Content uploaded by Andrew Winton
Author content
All content in this area was uploaded by Andrew Winton on Dec 23, 2020
Content may be subject to copyright.
THE JOURNAL OF FINANCE •VOL. LXX, NO. 6 •DECEMBER 2015
Risk Overhang and Loan Portfolio Decisions:
Small Business Loan Supply before and during
the Financial Crisis
ROBERT DEYOUNG, ANNE GRON, GӦKHAN TORNA, and ANDREW WINTON∗
ABSTRACT
We estimate a structural model of bank portfolio lending and find that the typical U.S.
community bank reduced its business lending during the global financial crisis. The
decline in business credit was driven by increased risk overhang effects (consistent
with a reduction in the liquidity of assets held on bank balance sheets) and by reduced
loan supply elasticities suggestive of credit rationing (consistent with an increase
in lender risk aversion). Nevertheless, we identify a group of strategically focused
relationship banks that made and maintained higher levels of business loans during
the crisis.
SMALL BUSINESSES,DEFINED AS HAVING fewer than 500 employees, employ about
one-half of the U.S. labor force and create nearly two-thirds of net new private
sector jobs in the United States annually (U.S. Small Business Administration
(2014)). Virtually all of these small firms are privately held and lack access to
public capital markets. To ensure access to credit, these informationally opaque
businesses establish close borrower–lender relationships with small “commu-
nity banks” (e.g., Petersen and Rajan (1994), Berger et al. (1997), Berger et al.
(2005)). This confluence of small firms and small banks is important for macroe-
conomic growth both in the United States and elsewhere: Berger, Hasan, and
Klapper (2004) find a strong positive link between a large, healthy small bank-
ing sector and macroeconomic growth across 49 developed and developing
nations.
The financial crisis took a toll on the U.S. small banking sector. About 6%
of all commercial banks and thrift institutions failed between 2007 and 2012,
∗DeYoung is at Kansas University. Gron is at NERA Economic Consulting. Torna is at State
University of New York at Stony Brook. Winton is at University of Minnesota. The opinions
expressed in this paper do not necessarily reflect the views of NERA Economic Consulting. We
thank three anonymous referees, Allen Berger, Lamont Black, Paolo Fulghieri, Ted Juhl, Evren
Ӧrs, Michael Roberts (the Editor), Elu von Thadden, Greg Udell, and seminar participants at
Bangor University, the Bank of Canada, the Federal Deposit Insurance Corporation, the Federal
Reserve Bank of Chicago, the Federal Reserve Bank of New York, the Financial Intermediation
Society, the University of Groningen, the University of Kansas, the University of Limoges, and the
University of Mannheim for their insightful comments and suggestions. The opinions expressed in
this paper do not necessarily reflect the views of NERA Economic Consulting. The authors have no
material financial or non-financial interests related to this research, as identified in the Journal
of Finance’s disclosure policy.
DOI: 10.1111/jofi.12356
2451
2452 The Journal of Finance R
and 411 of the 478 insolvencies were small institutions with assets less than
$1 billion.1As the FDIC fashioned resolutions for these insolvent banks, it is
understandable that some of their small business clients would suffer interrup-
tions to, reductions in, or even outright loss of their credit lines.2But whether
the stress of the financial crisis caused healthy banks in the United States
to reduce credit to small and medium enterprises (SMEs) remains an open
question. A reduction in SME credit supply by healthy banks—that is, a credit
crunch or credit rationing—would have had procyclical effects, exacerbating
the economic downturn by denying firms the short-term credit necessary to
finance increased inventories and retain workers. This would be antithetical to
the idea of small business relationship banking, which carries with it the pre-
sumption that additional credit will be available when needed. In this paper,
we investigate whether, how, and why small U.S. banks reduced small business
credit supply during the financial crisis.
Some evidence has emerged from European economies, where credit reg-
istries provide researchers highly detailed data on loans and loan applications,
that the financial crisis was accompanied by reduced credit supply to SMEs
(e.g., Popov and Udell (2012), Cotugno, Monferra, and Sampagnaro (2012),
Jimen´
ez et al. (2012)). These studies document that credit supply declined
more during the crisis as banks experienced financial stress (low levels of eq-
uity capital, poorly performing loan portfolios) but to a lesser extent for SMEs
with strong bank–borrower relationships. While this evidence is informative,
it remains incomplete. First, none of the extant studies examine SME credit in
the United States, where research has focused on the syndicated loan supply
to large firms during the crisis (e.g., Ivashina and Scharfstein (2010a,2010b),
Chodorow-Reich (2014)) due to a lack of systematic loan-level data for SMEs.
Given that the business, banking, and financial environments in the United
States and Europe are substantially different, one cannot assume that small
business credit supply behaved similarly on both sides of the Atlantic. Sec-
ond, while the studies cited above find well-identified statistical associations
between macroeconomic conditions and SME credit supply, they stop short of
modeling the underlying lender behaviors that drive these empirical associa-
tions and the channels through which these associations might occur. Third,
these studies typically do not differentiate banks by their business strategies,
an important distinction that could drive banks’ lending behaviors during eco-
nomic downturns.
We derive a theoretical loan supply function from a model of loan portfolio
optimization with market imperfections, following Froot and Stein (1998). In
our model, a capital-constrained bank (i.e., one facing imperfect capital mar-
kets) might reject an otherwise profitable lending opportunity if it cannot make
1Data from the Federal Deposit Insurance Corporation website.
2The FDIC arranged “purchase and assumption” resolutions for 427 of these failed banks. In
these transactions, the FDIC arranges for a healthy bank to acquire all of the assets of the failed
bank, so clients of these failed banks were unlikely to fully lose access to credit. In the other 51
bank insolvencies, the FDIC seized the failed bank’s assets and disposed of them piecemeal over
time; clients of these banks were more likely to fully lose access to new credit.
Risk Overhang and Loan Portfolio Decisions 2453
room on its balance sheet by cheaply selling off any of its existing loans (i.e.,
imperfect loan markets) and the returns on its existing loans covary positively
with the returns on the new lending opportunity. The illiquidity of its existing
loans, which we characterize as “loan overhang,” combined with costly external
capital effectively makes the bank risk averse, that is, less likely to make risky
but positive NPV loans. Because capital and credit markets tend to become
more imperfect during recessions, our theory predicts increased lender risk
aversion and reduced SME loan supply during the financial crisis. Thus, in the
course of testing empirically whether U.S. banks reduced and/or rationed credit
to SMEs during the financial crisis, we also perform an important empirical
test of financial intermediation theory.
We estimate the model using quarterly observations of community banks
with assets less than $2 billion (2010 dollars) operating in U.S. metropoli-
tan markets between 1991 and 2010.3These banks are too small to make or
participate in loans to large firms, and they tend to hold mixed portfolios of
consumer loans, real estate loans, and SME loans. Importantly, these banks
are typically consistent with the maintained assumptions of our theory model:
SME loans are illiquid assets and must be held in portfolio where they lock
up scarce equity capital; small banks are seldom publicly traded, rarely have
public credit ratings, and face relatively inelastic deposit markets, all of which
make external capital expensive; and the owners of these banks tend to invest
a disproportionate share of their (or their family’s) wealth in the bank, and this
lack of diversification further encourages risk-averse business practices (Spong
and Sullivan (2007)).
Our empirical results largely confirm the predictions of our theory. We find
strong evidence of loan overhang effects. All else equal, banks make fewer
new business loans when their portfolios contain large amounts of preexisting
business loans, and make more new business loans when their portfolios con-
tain large amounts of loans to other sectors (e.g., consumer loans) that covary
negatively with business loans. Consistent with increases in loan illiquidity
and lender risk aversion during times of economic uncertainty, the negative
impact of overhanging business loans on new business lending is substantially
stronger during the financial crisis. Regarding the main question of our inves-
tigation, we find that new SME lending declined during the financial crisis for
the average bank in our data, and this finding looks like credit rationing: prior
to the crisis there is a strong positive relationship between business loan sup-
ply and the expected returns on business loans, but during the crisis business
loan supply became insensitive (perfectly inelastic) to expected loan returns.
This new evidence complements the finding of Montorial-Garriga and Wang
(2012) that small business borrowers were more likely to be rationed out of the
bank loan market during the financial crisis than large firms. Importantly, we
identify a small set of banks that strategically focused on making illiquid loans
3For decades, both bank regulators and bank researchers used $1 billion as a convenient upper
size threshold to define the U.S. community bank sector (DeYoung, Hunter, and Udell (2004)). Our
$2 billion threshold is similarly convenient, but recognizes several decades of inflation.
2454 The Journal of Finance R
to commercial entities, and we find that these banks increased their supply of
SME loans during 2008 (the first full year of the crisis) and maintained this
higher level of SME credit supply during both 2009 and 2010. This result sug-
gests that borrower–lender relationships help mitigate credit supply shocks
to small businesses, in line with the findings of non-U.S. studies based on
loan-level data (Cotugno, Monferra, and Sampagnaro (2012) for Italian banks,
Liberti and Sturgess (2012) for a single multinational lender).
While our results confirm the primary findings of recent investigations of
SME lending in Europe during the global financial crisis, namely, that supply-
side phenomena were important drivers of the reduced credit availability for
small businesses, our findings extend this literature in important ways. First,
we show that the financial crisis induced reductions in SME lending not only
in the bank-based financial systems of Europe, but also in the United States,
where SMEs still rely on bank finance even though larger firms are able to
access capital markets. Second, we devise well-identified empirical tests of the
impact of the financial crisis on SME lending even in the absence of loan-level
data or a convenient natural experiment. Third, our empirical specifications
are informed directly by a theory of loan supply with underlying assumptions
that closely match the real-world conditions facing small bank lenders. As
such, our empirical tests reveal information about the underlying drivers of
lender behavior such as loan illiquidity, loan portfolio composition, loan perfor-
mance covariances, and lender risk aversion. Fourth, we find that the impact
of these underlying drivers (and of economic downturns) on bank loan sup-
ply is conditional on banks’ ex ante lending strategies. In particular, we show
that bank–business relationship lending—characterized here as a strategic
pre-commitment made by bank lenders—can partially mitigate procyclical re-
ductions in SME loan supply. Fifth, our results suggest that the reduction in
the quantity of credit to SMEs during the crisis was caused not by increased
pricing of credit risk, but rather by quantity rationing of credit in the face of
increased economic uncertainty.
The organization of the paper is as follows. In Section I, we derive a theoreti-
cal loan supply function. In Section II, we show that the characteristics of U.S.
community banks are in line with the maintained assumptions of our theory
model, and as such provide a natural venue for testing the model’s predictions
for SME loan supply. In Section III, we present the data and variables used in
our regression models. In Section IV, we discuss our main testable hypotheses
and describe our empirical identification scheme. In Section V, we present our
main empirical results. In Section VI, we discuss the implications of our results
for policy.
I. Loan Supply with Capital Market Imperfections: Theory
Froot, Scharfstein, and Stein (1993) predict that, when external finance is
costly, value-maximizing firms make investment decisions in a risk-averse
manner: they base decisions not only on the expected returns from the in-
vestment opportunity in question, but also on the new investment’s impact on
Risk Overhang and Loan Portfolio Decisions 2455
the risk of their business and on their stock of available investment capital.
These considerations increase a firm’s expected profits by reducing the prob-
ability that it will forgo a valuable future investment opportunity when the
return on the prospective investment does not justify the costs of raising addi-
tional external capital—either because the firm has too little internal capital
to make the investment or it is unable to free up internal capital by selling off
lower-yielding assets because they are illiquid. Froot and Stein (1998) apply
this theory to banks. In their role as delegated monitors, banks have private
information that makes their loans relatively or completely illiquid. If existing
loans cannot be cheaply sold off, and if the returns on these existing exposures
covary positively with the returns on new lending opportunities, then capital
constrained banks will make fewer new loans. Similarly, if a bank with largely
illiquid existing loans suffers a reduction in its equity capital, then the bank
will also make fewer new illiquid loans. Gron and Winton (2001) apply this
model to data from insurance markets, and find that outstanding illiquid risk
exposures from long-term insurance policies—which they call risk overhang—
can reduce the current supply of new insurance policies.4
Following this literature, we develop a portfolio model of bank loan supply
when external finance is costly due to imperfections in capital markets and
credit markets. We begin with a representative bank that has lending oppor-
tunities in several sectors. Loans can be funded out of net internal capital W
or external funds F, where external funds are assumed to be more costly than
internal funds. This additional cost reflects information asymmetries and con-
flicts between the firm and outside investors (e.g., Myers and Majluf (1984),
Holmstr¨
om and Tirole (1997), DeMarzo and Duffie (1999)), as well as other
transaction costs in accessing public markets. In addition to current-period
loans, the bank may be able to make profitable loans in future periods. As
shown by Froot, Scharfstein, and Stein (1993), profitable future investment op-
portunities combined with costly external funds and stochastic internal funds
cause the firm’s objective function to be increasing and generally concave in
the stock of internal funds. Intuitively, more internal funds lessen the extent
to which a bank must rely on costly external funds, but this benefit is generally
decreasing because, at the margin, there are fewer profitable uses for these
funds. Denoting the indirect form of the bank’s objective function as P(W), we
have PW>0andPWW <0, where the subscript denotes the partial derivative.
The bank begins period twith Wt–1in net internal funds, Lt– 1,i in outstanding
loans in each sector i, and net external finance of Ft–1=ࢣi(Lt– 1,i)–Wt–1>
0. For simplicity, we assume that all external finance takes the form of debt.5
4Other applications to insurance markets include Froot and O’Connell (1999) and Froot (2007).
The former use data from catastrophe reinsurance markets and show that financial market im-
perfections can lead to both costly reinsurer capital and reinsurer market power. The latter adapts
the existing theory to describe insurance companies and introduces product market imperfections
and risk hedging to the model.
5Regardless of its form, external finance is costly for banks. Because banks face minimum
equity-to-asset rules, issuing a substantial amount of external debt requires that they also raise
new equity finance. Indeed, Berger et al. (2008) show that, when commercial banks fall closer
2456 The Journal of Finance R
For the moment, assume that all of the bank’s outstanding loans are illiquid
and cannot be sold due to the bank’s private information on loan quality. Since
the bank must bear the risk of Lt–1,i loans in each sector iregardless of its
subsequent decisions in period t,Lt–1,i is the bank’s risk overhang in sector iin
period t.
During period tthe bank can make new loans NLt,i ࣙ0 to each sector i,
resulting in end-of-period outstanding debt of Ft=ࢣi(Lt–1,i +NLt,i)–Wt– 1.
The gross per dollar cost of debt funding is 1+rt, which includes any costs of
accessing external markets rather than using internal capital. During period
t, the bank realizes the gross per dollar return of ˜
Rt,i/t−1on loans to sector
ithat were originated in period t–1. The return ˜
Rt,i/t−1equals 1+rt+pt–1,i –
˜ηt,i, where pt– 1,i is the per dollar credit spread or markup charged on sector i
loans that originated in period t–1, and ˜ηti is the random per dollar loan losses
on sector iloans in period t. Similarly, the bank realizes the gross per dollar
return ˜
Rt,i/t=1+rt+pt,i –˜ηt,i on the new loans to sector ioriginated in period
t, where pt,i is the per dollar credit spread on these loans. For simplicity, we
assume that all losses on loans to sector iborrowers in period tare perfectly
correlated, regardless of when the loan was made. Current-period loan losses
are assumed to be normally distributed: ˜ηt,i∼N(μt,i,σ
t,i),where μt,i and σt,i
depend on the sector’s economic outlook at the start of period t.6Both μt,i and
σt,i are decreasing in the sector’s economic outlook: when borrowing firms have
better prospects, both ex ante credit risk and ex post realized loan losses are
lower because the borrowing firms’ chances of default are reduced. Given these
assumptions, it follows that the bank’s net capital at the end of period tis
˜
Wt=
n
i=1
[Lt−1,i˜
Rt,i/t−1+NL
t,i˜
Rt,i/t]−Ft(1 +rt)
=W0(1 +rt)+
n
i=1
[Lt−1,i(pt−1,i−˜ηt,i)+NL
t,i(pt,i−˜ηt,i)],
(1)
where we make use of the definitions of ˜
Rt,i/t−1,˜
Rt,i/t,andFt.
to their regulatory minimums, they actively manage their capital to return quickly to their in-
ternal capital targets. Issuing new equity involves significant transaction and information costs,
especially for banking companies that are not publicly traded (the majority of the industry). For
banks that are not too-big-too-fail (again, the majority of the industry), issuing subordinated debt
or large-denomination deposit contracts also entails such costs. Moreover, federally insured retail
deposit contracts are not perfect, costless substitutes for uninsured debt. Billett, Garfinkel, and
O’Neal (1998) find that total debt finance (insured plus uninsured liabilities) declines at large
banks following downgrades of their publicly traded debt, consistent with increased external costs
of debt finance. Further support that external funding is costly for banks comes from Jayaratne
and Morgan (2000), who find that banks finance an unusually large portion of their assets with
internal funds. Finally, Amel and Hannan (1999) show that insured deposits are relatively price
inelastic, so that banks wanting to raise large additional amounts of these funds must pay higher
rates.
6The assumption of normality allows us to give a simple, tractable analytic solution to the
bank’s portfolio choice problem. In reality, loan losses are skewed to the right due to infrequent
but concentrated episodes of correlated borrower distress during economic downturns.
Risk Overhang and Loan Portfolio Decisions 2457
The bank chooses new loan amounts NLt,i that maximize expected profit
E[P(˜
Wt)] given the financing constraints. This leads to the first-order condition
for each sector i
0=EPW
∂˜
Wt
∂NL
t,i=EPW(pt,i−˜ηt,i)=EPW(pt,i−μt,i)−Cov(PW,˜ηt,i),(2)
where we make use of (1) and the identity E(xy)=E(x)E(y)+Cov(x,y). Since
loan losses ˜ηt,iand the level of internal funds ˜
Wtare both normally distributed,
we can apply Stein’s Lemma and the definition of covariance to derive the
bank’s supply of new loans NLS
t,ito sector i
NLS
t,i=−
j=i
NLS
t,j
σij
σii
−Lt−1,i−
j=i
Lt−1,j
σij
σii
+1
G·pt,i−μt,i
σii
,(3)
where σii is the variance of loan losses in sector iover time, σij is the covariance
of loan losses across sectors iand jover time, and G=−E[PWW]
E[PW]measures the
bank’s effective risk aversion induced by the costs of external finance (we refer
to its reciprocal, 1/G, as the bank’s risk tolerance).7For convenience, we sup-
press the time subscripts on each of these terms. The bank’s supply of new loans
to sector iis determined by several factors on the right-hand side of equation
(3). The first term is the effect of covariance-adjusted lending opportunities in
other sectors jࣔiattimet. The second term is the preexisting portfolio expo-
sure in sector i, that is, the overhang of outstanding loans in sector iat time t.
The third term is the effect of the covariance-adjusted loan overhangs in other
sectors jࣔi. The final term is the bank’s risk tolerance 1/Gmultiplied by the
risk-adjusted returns ratio (pt,i –μt,i)/σii .
We complete the model by relaxing our assumption that loans are completely
illiquid. As shown by Froot and Stein (1998), under optimal portfolio allocation
with imperfect capital markets, banks will shed any existing loan that can be
sold at fair value. For some loans, however, the market price may be less than
banks’ expected (fair) value due to information asymmetries or transactions
costs—the bank will hold rather than sell these illiquid loans. Let δt–1,i ∈(0,1)
be the illiquid portion of the outstanding loans at the beginning of period t(end
of period t–1). Since only illiquid loan stocks will affect new lending (liquid loan
stocks can be sold off at no cost to make room for new loans), we can rewrite
equation (3) as
NLS
t,i=−
j=i
NLS
t,j
σij
σii
−δt−1,iLt−1,i−
j=i
δt−1,jLt−1,j
σij
σii
+1
G·pt,i−μt,i
σii
,(4)
where δt–1,i Lt–1,i and δt– 1,j Lt–1,j represent the stocks of illiquid outstanding (i.e.,
overhanging) loans.
7Stein’s lemma implies Cov(PW,˜ηt,i)=E[PWW ]Cov( ˜
Wt,˜ηt,i). We also use Cov( ˜
Wt,˜ηt,i=
−j(Lt−1,j+NL
t,j)σi,j.
2458 The Journal of Finance R
Equation (4) provides the following predictions. First, the supply of new loans
to sector iis decreasing in the outstanding illiquid loans δt–1,i
Lt–1,i in that
sector. Second, the supply of new loans to sector iis decreasing (increasing)
in the outstanding illiquid loans δt–1,j
Lt–1,j in sector jif the covariance σij is
positive (negative). Similarly, the supply of new loans to sector iis decreasing
(increasing) in the supply of new loans NLS
t,jto sector jif the covariance σij is
positive (negative). Finally, the supply of new loans to sector iis increasing in
both sector irisk-adjusted returns (pt,i –μt,i)/σii and the bank’s risk tolerance
1/G(assuming reasonably that (pt,i –μt,i)/σii >0). Given that these two final
terms are multiplicative, we can say that the expected return elasticity of new
loan supply is increasing in bank risk tolerance.
II. Market Imperfections and Small Bank Lenders
Our model is especially descriptive of the business lending environment faced
by small commercial banks. These banks are too small to make or participate
in business loans to large publicly traded firms; instead, they provide credit
to small, privately held businesses that are opaque to public capital markets.
These loans are typically based on close bank–business relationships that per-
mit the bank to observe soft (i.e., not quantifiable) information about the bor-
rower’s creditworthiness (Stein (2002)). Although relationship loans are not
based solely on soft information—for example, banks usually require collateral
for which a hard value can be determined—these loans remain far less liquid
than loans based solely upon quantifiable information.8Relationship loans are
not securitizable and can be sold to other banks only at large discounts, because
the informational value of the borrower–lender relationship cannot be credi-
bly conveyed to outside investors. When a bank makes a relationship loan it
knows that such an option does not exist, and this illiquidity discourages larger
hard information-based lenders from making these types of loans. Berger et al.
(2005) find evidence consistent with this description.
The real estate loans and consumer loans made by small banks also tend
to be less liquid than those made by larger banks. Large banks originate with
the intent to securitize large portions of their real estate loans (e.g., residential
mortgages, home equity lines of credit) and consumer loans (e.g., auto loans,
student loans, credit card receivables). Although the originate-and-securitize
production process generates additional costs not present in portfolio lend-
ing (e.g., legal and credit rating agency fees, establishing a reputation in the
asset-backed securities market, providing credit enhancements to buyers of the
8While a large portion of these loans have short maturities, confusing maturity with liquidity
belies the nature of the long-term borrower-lender relationship at the core of small banks’ business
lending strategies. All else equal, community banks will be reticent to allow these loans to roll off
their balance sheets, as this represents a loss of intangible relationship value in which the bank
has invested. Moreover, as in Rajan (1992), the borrowers are likely to face informational lock-in
costs if they try to repay their lender by seeking other sources of finance. Thus, the short (usually
one-year) contractual maturities of small business credit lines are better interpreted as a risk
management tool that provides a periodic opportunity for adjusting loan terms and prices.
Risk Overhang and Loan Portfolio Decisions 2459
asset-backed securities), these additional expenses can be offset by reduced ex-
penses for credit screening, increased revenue from mortgage origination and
servicing fees, and various scale economies associated with this production
process. Loan securitization is antithetical to the small bank business model:
high loan origination volume is necessary to run this process efficiently, and
selling off rather than holding loans runs counter to close bank–borrower rela-
tionships. Small banks retain the lion’s share of the real estate and consumer
loans that they originate; the principle exception is conforming home mortgage
loans sold to government-sponsored enterprises such as Fannie Mae, Freddie
Mac, and Ginnie Mae.
Small banks lack access to public funding markets; this increases their cost
of external financing, which in turn magnifies the consequences of new lending
decisions. Credit derivatives are not a viable hedging strategy—credit default
swaps (CDS) do not exist for small business loans, and using CDS to hedge these
loans would in any case entail extreme basis risk—so small banks manage
credit risk by limiting credit concentrations, adjusting portfolio loan shares,
and holding equity capital cushions. Small bank CEOs are often placing their
own undiversified family capital at risk when making lending decisions, so risk-
averse lending behavior should be relatively free of potentially confounding
principal-agent effects (Spong and Sullivan (2007)). For all these reasons, small
banks should be especially sensitive to the frictions that lie at the core of our
theoretical model.
Consistent with the above discussion, we adopt the following broad definition
of loan illiquidity: any friction that (a) prevents the bank from selling a loan
prior to its contractual maturity, thus reducing its current ability to redeploy
capital, or (b) causes the bank to retain a loan beyond its contractual maturity,
thus disrupting its future plans for redeploying capital. The former condition
is consistent with the more usual concept of loan illiquidity in which a bank is
unable to sell loans in secondary (securitized or syndicated) loan markets. The
latter condition is consistent with illiquidity that arises from a bank’s long-run
lending strategies or from frictions in related markets. Condition (b) could oc-
cur, for example, when a bank renegotiates the terms of a delinquent business
loan by extending its contractual loan maturity in an attempt to salvage the
long-run value of a lending relationship. Alternatively, condition (b) could occur
when a bank forecloses on a delinquent business loan and seizes loan collateral
that, due to its illiquidity, may have to be carried on the bank’s books beyond
the maturity date of the loan. Still another example of (b) would be a situation
in which the bank had assumed other lenders would be willing to refinance
the borrower at the loan’s maturity, but a deterioration in the borrower’s con-
dition or the general business climate makes this less likely (Rajan (1992)). In
each of these cases, banks must hold assets beyond their planned-for maturity
dates; as the likelihood of these illiquidity events increases, so should loan
overhang effects. Historical correlations of loan default frequencies with the
business cycle (Furlong and Knight (2010)) imply that condition (b) illiquidity
events should occur more often for business and commercial real estate loans,
especially during economic downturns.
2460 The Journal of Finance R
Data from the fed funds market—a major source of short-run liquidity for
small and large banks alike—suggest that banks were experiencing tighter
liquidity conditions during the financial crisis. Small U.S. banks (defined here
as banks with less than $2 billion in assets) tend to be deposit-rich, while large
U.S. banks (defined here as banks with more than $50 billion in assets) tend
to be deposit-poor, and in normal times the fed funds market transfers excess
liquidity from small banks to large banks. Prior to the crisis, fed funds sold
fluctuated between 3% and 5% of small bank assets, but fell to only about 2% of
small bank assets during the crisis. Similarly, fed funds purchased fluctuated
between 10% and 12% of large bank asset funding prior to the crisis, but
plunged to about 5% during the crisis. These two developments were strongly
linked: the quarterly time series correlation between small bank fed funds sold-
to-assets and large bank fed funds purchased-to-assets during the 17 years
leading up to the crisis was –0.03, but was 0.57 during the crisis years (2008
to 2010). (All calculations in this paragraph were performed by the authors
using data from the call reports.) Hence, both large banks and small banks
experienced unusual liquidity pressure during the crisis: small banks felt it
necessary to hold higher stores of precautionary liquidity, resulting in a reduced
supply of liquidity to large banks.
III. Data and Variables
We estimate our model using quarterly financial statement data on U.S. com-
mercial banks, taken from the Reports of Condition and Income (call reports),
from the fourth quarter of 1991 through the fourth quarter of 2010. The first 64
quarters of these data comprise the pre-crisis period (1991:Q4 though 2007:Q3)
and the final 13 quarters comprise the crisis period (2007:Q4 through 2010:Q4).
Because our main objective is to determine whether, how, and why bank SME
lending changed during the global financial crisis, we choose the beginning and
end dates of the crisis period based on available information about the lending
behaviors of U.S. banks. The Federal Reserve’s Senior Loan Officer Opinion
Survey on Bank Lending Practices (SLOOS) is administered four times each
year to a relatively stable set of around 55 large and medium-sized U.S. com-
mercial banks. Among other questions, the survey asks each bank whether its
credit standards for approving small business loan applications have eased,
remained unchanged, or tightened over the past three months. The net per-
centage of banks tightening their small business lending standards exceeded
10% for the first time in the January 2008 SLOOS, so we mark 2007:Q4 as the
beginning of the crisis period. The net percentage of banks easing their small
business lending standards exceeded 10% for the first time in the April 2011
SLOOS, so we mark 2010:Q4 as the final quarter of the crisis period.9
9In the January 2008 SLOOS, 17 banks tightened standards, 39 did not change their standards,
and 0 eased their standards. Thus, the net percentage of banks that tightened standards =(17–
0)/56 =30.4%, up from just 9.6% in the previous survey. In the April 2011 SLOOS, 0 banks
tightened standards, 45 did not change their standards, and 7 eased their standards. Thus, the net
Risk Overhang and Loan Portfolio Decisions 2461
We apply three primary filters to select the banks to include in our data set.
First, we only include so-called community banks with less than $2 billion in
assets in real 2010 dollars. Second, we only include banks located in metropoli-
tan statistical areas (MSAs); banks located in rural areas face a different set of
lending opportunities from urban banks, which results in different exposures
to loan overhang and different incentives for dealing with this risk.10 Third, we
only include banks that make relatively balanced amounts of business loans,
real estate loans, and consumer loans—the three main types of loans as cate-
gorized in the call reports. We define these “nonspecialist” lenders as follows
in each quarter of the data: the dollar value of their sector iloans must be no
more than 10 times, and no less than one-tenth, the dollar value of their sector
jloans (iࣔj).
We make a number of additional adjustments to mitigate the potential effects
of data errors, merging banks, or banks with abrupt changes in lending strate-
gies. We delete bank-quarter observations in which the assets of another bank
are acquired, bank-quarters when banks are less than five years old or less
than $25 million in assets, all observations for banks that lend out fewer than
25% of their assets, and all observations for banks not present in the data for
at least five consecutive quarters. We delete bank-quarter observations when
the ratio of nonperforming loans to beginning-of-period loans, the ratio of net
lending change to beginning-of-period assets, the quarterly change in assets,
or the quarterly change in equity capital are over the 99th percentile or below
the 1st percentile of the sample distributions. Similarly, we omit bank-quarter
observations when the expected profit variable in any of the three loan sec-
tors is less than the 0.5th or greater than the 99.5th percentile of the sample
distribution.
In our empirical model, the portfolio Lof existing loans at the end of quarter
t–1 consists of business loans (BUS), real estate loans (RE), and consumer loans
(CON). Business loans, BUS, include all commercial and industrial loans. Real
estate loans, RE, include all loans secured by a lien on real estate: construc-
tion and development loans, first and second mortgages on single family and
multi-family residential properties, and mortgages on commercial properties.
Consumer loans, CON, include all revolving, installment, or single-payment
loans to individuals (e.g., auto loans, student loans, personal lines of credit),
with the exception of credit card loans, which we exclude because they are rel-
percentage of banks that eased standards =(7–0)/52 =13.5%, up from just 1.9% in the previous
survey.
10 Rural banks typically have local market power; with greater rents at stake, their ability and
willingness to absorb risk overhang may differ markedly from those of urban banks. The extreme
localness, or “ruralness,” of these banks influences the manner in which they underwrite loans and
results in lower levels of credit risk (DeYoung et al. (2012)). Rural banks hold relatively low levels
of total loans, high levels of marketable securities, and high levels of equity compared to simi-
larly sized urban banks (DeYoung, Hunter, and Udell (2004)), consistent with a less sophisticated
approach to risk management.
2462 The Journal of Finance R
Tab le I
Descriptive Statistics
This table summarizes the main variables in our sample of small U.S. commercial banks be-
tween 1991:Q4 and 2010:Q4. Unbalanced panel includes 66,798 quarterly observations from 3,210
separate banks.
NMean Median Std Dev Min Max
Number of quarters per bank – 39.6 37.0 20.2 5 88
Bank assets (millions of 2010 $) 66,798 212.6 110.3 267.1 18.1 1975.2
Structural variables
NEW BUS 66,798 0.0027 0.0018 0.0111 −0.1448 0.1008
NEW RE 66,798 0.0089 0.0068 0.0169 −0.0602 0.1549
NEW CON 66,798 0.0012 0.0005 0.0079 −0.1684 0.1868
BUS 66,798 0.1183 0.1034 0.0635 0.0154 0.5420
RE 66,798 0.3387 0.3375 0.1093 0.0278 0.7216
CON 66,798 0.1019 0.0838 0.0631 0.0160 0.7037
RAR 66,798 0.8139 0.3115 1.5929 −0.0624 12.4435
Demand Shifters (DS)
Per Capita Income ($1,000s) 66,798 26.2944 25.2525 6.0509 14.1412 56.8060
%࢞Unemployed Persons 66,798 0.0044 −0.0074 0.1312 −0.5031 0.8057
Unemployment Rate 66,798 5.3150 5.1000 1.5380 1.7000 14.3000
Instruments
Personal Income Tax Rate 66,798 0.1466 0.1514 0.0152 0.1021 0.1822
Traffic Fatalities (per driver) 66,798 0.0003 0.0002 0.0001 0.0001 0.0006
Rental Vacancies (% vacant) 66,798 8.7295 8.4000 2.7065 2.7000 18.1000
Unexpected Snowfall (feet) 66,798 0.2367 0 0.8167 −6.2385 15.6
Variables capturing exogenous variation
COMM 66,798 0.1253 0 0.331 0 1
CRS 66,798 0.0522 0 0.2224 0 1
atively unimportant for small banks.11 While the latter two loan categories are
quite broad, this high level of aggregation is unavoidable given the structure
of the call reports.12 We normalize BUS,RE,andCON by end-of-quarter t–1
bank assets to control for bank size. Table Ishows that, on average, business
loans, real estate loans, and consumer loans account for about 12%, 34%, and
10% of total assets, respectively, for the banks in our sample.
As shown in Figure 1, the average asset share of real estate loans approx-
imately doubled during our sample period before declining somewhat during
the financial crisis. The asset share of consumer loans declined by more than
half during our sample period, while the asset share of business loans remained
relatively stable over time.
11 Small banks exited credit card lending with the development of loan production processes
(i.e., credit scoring and loan securitization) that exhibited huge scale economies. For the banks in
our data, credit card loans never exceeded 1% of bank assets on average during our sample period.
12 While the call reports disaggregate BUS,RE,andCON loan volumes into various subcate-
gories, the call reports do not report separately the interest revenue generated by the loans in those
subcategories. This prevents us from calculating the covariances σij for the loans in those subcat-
egories, which would in turn prevent us from predicting the signs of the estimated coefficients φ
and ρin equation (6) below.
Risk Overhang and Loan Portfolio Decisions 2463
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
Consumer C&I Real Estate
Figure 1. Cross-sectional means of three loan categories. The figure shows the quarterly
cross-sectional means for real estate loans, C&I (business) loans, and consumer loans as a percent-
age of assets for all U.S. commercial banks with assets less than $2 billion (2010 dollars) between
1987 and 2010.
The data displayed in Table II indicate that business loans tend to be rela-
tively risky. On average, business loans have the largest loan charge-off ratio
(0.65%), followed by consumer loans (0.51%) and real estate loans (0.10%). The
low average charge-off ratios for real estate loans is driven by residential real
estate loans; as shown in Figure 2, residential mortgage borrowers were the
least likely of all borrowers to default prior to the financial crisis. Table II also
indicates that business loans tend to be relatively illiquid. We use the very
limited data available in the call reports starting in 2001 to construct an av-
erage loan liquidity ratio as “outstanding principal balances of assets sold and
securitized by the reporting banks with serving retained or with recourse or
other seller-provided credit enhancements” plus “assets sold with recourse of
other seller-provided credit enhancements and not securitized by the reporting
bank” divided by loan balances. Business loans are the least liquid (0.07%),
followed by consumer loans (1.01%), and then real estate loans (1.29%). Note
that adding the nonspecialist lenders to the averages increases both credit risk
and loan liquidity: because they are less diversified, specialist banks operate
with concentrations of credit risk and must rely relatively more on loan sales
and securitization to manage their risk profiles.
New loan supply NLSis not directly observable in the call report data; we
only know the outstanding stock of loans Lat the end of each accounting
period. We calculate the quarter-to-quarter net lending change NLC =Lt,i −
2464 The Journal of Finance R
Tab le I I
Characteristics of Community Bank Sample (Nonspecialist Banks)
and Community Bank Population (Specialist and Nonspecialist
Banks)
This table describes the credit risk and loan liquidity for the sample banks (first column) and the
population of banks with less than $2 billion in assets (second column). The mean values for the
loan charge-off ratios reported in item 1 are computed using bank-quarter observations during
the 1991 to 2010 sample period for the three loan variables. BUS is business loans. RE is real
estate loans. CON is consumer loans. The mean aggregate values for the ratio of loans sold or
securitized reported in item 2 are the average of the quarterly sample aggregate ratios, and are
based on the sum of two call report items: “Outstanding principal balances of assets sold and
securitized by the reporting banks with serving retained or with recourse or other seller-provided
credit enhancements” plus “Assets sold with recourse of other seller-provided credit enhancements
and not securitized by the reporting bank.”
Community Bank Sample Community Bank Population
1. Loans charged off, % of total loans (means of bank-quarter observations):
BUS 0.65% 0.97%
RE 0.10% 0.18%
CON 0.51% 0.62%
2. Loans sold or securitized for which banks have existing recourse exposure, % of total loans
(means of quarterly aggregate ratios):
BUS 0.07% 0.19%
RE (excluding commercial real estate) 1.29% 3.25%
CON 1.01% 0.45%
Lt–1,i, and use this as our proxy for NLS. This measure is a close proxy for
the actual new loan supply when banks make illiquid loans and retain those
loans on their balance sheets, conditions that are characteristic of the small
banks in our sample.13 We also account for loan quality and bank size in our
calculations: net lending change equals end-of-quarter tloans minus end-of-
quarter t–1 loans, plus net loan charge-offs during quarter t(loans charged off
minus loans recovered), divided by bank assets at t–1. We calculate NLC for
each of the three loan categories—net new business loans (NEW BUS), net new
real estate loans (NEW RE), and net new consumer loans (NEW CON). During
our sample period, business loans, real estate loans, and consumer loans grew
at average quarterly rates of 27 bps, 89 bps, and 12 bps of assets, respectively.
Constructing an empirical analog for risk-adjusted returns (pt,1–μt,1 )/σ11 is
an imperfect exercise and requires that we make some choices. In our main
tests, we define risk-adjusted returns on business loans (RAR) as bank i’s
13 Let δt,i ∈(0,1) be the fraction of outstanding loans from period t–1 that are illiquid and let
τt,i ∈(0,1) be the fraction of liquid outstanding loans from period t−1 that the bank retains in
period t. This allows us to write Lt,i as NLSt,i +(δt,i +τt,i(1–δt,i ))Lt–1,i . Substituting this for Lt,i in
our definition of net lending change yields NLCt,i=NL
s
t,i−[(1 −τt,i)(1 −δt,i)]Lt−1,i.Hence,NLC
equals actual NLSminus the portion of liquid loan stocks that are actually sold. As either illiquidity
δor retention τapproaches one, NLC approaches NLS.
Risk Overhang and Loan Portfolio Decisions 2465
0.0%
1.0%
2.0%
3.0%
4.0%
5.0%
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
Residential Real Estate Commercial Real Estate Consumer Business
Figure 2. Nonperforming loan ratios. The figure shows the percentage of various types of
loans that were nonperforming for U.S. commercial banks in each quarter during the 1991 to 2010
period. Data are seasonally adjusted industry aggregates. NBER-defined recessions are indicated
by grey bands.
expected returns on business loans in quarter tdivided by the market-specific
variance of these returns over the preceding 20 quarters. In this ratio, the
numerator is the expected percent return (the bank’s interest and fee income
from business loans during period t, divided by its stock of accruing business
loans at the beginning of period t), multiplied by the expected performance
of business loans (the within-state percentage of performing loans averaged
over the preceding 20 quarters), minus the average deposit rate paid by the
bank (the bank’s interest expense on deposits during period tdivided by its
average volume of deposits during period t). The denominator is the preceding
20-quarter variance of the mean expected returns on business loans in the state
in which the bank’s main office is located. By measuring loan performance and
expected return variance at the market level, we capture the average risk in
the pool of small businesses from which the bank is drawing its loans.14
14 In addition to this RAR measure of “expected” loan returns, we also constructed two alter-
native RAR concepts: “realized” loan returns and “perfect foresight” loan returns. The Internet
Appendix, available in the online version of the article on the Journal of Finance website, contains
detailed definitions of these alternative RAR concepts as well as robustness tests that compare
the three alternative measures. There is no evidence that the alternative measures perform better
than our preferred “expected” loan returns measure.
2466 The Journal of Finance R
Descriptive statistics for all of the variables described here, as well as for all
other variables used in our regression tests, are reported in Table I. Additional
details are provided in the Internet Appendix.
IV. Estimation and Identification
The following parameterization transforms our theoretical loan supply equa-
tion (4) into an estimable business loan supply equation:
NLCt,1=−
i=2,3
ϕiNLCt,i−β1Lt−1,1−
i=2,3
ρiLt−1,i+χ1
pt,1−μt,1
σ11
,(5)
where subscript iindexes the three loan sectors in our data (business loans =
1; real estate loans =2; consumer loans =3). The coefficients φ,β,ρ,andχ
are parameters to be estimated. The φicoefficients absorb the loan covariance–
variance terms σij/σii . The coefficient β1absorbs the loan illiquidity term δ1.
The ρicoefficients will absorb both the loan covariance–variance terms σij/σii
and the loan illiquidity terms δj. The coefficient χ1absorbs the risk tolerance
term 1/G. Fully specified, our empirical business loan supply equation is
NEW BUSi,t=α+ϕRE ·NEW REi,t+ϕCON ·NEW CONi,t
+ρRE ·REi,t−1+βBUS ·BUSi,t−1+ρCON ·CONi,t−1
+χ·RARi,t+ω·DSi,t+φ·Bi+γ·Tt+εi,t,(6)
where iindexes banks, tindexes time in quarters, DS is a vector of business loan
demand shifters (described below), Bindicates bank fixed effects, Tindicates
time fixed effects, and the error term εis normally distributed around zero. We
refer to equation (6) as our baseline regression model.
Based on the discussion above, we can make the following predictions about
the estimated coefficients in equation (6):
rSame-sector overhang effect: Net new business lending is negatively related
to overhanging business loans (βBUS <0). This effect is stronger when
business lending is less liquid.
rCross-sector overhang effect: Net new business lending is positively related
to overhanging real estate and consumer loans (ρRE >0, ρCON >0) if loan
performance in these two sectors covaries negatively with business loans.
If these sectors covary positively with business loans, this effect is negative
(ρRE <0, ρCON <0). In either instance, the magnitude of these effects is
stronger when real estate lending or consumer lending is less liquid.
rCross-sector new lending effect: Net new business lending is positively
related to net new real estate and consumer loans (φRE >0, φCON >0) if
loan performance in these two sectors covaries negatively with business
loans. If these sectors covary positively with business loans, this effect is
negative (φRE <0, φCON <0). Regardless, these coefficients should carry
the same signs as their related cross-sector overhang effects.
Risk Overhang and Loan Portfolio Decisions 2467
Table III
Expected Profit Covariances
This table presents the number and percentage of banks for which the expected quarterly returns
from business loans (BUS) covary negatively with the expected quarterly returns from real estate
loans (RE) or consumer loans (CON) during various sample and sub-sample periods. ***, **, and *
indicate statistically different from zero at the 1%, 5%, and 10% level of significance, respectively.
Full sample, 1991:Q4 to 2010:Q4
Cov(BUS, RE)Cov(BUS, CON)
# of negative covariances 1,809 1,729
% of negative covariances 56.4%*** 53.9%***
Number of banks 3,210 3,210
Pre-crisis sub-sample, 1991:Q4 to 2007:Q3
Cov(BUS, RE)Cov(BUS, CON)
# of negative covariances 1,830 1,733
% of negative covariances 57.4%*** 54.3%***
Number of banks 3,189 3,189
Crisis sub-sample, 2007:Q4 to 2010:Q4
Cov(BUS, RE)Cov(BUS, CON)
# of negative covariances 247 244
% of negative covariances 49.8% 49.2%
Number of banks 496 496
rRisk-adjusted return effect: Net new business lending is positively related
to the expected returns on business loans (χ>0). This coefficient captures
the expected return elasticity of new business loan supply. This effect is
stronger for higher levels of lender risk tolerance.
Making a definitive prediction for the cross-sector overhang and cross-sector
new lending effects requires that we know whether business loan performance
covaries positively or negatively with the performance of real estate loans
and consumer loans in the data. We calculate the covariances σij separately
for each bank based on quarterly observations of their expected loan returns
(pt–μt) for each type of loan.15 As shown in Table III, these covariances
tend to be negative for the full sample—Cov(BUS,RE) is negative for 56.4%
of the banks and Cov(BUS,CON) is negative for 53.9% of the banks, both
statistically different from 50%—which indicates potential diversification gains
15 Although our theoretical loan supply function is expressed in terms of nonperforming loan
comovements, we use loan return comovements in our empirical implementation. Banks have
incentives to delay reporting reductions in loan quality because doing so requires them to make
accounting provisions that reduce stated net income. Even short delays in making these accounting
adjustments could be problematic given the quarterly frequency of our data. Moreover, our measure
of expected loan returns (pt–μt) does not rely on individual bank accounting discretion because
we use historical average data on state-level performing loans to measure μt.
2468 The Journal of Finance R
from combining business loans with other types of loans.16 Hence, our theory
unambiguously predicts positive signs on φRE and φCON. If a large portion of
outstanding real estate and consumer loans is illiquid, then our theory also
predicts positive signs on ρRE and ρCON (recall that the ρicoefficients absorb
both σij/σii and δj). However, given that residential mortgage loans were easily
securitizable during most of our sample period, the ρRE >0 prediction is a
relatively weak one. By contrast, these potential diversification effects are not
present during the crisis subperiod—the percentages of negative Cov(BUS,RE)
and Cov(BUS,CON) are not statistically different from 50%—so our theory
predicts values of zero for all of these coefficients during the crisis.17
A. Business Loan Demand
The demand shifters DS are included to absorb changes in small business
loan demand, and thus purge demand-side effects from our estimated supply-
side test coefficients. To construct DS, we collect (from the Bureau of Economic
Analysis and the Bureau of Labor Statistics) quarterly data on the unemploy-
ment rate, the percent change in unemployed persons, and personal income per
capita in each state. All microeconomics textbooks list household income and
employment as key drivers of demand for goods and services, which in turn
will drive activity at SMEs and hence their demand for loans. Unfortunately,
bank SME loan supply will be similarly related to household income and em-
ployment. We use regression analysis to separate the demand effects from the
supply effects. In the SLOOS, the Federal Reserve administers a battery of
questions each quarter to U.S. commercial banks in each of the 12 Federal Re-
serve Districts. Question 4b asks “Apart from normal seasonal variation, how
has demand for C&I loans changed over the past three months, from small
firms with annual sales of less than $50 million?” The surveyed loan officers
have five choices—substantially stronger, moderately stronger, about the same,
moderately weaker, or substantially weaker—and the Federal Reserve makes
public the net percentage of loan officers reporting stronger loan demand each
quarter. For each of the 50 states in our data, and for each of our three eco-
16 Given that the locally focused banks in our sample are not diversified across regional business
cycles, one might expect largely positive loan performance covariances. There are a number of
reasons for negative covariances in loan performance. Historically, households under stress have
tended to default on consumer loans (auto loans, credit cards) relatively early in a recession while
continuing to service their home mortgage loans (Andersson et al. (2013)). Small banks have local
geographic focus in business lending, but it is not unusual for them to make out-of-market real
estate loans. The financial health of the average local household will be more closely related to local
economic conditions, but the financial health of local businesses that export goods and services to
other geographic markets will be exposed to nonlocal economic conditions. The risk spreads pt
reflect local conditions for business loans, but follow economy-wide conditions for mortgage and
consumer loans.
17 The small number of sample banks operating during the crisis (only 496) is due to two
secular trends during our sample period: banking industry consolidation, which reduced the overall
number of community banks, and the diverging paths of real estate loan shares and consumer loan
shares (see Figure 1), which reduced the number of nonspecialist banks in our data.
Risk Overhang and Loan Portfolio Decisions 2469
nomic conditions variables, we regress the quarterly time series of seasonally
adjusted state-level economic conditions on the quarterly SLOOS measure of
net business loan demand change (150 regressions in all). We use the quar-
terly fitted values of these regressions as our demand shifters DS. These fitted
demand shifters, which contain both cross sectional and time series variation,
provide us with information on local economic conditions that is related directly
to small business loan demand.
B. Endogenous Right-Hand-Side Variables
Because banks make their lending decisions in each loan sector isimulta-
neously, NEW RE and NEW CON are endogenous in equation (6). If commu-
nity banks have market power in small business lending (Petersen and Rajan
(1995)), then the right-hand-side variable RAR in equation (6) is also potentially
endogenous. We use standard two-stage least squares instrumental variables
(2SLS-IV) estimation methods to address these endogeneity issues.
We select four instruments, all of which vary across states and over time:
unexpected (actual minus historical median) snowfall, the percentage of vacant
rental units, the average marginal tax rate (federal plus state) on personal in-
come, and traffic fatalities per licensed drivers. These instruments are clearly
exogenous to the banks in our data. These instruments can also safely be
excluded from the second-stage regression: NEW BUS should be largely un-
affected by weather conditions because it excludes loans for weather-sensitive
construction or agriculture projects, unrelated to rental vacancies because it
excludes loans that are secured by real estate, and relatively unrelated to per-
sonal tax rates because the large majority of banks in our sample are taxed
at the corporate level.18 Finally, there are plausible reasons to expect these
instruments to be correlated with the endogenous right-hand-side variables.
Extreme winter weather conditions can affect construction schedules and hence
are related to the supply of real estate loans; weather can also affect consumer
purchase behavior and hence the supply of consumer loans.19 Changes in rental
vacancies should be related to the number and size of residential housing loans,
while personal tax rates should be related to the number and size of consumer
loans. Traffic fatality data are related to a number of primary factors—for
example, commuting distances (and hence the real estate rent gradient and
automobile longevity), highway conditions (and hence investment in infras-
tructure), destroyed vehicles—that are correlated with variation in real estate
and consumer loan supplies.20 By similar reasoning, these four instruments
18 Only 13.7% of the banks in our sample were organized as Subchapter S corporations; these
banks are exempt from paying corporate income taxes, in exchange for making large earnings
distributions to their shareholders so that bank earnings are fully taxed at the household level.
19 Construction company contracts include clauses that extend deadlines if rainfall or snow
totals exceed amounts that will make it difficult to deliver their projects on time, especially in
winter.
20 Traffic fatalities are projected to be the fifth leading cause of death by 2030 (World Health
Organization (2009)). The economic damage done by traffic accidents and fatalities—for example,
2470 The Journal of Finance R
should also be correlated with bank profitability and risk (RAR). When we esti-
mate our models, we conduct diagnostic tests to verify that our instruments are
not correlated with the dependent variable while still being strongly correlated
with the explanatory variables being instrumented.
C. Exploiting Exogenous Variation
Recent studies of bank loan supply in Europe gain identification by exploit-
ing natural experiments, detailed loan-level credit registry databases, and
borrower-level survey data (e.g., Popov and Udell (2012), Puri, Rocholl, and
Steffen (2011), Jimen´
ez et al. (2012)). Our theory of bank loan supply, in which
the composition and characteristics of bank loan portfolios drive the results,
is more appropriately tested using bank-level data. We do not have access to
a convenient natural experiment. We instead gain identification by exploiting
two sources of exogenous variation: the financial crisis itself (which provides
time-series variation) and differences in the long-run ex ante lending strategies
of the banks in our data (which provides cross-sectional variation).
We define the financial crisis dummy variable CRS that is equal to one for all
quarterly observations from 2007:Q4 through 2010:Q4, and we interact CRS
with each of the main test variables BUS,RE,CON,andRAR in equation
(6). The CRS dummy is doubly useful: not only was the financial crisis an
exogenous shock, but it was also associated with conditions that should result
in strong risk overhang effects: loan write-downs that consumed internal bank
equity capital, stock price declines that made external equity capital more
expensive, and general economic uncertainty that reduced the liquidity of loan
secondary markets and securitizations. For example, by adding the following
terms to the right-hand side of (6), we can test not only whether the financial
crisis was associated with a reduction in SME loan supply, but also whether
our hypothesized same-sector loan overhang effect played a role in any such
reduction:
λ·CRSt+ξ1·BUSi,t−1+ξ2·BUSi,t−1·CRSt.(7)
The impact of the crisis on SME loan supply is given by λ+ξ2·BUSi,t−1,with
ξ2<0 indicating that loan overhang effects grew stronger during the crisis.
The approach in (7) is useful but provides only weak identification of the
hypothesis in question: the estimates of λand ξ2capture all of the changes
in the bank lending environment contemporaneous with the crisis, not just
the crisis-related changes associated with the hypothesized same-sector loan
overhang effect. We strengthen the specification by adding a second exogenous
interaction term that identifies banks with especially illiquid loan portfolios—
according to our theory, banks with more illiquid loan portfolios should exhibit
the destruction of vehicles, the loss of goods being shipped by trucks, the death of employees—has
been estimated as equivalent to 1% of national output in the typical country (Fouracre and Jacobs
(1976)). Replacing these lost resources is likely to have nontrivial effects on financial institution
profitability, via additional loans and insurance policies.
Risk Overhang and Loan Portfolio Decisions 2471
stronger risk overhang effects both before and during the crisis. We argue
that banks that concentrate in SME loans and/or commercial real estate loans
should be the best place to test this hypothesis, because these loans are espe-
cially illiquid.
The illiquidity of SME loans is a well-established fact (Berger and Frame
(2007)). Apart from some very small loans guaranteed by the U.S. Small Busi-
ness Administration (SBA loans) and some credit card–like revolving loans
issued to very small proprietorships, bank loans to SMEs are not easily se-
curitizable or otherwise sellable. Commercial real estate loans also tend to be
very illiquid: Levitin and Wachter (2012) report that approximately 80% of U.S.
commercial real estate debt in 2011 was held in portfolio rather than securi-
tized, and even higher percentages were held in portfolio during the 1990s and
early 2000s. Moreover, the findings of Black, Krainer, and Nichols (2014)sug-
gest that both SME loans and commercial real estate loans are consistent with
condition (b) in our broad definition of loan illiquidity: the authors conclude
that borrowers with difficult-to-assess projects (e.g., small businesses) tend to
be matched with lenders that have comparative advantages in loan renegotia-
tion, and commercial real estate loans with high default probabilities are more
likely to be renegotiated to extend their contractual maturities. Finally, given
the strong procyclicality of nonperforming business loans and commercial real
estate loans (see Figure 2), these loans are likely to become even more illiquid
during economic downturns (Furlong and Knight 2010).
We define commercial loans as BUS plus the portion of RE comprised by
commercial real estate loans, and we define retail loans as CON plus the re-
maining portion of RE. We further define a commercial focus dummy variable
COMMi,t that equals one at time tif a bank was in both the highest quartile of
commercial loans and the lowest quartile of retail loans in each of the past 10
quarters (t–11 through t–1). We choose this 10-quarter time threshold based on
the results of the Kaplan–Meier hazard estimation in Figure 3, which shows
that the probability of a bank switching away from a commercial-focus strategy
falls below 1% after it has engaged in this strategy for 10 consecutive quarters.
Thus, COMM captures exogenous cross-sectional variation in banks’ business
models and also reflects differences in the illiquidity of banks’ existing loan
stocks. About 13% of the observations in our sample have commercial focus.
For ease of exposition, we refer to the remaining 87% of the observations in
our sample as having retail focus. Table IV compares the average composition
of the loan portfolios of these two sets of banks.
We now specify a stronger set of tests by expanding (7) to include the following
terms:
λ·CRSt+η·COMMi,t+ξ1·BUSi,t−1+ξ2·BUSi,t−1·CRSt
+ξ3·BUSi,t−1·COMMi,t+ξ4·BUSi,t−1·CRSt·COMMi,t.(8)
This specification employs a difference-in-difference logic in which CRS
serves as a treatment variable and COMM serves as a control variable. The
2472 The Journal of Finance R
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
0.11
0.12
13579111315171921232527303234374245515365
Quarters
Figure 3. Kaplan–Meier estimates. This figures plots the hazard probability that a commercial-
focus bank (COMM =1) will abandon its strategy after having practiced it for a given number of
consecutive quarters (horizontal axis).
impact of the crisis on SME loan supply is now given by λ+ξ2·BUSi,t−1+
ξ4·BUSi,t−1·COMM
i,t,withξ4<0 indicating that loan overhang effects
grew stronger during the crisis for banks with especially illiquid portfolios
(i.e., commercial-focus banks). Applying this double-interaction identification
scheme to each of the main test variables—not just to BUS, but also to RE,
CON,andRAR—provides strong and internally consistent identification for
each of the main predictions of our theory.
As defined here, commercial-focus banks are highly likely to be relationship
lenders. Business loans made by small banks are the quintessential relation-
ship loan, and commercial real estate loans made by small banks tend to be
held in portfolio because they are too idiosyncratic or information-opaque to
be securitized (i.e., relationship loans). Because bank–borrower relationships
have value beyond the current loan contract, banks act to preserve the rela-
tionship by holding these loans on balance sheet rather than selling them, and
by renegotiating delinquent loans rather than simply writing them off (Minton,
Stulz, and Williamson (2009)).21 By choosing a long-run relationship lending
strategy, commercial-focus banks commit to making their loans even more illiq-
uid. A negative value for ξ3(the coefficient on BUS*COMM) is consistent with
the illiquidity of relationship loans during normal times.
21 Minton, Stulz, and Williamson (2009p. 11) provide three reasons why relationship lenders
are more likely to retain loans on the balance sheet: (1) “the borrower may not want the loan to be
sold since it would be harder to negotiate with a lender who has no experience with the borrower;
(2) lender may want to protect the relationship with the borrower; (3) relationship-based lending
can involve implicit commitments on both parties that become worthless if the loan is sold.”
Risk Overhang and Loan Portfolio Decisions 2473
Tab le I V
Loan Portfolio Composition
This table presents loan portfolio compositions, expressed as a percentage of total loans. The com-
mercial focus dummy (COMM) equals one if a bank’s commercial loan share is in the upper quartile
and its noncommercial loan share is in the bottom quartile of the quarterly sample distribution
over the previous 10 quarters.
(1) (2) (3)
Full Sample COMM =0COMM =1
Business loans (BUS)/total loans 21.5% 19.5% 32.8%
Consumer loans (CON)/total loans 18.3% 19.5% 11.7%
Real estate loans (RE)/total loans 60.2% 61.0% 55.5%
100.0% 100.0% 100.0%
Components of real estate loans:
Residential mortgage 29.7% 31.8% 14.6%
Home equity line of credit (HELOC) 2.4% 2.5% 1.9%
Construction & land development 6.0% 5.5% 9.4%
Nonfarm, nonresidental loans 20.2% 19.0% 28.5%
Number of observations 66,798 58,430 8,368
V. Re s u l t s
The results from our baseline model (6) are displayed in Table V.Column
6 shows our preferred specification; the other five columns show estimations
without full sets of instruments and/or without fitting the demand shifters.
Standard errors are clustered at the bank level in all six regressions, and in all
of the regressions that follow.
The coefficients on the demand shifters carry economically intuitive signs
throughout, but the coefficients on the fitted demand shifters (columns 4 to
6) are both larger and more precise than the coefficients on the raw demand
shifters (columns 1 to 3). The underidentification tests (where we seek to re-
ject the null) and overidentification tests (where we seek not to reject the
null) are strong throughout. Although the weak identification tests are at
best border line (just below the critical Fvalue in column 6), we note that
this critical threshold is cleared by a wide margin in all of the remaining
regression tests that follow.22 Instrumental variable estimation also results
in stronger and/or more reasonable coefficients for the endogenous variables.
The economically non intuitive negative coefficient on NEW RE disappears
(e.g., compare columns 4 and 5), an economically intuitive positive coefficient
emerges for NEW CON (again, compare columns 4 and 5), and the economi-
cally intuitive positive coefficient on RAR increases by an order of magnitude
22 We present additional diagnostics in the Internet Appendix. Each of the four instrumental
variables is statistically significant at least once in the first-stage regressions, with economically
sensible signs. When we include the four instruments as right-hand-side regressors in OLS versions
of our main model, none of the instruments carry statistically significant coefficients.
2474 The Journal of Finance R
Tab le V
Baseline Model
This table presents the parameter values for equation (6) estimated using the full 1991:Q4 to
2010:Q4 sample. The dependent variable is new business lending, NEW BUS. NEW RE and
NEW CON are new lending in the real estate and consumer loan sectors. RE, BUS, and CON
are the preexisting loan stocks in the real estate, business, and consumer loan sectors. RAR is
risk-adjusted returns. Per Capita Income,%࢞Unemployed Persons, and Unemployment Rate are
state-level personal income per capita, the percent change in unemployed persons, and the un-
employment rate in each state, respectively. Clustered standard errors appears in parentheses.
***, ** and * indicate statistically different from zero at the 1%, 5%, and 10% level of significance,
respectively.
(1) (2) (3) (4) (5) (6)
Model: Panel OLS IV-2SLS IV-2SLS Panel OLS IV-2SLS IV-2SLS
NEW RE −0.0517*** −0.0579 −0.2032 −0.0513*** −0.0617 −0.2351
(0.0027) (0.1606) (0.1640) (0.0027) (0.1526) (0.1675)
NEW CON −0.0026 0.8294* 0.2163 −0.0022 0.8247* 0.3949
(0.0057) (0.4863) (0.3705) (0.0057) (0.4211) (0.3542)
RE −0.0022** −0.0006 −0.0034 −0.0022** −0.0007 −0.0034
(0.0009) (0.0028) (0.0026) (0.0009) (0.0025) (0.0027)
BUS −0.0366*** −0.0379*** −0.0317*** −0.0365*** −0.0378*** −0.0306***
(0.0015) (0.0040) (0.0049) (0.0015) (0.0038) (0.0051)
CON 0.0047*** 0.0118** 0.0069* 0.0046*** 0.0116*** 0.0085**
(0.0014) (0.0049) (0.0037) (0.0014) (0.0045) (0.0038)
RAR 0.0002** 0.0002* 0.0019* 0.0002*** 0.0002** 0.0024**
(0.0001) (0.0001) (0.0026) (0.0001) (0.0001) (0.0011)
Per Capita Income 0.0002*** 0.0001 0.0002 0.0004*** 0.0002 0.0003*
(0.0001) (0.0001) (0.0001) (0.0001) (0.0002) (0.0002)
%࢞Unemployed
Persons
−0.0020*** −0.0021*** −0.0023*** −0.0034*** −0.0037*** −0.0041***
(0.0005) (0.0006) (0.0006) (0.0006) (0.0009) (0.0010)
Unemployment Rate −0.0001 −0.0000 −0.0000 −0.0008*** −0.0009*** −0.0007**
(0.0001) (0.0001) (0.0001) (0.0002) (0.0003) (0.0003)
Bank fixed effects Yes Yes Yes Yes Yes Yes
Quarter fixed effects Yes Yes Yes Yes Yes Yes
Clustered standard
errors (bank)
No Yes Yes No Yes Yes
Instruments for
NEW RE and
NEW CON
No Yes Yes No Yes Yes
Instruments for RAR No No Yes No No Yes
Demand shifters fitted No No No Yes Yes Yes
F-test for demand
shifters
10.34 10.25 14.59 21.96 34.38 38.15
Underidentification
test (p-value)
– 0.03 0.00 – 0.00 0.00
Overidentification test
(p-value)
– 0.10 0.34 – 0.17 0.55
Weak identification
test (Wald F=7.56
at 10% maximal IV
relative bias
– 5.06 5.83 – 7.28 7.40
Number of
observations
66,798 66,798 66,798 66,798 66,798 66,798
Number of banks 3,210 3,210 3,210 3,210 3,210 3,210
Risk Overhang and Loan Portfolio Decisions 2475
(compare columns 5 and 6). As the specification in column 6 is the strongest,
the remainder of our discussion focuses on those results.
As expected, the same-sector overhang effect (BUS) is negative, statistically
significant, and economically large. A 10% increase in overhanging business
loans is associated with a next-quarter decline in new business lending equal
to about 2.5% of a bank’s outstanding business loan balances.23 This substan-
tial one-quarter effect understates the eventual impact of loan overhang, as the
initial “shock” will result in additional (though diminishing) quarterly reduc-
tions in the quarters that follow. If we make the relatively reasonable assump-
tions that business loans are illiquid, are originated uniformly across quarters,
and have one-year maturities, then the cumulative reduction caused by a 10%
increase in overhanging business loans equals roughly 6.5% of outstanding
business loan balances.24
The estimated cross-sector overhang effect for consumer loans (CON)is
positive. Given the negative BUS-CON covariance result in Table III,this
result is consistent with the predictions of our theory model. A 10% quar-
terly increase in overhanging consumer loans is associated with an increase
in net new business lending of about 1.8% over the next four quarters. Al-
though not statistically significant in column 6, the cross-sector net lending
change effect (NEW CON) is also positive as predicted—an additional dollar
of new consumer lending is associated with a $0.39 increase in new business
lending.
The coefficients on RE and NEW RE are both statistically zero. Given the
negative BUS-RE covariance result in Table III, our model predicts positive
coefficients. However, those predictions are based on the assumption of perfect
loan illiquidity, and over half of the real estate loans held by our sample banks
were highly liquid residential mortgage and home equity loans (see Table IV).
We account for such a scenario in equation (4), where the loan overhang effects
in sector jdecline toward zero as a bank’s holdings in sector jloans become
more liquid (i.e., smaller values of δt–1,j ), regardless of the value of σij. Hence,
the statistically insignificant coefficients on RE and NEW RE are consistent
with our model, in which loan liquidity provides an option to sell that defuses
any RE overhang effects.
As expected, the risk-adjusted return variable RAR carries a positive and
statistically significant coefficient throughout. While we are not estimating
a formal loan supply function, this result is strongly consistent with a loan
supply relationship: when the expected returns from making business loans
increase, banks supply more net business loans. A one-standard-deviation in-
crease in RAR is associated with a next-quarter increase in net new busi-
ness lending equal to about 3.2% of a bank’s outstanding business loan
balances.25
23 The result is calculated as follows: –0.0306*0.10/0.1183 =0.0259.
24 The result is calculated as follows: –0.0306*(0.10 +0.075 +0.05 +0.025)/0.1183 =0.0647.
25 The result is calculated as follows: 0.0024*1.59/0.1183 =0.0323.
2476 The Journal of Finance R
A. Impact of the Financial Crisis
We now address our main questions: Did U.S. community banks reduce SME
loan supply during the financial crisis? Was the reduction in loan supply related
to risk overhang frictions, as our theory predicts? Were these effects stronger
at banks specializing in illiquid loans, again as our theory predicts?
We begin by augmenting the baseline model as in (7) to include the CRS
dummy variable. The results are displayed in Table VI and provide affirma-
tive answers to the first and second of the above questions. The derivative
࢚NEW BUS/࢚CRS indicates that net new business lending per dollar of assets
declined by 24 bps per quarter during the crisis on average; this is equivalent
to a 2.03% quarterly reduction in existing business loans at these banks.26 The
negative and significant coefficient on BUS*CRS indicates that the same-sector
loan overhang effects increased during the crisis by about 56% (0.0202/0.0359),
consistent with a substantial reduction in liquidity in SME loan markets during
the crisis. In contrast, we find no evidence that the cross-sector overhang effects
changed during the crisis, as neither NEW RE*CRS nor NEW CON*CRS carry
statistically significant coefficients. The results are somewhat more suggestive
of credit rationing, or at least much less return-elastic BUS loan supply, during
the crisis: the derivative ࢚NEW BUS/࢚RAR declines by half during the crisis,
from a statistically significant 0.0004 to a statistically insignificant 0.0002.
We further augment the baseline model as in (8) to include both the CRS and
COMM dummy variables. The results are displayed in Table VII.Ourmain
test comes from evaluating the derivative ࢚BUS NEW/࢚CRS separately for
commercial-focus and retail-focus banks. At retail-focus banks, net new busi-
ness lending per dollar of assets declined on average by 31 bps per quarter
during the crisis, equivalent to a 2.86% quarterly reduction in existing busi-
ness loans at these banks. In contrast, new business lending per dollar of
assets increased on average by 150 bps per quarter during the crisis for the
commercial-focus banks, equivalent to an 8.00% quarterly increase in existing
business loans at these banks. This result is stark: the typical community bank
(retail focused) effectively reduced its ongoing new supply of credit to SMEs
by about 10.86% (2.86% +8.00%) during the crisis relative to banks that were
strategically dedicated to making and holding illiquid SME loans (commercial
focused).
The volume of business loans outstanding at U.S. banks increased well into
the financial crisis—banks were not necessarily originating new loans, but for
a time firms were able to draw down their existing credit lines (Berrospide,
Meisenzahl, and Sullivan (2012)). According to the National Bureau of Eco-
nomic Research (NBER), the recession in the United States lasted from De-
cember 2007 to June 2009; however, total commercial and industrial (C&I)
loan balances at U.S. commercial banks did not peak until 2008:Q3, and small
business loan balances (defined as C&I loans with principle amounts less than
26 The former result is simply the derivative with respect to CRS. Dividing this derivative by
the sample mean for BUS/ASSETS yields the latter result.
Risk Overhang and Loan Portfolio Decisions 2477
Tab le V I
Risk Overhang Effects during the Pre-crisis and Crisis Periods
This table reports the test for risk overhang effects using exogenous variation in macroeconomic
regimes. The dependent variable is new business lending. NEW BUS,NEW RE and NEW CON are
new lending in the real estate and consumer loan sectors. RE, BUS, and CON are the preexisting
loan stocks in the real estate, business, and consumer loan sectors. RAR is risk-adjusted returns.
CRS is a dummy equal to one for bank-quarter observations between 2007:Q4 and 2010:Q4. The
baseline 2SLS-IV model of equation (6) is augmented to include interactions of CRS with the main
variables of interest. The model is estimated for the full sample (1991:Q4 to 2010:Q4). Clustered
standard errors appear in panrentheses. ***, **, and * indicate statistically different from zero at
the 1%, 5%, and 10% level of significance, respectively.
(1) (2) (3)
Estimated partial derivatives evaluated
for different values of CRS
࢚Y/࢚Xat CRS =0࢚Y/࢚Xat CRS =1
NEW RE −0.0515
(0.1364)
NEW CON 0.3233
(0.2373)
RE −0.0028 −0.0028 −0.0009
(0.0021) −0.0021 −0.0036
RE*CRS 0.0019
(0.0025)
BUS −0.0359*** −0.0359*** −0.0561***
(0.0034) −0.0034 −0.0069
BUS*CRS −0.0202***
(0.0063)
CON 0.0087*** 0.0087*** 0.0079
(0.0028) −0.0028 −0.0070
CON*CRS −0.0008
(0.0056)
RAR 0.0004** 0.0004** 0.0002
(0.0002) −0.0002 −0.0003
RAR*CRS −0.0003
(0.0002)
CRS −0.0010
(0.0016)
Bank fixed effects Yes
Quarter fixed effects Yes
Clustered standard errors (banks) Yes
Fitted demand shifters Yes
F-Test for demand shifters 9.20
Instruments for NEW RE,
NEW CON, RAR
Yes
Underidentification test (p-value) 0.00
Overidentification test (p-value) 0.29
Weak identification test (Wald
F=7.56 at 10%
maximal IV relative bias) 9.98
࢚NEW BUS/࢚CRS (at means of the
data)
−0.0024*
Number of observations 66,798
Number of banks 3,210
2478 The Journal of Finance R
Tab le V II
Risk Overhang Effects Using Exogenous Variation in Bank Business
Models and Macroeconomic Regimes
This table reports the test for risk overhang effects using exogenous variation in bank business
models and macroeconomic regimes. The dependent variable is new business lending, NEW BUS.
NEW RE and NEW CON are new lending in the real estate and consumer loan sectors. RE, BUS,
and CON are the preexisting loan stocks in the real estate, business, and consumer loan sectors.
RAR is risk-adjusted returns. COMM is a dummy equal to one for banks with commercial lending
focus, where COMM equals one if a bank’s commercial loans share is in the upper quartile and its
noncommercial loans share is in the bottom quartile of the quarterly sample distribution over the
previous 10 quarters. CRS is a dummy equal to one for bank-quarter observations between 2007:Q4
and 2010:Q4. The baseline 2SLS-IV model of equation (6) is augmented to include interactions of
COMM and CRS with the main variables of interest. The model is estimated for the full sample
(1991:Q4 to 2010:Q4). Clustered standard errors appear in parentheses. ***, **, and * indicate
statistically different from zero at the 1%, 5%, and 10% level of significance, respectively.
(1) (2) (3) (4) (5)
Estimated partial derivatives evaluated for different
values of the CRS and COMM dummy variables:
CRS =0CRS =1CRS =0CRS =1
COMM =0COMM =0COMM =1COMM =1
NEW RE −0.0426
(0.1362)
NEW CON 0.3169
(0.2346)
RE −0.0021 −0.0021 −0.0012 −0.0084** −0.0116
(0.0020) (0.0020) (0.0037) (0.0036) (0.0172)
RE*CRS 0.0009
(0.0025)
RE*COMM −0.0062**
(0.0029)
RE*CRS*COMM −0.0041
(0.0161)
BUS −0.0333*** −0.0333*** −0.0539*** −0.0474*** −0.0671***
(0.0036) (0.0036) (0.0071) (0.0054) (0.0137)
BUS*CRS −0.0206***
(0.0065)
BUS*COMM −0.0141***
(0.0053)
BUS*CRS*COMM 0.0008
(0.0151)
CON 0.0078*** 0.0078*** 0.0080 0.0242** 0.2484***
(0.0028) (0.0028) (0.0068) (0.0102) (0.0864)
CON*CRS 0.0001
(0.0056)
CON*COMM 0.0164
(0.0105)
CON*CRS*COMM 0.2241**
(0.0876)
RAR 0.00043** 0.00043** 0.00028 0.00064*** −0.0036
(0.00021) (0.00021) (0.00031) (0.00021) (0.00240)
RAR*CRS −0.0002
(0.0002)
RAR*COMM 0.0002
(Continued)
Risk Overhang and Loan Portfolio Decisions 2479
Tab le V II—Continued
(1) (2) (3) (4) (5)
Estimated partial derivatives evaluated for different
values of the CRS and COMM dummy variables:
CRS =0CRS =1CRS =0CRS =1
COMM =0COMM =0COMM =1COMM =1
(0.0001)
RAR*CRS*COMM −0.0041***
(0.0014)
COMM 0.0024*
(0.0013)
CRS −0.0009
(0.0016)
Bank fixed effects Yes
Quarter fixed effects Yes
Clustered standard errors
(banks)
Yes
Fitted demand shifters Yes
F-test for demand shifters 8.62
Instruments for NEW RE,
NEW CON and RAR
Yes
Underidentification test
(p-value)
0.00
Overidentification test
(p-value)
0.24
Weak identification test
(Wald F=7.56 at 10%
maximal IV relative bias)
10.01
࢚NEW BUS/࢚CRS(COMM
=0)
−0.0031**
࢚NEW BUS/࢚CRS(COMM
=1)
0.0150***
Bank-quarter observations 66,798
Number of banks 3,210
$1 million) did not peak until 2008:Q4. Both of these peaks were followed by
sharp and long-lasting declines in outstanding loans.27 The data displayed in
Table VIII come from three alternative versions of the Table VII model in which
the CRS dummy is replaced with a dummy equal to one in either the first year
of the recession (2007:Q4 through 2008:Q4), the second year of the recession
(2009), or the third year of the recession (2010).
27 Total C&I loans at U.S. banks peaked at $1.503 trillion in 2008:3Q, bottomed out at about
$1.165 trillion in 2010:Q2-Q3, and then began a steady quarterly increase. Small business loans
peaked at $0.336 at the end of 2008, but did not hit bottom until 2011, and since then have
fluctuated between $0.278 and $0.285 trillion with no clear increasing trend (as of the date of this
draft). The figures are based on data from various editions of the FDIC Quarterly Banking Profile.
See http://www2.fdic.gov/qbp/index.asp.
2480 The Journal of Finance R
Table VIII
The Impact of the Financial Crisis (CRS) on Net New Business
Lending
This table presents results of an augmented version of the model in Table VII where the single
CRS dummy is replaced by three single-year dummy variables. Reported are the values of the
derivative ࢚BUS NEW/࢚CRS evaluated in each of the three years of the crisis (columns 1, 2, and
3), for either value of the strategic lending focus variable COMM (rows 1 and 2), and otherwise at
the means of the data. COMM equals one if a bank’s commercial loans share is in the upper quartile
and its noncommercial loan share is in the bottom quartile of the quarterly sample distribution
over the previous 10 quarters. ***, **, and * indicate statistically different from zero at the 1%,
5%, and 10% level of significance, respectively.
(1) (2) (3)
࢚NEW BUS/࢚CRS
1st Crisis Year
(2007:Q4-
2008:Q4)
2nd Crisis Year
(2009:Q1-
2009:Q4)
3rd Crisis Year
(2010:Q1-
2010:Q4)
Evaluated at COMM =0 (reflects net new
business lending at 58,430 quarterly
observations of banks with retail lending
focus)
−0.0036*** −0.0049*** −0.0035*
Evaluated at COMM =1 (reflects net new
business lending at 8,368 quarterly
observations of banks with commercial
lending focus)
0.0211*** 0.0019 0.0119
The results of these three models indicate that new business loan supply
from retail-focus banks declined during all three years of the recession, while
new business loan supply from commercial-focus banks increased during the
first year of the recession and held steady during the second and third years
of the recession. For commercial-focus banks, these results imply that the
value of preserving bank–borrower relationships offset any crisis-induced re-
duction in loan liquidity or scarcity in equity capital. This result suggests
that relationship-based business lending strategies can partially mitigate pro-
cyclical reductions in SME loan supply, complementing the findings of Garcia-
Appendini and Montoriol-Garriga (2013) that suppliers provide increased trade
credit to their clients during recessions.
For pre-crisis banks (CRS =0) with retail-focus banking strategies (COMM =
0)—that is, banks with relatively low loan illiquidity during times of relatively
plentiful and inexpensive equity capital—the results in Table VII are similar in
sign, size, and statistical significance to the baseline results in Table V: negative
same-sector overhang effects for business loans, positive cross-sector overhang
and new lending effects for consumer loans, zero cross-sector effects for real
estate loans, positive risk tolerance effects, and positive risk-adjusted return
effects. Note, however, that the same-sector overhang effect (BUS) increases
both for commercial-focus banks (coefficient on BUS*COMM) and during the
financial crisis (coefficient on BUS*CRS). Indeed, for the most illiquid banks
during the most illiquid time period, the same-sector overhang effect doubles
its baseline magnitude (compare columns 2 and 5). The cross-sector overhang
Risk Overhang and Loan Portfolio Decisions 2481
effect for consumer lending (CON) is also larger for commercial-focus banks
both before and during the crisis, but the cross-sector overhang effect for real
estate lending (RE) remains near zero in all four cases.28
The risk-adjusted return effects (RAR) are consistent with credit rationing
during the financial crisis. Prior to the crisis we find the expected positive
effect of RAR on new business loan supply for both retail- and commercial-
focus banks (columns 2 and 4), but this effect disappears for both sets of banks
during the crisis years (columns 3 and 5). This result is consistent with a
quantity rationing explanation, rather than a price rationing explanation, for
the reduction in small business loan supply during the crisis.
B. Cross-Study Comparisons
It is instructive to compare the economic magnitudes of our results to those
found in other studies. Because our empirical approach differs from those
used in other studies of SME loan supply during the financial crisis, direct
comparisons are not always possible. We restrict our comparisons to studies
that estimate loan supply regressions using a right-hand-side financial crisis
variable.
Jimen´
ez et al. (2012) estimate a model of SME loan application acceptance
rates in Spain during the crisis. Using their regression estimates, it is straight-
forward to calculate that the acceptance rate attributable to supply-side con-
ditions declined by 554 bps during the crisis. This change represents about
a 14% reduction in the average acceptance rate.29 Puri, Rocholl, and Steffen
(2011) estimate a model of consumer loan application acceptance rates at Ger-
man savings and loans during the crisis. At lenders that were exogenously
exposed to the financial crisis, acceptance rates for mortgage applications de-
clined by about 1,150 bps, compared to an increase in acceptance rates of
70 bps at comparable lenders not exposed to the crisis. The relative difference
represents a 12.5% reduction in the average acceptance rate.30
We find a similar average reduction in SME loan supply of 10.9% during the
crisis (relative to the benchmark commercial-focus banks from which we gain
identification). Our estimate may be smaller because the dependent variable in
our bank-level study measures lending from newly approved loan contracts plus
new drawdowns on lending lines approved in the past. Increases in the latter
(drawdowns of existing credit lines) partially offset reductions in the former
(new originations) during the crisis, resulting in relatively smaller estimated
crisis-induced reductions in SME lending.
28 The outlying result here is the coefficient on RE*COMM, which is relatively small but statis-
tically negative. This likely arises because real estate lending increases at COMM=1 banks will
be heavily weighted toward commercial real estate loans, which are relatively illiquid loans that
perform similarly to BUS loans across the business cycle.
29 We performed these calculations using the linear probability estimates in Jimen´
ez et al. (2012,
Table III) along with the descriptive statistics from the same study (Table I).
30 The reduction in the average consumer loan application acceptance rate was smaller, at about
7.5%.
2482 The Journal of Finance R
C. Additional Tests
We end our analysis with a series of robustness tests. Each of these tests
is based on the model in Table VII and either augments that specification or
reestimates it using a data subsample. We provide only a brief description of
the robustness tests here; complete results and more detailed discussion are
available in the Internet Appendix.
First, we test whether financially distressed banks may be influencing our
results. We reestimate the model after dropping all quarterly observations of
banks seized and resolved out of existence (via forced mergers, asset liqui-
dation, or depositor payouts) by the FDIC during our 20-year sample period.
The results are qualitatively unchanged and quantitatively very similar to the
full-sample results. Second, we test whether our findings vary by bank size.
We replace the COMM dummy with a LARGE dummy equal to one for banks
above the median value of bank assets during each quarter. The coefficient on
LARGE, the coefficients on all 10 of the variables with which LARGE is inter-
acted, and the derivative ࢚BUS NEW/࢚LARGE are all statistically zero in this
specification. Third, we test whether a broader geographic footprint, and hence
potential portfolio diversification effects, reduced banks’ sensitivities to mar-
ket imperfections. We replace the COMM dummy with a GEO dummy equal to
one for banks whose deposits were well-diversified across counties (i.e., in the
lower quartile of banks based on a Herfindahl index of geographic deposit con-
centration). As with bank size, none of the coefficients related to GEO,northe
derivative ࢚BUS NEW/࢚GEO, is statistically significant in this specification.
Hence, our main findings are not driven or influenced by variation in financial
stress, asset size, or geographic dispersion across the community banks in our
sample.
Fourth, we investigate whether bank business lending behavior during the
financial crisis is indicative of bank behavior during more mild recessions. We
reestimate the model using a pre-crisis subsample (1991:Q4 through 2007:Q3)
and replaced the CRS dummy with a REC2001 dummy equal to one for all quar-
terly observations from 2000:Q2 through 2003:Q2. We chose these dates based
on banks’ responses to the “demand for small business lending” question in the
quarterly SLOOS surveys surrounding the 2001 recession. The baseline results
for same-sector loan overhang, cross-sector loan overhang, and risk-adjusted
loan return are similar in sign, magnitude, and statistical significance to those
in Table VII. However, ࢚NEW BUS/࢚REC2001 is statistically zero and only
one of the coefficients on the REC2001 interaction terms is statistically signif-
icant. We conclude that the mild 2001 recession did not interrupt community
banks’ small business loan supply, nor did it result in stronger loan overhang
or risk-adjusted loan return effects.
In our last robustness test, we test whether our estimates of bank lending be-
havior vary with equity capital levels. It is well known that banks with very low
capital levels may engage in moral hazard via risk-shifting, possibly by over-
aggressive lending in the presence of underpriced deposit insurance (e.g., Mer-
ton (1977), Marcus (1984)). By contrast, banks with higher capital levels may
Risk Overhang and Loan Portfolio Decisions 2483
become more conservative lenders when their capital levels fall, as in Besanko
and Kanatas (1996), Thakor (1996), Holmstrom and Tirole (1997), and Diamond
and Rajan (2000). We replace the COMM dummy with a LOWEQ dummy equal
to one for banks with EQ ࣘ8%.31 The derivative ࢚NEW BUS/࢚CRS indicates
similar crisis-induced reductions in new business lending for both low-equity
and high-equity banks. We find evidence consistent with credit rationing for
both sets of banks: during the crisis, business lending became insensitive to
RAR for both high-equity and low-equity banks. Notably, the same-sector loan
overhang effect grew weaker for low-equity banks during the crisis. While this
result is consistent with risk-seeking behavior at poorly capitalized banks, one
cannot draw a strong conclusion without performing a closer investigation.
VI. Conclusion
Small businesses are especially reliant on bank finance. But during reces-
sions, credit can become less available to small firms if bank lenders—who
face declining loan quality, potential or actual reductions in the equity capital
necessary to back new loans, and illiquid asset markets that make it difficult
to raise funds via loan sales—take risk-averse actions to conserve equity capi-
tal. Such behavior by banks can exacerbate economic downturns by restricting
credit to job-creating small businesses. We estimate a structural econometric
model of new business lending by small U.S. banks between 1991 and 2010,
paying special attention to the impact of the global financial crisis on small
business credit supply over the 2007 to 2010 period. The empirical loan sup-
ply equation is derived from a model of portfolio lending in which lenders are
risk averse, originate and hold illiquid loans, and have costly external cap-
ital (Froot, Scharfstein, and Stein (1993), Froot and Stein (1998), Gron and
Winton (2001)), conditions that are especially descriptive of community bank
lenders. The model predicts that a bank’s small business lending decisions
will be constrained by the risk-adjusted returns on small business lending, the
composition of the preexisting (overhanging) loans on its balance sheet, the
covariances of the preexisting loans with small business lending opportunities,
and the bank’s own tolerance for taking risk. Thus, our study not only extends
the literature on small business credit supply during the global financial cri-
sis (e.g., Popov and Udell (2012), Jimenez et al. (2012), Cotugno, Monferra
and Sampagnaro (2012)) to include the U.S. experience, but also provides an
important test of financial intermediation theory.
We have two primary results. The first result is consistent with the prior
studies of small business lending in Europe: on average, U.S. community banks
reduced their credit supply to SMEs by a nontrivial amount during the financial
crisis. This suggests important similarities in U.S. and European SME lending
31 The median sample value of EQ is 0.0864, so LOWEQ equals one for slightly less than half of
the banks in the sample. Using lower threshold values (e.g., 7%) does not meaningfully change our
results. Banks with EQ below the median are bunched tightly between 6% and 8%; most banks
whose equity ratios dip below this range quickly exit the industry via failure or acquisition.
2484 The Journal of Finance R
markets, despite substantial institutional and regulatory differences across
these two sets of markets. The second result is new and provides an important
contrast: we find that new loan supply increased during the financial crisis
at a small segment of U.S. community banks (about 13% of our sample) that
strategically precommitted to making illiquid loans to commercial entities.
This result documents a central characteristic of small business lending that
has both micro and macroeconomic implications: in a true relationship lending
context, credit is made available when it is most needed.
Our empirical results offer strong support for the predictions of our theory.
Overall, the results indicate that business lending by small U.S. banks is con-
sistent with portfolio management practices whereby current credit is allocated
efficiently and scarce capital is conserved for future profitable lending opportu-
nities. Some borrowers will face tighter credit supply during short-run periods
of heightened bank risk aversion, when lender balance sheets exhibit unusu-
ally high loan illiquidity and/or when banks experience internal or external
pressure to increase their equity capital. But in the long run, these risk-averse
lenders are more likely to be around to provide funding and other financial
services, thus making long-run bank–borrower relationships possible. During
the pre-crisis period (1991 through 2007:Q3), we find that small U.S. banks
allocated additional capital to business loans when the expected returns to
making those loans was high and when the expected returns on business loans
covaried negatively with the preexisting illiquid (overhanging) loans in their
portfolios. During the financial crisis (2007:Q4 through 2010), new business
lending became more sensitive to overhanging loans but became less sensitive
to expected loan returns, implying that banks became less tolerant of risk dur-
ing the downturn as overhanging loans became less liquid and external equity
capital became more expensive. We also find evidence of lending inefficiencies
during the crisis, when new business lending became statistically insensitive to
expected returns on business loans, a result that is indicative of credit rationing
behavior.
It is worth thinking about how the small and relatively less sophisticated
banks in our data have been able to manage loan portfolio risk with the de-
gree of efficiency suggested by our estimates. These banks probably do not
make calculations on cross-sector loan performance covariances, but they may
approximate modern portfolio theory (dating to Markowitz (1952)) by using
rules of thumb or crude risk management tools. One such tool is loan concen-
tration limits, which, when binding, mimic modern portfolio management by
preventing banks from making new loans to small businesses in industries
or geographic markets that are already well represented in their existing loan
portfolios (i.e., new business loans that will covary positively with existing busi-
ness loans). Of course, the secular build-up of real estate loans in banks’ loan
portfolios during the 1990s and early 2000s (see Figure 1) indicates that the
typical small bank did not tightly apply concentration limits at the loan-sector
level, often to its detriment.
The procyclical loan overhang and procyclical credit rationing results that we
find here suggest that bank lending behaviors may exacerbate macroeconomic
Risk Overhang and Loan Portfolio Decisions 2485
cycles. In sum, our findings are consistent with the following economic tableaus.
As bank lending becomes more profitable due to an economic or sector-specific
expansion, banks’ equity capital, lending capacity, and tolerance for risk will
all increase. The resulting increase in loan supply will be further enhanced by
the relatively liquid nature of loans that are performing well. As the expansion
continues, at some point banks will need to compete for new business by provid-
ing loans to riskier borrowers and/or by providing loans at lower interest rates.
When the expansion inevitably ends, defaulting loans will reduce bank capital,
lending capacity, and risk tolerance; increased credit risk and loan illiquidity
will bolster already existing loan overhang effects; and the resulting reduction
in loan supply will exacerbate the macroeconomic downturn. These effects will
be moderated if banks hold precautionary balances of equity capital and liquid
assets, if banks hold significant portions of their loans in uncorrelated sectors
or in sectors that are robust to economic downturns, or if (as highlighted in our
results) banks are strategically committed to relationship-based small busi-
ness lending. As such, our findings suggest that the procyclical capital buffers
and liquidity coverage requirements included in Basel III will result in social
welfare gains.
With regard to the banks in our data set, a caveat is in order. We have focused
on the behavior of small and relatively diversified banks; large banks or more
specialized banks may behave differently. For example, large banks may be able
to use alternative risk management techniques such as credit derivatives to
reduce overhang effects. Similarly, specialized banks’ loan performance may be
better than that of diversified banks due to the expertise derived from greater
lending focus, which might lead to improved risk-bearing ability in downturns.
Alternatively, the absence of diversification effects at specialized banks may
make them behave in a more procyclical way, exacerbating the effects we find
here.
Initial submission: August 3, 2012; Final version received: June 25, 2015
Editor: Michael Roberts
REFERENCES
Amel, Dean, and Timothy Hannan, 1999, Establishing banking market definitions through esti-
mation of residual deposit supply equations, Journal of Banking and Finance 23, 1667–1690.
Andersson, Fredrik, Souphala Chomsisengphet, Dennis Glennon, and Feng Li, 2013, The changing
pecking order of consumer defaults, Journal of Money, Credit and Banking 45, 251–275.
Berger, Allen, Robert DeYoung, Mark Flannery, David Lee, and ¨
Ozde ¨
Oztekin, 2008, How do large
banking companies manage their capital? Journal of Financial Services Research 34, 123–149.
Berger, Allen, and W. Scott Frame, 2007, Small business credit scoring and credit availability,
Journal of Small Business Management 45, 5–22.
Berger, Allen, Iftekhar Hasan, and Leora Klapper, 2004, Further evidence on the link between
finance and growth: An international analysis of community banking and economic perfor-
mance, Journal of Financial Services Research 25, 169–202.
Berger, Allen, Nathan Miller, Mitchell Petersen, Raghuram Rajan, and Jeremy Stein, 2005, Does
function follow organizational form? Evidence from the lending practices of large and small
banks, Journal of Financial Economics 76, 237–269.
2486 The Journal of Finance R
Berger, Allen, Anthony Saunders, Joseph Scalise, and Gregory Udell, 1997, The effects of bank
mergers and acquisitions on small business lending, Working paper, Wharton School of Busi-
ness.
Berrospide, Jose, Ralf Meisenzahl, and Briana Sullivan, 2012, Credit line use and availability in
the financial crisis: The importance of hedging, Federal Reserve Board of Governors, Finance
and Economics Discussion Series 2012–2027.
Besanko, David, and George Kanatas, 1996, The regulation of bank capital: Do capital standards
promote bank safety? Journal of Financial Intermediation 5, 160–183.
Billett, Matthew T., Jon A. Garfinkel, and Edward S. O’Neal, 1998, The cost of market versus
regulatory discipline in banking, Journal of Financial Economics 48, 333–358.
Black, Lamont, John Krainer, and Joseph Nichols, 2014, From origination to renegotiation: A
comparison of portfolio and securitized commercial real estate loans, Working paper, DePaul
University.
Chodorow-Reich, Gabriel, 2014, The employment effects of credit market disruptions: Firm-level
evidence from the 2008–09 financial crisis, Quarterly Journal of Economics 129, 1–59.
Cotugno, Matteo, Stefano Monferra, and Gabriele Sampagnaro, 2012, Relationship lending, hier-
archical distance and credit tightening: Evidence from the 2007–09 financial crisis, Working
paper, Bocconi School of Management.
DeMarzo, Peter, and Darrell Duffie, 1999, A liquidity-based model of security design, Econometrica
67, 65–99.
DeYoung, Robert, Dennis Glennon, Peter Nigro, and Kenneth Spong, 2012, Ruralness and the
quality of small business credit, CBE Research Paper 2012-1, University of Kansas.
DeYoung, Robert, William C. Hunter, and Gregory F. Udell, 2004, The past, present, and probable
future for community banks, Journal of Financial Services Research 25, 85–133.
Diamond, Douglas, and Raghuram Rajan, 2000, A theory of bank capital, Journal of Finance 55,
2431–2465.
Fouracre, P.R., and G.D. Jacobs, 1976, Comparative accident costs in developing countries, TRRL
Supplementary Report 206, Transportation Research Laboratory, Crowthome, United King-
dom.
Froot, Kenneth, 2007, Risk management, capital budgeting, and capital structure policy for insur-
ers and reinsurers, Journal of Risk and Insurance 74, 273–299.
Froot, Kenneth, and Paul O’Connell, 1999, The pricing of U.S. catastrophe reinsurance, in Kenneth
Froot ed.: The Financing of Catastrophe Risk (University of Chicago Press, Chicago).
Froot, Kenneth, David Scharfstein, and Jeremy Stein, 1993, Risk management: Coordinating
corporate investment and financing policies, Journal of Finance 48, 1629–1658.
Froot, Kenneth, and Jeremy Stein, 1998, Risk management, capital budgeting, and capital struc-
ture policy for financial institutions: An integrated approach, Journal of Financial Economics
47, 55–82.
Furlong, Fred, and Zena Knight, 2010, Loss provisions and bank charge-offs in the financial crisis:
Lesson learned, Federal Reserve Bank of San Francisco Economic Letter 2010–16.
Garcia-Appendini, Emilia, and Judit Montoriol-Garriga, 2013, Firms as liquidity providers: Evi-
dence from the 2007–2008 financial crisis, Journal of Financial Economics 109, 272–291.
Gron, Anne, and Andrew Winton, 2001, Risk overhang and market behavior, Journal of Business
74, 591–612.
Holmstrom, Bengt, and Jean Tirole, 1997, Financial intermediation, loanable funds, and the real
sector, Quarterly Journal of Economics 112, 663–691.
Ivashina, Victoria, and David Scharfstein, 2010a, Loan syndication and credit cycles, American
Economic Review, Papers and Proceedings 100, 1–8.
Ivashina, Victoria, and David Scharfstein, 2010b, Banking lending during the financial crisis of
2008, Journal of Financial Economics 97, 319–338.
Jayaratne, Jith, and Donald P. Morgan, 2000, Capital market frictions and deposit constraints at
banks, Journal of Money, Credit and Banking 32, 74–92.
Jimen´
ez, Gabriel, Steven Ogena, Jos´
e-Luis Peydr´
o, and Jes ´
us Saurina, 2012, Credit supply versus
demand: Bank and firm balance-sheet channels in good and bad times, European Banking
Center Discussion Paper 2012-003, Tilburg University.
Risk Overhang and Loan Portfolio Decisions 2487
Levitin, Adam J., and Susan M. Wachter, 2012, The commercial real estate bubble, Business,
Economics and Regulatory Policy Working Paper Series No. 12-005, Georgetown University.
Liberti, Jose, and Jason Sturgess, 2012, The anatomy of a credit supply shock: Evidence from an
internal credit market, Working paper, DePaul University and Tilburg University.
Marcus, Alan, 1984, Deregulation and bank financial policy, Journal of Banking and Finance 8,
557–565.
Markowitz, Harry M., 1952, Portfolio selection, Journal of Finance 7, 77–91.
Merton, Robert C., 1977, An analytic derivation of the cost of deposit insurance loan guarantees,
Journal of Banking and Finance 1, 3–11.
Minton, Bernadette A., Ren´
e Stulz, and Rohan Williamson, 2009, How much do banks use credit
derivatives to hedge loans? Journal of Financial Services Research 35, 1–31.
Montorial-Garriga, Judit, and Christina Wang, 2012, The great recession and bank lending to
small businesses, Working paper 11–16, Federal Reserve Bank of Boston.
Myers, Stewart C., and Nicholas S. Majluf, 1984, Corporate financing and investment decisions
when firms have information that investors do not have, Journal of Financial Economics 13,
187–221.
Petersen, Mitchell, and Raghuram Rajan, 1994, The benefits of lending relationships: Evidence
from small business data, Journal of Finance 49, 3–37.
Petersen, Mitchell, and Raghuram Rajan, 1995, The effect of credit market competition on lending
relationships, Quarterly Journal of Economics 110, 407–443.
Popov, Alexander, and Gregory F. Udell, 2012, Cross-border banking and the international trans-
mission of financial distress during the crisis of 2007–2008, Journal of International Economics
87, 147–161.
Puri, Manju, J¨
org Rocholl, and Sascha Steffen, 2011, Global retail lending in the aftermath of the
U.S. financial crisis: Distinguishing between supply and demand factors, Journal of Financial
Economics 100, 556–578.
Rajan, Raghuram, 1992, Insiders and outsiders: The choice between informed and arm’s-length
debt, Journal of Finance 47, 1367–1400.
Spong, Kenneth, and Richard J. Sullivan, 2007, Manager wealth concentration, ownership struc-
ture, and risk in commercial banks, Journal of Financial Intermediation 16, 229–248.
Stein, Jeremy C., 2002, Relevance of soft information in credit rating: The case of a social bank
financing small businesses, American Economic Review 71, 393–410.
Thakor, Anjan, 1996, Capital requirements, monetary policy, and aggregate bank lending: Theory
and empirical evidence, Journal of Finance 51, 279–324.
U.S. Small Business Administration, 2014. Frequently asked questions, Office of Advocacy and
Small Business Data, Washington, DC.
World Health Organization, 2009, Global status report on road safety: Time for action, Geneva,
Switzerland.
Supporting Information
Additional Supporting Information may be found in the online version of this
article at the publisher’s website:
Appendix S1: Internet Appendix.
2488 The Journal of Finance R
