured by the statue, is Logan Square,
one of the five squares city founder William Penn and his surveyor, Thomas Holme,
included in their 1687 city plan. In the distance looms City Hall, a testament to
the post-Civil War boom; it was the tallest building in the world upon its completion
in 1894. In the near distance is Rudolf Siemering’s 1897 Washington Monument.
The monument and City Hall are connected by the Benjamin Franklin Parkway, the
City’s post-World War I attempt to tame the unruly, industrial metropolis with
open vistas, museums, and classical architecture. Starting in the 1970s, a downtown building boom led to the construction of numerous skyscrapers, visible
to the right. Few images so fully convey Philadelphia’s complex and varied history,
culture, and economy.
Photograph by Brendan Barry.

Q&A…

with Jeffrey Brinkman,
a senior economist here
at the Philadelphia Fed.

What led you to study engineering
in college?
I was more interested in physics, but engineering seemed like a practical form
of physics, with a job waiting at the end of
college. But my first job after college,
I wasn’t doing research and design. I was
in quality control. That environment
was less interesting to me.

Is designing a model in economics
similar to the kind of work you were
hoping to do in engineering?
Yes. A lot of people don’t realize that we
do a lot of math and computational modelling in economics. When I got to grad
school in economics, I just had to learn
a new vocabulary. The actual work is very
similar to engineering.

Jeffrey Brinkman
Senior economist Jeffrey Brinkman grew
up outside Columbus, Ohio, the son
of an engineer and a high school math
teacher. He studied electrical engineering
at The Ohio State University before
switching to public policy and then
economics, earning his doctorate from
Carnegie Mellon University in 2011.
For the past 11 years, he has researched
and written about urban economics and
the local consequences of policymaking
for the Philadelphia Fed.

It sounds like what drew you to both
engineering and economics was the
opportunity to solve problems.
Yes. There’s nothing better than writing
down a mathematical model and trying
to solve it on the computer. It’s a very
focused activity.

You’ve lived in Los Angeles, Pittsburgh,
Detroit, and Philadelphia. How did living in these different cities shape your
thinking about urban economics?
Before I moved to Los Angeles, I thought,
if you want walkable neighborhoods,
all you need is density. Well, Los Angeles
is one of the densest cities in the country,
and yet it’s very auto dependent. There
are other dimensions for cities besides
density—things like, how the streets are
laid out, whether the city was built in the
19th or 20th century. Cities built today
tend to have less transit infrastructure
because now we have cars. All these dimensions matter. A city isn’t just your standard
model with a central business district
surrounded by residences.

Models are supposed to be applicable
to different situations, but you’re also
pointing out that every city is unique.
How do you reconcile your models
with all these differences among cities?
Models should make our thinking more
Q&A

2022 Q1

concrete, so that we all know what we’re
talking about, but they should also allow
us to measure differences. Like in our
article about freeways. Our model helped
us measure the size of the negative effects
of freeways on central cities. As our models get more sophisticated, they capture
that heterogeneity, but more sophistication
means more complication and maybe less
clarity of what they’re trying to tell us.
So, there’s a tradeoff between “let’s try to
model everything” and “let’s have a simpler model where I can get intuition about
what’s going on.”

For your article, you applied costbenefit analysis to the proposal to cap
I-95 through central Philadelphia. Did
local transportation authorities use
cost-benefit analysis when designing
these freeways in the first place?
They did, but it was all about, what are the
transportation benefits of these highways?
How do we get people into and out of
the city? How do we move goods? They
weren’t considering these big negative
effects on central cities. Even urban mayors at the time were like, this is going to
revive the city, this is going to bring people
into the city. But the exact opposite
happened. The highways took people out
of the city. They allowed people to live
farther away, and because there are these
big negative amenity effects for the neighborhoods nearby, that pushed people out
of the city, too. People quickly realized
that this was a problem. It led to protests
everywhere.

It sounds like the public blowback was
in part a response to the unquestioned
assumptions of the planners in their
modelling and cost-benefit analysis.
That’s one of the things I enjoy. I love
identifying unintended consequences.
People have been yelling about this for
years, but I think we’re among the first
economists to quantify these freeway
disamenities. Learning how to look at the
data is important. It’s not just, “locations
near freeways declined.” It’s, “locations in
central cities near freeways declined.”
You have to get into the model and think
about the economics of it to understand
how to look at the data.
Federal Reserve Bank of Philadelphia
Research Department

1

Photo: fizkes/iStock

Banking Trends

Discrimination in
Mortgage Markets

Automated underwriting may reduce but likely
does not end discrimination against racial minorities.
Edison Yu
Economic Advisor
and Economist
Federal Reserve Bank
of Philadelphia
The views expressed
in this article are not
necessarily those of
the Federal Reserve.

2

R

acial discrimination has haunted mortgage
markets for decades, prompting legislation and public policy debates that
continue to shape how all of us get our mortgages. A new technology may help reduce this
discrimination.
Previous research into mortgage markets has
shown evidence of a century’s worth of racial
discrimination, including redlining, unequal
mortgage access, and differences in mortgage
costs. Federal, state, and local governments have
responded to this discrimination by enacting
laws such as the federal Fair Housing Act of 1968.1
These laws are not relics of a vanished era.
Many banks were recently fined a large amount
of money due to antidiscrimination lawsuits.
For example, in 2019, Wells Fargo Bank wrote
the City of Philadelphia a check for $10 million
to settle a lawsuit alleging that the bank engaged in discriminatory lending practices.2
Some analysts argue that algorithmic or automated underwriting (AU), which has become
increasingly popular in mortgage markets,

Federal Reserve Bank of Philadelphia
Research Department

renders lenders less likely to discriminate
because it does not use race as an input and
presumably bases loan decisions only on the
applicant’s financial data, limiting the discretionary judgement of human decisionmakers. If
that’s the case, then AU may help antidiscrimination efforts. This is why we need to study
AU’s impact on the mortgage lending business.
Researchers disagree as to how best to
measure racial discrimination in the mortgage
markets, so, before studying AU’s impact on
antidiscrimination efforts, I survey the history of
the statistical methods used to identify racial
discrimination. Although it may seem straightforward, identifying racial discrimination is
challenging since researchers cannot observe
all the information used by loan officers or
borrowers. I then describe how, in a previously
published working paper, my coauthors
and I addressed this challenge by using highfrequency data. I conclude by exploring
preliminary evidence on the effects of AU on
racial discrimination in the mortgage markets.

Banking Trends: Discrimination in Mortgage Markets
2022 Q1

The Mortgage Application Process

rely on the AU-generated recommendation; a human underwriter
In the U.S, a mortgage application typically starts with a borrower still makes the final loan decision. AU decisions are only recomcontacting a potential lender to inquire about mortgage products. mendations. So, the final loan decision may still be biased.
The borrower usually has an initial conversation with a loan
officer, who is the front person within the lender organization
responsible for communicating with the borrower. The loan officer Identification of Racial Discrimination
can gauge the borrower’s initial eligibility by using the borrower’s
Using Observational Data
basic financial information, such as the loan amount needed and Social scientists have documented racial disparities in a wide
credit scores, and guide the borrower to select a mortgage
range of areas, including in labor markets, credit markets, and
product. Once a mortgage product is selected, the borrower can
the legal system. Because of the limitations of empirical tests,
submit a formal mortgage application.
social scientists disagree as to whether or not these disparities
Increasingly in recent years, borrowers have contacted poare the result of discrimination by economic decision makers.
tential lenders over the Internet.3 Borrowers search for potential
The two main types of tests used to identify racial discriminamortgage products on mortgage-shopping platforms such as
tion are benchmarking tests and outcome tests. Benchmarking
Zillow or with fintech lenders such as Quicken Loans.4 After the
tests (also known as audit tests) use observational data in
initial search, a borrower is contacted by the lender, usually
a straightforward way to test for discrimination. In the context
by the lender’s loan officer. Even at fintech lenders, human loan
of mortgage lending, if, after adjusting for credit risks, a minority
officers are involved in the application process.
group receives a lower average approval rate or higher interest
The loan officer’s involvement does not end with the initial
rate for the same mortgage product, then the test has identified
contact. After a borrower submits a formal application, the loan
discrimination. Benchmarking tests are useful because they
officer ensures that the borrower has submitted all the necessary can be executed in real time.8 However, benchmarking tests are
documentation, such as verification of
vulnerable to omitted-variable bias. The
income and employment, credit reports,
omitted variables are the differences in
and property appraisal reports. The officer
group characteristics that the researcher
also ensures the accuracy of the infordoes not observe but that can cause
mation in the application. Once the
differences in evaluations. For example,
application is complete, the loan officer
minority applicants might have riskier
sends it to a loan underwriter, who
financial profiles, and the approval rate gap
makes the final credit decision based on
seen in the data could reflect differences
the application and supporting docuin credit risk rather than discrimination.
ments. Although loan officers do not make
One common approach for solving this
final credit decisions, they influence the
problem is to include additional variables
potential credit decision by nudging
as control variables in the analysis.
the borrower to provide timely and accurate documentation. Thus, For example, one can look at the mortgage approval gap when
a loan officer’s racial bias can still affect the outcome of a mortcomparing minority and majority applicants while controlling for
gage application. For example, a loan officer might not inform
financial variables such as income and credit risk. However, no
a minority applicant of an incomplete application in a timely
researcher has data on all the variables observed by decision
manner, leading to rejection of the application by the underwriter. makers or borrowers. For example, researchers may not have the
Many lenders increasingly rely on AU, which was used in
“soft” information that loan officers have from their interactions
about 56 percent of mortgage applications in 2019. An AU system with borrowers. In antidiscrimination lawsuits where minority
processes an applicant’s financial information and recommends
applicants charge that they received a higher interest rate or
whether to approve the loan. This recommendation is generated
larger fees, an oft-used counterargument is that minority appliby a computer, not a human. These underwriting systems do
cants do not shop around as much as majority applicants. That
not use race as an input and presumably base loan decisions
kind of data is not generally available to researchers, regulators,
only on the applicant’s financial data. Antidiscrimination regula- or even lenders. So, although it helps to include more control
tions allow lenders, when making loan decisions, to use variables
variables, they cannot eliminate the omitted-variable bias problem.
directly related to credit history and risks, but they prohibit
An alternative is to use an outcome test. Instead of comparing
lenders from discriminating on the basis of race, religion, color,
differences in how groups are evaluated, such as approval
national origin, sex, familial status, or disability status.5
rates, outcome tests compare the subsequent performance of
However, there are a few reasons why the final credit decision successful applicants through, for example, default rates.9
might still be biased. Algorithms may produce decisions that
Suppose there is a cutoff of application quality, below which
unintentionally correlate with impermissible variables (such as
the application will be rejected. The applicant who just barely
race and gender).6 Since the algorithms can be quite complicated, meets the cutoff is the marginal applicant. If there is discriminadecision makers may not understand that the algorithms are
tion, the minority group will face a higher threshold for inclusion,
making biased underwriting choices.7 In addition, a loan officer’s and the marginal minority applicant will thus have better
bias may affect the timeliness and completeness of the informaex post outcomes (for example, a lower default rate) than the
tion input into the AU system. Finally, lenders do not completely
marginal majority applicant.

AU may help antidiscrimination efforts.
This is why we need
to study AU’s impact
on the mortgage
lending business.

Banking Trends: Discrimination in Mortgage Markets

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

3

Though intuitively appealing, outcome tests are notoriously
difficult to implement. We rarely have data that enable us to
identify the marginal applicant, so researchers usually use average
quantities (such as the average default rate) instead. But the
average difference in ex post outcomes can be a poor approximation of the difference in marginal outcomes. Since we observe
only average performance, an observation that the default rate
among African Americans is the same as among whites, for
example, should not be interpreted as evidence of nondiscrimination.10 Studies using this approach are less likely to find evidence
of discrimination.11 Thus, implementing an outcome test is very
challenging, and many papers in the literature have pointed out
its limitations.12 For that reason, I focus on benchmarking tests.

Evidence of Racial Discrimination in the
Mortgage Markets

There is a long history of attempts to identify racial discrimination
in the mortgage markets using observational data. Many of these
papers try to solve the omitted-variable problem by including
more control variables in the statistical analysis.
This literature can be traced back at least as far as the 1996
work of the former director of research at the Federal Reserve
Bank of Boston, Alicia Munnell, and her coauthors. In that paper,
the authors use the 1990 Home Mortgage Disclosure Act (HMDA)
data, but they also use information from a survey that collects
additional information from lending institutions operating in the
Boston metropolitan statistical area.13 This additional information
includes financial, employment, and property characteristics
relevant to a lending decision but missing from the HMDA data.
The survey sample covers all applications for conventional mortgage loans made by African American and Hispanic American
applicants and a random sample of 3,300 applications made by
white applicants. When using the HMDA data alone, the paper
finds that the rejection rate of minority applicants is 18 percentage
points higher than that of white applicants. When the researchers
controlled for the additional information from the survey, the
disparity between the rejection rates of minority and white
applicants declined to just over 8 percentage points. These results
show the importance of controlling for relevant variables absent
from the HMDA data set. Still, the rejection rate gap remains
large even after adding the controls. Many early papers in the
literature show similar results.14
In a recent paper, University of California, Berkeley, law professor Robert Bartlett and his coauthors examine whether
African Americans and Hispanic Americans pay higher mortgage
interest rates than white Americans, and whether this pricing
differential remains when the origination is automated. To
address the omitted-variable problem, they merged 2009–2015
HMDA data with other data sets that include information about
interest rates, the names of lenders, and loan performance. In
addition, by using a sample of mortgages insured by a governmentsponsored enterprise (GSE), they filter out the default and
prepayment risks borne by the lenders. Thus, any disparity in
the interest rates paid by minority and majority borrowers should
reflect racial discrimination, not credit risk. They find that
Hispanic American and African American borrowers collectively

4

Federal Reserve Bank of Philadelphia
Research Department

pay an additional 7.9 and 3.6 basis points in interest rates for,
respectively, purchase mortgages and refinance mortgages. They
also find a 40 percent lower level of price discrimination if the
lender is a fintech firm, which suggests that AU reduces but does
not eliminate discrimination.
However, in their 2021 article, Federal Reserve economists
Neil Bhutta and Aurel Hizmo account for more pricing variables,
such as discount points and fees, and find no evidence that
minorities pay more in mortgages. They supplement the 2014–2015
HMDA data with administrative data from the Federal Housing
Administration (FHA) on all FHA-insured mortgages, and with
information on points and fees from Optimal Blue, a leading
provider of secondary marketing solutions and data services
in the mortgage industry. They find a statistically significant gap in
interest rates paid by race, but the gap is offset by differences
in discount points. They argue that the differences in interest
rates across racial groups found in the earlier papers are a result
of African American borrowers choosing mortgage products
with higher interest rates but lower points, potentially because
minority borrowers may find it more difficult to put funds up
front. This finding is restricted to FHA mortgages; whether
the results generalize to other samples, such as the more GSEdominated sample used by Robert Bartlett and his coauthors,
remains unclear. Notably, lenders who provide FHA mortgage
products tend to serve lower-income and minority communities
and thus may be less biased.
In another recent paper, Penn State professor of real estate
Brent Ambrose and his coauthors find that pricing disparities
in mortgage contracts are influenced by whether the borrower
and broker are of the same race. They used a novel data set
that covers all mortgages approved and funded by New Century
Financial Corporation, a now-defunct real estate investment
trust, between 2003 and 2007. These loans are representative of
the overall subprime market before the Great Recession. This
data set comprises more than 300,000 mortgages originated
by 124,736 individual brokers, and it contains a rich set of control
variables. In addition, the data set includes the names of the
brokers. When the authors used a surname-geocoding algorithm
to infer each broker’s race, they found that minorities pay more
in fees than similarly qualified whites, but the premium paid
by minorities depends on whether the broker shares their race.
For example, African American borrowers who obtain a loan
through a white mortgage broker pay 14 percent more than
white borrows who work with a white broker, but this premium
is lowered to 6 percent when the broker is African American.
For a 2021 working paper, Neil Bhutta and his coauthors studied
the impact of AU on racial discrimination using a newly available
data set from HMDA. This 2018–2019 HMDA data set provides
a longer list of variables—including credit scores, debt-to-income
ratios, and AU recommendations—than did earlier HMDA data
sets. When they focused on the sample of loans that utilized the
AU system, they found that, by controlling for AU recommendations and using these new loan-level variables, the racial gap in
mortgage denial rates fell to about 1–2 percentage points. They
argue that the remaining gap might be explained by unobserved
characteristics of the borrowers, which suggests a more limited
role for racial discrimination in mortgages that use AU.

Banking Trends: Discrimination in Mortgage Markets
2022 Q1

All of these researchers find that the
racial gap in mortgage approval rates and
costs is very large in the data, but some of
this gap can be explained by factors such
as credit risk. The question is whether the
remaining racial gap is caused by racial
bias or by insufficient control of omitted
variables. Many of these papers attempt
to reduce the problem of omitted variables
by adding control variables to the analysis.
However, it is difficult to know whether
the additional variables eliminate the
bias. In addition, these and other papers
use samples across different data sets or
cover different time periods. This makes
it difficult to compare results.

Testing for Discrimination
Using High-Frequency Data

In a recent research paper, my coauthors
and I took a different approach to address
the challenges of identifying racial discrimination.15 With some assumptions, this
approach avoids the problem of omitted
variables by employing high-frequency
data. We used time variation in loan
officers’ loan approval decisions to draw

inferences about likely discriminatory behavior. We also used the entire HMDA data
set from 1994 through 2019, which covers
most mortgage applications in the U.S.
during those 25 years, making our sample
more comprehensive than samples used
in earlier work. After discussing this new
approach, I will show how we used this
new approach to ascertain the impact of
AU on discrimination.
First, we find that the volume of
mortgage originations increases over the
course of a calendar month (Figure 1).16
The number of loans originated on the
last day of a month is almost twice as high
as on the first day of the month. There
is no similar pattern in application volume.
This bunching pattern in originations is
likely caused by loan officers’ incentive to
meet their month-end quota. Loan officers
tend to receive a commission that equals
a percentage of the total dollar amount
they originate during the month. They can
also receive a bonus for meeting their
monthly origination target. Loan officers
who fail to meet volume targets can be
disciplined and risk getting fired. Our
key insight is that this pressure to meet

month-end quotas makes it costlier for
loan officers to discriminate at the end
of a month.
At the same time, we observe that the
mortgage approval gap between white
and African American applicants shrinks
over the course of the month (Figure 2).
In the first seven days of a calendar
month, the approval rate gap is close to
20 percent. The gap shrinks in the last
days of the month and reaches the lowest
point of around 10 percent on the last day
of the month. When we control for many
observable variables, the approval rate
gap shrinks to almost zero on the last day
of the month (Figure 3).
The higher-frequency daily data help us
address the omitted-variable bias. In our
paper, we discuss a number of potential
omitted variable issues. The reduction
in the approval gap within a month, as
seen in Figure 3, might be attributed to an
unobserved within-month movement of
application quality rather than changes in
discrimination. For example, the gap
would be explained without reference to
discrimination if the quality of African
American applications is higher toward

FIGURE 1

FIGURE 2

Originations Surge as the Month Ends

The Approval Rate Gap Shrinks
Toward the End of the Month

Loan officers approve more applications toward the end of the month,
most likely to meet their monthly quotas.

Daily loan applications and originations, as a percentage of loan applications and originations
on the first day of the month, 1994–2019
First of the month
End of the month
Beginning of the month

200%

Approval rate for African American applicants minus
rate for white applicants in the seven days preceding
and succeeding the first of the month, 1994–2019

0.00

End of the month
Beginning of…
−7
0 1
7

Origination
150%

−0.05

100%

−0.10

50%

−0.15
Application

0%
−0.20
−50%

−7

−6

−5

−4

−3

−2

−1

0

1

2

3

4

5

6

7

−0.25

Source: Home Mortgage Disclosure Act (HMDA) data set, Board of Governors of the Federal Reserve System.
Note: The number 0 on the horizonal axis indicates the last day of a calendar month, the positive numbers
indicate the first seven days of the month, and the negative numbers indicate the last days of the month.

Banking Trends: Discrimination in Mortgage Markets

2022 Q1

Source: Home Mortgage Disclosure Act (HMDA) data
set, Board of Governors of the Federal Reserve System.
Note: This figure does not consider observable factors
that might affect the approval decision.

Federal Reserve Bank of Philadelphia
Research Department

5

the end of a month.17 But we do not see
a within-month bunching pattern in application composition and observed applicant
quality in the data. Minority borrowers do
not seem more likely to submit applications toward the end of the month. Nor
do we find evidence that omitted variables—
including application quality—change
within the month. For example, the share
of applicants with an income lower than
the county median is stable over the
course of the month. This income test
serves as a proxy for other potential
differences between applicants. Furthermore, the ex post default rate gap doesn’t
vary over the course of the month. Our
findings suggest that the shrinking
approval rate gap is likely not caused by
application- or applicant-related factors.18
Therefore, using the high-frequency
data allows us to attribute the decline in
discrimination to loan officers rushing to
meet their monthly quotas.

We find that the time-varying discrimination explains about 3.5–5 percentage
points of the approval rate gap, which is
about half of the unexplained approval
rate gap of 7 percent after controlling for
observable loan-level characteristics.
Our research also enables us to test the
theory that AU reduces discrimination
in the mortgage markets. We find that
the gap in AU recommendations is nearly
constant over the course of the month,
which suggests less racial bias in AU
decisions. Nonetheless, the approval rate
gap of human-made decisions decreases
for lenders that use an AU system (though
not as much as for lenders that do not).19
This implies that there can still be racial
bias when a human is making the approval
decision, even after receiving an AU
recommendation (Figure 4). Consistent
with previous studies, our research shows
that AU seems to reduce but does not
eliminate the racial gap in approval rates

FIGURE 3

FIGURE 4

The Approval Rate Gap Shrinks to
Nearly Zero When We Control for
Other Variables

Even When Loan Officers Use AU,
They Tend to Reject African
American Applicants More Often
Than AU Recommends

Approval rate for African American applicants minus
the rate for white applicants in the seven days
preceding and succeeding the first of the month, controlling for several observable variables, 1994–2019
End of the month
Beginning of…
0 1
7
−7
0.00
with controlling for
observable variables

AU’s approval rate gap between African American
and white applicants versus the actual approval rate
gap among lenders who use AU, 2018–2019
End of the month
Beginning of…
0 1
7
−7
0.00

−0.05

−0.05

−0.10

−0.10

−0.15

without controlling for
observable variables

−0.15

−0.20

−0.20

−0.25

−0.25

Source: Home Mortgage Disclosure Act (HMDA) data
set, Board of Governors of the Federal Reserve System.

AUS acceptance gap

Approval gap

Source: Home Mortgage Disclosure Act (HMDA) data
set, Board of Governors of the Federal Reserve System.

in the mortgage markets. One possible
criticism of our methodology is that the
within-month variation in the approval
gap merely reflects differences in how
long it takes to complete the origination
process. For example, the shrinking
approval rate gap can be a result of
African American borrowers being more
likely to settle their housing transactions
(and hence mortgage applications) at
the end of a month. But we find little
evidence that racial differences in the time
between application and origination vary
within the month.

Conclusion

Researchers have long documented racial
discrimination in the mortgage markets, and that literature is growing as AU
prompts them to study its impact on
antidiscrimination efforts in the mortgage-lending business. In this article,
I summarize work by some earlier and
more recent researchers who studied
racial discrimination in the mortgage
markets. Except for Bhutta and Hizmo’s 2021 article, most papers find that
there is at least some racial bias in the
mortgage markets.
Papers in the literature attempt to reduce the problem of omitted variables
largely by adding more control variables to
the analysis. However, it is difficult to know
whether the additional variables eliminate
the bias. My research shows an approach
that could solve the problem of omitted
variables by using high-frequency data.
A 2021 working paper by Bhutta, Hizmo,
and Federal Reserve economist Daniel
Ringo, as well as my research using the
high-frequency data approach, both show
that AU seems to reduce but not eliminate
the racial approval rate gap in the mortgage markets. Based on these findings,
policymakers might want to encourage the
use of AU to help reduce racial discrimination. However, data are available only for
the last few years, so research in this area
is still relatively new. Further research is
needed to confirm our findings.

Note: Figure 3 plots the average approval rate residual
gap from a regression of approval rates on loans and
application characteristics from HMDA, such as loan
amount, applicant income, and whether the loan was
for purchase or refinance.

6

Federal Reserve Bank of Philadelphia
Research Department

Banking Trends: Discrimination in Mortgage Markets
2022 Q1

Notes
1 Some analysts argue that such regulations are unnecessary
and that market competitive pressures undermine the desire
to discriminate. See, for example, Becker (1957).

when the ex post default rate is higher for the minority group
while there is discrimination, contrary to what the outcome
test would suggest if we observe decision thresholds for the
marginal borrowers.

2 See McCabe (2019).
11 See, for example, Berkovec et al. (1998).
3 See Buchak et al. (2018) and Fuster et al. (2019) for
examples.
4 Fintech lenders use innovative technology designed to
outperform traditional financial methods in the delivery of
financial services.
5 According to the law, both “taste-based” and “statistical”
discrimination are illegal. In economics, taste-based discrimination refers to discrimination as a result of prejudice,
while statistical discrimination refers to decisions that unintentionally correlate with impermissible variables.

12 See Ayres (2002) and Canay et al. (2020) for examples.
13 HMDA data are among the earliest and most comprehensive mortgage application data sets in the U.S. One
of Congress’s goals in enacting the HMDA in 1975 was to
identify possible discriminatory lending patterns in the
data collected. HMDA data are also used for Community
Reinvestment Act bank exams.
14 See Ladd (1998) for a survey of the older literature.
15 See Giacoletti et al. (2021).

6 This would constitute statistical discrimination.
7 See Fuster et al. (forthcoming) for an example.
8 This is in contrast to results from experimental studies,
which are more difficult to implement.
9 See Becker (1957).
10 Suppose that there are two, easily distinguishable types of
white mortgage applicants: those who have a 1 percent
chance of defaulting on a mortgage, and those who have
a 50 percent chance. Similarly, assume that African American
applicants have either a 5 percent or 50 percent chance of
defaulting. If lenders are biased and approve white applicants
who have a default rate of no more than 10 percent and
African American applicants who have a default rate of no
more than 5 percent, these decisions will generate observed
ex post average default rates of 1 percent for white borrowers
and 5 percent for African American borrowers. This is a case

16 The figure shows the average loan origination number by
calendar days of month. We can also restrict the end of the
month to be immediately before the beginning of the month,
but the results would look very similar.
17 Similarly, another example that can explain the timevarying approval rate gap is changing underwriting standards.
But as shown later in this article, there is no evidence of
changes in underwriting standards or application quality
within a month.
18 Some regressions in our paper control for additional
variables, such as credit scores and low-documentation
status, by using a sample that merges HMDA and Black
Knight McDash data. The results are similar.
19 The acceptance rate gap of AU shrinks by about 1–2
percentage points within-month, and the approval rate gap
decreases by about 6–7 percentage points within-month.

References
Ambrose, Brent, James Conklin, Luis Lopez. “Does Borrower
and Broker Race Affect the Cost of Mortgage Credit?” Review
of Financial Studies, 34:2 (2021), pp. 790–826, https://doi.org/
10.1093/rfs/hhaa087.
Ayres, Ian. “Outcome Tests of Racial Disparities in Police
Practices,” Justice Research and Policy, 4:1-2 (2002), pp.
131–142, https://doi.org/10.3818%2FJRP.4.1.2002.131.
Becker, Gary. The Economics of Discrimination, 2nd edition.
Chicago: The University of Chicago Press, 1957.
Banking Trends: Discrimination in Mortgage Markets

2022 Q1

Berkovec, James, Glenn Canner, Stuart Gabriel, and Timothy
Hannan. “Discrimination, Competition, and Loan Performance
in FHA Mortgage Lending,” Review of Economics and
Statistics, 80:2 (1998), pp. 241-250, https://doi.org/10.1162/
003465398557483.
Bartlett, Robert, Adair Morse, Richard Stanton, and Nancy
Wallace. “Consumer-Lending Discrimination in the Fintech
Era,” National Bureau of Economic Research Working Paper
25943 (2019), https://doi.org/10.3386/w25943.

Federal Reserve Bank of Philadelphia
Research Department

7

Bhutta, Neil, and Aurel Hizmo. “Do Minorities
Pay More for Mortgages?” Review of Financial
Studies, 34:2 (2021), pp. 763–789, https://doi.
org/10.1093/rfs/hhaa047.
Bhutta, Neil, Aurel Hizmo, and Daniel Ringo.
“How Much Does Racial Bias Affect Mortgage
Lending? Evidence from Human and Algorithmic
Credit Decisions,” working paper (2021).
Buchak, Greg, Gregor Matvos, Tomasz Piskorski,
and Amit Seru. “Fintech, Regulatory Arbitrage,
and the Rise of Shadow Banks,” Journal of
Financial Economics, 130:3 (2018), pp. 453–483,
https://doi.org/10.1016/j.jfineco.2018.03.011.
Canay, Ivan, Magne Mogstad, and Jack
Mountjoy. “On the Use of Outcome Tests for
Detecting Bias in Decision Making,” National
Bureau of Economic Research Working Paper
27802 (2020), https://doi.org/10.3386/w27802.
Fuster, Andreas, Matthew Plosser, Philipp
Schnabl, and James Vickery. “The Role of
Technology in Mortgage Lending,” Review of
Financial Studies, 32:5 (2019), pp. 1854–1899,
https://doi.org/10.1093/rfs/hhz018.
Fuster, Andreas, Paul Goldsmith-Pinkham, Tarun
Ramadorai, and Ansgar Walther. “Predictably
Unequal? The Effects of Machine Learning on
Credit Markets,” Journal of Finance, forthcoming.
Giacoletti, Marco, Rawley Heimer, and Edison Yu.
“Using High-Frequency Evaluations to Estimate
Discrimination: Evidence from Mortgage Loan
Officers,” Federal Reserve Bank of Philadelphia
Working Paper 21-04/R (2021), https://doi.org/
10.21799/frbp.wp.2021.04.
Ladd, Helen. “Evidence on Discrimination in
Mortgage Lending,” Journal of Economic
Perspectives, 12:2 (1998), pp. 41–62, https://doi.
org/10.1257/jep.12.2.41.
McCabe, Caitlin. “Wells Fargo to Pay Philly $10
million to Resolve Lawsuit Alleging Lending
Discrimination Against Minorities,” Philadelphia
Inquirer, December 16, 2019.
Munnell, Alicia, Geoffrey Tootell, Lynn Browne,
and James McEneaney. “Mortgage Lending in
Boston: Interpreting HMDA Data,” American
Economic Review, 86:1 (1996), pp. 25–53.

8

Federal Reserve Bank of Philadelphia
Research Department

Banking Trends: Discrimination in Mortgage Markets
2022 Q1

Photo: LaylaBird/iStock

Meet the New Borrowers

Credit history is critical for credit access, and it’s more than just
a history of repayment.

Igor Livshits
Economic Advisor
and Economist
Federal Reserve
Bank of Philadelphia
The views expressed
in this article are not
necessarily those of
the Federal Reserve.

F

inancially responsible households benefit from
access to credit.1 The ability to borrow against
future income helps these households buy
their homes, invest in their education, and maintain their preferred level of consumption despite the
occasional shock to their income.2
Credit access is an important part of the conversation when it comes to social inequality and
discrimination. When certain social groups can’t access credit, it contributes to and perpetuates inequality
in overall economic outcomes. For example, exclusion
from (affordable) mortgages is a barrier to homeownership and geographic mobility, both of which in
turn affect children’s educational outcomes and
social mobility. Consequently, any discrimination in
access to credit can have a long-lasting detrimental
effect on communities subject to such discrimination.
Access to credit can be determined in no small part
by one’s credit history, which is often summarized
Meet the New Borrowers

2022 Q1

by a credit score (such as the FICO score). These
scores are easy to read and compare.3 For example,
discrimination can simply be defined as different
treatment of two individuals with identical credit
scores. However, when investigating the presence of
discrimination in the marketplace, it may not be
enough to check whether individuals from different
social groups are treated equally conditional on their
credit score. Sometimes, underprivileged borrowers
fail to achieve a good credit score in the first place
because of their inability to build a credit history.4
The traditional view is that a credit history is
a history of repayment. But new borrowers (who are
the focus of this article) have had little time to
establish such a record of paying on time. This
brings us to our Catch-22: You need a credit history
to get credit, and you need credit to build a credit
history. This Catch-22 is particularly pronounced at
the initial stage of the credit life cycle, which is the
Federal Reserve Bank of Philadelphia
Research Department

9

subject of this article. And it is likely to be especially
pronounced for individuals from an underprivileged
background because these individuals cannot “piggyback” on their parents’ credit histories.
With that in mind, I will highlight an additional role
of credit history. A credit history is also a record of
borrowing, that is, of loan approvals. In this article,
I examine the initial stage of the credit life cycle to
better understand how inequality manifests itself in
individuals’ gaining access to credit, and I examine
how emerging borrowers overcome our Catch-22.

The Unscored and the Invisible

Until recently, there was little academic research into
the early stages of the credit life cycle and the dynamics of access to credit among emerging borrowers.
But that’s changing thanks in part to researchers
gaining access to anonymized credit records data. By
using such data in their 2017 article, economists
Kenneth P. Brevoort and Michelle Kambara, both then
at the Consumer Finance Protection Bureau (CFPB),
answer some of my key questions. (Later in this article
I refer to their findings alongside my own.)
In important related work that also uses anonymized credit records data, Brevoort, Kambara, and
economist Philipp Grimm, who was also at the CFPB
at the time, studied “credit invisibles” and “unscorable”
individuals. “Credit invisibles” are individuals who
have no record with one of the three major creditreporting agencies.5 It is difficult to study those you
cannot see. But we can compare the population
in credit bureau files with the population in the U.S.
Census to figure out who is missing from the former.
Unlike the credit invisibles, people who are unscorable
have a record with the credit-reporting agency, but
their file is “too thin” to generate a reliable credit
score. Brevoort, Grimm, and Kambara rightfully refer
to these individuals as “unscored,” highlighting the
conceptual possibility of assigning a score, especially
if alternative data sources are permitted.
A key takeaway from Brevoort, Grimm, and Kambara’s 2016 article is that neighborhoods with a greater
share of underprivileged individuals have a greater
prevalence of credit invisibility. Perhaps most notably,
these economists also find that minority consumers
are less likely to be credit visible, even when researchers control for (relative) income. This supports the
idea that solely controlling for credit scores is not
sufficient for identifying unequal access to credit.

Shortcuts to Credit History

For new borrowers who lack a history of repayment,
there are two shortcuts to acquiring credit and
building a credit history: secured credit cards and
“piggybacking.”

10

Federal Reserve Bank of Philadelphia
Research Department

In his 2016 discussion paper, Philadelphia Fed
economist Larry Santucci highlights the importance of
secured credit cards as a gateway product and investigates how consumers graduate from secured to
unsecured credit cards. Secured cards are a rather
unusual credit product, because borrowers end up
largely borrowing from themselves—a “lender” is
secured by a cash deposit (or a locked savings account)
that often matches the credit line on the card. Yet,
since the card is reported to the credit bureaus, this
product helps borrowers establish or repair a credit
record. My findings confirm Santucci’s insight: The
probability of having a secured card is 17 percent
among new borrowers with credit cards, but less than
a tenth of that for established borrowers.6
In their 2010 working paper, Brevoort, Federal
Housing Administration economist Robert Avery, and
Federal Reserve economist Glenn Canner point to
another way for a new borrower to quickly establish
a credit history: piggybacking, which refers to the
practice of adding a new borrower to an existing (and
established) credit card account, often that of a parent.
This allows the new borrower to add the established
card to their credit record. In their 2017 article,
Brevoort and Kambara report that a quarter of new
borrowers enter the credit market with someone’s
help. (Fifteen percent enter with a joint account, and
another 10 percent enter as authorized users—that is,
they piggyback.) Importantly, if unsurprisingly, this
number is smaller in poor neighborhoods. Although
I cannot directly observe the prevalence of authorized
users in the data, I can approximate their prevalence
by seeing how often an old card—that is, a credit
card more than nine months old—first appears on
a borrower’s credit record.7 Surprisingly, piggybacking did not become more common following the
implementation of the Credit Card Accountability
Responsibility and Disclosure (CARD) Act of 2009,
which made it harder for young people to get a credit
card independently.

The New Borrowers

Like Brevoort and Kambara, I use anonymized credit
bureau data to study new borrowers. However,
the FRBNY Consumer Credit Panel/Equifax (CCP) data
I use is distinct from the data set employed by Brevoort and Kambara.8 One peculiar aspect of the CCP
data I use is that the sample expands unevenly
over time (likely due to the household aspect of
the data by design). As a result, I can’t safely define a new borrower as someone appearing in the
data set for the first time. Instead, I define a new
borrower as someone whose oldest credit trade
(credit product) is no more than three months old.9
(The two data sets also categorize credit products
slightly differently.)

Meet the New Borrowers
2022 Q1

New

Borrowers

28

Average age of person

$3,874

Average nonmortgage credit line

$1,256

Average credit card limit

Established
Borrowers

51

Average age of person

$37,000

Average nonmortgage credit line

$23,000

Average credit card limit

Source: Author’s calculations based on FRBNY
Consumer Credit Panel/
Equifax (CCP) data.
Note: New borrower is
one whose oldest credit
product is less than three
months old.

Not surprisingly, new borrowers are
much younger than established borrowers.
The average new borrower is approximately 28 years old, while the average age
of a person with a credit record is almost
51. And new borrowers’ credit lines are
a fraction of established borrowers’ credit
lines. The average nonmortgage credit limit of a new borrower is a mere $3,874—just
over a tenth the average nonmortgage
credit limit of established borrowers. The
average credit limit of a new borrower’s
credit cards is $1,256, which is almost 18
times smaller than the average for established borrowers.
More surprisingly, new borrowers
disproportionately live in poorer neighborhoods. This observation coexists with
Brevoort, Grimm, and Kambara’s observation that poorer neighborhoods have
a greater share of credit invisibles. This
could reflect either systematic differences
in age composition across neighborhoods
or a tendency of people who gain access
to credit to move out of poorer neighborhoods (Figures 1 and 2).10
To determine whether households in
disadvantaged neighborhoods struggle
more to access credit, I compared the
average age of new borrowers across neighborhoods. There is basis for concern:
Individuals in disadvantaged neighborhoods gain credit visibility (and credit
access) later in life than their counterparts
in more privileged areas. When I conducted multivariate linear regressions of the
average age of new borrowers on a set
of neighborhood characteristics, I found
that the regressions yielded positive and
strongly statistically significant coefficients
on the percentage of the neighborhood’s
population that belong to a racial minority,
the percentage of the population that are
noncitizens, and the percentage living below the poverty line; these coefficients get
larger when the regression controls for the
age composition of the neighborhood.11
In other words, people in disadvantaged
neighborhoods get credit access later
in life than their peers in more privileged
neighborhoods.
We can also
See Credit Card:
analyze the credit
The Gateway
products new borProduct.
rowers use to enter
the credit market. Again, Brevoort and
Kambara have already looked into this.12

FIGURE 1

Credit Invisibles Disproportionately Live in Disadvantaged Neighborhoods
Share of credit visibles in neighborhoods divided into quartiles, by ethnicity (left) and poverty (right), from
lowest (1st quartile) to highest (4th quartile)
1.10

1.10

1.05

1.05

1.00

1.00

0.95

0.95

0.90

1st

2nd

3rd

4th

0.90

1st

Quartile of Neighborhoods
by Proportion of People
Who Belong to an Ethnic Minority
←Fewer

2nd

3rd

4th

Quartile of Neighborhoods
by Proportion of People
Living Below the Poverty Line

More →

←Fewer

More →

Source: Author's calculations based on FRBNY Consumer Credit Panel/Equifax (CCP) and U.S. Bureau of Census
(Census) data. Ethnicity data are from Census.
Note: This figure depicts the ratio of each quartile's average of the statistic to the overall average of the statistic.

FIGURE 2

New Borrowers Disproportionately Live in Disadvantaged Neighborhoods
Share of new borrowers in neighborhoods divided into quartiles, by ethnicity (left) and poverty (right), from
lowest (1st quartile) to highest (4th quartile)
1.3

1.3

1.2

1.2

1.1

1.1

1.0

1.0

0.9

0.9

0.8

0.8

0.7

1st

2nd

3rd

4th

Quartile of Neighborhoods
by Proportion of People
Who Belong to an Ethnic Minority
←Fewer

More →

0.7

1st

2nd

3rd

4th

Quartile of Neighborhoods
by Proportion of People
Living Below the Poverty Line
←Fewer

More →

Source: Author's calculations based on FRBNY Consumer Credit Panel/Equifax (CCP) and U.S. Bureau of Census
(Census) data. Ethnicity data are from Census.
Note: This figure depicts the ratio of each quartile's average of the statistic to the overall average of the statistic.

Meet the New Borrowers

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

11

I find that credit cards are even more important for initiating
credit records than Brevoort and Kambara suggest (Figure 3), even
though I omit piggybackers from my definition of new borrowers.
Overall, credit cards account for about half of all credit market
entries. Student loans, retail credit, and auto loans are the other
important contributors. Not surprisingly, mortgages account
for just a small fraction of new entries, since a typical first-time
homebuyer has a well-established credit history.
Figure 3 further illustrates another important point: Credit
market entry is very sensitive to aggregate economic conditions.
An economic downturn (such as the Great Recession) leads to
tighter lending standards that dramatically curtail the entry of
new borrowers. A notable exception to that rule is student loans,
which may boom in downturns as more people choose to pursue
a formal education while labor markets slump.
I conclude with a few key observations regarding the evolution
of credit access for new borrowers in the first year after entering
the credit market. As already noted, their average initial credit
limit is a fraction of that of established borrowers. But this average
credit limit grows rapidly.13 The average (total) credit line of
new borrowers’ credit cards more than doubles in the first year.
The growth of credit access is particularly dramatic among
borrowers who gain additional credit cards: Their aggregate credit
line quadruples in the first year. This is unsurprising—more
credit cards typically mean more available credit. What is surprising is that a significant share of this expansion of credit comes
from an increase in the credit limit of their original credit card. In
2021, Arizona State University economist Natalia Kovrijnykh,
Carnegie Mellon University economist Ariel Zetlin-Jones, and
I documented this fact using a distinct (customized) anonymized
data set from a different credit reporting agency. (We ran a regression analysis to confirm the statistical significance and robustness
of this observation.) Notably, although the observation is robust
among new borrowers, it is not statistically significant among
established borrowers. These facts point to the importance of
borrowing from multiple lenders, particularly for new borrowers.

New Borrowers and Borrowing from
Multiple Lenders

My ongoing research with Kovrijnykh and Zetlin-Jones starts with
the question: Why do new borrowers who obtain an additional
card see a disproportionate increase in the credit line from their
original lender? This goes against conventional wisdom, which
states that the original lender should be concerned about debt
dilution (where an additional loan decreases the value of preexisting debt). The new loan increases the overall repayment
burden of the borrower and should thus lead to a greater likelihood of default. Why then would the original lender extend the
credit line even further?
This increase in aggregate credit is not driven solely by the
demand channel (that is, by a borrower’s request for more
credit from all lenders, old and new). New borrowers who try
but fail to get an additional loan do not see the large increase
in the credit line of their original credit card.
It appears that incumbent lenders interpret the fact that
a new borrower obtains additional credit as a positive signal

12

Federal Reserve Bank of Philadelphia
Research Department

about the borrower’s underlying risk (that is, the borrower’s
quality). This signal appears to be particularly important when
it applies to a new borrower. Incumbent lenders respond
positively to new credit when a borrower has a very short
credit history, but this response is not evident when it comes to
established borrowers. In other words, these signals from other
lenders appear to be particularly valuable when a long history
of repayment behavior is absent from a credit record.

Aggregating Information Across Lenders
by Building Credit History

Because lenders seem to interpret other lenders’ decisions about
an individual borrower as informative, I focus on the signaling
component of credit histories. Although repayments are an
important component of credit histories, so too is the informationaggregation aspect of these records, particularly as it applies to
emerging borrowers. Lenders appear to respond to the granting
of credit to a new borrower before the borrower establishes
any pattern of repayment. That’s strong evidence in support of
the information-aggregation mechanism.
For the theoretical portion of our research, we put forward
a simple model that captures this information-aggregation
mechanism (Figure 5). In the model, as in real life, lenders are
heterogeneously informed—that is, they differ in what information they have about a borrower, or in how they interpret the
information available to them. An example of differing access
to information: My first credit card came from a lender that
verified my enrollment as a university student—information not
directly available to other lenders. An example of differing interpretations of information: the proprietary credit-scoring models
employed by credit card lenders. However, our theoretical
analysis does not distinguish between these two sources of
information dispersion. We simply model multiple lenders
receiving separate informative signals regarding a borrower’s
underlying risk type (that is, their likelihood of being able to repay
loans of various sizes). Lenders in the model recognize the
fact that their competitors receive additional information that
is useful above and beyond the signal they received themselves.
Consequently, lenders have a reason to read into their competitors’ credit approval decisions, as these may reflect the
competitors’ information about the borrower.
In order to capture the mechanism described above, the model
features borrowing over multiple stages. Early-stage loans are
recorded in a publicly visible credit history. This credit history
then affects the loan offers a borrower receives in the late stage of
borrowing. Our theoretical analysis abstracts from learning from
repayment: All of the loans are advanced before any repayment
takes place. Yet borrowing over multiple stages permits the model
to capture both credit-history building and debt dilution.
As intended, the model yields credit-history building. Specifically, only lenders with positive signals about a borrower offer
a loan to that borrower at the early stage of borrowing. An earlystage loan thus informs other lenders of the positive signal the
early lender has received. As a result, the dispersed information
across lenders is aggregated in the late borrowing stage. That
is, late-stage loan contracts reflect both the information (signal) of

Meet the New Borrowers
2022 Q1

FIGURE 3

Credit Cards Are the Dominant Form of Credit Market Entry

When it comes to initiating a credit record, credit cards are the most important product.
Estimated count of borrowers by first credit product, thousands, 2003–2019
2,500

Credit Cards

Student Loans

Auto Loans

Department Store

More Than One

Mortgage

2,000
1,500
1,000
500
0

2003

2019

2003

2019

2003

2019

2003

2019

2003

2019

2003

2019

Source: Author’s calculations based on FRBNY Consumer Credit Panel/Equifax (CCP) data.

Credit Card: The Gateway Product
In their 2017 article, Brevoort and Kambara
highlight the increasing importance of credit
cards as the gateway product, but as they
point out, this observation does not apply to
borrowers under 25. They point to the
restrictions imposed by the Credit CARD Act
of 2009 as the likely explanation for this
dichotomy. The Credit CARD Act explicitly
restricts the marketing of credit cards to
college students and individuals under 21.
Unsurprisingly, the average age of a new
borrower increased significantly following
the act’s implementation (Figure 4). In their
2021 working paper, Boston Fed economists
Daniel Cooper and María José LuengoPrado and University of Delaware economist
Olga Gorbachev also document this fact.
They argue that this restriction of credit has
slowed the growth in overall consumption.
Figure 4 also illustrates the seasonality of
credit market entry. The average age of entrants plunges in the third quarter of every
year, just when the entry rate spikes (for
both student loans and credit cards).15 In
their 2013 discussion paper, the Philadelphia
Fed’s Keith Wardrip and Robert M. Hunt
suggest that the composition of new borrower cohorts is also affected by the business
cycle, as lending standards tend to tighten
during recessions.

FIGURE 4

The Average Age of a New Credit Card Borrower Increased Significantly
Following the Credit CARD Act of 2009
The act restricts the marketing of credit cards to college students and to borrowers
under 21 years old.
Average age of new borrowers whose first credit product is a credit card, 2003–2020
36
35
34
33
32
31
30
29
28

2003
Q1

2005

2010

2015

Q1

Q1

Q1

2019
Q4

Source: Author’s calculations based on FRBNY Consumer Credit Panel/Equifax (CCP) data.
Note: A new credit card borrower is one whose oldest credit card and oldest credit trade are at most two
months old.

Meet the New Borrowers

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

13

the late lender and the signal of the early
lender. We thus have an environment
in which the “best” new borrowers build
a credit history by taking on (rather than
repaying) an early-stage loan. Doing so
facilitates information aggregation across
lenders—that is, it convinces late-stage lenders of their creditworthiness.
But there are costs associated with
building a credit history. One such cost is
having to pay inflated interest rates at the
early stage of borrowing. This is a result of
cross-subsidization, as the (best) borrowers
who are building their credit histories
are pooled with riskier borrowers who are
taking advantage of an interest rate that
does not fully reflect their true risk of
default. Another possible cost of credithistory building is overborrowing—that is,
ending up with a larger-than-optimal
loan. Borrowing over multiple stages (and
from multiple lenders) gives rise to debt
dilution, which is familiar from both corporate and international finance literature,
though it is rarely emphasized in consumer

credit literature. In our theoretical environment, borrowers’ inability to commit
not to overborrow, combined with the
debt dilution motive at a late stage, may
result in the best borrowers taking on
inefficiently large loans.14
Yet, despite the costs associated with
building a credit history, the best borrowers still find it worthwhile to take early
loans in order to facilitate information
aggregation across lenders. That is, they
still use the early loan to signal to their
later lenders the favorable signal of the
early lenders. The alternative to building
a credit history in this way is either a smaller overall loan or one at a less-favorable
interest rate.
Our model highlights the importance
(and the favorable side) of borrowing
from multiple lenders. This is in contrast
to how this is normally viewed in the
literature: as simply debt dilution.
This theory also has a surprising implication: More dilution is better. In the
late stage, the early lender would rather

see a larger top-up loan than a smaller topup loan. (A top-up loan is a loan added
to a preexisting loan.) In the model, this
loan comes from a later lender. And that
later lender has an additional piece of
information (signal) about the borrower,
beyond what was available to the early
lender. As a result, the size of the top-up
loan is informative. Although a larger loan
to a given type of borrower is bad news
for lenders (because it increases the probability of default in the repayment period),
a counteracting force dominates in
our model: Only the best borrowers get
a large top-up loan, while borrowers with
smaller top-up loans have less-favorable
signals. This selection effect dominates
the dilution effect explained above.
Strikingly, we find that this model
prediction is borne out in the data. Delinquency rates are indeed lower among
new borrowers with larger top-up loans
than among new borrowers with smaller
top-up loans. Notably, this observation
does not apply to established borrowers,

FIGURE 5

A Simple Model of the Information-Aggregation Mechanism
Stage 1

Stage 2

Different lenders may have
different information.

Lender A Lender B

A

Lender A Lender B
Lenders may interpret
information differently.

A

This leads to different loan
terms, or no loan offer at all.
Loan 1

$

Borrower accepts at most
one loan offer.

Borrower

14

Federal Reserve Bank of Philadelphia
Research Department

Loan 2

$$$

Borrower

Meet the New Borrowers
2022 Q1

Lenders know that other
lenders use different
information or interpret
information differently.
Hence, Lender A's loan
informs Lender B, who
now offers a loan with
more favorable terms.

Borrower accepts one
more loan.

Notes

which suggests that the informational content (or
rather, spillover) of a lending relationship is less
important for that group.

1 The pricing of loans may be just as important
as their availability. Thus, “access to credit”
really means “access to affordable credit.”

Conclusion

2 On the other hand, there are concerns about
excessive indebtedness. These concerns
motivated some of the restrictions in the Credit
CARD Act, which explicitly limit the marketing
of credit to young (new) customers, thus
limiting their access to credit. In this article,
I focus on the positive (in every sense of the
word) aspects of credit access, including
positive analysis of the effects of the Credit CARD
Act. Normative concerns, as well as the
specific behavioral biases that lead to financial
mistakes, are the subject of my 2020 Economic
Insights article.

To analyze and address issues related to (unequal)
access to credit, we need to understand how new
borrowers gain and expand their access to credit,
and how policies and external circumstances affect
their ability to do so. My research highlights the
importance of aggregating dispersed information
regarding borrowers’ creditworthiness and the role
played by credit history in that aggregation.
These insights should complement the recent findings of researchers who document that lenders can
benefit from alternative data (ranging from noncredit
bill payment to social media behavior) in loan
underwriting decisions. For example, in their 2020
article, Frankfurt School of Finance & Management
economists Tobias Berg and Ana Gombović, Humboldt
University (Berlin) economist Valentin Burg, and
Duke University economist Manju Puri point out that
the digital footprint of an online shopper can be
as informative about their future default rate as the
information contained in their credit records.
But information improvements may not be a “free
lunch.” The same information that convinces lenders
that a subset of borrowers is creditworthy likely leads
them to reject other potential new borrowers.

3 Theoretical analysis conducted by Kovrijnykh,
Livshits, and Zetlin-Jones (2019) highlights an
important distinction between building one’s
credit history and improving one’s credit score.
Borrowers with the most favorable income
prospects build their credit history in order
to convince lenders to grant them large (and
riskier) loans. Thus, these “best” borrowers
end up with a higher probability of default
(and a lower credit score) than borrowers
without a credit history, who do not qualify for
the riskiest loans.
4 Both this distinction and the critical importance of the credit history are highlighted in
a new article by Stanford University economist
Laura Blattner and University of Chicago economist Scott Nelson.
5 Although mainstream lenders, as a rule, report
their loans (and their repayment) to credit bureaus, some fringe lenders do not. Payday loans,
for example, are typically not reflected in
the credit records that we consider here. This
explains why a nontrivial fraction of credit
records begin with a record of a collection, even
though the loan that led to that collection was
not itself in the credit record.
6 The data set used is the FRBNY Consumer
Credit Panel/Equifax data (CCP).
7 I do not observe whether an account is jointly
opened with another borrower. This is another
way to facilitate credit market entry, as pointed
out by Brevoort and Kambara (2017).

Meet the New Borrowers

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

15

8 I use data from the first quarter of 2019 for my regression analysis,
whereas for the summary statistics I aggregate a full year’s worth of
data (ending in that quarter) because the characteristics of new borrowers
have pronounced seasonal fluctuations.
9 In contrast to Brevoort and Kambara (2017), I exclude from the
definition of “new borrowers” consumers whose first entry with a credit
bureau is when they are in collections. My definition also excludes
borrowers who enter the credit market by piggybacking on older cards.
10 These statistics are computed based on representation in the CCP
compared with the adult populations of census tracts reported in the
American Community Survey (ACS). These levels of credit visibility are
implausibly high across the board (likely due to the household aspect
of the data by design), but our focus is the comparison across different
clusters of neighborhoods. These comparisons are robust to various
ways of addressing bias in the levels of visibility.
11 The table of regression results is available from the author upon
request: igor.livshits@phil.frb.org.
12 Our findings differ slightly, most likely due to our different definitions
of new borrowers, our different data sets, our different time periods, or
a combination thereof.
13 Another aspect in which new borrowers differ from established ones
is their riskiness — that is, the probability that they will fail to pay their
debts in a timely manner. When new borrowers first appear in the data
set, their delinquency rate is very low, but that is largely mechanical—
they haven’t had debts long enough to miss many payments. Within
a couple of years, the delinquency rate of a cohort of new borrowers
overtakes that of established borrowers. That means lenders may have
good reason to be reluctant to advance credit to individuals without an
established credit history.
14 See Eyigungor (2013) for an excellent discussion of debt dilution.
15 Because the entry rate is seasonally volatile, I plot Figure 4 at the
annual frequency.

References
Avery, Robert B., Kenneth Brevoort, and Glenn B. Canner. “Credit Where
None Is Due? Authorized User Account Status and Piggybacking
Credit,” Finance and Economics Discussion Series (FEDS) Working Paper
2010-23 (2010).
Berg, Tobias, Valentin Burg, Ana Gombović, and Manju Puri. “On the Rise
of Fintechs: Credit Scoring Using Digital Footprints,” Review of Financial
Studies, 33:7 (2020), pp. 2845–2897, https://doi.org/10.1093/rfs/hhz099.
Blattner, Laura, and Scott Nelson. “How Costly Is Noise? Data and
Disparities in Consumer Credit,” unpublished manuscript (2021).
Brevoort, Kenneth P., Philipp Grimm, and Michelle Kambara. “Credit
Invisibles and the Unscored,” Cityscape, 18:2 (2016), pp. 9–34.
Brevoort, Kenneth P., and Michelle Kambara. “Becoming Credit Visible,”
Consumer Financial Protection Bureau Data Point Series 17-1 (2017).
Cooper, Daniel, Olga Gorbachev, and María José Luengo-Prado.
“Consumption, Credit, and the Missing Young,” working paper (2021).
Eyigungor, Burcu. “Debt Dilution: When It Is a Major Problem and How
to Deal with It,” Federal Reserve Bank of Philadelphia Business Review
(Fourth Quarter 2013), pp. 1–8.
Kovrijnykh, Natalia, Igor Livshits, and Ariel Zetlin-Jones. “Building Credit
History with Heterogeneously Informed Lenders,” Federal Reserve Bank
of Philadelphia Working Paper 19-17 (2019), https://doi.org/10.21799/
frbp.wp.2019.17.
Kovrijnykh, Natalia, Igor Livshits, and Ariel Zetlin-Jones. “Building Credit
Histories,” unpublished (2021).
Livshits, Igor. “Regulating Consumer Credit and Protecting (Behavioral)
Borrowers,” Federal Reserve Bank of Philadelphia Economic Insights
(First Quarter 2020), pp. 14–19, https://www.philadelphiafed.org/
consumer-finance/consumer-credit/regulating-consumer-credit-andprotecting-behavioral-borrowers.
Santucci, Larry. “The Secured Credit Card Market,” Federal Reserve Bank
of Philadelphia Payment Cards Center Discussion Paper 16-3 (2016),
https://www.philadelphiafed.org/consumer-finance/consumer-credit/
the-secured-credit-card-market.
Wardrip, Keith, and Robert M. Hunt. “Residential Migration, Entry, and
Exit as Seen Through the Lens of Credit Bureau Data,” Federal Reserve
Bank of Philadelphia Community Development Studies and Education
Discussion Paper 13-01 (2013).

16

Federal Reserve Bank of Philadelphia
Research Department

Meet the New Borrowers
2022 Q1

Photo: Brendan Barry

The Costs and Benefits
of Fixing Downtown Freeways
Urban freeways spurred our suburban boom.
Can burying them do the same for the urban core?

F

reeways are conspicuous features
of urban landscapes.1 Highway construction represented a massive
infrastructure investment in the 20th century, and it improved access, commuting,
and trade. Nonetheless, it has long been
recognized that there were negative effects
for nearby neighborhoods, particularly
in central cities. Today, many cities are mitigating some of the negative effects of
freeways through expensive measures to
cap or bury sections of freeway. Do these
projects justify the costs? Could future
infrastructure investments benefit from consideration of neighborhood disamenities?
In this article, we summarize evidence

of freeways’ effects on quality of life and
discuss the potential benefits of real-world
policy interventions in Philadelphia.

Construction of the Interstate
Highway System

Discussion of a national system of interstate highways, which had been gaining
momentum since the 1930s, culminated
with the signing of the Federal-Aid Highway Act of 1956 by President Eisenhower.
This act authorized the construction of
41,000 miles of freeways over a 10-year
period. To gain popular support for the
highway system, Eisenhower emphasized

The Costs and Benefits of Fixing Downtown Freeways

2022 Q1

Jeffrey Brinkman
Senior Economist
Federal Reserve Bank
of Philadelphia

Jeffrey Lin
Vice President and Economist
Federal Reserve Bank
of Philadelphia
The views expressed in this article
are not necessarily those of the
Federal Reserve.

Federal Reserve Bank of Philadelphia
Research Department

17

its advantages for national defense. However, the economic
benefits were the primary motivation for supporters of the plan,
and boosters of freeway construction touted the reduced transportation costs associated with freeways.2 Mayors of major cities
broadly supported construction, believing that new freeways
would reduce congestion and improve the local economy.
Economic development was an important rationale for freeway
construction, and while there are clear benefits for a region, the
effects of freeways can be either positive or negative for an
individual neighborhood. The new interstate system improved
commerce and trade by connecting major cities and reducing
travel times,3 but as University of Toronto economist Nathaniel
Baum-Snow shows, freeways also accelerated suburbanization and
exacerbated the population decline in central cities. This population decline near downtowns was partially driven by reduced
transportation costs that improved access and made suburban
areas relatively more attractive. Freeways further worsened the
decline by reducing the quality of life in central neighborhoods.
As freeway construction began rapidly in the late 1950s, residents
came to recognize these negative effects, and protests against
construction appeared in most large U.S. cities.

Amenities for Some, Disamenities for Others

The construction of freeways brought broad changes to urban
areas, but the costs and benefits of the freeways were not the same
for all neighborhoods. When a freeway is built through a city,
access to regional amenities such as job centers improves in
neighborhoods near the freeway due to reduced travel times. This
is particularly true for outlying areas located a long distance from
the jobs and services that are often concentrated in central cities.
Therefore, when freeways were constructed, neighborhoods
in suburbs far from the central business district grew rapidly.
However, freeways also negatively affect the quality of life
for nearby neighborhoods. These disamenities include noise,
pollution, and barrier effects, whereby a newly constructed freeway limits access to amenities and services located on the other
side of the freeway. For neighborhoods that do not benefit
significantly from improved regional access, these negative effects
can lead to neighborhood decline. For example, locations near
central business districts already have access to jobs and other
regional amenities and thus do not gain much from freeways. In
these neighborhoods, the negative effects of freeways dominate,
and the net result is population loss.
In a recent working paper, we provide evidence that freeways reduced the quality of life in nearby neighborhoods by
looking at long-run changes in population and other variables.
What we find is as expected: Suburban neighborhoods near
freeways grew rapidly after freeways were constructed, while
central neighborhoods near freeways declined. Using fine geographic data covering 1950 to 2010, we studied long-run changes
in neighborhoods before and after the interstate highway
system was built (Figure 1). We find that in the group of centralcity neighborhoods closest to freeways, population declined
by 32 percent, while in the group of central neighborhoods
more than 2 miles from freeways, population actually grew
by 56 percent.

18

Federal Reserve Bank of Philadelphia
Research Department

Much of the negative effect on local amenities from freeways is
due to barrier effects. Freeways often block local streets and
limit the passthrough of cars and pedestrians. When a freeway
cuts off a neighborhood from nearby amenities, the neighborhood becomes less desirable, and people relocate to other
neighborhoods. Using data from historic travel surveys before
and after freeways were constructed, we find that people were
less likely to travel to the other side of a freeway locally, and if
they did, the travel time was longer. In other words, although
freeways improve overall regional access, they reduce access to
nearby neighborhood amenities.4 This research suggests that
construction of the interstate highway system incurred significant
external costs, and policymakers should consider these costs
when assessing the value of urban freeway projects.

Quantifying Neighborhood Amenities

When measuring the effects of freeways, it is often hard to disentangle negative quality-of-life effects from the benefits accrued
thanks to greater access to jobs and other regional resources. One
way to identify quality-of-life amenities is to use proximity to
a city’s central business district as a proxy for job access before
construction of the highway. But cities are more complex than
suggested by that simple proxy. For example, cities often exhibit
multiple job centers. For this reason, we prefer measures of job
access that help us study cities with real-world geographies.

FIGURE 1

In Central Neighborhoods, Population Declines Are
Greatest in Census Tracts Nearest to Freeways

Average population change, 1950–2010, in bins of neighborhoods within 2.5
miles of the city center, plotted against the distance to the closest freeway, for
a sample of 2,312 neighborhoods in 64 metro areas in the U.S.
Tract Population in 2010 / Tract Population in 1950
60%
50%
40%
30%
20%
10%
0%
−10%
−20%
−30%
−40%
−50%
−60%
0.0 mi. 0.5 mi.
1.0 mi.
Miles to the nearest highway

1.5 mi.

Sources: Brinkman and Lin (2019); U.S. Census Bureau.

The Costs and Benefits of Fixing Downtown Freeways
2022 Q1

2.0 mi.

2.5 mi.

FIGURE 2

Many People Highly Value Living Near Jobs

Jobs are very dense in the central business district,
and many residents live near those jobs.

Employment density (top panel) and employed residential population density
(bottom panel) for census tracts in central Philadelphia on the same scale, 2000.

ard
Girve
A

676

Market St

Employment Density
Jobs per square mile
100,000+
50,000–100,000
10,000–50,000
5,000–10,000
1,000–5,000
1–1,000
0

76

Broad St

95

Snyder Ave

0
0.5
Miles

1

1.5

2

2.5

1.5

2

2.5

ard
Girve
A

676

Market St

Residential Density
People per square mile
20,000+
10,000–20,000
5,000–10,000
1,000–5,000
1–1,000
0

76
95

Broad St

An emerging literature in urban economics uses the spatial distribution of jobs,
residences, land prices, and wages to separately quantify the value of locations for
production (productivity) and the value
of locations for consumption (residential
amenities). The value of a residential
location can arise from a variety of characteristics, including good schools,
entertainment options, and natural amenities such as ocean views. Likewise,
locations vary in their value for production
due to natural advantages such as proximity to natural resources, or proximity to
customers, suppliers, or employees.
Finally, these locations are all connected,
given that people consider the time and expense of traveling to work when choosing
where to live. In addition, firms consider
access to a pool of employees when considering where to locate. Since people can
usually choose where to live and where
to work, the spatial distributions of population and employment in cities provide
evidence of the value of locations for
different activities.
Employment and residences are distributed very differently within urban areas.
This suggests that locations vary in their
value for production versus residential
uses. We find that there is an extremely
high density of jobs in the central business
district of Philadelphia, with employment
densities exceeding 200,000 jobs per
square mile for several census tracts
(Figure 2, top panel). Jobs are highly
concentrated in the central business
district even though land prices there are
extremely high. It is common for per-acre
land prices in American cities to be at
least 10 times higher in the central business
district than in suburbs just 10 miles
away.5 The concentration of jobs and the
willingness to pay such high prices is clear
evidence that business districts provide
advantages for the production of goods
and services. Researchers have shown that
these efficiencies can arise from access
to a pool of employees, input sharing, and
information spillovers (that is, information
about one thing generating information
about seemingly unrelated things).6
However, although residences are not
as spatially concentrated as jobs, there are
still big differences in density across space
(Figure 2, bottom panel). For example,
the neighborhoods directly south of the

Snyder Ave

0
0.5
Miles

1

Sources: Brinkman and Lin (2019); American Association of State Highway and Transportation Census Transportation Planning Products (CTPP) program; U.S. Census Bureau.

The Costs and Benefits of Fixing Downtown Freeways

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

19

central business district exhibit residential population densities
as high as 25,000 employed workers per square mile. Again,
given that these locations are also very expensive, it is clear that
people particularly value living in these locations. Less obvious
is whether they derive this value from proximity to jobs or from
residential amenities.
Recently, some economists have developed quantitative models
to disentangle how much people value different characteristics
of a location, including access to jobs and residential amenities.7
By using the observed spatial distribution of jobs and workers,
and also by incorporating information on rents, wages, and travel
times between locations, these economists can identify the mechanisms that guide the spatial layout of cities and the colocation
patterns of firms and workers in cities. In particular, their models
separate the value of job access from the quality-of-life benefits
of neighborhoods. These models also allow for the analysis of
real-world policies. For example, Philadelphia Fed economist
Christopher Severen uses one such model to study the effects of
subway construction in Los Angeles. By using such a model for
our working paper, we find that the quality-of-life effects of freeways play an important role in decentralization and significantly
affect overall welfare.

Mitigating Freeway Disamenities

Many cities have implemented or considered projects to mitigate
disamenity effects by burying or capping freeways through city
centers. The goal of these projects is to reconnect streets and
neighborhoods, reduce noise, and reclaim land for other urban
uses. These projects continue to move forward despite high
construction costs. Costs vary depending on project details but
can range from $300 million to $700 million per mile of freeway.
Freeway construction costs have increased drastically since
construction of the interstate system.8 Therefore, it is important to
know whether the benefits of these projects are worth the costs.
In Philadelphia, several projects have partially capped small
parts of freeways. Parts of I-676 though Center City were partially
capped to create small parks near the Benjamin Franklin Parkway,
the scenic, tree-lined boulevard connecting City Hall with the
Philadelphia Museum of Art. The costs were modest given that
the freeway was already below grade, and construction was
done as part of a project to reconstruct existing bridges crossing
the freeway. Another project would extend an already existing
cap over I-95, which closely follows the Delaware River through
the city, to better connect the city to the riverfront. The new
project covers only an additional one-tenth of a mile of freeway
but involves development of a large urban park. Despite this
improvement, large sections of the Philadelphia waterfront will
remain cut off by I-95. When the freeway was first built, much
of the waterfront was a declining industrial zone. Planners saw
this zone as the logical route for the new north/south interstate
through Philadelphia. However, 60 years later, the Philadelphia
waterfront remains underutilized, and I-95 is the obvious obstacle preventing redevelopment.
We estimate the benefits of a more ambitious project that
would reconnect a much larger portion of central Philadelphia
to the Delaware River waterfront (Figure 3). Using quantitative

20

Federal Reserve Bank of Philadelphia
Research Department

methods developed in urban economics, we simulate the effects
of burying a section of I-95 from Snyder Avenue to Girard Avenue.
This roughly 4.5-mile stretch of freeway starts in South Philly
and traverses the riverfront neighborhoods of Pennsport, Queen
Village, Society Hill, Old City, Northern Liberties, and Fishtown.
The proximity of these neighborhoods to the central business
district and their high population density suggest that this might
be an ideal setting for such an intervention.
We conduct the analysis using data on the location of population and employment, as well as data on commuting travel
times between locations. We input these data into a quantitative
model to estimate the amenities and productivities of different
neighborhoods. Intuitively, amenities are estimated through the
model by comparing neighborhoods in terms of job access and
population density. If a neighborhood has superior job access
but low population density, this is evidence of fewer amenities.
For Philadelphia, we find that neighborhood amenity values are
roughly 11 percent lower immediately next to a freeway compared to locations far away. In addition, these effects decline but
persist out to at least a mile from a freeway. In other words,
people would be willing to pay roughly 11 percent of their income
to avoid living directly next to a freeway, holding everything else
constant (including access to jobs). In analysis conducted for
a recent working paper, we find an even larger effect of 17 percent
in Chicago. These estimates suggest that negative quality-of-life
effects from freeways are quantitatively important.
Next, we use these estimates of disamenities and quantitative
modeling techniques to analyze the effect of burying I-95 in
central Philadelphia. We simulate a counterfactual economy
where the transportation benefits of the freeway remain but the
negative effects to nearby neighborhoods would be fully mitigated.
The improvement to nearby neighborhoods would be accomplished by reconnecting streets, reducing noise and pollution,
and reclaiming land for other uses. We do not consider removing
the freeway altogether, given that this would require calculating
changes in travel patterns throughout the region. This is harder to
simulate, but techniques have been developed to account for
the effect of changes in transport networks on travel. Removal
of the freeway would likely have muted benefits relative to the
mitigation experiment we present here.
The first result of the experiment is that population near the
freeway increases drastically, with population densities of employed individuals in neighborhoods near the freeway increasing
by as much as 2,840 people per square mile after the intervention.
Overall, for neighborhoods within one mile of the freeway
project, population increases by 7 percent in this scenario. Land
prices in these same neighborhoods increase by 2.4 percent.
With this simulation, we can roughly estimate the benefits of
such a project. The simulation provides an estimate of the overall
increase in welfare for the entire regional population. This benefit is derived from the improved amenities in neighborhoods
near the freeway project, and it accounts for general equilibrium
effects that lead to changes in population and employment
throughout the city. Overall, we find that this project alone
leads to the quality-of-life equivalent of a 0.05 percent increase
in income, or roughly $245 million every year for the entire
Philadelphia metropolitan area. Using a discount rate of 7 percent,

The Costs and Benefits of Fixing Downtown Freeways
2022 Q1

this suggests the total lifetime value for the
project is around $3.5 billion.9 This notable
result shows that the benefits of these
projects are on the same order of magnitude as the costs. Projects of this sort
often cost around $500 million per mile,
so the total cost of this project is around
$2.25 billion. Based on these rough estimates, this particular project would pass
a cost-benefit test.
A project like this could be funded using
general tax revenue from the city, state,
or federal government. However, the benefits of the project would accrue mostly
to the surrounding neighborhoods. New
York University professor of finance Arpit
Gupta and his coauthors find that the
Second Avenue subway in New York created value for nearby property owners in
excess of the construction costs. Improvements in local amenities are capitalized
into higher property prices. This suggests

that a targeted tax or assessment could be
used to finance improvements such as the
one proposed here.10
There is significant uncertainty surrounding these estimates. These results
are conservative estimates of quality-oflife benefits. The results change depending
on the assumptions, modeling choices,
and setting. In particular, estimates of parameters that describe how people value
neighborhood amenities vary in the existing literature yet have significant effects on
welfare calculations. If we use values at
the high end of existing estimates, the
benefits of mitigation can increase by 100
percent, whereas low-parameter estimates
can reduce the benefits by about 30
percent. Additional work and more development of quantitative modeling would
improve the precision of these estimates.
Nonetheless, negative quality-of-life effects
are quantitatively important, and targeted

projects like the ones being proposed
or implemented in cities all over the
U.S. may provide important benefits for
central cities.

Conclusion

Economic development was an important
rationale for freeway construction, but not
everyone benefited from the new freeways.
That’s because freeways bring amenities
to some neighborhoods by increasing
access but disamenities to others by reducing the quality of life. Using techniques
developed in recent economic research,
we can quantify neighborhood amenities
and thus the costs and benefits of freeway
construction for individual neighborhoods
and for an entire metro area. Many cities,
including Philadelphia, could benefit from
mitigation of freeway disamenities by covering or capping central city highways.

FIGURE 3

Burying a Portion of I-95 in Philadelphia Would Likely Lead
to a Population Boom in Neighboring Census Tracts

Change in employed residential population density for census tracts in central Philadelphia
if negative neighborhood effects were mitigated for I-95.

ard
Girve
A

676

Market St

Change in Residential Density
People per square mile

76

Broad St

95

Snyder Ave

> 1,000
500 to 1,000
100 to 500
50 to 100
0 to 50
0
0 to −50
−50 to −100
−100 to −500
−500 to −1,000
<−1,000

0
0.5
Miles

1

1.5

2

2.5

Source: Authors’ calculations; U.S. Census Bureau.

The Costs and Benefits of Fixing Downtown Freeways

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

21

Notes

References

1 “Freeway” generally refers to a limited-access highway built for highspeed automobile travel. We use the terms “freeway” and “highway”
interchangeably to refer to these types of limited-access roads.

Ahlfeldt, Gabriel M., Stephen J. Redding, Daniel M. Sturm, and Nikolaus
Wolf. “The Economics of Density: Evidence from the Berlin Wall,”
Econometrica, 83:6 (2015), pp. 2127–2189, https://doi.org/10.3982/
ECTA10876.

2 See Weingroff (1996) for an extensive history of highway building in
the U.S.
3 See Duranton, Morrow, and Turner (2014).
4 Highways reduce the quality of life through other margins, too—for
example, noise or pollution. Given the extensive literature estimating
these effects, we don’t attempt to quantify them. But the spatial scale
of barrier effects seems to exceed the spatial scale of noise or pollution
effects by a large degree.
5 Haughwout et al. (2008) find that average land prices in Manhattan
can be hundreds of times higher per square foot than in suburban
locations just 30 miles away.

Baum-Snow, Nathaniel. “Did Highways Cause Suburbanization?”
Quarterly Journal of Economics, 122:2 (2007), pp. 775–805, https://doi.
org/10.1162/qjec.122.2.775.
Brooks, Leah, and Zachary D. Liscow. “Infrastructure Costs,” working
paper (2019).
Brinkman, Jeffrey, and Jeffrey Lin. “Freeway Revolts!” Federal Reserve
Bank of Philadelphia Working Paper 19-29 (2019), https://doi.org/10.21799/
frbp.wp.2019.29.
Carlino, Gerald A. “Three Keys to the City: Resources, Agglomeration
Economies, and Sorting,” Federal Reserve Bank of Philadelphia Business
Review (Third Quarter 2011), pp. 1-13.

6 For more on this topic, see Carlino (2011).
Duranton, Gilles, Peter M. Morrow, and Matthew A. Turner. “Roads and
Trade: Evidence from the U.S.,” Review of Economic Studies, 81:2 (2014),
pp. 681–724, https://doi.org/10.1093/restud/rdt039.

7 See, for example, Ahlfeldt et al. (2015).
8 See Brooks and Liscow (2019).
9 A discount rate is used to calculate the present value of a project that
will have benefits in the future. The federal Office of Management and
Budget recommends using a discount rate of 7 percent for public infrastructure investments, although state transportation departments often
use lower values, which increases the estimated benefits of a project.
10 Gupta, Van Nieuwerburgh, and Kontokosta (2022) study value capture
and the potential of targeted property taxes.

Gupta, Arpit, Stijn Van Nieuwerburgh, and Constantine Kontokosta. “Take
the Q Train: Value Capture of Public Infrastructure Projects,” Journal of
Urban Economics, 129 (2022) https://doi.org/10.1016/j.jue.2021.103422.
Haughwout, Andrew, James Orr, and David Bedoll. “The Price of Land
in the New York Metropolitan Area,” Current Issues in Economics and
Finance, 14:3 (2008).
Severen, Christopher. “Commuting, Labor, and Housing Market Effects
of Mass Transportation: Welfare and Identification,” Federal Reserve
Bank of Philadelphia Working Paper 18-14 Revised (2019), https://doi.org/
10.21799/frbp.wp.2018.14.
Weingroff, Richard F. “Federal-Aid Highway Act of 1956: Creating the
Interstate System,” Public Roads, 60:1 (1996), pp. 10–17.

22

Federal Reserve Bank of Philadelphia
Research Department

The Costs and Benefits of Fixing Downtown Freeways
2022 Q1

Research Update

These papers by Philadelphia Fed economists,
analysts, and visiting scholars represent
preliminary research that is being circulated
for discussion purposes.

The views expressed in these papers are
solely those of the authors and should not
be interpreted as reflecting the views of
the Federal Reserve Bank of Philadelphia
or Federal Reserve System.

Dynamic Pricing of Credit Cards and
the Effects of Regulation

Individual and Local Effects of Unemployment on
Mortgage Defaults

We construct a two-period model of revolving credit with asymmetric
information and adverse selection. In the second period, lenders exploit an informational advantage with respect to their own customers.
Those rents stimulate competition for customers in the first period.
The informational advantage the current lender enjoys relative to its
competitors determines interest rates, credit supply, and switching
behavior. We evaluate the consequences of limiting the repricing of
existing balances as implemented by recent legislation. Such restrictions increase deadweight losses and reduce ex-ante consumer
surplus. The model suggests novel approaches to identify empirically
the effects of this law. We find the pattern of changes to interest
rates and balance transfer activity before and after the CARD Act are
consistent with the testable implications of the model.

Using survey data from the Panel Study of Income Dynamics, we
document descriptively that unemployment has a relatively large
effect on individual mortgage default rates: The average default rate
for the employed is 2.4 percent; whereas for the unemployed, it is
8.5 percent. Once several other characteristics are controlled for, the
unemployed have default rates that are 4 percentage points larger
than those of the employed; and when endogeneity is additionally
accounted for, the unemployment effect on default rates declines to
3 percentage points. Moreover, we find that more granular metrics
for unemployment entail lower comparable effects of unemployment
on default rates. That is, the comparable effect of individual unemployment on mortgage defaults is rather lower than the effect of
state or county unemployment rates. This finding suggests that local
metrics of unemployment, rather than attenuating possibly large
individual unemployment effects on defaults, indeed contain more
information than the aggregation of these individual effects.

WP 21-38. Suting Hong, Shanghai Tech University; Robert M. Hunt,
Federal Reserve Bank of Philadelphia Consumer Finance Institute;
Konstantinos Serfes, Drexel University and Federal Reserve Bank of
Philadelphia Consumer Finance Institute Visiting Scholar.

WP 21-39. Silvio Rendon, Federal Reserve Bank of Philadelphia
Supervision, Regulation, and Credit Department; Kevin Bazer, Federal
Reserve Bank of Philadelphia Supervision, Regulation, and Credit
Department.

Research Update

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

23

Corporate Disclosure: Facts or Opinions?

Financial Consequences of Severe Identity Theft
in the U.S.

A large body of literature documents the link between textual communication (e.g., news articles, earnings calls) and firm fundamentals,
either through predefined “sentiment” dictionaries or through
machine learning approaches. Surprisingly, little is known about why
textual communication matters. In this paper, we take a step in that
direction by developing a new methodology to automatically classify
statements into objective (“facts”) and subjective (“opinions”) and
apply it to transcripts of earnings calls. The large-scale estimation
suggests several novel results: (1) Facts and opinions are both prominent
parts of corporate disclosure, taking up roughly equal parts, (2) higher
prevalence of opinions is associated with investor disagreement, (3)
anomaly returns are realized around the disclosure of opinions rather
than facts, and (4) facts have a much stronger correlation with contemporaneous financial performance, but facts and opinions have an
equally strong association with financial results for the next quarter.
WP 21-40. Shimon Kogan, IDC Herzliya and the University of
Pennsylvania; Vitaly Meursault, Federal Reserve Bank of Philadelphia
Research Department.

24

Federal Reserve Bank of Philadelphia
Research Department

We examine how a negative shock from severe identity theft affects
consumer credit market behavior in the United States. We show that
the immediate effects of severe identity theft on credit files are
typically negative, small, and transitory. After those immediate
effects fade, identity theft victims experience persistent increases in
credit scores and declines in reported delinquencies, with a significant
proportion of affected consumers transitioning from subprime-toprime credit scores. Those consumers take advantage of their improved
creditworthiness to obtain additional credit, including auto loans and
mortgages. Despite having larger balances, these individuals default
on their loans less than they did prior to the identity theft incident.
WP 21-41. Nathan Blascak, Federal Reserve Bank of Philadelphia
Consumer Finance Institute; Julia Cheney, Federal Reserve Bank of
Philadelphia Consumer Finance Institute; Robert Hunt, Federal
Reserve Bank of Philadelphia Consumer Finance Institute; Vyacheslav
Mikhed, Federal Reserve Bank of Philadelphia Consumer Finance
Institute; Dubravka Ritter, Federal Reserve Bank of Philadelphia
Consumer Finance Institute; Michael Vogan, Ally Bank.

Research Update
2022 Q1

Consumer Credit with Over-Optimistic Borrowers
Do cognitive biases call for regulation to limit the use of credit? We
incorporate over-optimistic and rational borrowers into an incomplete
markets model with consumer bankruptcy. Over-optimists face
worse income risk but incorrectly believe they are rational. Thus, both
types behave identically. Lenders price loans forming beliefs—type
scores—about borrower types. This gives rise to a tractable theory of
type scoring. As lenders cannot screen types, borrowers are partially
pooled. Over-optimists face cross-subsidized interest rates but make
financial mistakes: borrowing too much and defaulting too late.
The induced welfare losses outweigh gains from cross-subsidization.
We calibrate the model to the U.S. and quantitatively evaluate policies
to address these frictions: financial literacy education, reducing
default cost, increasing borrowing costs, and debt limits. While
some policies lower debt and filings, only financial literacy education
eliminates over-borrowing and improves welfare. Score-dependent
borrowing limits can reduce financial mistakes but lower welfare.
WP 21-42. Florian Exler, University of Vienna; Igor Livshits, Federal
Reserve Bank of Philadelphia Research Department; James MacGee,
Bank of Canada; Michèle Tertilt, University of Mannheim.

Assessment Frequency and Equity of the Real
Property Tax: Latest Evidence from Philadelphia
Philadelphia’s Actual Value Initiative, adopted in 2013, creates a unique
opportunity for us to test whether reassessments at short intervals
to true market value and taxing by such values improve equity. Based
on a difference-in-differences framework using parcel-level data
matched with transactions in Philadelphia and 15 comparable cities,
this study finds positive evidence on equity outcomes from more
regular revaluations. The quality of assessment, as measured by the
coefficient of dispersion, improves substantially after 2014, although
the extent of improvement varies across communities. Vertical equity,
measured by price-related differential, also improved, although it was
still above the standard threshold. Cross-city comparisons confirm
Philadelphia’s improvement in quality and equity of assessments after
adopting the initiative. These results highlight the importance of
regular reassessment in places where property values increase quickly,
and they shed light on the disparate impacts of reassessment across
income, property value, race, and gentrification status. The paper
makes the case that the property tax, if designed well, can be an
equitable tax instrument.
WP 21-43. Yilin Hou, Maxwell School, Syracuse University; Lei Ding,
Federal Reserve Bank of Philadelphia Community Development and
Regional Outreach; David J. Schwegman, School of Public Affairs,
American University; Alaina G. Barca, Federal Reserve Bank of Philadelphia Community Development and Regional Outreach.

Research Update

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

25

Democratic Political Economy of
Financial Regulation

Visualization, Identification, and Estimation
in the Linear Panel Event-Study Design
Linear panel models, and the “event-study plots” that often accompany
them, are popular tools for learning about policy effects. We discuss
the construction of event-study plots and suggest ways to make
them more informative. We examine the economic content of different
possible identifying assumptions. We explore the performance of
the corresponding estimators in simulations, highlighting that a given
estimator can perform well or poorly depending on the economic
environment. An accompanying Stata package, xtevent, facilitates
adoption of our suggestions.
WP 21-44. Simon Freyaldenhoven, Federal Reserve Bank of Philadelphia Research Department; Christian Hansen, University of Chicago
and Visiting Scholar, Federal Reserve Bank of Philadelphia Research
Department; Jorge Pérez Pérez, Banco de México; Jesse M. Shapiro,
Brown University, NBER, and Visiting Scholar, Federal Reserve Bank
of Philadelphia Research Department.

This paper offers a simple theory of inefficiently lax financial regulation
arising as an outcome of a democratic political process. Lax financial
regulation encourages some banks to issue risky residential mortgages.
In the event of an adverse aggregate housing shock, these banks
fail. When banks do not fully internalize the losses from such failure
(due to limited liability), they offer mortgages at less than actuarially
fair interest rates. This opens the door to homeownership for young,
low-net-worth individuals. In turn, the additional demand from these
new homebuyers drives up house prices. This leads to a nontrivial distribution of gains and losses from lax regulation among
households. On the one hand, renters and individuals with large
nonhousing wealth suffer from the fragility of the banking system.
On the other hand, some young, low-net-worth households are able
to get a mortgage and buy a house, and current (old) homeowners
benefit from the increase in the price of their houses. When these
latter two groups, who benefit from the lax regulation, constitute
a majority of the voting population, then regulatory failure can be an
outcome of the democratic political process.
WP 22-01. Igor Livshits, Federal Reserve Bank of Philadelphia Research
Department; Youngmin Park, Bank of Canada.

Concentration in Mortgage Markets:
GSE Exposure and Risk-Taking in Uncertain Times
When home prices threaten to decline, large mortgage investors can
benefit from fostering new lending that boosts demand. We ask
whether this benefit contributed to the growth in acquisitions of risky
mortgages by the government-sponsored enterprises (GSEs) in the
first half of 2007. We find that it helps explain the variation of this
growth across regions. The growth predicted by this benefit is on top
of the acquisition growth caused by the exit of private-label securitizers.
We conclude that the GSEs actively targeted their acquisitions to
counter home-price declines.
WP 20-04 Revised. Ronel Elul, Federal Reserve Bank of Philadelphia
Research Department; Deeksha Gupta, Carnegie Mellon University
and Visiting Scholar, Federal Reserve Bank of Philadelphia Research
Department; David Musto, University of Pennsylvania and Visiting
Scholar, Federal Reserve Bank of Philadelphia Research Department.

26

Federal Reserve Bank of Philadelphia
Research Department

Research Update
2022 Q1

Inequality in the Time of COVID-19: Evidence
from Mortgage Delinquency and Forbearance
Using novel data, we show that during the COVID-19 pandemic minority
and lower-income borrowers experienced significantly more financial
distress. We quantify how much the pandemic has exacerbated
inequalities with a difference-in-differences analysis. We then show
that forbearance programs mitigated inequalities as minority and lowerincome borrowers took up forbearances at higher rates, reducing their
delinquency rates more than White and higher-income borrowers in
2020. Finally, we show that minority and lower-income borrowers are
more likely to fall into delinquency and default after exiting forbearance and that fast-tracking FHA modifications with 40-year
terms could best help these borrowers obtain longer-term debt relief.
WP 21-09 Revised. Xudong An, Federal Reserve Bank of Philadelphia
Supervision, Regulation, and Credit Department; Larry Cordell,
Federal Reserve Bank of Philadelphia Supervision, Regulation, and
Credit Department; Liang Geng, Federal Reserve Bank of Philadelphia
Supervision, Regulation, and Credit Department; Keyoung Lee,
Federal Reserve Bank of Philadelphia Supervision, Regulation, and
Credit Department.

Why Are Residential Property Tax Rates Regressive?

Bond Insurance and Public Sector Employment
This paper uses a unique data set of local governments’ bond issuance,
expenditure, and employment to study the impact of the monoline
insurance industry’s demise on local governments’ operations. To show
causality, I use an instrumental variable approach that exploits
persistent insurance relationships and the cross-sectional variation in
insurers’ exposure to high-quality residential mortgage-backed
securities. Governments associated with ailing insurers issued less debt,
cut expenditures, and hired fewer workers. These effects are persistent.
Partial equilibrium calculations show that affected governments’
aggregate expenditures and employment levels in 2017 would
have been 6 percent to 10 percent higher if bond insurance had
remained available.
WP 22-03. Natee Amornsiripanitch, Federal Reserve Bank of
Philadelphia Supervision, Regulation, and Credit Department.

Among single-family homes that enjoy the same set of property taxfunded amenities and pay the same statutory property tax rate, owners
of inexpensive houses pay almost 50 percent higher effective tax rates
than owners of expensive houses. This pattern appears throughout
the U.S. and is caused by systematic assessment regressivity—
inexpensive houses are overassessed relative to expensive houses. I use
an instrumental variable approach to show that a large portion of this
pattern can be attributed to measurement error in sale prices. Sixty
percent of the remaining regressivity can be explained by tax assessors’ flawed valuation methods that ignore variation in priced house
and neighborhood characteristics and 40 percent by infrequent
reappraisal. A simple valuation method can alleviate assessment
regressivity and increase poor homeowners’ net worth by more
than 10 percent.
WP 22-02. Natee Amornsiripanitch, Federal Reserve Bank of
Philadelphia Supervision, Regulation, and Credit Department.

Research Update

2022 Q1

Federal Reserve Bank of Philadelphia
Research Department

27

The Great Migration and Educational Opportunity
This paper studies the impact of the First Great Migration on children.
We use the complete-count 1940 Census to estimate selectioncorrected place effects on education for children of Black migrants. On
average, Black children gained 0.8 years of schooling (12 percent) by
moving from the South to the North. Many counties that had the
strongest positive impacts on children during the 1940s offer relatively
poor opportunities for Black youth today. Opportunities for Black
children were greater in places with more schooling investment,
stronger labor market opportunities for Black adults, more social capital,
and less crime.
WP 22-04. Cavit Baran, Northwestern University; Eric Chyn, NBER
and Dartmouth College; Bryan A. Stuart, Federal Reserve Bank of
Philadelphia Research Department.

Net Income Measurement, Investor Inattention,
and Firm Decisions
When investors have limited attention, does the way in which net
income is measured matter for firm value and firms’ resource allocation
decisions? This paper uses the Accounting Standards Update (ASU)
2016-01, which requires public firms to incorporate changes in
unrealized gains and losses (UGL) on equity securities into net income,
to answer this question. We build a model with risk-averse investors
who can be attentive or inattentive and managers who choose how
much to invest in financial assets to maximize firms’ stock prices. The
model predicts that, with inattentive investors, stock prices react
more to changes in UGL from equity securities under the new regime
and, under certain conditions, investors assign larger price discounts.
Managers respond to such discounts by cutting financial asset
holdings. We use insurance company data to test these predictions.
Prices of stocks with low analyst coverage react more to changes
in UGL from equity securities, highlighting the role of investor inattention. Using a difference-in-differences approach, we find that
by 2020, publicly traded insurance companies cut investments in
public stocks by $23 billion.
WP 22-05. Natee Amornsiripanitch, Federal Reserve Bank of
Philadelphia Supervision, Regulation, and Credit Department;
Zeqiong Huang, Yale School of Management; David Kwon, Yale
School of Management; Jinjie Lin, Yale School of Management.

28

Federal Reserve Bank of Philadelphia
Research Department

Research Update
2022 Q1

Data in Focus

ATSIX

The Philadelphia Fed collects, analyzes, and shares useful data
about the Third District and beyond. Here’s one example.

T

he Federal Open Market Committee
(FOMC) responded to the Great
Recession of 2007–2009 with unprecedented measures, cutting interest rates
to low levels and buying financial assets.
But some wondered at the time, might
those measures induce higher inflation?
The public’s expectations for higher
inflation can be self-fulfilling, so, if Americans believe that these measures will
induce higher inflation, higher inflation
might just occur, and that in turn can
alter the effectiveness of the FOMC’s target
interest rates.
Generally, the best way to gauge
inflation expectations is by surveying economists and consumers, but those surveys
typically ask respondents to give their
inflation expectations for specific dates in
the future. What policymakers need is
a continuous curve of inflation expectations, not just expectations for specific
future dates.
That’s why Philadelphia Fed visiting
scholar S. Borağan Aruoba of the University
of Maryland developed the Aruoba Term
Structure of Inflation Expectations (ATSIX),
a smooth, continuous curve of inflation
expectations three to 120 months ahead.1
Aruoba’s model optimally combines the
Philadelphia Fed’s Survey of Professional
Forecasters2 with two surveys published by
Wolters Kluwer Law & Business. Aruoba
found that his model’s inflation expectations “track realized inflation quite well,
and in terms of forecast accuracy, they
are at par with or superior to some popular alternatives.”3
With the FOMC once again taking extraordinary measures (this time to counter
a devastating economic shock resulting
from the COVID-19 pandemic), now is
a good time to revisit ATSIX and see what
it tells us about expectations for future
inflation.

ATSIX: Term Structure of Inflation Expectations
Average Annualised Expected CPI Inflation, percentage points
3.4
3.2
3.0
2.8
2.6
2.4

Feb 2022
Jan 2022
Feb 2021

2.2
2.0

0
24
Months ahead

48

72

96

120

Source: Research Department, Federal Reserve Bank of Philadelphia.

Notes
1 The ATSIX methodology was developed from
research initially funded by the Federal Reserve
Bank of Minneapolis.
2 https://www.philadelphiafed.org/surveysand-data/real-time-data-research/survey-ofprofessional-forecasters
3 S. Borağan Aruoba, “Term Structures of
Inflation Expectations and Real Interest Rates,”
Philadelphia Federal Reserve Working Paper
16-09/R (2016).

Learn More
Online: https://www.philadelphiafed.
org/surveys-and-data/real-time-dataresearch/atsix
E-mail: tom.stark@phil.frb.org

PRESORTED STANDARD
U.S. POSTAGE
PAID
PHILADELPHIA, PA
PERMIT #583

Ten Independence Mall
Philadelphia, Pennsylvania 19106-1574

ADDRESS SERVICE REQUESTED

From Philadelphia, Pennsylvania, the birthplace of American finance, comes Economic Insights.

2017 Q2

2018 Q1

2019 Q1

2020 Q1

Featuring articles about how Americans work, invest, spend…and so much more.

2021 Q1

2021 Q2

2021 Q3

2021 Q4