Loan Guarantees in a Crisis: An Antidote to a Credit Crunch
Summary
Federal Reserve Bank of Kansas City working paper RWP 21-03 by W. Blake Marsh and Padma Sharma, June 2021, updated April 2022, titled Loan Guarantees in a Crisis: An Antidote to a Credit Crunch, which says it supersedes the earlier version titled Government Loan Guarantees during a Crisis: The Effect of the PPP on Bank Lending and Profitability. The abstract says the authors study the Paycheck Protection Program and find that it averted a historic credit crunch at a time when banks were unlikely to meet firm credit needs by risking their own capital. It says risk-aversion, rather than profitability, motivated bank participation, and that the program boosted loan growth among participants but attenuated profitability. The paper states that the views are the authors' own.
Summary drafted by a model from the document's text below and checked by script against that text before publication. It is a navigation aid, not a reading of what the document proves. Where AI is used
- Issuer
- Source documents
- Document type
- Government Loan Guarantees During A Crisis The Effect Of The Ppp On Bank Lending And Profi
- Case
- Government Loan Guarantees During A Crisis The Effect Of The Ppp On Bank Lending And Profitability
Cite this document
“Loan Guarantees in a Crisis: An Antidote to a Credit Crunch.”
Original at doi.org. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/government-loan-guarantees-during-a-crisis-the-effect-of-the-ppp-on-bank-lending-and-profitability/
, SHA-256 4aa7a4834c81e0368659f861c09752d74c362f68389ee33064c712aea0d26e44.Full text
ISSN 1936-5330
Loan Guarantees in a Crisis:
An Antidote to a Credit
Crunch
W. Blake Marsh and Padma Sharma
June 2021; updated April 2022
RWP 21-03
http://doi.org/10.18651/RWP2021-03
This paper supersedes the old version:
"Government Loan Guarantees during a
Crisis: The Effect of the PPP on Bank
Lending and Profitability"
Loan Guarantees in a Crisis: An Antidote to a
Credit Crunch
W. Blake Marsh∗ Padma Sharma∗†
April 11, 2022
Abstract
Credit contractions are costly, but policymakers have limited tools to counter
them. In this paper, we examine the efficacy of public credit guarantees as anti-
dotes to a credit crunch by studying the Paycheck Protection Program (PPP).
We find that the program averted a historic credit crunch at a time when
banks were unlikely to meet firm credit needs by risking their own capital.
Our evaluation incorporates selection effects emanating from banks’ partici-
pation decision on both the extensive and intensive margins. Risk-aversion,
rather than profitability, motivated bank participation in the program. Indeed,
even as the program boosted loan growth among participants, it attenuated
profitability.
JEL Codes: C11, G21, G28, H12
Keywords: Paycheck Protection Program, government loan guarantees, credit
crunch, bank liquidity, net interest margins, Bayesian modeling
∗
A previous version was titled “Government Loan Guarantees in a Crisis: Bank Protections from
Firm Safety Nets”. We thank Jacob Dice for excellent research assistance. We thank Sriya Anbil,
Brent Bundick, Christoffer Koch, Paolina Medina, Camelia Minoiu, Anna-Leigh Stone, Greg Udell,
and Angela Vossmeyer for helpful comments and discussion. We also thank seminar participants
at the 2021 Community Banking in the 21st Century Conference, 2021 IFABS Conference, 2021
Financial Management Association Annual Meetings, 2021 Women in System Economic Research
Conference, 2022 Midwest Finance Association Annual Meetings, the Federal Deposit Insurance
Commission, the Federal Reserve Bank of Kansas City, and Monash University for helpful feedback.
The views expressed in this paper are those of the authors and do not necessarily represent those of
the Federal Reserve Bank of Kansas City or the Federal Reserve System. All errors and omissions
are our own. Research Division, Federal Reserve Bank of Kansas City, 1 Memorial Drive, Kansas
City, MO 64198. E-mails: Blake.Marsh@kc.frb.org; Padma.Sharma@kc.frb.org
†
Corresponding Author.
1 Introduction
Credit contractions are costly. Indeed, loan supply reductions were a principal con-
tributor to both the Great Depression and the more recent Global Financial Crisis–
two of the deepest recessions in U.S. history. Despite the potentially severe costs
of credit contractions, however, policymakers have few tools available to boost tight
loan supply in the depths of a crisis.
In this paper, we examine a policy tool that is seldom used as an emergency mea-
sure during economic crises: the public credit guarantee. Bachas, Kim, and Yannelis
[2021] exhaustively uncovered the favorable effect of credit guarantees on bank credit
supply during normal times, and prior empirical studies have illuminated the role of
credit guarantees in supporting lending when banks are capacity constrained due to
impaired balance sheets [Ono, Uesugi, and Yasuda, 2013; Wilcox and Yasuda, 2019;
Boeckx, de Sola Perea, and Peersman, 2020]. However, the efficacy of guarantees as
a crisis-response measure when banks are not under duress is still an open question.
More specifically, we ask do guarantees preserve bank lending in a crisis whose ori-
gins were exogenous to the banking system? In theory, public credit guarantees are
well suited to such situations because they can help avert a credit market collapse
during a crisis that emerges from deteriorating risk perceptions [Mankiw, 1986]. The
answer to this question clarifies to policymakers the broader role of guarantees be-
yond serving as an intervention in banking crises and redressing credit rationing in
normal times. But empirical evidence is scant owing to limited instances in which
guarantees are used outside of banking crises.
We quantify the effects of public credit guarantees under these circumstances by
examining the U.S. Paycheck Protection Program (PPP), a credit guarantee program
that was introduced in response to the COVID-19 pandemic– an exogenous health
crisis. At its onset, the pandemic and associated containment measures stifled eco-
nomic activity and undermined business revenue and profit expectations across the
United States. U.S. banks, in aggregate, were stable and healthy, but began actively
tightening lending standards and setting aside additional provisions in anticipation of
credit losses from the pandemic’s economic ravages. We find that the PPP, however,
1
offset a potentially large credit crunch through two channels– by stimulating growth
in loans directly covered by the guarantee, and by providing a backstop that pre-
vented contractions of non-guaranteed loans on bank balance sheets. Our results also
shed light on bank incentives to participate in the program. Risk-aversion, rather
than profitability, induced program participation, both along the intensive and ex-
tensive margins. Relatedly, we find that the program did not did not enhance bank
profitability– PPP loans compressed net interest margins relative to pre-pandemic
levels among participating banks– but instead forestalled a contraction in their loan
books and averted an even larger decline in their margins.
We design our evaluation of the PPP by recognizing that banks likely selected
into the program strategically based on its features[Joaquim and Netto, 2021]. A
unique feature of the PPP was that it relied on an initial private capital investment
by financial intermediaries that could later be reimbursed with federal funding if the
loans were forgiven or defaulted. This feature of the PPP entailed inherent trade-
offs for banks. First, banks had to balance risk and profitability concerns. Although
PPP loans carried no credit risk, they yielded a low interest rate of one percent, and
processing fees that accrued only over the life of the loan. Second, banks incurred
the opportunity cost of funding new loans using their own capital until the loans
were fully reimbursed upon forgiveness or default. Participating banks would likely
evaluate whether to extend PPP loans at the cost of reducing other non-guaranteed
lending or to simultaneously expand both types of loan portfolios. In other words,
extending risk-free, but low-yielding PPP loans may have induced banks to rebalance
the composition of their risky assets with a view toward maintaining profits.
Addressing bank trade-offs in evaluating the PPP gives rise to two empirical
challenges– selection effects and simultaneity in the determination of PPP intensity
and bank lending and profits. To address these empirical challenges, we jointly model
the decision to participate, the intensity of participation, and bank outcomes in a
Bayesian framework. Importantly, the joint model addresses selection effects inherent
in a bank’s program participation decision and their choice of the extent of partici-
pation. In other words, we account for selection effects from both the extensive and
intensive margins of participation. We additionally address the simultaneous deter-
2
mination of bank decisions and outcomes by specifying covariances across equations.
Identification requires variables that affect PPP participation and intensity without
directly affecting bank outcomes. Accordingly, we require variables that influence
the bank’s PPP participation decisions but are otherwise uncorrelated with lending
or profitability outcomes.
We use pre-pandemic technological efficiency to measure exogenous variation in
the probability of bank participation. Banks’ technology platforms were especially
salient in processing PPP loans, in no small part due to the SBA’s digital application
processing requirements. In addition, high borrower demand for PPP loans required
banks to process large application volumes expeditiously. Banks reliant on manual
application processing would have likely found their processes to be inadequate for
participation. Once the participation decision was made, however, technological
efficiency did not directly influence participation intensity because banks may book
large- or small-denominated loans with the same processing platform. One may also
be concerned that the state of a bank’s technology may be correlated with long-
term bank profitability and balance sheet characteristics as well. We address these
correlations by controlling for pre-pandemic profitability and balance sheet measures.
We next address a bank’s choice of the intensity of participation in the PPP. In
this case, we use the deposit-weighted share of employment in COVID-affected in-
dustries, such as hospitality and retail, to isolate demand for PPP loans from supply
effects. We turn to the core principle underlying the PPP’s enactment– to protect
jobs at small businesses– in measuring the demand for PPP loans. Because jobs in
contact-intensive sectors were most at risk from pandemic-abatement measures such
as lockdowns and social-distancing, firms from these sectors faced an immediate
need for support under the program to retain workers and to survive the economic
disruption. Indeed, studies that evaluated the design and roll-out of the program
noted that firms in COVID-affected sectors were over-represented in the pool of PPP
applicants [Bartik, Cullen, Glaeser, Luca, Stanton, and Sunderam, 2020b; Balyuk,
Prabhala, and Puri, 2021]. Our measure of the predominance of COVID-affected
sectors in a bank’s region of operation is thereby closely related to the pandemic-
induced demand for PPP loans. Moreover, our measures of bank outcomes, lending
3
and profits, become directly related to these exposures only if the share of exist-
ing bank loans to these sectors mirror their share in a bank’s operating region. In
this setting, participation in the PPP, lending growth, and profitability would si-
multaneously increase with the presence of contact-sensitive sectors in the bank’s
operating area. However, sector-specific lending specialization by banks [Blickle,
Parlatore, and Saunders, 2021; Paravisini, Rappoport, and Schnabl, 2015] precludes
such a one-to-one mapping between geographic concentration of industries and their
respective balance sheet loan share, breaking the direct relationship between our
COVID-affected employment measure and bank outcomes.
Our results indicate that selection effects were present despite widespread partic-
ipation in the program. We find that larger and more profitable lenders, who were
consequently better positioned to originate and hold loans until forgiveness, were
more likely to participate in the program. These findings have been confirmed by
others in the emerging PPP literature [Anbil, Carlson, and Styczynski, 2021; Lopez
and Spiegel, 2021], however, our model additionally highlights that risk-aversion
emerged as the more dominant concern when banks weighed risk and profitability
considerations. Banks with larger C&I loan portfolio concentrations and greater
undrawn commitments, and thus facing greater loan loss risk from the economic
downturn, participated in the PPP more actively. Consistent with this risk-aversion
explanation of bank participation, we find that banks that were riskier ex-ante, as
measured by lower leverage capital ratios, were both more likely to participate and
originate more PPP loans relative to the size of their total lending portfolio.
Our observation that banks were driven to participate in the PPP by risk-aversion
rather than profitability is further confirmed by our results on bank outcomes. We
estimate that a one percentage point increase in PPP participation intensity resulted
in a 4 basis point decline in NIM, or a 10 percent decline relative to 2019 levels for
the average bank, during the quarters when the program was active. This result is
driven by the fact that the PPP was not immediately profitable for participating
banks because the loans offered low interest rates and deferred fee recognition. We
also find that participating banks expanded their business loan portfolios due to PPP
lending, but not outside the program. Banks grew their overall C&I loan portfolio by
4
10 percent but their non-PPP C&I portfolio declined modestly on a year-over-year
basis per percentage point increase in PPP intensity. Incremental participation in
the PPP also did not induce risk-taking in the form of growth in Commercial Real
Estate (CRE) loans.
Counterfactual results from our joint model further clarify the PPP’s effects on
bank lending and profits by depicting participants’ outcomes in the event they had
not participated in the program. Using counterfactual lending growth, we evaluate
whether the program crowded out private capital or provided an additional boost to
business lending. We estimate that absent PPP lending, C&I loan growth would have
contracted during the second half of 2020. Our results suggest that the counterfactual
average decline of 78 percent in commercial loans, would have been similar to the
runoff rates observed for the worst hit loan portfolios during the Global Financial
Crisis. We additionally find some evidence of spillover effects across the loan book
with PPP participation precluding a large runoff in commercial real estate loans. In
total, our results show that participation in the PPP forestalled a large contraction
in business lending outside the program, and by doing so, provided a modest boost
to net interest margins at participating banks.
Most importantly, our counterfactual results indicate that the PPP helped avert
a credit crunch primarily by facilitating lending by riskier banks. Chodorow-Reich,
Darmouni, Luck, and Plosser [2021] document that the PPP offset a pandemic driven
credit crunch among commercial borrowers, particularly among small firms obtain-
ing capital from large banks. We complement these findings by both quantifying
how severe that credit crunch might have been and identifying the lender attributes
that drove the contraction in private credit. Indeed, a decomposition of our loan
growth counterfactual result shows a clear pattern of risk aversion as the main driver
of a potential credit crunch. The PPP’s effectiveness lies in inducing participation
among such risk-averse banks that might have otherwise contracted lending. Specifi-
cally, larger and less capitalized banks would have contracted lending by the greatest
amount. These are precisely the banks we find were more likely to participate in the
PPP program. That said, our results also suggest a role for the importance of rela-
tionship lending at community banks– a role emphasized by Li and Strahan [2021]
5
and Balyuk et al. [2021]. Banks with larger C&I loan concentrations, and thus those
with more prior relationships, were more likely to participate in the PPP, and to do
so more intensively. Our counterfactual results, however, illustrate the limitations
of relationship lending and the importance of supplementary loan guarantees during
a crisis. In the absence of the PPP’s loan guarantees, banks would have retrenched
lending during the unprecedented pandemic-led crisis, even if they had long-standing
relationships with borrowers.
Our results point to the interventionary role for credit guarantees following a large
economic shock. Previous work has shown that credit rationing is costly and credit
guarantees can be welfare-improving when they increase lending to optimal levels in
normal times [Stiglitz and Weiss, 1981; Gale, 1990; Mankiw, 1986]. We demonstrate
the effectiveness of credit guarantees as emergency measures during crises. The ex-
periences from the PPP illustrate that public guarantees can be especially effective
in forestalling a credit crunch during a crisis that sharply exacerbates banks’ un-
certainty about borrowers’ ability to repay. Avoiding credit rationing during such a
crisis supports the eventual recovery by limiting the severity of the economic down-
turn [Bernanke, 1983; Chodorow-Reich, 2014]. Consequently, in crises that emerge
outside the banking sector, guarantees support credit supply by fostering banks’
willingness to lend, even when their ability to lend is not directly impaired.
The rest of the paper proceeds as follows: Section 2 describes the PPP program
parameters, the Bayesian model setup, and the data and identification restrictions
needed to estimate the model. Section 3 examines the question of which bank charac-
teristics predicted PPP participation. Section 4 examines the effect of PPP lending
on bank balance sheets and income. Section 5 estimates lending counterfactuals ab-
sent PPP program participation to assess whether PPP lending crowded out other
borrowing. Section 6 provides robustness exercises, and Section 7 concludes.
6
2 Institutional Background and Methodology
2.1 The Paycheck Protection Program
At the onset of the COVID-19 pandemic, businesses faced the prospect of unprece-
dented revenue shortfalls and potential closure from the direct health effects of the
pandemic as well as public health measures that were introduced to curtail its spread.
Employment at firms with fewer than five hundred employees fell by 18 percent dur-
ing the first quarter of 2020, and small firms had cash reserves to cover only two weeks
of expenses early in the crisis [Bartik, Bertrand, Cullen, Glaeser, Luca, and Stanton,
2020a]. The Paycheck Protection Program was designed to enable small businesses
to endure these economic ravages without having to downsize their workforce.1
The PPP provided forgivable, low-cost, government guaranteed loans via the
Small Business Administration (SBA) to enhance small businesses’ survival prospects
and help retain workers. PPP loans were generally to U.S. small businesses with 500
or fewer employees, however, eligibility was determined based on the SBA’s firm size
standards which can vary across industries. Forgiveness rules required the bulk of
the funds be used to cover payroll expenses, but payments toward utilities, state and
local taxes, interest and rent on existing mortgages and leases were also permitted
to a limited extent. Overall, the PPP was a fiscal tool designed to use forgivable
loans to preserve employment at small businesses that may have otherwise been lost
to the economic disruption from the pandemic.
Banks and other financial institutions funded and disbursed PPP loans to busi-
nesses, and participation in the program was entirely subject to lenders’ discretion.
Because PPP loans were risk-free, lenders received a modest interest rate of 1 per-
cent.2 In addition, banks received processing fees on a sliding scale, which were
entirely paid by the SBA and were recognized as interest income over the term of
the loan. During most of the program, loans had five year maturities, and payments
1
The Paycheck Protection Program (PPP) was created as part of broader legislation that pro-
vided fiscal support program enacted under the Coronavirus Aid, Relief, and Economic Security
(CARES) Act passed on March 27, 2020.
2
Initially, the interest rate on PPP loans was set at 50 basis points but was increased to en-
courage greater bank participation. See Hayashi [2020] for details.
7
were deferred until forgiveness or approximately sixteen months after disbursal if bor-
rowers did not apply for forgiveness in the intervening period. The revenue stream
on PPP loans therefore depended on borrowers’ choice of timing in applying for
forgiveness and the speed with which the SBA processed forgiveness applications. 3
Because banks funded and held PPP loans on their balance sheet until repayment
or forgiveness, the Federal Reserve established the Paycheck Protection Program
Liquidity Facility (PPPLF) to address lenders’ liquidity and capital constraints. The
PPPLF provided low cost liquidity to banks by lending funds at 35 basis points
against whole PPP loans as collateral. The facility also eased capital constraints for
banks by excluding loans used as collateral in the PPPLF from the calculation of
the leverage ratio. In setting aside capital for PPP loans, banks were likely to be
constrained by the leverage ratio rather than the risk-based capital ratio because
PPP loans were assigned a zero risk weight under the CARES Act. Analogous to
the PPP, participation in the PPPLF was voluntary. This facility was established
to expand participation by banks and other financial institutions that may not have
had sufficient liquidity or capital space to disburse PPP loans.
2.2 Bayesian Joint Model Setup
We measure the effect of the PPP on bank loan growth and profits by controlling for
changes in these measures that arise from bank decisions rather than the program’s
features. Banks decided whether and how much to participate in the PPP due to
the voluntary nature of the program. These decisions introduce a selection bias,
and render the main treatment variable– PPP loans to total loans– endogenous. We
address these econometric challenges by modeling banks’ decisions to participate in
the PPP, and their choice of PPP intensity, jointly with the effect of PPP on bank
outcomes. Figure 1 illustrates the structure of our statistical model. Equation (1)
defines the decision to participate in the PPP, equation (2) models the intensity
of participation, equation (3) models the financial outcomes of participants, and
equation (4) models the outcomes of non-participants.
3
PPP loan terms are described in detail in Appendix B.
8
Figure 1: Joint model of PPP participation, intensity, and bank outcomes
Selection into PPP - all banks (eq. 1)
Participants Non-participants
v
PPP intensity (eq. 2)
v v
Bank outcomes (eq. 3) Bank outcomes (eq. 4)
By modeling participation and intensity in two separate equations, we allow for
the possibility that bank characteristics may distinctively affect the two decisions.
Our model follows the multivariate structure in Vossmeyer [2016] for sample selection
and treatment, but specifies a continuous treatment instead of a censored treatment
specified in the former study. An alternative modeling structure consists of collapsing
equations 1 and 2 into a single equation and specifying PPP intensity as a censored
outcome that takes values of zero among non-participants. This setup assumes that
bank propensity to participate and extent of participation respond identically to
balance sheet constraints and program characteristics. Instead, we adopt a more
general modeling approach where we estimate distinct parameters associated with
PPP participation and intensity.
The model structure in Figure 1 is formally represented by Equations | - 4. The
control variables x; are common to all four equations and consist of pre-determined
bank and operating region characteristics discussed below. We specify a multivariate
normal distribution, N4(0,Q), for the errors €; = (€i1, €:2, 3, €:4).
Selection into PPP - all banks: yi, = x}814+ 2171+ eu, (1)
PPP intensity - participants: yj = x,Ge+ zi2y2 + €2, (2)
Bank outcomes - participants: yj = xi33 + yi2d + &3, (3)
Bank outcomes - non-participants: yj = x,G44+ ea. (4)
Equation 1 represents the probability of each bank participating in the PPP.
The outcome y7, in this equation is a continuous latent variable that represents
bank 2’s underlying net utility from participating in the program relative to non-
participation. Bank participation signifies that this net utility is positive. The
observed binary outcome, y;1, is accordingly related to the latent variable through
the indicator operator, y;1 = 1(y;, > 0). The covariates in this equation are bank-
level controls, x{,, and a covariate, 2:1, that is correlated with particpation, but not
with the errors, PPP intensity, or bank outcomes given the covariates.
Equation 2 describes banks’ decision rule to determine the intensity of partic-
ipation in the PPP, yj;2, measured as the ratio of PPP loans to total loans. The
covariates consist of bank-level controls, x{,, and an instrument, zj2, that is indepen-
dent of the errors and bank outcomes but related to the treatment, yj2, as specified
in Li and Tobias [2014] and Greenberg [2012].
Equation 3 measures the main treatment effect of interest, namely the effect of
incremental participation in PPP, yj2, on bank outcomes of participants, yi3, given
bank-level controls. yj; enters this equation as an endogenous variable and its co-
efficient is the treatment effect, 6. Finally, Equation 4 measures outcomes for non-
participating banks.
One of the main advantages of undertaking a joint modeling approach is that it al-
lows us to incorporate covariances between outcomes across Equations 1 - 4. The co-
10
variance matrix Q depicts the relationships between unobservables €; = (€;1, €i2, €i3, €:4)-
Lo Qyg Qi Qi
Qi Qe Qa3
031 O32 OQ33
Qy, » Qa
Q=
The term QQ). measures the covariance between unobservables underlying the de-
cision to participate and the intensity of participation. The covariance terms (013
and (4,4 record the relationship between unobservables pertaining to the decision to
participate and bank outcomes for participants and non-participants, respectively.
The covariance term (293 records the effect of unobservables across the intensity of
participation in the PPP and bank-level outcomes. The elements (24 and (34 are
not identified as they correspond to covariances across outcomes for participants and
non-participants, which are mutually exclusive.
We obtain the likelihood for this model by partitioning the equations into out-
comes and covariates pertaining to participants and non-participants, as described
in Appendix D. We denote N, and N,,, as the set of participant and non-participant
banks in the sample. The complete-data likelihood function for the full sample of
observations combines the elements pertaining to each group of banks,
FY, Vili, A, QA, Qnp) = I] [fv Yipl Hip, p))] I] [Ev (Yinp| inp» Qnp)]. (5)
i€Np i€Nnp
We assign independent multivariate normal priors to the coefficients f(@) =
fr (0|8o0, To), where 6 = [71, 72,6, 8], and B = {(41, Bo, 63, b4}. The covariance ma-
trices Q, and 2, are assigned Inverse Wishart priors, f(Q,) = fow(Qp|Y%p, Qp), and
f(Qnp) = forw(Qnp|Ynp; Qnp), Which are independent of priors assigned to the co-
efficients. On combining the complete-data likelihood, and priors, we obtain the
augmented posterior as follows.
F(A, Q, Qnps ViIY) & F(Ys YL A, Mp, Qnp) F(9) F(Qp) F(Qnp) (6)
11
The Markov Chain Monte Carlo (MCMC) algorithm used to estimate this model
utilizes the estimation approach for incorporating multiple selection mechanisms in
Li [2011] and Vossmeyer [2016]. The steps underlying the algorithm and the results
from simulation exercises are provided in Appendix D. To evaluate the effect of the
exclusion restriction in Equation 1, our simulation exercises consider a setting in
which we specify an instrument in the selection equation and another, when there
is no instrument. We recover the true values of parameters within a 95 percent
posterior credibility interval in both cases.
2.2.1 Data and Identification
We require data on bank balance sheets, PPP lending activity, and various local
measures of both the pandemic’s impact as well as the economic well-being of the
local area. Data on bank balance sheets and PPP lending are drawn from the FFIEC
call reports.4 We determine a bank’s local market using the Summary of Deposits
(SOD) data, which is an annually collected FDIC data set that reports the location
and holdings of bank branches and their booked deposits. We also collect data
on local macroeconomic and health conditions. Data on local COVID case counts
are collected from Johns Hopkins University’s COVID database.Local employment
by industry is drawn from the Current Employment Statistics database from the
Bureau of Labor and Statistics.
Our bank sample considers only community banks, defined as banks with less
than $10 billion in total assets. The majority of the nearly 5,000 banks operating in
the U.S. are below this asset level, providing the widest possible source of variation.
The community bank focus also provides us with a set of banks focused on business
lending as a core activity. Larger banks often have less uniform business models
4
Call reports are collected by federal banking regulators on all supervised institutions at the
end of each calendar quarter. The Call Reports contain a wide variety of items on bank balance
sheets, income, and regulatory capital. As of the second quarter of 2020, these forms also collect
quarter-end balances and counts of PPP loans outstanding. PPP loans pledged to the PPPLF are
also reported. Specifically, an item reporting the quarterly average balance of PPP loans pledged
to the PPPLF allows adjustments to the leverage ratio in each quarter the PPPLF was active. We
use these items in our PPPLF analysis reported in the appendix.
12
that can be more complex or specialized. Inclusion of these banks would complicate
the analysis.5 Overall program participation was broad within the community bank
space. On average, about 85 percent of community banks reported at least one PPP
loan outstanding at quarter end on the Call Report between 2020:Q2 and 2020:Q4.
Participation across all quarters among community banks was slightly higher, with
about 87 percent of community banks reporting a PPP loan at the end of any quarter
in 2020.
We achieve identification of our model by specifying exclusion restrictions in the
selection and intensity equations. In principle, we can obtain identification in the
selection step by relying on the functional form of the normal distribution.6 But
identification is enhanced by the exclusion restriction. Leung and Yu [1996] show
that the absence of an exclusion restriction in the sample selection model primarily
exacerbates collinearity and results in a deterioration in prediction rather than in
parameter biases. Our main interest from the estimation of equations 1 through 4
lies in the estimates of the parameters rather than predicted outcomes. Moreover,
results from our simulation exercises in Appendix D show that we recover true values
of parameters both with and without the exclusion restriction, but arrive at more
precise estimates in the former case.
We use two main instruments to estimate equations 1 - 4, corresponding to the
variables denoted as zi1 and zi2 in equations 1 and 2. The variable for zi1 should
be related to PPP lending through its impact on participation while zi2 should be
related to bank outcomes through its effect on intensity.
We specify technical expenses relative to assets as the exogenous variable zi1 in the
equation for participation that is excluded from subsequent equations. Specifically,
5
We also exclude non-deposit trusts from our sample due to their unique business model. Non-
deposit trusts do not operate as typical deposit banks and instead primarily conduct fiduciary
business and hold only limited deposit types. See U.S. DOL’s or more information. Non-deposit
trusts dis not participated in the PPP.
6
Identification by functional form can be understood by contrasting the sample selection model
with Heckman’s two-step procedure, a classical alternative that requires an exclusion restriction.
Because the difference between the two methods is the specification of the normal likelihood, identi-
fication without an exclusion restriction arises mainly from this distributional assumption [Cameron
and Trivedi, 2005].
13
we consider pre-pandemic payments toward data processing and telecommunications
expenses to third-party vendors. Higher expenses per dollar of assets represent likely
inefficiencies from lower prior investments on technology and the inability to develop
in-house technical platforms.7 Technical efficiency likely enhanced bank PPP par-
ticipation. We hypothesize that a key driver of banks’ PPP participation was an
ability to build loan processing platforms internally. Loan processing ability was
particularly important given high borrower demand and the quickness with which
competitor banks originated loans. Pandemic-related branch closures and the re-
duced willingness of borrowers to visit bank branches that were opened also likely
required banks to have at least some ability to reach customers virtually. For these
reasons, banks already equipped with efficient technical platforms would have had the
capability to participate in the program. That said, technical efficiency should not
affect lending intensity. Participating banks could book large or small-denominated
loans with the same processing platform. Indeed, the FDIC’s Quarterly Report noted
that “banks with greater technology investment made a larger share of loans of all
sizes” [FDIC, 2021]. Because technology expenses directly relate to participation,
and only indirectly to loan amounts through participation, we exclude this variable
from the equation for participation intensity.
We implement the exclusion restriction on zi2 by defining it as a measure of
pandemic-induced demand for the PPP, unrelated to bank supply considerations.
In particular, this variable is the deposit-weighted share of COVID-affected employ-
ment in a bank’s region of operation, which represents an aggregate measure of
potential pandemic-related demand for PPP loans. We construct this variable as
shown by equation 7 where Empj is the employment in COVID-affected industries
in county j, and di,j is the total amount of bank i0 s deposits in county j as reported
in the SOD data. The main exclusion restriction is that the share of employment in
contact-sensitive sectors such as hospitality and retail does not directly affect bank
profitability and loan growth outside of the PPP, which are outcomes in subsequent
7
Call report instructions for these items are open to banks’ interpretation. However, discussions
with bank examination teams reveal that these items mostly capture outsourcing of technology needs
to third-parties, consistent with our requirements to measure lack of in-house technological ability.
Examples provided were expenses to build and maintain a website or to cover general IT needs.
14
equations. This measure disrupts the simultaneity in the determination of bank out-
comes and PPP intensity by isolating the variation in participation intensity that
arises from firm demand for loans under the program. Indeed, in their evaluation of
the PPP, Bartik et al. [2020b] and Balyuk et al. [2021] noted that firms in COVID-
affected sectors were over-represented in the pool of PPP applicants. Glancy [2021]
found that PPP loans predominantly flowed to firms located close to bank branches.
Taken together, the predominance of COVID-affected sectors in a bank’s region of
operation is closely related to the pandemic-induced demand for PPP loans.
PJ
j=1 Empj di,j
Zemp,i = PJ , (7)
j=1 di,j
Exposure to pandemic-affected sectors is only indirectly associated with banks’
profits and loan growth through PPP intensity. Bank outcomes become directly
related to exposure to pandemic-affected sectors if the share of pre-existing bank
loans to these sectors mirror their share in a bank’s operating region. In this setting,
the presence of contact-sensitive sectors in a bank’s operating area would simultane-
ously determine its participation in the PPP, lending growth and profitability. These
co-movements would arise from strategic bank objectives to preserve the health of
existing borrowers by issuing them government-guaranteed loans. However, studies
on sectoral specialization in bank lending [Blickle et al., 2021; Paravisini et al., 2015]
have shown that banks typically specialize in lending to specific sectors, that may not
necessarily conform to the sector’s employment share in a bank’s operating region.
Since the specialized nature of bank lending disrupts the direct relationship between
our measure of COVID-affected employment and bank outcomes, we exclude this
measure from subsequent equations for bank lending growth and interest margins.
The control variables specified by the vector x0i1 include bank controls– asset size,
share of business loans, capital and liquid asset ratios, and profitability– and local
health conditions measured by COVID cases per capita. Local health conditions are
constructed based on local bank deposit weights similar to the local employment
measure using equation 7.8
8
COVID case data are reported daily at the county level. We average these daily counts by
15
2.3 Summary Stats
Table 1 shows the summary statistics for our core Call Report sample. The table is
divided into participants with PPP to loan shares above the median and those below
the median share. Non-participants, defined as those that do not report any PPP
loans outstanding on their Call Reports, are shown in the far right columns. Overall,
banks with larger PPP loan shares have larger C&I loan concentrations, more unused
C&I loan commitments, more core deposit funding and liquid assets, and are slightly
larger than banks with lower PPP loan shares. High share banks also have slightly
lower capital ratios but are somewhat more profitable prior to the pandemic. Banks
with higher PPP shares also underwent larger technical expenses to assets, and were
more exposed to COVID-affected employment share. Post pandemic, we see that
high participating banks had slightly lower net interest margins (NIMs) and had
a larger drop in NIMs from their pre-pandemic averages. C&I growth overall was
higher, which includes the impact of the PPP loans, but was lower for non-PPP loans.
CRE growth was higher for banks with larger PPP loan shares. Non-participating
banks were less profitable than both participating groups, but made significantly
more C&I loans compared to the C&I growth rate of participating banks less PPP
loans. Non-participants grew their CRE portfolios less than the two participating
groups.
3 Who Participated in the PPP and How Much?
The first step in our empirical model, and the PPP loan origination process consists of
banks deciding whether they want to participate in the program. In the second step,
banks that opted to lend PPP loans determine the intensity of participation. We
show that the selection effects from these bank decisions were non-ignorable. Banks’
risk-aversion, and their capacity to fund PPP loans determined their participation
on the intensive and extensive margins.
Table 2 reports the determinants of bank participation and intensity of participa-
county to determine quarterly exposure rates that can be linked to the quarterly bank data.
16
Table 1: Summary Stats By PPP Lending Intensity
High PPP Low PPP Non-Participants
Mean Std. Dev Mean Std. Dev Mean Std. Dev
Pre-pandemic Averages
Tech Exp. to Assets 0.20 (0.13) 0.18 (0.14) 0.21 (0.19)
COVID-affected emp. share 19.69 (6.99) 17.05 (8.38) 18.33 (10.12)
C&I to Assets 10.85 (6.93) 7.57 (5.33) 8.27 (9.81)
C&I Commitments to Assets 15.42 (9.78) 9.84 (6.69) 10.09 (11.00)
Unused C&I Commitments to Assets 4.57 (3.87) 2.26 (2.32) 1.83 (2.96)
Small C&I to Assets 6.22 (4.00) 5.31 (3.81) 6.42 (8.42)
Core Deposits to Assets 71.62 (10.29) 68.09 (10.45) 67.50 (13.25)
Liquid Assets to Total Assets 20.63 (11.90) 19.09 (11.38) 25.17 (15.21)
ALLL to Total Loans 1.32 (0.64) 1.34 (0.59) 1.50 (1.21)
Total Assets ($ Millions) 0.68 (1.02) 0.42 (0.87) 0.23 (0.63)
ln(Total Assets) 12.78 (1.10) 12.20 (1.09) 11.59 (1.05)
Leverage Ratio 10.90 (2.20) 11.85 (3.21) 12.77 (4.44)
Tier 1 Ratio 15.60 (5.80) 17.57 (7.05) 21.49 (10.36)
ROA2019 Avg 1.19 (0.61) 1.19 (0.57) 0.96 (0.70)
Post-Pandemic Outcomes
PPP Share 13.15 (6.98) 3.91 (1.83) 0.00 (0.00)
NIM 3.46 (0.59) 3.49 (0.62) 3.38 (0.78)
∆NIM -50.06 (49.65) -39.57 (38.07) -48.65 (47.38)
CI Gwth 129.97 (118.09) 51.47 (62.72) 10.14 (36.46)
CI Gwth Less PPP -3.70 (22.15) -2.64 (25.11) 10.14 (36.46)
Total Banks 1,824 1,689 378
Notes: Pre-pandemic outcomes are averaged over all of 2019. High PPP banks are those with exposures greater than
the median PPP loans to total loans share. Banks with low exposures are those with PPP loans to total loans shares
less than the median. Non-participants are banks that did not report holding any PPP loans over 2020:Q2 or 2020:Q3.
tion in the PPP. The results represent estimates for equations 1 and 2, respectively.
The results are based on the specification where change in NIM relative to 2019 levels
is the outcome variable. The results from the remaining specifications, where the
outcomes are C&I growth, C&I growth outside of the PPP, and CRE growth, are
qualitatively similar and are provided in Appendix D.3.
Column (1) reports the determinants of a bank’s decision to participate in the
PPP. Technologically efficient banks were more likely to opt into the program in that
participation is negatively associated with technology expenses relative to assets.
The posterior distribution shows that this coefficient takes negative values with a
probability of 90 percent. As discussed in Section 2, because these expenses denote
17
Table 2: PPP Participation and Intensity Determinants
Participation Intensity
(1) (2)
Tech expenses to assets -0.08
[-0.2, 0.04]
COVID-affected employment share 0.04
[0.03, 0.05]
ln Assets 0.17 1.15
[0.14, 0.19] [1.01, 1.29]
ROA 0.09 0.32
[0.04, 0.14] [0.04, 0.6]
C&I to assets 0.03 0.30
[0.03, 0.04] [0.28, 0.33]
Leverage Ratio -0.05 -0.33
[-0.06, -0.04] [-0.38, -0.27]
ALLL to Total Loans -0.02 0.13
[-0.06, 0.02] [-0.12, 0.38]
Liquid Assets to Assets 0.01 0.07
[0, 0.01] [0.06, 0.09]
Cases Per 100k 0.03 0.13
[-0.01, 0.06] [-0.05, 0.31]
Constant -0.98 -8.70
[-1.34, -0.62] [-10.63, -6.79]
Note: The reported values are posterior means of the parameters, and
95% credibility intervals in brackets. The results are based on 55,000
MCMC draws with a burn-in of 5000. The results are based on the
specification that uses C&I loan growth as the main outcome variable.
payments to third-party vendors, this ratio is lower for banks that have already
invested in efficient, internal technology divisions. Bank funding capacity served
as a constraint on PPP participation–larger and more profitable banks were more
likely to participate in the program. These relationships are statistically important
as the 95 percent credibility intervals for bank size and ROA are entirely on the
positive real line. An additional factor that induced PPP participation was bank
risk aversion. Banks with larger exposures to business loan risks, as determined
18
by more concentrated shares of total C&I loans, were more likely to participate
in the PPP. Relatedly, banks with lower leverage capital ratios were more likely to
participate in the PPP, providing evidence of a risk-neutralizing effect of the program.
Banks that were more vulnerable to coming up against capital requirements likely
viewed the PPP as a means to expand lending without taking on credit risk. Other
bank controls consisting of reserved loan loss allowance relative to total loans and
liquid asset shares are not statistically important. Finally, exposure to the pandemic
in the bank’s operating region was not statistically important in explaining PPP
participation.
Column (2) represents the determinants of participation intensity. Banks re-
sponded to local demand factors by lending a larger share of PPP loans in the pres-
ence of greater COVID-affected employment in their operating region. The coefficient
of 4 basis points implies that at the average level of COVID-affected employment
of 19.7 percent, this demand-based measure explained nearly 80 basis points of the
participation intensity in the PPP. Bank risk aversion and funding capacity, which
were important drivers of bank participation also influenced participation intensity.
Larger and more profitable banks were more likely to make PPP loans at greater
relative volumes. Additionally, banks facing greater exposure to C&I lending losses
and those with lower capital ratios participated more intensively, on average. These
findings suggest that PPP loans likely allowed capital-constrained banks to transfer
risks to the government–a PPP loan could meet borrower loan demand while pre-
venting the need for the borrower to draw on existing lines of credit, which would put
additional bank capital at risk. Banks with larger shares of liquid assets participated
more intensively, which further underlines the role of pre-existing bank funding ca-
pacity. Loss reserves and COVID cases in banks’ operating region did not influence
intensity in a statistically important manner.
In related work, Li and Strahan [2021] also find a strong association between
existing C&I lending and relative PPP lending volumes similar to our own results.9
9
Estimation methods also differ across these two papers. We model PPP lending intensity using
a multi-stage model that instruments for participation using expenses on technology while Li and
Strahan [2021] use a reduced form approach.
19
However, our results differ in subtle, but important ways. Namely, our joint model
uncovers a stronger positive association between bank capital and PPP intensity
than Li and Strahan found for small banks. When combined with our C&I exposure
results, we interpret our findings as a risk-aversion channel–banks participated in the
PPP more intensively when they were already exposed to potential losses on business
loans, and when they faced capital constraints. PPP loans permitted banks to extend
loans without risking capital to future credit losses. We rely on this evidence to
highlight that risk-aversion, rather than relationship lending, was a stronger driver
of PPP participation. Our results, nevertheless reconcile with results from Li and
Strahan as banks likely relied on their existing relationships to identify potential
PPP borrowers.
3.1 The PPPLF Alleviated Capital Constraints
Our results on participation and intensity of participation reveal the remarkable find-
ing that more leveraged banks participated in the PPP and disbursed larger shares of
loans. A direct explanation for this behavior, as we previously discussed, is that PPP
loans offered a channel to transfer risks from capital-constrained banks to the federal
government. In this subsection, we show that the features of the Federal Reserve’s
PPPLF, which was established to enhance PPP participation, also contributed to
increased participation by leveraged banks. Whereas PPP loans carried a zero risk
weight and thereby did not weigh on risk-based capital ratios, they depressed issu-
ing banks’ leverage ratio. In this context, the PPPLF alleviated potential capital
constraints. PPP loans pledged to the facility were exempt from the bank’s leverage
ratio. Other studies of the PPPLF have examined the role of the PPPLF in allevi-
ating liquidity consitraints [Anbil et al., 2021]. Our analysis highlights the PPPLF’s
role in fostering participation in the PPP by alleviating capital constraints.
Table 3 shows the determinants of banks’ participation decision in the PPPLF
based on estimates of logit models. We focus on the effects of leverage ratio and C&I
exposure, but have considered a range of other controls related to liquidity, prof-
itability, and provisions. Exemption on the leverage ratio was an important driver
20
of bank participation in the PPPLF. The estimate in column (1) shows that banks
with lower leverage ratios were more likely to participate in the PPPLF. This effect
remains statistically significant even after controlling for measures of C&I exposure
in column (5). In this specification, both the share of undrawn commitments and
on-balance sheet exposures drive PPPLF participation higher. Banks with substan-
tial C&I exposure could potentially forestall additional drawdowns and credit losses,
both sources of pressure on capital, by lending PPP loans. Exposed banks could
then maintain their capital level by financing PPP loans through the PPPLF. To-
gether, the coefficients on leverage and C&I lending suggest that the exclusion of
loans pledged to the PPPLF from leverage ratio were important in driving bank
participation in the PPP.
Table 3: PPPLF Participation Determinants
(1) (2) (3) (4) (5)
Leverage Ratio -0.116∗∗∗ -0.114∗∗∗
(0.020) (0.020)
C&I to assets 0.058∗∗∗ 0.034∗∗∗
(0.005) (0.009)
Small C&I to assets 0.078∗∗∗ 0.013
(0.009) (0.013)
Unused C&I Commitments to Assets 0.103∗∗∗ 0.060∗∗∗
(0.010) (0.013)
Observations 7,022 7,022 7,022 7,022 7,022
Loglik -2,630.09 -2,583.72 -2,612.38 -2,596.61 -2,551.76
Bank controls Yes Yes Yes Yes Yes
Psuedo R2 0.07 0.08 0.07 0.08 0.10
Notes: Dependent variable is an indicator for PPP loans pledged to the PPP Liquidity Facility at the end of
the quarter. Sample is 2020:Q2 and 2020:Q3. Regressor balance sheet variables are measured as four quarter
averages from 2019.
t statistic in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
We next turn to the determinants of how much banks decided to rely on the
PPPLF in Table 4. The results represent estimates from a linear regression of the
share of PPP loans pledged to the PPPLF on bank characteristics. Our findings on
PPPLF intensity mirror those on the PPPLF participation decision. Columns (1)
through (5) show that capital constrained banks and those with a larger exposure to
C&I loans pledged larger shares of loans to the PPPLF facility.
21
Table 4: PPPLF Participation Intensity Determinants
(1) (2) (3) (4) (5)
Leverage Ratio -0.609∗∗∗ -0.541∗∗∗
(0.098) (0.097)
C&I to assets 0.518∗∗∗ 0.598∗∗∗
(0.059) (0.114)
Small C&I to assets 0.581∗∗∗ -0.199
(0.090) (0.144)
Unused C&I Commitments to Assets 0.605∗∗∗ -0.017
(0.103) (0.136)
Observations 6,940 6,940 6,940 6,940 6,940
Adjusted R2 0.041 0.056 0.046 0.044 0.059
Bank controls Yes Yes Yes Yes Yes
Notes: Dependent variable is the share of PPP loans pledged to the PPP Liquidity Facility in 2020:Q2
and 2020:Q3. Regressor balance sheet variables are measured as four quarter averages from 2019. COVID
cases are county level case counts averaged over counties where the bank operates a branch according to the
Summary of Deposit data. Daily county-level COVID case counts are drawn from John Hopkins.
t statistic in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Absent the PPPLF, many banks may have found PPP loans unprofitable given
their low interest rate relative to funding costs or may been concerned about the
capital space used by PPP loans that might put them too close to their regulatory
minimums. By alleviating these concerns, the PPPLF made the PPP more successful
among community banks than it otherwise would have been.
4 How did the PPP Affect Bank Lending and
Profits?
We quantify the effect of the PPP on bank profits and lending after conditioning
on the selection effects modeled in the previous section. Notably, incremental par-
ticipation in the PPP diluted bank profitability on account of its low interest rates
and gradual recognition of fees. Nearly all of the lending growth the PPP generated
was directly within the program. The program did not induce non-PPP C&I lending
or risk-taking under CRE categories. The countervailing effect of the PPP on bank
profits, and negligible effects on risk-taking reinforce our finding from the previous
section–risk-aversion was salient in motivating banks to participate in the program.
22
Table 5 reports the estimates for equation 3, which models bank outcomes for
PPP participants. The first row reports the main treatment effects of incremental
participation in the PPP on profitability and loan growth. In column (1), a one
percent higher participation intensity lowered NIM by 4.3 basis points relative to
2019 levels. At the mean PPP intensity rate of 8.5 percent, this estimate explains
the full decline of 33 basis points in NIM experienced by participating banks since
2019. C&I lending grew by 10.5 percent on a year-over-year basis in response to a one
percent higher intensity of PPP lending as shown in column (2). However, all of the
lending growth emanated from within the program–we find no statistically important
effect on lending growth outside the PPP in column (3). Similarly, incremental
participation in the PPP did not result in statistically important effects on risk-
taking as measured by growth in CRE loans in column (4).
We report the estimates for bank controls that reflect funding capacity and ex-
posure to losses, namely bank size and loan and loan loss allowances in Table 5.10
We trace the effect of these characteristics on bank profits and loan growth as fund-
ing capacity and exposure to losses were salient in determining participation and
intensity decisions. The PPP was more profitable and generated higher C&I growth
among larger community banks. Previously, we had found that larger banks were
more likely to participate in the program, and with greater intensity. Taken together,
these findings suggest that larger banks that were able to participate materially in
the program were likely overcame the limiting effect of the program’s low interest
rates and generated profits from originating large volumes of loans.
Banks that had reserved larger shares of their loan portfolios for loan loss al-
lowances underwent larger declines in profits and loan growth. When banks set aside
larger reserves to meet expected losses, they have a diminished pool of funds to ex-
pand lending, which likely led to lower margins depicted in column (1). Banks with
larger ALLL shares also face greater risk in their loan portfolios, possibly leading
them to pull back lending more than peers facing less risk. While lending declined
across C&I and CRE categories in columns (2)-(4), much of this decline was concen-
10
We report the estimates for the full set of control variables in Tables D.3 and D.4 in Appendix
D.3
23
Table 5: Profitability and Loan Growth Outcomes at Participant Banks
∆NIM C&I Non-PPP CRE
(bps) Growth(%) C&I Growth(%) Growth(%)
(1) (2) (3) (4)
PPP Loans to Total Loans -4.27 10.52 -0.46 0.23
[-6.03, -2.7] [9.26, 11.87] [-1.46, 0.57] [-0.54, 1.01]
ln Assets 3.98 6.20 0.13 0.36
[2.72, 5.33] [5.18, 7.19] [-0.82, 1.04] [-0.48, 1.2]
ALLL to Total Loans -3.29 -3.67 -2.72 -0.53
[-5.28, -1.32] [-6.2, -1.15] [-3.65, -1.78] [-1.23, 0.18]
Table 6: Profitability and Loan Growth Outcomes at Non-Participant Banks
∆NIM C&I Non-PPP CRE
24 (bps) Growth(%) C&I Growth(%) Growth(%)
(1) (2) (3) (4)
ln Assets -0.91 -7.47 -6.64 -3.72
[-4.99, 3.55] [-8.46, -6.49] [-7.57, -5.72] [-4.41, -3.02]
ALLL to Total Loans -9.60 -2.06 0.02 -0.93
[-12.19, -6.98] [-4.6, 0.44] [-2.42, 2.44] [-2.55, 0.68]
Note: The reported values are posterior means of the parameters, and 95% credibility
intervals in brackets. The results are based on 55,000 MCMC draws with a burn-in of
5000.
trated in non-PPP C&I where banks would be fully exposed to inherent credit risks.
This finding is consistent with the heightened risk aversion among banks exposed to
losses from business loans, which we encountered in their participation and intensity
decisions.
Table 6 reports the coefficient estimates for non-participants. Notably, the direc-
tion of coefficients associated with bank size is the opposite to that of participants.
Larger banks experienced a steeper decline in margins and a sharper contraction
in lending. This suggests that smaller non-participants expanded their loan port-
folio and exposed their capital to potential credit losses even as larger banks cur-
tailed lending, likely reflecting pandemic support from the smallest banks based on
relationship-lending. As in the case of participants, increased loan loss allowances at
non-participants was associated with a larger decline in NIM as well as a contrac-
tion in C&I lending. Since these allowances represent higher expected credit losses,
they limited lenders’ ability to use their resources to disburse risky loans and to
maintain profits. Table 7 characterizes the direction and magnitude of covariances
across the steps in the joint model–bank participation, intensity and outcomes on
profitability and lending. 11 Overall, the covariances illustrate that selection effects
from bank decisions were salient–banks that were better positioned to expand their
loan portfolios and to earn higher net interest margins strategically opted into the
program. The first row shows that participation was positively related to intensity–
banks that were most likely to participate in the PPP expended larger shares of their
loanable funds on the program. The second and third rows show that unobserved
factors related to bank participation, and intensity were positively related to the
unobserved component of bank profitability and loan growth. Finally, in line with
our expectations, the relationships between non-participation and bank outcomes in
the bottom row move in an approximately opposite direction to those in the second
row between participation and outcomes. These results support our hypothesis that
bank decisions related to the PPP were endogenous – banks that were more likely to
expand their loan portfolio and earn higher profits participated more intensively in
the program. Banks less likely to expand loans and profit from the program withheld
11
The underlying numerical estimates are presented in Table E.1 in Appendix E.
25
participation.
Table 7: Covariance Estimates from the Bayesian Joint Model
∆NIM C&I Non-PPP CRE
Gwth C&I Gwth Gwth
(1) (2) (3) (4)
COV(participation, intensity) + + + +
COV(participation, bank outcome) + + + –
COV(intensity, bank outcome) + + + –
COV(non-participation, bank outcome) – – – –
This table characterizes the direction and magnitude of covariances estimated
from the Bayesian joint model. Positive and negative relationships that are
statistically important are depicted in blue and red symbols respectively. Esti-
mates that are not statistically different from zero are represented in grey. The
numerical estimates underlying this table are in Table E.1 in the Appendix.
We investigate the timing of the PPP’s effects on bank outcomes in Table 8.
While early access to the program was critical for firms to withstand the disruptions
from the pandemic[Balyuk et al., 2021; Doniger and Kay, 2021], we find that the
timing of PPP lending was also relevant to banks’ profitability. We split our sample
period into subsamples for 2020:Q2 and 2020:Q3, the quarters when the program was
most active. 12 The first column shows that the decline in NIM we had previously
reported occurred almost entirely in the second quarter of 2020. Lending growth,
however, did not vary across the two quarters. C&I lending consistently grew year-
over-year in both quarters as shown by column (2). Similarly, columns (3) and (4)
show that the program had only a modest effect on lending outside of the program in
both quarters. Our results suggest that banks’ margins declined when they swiftly
disbursed a large number of loans early in the pandemic, when economic activity
was restricted, and financial markets were in turmoil. But, as economic activity
increased and demand for loans stabilized, banks likely developed the capacity to
process applications by assessing their impact on profit margins. In addition, the
12
Additional details about the individual quarter estimates are available in Appendix G. Results
for the fourth quarter of 2020, which is not considered in our main sample, are presented in Appendix
I.
26
latter rounds of the PPP involved smaller loan amounts, which carried a higher
percent of fees, further supporting bank margins.
Table 8: Quarterly Treatment Effects by Outcome
∆NIM(bps) C&I Gwth(%) Non-PPP C&I Gwth(%) CRE Gwth(%)
(1) (2) (3) (4)
Baseline -4.27 10.52 -0.46 0.23
[-6.03, -2.7] [9.26, 11.87] [-1.46, 0.57] [-0.54, 1.01]
Q2 2020 -6.91 10.72 0.36 0.20
[-9.15, -4.92] [8.65, 12.92] [-0.89, 1.71] [-0.71, 1.09]
Q3 2020 -0.19 9.53 -0.33 0.41
[-2.54, 2.39] [7.18, 12.04] [-2.33, 1.54] [-0.76, 1.61]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in
brackets. The results are based on 55,000 MCMC draws with a burn-in of 5000.
Finally, we reevaluate treatment effects by controlling for the sharp rise in credit
line drawdowns that banks experienced in the first quarter of 2020. Table 9 reports
our baseline result for PPP intensity from Table 5 as well as estimates from a model
that includes an indicator for banks in the top quartile of C&I growth in 2020:Q1.
We find that on controlling for large drawdowns, the negative effect of the PPP on
profit margins is only slightly attenuated. This finding suggests that the decline in
NIM from the PPP was not inadvertently reporting the effect of a sudden expansion
of the asset base from drawdowns. Growth in total C&I lending was higher in the
specification that included large C&I drawdowns. Columns (3) and (4) show that
the effect of the PPP on lending outside of the program remained muted even after
controlling for credit line drawdowns. Therefore, the effects of the PPP on lending
growth remain broadly unchanged even after addressing the conversion of off-balance
sheet commitments into on-balance sheet exposures from firm usage of credit lines.
13
13
Appendix K summarizes the direct effect of drawdowns on participation, intensity, and bank
outcomes controlling for other characteristics. We find that banks that experienced large drawdowns
were more likely to participate in the PPP, as well as undergo larger growth in C&I loans and larger
change in NIM relative to 2019.
27
Table 9: C&I Loan Draw Effects
∆NIM(bps) C&I Gwth(%) Non-PPP C&I Gwth(%) CRE Gwth(%)
(1) (2) (3) 4)
Baseline -4.27 10.52 -0.46 0.23
[-6.03, -2.7] [9.26, 11.87] [-1.46, 0.57] [-0.54, 1.01]
Baseline + CI gwth top qrtile -3.92 12.13 0.20 0.29
[-5.45, -2.37] [10.67, 13.61] [-0.78, 1.17] [-0.46, 0.99]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets.
The results are based on 55,000 MCMC draws with a burn-in of 5000.
5 Did the PPP Crowd Out Lending or Avert A
Credit Crunch?
The PPP will ultimately be deemed successful if the program met small businesses’
demand for funds that would have otherwise been unaddressed by the private sec-
tor. Alternatively, the program would be considered inefficient if it instead crowded
out private lending by banks and hindered their prospects for growth and earnings.
Would banks have been willing to extend loans at a time when firms’ need for funds
was intense, but ability to repay was uncertain? To answer this question, we generate
counterfactual rates at which lending would have grown at participating banks in the
event they had not participated in the PPP. This counterfactual analysis is different
from the potential outcome of banks had the PPP itself not been introduced. The lat-
ter is not estimable because the PPP was an unprecedentedly large support program
that was available to a broad range of financial institutions including banks, thrifts,
credit unions, and fintechs. The counterfactual in the PPP’s absence could have
been estimated if any sub-category of institutions had been ineligible to participate
in the program.
Since no major sub-class of financial intermediaries were excluded from participat-
ing in the PPP, we evaluate counterfactuals by using outcomes of banks that were un-
affiliated with the program by their own choice–namely, non-participant banks. Ac-
cordingly, our estimates represent lending and profits that participating banks would
have incurred, had they chosen not to participate in the PPP. To evaluate counter-
factuals, we multiply the estimates that correspond to outcomes for non-participants
28
(Equation 4) with the financial characteristics pertaining to participants(xi ).
Figure 3a shows that had participating banks not engaged in the PPP, they
would have contracted their loan portfolio substantially. The blue and green lines
respectively denote the average realized levels of C&I loan growth for participants
and non-participants. The yellow density represents the posterior distribution of the
counterfactual C&I growth for PPP participants. Participating banks experienced
exceptional year-over-year growth of 92%, whereas non-participants grew their loan
books by a more modest, but notable rate of 10%. If participating banks had not
associated with the program, their C&I loans would have likely contracted by 78
percent. This contraction is statistically important since the full density lies on the
negative real line. Participating banks would have substantially reduced lending and
allowed portfolio runoffs if they had not entered the PPP. This finding reveals that
the PPP served as a backstop to the banking system by preventing a contraction in
bank lending.
Our counterfactual estimate of a decline of nearly 80 percent in C&I loan books
is striking. However, Figure 2 demonstrates that our estimates align with declines
across loan portfolios during the Global Financial Crisis (GFC), especially among
loan categories directly afflicted by the crisis. In that crisis, community banks expe-
rienced shocks to commercial real estate portfolios, and specifically to construction
and land development loans [Bassett and Marsh, 2017; Friend, Glenos, and Nichols,
2013]. Figure 2 shows that peak-to-trough declines in C&I and CRE loans at com-
munity banks registered at about 15 and 20 percent respectively. Most of the run-off
in CRE loans was due to declines in the most affected portfolio, construction and
land development (CLD), which contracted by nearly 70 percent in total. Notably,
the declines in loan balances that began during the GFC persisted for several years
after the NBER recession ended.
While our results suggest that lending would have contracted more precipitously
than during the GFC, a larger and more rapid contraction in lending is consistent
with the nature of the COVID shock. The onset of the pandemic induced widespread
panic in financial markets[Acharya and Steffen, 2020], and a sudden and deep reces-
sion that was ultimately short-lived. At its genesis, however, the pandemic looked
29
Figure 2: GFC Growth Rates at Community Banks
0 0
−10 −10
−20 −20
−30 −30
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
C&I CRE CLD
Chart shows peak-to-trough growth rates of selected loan portfolios at community banks from the
Global Financial Crisis. Peak date is 2007:Q3, the start of the NBER recession period, for all
portfolios. Trough dates are the last quarter when aggregate growth was negative. Trough dates
by portfolio are: C&I, 2011:Q1; CRE, 2012:Q3; and CLD, 2013:Q1.
Source: Call Reports.
likely to rival and even surpass the Global Financial Crisis in terms of economic
costs.14 Due to the starkly weaker outlook at the onset of the pandemic, we should
expect that loan growth would decline suddenly and dramatically in the ensuing
period as our estimates suggest, rather than over a period of several years as seen
in the Global Financial Crisis. Accordingly, our findings are in line with the results
from Federal Reserve’s Senior Loan Officer Opinion Survey (SLOOS) that provided
evidence of a severe and sudden tightening of lending standards.15 These predic-
tions also align with the actions of banks with respect to other loan categories that
were not supported by government credit programs such as consumer loans. Banks
14
The U.S. economy shed about 8 million jobs during the Great Recession of 2008, and over 20
million jobs during the COVID-19 recession. Source: Authors’ calculations from Total Nonfarm
Employees from the U.S. Bureau of Labor Statistics.
15
See SLOOS results for April 2020 https://www.federalreserve.gov/data/sloos/
sloos-202004.htm and July https://www.federalreserve.gov/data/sloos/sloos-202007.
htm
30
substantially reduced consumer lending during 2020 to the extent that households,
particularly those in the subprime category, reported lack of access to credit during
this period [Horvath, Kay, and Wix, 2021].
The PPP not only forestalled contraction in C&I lending, the program also pre-
empted lending declines in the commercial real estate portfolio. Figure 3b shows
that CRE lending would have declined 52 percent if participants had forgone the
PPP program compared to a realized growth rate of 7 percent. This substantial
difference between realized and counterfactual growth further supports our finding
that the PPP served as a backstop to bank lending.
Consistent with the loan declines the PPP averted, we find that, on average,
bank profitability was also supported by program participation. Figure 3c shows that
participants had a realized decline in net interest margins of 33 basis points. Absent
program participation however, average net interest margins were projected to decline
more than 40 basis points. Although the posterior interval around this estimate
spans zero, the mass of the distribution lies on the negative real line. These findings
reconcile easily with results from Section 4. Even though incremental participation
compressed margins due to low interest rates, the program averted a larger decline
in NIM that would have arisen from a freeze in bank lending and runoffs in loan
portfolios. Undoubtedly, the PPP helped to minimize credit losses, which supported
margins, but it also upheld margins by inducing banks to continue, rather than
retrench lending.
In Table 10, we examine the main drivers underlying the remarkable decline
in counterfactual lending and margins for program participants.16 To this end, we
evaluate a measure of the average contribution of each covariate to the counterfactual,
namely, the product of the mean value of the covariate across participants with
the estimated coefficient of that covariate for non-participants. This denoted as
(g)
x̄p,j β4j , g = 1, 2, ..., 50, 000, where x̄p,j is the mean value of each covariate j across
(g)
participants and β4j is the g th posterior draw of the associated coefficient from
equation 4. The table reports the mean and 95 percent credibility interval of this
product across the 50,000 posterior draws.
16
We provide the full set of results from the decomposition in Table H.1 in Appendix H.
31
Table 10: Decomposition of predicted counterfactual outcomes for participants
∆NIM C&I Gwth CRE Gwth
(1) (2) (3)
ln Assets -32.69 -91.32 -44.71
[-68.92, 10.89] [-103.66, -79.68] [-53.63, -36.27]
ALLL to Total Loans -12.66 -2.95 -1.1
[-16.05, -9.32] [-6.25, 0.29] [-3.21, 1.01]
Note: The reported values are posterior means of the product of covariates
and parameters, and 95% credibility intervals in brackets. The results are
based on 55,000 MCMC draws with a burn-in of 5000.
Our decomposition of the counterfactual outcomes reveals two important find-
ings. First, bank size and loan loss allowances predominantly explain the full magni-
tude of counterfactual lending and margins. Second, the relationship between bank
characteristics and outcomes is fundamentally different across participants and non-
participants. Together with the differences in the pre-pandemic characteristics across
the two groups of institutions, these differences in relationships explain why partic-
ipating banks would have contracted their loan portfolios and faced greater margin
compression if they had not engaged in the PPP, compared to the banks that actually
opted out of the program.
Bank size largely explains the sharply lower counterfactual margins and loan
growth for participants relative to their realized outcomes. Contrary to participants,
margins and lending growth declined with size among non-participants. Moreover,
participating banks are larger than non-participants. Taken together, the model
predicts that the mean participants’ asset size would have contributed to a decline
in NIM of 32 basis points relative to 2019, and a reduction in C&I and CRE loan
portfolios by 91 percent and 45 percent respectively, had they not participated in
the PPP. This result points to the differences in the way small and large community
banks responded to the crisis. Smaller banks, which were less likely to participate in
the program, continued to lend loans using their own capital. The larger banks were
more likely to participate in the PPP, and those that did not engage in the program
32
Figure 3: Estimated Counterfactual Values
0.15
Participant Mean = 91.45
Counterfactual Mean = -78.03
0.1
0.05 Non-participant Mean = 10.35
0
-100 -80 -60 -40 -20 0 20 40 60 80 100
(a) C&I Loan Growth
0.16
0.14 Participant Mean = 6.99
Counterfactual Mean = -52.45
0.12
0.1 Non-participant Mean = 3.5
0.08
0.06
0.04
0.02
0
-70 -60 -50 -40 -30 -20 -10 0 10
(b) CRE Loan Growth
0.014
0.012 Participant Mean = -33.3
0.01
Counterfactual Mean = -41.49
0.008
0.006
0.004 Non-participant Mean = -44.48
0.002
0
-140 -120 -100 -80 -60 -40 -20 0 20 40 60
(c) ∆NIM
Charts show the posterior density of counterfactual YoY growth in C&I loans, average YoY growth
in C&I loans outside of the PPP, average YoY growth in CRE loans outside of the PPP, and average
change in NIM, respectively, for participating banks in the event of non-participation. The blue
and green lines represent the realized average values for these categories for participants and non-
participants, respectively, over year ending in Q2 and Q3 2020.
Source: Authors’ calculations.
33
curtailed lending. Therefore, participating banks, which were relatively larger in size,
would have retrenched lending and undergone further compression in margins if they
had not engaged in the PPP.
The second salient factor underlying participants’ counterfactuals is the ratio of
loan loss allowances to total loans. Participating banks with average ALLL ratios
were likely to undergo a 13 basis point decline in NIM relative to 2019, and declines
of 3 and 1 percentage points in C&I and CRE loan growth respectively. Loan loss
allowances, were on average higher among non-participants, and were associated
negatively with loan growth and margins among participants and non-participants.
Our results show that if participants had foregone the PPP, banks already exposed
to potential losses would have withheld lending to mitigate further losses, even at
the cost of additional margin compression.
The counterfactual values for bank lending are consistent with our finding that
bank participation in the PPP was motivated by risk-aversion. Banks that partici-
pated in the PPP would not have expanded lending to address business credit needs
had they not participated in the program. This means the PPP, crucially, did not
crowd out private lending. Instead, the PPP offset what would have been a sharp
decline in bank lending. Notably, the PPP served as a backstop to bank lending.
Minoiu, Zarutskie, and Zlate [2021] identified that the Main Street Lending Progam
(MSLP) functioned as a backstop by stimulating lending outside of the program.
The PPP, instead, provided a backstop by preventing declines in non-PPP loan
portfolios. Further, our counterfactual results highlight the limitation of relationship
lending during crises. Our counterfactual results confirm that, in fact, absent the
PPP, bank lending to businesses would have contracted substantially. While existing
relationships measured by C&I lending exposures would have mitigated this effect
to some degree, they do not fully offset the credit crunch that would have occurred
absent the PPP.
34
6 Robustness
We verify that our results are not sensitive to model misspecification issues by gener-
ating related, but distinct, exogenous variables for PPP intensity. This corresponds
to reestimating the joint Bayesian model by replacing the variable zi,2 with alterna-
tives. We consider three additional variables. First, we calculate the share of small
firm employment from the QWI data. This variable is constructed in the same fash-
ion as the variable measuring COVID-affected employment share. We weight the
share of firms with less than 500 employees in a county by a bank’s deposit share
in that county. The 500 employee cutoff roughly corresponds to the firm eligibility
criteria in the PPP. This variable acts as an alternative to the PPP demand instru-
ment created using COVID-affected sector employment. Second, we use a bank’s
core deposits to assets ratio as the exogenous variable. This variable measures a
bank’s ability to fund PPP loans without additional capital market activity. Finally,
we consider the share of unused commitments to total assets. This variable captures
a bank’s exposure to potential pandemic-induced drawdowns. In our risk-aversion
interpretation of the model results, banks should be more willing to generate PPP
loans during periods of heightened uncertainty for customers that have unused com-
mitments because they can fully capture the benefits of the government guarantee
while retaining their customer relationships without risking their own capital.
While we find these exogenous variables acceptable, they are not preferable to our
main excluded variable. COVID-affected sector employment is exogenous to both
the broader economy and bank lending. COVID’s differential effect across sectors
was unrelated to economic activity and firm access to bank funding. For these
reasons, we find this variable to be most suitable for identification. Alternatively,
while we have no reason to believe PPP’s employment eligibility cutoff was not set
arbitrarily, it could have been set in response to financing needs or COVID’s expected
impact. While we still find the share of eligible firms to be a convincing variable
for identification, it is not as strongly exogenous as COVID-affected employment.
Finally, the bank measures are least suitable as exogenous variables. First, balance
sheet characteristics will affect both profitability and lending outcomes. Second,
35
while we measure these variables as-of 2019, or during the pre-pandemic period, bank
balance sheets changed dramatically during the pandemic which also affects outcomes
for banks. For this reason, it is unclear, ex-ante, that pre-pandemic balance sheet
measures are useful indicators for predicted decision making during the pandemic,
especially when considering the sizable changes in deposit and loan growth beginning
in 2020:Q1.
Table 11 shows the standardized coefficient estimates for the PPP intensity equa-
tion using the COVID-affected employment variable and the three alternative vari-
ables. Small firm employment is negatively related to PPP intensity suggesting that
firm size is not an adequate demand control. The bank balance sheet measures, how-
ever, are consistent with our expectations. Banks with greater core deposit funding
and more unused commitments as a share of assets were more intensive PPP lenders.
These results reemphasize the selection effects across banks and show that banks bet-
ter positioned to lend, or those that faced higher loss probabilities, were more likely
to participate in the program. All these estimates are statistically important.
Table 11: Alternative Instrument Effects
COVID-affected Small firm Core Deposit Unused C&I Cmmt
Employment Employment Ratio Ratio
(1) (2) (3) (4)
Mean 0.093 -0.135 0.106 0.263
[0.07, 0.11] [-0.16, -0.11] [0.09, 0.13] [0.24, 0.29]
Note: Table shows standardized coefficients for each exogenous variable on
PPP intensity. Coefficients are estimated using the Bayesian joint model shown
in equations 1 - 4. 95% credibility intervals are shown in brackets.
Figure 4 shows the coefficients on PPP intensity, namely, the treatment effects of
the PPP on participants using each of the instruments. We find that the point esti-
mates for the change in net interest margins range from about -1 to -8 basis points
across each set of instruments. Most all the point estimates from our main speci-
fication are within the range of estimates generated using alternative instruments,
suggesting that our main findings are robust to model specification changes. For
example, all the net interest margin estimates are negative, statistically important,
36
and closely clustered, regardless of the exogenous variable used. C&I loan growth
estimates are similarly grouped. C&I lending increased between 6 and 12 percent,
and the estimates are statistically important across all model specifications. We find
only small effects for non-PPP C&I lending and CRE growth across specifications.
These estimates are not always statistically important either. This reconfirms our
result that spillovers across portfolios were limited.
Figure 4: Treatment effects by instrument
15 15
12 12
9 9
6 6
3 3
0 0
−3 −3
−6 COVID−Affected Employment −6
Small Firm Employment
Core Deposit Share
−9 Unused Commitment Share −9
−12 −12
−15 −15
∆NIM C&I Non−PPP C&I CRE
Growth Growth Growth
Source: Call Reports.
As an alternative to our Bayesian joint estimation, we use our exogenous variable
set to estimate outcomes using a classical frequentist two-stage least square approach.
In this setup, we model PPP intensity in the first stage and the outcome in the second
stage. We instrument PPP intensity using our preferred instrument, COVID-affected
employment share in a bank’s local market. The two stage least square results are
37
shown in Table 12. The first row reports the coefficients on PPP intensity on various
bank outcomes from the baseline, Bayesian joint model. Our OLS and IV estimates
of the PPP intensity coefficient are reported in rows 2 and 3 respectively.17 We find
that using OLS or instrumenting PPP intensity with COVID-affected employment
both show negative effects on NIMs and strong positive effects on total C&I lending.
The IV estimates, which account for selection of PPP intensity, are larger than the
OLS estimates in both cases. Bayesian estimates account for both, selection into
the PPP and intensity of participation. The magnitude of the treatment effect on
profitability in the Bayesian setting is larger than the IV estimate. These results
reveal the salience of selection effects and the need to appropriately model them to
generate valid estimates of treatment effects. Treatment effects on non-PPP lending
portfolios are small, limited, and not always statistically important. These results
are consistent with our Bayesian model estimates and suggest that the effects of the
program predominantly operated through the C&I portfolio.
Table 12: OLS and Two-stage Least Squares Estimation
∆NIM(bps) C&I Gwth(%) Non-PPP C&I Gwth(%) CRE Gwth(%)
(1) (2) (3) (4)
Baseline -4.27 10.52 -0.46 0.23
[-6.03, -2.7] [9.26, 11.87] [-1.46, 0.57] [-0.54, 1.01]
OLS -1.22∗∗∗ 11.26∗∗∗ -0.10∗ 0.18∗∗∗
(-5.00) (47.74) (-2.10) (4.41)
IV -3.25∗∗∗ 15.07∗∗∗ 0.77∗ 0.26
(-4.61) (15.15) (2.15) (0.87)
Notes: Table shows estimates of PPP intensity on bank profitability and balance sheet
outcomes from the Bayesian joint model (“Baseline”) as well as a standard OLS and
a two-stage least squares model. The two-stage least squares model uses the share
of COVID-affected employment in a bank’s local market as the instrument. For the
baseline model, 95% credibility intervals are shown in brackets. T-statistics are shown
in parenthesis for the OLS and two-stage least squares estimates.
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
17
Additional details on the logit, IV and OLS results are presented in Appendix J.
38
7 Discussion and Policy Implications
Under the Paycheck Protection Program (PPP), banks issued business loans that
could later be fully forgiven and reimbursed with federal funding. Banks—especially
small community banks—participated extensively, with PPP loans representing nearly
all new lending in 2020. However, the program’s effects on the balance sheets of com-
munity banks have not been fully understood.
Our results show that the PPP not only supported businesses, but also the banks
that disbursed the loans. Although the PPP carried a low interest rate, the program
ensured a modest revenue stream for participating banks when safe and profitable
lending opportunities were scarce. At the same time, by guaranteeing credit ex-
tensions, the program was able to avoid a credit crunch to small and mid-sized
businesses while revenues were falling quickly. Overall, the PPP indirectly provided
crucial support to community banks in the form of income and credit growth and
likely protected banks from business-related credit losses during the height of the
pandemic. In addition, we find that community banks with ample funding—namely,
larger and more profitable banks—were more likely to participate in the PPP; how-
ever, participating community banks with weak capital originated more PPP loans
relative to their size. This suggests that the PPP helped mitigate risk for weaker
banks at a time of high economic and financial uncertainty.
The PPP highlights a few important lessons for structuring government lending
programs in the future. First, government guarantees serve as an antidote to a credit
crunch in times of severe economic uncertainty. We generate counterfactual analyses
that show that small businesses would have likely faced steep constraints in accessing
credit during the pandemic in the absence of the PPP.
Second, the benefits of large-scale credit guarantee programs likely outweigh their
costs in the event of large, exogenous shocks like the COVID-19 pandemic but may
be less effective in a financial crisis. The PPP elicited more intensive participation
among banks that were relatively weakly capitalized. If such a program were to
be offered following a financial shock such as the Global Financial Crisis, weakly
managed banks potentially at the risk of failure may have used the program to
39
gamble for resurrection and transfer substantial risks to the federal government.
Third, the parameters of the guarantee program must balance incentives for par-
ticipation with those for underwriting. One of the reasons for the widespread take-
up of the PPP was likely the broader guarantee implicit in the program relative to
standard loan guarantee programs. European guarantee programs that were set up
contemporaneously with the PPP, and prior programs in Japan did not convert loans
to grants by means of a forgiveness procedure [Ono et al., 2013; Core and De Marco,
2021]. Indeed, other programs provided only partial guarantees for loans above a
threshold. Despite the more generous nature of the PPP’s loan guarantees, which
likely induced greater bank and firm participation, other features of the program
placed costs on participating banks and checked excessive transfer of risk to the fed-
eral government. Our findings show that bank interest margins declined with the
intensity of participation in the PPP. Low interest rates and deferral of fees until
forgiveness likely diluted margins, but also curtailed incentives for originating poor-
quality loans that may have later been deemed ineligible for forgiveness. Similarly,
requiring banks to initially use their own capital to lend these loans also likely served
to check moral hazard incentives.
Overall, the PPP serves as a new tool that may be used in times of a large,
exogenous shock to the economy. Future uses of this program may require adjusting
loan terms to ensure credit support while disincentivizing moral hazard.
40
References
V. V. Acharya and S. Steffen. The Risk of Being a Fallen Angel and the Corporate
Dash For Cash. COVID Economics: A Real Time Journal, 2020.
S. Anbil, M. Carlson, and M.-F. Styczynski. The Effect of the PPPLF on PPP
Lending By Commercial Banks. Board of Governors of the Federal Reserve System,
FEDS Working Paper, No. 2021-030, May, 2021.
N. Bachas, O. S. Kim, and C. Yannelis. Loan guarantees and credit supply. Journal
of Financial Economics, 139(3):872–894, 2021.
T. Balyuk, N. Prabhala, and M. Puri. Small Bank Financing and Funding Hesitancy
in a Crisis: Evidence from the Paycheck Protection Program. Working Paper,
2021.
A. W. Bartik, M. Bertrand, Z. Cullen, E. L. Glaeser, M. Luca, and C. Stanton. The
Impact of COVID-19 on Small Business Outcomes and Expectations. Proceedings
of The National Academy of Sciences, 117(30):17656–17666, 2020a.
A. W. Bartik, Z. B. Cullen, E. L. Glaeser, M. Luca, C. T. Stanton, and A. Sunderam.
The Targeting and Impact of Paycheck Protection Program Loans to Small Busi-
nesses. National Bureau of Economic Research Working Paper No. 27623, 2020b.
W. F. Bassett and W. B. Marsh. Assessing Targeted Macroprudential Financial
Regulation: The Case of the 2006 Commercial Real Estate Guidance for Banks.
Journal of Financial Stability, 30:209–228, 2017.
A. N. Berger and G. F. Udell. Relationship Lending and Lines of Credit in Small
Firm Finance. Journal of Business, pages 351–381, 1995.
B. S. Bernanke. Nonmonetary Effects of the Financial Crisis in the Propagation of
the Great Depression. The American Economic Review, 73(3):257–276, 1983.
K. Blickle, C. Parlatore, and A. Saunders. Specialization in Banking. FRB of New
York Staff Report No. 967, 2021.
J. Boeckx, M. de Sola Perea, and G. Peersman. The transmission mechanism of
credit support policies in the euro area. European Economic Review, 124:103403,
2020.
41
N. Boyarchenko, A. Kovner, and O. Shachar. It’s What You Say and What You
Buy: A Holistic Evaluation of Corporate Credit Facilities. Federal Reserve Bank
of New York Working Paper, No. 935, 2020.
A. C. Cameron and P. K. Trivedi. Microeconometrics: Methods and Applications.
Cambridge university press, 2005.
S. Chib, E. Greenberg, and I. Jeliazkov. Estimation of semiparametric models in
the presence of endogeneity and sample selection. Journal of Computational and
Graphical Statistics, 18(2):321–348, 2009.
G. Chodorow-Reich. The Employment Effects of Credit Market Disruptions: Firm-
level Evidence from the 2008–9 Financial Crisis. The Quarterly Journal of Eco-
nomics, 129(1):1–59, 2014.
G. Chodorow-Reich, O. Darmouni, S. Luck, and M. Plosser. Bank Liquidity Provision
Across the Firm Size Distribution. Journal of Financial Economics, forthcoming,
2021.
F. Core and F. De Marco. Public guarantees for small businesses in italy during
covid-19. 2021.
J. G. Cragg and S. G. Donald. Testing Identifiability and Specification in Instru-
mental Variable Models. Econometric Theory, pages 222–240, 1993.
C. Doniger and B. S. Kay. Ten Days Late and Billions of Dollars Short: The Em-
ployment Effects of Delays in Paycheck Protection Program Financing. Working
Paper, 2021.
FDIC. FDIC Quarterly. Volume 15, Number 3, 2021.
K. Friend, H. Glenos, and J. B. Nichols. An Analysis of the Impact of the Commercial
Real Estate Concentration Guidance. Federal Reserve Board and Office of the
Comptroller of the Currency, 2013.
W. G. Gale. Federal Lending and the Market for Credit. Journal of Public Economics,
42:177–193, 1990.
D. Glancy. Bank Relationships and the Geography of PPP Lending. Avail-
able at SSRN: https: // papers. ssrn. com/ sol3/ papers. cfm? abstract_
id= 3839671 , 2021.
42
E. Greenberg. Introduction to Bayesian Econometrics. Cambridge University Press,
2012.
Y. Hayashi. Government Doubles Interest Rate on SBA Loans After Banks Complain.
Wall Street Journal, 2020. April 2.
A. Horvath, B. Kay, and C. Wix. The COVID-19 Shock and Consumer Credit:
Evidence from Credit Card Data. Board of Governors of the Federal Reserve
System, FEDS Working Paper, No. 2021-008, February, 2021.
G. Joaquim and F. Netto. Bank Incentives and the Impact of the Paycheck Protection
Program. Available at SSRN: https://ssrn.com/abstract=3704518, 2021.
S. F. Leung and S. Yu. On The Choice Between Sample Selection and Two-Part
Models. Journal of Econometrics, 72(1-2):197–229, 1996.
L. Li and P. E. Strahan. Who Supplies PPP Loans (and Does It Matter)? Banks,
Relationships, and the COVID Crisis. Journal of Financial and Quantitative Anal-
ysis, 56(7):2411–2438, 2021.
L. Li, P. Strahan, and S. Zhang. Banks as Lenders of First Resort: Evidence from the
COVID-19 Crisis. The Review of Corporate Finance Studies, forthcoming, 2020.
M. Li and J. L. Tobias. Bayesian Analysis of Treatment Effect Models. Bayesian
Inference in the Social Sciences, pages 63–90, 2014.
P. Li. Estimation of Sample Selection Models with Two Selection Mechanisms. Com-
putational Statistics & Data Analysis, 55(2):1099–1108, 2011.
J. A. Lopez and M. M. Spiegel. Small Business Lending Under the PPP and PPPLF
Programs. Federal Reserve Bank of San Francisco Working Paper 2021-10, April,
2021.
N. G. Mankiw. The Allocation of Credit and Financial Collapse. The Quarterly
Journal of Economics, 101(3):455–470, 1986.
C. Minoiu, R. Zarutskie, and A. Zlate. Motivating Banks to Lend? Credit Spillover
Effects of the Main Street Lending Program. Available at: https://ssrn.com/
abstract=3773242, 2021.
A. Ono, I. Uesugi, and Y. Yasuda. Are Lending Relationships Beneficial or Harmful
for Public Credit Guarantees? Evidence from Japan’s Emergency Credit Guaran-
tee Program. Journal of Financial Stability, 9(2):151–167, 2013.
43
D. Paravisini, V. Rappoport, and P. Schnabl. Specialization in Bank Lending: Ev-
idence from Exporting Firms. National Bureau of Economic Research Working
Paper No. 21800, 2015.
D. J. Poirier. Intermediate Statistics and Econometrics: A Comparative Approach.
Mit Press, 1995.
P. Sharma. Risk-Shifting, Regulation, and Government Assistance. Federal Reserve
Bank of Kansas City Working Paper, RWP 19-70, 2019.
J. E. Stiglitz and A. Weiss. Credit Rationing in Markets with Imperfect Information.
The American Economic Review, 71(3):393–410, 1981.
J. H. Stock and M. Yogo. Testing for Weak Instruments in Linear IV Regression.
Identification and Inference for Econometric Models: Essays in Honor of Thomas
Rothenberg, 80(4.2):1, 2005.
A. Vossmeyer. Sample Selection and Treatment Effect Estimation of Lender of Last
Resort Policies. Journal of Business & Economic Statistics, 34(2):197–212, 2016.
J. A. Wilcox and Y. Yasuda. Government Guarantees of Loans to Small Businesses:
Effects on Banks’ Risk-Taking and Non-Guaranteed Lending. Journal of Financial
Intermediation, 37:45–57, 2019.
A. Zellner. An Efficient Method of Estimating Seemingly Unrelated Regressions and
Tests for Aggregation Bias. Journal of the American Statistical Association, 57
(298):348–368, 1962.
44
ONLINE APPENDIX MATERIALS
45
A Key Paycheck Protection Program Dates
Table A.1 summarizes the key funding developments in the PPP program through
2020 and 2021. Round one funding appropriated by the CARES Act was $349 billion.
The program was scheduled to run from the earliest possible date following passage
of the act until June 30, 2020. The SBA began making loans just a few days after
CARES Act passage and the funding was quickly exhausted. By April 15, less than
three weeks after the CARES Act was signed, the SBA announced that the initial
funds were exhausted. In response, Congress approved an additional $321 billion
in appropriations to continue making loans though the program end date remained
June 15. During this time, government provided support via fiscal and monetary
agents began to stabilize the economic situation and financial markets. Consumers
and businesses also began to adapt to social distancing restrictions that allowed
economic activity to increase substantially from their early pandemic levels. Due
to this rise in economic activity and the stabilization of financial markets, demand
for PPP loans likely waned during the later part of the program. Thus, fund use
slowed and funds remained available as the original expiration date of the program
approached, spurring Congress to extend the program by several weeks in July 2020.
Finally, in late 2020, COVID cases again began to rise in the United States,
prompting concerns that economic activity would again decline. In response, Congress
appropriated an additional $284 billion in funding for a renewed PPP program for
the first quarter of 2021. The legislation also rescinded the remaining $146.5 billion
in unused funds from the program’s second round.
46
Table A.1: Key Paycheck Protection Program Dates
Enactment Date Appropriations End Date
CARES Act 03/27/2020 $349 billion 06/30/2020
PPP and Health Care Enhancement Act 04/24/2020 $321 billion 06/30/2020
S.4116 07/04/2020 – 08/08/2020
Consolidated Appropriations Act 12/21/2020 $284 billion 03/31/2021
Notes: Funds originally appropriated by the CARES Act were authorized for use until June 30, 2020. However,
funds were exhausted on April 15, 2020. Second round funds appropriated by the Paycheck Protection Program
and Health Care Enhancement Act were also to be used by June 30, 2020. This deadline was later extended by
S. 4116 to August 8, 2020. Third round funding was appropriated under the Consolidated Appropriations Act for
use through March 31, 2021. While allocating new funds, the act also recinded $146.5 billion in unsed funds from
round 2 and placed them into the Treasury General Account.
47
B Paycheck Protection Program Loan Terms
Table B.1 describes the PPP loan terms.
Table B.1: Paycheck Protection Program Loan Details
Category Details
Program Dates: Rounds 1-2: 2/15/20 - 8/8/20
Round 3: 01/11/20 - 3/15/20
Eligibility: Less than 500 U.S. employees
meets SBA’s small business concern definition or,
tax-exempt nonprofit org
operating before 2/15/20
Loan Amount: lesser of,
- 2.5 times avg monthly payroll costs up to $100k per employee
plus any oustanding EIDL loans
- $10 million
Maturity: 2 years if originated before 06/05/20
5 years otherwise
Covered expenses payroll costs:
- employee compensation
- employee leave payments
- health and retirement benefits costs
- state and local taxes assessed on compensation
mortgage interest and rent
utility payments
previously incurred interest on debt
Rate and Fees: 1 percent
No borrower paid fees
Payment: Deferral up to 10 months (originally 6 months)
Interest accrues
Forgiveness: Generally requires that 75% of fund use is attributed to payroll costs
Notes: Third round PPP appropriations made a number of changes to the original PPP program terms
including additional eligible expenses including property damage and certain worker protection costs. The
third round also allowed modifications of exisitng loan amounts as well as second draw loans. Second draw
loans were limited to firms with less than 300 employees that had same quarter, year-over-year income
reductions of 25 percent or more in 2019 and 2020.
Sources: SBA, Federal Register.
48
C PPP Liquidity Facility Loan Terms
Table C.1 describes the terms of the PPP Liquidity Facility (PPPLF) program as
well as the capital treatment on PPP loans and PPP loans pledged to the PPPLF.
Table C.1: Paycheck Protection Program Liquidity Facility Terms
Eligibility All DIs originating PPP Loans
Collateral Whole PPP loans
Maturity Equals maturity of the pledged PPP loan
Principal Equals principal amount of the pledged PPP loan
Rate 35 bps
Fees No Fees
Regulatory Capi- Risk weights on PPP loans equal 0%
tal Treatment Loans pledged to PPPLF excluded from leverage ratio assets
Sources: Federal Reserve Board.
49
D Estimation of the Bayesian Joint Model
This appendix presents the Markov Chain Monte Carlo (MCMC) algorithm used to
estimate the Bayesian Joint Model and the results from a simulation study.
To implement the estimation algorithm, we partition the full set of outcomes into
those that pertain to participants and non-participants, yi,p , and yi,np , respectively,
where, ∗
yi1 ∗
!
yi1
yi,p = yi2 , yi,np = . (8)
yi4
yi3
The marginal mean of each set of outcomes based on equations 1 - 4 is obtained from
the following expressions.
0
xi β1 + zi1 γ1 !
x0i β1 + zi1 γ1
µi,p = x0i β2 + zi2 γ2 , µi,np = . (9)
x0i β4
x0i β3 + yi2 δ
We consider the elements of the covariance matrix pertaining to participants and
non-participants separately and label them Ωp and Ωnp , respectively. Accordingly,
the two covariance matrices are defined as,
1 Ω12 Ω13 !
1 Ω14
Ωp = Ω21 Ω22 Ω23 , Ωnp = . (10)
Ω41 Ω44
Ω31 Ω32 Ω33
Subsequently, we rearrange the data in a Seemingly Unrelated Regressions setup
[Zellner, 1962]. The rearranged covariate matrices are,
x0 zi1 0 0 0 0 x0 zi1 0 0 0 0
i i
0 0 x0i zi2 0
0 , Xi,np = 0 0 0 0 0 0
Xi,p = .
0 0 0 0 x0 0 0 0 0 0 0 0
i
0 0 0 0 0 0 0 0 0 0 0 x0i
50
The outcomes are stacked into vectors Yi,p and Yi,np ,
∗ ∗
yi1 yi1
yi2
, Yi,np = 0 .
Yi,p =
y 0
i3
0 yi4
D.1 Markov Chain Monte Carlo Algorithm
The likelihood and priors we have specified generate conditional conjugacy. We
thereby develop the following Gibbs sampler to estimate the model.18
1. Sample Ω from Ω|θ, y, y1∗ in one block by partitioning into sub-matrices, where
θ = [β, γ1 , γ2 , δ]0 .
2. Sample θ from the distribution θ|Ω, y, y1∗ .
∗ ∗
3. Sample yi1 from yi1 |θ, y, Ω for i = 1, 2, ..., n.
The details underlying each step of the algorithm are discussed in the following
subsections.
D.1.1 Sampling Ω
We sample the elements in Ωp and Ωnp separately using the algorithm in Chib,
Greenberg, and Jeliazkov [2009], as applied in Vossmeyer [2016] and Sharma [2019].
The conditional distributions consist of inverse Wishart and matrix-variate normal
distributions.
18
The trace plots for the results in Section 3 and the simulation study are available upon request.
51
To specify the sampling steps, define 7,, 1, Rp, and Rp, as,
™ = (vi) — (x) 61 + Z1pV1) Y2— (x! G2 + 2272) Y3 — (xi, 6s + y2d)) ;
™np = (Vin a (xp G1 + zp) Ya — (xp) ’
Qiu Qi Q13
Rp = | Qar Q22 Qaz | + Mp;
Q31 Qs32 Q33
Qi Qi
Rip = + Mhaplnp-
Qa Qaa
Finally, define,
Qui = QW — MOAT! Qu,
Bu = 07 Qu.
Expressions for Ry. are analogous to the expression for Qy;. Using these elements,
we sample each term of (2 as follows.
ay
- O22.1)8,y, YE ~ TW (v + np, Rp,22.1)
2. By2|O,y, yf, Qo21 ~ N (Rpt Rpar, Ry 1222.1)
3. Define 2, = PO.
Qo, Qe
A. O33.u19, Y, Yp, Q33.u ~ LW (v + np, Rp33.u)
5. BuslO,y, yp ~ MN (R7* Rus, Q33.u ® Ru)
6. Q441|0,y, yt ~ ZW (v + np, Rnp22.1)
7. BralO,y, Yi, Qaat ON (Rip Rnp.21s Rip 244.1)
52
D.1.2 Sampling 6
We sample the elements of @ in one step by stacking the outcomes and covariates in
Equations 1-4 in a SUR setup as described above. The conditional distribution of
theta is multivariate normal, (6,7), where
p>
l
T (Ty 100 + X'isp (Inp ® %!) Yip +X inp (Tnnp ® Qn) Yinp)
(Ty + Xap (In, @ MH") Kip +X inp Tony @ Qnp) Kisnp) +
>
I
D.1.3. Sampling yj
We sample the latent variables y;, for i = 1,2,...,n from a truncated normal distri-
bution whose bounds are (—oo,0) for non-participants and (0,00) for participants.
Accordingly, y4|0,y,Q2 ~ TN (~co,0) (Hinp|\1sQnp)\1) for i € Nnp and y|0,y,Q ~
TN (0,00) (Mip|\1; Qp)\1) for i € Np. The parameters in the conditional distributions of
y;,|0, y,® are the standard conditional moments from a Normal distribution where
the conditioning is on all except for the first element in the vectors y;,p and Yj np.
53
D.2 Simulation Study
Table D.1 presents the results of the simulation study. We set the following priors
under the two specifications: θ ∼ N (0, 10 × I), Ωp ∼ IW(7, 3 × I4 ), and Ωnp ∼
IW(7, 3 × I3 ) where θ = [γ1 , γ2 , δ, β], and β = {β1 , β2 , β3 , β4 }.
Table D.1: Simulation Results
No exclusion Exclusion
True values 95% credibility interval True values 95% credibility interval
β11 -0.1 [-0.23, 0.08] -0.1 [-0.16, -0.06]
β12 -0.2 [-0.37, -0.14] -0.2 [-0.21, -0.1]
β13 0.1 [-0.07, 0.15] 0.1 [0.08, 0.13]
β14 0.2 [0.06, 0.27] 0.2 [0.16, 0.22]
β21 1 [0.8, 1.95] 1 [0.88, 1.12]
β22 0.5 [0.43, 0.72] 0.5 [0.47, 0.51]
β23 -0.6 [-0.67, -0.43] -0.6 [-0.62, -0.57]
β24 -1 [-1.13, -0.88] -1 [-1.03, -0.96]
β31 2 [1.37, 2.58] 2 [1.89, 2.12]
β32 -3 [-3.21, -2.66] -3 [-3.05, -2.99]
β33 2.5 [2.31, 2.69] 2.5 [2.46, 2.52]
β34 4 [3.77, 4.29] 4 [3.94, 4.03]
β41 -2 [-2.58, -1.67] -2 [-2.37, -1.66]
β42 1.5 [1.42, 1.65] 2 [1.94, 2.02]
β43 -3 [-3.09, -2.85] -3 [-3.08, -2.95]
Ω12 0.5 [-0.69, 0.6] 0.5 [0.33, 0.57]
Ω22 0.8 [0.57, 1.06] 0.8 [0.69, 0.86]
Ω13 0.5 [-0.34, 1.08] 0.5 [0.45, 0.67]
Ω23 -0.1 [-0.82, -0.12] -0.1 [-0.14, -0.04]
Ω33 0.75 [0.7, 1.53] 0.75 [0.69, 0.87]
Ω14 -0.2 [-0.82, 0.5] -0.2 [-0.72, 0.3]
Ω44 0.8 [0.74, 1.28] 0.8 [0.77, 1.11]
Note: The 95% credibility intervals in brackets. The results are based on 11,000 MCMC draws
with a burn-in of 1000. The specification of “Exclusion” consists of an instrument in the selection
equation. The specification of “No exclusion” consists of no instruments in the selection equation.
54
D.3 Additional Results for Participation and Intensity
See Table D.2.
55
Table D.2: Results for participation and intensity from the Bayesian joint model
C&I Gwth Non-PPP C&I Gwth CRE Gwth
(1) (2) (3) (4) (5) (6)
Part. Intensity Part. Intensity Part. Intensity
Tech exp. to assets -0.17 -0.09 0.02
[-0.26, -0.07] [-0.17, -0.01] [-0.07, 0.11]
COVID-affected employment share 0.08 0.03 0.03
[0.06, 0.1] [0.02, 0.04] [0.02, 0.04]
ln Assets 0.14 0.81 0.15 1.12 0.19 1.25
[0.12, 0.16] [0.68, 0.95] [0.13, 0.17] [0.99, 1.24] [0.17, 0.21] [1.12, 1.38]
CI to assets -0.02 0.38 0.04 0.31 0.03 0.30
[-0.03, -0.02] [0.36, 0.41] [0.03, 0.04] [0.28, 0.33] [0.03, 0.04] [0.27, 0.32]
Leverage Ratio -0.02 -0.26 -0.04 -0.30 -0.04 -0.30
[-0.03, -0.01] [-0.32, -0.21] [-0.05, -0.03] [-0.36, -0.25] [-0.05, -0.03] [-0.36, -0.25]
Liquid Assets to Assets 0.00 0.09 0.01 0.07 0.01 0.07
[0, 0] [0.08, 0.1] [0, 0.01] [0.06, 0.09] [0, 0.01] [0.06, 0.09]
56 ALLL to Total Loans 0.00 0.45 -0.03 0.09 -0.02 0.12
[-0.03, 0.04] [0.19, 0.7] [-0.07, 0.01] [-0.15, 0.33] [-0.06, 0.03] [-0.12, 0.37]
ROA 0.07 0.10 0.08 0.28 0.07 0.25
[0.03, 0.12] [-0.16, 0.37] [0.03, 0.12] [0.02, 0.54] [0.03, 0.12] [-0.01, 0.52]
Cases Per 100k 0.03 0.11 0.01 0.09 0.02 0.11
[0, 0.06] [-0.06, 0.28] [-0.02, 0.04] [-0.08, 0.27] [-0.02, 0.05] [-0.07, 0.28]
Constant -0.41 -6.75 -0.82 -8.38 -1.25 -10.01
[-0.67, -0.15] [-8.59, -4.91] [-1.07, -0.56] [-10.07, -6.67] [-1.54, -0.96] [-11.77, -8.24]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based
on 55,000 MCMC draws with a burn-in of 5000.
Table D.3: Profitability and Loan Growth Outcomes at Participant Banks
∆NIM(bps) CI Gwth(%) Non-PPP CI Gwth(%) CRE Gwth(%)
(1) (2) (3) (4)
PPP Loans to Total Loans -4.27 10.52 -0.46 0.23
[-6.03, -2.7] [9.26, 11.87] [-1.46, 0.57] [-0.54, 1.01]
ln Assets 3.98 6.20 0.13 0.36
[2.72, 5.33] [5.18, 7.19] [-0.82, 1.04] [-0.48, 1.2]
C&I to assets 0.42 -7.18 0.22 0.10
[-0.09, 0.97] [-7.77, -6.63] [-0.1, 0.54] [-0.14, 0.34]
Leverage Ratio -2.48 -0.23 -0.09 0.28
[-3.33, -1.71] [-0.94, 0.49] [-0.5, 0.34] [-0.04, 0.59]
57 Liquid Assets to Assets -0.05 -0.80 0.07 -0.03
[-0.2, 0.1] [-0.99, -0.61] [-0.02, 0.15] [-0.1, 0.03]
ALLL to Total Loans -3.29 -3.67 -2.72 -0.53
[-5.28, -1.32] [-6.2, -1.15] [-3.65, -1.78] [-1.23, 0.18]
ROA -10.48 1.38 -2.12 -1.84
[-12.56, -8.3] [-1.24, 4.03] [-3.14, -1.09] [-2.61, -1.08]
Cases Per 100k -8.42 2.79 -0.07 0.02
[-9.76, -7.05] [0.98, 4.6] [-0.7, 0.56] [-0.45, 0.5]
Constant 0.03 -0.64 2.55 0.03
[-5.94, 6.06] [-6.61, 5.35] [-3.27, 8.41] [-5.89, 5.91]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in
brackets. The results are based on 55,000 MCMC draws with a burn-in of 5000.
Table D.4: Profitability and Loan Growth Outcomes at Non-Participant Banks
∆NIM(bps) CI Gwth(%) Non-PPP CI Gwth(%) CRE Gwth(%)
(1) (2) (3) (4)
ln Assets -0.91 -7.47 -6.64 -3.72
[-4.99, 3.55] [-8.46, -6.49] [-7.57, -5.72] [-4.41, -3.02]
C&I to assets -0.10 1.42 -1.86 -1.07
[-1.13, 1.04] [1.06, 1.79] [-2.17, -1.56] [-1.28, -0.87]
Leverage Ratio -0.26 0.84 1.98 1.06
[-1.54, 0.92] [0.21, 1.46] [1.36, 2.61] [0.67, 1.45]
Liquid Assets to Assets -0.41 -0.06 -0.47 -0.24
[-0.68, -0.12] [-0.23, 0.11] [-0.64, -0.3] [-0.36, -0.13]
ALLL to Total Loans -9.60 -2.06 0.02 -0.93
58 [-12.19, -6.98] [-4.6, 0.44] [-2.42, 2.44] [-2.55, 0.68]
ROA -1.54 -0.03 -0.39 0.27
[-5.5, 2.45] [-3.16, 3.11] [-3.48, 2.74] [-1.87, 2.4]
Cases Per 100k -3.77 -2.46 -1.74 -0.43
[-6.83, -0.69] [-4.75, -0.21] [-4.02, 0.54] [-2.03, 1.13]
Constant -0.60 -1.27 0.68 -2.00
[-6.7, 5.48] [-7.35, 4.8] [-5.32, 6.71] [-7.84, 3.87]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in
brackets. The results are based on 55,000 MCMC draws with a burn-in of 5000.
E Covariances from the Bayesian Joint Model
Estimated covariances from the Bayesian Joint Model are presented in Table E.1.
59
Table E.1: Covariance estimates from the Bayesian joint model
∆NIM C&I Non-PPP CRE
Gwth C&I Gwth Gwth
COV(participation, intensity) 6.87 3.72 6.89 6.89
[6.74, 6.99] [3.38, 4.04] [6.77, 7.01] [6.77, 7.01]
COV(participation, bank outcome) 17.50 64.72 3.35 -0.09
[6.48, 29.58] [59.16, 69.6] [-3.83, 10.19] [-5.53, 5.22]
COV(intensity, bank outcome) 143.31 63.71 18.49 -2.11
[67.09, 228.09] [6.64, 116.31] [-31.82, 66.78] [-40.18, 35.36]
COV(non-participation, bank outcome) -0.13 -61.53 -62.38 -36.18
[-33.01, 35.67] [-65.52, -57.74] [-66.37, -58.52] [-39.21, -33.22]
60
E.1 Interpreting Marginal and Conditional Estimates
To interpret the differences in the estimated coefficients of C&I loans to assets across
columns (4) and (6) in Table 5 and 6, we must consider the implications of using
a joint modeling structure represented in Equations 1-4. First, estimates for each
equation represent moments from the marginal distribution, after marginalizing out
the remaining outcomes from the remaining equations. Second, covariances between
selection, and outcomes for participants and non-participants introduce dependence
between estimates for the two groups. Below, we consider conditional estimates in-
stead of marginal estimates for non-participants, to perform a more direct comparison
across specifications in columns (4) and (6).
Consider the outcomes, mean and covariances for non-participants represented
in Equations 8, 9, and 10. The joint model for the two outcomes pertaining to
non-participants is represented by,
! ! !!
∗
yi1 x0i β1 + zi1 γ1 1 Ω14
∼N , . (11)
yi4 x0i β4 Ω41 Ω44
This joint Gaussian distribution results in the following expression for the mean of
∗
bank outcomes yi4 conditional on non-participation[Poirier, 1995].
∗
E [yi4 |yi1 , β, Ω] = x0i (β4 − Ω14 β1 ) + Ω14 yi1
∗
− Ω14 zi1 γ1 . (12)
We obtain the posterior conditional mean of the coefficient of C&I loans to assets and
95 percent credible intervals by using the estimates of β1 , β4 , and Ω14 from Tables 2,
5, 6, and E.1 in Equation 12.
The conditional moments of the coefficient of C&I assets are qualitatively similar
when the outcome is C&I loans or C&I loans excluding PPP. C&I loans increase by
0.25 percentage points for a percent increase in C&I concentration. The 95 percent
probability intervals for this estimate are -0.10 and 0.62 percentage points. Under
the specification with C&I loans outside the PPP, the conditional mean estimate is
0.49 percentage points with a probability interval of 0.19 and 0.79. Therefore, the
61
F
−80 −70 −60 −50 −40 −30 −20 −10 0 10
Clothing and clothing accessories stores
This appendix presents the sorted declines in employment by NAICS sector bet ween
Scenic and sightseeing transportation
Personal and laundry services
Amusements, gambling, and recreation
Motion picture and sound recording industries
Furniture and home furnishings stores
Food services and drinking places
Accommodation
Sporting goods, hobby, book, and music stores
Performing arts and spectator sports
Transit and ground passenger transportation
Miscellaneous store retailers
Apparel
Textile product mills
Museums, historical sites, and similar institutions Jan 2020−Apr 2020
Transportation equipment
Furniture and related products
Textile mills
Rental and leasing services
Printing and related support activities
Motor vehicle and parts dealers
Ambulatory health care services
Repair and maintenance
January and April 2020. These sectors are used to determine pre-pandemic county
Administrative and support services
Chart shows percent change in employment over Jan - Apr 2020 across industries.
Miscellaneous durable goods manufacturing
seemingly large difference between the marginal estimates of 1.4 and -1.8 percentage
Social assistance
Couriers and messengers
Air transportation
Support activities for mining
Specialty trade contractors
Support activities for transportation
Construction of buildings
Health and personal care stores
Electronics and appliance stores
Plastics and rubber products
Machinery
Membership associations and organizations
Nonmetallic mineral products
Primary metals
Wood products
62
General merchandise stores
Fabricated metal products
Electronic markets and agents and brokers
Nondurable goods
Water transportation
level exposures to COVID as-of 2019:Q4. Bank-market specific COVID exposures
Food manufacturing
Durable goods
Nonstore retailers
Broadcasting, except Internet
Electrical equipment and appliances
Warehousing and storage
Truck transportation
Real estate
Gasoline stations
Mining, except oil and gas
Petroleum and coal products
Nursing and residential care facilities
Waste management and remediation services
Oil and gas extraction
Figure F.1: Change in employment
Rail transportation
Publishing industries, except Internet
Hospitals
Paper and paper products
Chemicals
Food and beverage stores
Telecommunications
Computer and electronic products
are assembled by weighting county exposures by bank deposits. The methodology is
Credit intermediation and related activities
Heavy and civil engineering construction
Pipeline transportation
Lessors of nonfinancial intangible assets
Data processing, hosting and related services
Source: CES data from the Bureau of Labor Statistics.
Insurance carriers and related activities
Other information services
Monetary authorities − central bank
Building material and garden supply stores
Percent
0
10
−80 −70 −60 −50 −40 −30 −20 −10
Categorization of COVID-sensitive industries
points are resolved upon evaluating their corresponding conditional estimates.
taken from Boyarchenko, Kovner, and Shachar [2020].
G Quarterly Results from the Bayesian Model
Tables G.1 and G.2 provide results for bank outcomes in the quarters 2020:Q2 and
2020:Q3, respectively. Columns (1), (3), (5), and (7) report the results for partici-
pation in the program. The results for each quarter are qualitatively similar to the
combined results. In particular, larger, and more liquid banks were more likely to
participate while more capitalized banks were less likely to participate across both
quarters. In Q2 2020, when the first round of the PPP was in operation, more prof-
itable banks were more likely to participate. This result continued to hold in Q3
2020, but the estimated effects were statistically weaker relatively to the previous
quarter.
63
Table G.1: Results for participation and intensity from the Bayesian joint model in Q2 2020
∆NIM C&I Gwth Non-PPP C&I Gwth CRE Gwth
(1) (2) (3) (4) (5) (6) (7) (8)
Part. Intensity Part. Intensity Part. Intensity Part. Intensity
Tech exp. to assets -0.128 -0.204 -0.145 -0.032
[-0.31, 0.05] [-0.33, -0.07] [-0.27, -0.02] [-0.18, 0.1]
COVID-affected employment share 0.039 0.086 0.04 0.039
[0.02, 0.06] [0.06, 0.11] [0.02, 0.06] [0.02, 0.06]
ln Assets 0.178 1.213 0.122 0.786 0.141 1.054 0.191 1.222
[0.14, 0.22] [1.04, 1.39] [0.09, 0.15] [0.6, 0.97] [0.11, 0.17] [0.87, 1.23] [0.16, 0.22] [1.04, 1.4]
CI to assets 0.033 0.294 -0.024 0.368 0.034 0.297 0.032 0.289
[0.03, 0.04] [0.26, 0.33] [-0.03, -0.02] [0.33, 0.4] [0.03, 0.04] [0.26, 0.33] [0.03, 0.04] [0.26, 0.32]
Leverage Ratio -0.046 -0.327 -0.025 -0.292 -0.042 -0.328 -0.044 -0.324
[-0.06, -0.03] [-0.4, -0.25] [-0.04, -0.01] [-0.37, -0.21] [-0.06, -0.03] [-0.41, -0.25] [-0.06, -0.03] [-0.4, -0.25]
Liquid Assets to Assets 0.006 0.074 -0.001 0.09 0.006 0.07 0.006 0.071
[0, 0.01] [0.05, 0.09] [0, 0] [0.07, 0.11] [0, 0.01] [0.05, 0.09] [0, 0.01] [0.05, 0.09]
ALLL to Total Loans -0.026 0.136 0.007 0.389 -0.038 0.051 -0.03 0.046
[-0.09, 0.04] [-0.21, 0.48] [-0.04, 0.06] [0.04, 0.74] [-0.09, 0.01] [-0.28, 0.38] [-0.09, 0.03] [-0.3, 0.39]
ROA 0.116 0.562 0.092 0.277 0.093 0.483 0.1 0.478
[0.04, 0.19] [0.19, 0.93] [0.03, 0.16] [-0.09, 0.65] [0.03, 0.16] [0.12, 0.85] [0.03, 0.17] [0.1, 0.85]
64 Cases Per 100k 0.026 0.034 0.042 -0.095 -0.004 -0.148 -0.015 -0.132
[-0.07, 0.13] [-0.47, 0.54] [-0.03, 0.12] [-0.6, 0.41] [-0.09, 0.08] [-0.64, 0.35] [-0.11, 0.08] [-0.65, 0.38]
Constant -1.096 -9.708 -0.159 -6.048 -0.624 -7.399 -1.233 -9.483
[-1.59, -0.6] [-12.01, -7.47] [-0.54, 0.21] [-8.51, -3.57] [-1.01, -0.23] [-9.8, -4.96] [-1.66, -0.81] [-11.91, -7.06]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC draws with
a burn-in of 5000.
Table G.2: Results for participation and intensity from the Bayesian joint model in Q3 2020
∆NIM C&I Gwth Non-PPP C&I Gwth CRE Gwth
(1) (2) (3) (4) (5) (6) (7) (8)
Part. Intensity Part. Intensity Part. Intensity Part. Intensity
Tech exp. to assets -0.108 -0.143 -0.083 0.048
[-0.29, 0.07] [-0.3, 0.01] [-0.26, 0.07] [-0.11, 0.2]
COVID-affected employment share 0.035 0.081 0.061 0.045
[0.02, 0.05] [0.06, 0.11] [0.01, 0.13] [0.02, 0.11]
ln Assets 0.161 1.106 0.147 0.739 0.173 0.84 0.188 1.074
[0.12, 0.2] [0.91, 1.3] [0.12, 0.18] [0.55, 0.93] [0.13, 0.23] [0.34, 1.24] [0.15, 0.24] [0.45, 1.33]
CI to assets 0.034 0.317 -0.02 0.398 0.018 0.346 0.029 0.321
[0.03, 0.04] [0.28, 0.35] [-0.03, -0.01] [0.36, 0.43] [-0.01, 0.04] [0.29, 0.41] [0, 0.04] [0.28, 0.4]
Leverage Ratio -0.043 -0.315 -0.02 -0.249 -0.036 -0.249 -0.041 -0.289
[-0.06, -0.03] [-0.4, -0.23] [-0.03, -0.01] [-0.33, -0.17] [-0.05, -0.02] [-0.36, -0.12] [-0.05, -0.03] [-0.38, -0.14]
Liquid Assets to Assets 0.006 0.073 -0.001 0.089 0.001 0.084 0.005 0.076
[0, 0.01] [0.05, 0.09] [0, 0] [0.07, 0.11] [-0.01, 0.01] [0.06, 0.12] [-0.01, 0.01] [0.05, 0.11]
ALLL to Total Loans -0.021 0.163 0.001 0.455 -0.022 0.267 -0.007 0.23
[-0.08, 0.04] [-0.2, 0.53] [-0.05, 0.05] [0.1, 0.81] [-0.08, 0.04] [-0.15, 0.72] [-0.07, 0.06] [-0.15, 0.64]
ROA 0.056 0.009 0.059 -0.058 0.08 -0.083 0.057 -0.007
[-0.02, 0.13] [-0.4, 0.41] [0, 0.12] [-0.42, 0.31] [0.01, 0.16] [-0.66, 0.43] [-0.01, 0.14] [-0.54, 0.41]
65 Cases Per 100k 0.022 0.132 0.03 0.144 0.015 0.148 0.01 0.142
[-0.02, 0.07] [-0.11, 0.37] [-0.01, 0.07] [-0.08, 0.37] [-0.03, 0.06] [-0.09, 0.39] [-0.04, 0.06] [-0.1, 0.38]
Constant -0.88 -8.062 -0.529 -5.927 -0.81 -6.05 -1.157 -8.159
[-1.39, -0.37] [-10.7, -5.4] [-0.92, -0.15] [-8.46, -3.39] [-1.21, -0.39] [-10.18, -1.04] [-1.59, -0.69] [-11.22, -2.52]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC draws with
a burn-in of 5000.
Columns (2), (4), (6), and (8) report the results for PPP lending intensity. The
results remain qualitatively similar with banks facing more C&I exposure typically
making more PPP loans. Large, more liquid and riskier banks– as measured by
leverage capital ratios–also participated more intensively across quarters. In Q2
2020, more profitable banks participated more intensively in the PPP, but, as in the
case of participation, this relationship was weaker in Q3 2020.
Tables G.3 and G.4 report results for Q2 and Q3 2020 respectively, under the
specifications presented in Tables 5 and 6. Columns (1), (3), (5), and (7) report
the results for participants. The results remain qualitatively similar with the results
combined across quarters. The estimates in the first row show that incremental
participation in the PPP diluted bank profitability. These results do not persist into
Q3. Banks experienced a larger decline in profitability during the first round when
firms rushed to obtain PPP funding and when banks processed large volumes of
applications. By the second round, profit margins were likely cushioned by fees and
interest accrued from the first round as well as the smaller size of loans relative to the
first round on account of larger firms gaining early access, and small firms gaining
access subsequently [Balyuk et al., 2021].19 Banks that participated more intensively
in the PPP experienced substantial growth in their overall C&I loan portfolio and
weaker growth in non-PPP C&I loans. Finally, incremental participation in the
PPP did not result in statistically important effects on risk-taking in either quarter.
The results across the remaining control variables are consistent with the combined
results across quarters, with one exception. Banks that were concentrated to a greater
extent in C&I loans experienced a statistically important increases in NIM relative
to 2019 in Q2 2020. This relationship reversed in Q3 2020, when banks with larger
concentrations in C&I loans underwent statistically important declines in the change
in NIM. This finding suggests that banks with a focus on C&I lending experienced a
larger decline in NIM relative to 2019 during the second round of the PPP, at a time
when lending was likely more targeted to firms that were affected by the pandemic
than during the first round.
19
The sliding scale in fees resulted in a larger percent of loan amount paid out as fees for small
loans compared to large loans. Details of fee structure available here.
66
Table G.3: Results for profitability and loan growth of participating and non-participating banks in Q2 2020
∆NIM(bps) CI Gwth(%) Non-PPP CI Gwth(%) CRE Gwth(%)
(1) (2) (3) (4) (5) (6) (7) (8)
(part.) (non-part.) (part.) (non-part.) (part.) (non-part.) (part.) (non-part.)
PPP Loans to Total Loans -6.914 10.717 0.363 0.202
[-9.15, -4.92] [8.65, 12.92] [-0.89, 1.71] [-0.71, 1.09]
ln Assets 5.278 -1.501 5.898 -6.707 -0.384 -6.081 0.444 -3.728
[3.73, 7.02] [-6.44, 3.69] [4.38, 7.35] [-7.93, -5.51] [-1.5, 0.67] [-7.28, -4.93] [-0.45, 1.34] [-4.63, -2.86]
CI to assets 1.446 -0.284 -7.209 1.308 0.009 -1.618 0.08 -1.032
[0.78, 2.19] [-1.54, 1] [-8.09, -6.35] [0.82, 1.81] [-0.41, 0.39] [-2.05, -1.21] [-0.19, 0.36] [-1.32, -0.76]
Leverage Ratio -3.736 0.249 -0.114 0.656 0.218 1.667 0.34 0.939
[-4.95, -2.65] [-1.23, 1.64] [-1.2, 1] [-0.12, 1.43] [-0.34, 0.83] [0.84, 2.49] [-0.07, 0.75] [0.42, 1.47]
Liquid Assets to Assets 0.267 -0.45 -0.759 -0.106 0.021 -0.446 -0.052 -0.19
[0.05, 0.51] [-0.8, -0.11] [-1.04, -0.48] [-0.33, 0.12] [-0.09, 0.13] [-0.68, -0.21] [-0.13, 0.03] [-0.34, -0.04]
ALLL to Total Loans -2.992 -8.652 -3.881 -1.711 -2.714 0.518 -0.259 -0.511
[-6.1, 0.14] [-11.93, -5.28] [-7.17, -0.59] [-4.8, 1.33] [-4.01, -1.39] [-2.56, 3.59] [-1.26, 0.73] [-2.69, 1.65]
ROA -6.035 -1.985 1.521 -0.204 -1.899 -0.666 -2.229 0.236
[-9.43, -2.48] [-6.75, 2.78] [-1.99, 5.04] [-4.02, 3.6] [-3.43, -0.43] [-4.47, 3.15] [-3.36, -1.1] [-2.55, 3.03]
Cases Per 100k -3.176 -2.896 5.757 -2.377 0.059 -1.562 -0.088 -1.184
67 [-7.16, 0.79] [-8.26, 2.52] [1.59, 9.9] [-7.11, 2.3] [-1.77, 1.91] [-6.37, 3.28] [-1.48, 1.29] [-5.23, 2.75]
Constant 1.904 -0.577 -0.137 -0.608 1.95 0.32 -0.878 -1.476
[-4.25, 8.07] [-6.74, 5.55] [-6.18, 5.89] [-6.69, 5.48] [-3.9, 7.75] [-5.77, 6.45] [-6.72, 5.03] [-7.51, 4.56]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC
draws with a burn-in of 5000.
Table G.4: Results for profitability and loan growth of participating and non-participating banks in Q3 2020
∆NIM(bps) CI Gwth(%) Non-PPP CI Gwth(%) CRE Gwth(%)
(1) (2) (3) (4) (5) (6) (7) (8)
(part.) (non-part.) (part.) (non-part.) (part.) (non-part.) (part.) (non-part.)
PPP Loans to Total Loans -0.185 9.529 -0.331 0.41
[-2.54, 2.39] [7.18, 12.04] [-2.33, 1.54] [-0.76, 1.61]
ln Assets 0.751 -1.128 6.672 -7.459 0.35 -8.115 0.279 -3.887
[-1.12, 2.46] [-5.33, 3.93] [5.17, 8.17] [-8.82, -6.14] [-0.63, 1.58] [-11.48, -5.49] [-0.67, 1.24] [-5.94, -2.78]
CI to assets -1.138 -0.213 -6.826 1.198 0.149 -0.91 0.058 -0.856
[-1.98, -0.37] [-1.32, 1.13] [-7.89, -5.81] [0.69, 1.71] [-0.52, 0.81] [-2.29, 0.89] [-0.42, 0.45] [-1.28, 0.2]
Leverage Ratio -0.781 -0.968 -0.585 0.769 -0.274 1.785 0.204 1.029
[-1.83, 0.33] [-2.41, 0.37] [-1.68, 0.5] [-0.15, 1.68] [-0.82, 0.22] [0.77, 2.78] [-0.23, 0.65] [0.46, 1.61]
Liquid Assets to Assets -0.465 -0.456 -0.764 -0.041 0.007 -0.103 -0.039 -0.199
[-0.67, -0.27] [-0.81, -0.09] [-1.06, -0.47] [-0.29, 0.21] [-0.22, 0.21] [-0.62, 0.55] [-0.21, 0.07] [-0.4, 0.21]
ALLL to Total Loans -3.169 -7.649 -2.558 -2.166 -2.738 -0.768 -0.886 -1.282
[-5.46, -0.89] [-11.17, -4.14] [-5.8, 0.75] [-5.52, 1.12] [-4.23, -1.27] [-4.36, 2.67] [-2.1, 0.2] [-3.88, 1.06]
ROA -12.066 -0.988 0.924 0.134 -1.768 -0.793 -1.232 -0.253
[-14.44, -9.69] [-5.66, 3.73] [-2.41, 4.28] [-3.85, 4.08] [-3.76, 0.44] [-5.68, 3.72] [-2.44, 0.59] [-3.5, 2.79]
Cases Per 100k -1.685 -1.322 2.678 -1.649 -0.038 -1.16 -0.023 -0.729
68 [-3.25, -0.15] [-5.08, 2.45] [0.33, 5.06] [-4.51, 1.18] [-1.02, 0.94] [-4.26, 1.88] [-0.71, 0.65] [-2.85, 1.44]
Constant -1.364 -0.157 -0.253 -1.046 0.337 -0.453 0.286 -1.328
[-7.45, 4.78] [-6.37, 5.99] [-6.31, 5.83] [-7.19, 5.02] [-5.81, 6.75] [-6.58, 5.69] [-7.13, 6.77] [-7.35, 4.68]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC
draws with a burn-in of 5000.
Columns (2), (4), (6), and (8) report the coefficient estimates for non-participants.
The results continue to be consistent with overall results in Tables 5 and 6. Larger
non-participants underwent declines in profitability as well as a decline in C&I and
CRE loan growth. The effects of other controls are also consistent with overall results
for both quarters. Tables G.5 and G.6 report the estimated covariances in Q2 and
Q3 2020 respectively. These results are qualitatively similar to the overall results
across the two quarters reported in Table E.1. The decision to participate, and the
intensity of participation are positively correlated across both quarters. Even though
participation, and intensity of participation are positively related to profitability as
measured by the change in NIM in the overall sample, these relationships become
negative in Q3 2020. This finding likely points to PPP lending that was less oppor-
tunistic, and more conservative in Q3 than in Q2 2020. The relationship between
participation intensity and C&I loan growth continues to be positive and statisti-
cally important. PPP participation intensity was weakly negatively associated with
growth in non-PPP C&I growth in Q2 2020. This relationship became statistically
important in Q3 2020, suggesting that banks that participated more intensively in
the second round cut back lending outside of the program. The decision to not par-
ticipate in the PPP is negatively related to bank profitability and loan growth, or is
only weakly positive across the two quarters.
69
Table G.5: Covariance estimates from the Bayesian joint model based on Q2 2020 outcomes
∆NIM C&I Non-PPP CRE
Gwth C&I Gwth Gwth
COV(participation, intensity) 6.70 3.56 6.76 6.75
[6.51, 6.88] [3.12, 4.01] [6.59, 6.93] [6.58, 6.92]
COV(participation, bank outcome) 34.58 64.82 -2.43 -0.09
[20.65, 50.25] [56.31, 72.31] [-11.6, 6.14] [-6.13, 6.08]
COV(intensity, bank outcome) 276.33 46.73 -23.83 -2.26
[182.62, 382.21] [-42.5, 132.84] [-87.54, 34.95] [-43.89, 40.44]
COV(non-participation, bank outcome) -3.11 -56.67 -58.07 -34.66
[-43.06, 38.25] [-61.98, -51.63] [-63.74, -52.74] [-38.85, -30.79]
70
Table G.6: Covariance estimates from the Bayesian joint model based on Q3 2020 outcomes
∆NIM C&I Non-PPP CRE
Gwth C&I Gwth Gwth
COV(participation, intensity) 6.92 3.72 3.99 5.98
[6.75, 7.1] [3.31, 4.13] [-0.73, 7.08] [-0.5, 7.09]
COV(participation, bank outcome) -12.46 67.92 14.16 1.67
[-30.2, 3.92] [57.82, 77.12] [-0.06, 23.38] [-7.27, 15.76]
COV(intensity, bank outcome) -71.08 112.42 17.75 -8.82
[-198.82, 44.93] [4.58, 214.17] [-65.61, 114.26] [-61.05, 47.27]
COV(non-participation, bank outcome) -2.99 -61.76 -63.08 -36.19
[-35.84, 38.18] [-67.6, -56.46] [-69.02, -57.32] [-40.38, -32.39]
71
H Decomposition of Counterfactual Results
We examine the drivers of the predicted counterfactuals for participants in Table
(g)
H.1. To this end, we evaluate x̄p,j β4j , g = 1, 2, ..., 50, 000, which is the product of
the mean value of each covariate j across participants and the posterior draws of the
associated coefficient from equation 4 for non-participants. The table reports the
mean and 95 percent credibility interval of this product across the 50,000 posterior
draws.
Bank size is the primary determinant of the lower counterfactual change in NIM
and loan growth for participants. Participants with asset size at the mean of the
group would have undergone a decline in NIM of 32 basis points relative to 2019, and
a reduction of the C&I loan portfolio by 91 percent had they not participated in the
PPP. These findings are driven by the negative coefficient on bank size among non-
participants. Small, non-participant banks continued to lend C&I and CRE loans
over the course of the pandemic while large non-participants curtailed such lending as
depicted in Tables 5 and 6. Large non-participant banks underwent greater declines
in NIM relative to small non-participants. Accordingly, if participants had used
the decision rules of non-participants, they would have largely cut back lending and
earned lower profits as they were, on average, larger than non-participants.
The second most important factor driving the predicted decline in counterfactual
profitability and loan growth is the ratio of loan loss allowances to total loans. Par-
ticipating banks with average levels of this ratio were likely to undergo a 13 basis
point decline in change in NIM, and a decline of nearly 3 percentage points to C&I
loan growth. Loan loss allowances, therefore, served as a constraint on loan growth
and earnings. Moreover, as banks set aside larger reserves when they expect bigger
losses, large loan loss allowances are suggestive of riskier loan portfolios. Our results
thereby reveal that banks exposed to potential losses would have likely withheld
lending on account of risk-aversion, which in the aggregate, would have contributed
to a credit crunch.
Unlike bank size and the ratio of loss allowances to total loans, other characteris-
tics do not predict as large a decline in margins and lending, but reflect the realities
72
of increased risk taking in a crisis. For example, better capitalized banks, which
have greater space to absorb additional losses, would have increased their lending by
more than their less capitalized peers. Banks with greater exposure to C&I loans,
and more liquid banks would have increased C&I lending, but curtailed CRE loan
growth. And finally, we find that COVID exposure, and thus greater loan risk, would
have further reduced C&I and CRE loan growth. Interestingly, previous profitability
was not statistically important in determining counterfactual outcomes, suggesting
that the precarious outlook for borrowers would have likely outweighed banks’ ability
to generate pre-pandemic earnings in driving lending declines. These results are all
consistent with elevated risk-aversion and the importance of sizable capital buffers
in a crisis.
73
Table H.1: Decomposition of predicted counterfactual outcomes for participants
∆NIM C&I Gwth Non-PPP C&I Gwth CRE Gwth
(1) (2) (3) (4)
ln Assets -32.69 -91.32 -127.43 -44.71
[-68.92, 10.89] [-103.66, -79.68] [-141.69, -113.81] [-53.63, -36.27]
CI to assets -4.48 12.52 5.02 -10.96
[-11.98, 4.2] [9.03, 16.04] [1.58, 8.45] [-12.95, -9.01]
Leverage Ratio 1.17 9.24 19.65 11.77
[-10.69, 12.12] [2.41, 16.05] [12, 27.42] [7.45, 16.24]
Liquid Assets to Assets -7.82 -1.29 6.05 -5.26
[-12.91, -2.59] [-4.75, 2.14] [2.01, 10.22] [-7.52, -3.02]
ALLL to Total Loans -12.66 -2.95 -1.71 -1.1
[-16.05, -9.32] [-6.25, 0.29] [-5.33, 1.94] [-3.21, 1.01]
74
ROA -1.78 -0.56 -4.48 0.86
[-6.23, 2.67] [-4.29, 3.15] [-8.62, -0.4] [-1.67, 3.37]
Cases Per 100k -3.24 -2.4 -2 -0.51
[-5.68, -0.85] [-4.27, -0.52] [-4.22, 0.14] [-1.78, 0.76]
Constant -0.57 -1.17 -0.24 -2.29
[-6.73, 5.54] [-7.06, 4.72] [-6.26, 5.81] [-8.16, 3.74]
Note: The reported values are posterior means of the product of covariates and parameters, and 95%
credibility intervals in brackets. The results are based on 55,000 MCMC draws with a burn-in of 5000.
I 2020:Q4 Results from the Bayesian Model
Table I.1 reports the results for participation and intensity from Q4 2020. PPP
balances in this quarter reflect total balances from previous quarters, and changes
due to forgiveness and repayments. Columns (1), (3), (5), (7), and (9) report results
for participation. Larger, and less capitalized banks continue to be associated with
greater intensity and participation. However, the relationship between profitability,
and PPP participation is weaker than in the main results. Participation in this
specification is based on participation in Q2 and Q3 of 2020, and is thereby identical
to the outcome in the main specification. The control variables are also identical
to the main specifications. Therefore, differences from the main results arise from
changes in PPP intensity, and final outcomes.
75
Table I.1: Results for participation and intensity from the Bayesian joint model in Q4 2020
∆NIM C&I Gwth Non-PPP C&I Gwth CRE Gwth
(1) (2) (3) (4) (5) (6) (7) (8)
Part. Intensity Part. Intensity Part. Intensity Part. Intensity
Tech exp. to assets -0.072 -0.082 -0.049 -0.048
[-0.22, 0.08] [-0.22, 0.05] [-0.18, 0.09] [-0.18, 0.09]
COVID-affected employment share 0.022 0.032 0.019 0.024
[0.01, 0.03] [0.01, 0.05] [0.01, 0.03] [0.01, 0.04]
ln Assets 0.196 1.167 0.157 0.967 0.166 1.053 0.182 1.072
[0.16, 0.23] [1.02, 1.32] [0.13, 0.19] [0.8, 1.13] [0.14, 0.2] [0.91, 1.2] [0.15, 0.21] [0.92, 1.23]
CI to assets 0.038 0.265 -0.001 0.32 0.038 0.267 0.036 0.263
[0.03, 0.04] [0.24, 0.29] [-0.02, 0.02] [0.29, 0.35] [0.03, 0.04] [0.24, 0.3] [0.03, 0.04] [0.23, 0.29]
Leverage Ratio -0.035 -0.206 -0.029 -0.203 -0.038 -0.222 -0.04 -0.23
[-0.05, -0.02] [-0.27, -0.14] [-0.04, -0.02] [-0.27, -0.13] [-0.05, -0.03] [-0.29, -0.16] [-0.05, -0.03] [-0.3, -0.16]
Liquid Assets to Assets 0.007 0.055 0.003 0.06 0.007 0.056 0.007 0.053
[0, 0.01] [0.04, 0.07] [0, 0.01] [0.04, 0.08] [0, 0.01] [0.04, 0.07] [0, 0.01] [0.04, 0.07]
ALLL to Total Loans 0.01 0.131 0.024 0.363 0.001 0.104 0.014 0.159
[-0.04, 0.06] [-0.16, 0.42] [-0.03, 0.07] [0.05, 0.67] [-0.05, 0.05] [-0.18, 0.38] [-0.04, 0.07] [-0.13, 0.45]
ROA -0.006 -0.229 0.053 0.088 0.053 0.112 0.043 0.058
[-0.07, 0.06] [-0.54, 0.09] [-0.01, 0.11] [-0.22, 0.4] [-0.01, 0.12] [-0.19, 0.42] [-0.02, 0.11] [-0.26, 0.38]
76 Cases Per 100k -0.041 -0.349 -0.036 -0.388 -0.048 -0.383 -0.047 -0.377
[-0.07, -0.02] [-0.46, -0.23] [-0.06, -0.01] [-0.5, -0.27] [-0.07, -0.02] [-0.5, -0.27] [-0.07, -0.02] [-0.49, -0.26]
Constant -1.365 -9.484 -0.639 -8.218 -0.991 -8.095 -1.144 -8.253
[-1.87, -0.87] [-11.66, -7.41] [-1.06, -0.23] [-10.41, -6.03] [-1.4, -0.58] [-10.18, -6.04] [-1.57, -0.72] [-10.42, -6.1]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC draws with a
burn-in of 5000.
Columns (2), (4), (6), (8), and (10) report results for the intensity of PPP partic-
ipation. These results are consistent with the main findings–larger, less capitalized
banks were associated with larger PPP loan shares. This suggests that this group of
banks retained greater shares of PPP loans even after forgiveness was initiated. As
in the case of participation, we find the weaker relationship between ROA and PPP
intensity in the Q4. This result entails that banks that were more profitable in 2019
participated more intensively in the earlier rounds, and booked loans that became
eligible for forgiveness earlier in the program.
Table I.2 reports the results for bank profitability, and loan growth for partici-
pants and non-participants. Columns (1), (3), (5), (7), and (9) report the results of
bank outcomes for participants. The most notable results are that change in NIM
was statistically larger, at 6.28 basis points for participants for every percentage
point increase in PPP share intensity. This shows that the downward pressures of
PPP on net interest margins were largely transitory. Banks that participated more
intensively in the program began to recover margins as forgiveness progressed. Non-
PPP C&I growth declined, and CRE growth increased with the share of PPP loans
to total loans. Banks that participated intensively in the program in earlier quarters
likely began to diversify their portfolio and engage in risk-taking by booking CRE
loans.
77
Table I.2: Results for profitability and loan growth of participating and non-participating banks in Q4 2020
∆NIM(bps) CI Gwth(%) Non-PPP CI Gwth(%) CRE Gwth(%)
(1) (2) (3) (4) (5) (6) (7) (8)
(part.) (non-part.) (part.) (non-part.) (part.) (non-part.) (part.) (non-part.)
PPP Loans to Total Loans 5.759 7.427 -1.924 0.922
[3.43, 8.26] [4.04, 9.63] [-3.55, -0.44] [-0.25, 2.17]
ln Assets -1.619 -0.869 6.188 -6.7 1.06 -5.537 0.086 -2.967
[-3.44, 0.09] [-4.87, 3.75] [4.72, 8.09] [-8.02, -5.38] [-0.08, 2.27] [-6.75, -4.35] [-0.95, 1.07] [-3.86, -2.09]
CI to assets -2.356 -0.323 -4.418 0.11 0.432 -1.599 -0.013 -1.116
[-3.11, -1.66] [-1.51, 1.05] [-5.18, -3.48] [-0.76, 0.89] [0.01, 0.89] [-1.99, -1.22] [-0.35, 0.3] [-1.39, -0.85]
Leverage Ratio 0.23 -1.835 -0.616 0.819 -0.359 1.16 0.383 0.431
[-0.73, 1.28] [-3.28, -0.43] [-1.59, 0.28] [-0.04, 1.66] [-0.93, 0.16] [0.39, 1.94] [-0.01, 0.81] [-0.1, 0.96]
Liquid Assets to Assets -0.966 -0.282 -0.341 -0.204 0.088 -0.332 -0.05 -0.209
[-1.19, -0.76] [-0.7, 0.14] [-0.55, -0.1] [-0.45, 0.04] [-0.02, 0.2] [-0.54, -0.13] [-0.13, 0.02] [-0.36, -0.07]
ALLL to Total Loans -3.176 -6.629 -3.176 -2.234 -2.183 -1.357 -1.291 -1.069
[-6.24, -0.07] [-10.48, -2.77] [-5.94, -0.4] [-5.2, 0.69] [-3.64, -0.74] [-4.04, 1.33] [-2.3, -0.28] [-3.02, 0.88]
ROA -13.664 -1.193 0.869 2.147 -0.811 2.166 -1.379 1.347
[-16.9, -10.35] [-6.04, 3.69] [-2.1, 3.81] [-1.56, 5.83] [-2.3, 0.7] [-1.34, 5.7] [-2.43, -0.34] [-1.32, 4]
Cases Per 100k 1.695 -1.22 -2.512 1.46 -1.144 1.756 0.057 0.966
78 [0.09, 3.41] [-4.22, 1.84] [-4.42, -0.93] [-0.36, 3.27] [-2.09, -0.28] [0.12, 3.39] [-0.58, 0.74] [-0.25, 2.16]
Constant -0.944 -0.512 0.543 -0.331 1.606 0.452 0.894 0.234
[-7.11, 5.18] [-6.66, 5.58] [-5.55, 6.56] [-6.46, 5.83] [-4.52, 7.71] [-5.66, 6.53] [-5.05, 6.92] [-5.76, 6.3]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC
draws with a burn-in of 5000.
Columns (2), (4), (6), (8), and (10) report the results for bank profitability, and
loan growth among non-participants. These findings are consistent with the results
from the main specification. Larger non-participants underwent larger declines in
profitability and loan growth. Non-participants with larger capital buffers, and con-
centrations of C&I loans underwent a growth in this category of loans, but a decline
in CRE loan growth. This suggests that non-participants specialized in C&I lend-
ing continued to extend this category of loans throughout the pandemic, and the
recovery.
Table I.3 summarizes the covariances across the four equations in our Bayesian
joint model. The estimates are broadly consistent with the main results. Partici-
pation and the intensity of participation are positively related, and are also broadly
positively related to bank outcomes. The main exception to this finding is that
change in NIM is negatively associated with participation, and the intensity of par-
ticipation. This likely reflects the effects of the forgiveness program, which resulted
in the reversal of previous relationships between participation and intensity, with
profitability. Banks that were able to access forgiveness and scale down their share
of PPP loans earned larger interest margins by recognizing fees along with interest.
Unobservables underlying non-participation were negatively related to unobservables
related to profitability, and loan growth as in the case of the main results.
79
Table I.3: Covariance estimates from the Bayesian joint model based on Q4 2020 outcomes
∆NIM C&I Non-PPP CRE
Gwth C&I Gwth Gwth
COV(participation, intensity) 5.73 4.81 5.74 5.73
[5.6, 5.86] [3.99, 5.45] [5.6, 5.88] [5.6, 5.87]
COV(participation, bank outcome) -41.93 43.78 11.32 -4.33
[-56.16, -28.67] [30.81, 54.74] [2.76, 20.77] [-11.58, 2.44]
COV(intensity, bank outcome) -236.32 114.25 62.91 -25.28
[-319.47, -157.9] [42.85, 223.33] [13.11, 117.37] [-67.35, 14.07]
COV(non-participation, bank outcome) -1.39 -52.70 -48.76 -32.07
[-32.89, 37.82] [-58.02, -47.45] [-54, -43.81] [-35.81, -28.49]
80
J Robustness: OLS and IV results
As a check on our Bayesian analysis, we estimate our key results using classical OLS
and two-stage least squares methods. We consider similar instruments to those used
in the joint Bayesian model as described below in the 2SLS model. Our participation
and intensity regressions are estimated separately using a logit and OLS framework,
respectively. These estimations may be more familiar to readers but require stronger
identification assumptions in some cases, such as the requirements on instruments
for 2SLS. Broadly, our results using these estimation procedures are similar to those
generated by the Bayesian joint model.
J.1 Logit, OLS, and TSLS Estimation Setup
To formally test for participation characteristics, we estimate a logit regression using
Call Report data for the second and third quarters of 2020. The dependent variable
takes a value of one if the bank reported having PPP loans outstanding as-of quarter
end and zero otherwise. We regress this variable on a set of bank characteristics that
capture capital levels, funding types, size, and business lending concentration.
We estimate the effect of bank characteristics on PPP participation intensity using
an OLS model. The dependent variable in this model is PPP loans outstanding as a
share of total loans. We consider similar bank characteristics in this estimation as in
the participation logit, namely pre-pandmeic levels of bank capital, funding types,
lending concentrations, and size.
Finally, we adopt an instrumental variable framework to address the issue of
endogeneity in the intensity of participation relative to observed bank outcomes.
The source of this endogeneity is the simultaneous determination of PPP intensity,
bank loan growth, and profitability. Banks are likely to have adjusted the size of
their loan portfolios and accordingly, determined the extent of participation in the
PPP with the ultimate objective of maximizing profits. We address this simultaneity
in the determination of bank outcomes and PPP intensity by using an instrument
that isolates variation in the intensity of bank participation due to firm demand for
loans rather than from bank supply decisions. Specifically, we intend to measure the
81
exogenous variation in firm demand for PPP loans induced by economic disruptions
due to the COVID-19 pandemic, and associated containment efforts.
We implement the instrumental variable approach by using two-stage least squares
(TSLS). The first stage of this approach estimates the relationship between PPP-
intensity and the instrument Zemp,i ,
P P Pi = Zemp,i π + Wi0 ψ + νi . (13)
The second stage estimates the effect of the extent of PPP lending that is explained
by the instrument on bank outcomes,
Yi = P Pˆ P i β + Wi0 γ + i . (14)
where Yi denotes net interest margins, change in net interest margins relative to 2019,
growth in C&I loans, growth in C&I loans outside of the PPP, and growth in CRE
loans in separate regressions for each outcome. P P Pi measures the share of PPP
loans to total loans and leases of bank i. Wi is a set of control variables consisting of
the share of business loans to assets, size, return on assets, leverage ratio, the share
of loss allowances to assets, liquid assets, and the deposit-weighted share of COVID
cases per 100,000 population in a bank’s region of operation.
The main exclusion restriction, E[i Zemp,i |Wi ] = 0, is that the deposit-weighted
share of employment in contact-sensitive sectors does not directly affect bank prof-
itability and loan growth outside of the PPP. This measure disrupts the simultaneity
in the determination of bank outcomes and PPP intensity by delineating the vari-
ation in participation intensity that arises from firm demand for loans under the
program. Bartik et al. [2020b] reported survey results that showed that firms in
COVID-affected sectors such as retail and hospitality constituted the largest shares
of applicants for PPP loans. Crucially, the survey responses indicate that approval
rates did not vary substantially by sector, which entails that these sectors were over-
represented among recipients of PPP loans. Therefore, the share of COVID-affected
sectors in a bank’s region of operation manifests demand rather than strategic supply
82
considerations of banks.
Banks’ existing loans to COVID-affected sectors expose a potential channel for
violating the exclusion restriction. When borrowers are unable to service existing
loans, bank profitability and loan growth decline, particularly if the loans remain
unpaid to the extent that they are charged off. In this context, the exclusion restric-
tion requires that the deposit-weighted employment in COVID-affected sectors does
not mirror the share of existing bank loans to such sectors. This requirement is met
as long as certain banks specialize more heavily than others in lending to sectors
such as retail and hospitality irrespective of the sectoral composition of firms in their
region of operation.
We construct an alternate set of instruments that exploit the terms of the PPP
to address endogeneity emanating from bank incentives for participation. We use
the fraction of firms with fewer than 500 employees per county weighted by bank
deposits to determine the share of eligible firms in a bank’s operating region. Other
instruments we consider are the share of unused commitments and the share of
core deposits to total assets. These measures approximate the presence of existing
relationships with firms that would have expedited the PPP application process for
borrowers. Li and Strahan [2021] found that both of these measures were important
predictors of PPP lending among small banks. This finding supports our use of these
measures as relevant instruments for explaining PPP lending. Berger and Udell [1995]
uncover the informational value of unused commitments to banks in that over time,
these products enable lenders to overcome the problems of asymmetric information
in lending to small firms. These instruments have the drawback that they absorb
bank incentives to preserve the quality of their loans by lending to existing borrowers.
We disentangle the effects of relationship lending and emergency pandemic lending
by assessing the variation in treatment effects across instruments.
J.2 Logit Participation Results
Table J.1 shows the results of the logit estimation. Column (1) shows no statisti-
cally significant association between C&I loan concentration, measured as the share
83
of commercial loans to assets, and PPP participation. Similarly, the relationship
between business lending and participation is also not significant when we consider
only small C&I loans– outstanding loans with original amounts less than $1 million–
in column (2). We do, however, find a strong and statistically significant relation-
ship in column (3) when we consider the share of committed but undrawn C&I loan
commitments. In this specification, a percentage point increase in unused C&I com-
mitments relative to assets increases the log odds ratio of participation by about 0.11
points. In column (4), we consider a model that includes all these C&I loan measures
which confirms that unused commitments on C&I loans are the best predictor of the
three for PPP participation.
Table J.1: PPP Participation Determinants
All Banks Banks < $1 billion
(1) (2) (3) (4) (5) (6) (7) (8)
CI to assets -0.000 -0.013 0.001 0.004
(0.008) (0.018) (0.008) (0.021)
Small CI to assets -0.009 -0.006 -0.010 -0.025
(0.009) (0.020) (0.009) (0.022)
Unused CI Commitments to Assets 0.110∗∗∗ 0.127∗∗∗ 0.110∗∗∗ 0.118∗∗∗
(0.027) (0.030) (0.027) (0.029)
ln Assets 0.723∗∗∗ 0.711∗∗∗ 0.636∗∗∗ 0.614∗∗∗ 0.749∗∗∗ 0.733∗∗∗ 0.668∗∗∗ 0.625∗∗∗
(0.046) (0.046) (0.051) (0.054) (0.049) (0.049) (0.053) (0.057)
ROA 0.307∗∗∗ 0.313∗∗∗ 0.321∗∗∗ 0.330∗∗∗ 0.297∗∗∗ 0.305∗∗∗ 0.310∗∗∗ 0.330∗∗∗
(0.063) (0.063) (0.064) (0.064) (0.063) (0.063) (0.064) (0.064)
Liquid Assets To Assets -0.011∗∗∗ -0.012∗∗∗ -0.008∗∗∗ -0.011∗∗∗ -0.010∗∗∗ -0.012∗∗∗ -0.008∗∗∗ -0.010∗∗∗
(0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Leverage Ratio -0.071∗∗∗ -0.072∗∗∗ -0.069∗∗∗ -0.071∗∗∗ -0.066∗∗∗ -0.068∗∗∗ -0.064∗∗∗ -0.067∗∗∗
(0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011)
ALLL to Total Loans -0.034 -0.025 -0.034 -0.010 -0.046 -0.033 -0.044 -0.021
(0.051) (0.047) (0.050) (0.043) (0.050) (0.045) (0.049) (0.042)
Cases Per 100k -0.020 -0.018 -0.024 -0.020 -0.017 -0.015 -0.022 -0.018
(0.041) (0.041) (0.041) (0.041) (0.042) (0.042) (0.041) (0.041)
Constant -5.789∗∗∗ -5.579∗∗∗ -5.100∗∗∗ -4.708∗∗∗ -6.139∗∗∗ -5.863∗∗∗ -5.512∗∗∗ -4.878∗∗∗
(0.584) (0.606) (0.611) (0.647) (0.608) (0.634) (0.631) (0.686)
Observations 7,786 7,786 7,786 7,786 6,853 6,853 6,853 6,853
Loglik -2,362.48 -2,361.80 -2,341.77 -2,337.91 -2,274.61 -2,273.69 -2,255.32 -2,251.48
Psuedo R2 0.12 0.12 0.12 0.13 0.10 0.10 0.11 0.11
Notes: Dependent variable is an indicator for PPP loan outstanding at the end of the quarter. Sample is 2020:Q2 and 2020:Q3. Regressor balance
sheet variables are measured as four quarter averages from 2019.
Robust standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Among other characteristics, Table J.1 shows that larger and more profitable
banks were more likely to participate. A one percentage point increase in bank
assets is associated with about an 0.8 point increase in the log odds ratio of PPP
84
participation across specifications. Similarly, a one percentage point increase in a
bank’s return on assets (ROA) increases the log odds ratio of participation by about
0.3 points across specifications.
Similar to our findings that lending concentrations were important drivers of
participation, we also find that banks with greater holdings of liquid assets were less
likely to participate. For each 1 percentage point increase in the share of liquid assets
to total assets, we find that the probability of participation declines by 0.01 log odd
points.
Somewhat counter to our findings that more financially viable banks were likely
to participate, we find the opposite result regarding capital and loan loss allowance.
In this case, we find that better capitalized banks as measured by higher leverage
ratios were less likely to participate. The log odds ratio of participation declines by
a somewhat modest 0.07 points for each 1 percentage point increase in the leverage
ratio, though this effect is statistically significant. Similarly, banks that have reserved
more allowance for loan losses as a share of total loans appear to have been less
likely to participate. For each percentage point increase in the allowance stock to
total loans, the log odds ratio of participation declines about 0.03 points. This
relationship is not statistically significant though.
The COVID crisis itself seems to have had little impact on a bank’s decision of
whether or not to participate in PPP lending. Across all specifications, the deposit-
weighted COVID case variable is statistically insignificant meaning that local COVID
cases in a bank’s operating area was not an important participation determinant.
Columns (5) - (8) confirm that our results hold for the smallest of community banks,
those with total assets less than $1 billion. Qualitatively, our results are similar
to the full sample with larger and more profitable banks more likely to contribute.
However, less capitalized banks and those with greater exposure to business line
draws were also more likely to participate. The parameter estimates across these
specifications are similar in magnitude to the full sample as well.
85
J.3 OLS Participation Intensity Results
We next turn to how much participants decided to participate. We use the share
of PPP loans outstanding to total loans outstanding to determine a bank’s PPP
participation intensity. These regressions tell us how the level of PPP participation
varied conditional on the set of bank characteristics. The results are shown in Table
J.2.
Table J.2: PPP Participation Intensity Determinants
All Banks Banks < $1 billion
(1) (2) (3) (4) (5) (6) (7) (8)
CI to assets 0.359∗∗∗ 0.119∗∗∗ 0.442∗∗∗ 0.253∗∗∗
(0.022) (0.038) (0.025) (0.047)
Small CI to assets 0.467∗∗∗ 0.174∗∗∗ 0.523∗∗∗ 0.078
(0.042) (0.054) (0.037) (0.058)
Unused CI Commitments to Assets 0.763∗∗∗ 0.568∗∗∗ 0.844∗∗∗ 0.549∗∗∗
(0.033) (0.050) (0.040) (0.057)
ln Assets 0.733∗∗∗ 1.522∗∗∗ 0.163∗∗ 0.463∗∗∗ 1.466∗∗∗ 2.308∗∗∗ 0.787∗∗∗ 1.071∗∗∗
(0.071) (0.086) (0.078) (0.105) (0.105) (0.119) (0.115) (0.135)
ROA -0.268 -0.452 -0.302 -0.317 -0.437 -0.635∗∗ -0.461 -0.438
(0.273) (0.283) (0.286) (0.281) (0.284) (0.299) (0.303) (0.291)
Liquid Assets To Assets 0.106∗∗∗ 0.097∗∗∗ 0.085∗∗∗ 0.107∗∗∗ 0.118∗∗∗ 0.105∗∗∗ 0.090∗∗∗ 0.118∗∗∗
(0.007) (0.008) (0.007) (0.007) (0.008) (0.008) (0.008) (0.008)
Leverage Ratio -0.211∗∗∗ -0.176∗∗∗ -0.226∗∗∗ -0.193∗∗∗ -0.171∗∗∗ -0.141∗∗∗ -0.207∗∗∗ -0.168∗∗∗
(0.034) (0.036) (0.033) (0.033) (0.034) (0.036) (0.033) (0.033)
ALLL to Total Loans 0.368∗∗ 0.409∗∗ 0.360∗∗ 0.334∗∗ 0.195 0.271 0.236 0.187
(0.164) (0.167) (0.176) (0.168) (0.165) (0.170) (0.183) (0.170)
Cases Per 100k 0.117 0.157∗ 0.209∗∗ 0.162∗∗ 0.192∗∗ 0.228∗∗ 0.283∗∗∗ 0.225∗∗∗
(0.084) (0.086) (0.083) (0.081) (0.089) (0.092) (0.089) (0.086)
Constant -3.826∗∗∗ -13.144∗∗∗ 4.541∗∗∗ -1.344 -13.654∗∗∗ -23.116∗∗∗ -3.165∗∗ -9.343∗∗∗
(1.096) (1.536) (1.148) (1.630) (1.523) (1.830) (1.592) (1.965)
Observations 7,022 7,022 7,022 7,022 6,104 6,104 6,104 6,104
Adjusted R2 0.152 0.112 0.169 0.196 0.190 0.134 0.184 0.226
Notes: Dependent variable is PPP loans as a share of total loans in 2020:Q2 and 2020:Q3. Regressor balance sheet variables are measured as four
quarter averages from 2019. COVID cases are county level case counts averaged over counties where the bank operates a branch according to the
Summary of Deposit data. Daily county-level COVID case counts are drawn from John Hopkins.
t statistic in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Contrary to the logit model for PPP participation, the concentration of C&I
lending on a bank’s balance sheet seems to more strongly predict how intensively the
bank participated in PPP lending. Columns (1) - (4) repeat the previous exercise of
considering each C&I loan exposure measure individually and then jointly. Column
(1) shows that a one percentage point increase in a bank’s C&I exposure as a share
of total assets is associated with about a 35 basis point increase in PPP lending
relative to total loans. Similarly, an increase in a bank’s share of small C&I lending–
86
often used as a proxy for small business lending– is associated with about a 44
basis point increase in relative PPP lending. This is slightly higher than the overall
C&I lending effect, suggesting that small loans may proxy for existing relationships
with eligible PPP customers, a result discussed by Li and Strahan [2021]. Finally,
column (3) shows that unused C&I commitments relative to total assets is also a
statistically significant predictor of PPP participation intensity and it is qualitatively
larger than the coefficients on C&I concentration ratios. We jointly consider all
these C&I loan measures in column (4). All the C&I lending concentration measures
remain statistically significant and positive, with unused C&I commitments being the
strongest predictor of PPP intensity as measured by coefficient size. We interpret
this as a signal of the risk-aversion channel that PPP provided because undrawn C&I
commitments are an ex-ante measure of C&I liquidity and credit risk.
Regarding other characteristics, similar to the logit regressions on participation,
we find that larger banks were more likely to make more PPP loans as a share of
total loans. However, more profitable banks were less likely to participate as inten-
sively though this effect is not statistically significant. Banks with larger liquid asset
holdings, however, did participate more intensively, contrary to the participation
results.
Our remaining regressors provide more evidence of a risk-aversion channel. Banks
with more allowance against loan losses were likely to participate more intensively,
contrary to our findings on participation alone, while capital ratios remain a negative
predictor of participation intensity. Both greater ALLL holdings relative to loans and
lower capital ratios provide measures of risk. For ALLL, banks are required to hold
larger ALLL stocks when larger losses are more likely. Similarly, banks with smaller
capital bases will be more threatened by outsized loan losses emanating from the
pandemic’s economic effects.
In the next to last row, we consider a bank’s local exposure to COVID cases.
Across specifications we find that banks with greater local exposure to COVID made
more PPP loans as a share of total loans across the period when the program was
active. However, in some specifications and samples the finding is only weakly signif-
icant or even insignificant. Nonetheless, this result provides yet more evidence that
87
PPP provided some protection from potential risks related to the pandemic.
Columns (5) - (8) report results for the same specification for banks with less
than $1 billion in total assets. The results here are qualitatively similar to the full
sample results. Specifically, banks with more C&I loan exposure or those facing more
drawdown risk were more likely to make more PPP loans. Similarly, larger banks
and those with more liquid asset holdings made relatively more loans. However,
more profitable banks and better capitalized banks made fewer loans as a share of
total lending. COVID case exposure is a slightly more significant predictor of lending
intensity among smaller community banks than it is for the sample as a whole.20
J.4 TSLS Balance Sheet Impact Results
We evaluate the effects of increased intensity of participation in the PPP on the
balance sheets of participating banks. We examine how participation affected the
net interest margins, change in net interest margin, and growth in C&I and CRE
loans relative to levels in 2019.
Table J.3 reports the results from the first stage regressions. Our main instrument
is the deposit-weighted share of employment in COVID-affected industries. This
instrument, Zemp,i satisfies the assumption of relevance π 6= 0. Column (1) shows that
the deposit-weighted share of COVID-affected employment is positively associated
with PPP intensity and that this relationship is statistically significant. A 1 percent
increase in this ratio is associated with a 10 basis point increase in the intensity of
PPP participation. The F-test in the last two rows of the table test the model fit
after including the instrument. We reject the null hypothesis of a weak instrument
at a 1 percent level of significance [Cragg and Donald, 1993; Stock and Yogo, 2005].
Columns (2)-(4) summarize the first stage results for the remaining three in-
struments that we have considered. Notably, PPP share has a significant negative
relationship with the deposit-weighted share of employment in small firms. Columns
(3) and (4) show that core deposit shares and unused commitments, which measure
20
In Appendix J, we report the results shown in table J.2 disaggregated across 2020:Q2 and
2020:Q3. The results are mostly qualitatively similar to the combined results though COVID cases
are a stronger predictor of participation in the later quarter.
88
preexisting relationships with small firms, have a positive and significant relationship
with PPP participation intensity. In all cases, we reject the null hypothesis of weak
instruments.
Table J.4 summarizes the results from the second stage of instrumental variable
regression based on the share of employment in COVID-affected sectors. The table
reports OLS results along with the Hausman test of endogeneity. In all cases, the
IV result is larger than the OLS result, indicating that the endogeneity of the OLS
estimate biases the effect toward zero. Furthermore, the F statistic from the Haus-
man test indicates that the OLS results are not consistent, that is we reject the null
that the coefficient on residuals generated from a regression of the instrument on all
the controls is zero when used in the baseline regression. In all cases, we find that
the estimated residuals improve the regression fit except for the case on the change
in NIMs.
Notably, PPP participation resulted in a statistically significant decline in the
change in NIM. The levels of NIM increased marginally and in line with our expecta-
tions, C&I growth increased substantially in response to increased PPP participation.
The results in column (2) show that higher participation in the PPP entailed a small,
statistically significant improvement in bank NIMs. A one percent increase in PPP
loans to total loans generated a 5 basis point rise in NIMs. At the mean level of PPP
participation of 8.5 percent, the estimated coefficient predicts a 40 basis point rise
in NIM. This is consistent with the terms of PPP loans, which bear an interest of
merely 1 percent and result in fee income to banks, which is only fully realized after
a loan is forgiven.
Column (4) shows a statistically and economically significant decline of 3.75 basis
points in ∆NIM, which the change in the level of NIM in 2020 relative to 2019.
Even though this finding may appear at odds with the estimated positive effect of
PPP participation on NIM, these two results can be reconciled by considering the
interpretation of the two coefficients. The first result indicates that on balance,
PPP loans resulted in a small positive increase in NIM. The second result shows
that despite the rise in NIMs emanating from participation in the PPP, margins
fell relative to levels in 2019. This shows that the growth in NIMs in response
89
to a marginal increase in the share of PPP loans was substantially smaller than
the growth in NIMs generated by the bank’s asset portfolio from the pre-pandemic
period. Because PPP loans displaced regular bank lending, potential growth to NIMs
from other loans was shut down.
In column (6), we find that PPP loans generate a statistically and economically
significant increase in the growth of C&I loans. A one percent rise in the share of PPP
loans to total loans generated 15 percent growth in C&I loans relative to 2019. This
outsized effect of PPP participation on C&I growth is explained by the fact that PPP
lending contributed directly to bank loan portfolios. In addition, other factors tied
to the pandemic also expanded C&I lending. For example, firms rapidly drew down
on their lines of credit in response to the panic in March 2020 and thereby converted
off-balance sheet commitments into lending reported on bank balance sheets [Li,
Strahan, and Zhang, 2020]. Indeed, the introduction of the PPP alleviated this trend
as firms that received PPP loans repaid larger shares of the amount that they had
drawn down relative to non-PPP recipients [Chodorow-Reich et al., 2021]. Indeed,
in column (8) we see that non-PPP C&I lending increased modestly for banks with
larger PPP portfolios. However, non-C&I lending, as measured by CRE growth, did
not expand with PPP lending which is shown in column (10). We find these result
even conditioning on the size of the C&I portfolio shown in the second row.
The third row of Table J.4 shows that bank size had differential effects on NIM,
the change in NIM and C&I growth. Larger banks underwent a decline in NIM
but a rise in the change in NIM relative to 2019 levels. This suggests that non-
PPP lending that was forgone during the pandemic had yielded larger margins for
smaller banks than larger banks. Finally, we find that both total and non-PPP C&I
growth declined with asset size. This result is primarily driven by base effects as
loan portfolios grew by a larger percent among small banks for a given change in
C&I lending.
Banks with higher shares of liquid assets earned lower NIM and underwent a
steeper decline in NIM relative to 2019 levels. Participation in the PPP Liquidity
Facility (PPPLF) provides a potential explanation for this observation. This facil-
ity carried a low interest rate and was likely tapped by banks that were liquidity-
90
constrained. The low cost of funds from the facility would have supported margins
from falling substantially for participating banks. The weakly negative relationship
between liquid assets shares and C&I growth suggests that banks with more liquid
assets were also likely more conservative and expanded their loan portfolios to a more
limited extent than small banks. Liquid asset shares are not significantly associated
with CRE lending.
Banks with larger pre-pandemic ROA earned larger NIM but experienced a larger
decline in NIM relative to 2019. This indicates that banks that were more profitable
pre-pandemic underwent a greater opportunity cost by forgoing their regular lending
activities and instead participating in the PPP. More profitable banks also had lower
growth in C&I lending, both overall, and outside the PPP, as well as lower growth
in CRE lending.
Likewise, better capitalized banks had greater reductions in net interest margins
but more total C&I lending growth. This result differs from our result on PPP
intensity because it measures C&I lending relative to the base period whereas our
PPP intensity result measures PPP as a share of all loans in that quarter. Thus,
better capitalized banks increased C&I lending more but they also increased other
lending more as shown by the positive and significant coefficient on CRE growth in
column (10).
Table J.10 reports the coefficients on PPP share from the second stage of the IV
regressions using other instruments in place of the share of employment in COVID-
affected industries. Across different instruments, we find that the NIM level effect is
inconsistent both in sign and statistical significance. The change in NIMs however is
consistently negative across different instruments and statistically significant except
for when we use unused commitments as an instrument. Similarly, we find that
PPP lending boosted total C&I lending in all specifications but had much smaller
or even negative effects on non-PPP C&I lending. Results on CRE growth are also
not consistent across instruments with some specifications showing a statistically
significant increase and other showing insignificant declines. Thus, we conjecture
that the most consistent result is that PPP increased C&I lending sharply but pushed
net interest margins down considerably for the lenders participating most intensively.
91
J.5 Quarterly Estimates
Tables J.5 and J.6 provide results for the specifications presented in Table J.1 for
the quarters 2020:Q2 and 2020:Q3, respectively. The results for each quarter are
qualitatively similar to the combined results. In particular, larger and more profitable
banks were more likely to participate while more capitalized banks were less likely
to participate. Tables J.7 and J.8 report specifications on PPP lending intensity but
by individual quarter. The results remain qualitatively similar with banks facing
more C&I exposure typically making more PPP loans. Large and riskier banks–
as measured by leverage capital ratios– were also more likely to participate across
quarters. However, one difference does emerge. COVID cases seem to be a better
predictor of PPP loan holdings, particularly for smaller banks, only for the third
quarter which corresponds to the end of the second funding round. This suggests
that loan targeting improved as the program progressed and more controls were
added.
J.6 2020:Q4 Bank Outcomes
Table J.9 reports the results for 2020:Q4. This quarter was not considered in our
primary sample because the PPP was not operational during this time. Thus, any
changes in PPP lending are due to sales, purchases, paydowns or forgiveness. On
net, we find that banks with greater PPP loan shares had higher levels of NIMs
and C&I loan growth in 2020:Q4. There was no statistically singificant change in
the decline in NIMs and there was a moderate increase in CRE lending growth for
large PPP lenders. While this is only a single cross-section, it suggests that some
of our key findings are transitory. Thus, the impact on profitability is likely to be
temporary as loans are forgiven. Moreover, revenue generated by PPP lending and
efforts to boost post-PPP lending profitability may increase risk-taking in the future.
Unfortunately, our data series is too short to make strong statements about these
impacts. The PPP began again in 2021:Q1 so 2020:Q4 remains our only quarter
since the pandemic began without a PPP program or financial crisis.
Finally, Table J.10 shows the results from the IV using all the possible instru-
92
ments. Just as in the main results, we find disagreement across the estimates on the
level of net interest margins. The change in NIM is always negative and mostly signif-
icant. PPP boost C&I lending unambiguosly but non-PPP lending results are mixed
across instruments. CRE lending increases in some specifications by is negative and
insignificant in at least one specification.
93
Table J.3: PPP Intensity Share Determinants
(1) (2) (3) (4)
COVID-affected employment share 0.105∗∗∗
(0.009)
Small firm employment share -0.098∗∗∗
(0.007)
Core Deposits To Assets 0.071∗∗∗
(0.009)
Unused CI Commitments to Assets 0.535∗∗∗
(0.050)
CI to assets 0.365∗∗∗ 0.340∗∗∗ 0.364∗∗∗ 0.206∗∗∗
(0.022) (0.021) (0.022) (0.031)
ln Assets 0.564∗∗∗ 0.427∗∗∗ 0.738∗∗∗ 0.248∗∗∗
(0.074) (0.077) (0.071) (0.080)
ROA -0.191 -0.052 -0.424 -0.265
(0.270) (0.272) (0.271) (0.277)
Leverage Ratio -0.205∗∗∗ -0.219∗∗∗ -0.144∗∗∗ -0.211∗∗∗
(0.033) (0.033) (0.035) (0.032)
ALLL to Total Loans 0.425∗∗∗ 0.505∗∗∗ 0.380∗∗ 0.336∗∗
(0.162) (0.160) (0.164) (0.168)
Liquid Assets To Assets 0.101∗∗∗ 0.104∗∗∗ 0.084∗∗∗ 0.103∗∗∗
(0.007) (0.007) (0.008) (0.007)
Cases Per 100k 0.141∗ -0.148∗ 0.194∗∗ 0.158∗
(0.084) (0.083) (0.084) (0.081)
Constant -3.866∗∗∗ 5.851∗∗∗ -9.107∗∗∗ 1.869
(1.081) (1.289) (1.355) (1.304)
Observations 7,022 7,022 7,022 7,022
F value 114.472 227.765 82.293 351.182
F p-value 0.000 0.000 0.000 0.000
Notes: Dependent variable is PPP loans as a share of total loans in 2020:Q2 and 2020:Q3.
Regressor balance sheet variables are measured as four quarter averages from 2019. CRA origi-
nation share is calculated using 2019 origination volumes reported on the banks CRA disclosure.
Small firm employment share is the share of firms with 500 or fewer employees operating in a
county according to the QWI database of the U.S. Census. Share of affected employment is de-
termined at the county level from the share of employment in the most affected industrial sectors.
Affected industries are defined as the bottom quartile of total employment change from January
to April 2020. See Boyarchenko et al. [2020] for more information. County level variables are
weighted by bank branch deposits in each county according to the Summary of Deposit data.
County employment shares are from the QCEW database of the Bureau of Labor Statistics.
Robust standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
94
Table J.4: Bank Outcomes IV Regression: Employment share in COVID-affected
industries
∆ NIM C&I Gwth Non-PPP C&I Gwth CRE Gwth
(1) (2) (3) (4) (5) (6) (7) (8)
OLS IV OLS IV OLS IV OLS IV
PPP Loans to Total Loans -1.223∗∗∗ -3.246∗∗∗ 11.257∗∗∗ 15.066∗∗∗ -0.104∗∗ 0.768∗∗ 0.176∗∗∗ 0.258
(0.244) (0.704) (0.236) (0.995) (0.049) (0.357) (0.040) (0.297)
CI to assets -0.534∗∗∗ 0.194 -7.821∗∗∗ -9.190∗∗∗ 0.118∗∗∗ -0.195 0.118∗∗∗ 0.088
(0.103) (0.275) (0.274) (0.448) (0.043) (0.132) (0.038) (0.112)
ln Assets 0.808 2.292∗∗∗ 1.942∗∗ -0.851 -0.574∗∗ -1.213∗∗∗ 0.403∗ 0.343
(0.542) (0.757) (0.786) (1.126) (0.260) (0.362) (0.213) (0.290)
ROA -12.211∗∗∗ -12.754∗∗∗ -3.564∗ -2.542 -2.294∗∗∗ -2.060∗∗∗ -1.872∗∗∗ -1.850∗∗∗
(2.053) (1.758) (2.006) (2.155) (0.597) (0.614) (0.536) (0.540)
Leverage Ratio -1.470∗∗∗ -1.897∗∗∗ 0.912∗∗ 1.715∗∗∗ -0.004 0.180 0.265∗∗ 0.283∗∗
(0.224) (0.277) (0.371) (0.443) (0.128) (0.145) (0.116) (0.134)
ALLL to Total Loans -3.918∗∗∗ -3.173∗∗ -5.345∗∗∗ -6.746∗∗∗ -2.839∗∗∗ -3.160∗∗∗ -0.520 -0.550
(1.411) (1.357) (2.006) (2.025) (0.489) (0.517) (0.646) (0.657)
Liquid Assets To Assets -0.295∗∗∗ -0.081 -0.655∗∗∗ -1.058∗∗∗ 0.038 -0.054 -0.030 -0.038
(0.058) (0.093) (0.110) (0.155) (0.029) (0.048) (0.023) (0.039)
Cases Per 100k -9.180∗∗∗ -8.943∗∗∗ 3.303∗∗∗ 2.856∗∗∗ -0.097 -0.199 0.026 0.016
(0.625) (0.634) (1.008) (1.040) (0.318) (0.330) (0.232) (0.233)
Constant 21.743∗∗∗ 14.000 53.560∗∗∗ 68.136∗∗∗ 9.764∗∗ 13.101∗∗∗ -0.189 0.125
(7.752) (8.595) (11.413) (13.190) (3.821) (4.113) (3.127) (3.238)
Observations 7,022 7,022 7,022 7,022 7,022 7,022 7,022 7,022
Hausman F value 0.70 44.48 8.47 5.83
Hausman p-value 0.40 0.00 0.00 0.02
Notes: Instrumental variable is employment share in COVID-affected industries. Sample is 2020:Q2 and 2020:Q3. Regressor balance sheet
variables are measured as four quarter averages from 2019. CRA origination share is calculated using 2019 origination volumes reported on
the banks CRA disclosure. COVID cases are county level case counts averaged over counties where the bank operates a branch according to
the Summary of Deposit data. Daily county-level COVID case counts are drawn from John Hopkins. COVID-affected employment share is
employment in industries that underwent the largest decline in employment averaged over counties where the bank operates a branch according
to the Summary of Deposit data. County-level employment share in COVID-affected industries is obtained from the QCEW databse of the
Bureau of Labor Statistics.
Robust standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
95
Table J.5: 2020:Q2 PPP Participation Determinants
All Banks Banks < $1 billion
(1) (2) (3) (4) (5) (6) (7) (8)
CI to assets 0.001 -0.003 0.001 0.009
(0.011) (0.026) (0.011) (0.029)
Small CI to assets -0.009 -0.016 -0.011 -0.029
(0.012) (0.029) (0.012) (0.032)
Unused CI Commitments to Assets 0.101∗∗∗ 0.113∗∗∗ 0.102∗∗∗ 0.108∗∗∗
(0.038) (0.041) (0.037) (0.040)
ln Assets 0.734∗∗∗ 0.721∗∗∗ 0.655∗∗∗ 0.623∗∗∗ 0.741∗∗∗ 0.725∗∗∗ 0.668∗∗∗ 0.622∗∗∗
(0.063) (0.064) (0.070) (0.076) (0.068) (0.069) (0.074) (0.080)
ROA 0.348∗∗∗ 0.355∗∗∗ 0.355∗∗∗ 0.370∗∗∗ 0.340∗∗∗ 0.349∗∗∗ 0.347∗∗∗ 0.369∗∗∗
(0.092) (0.092) (0.093) (0.094) (0.092) (0.092) (0.093) (0.094)
Liquid Assets To Assets -0.012∗∗∗ -0.013∗∗∗ -0.010∗∗ -0.012∗∗∗ -0.012∗∗∗ -0.013∗∗∗ -0.010∗∗ -0.012∗∗∗
(0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004)
Leverage Ratio -0.071∗∗∗ -0.072∗∗∗ -0.069∗∗∗ -0.071∗∗∗ -0.070∗∗∗ -0.071∗∗∗ -0.068∗∗∗ -0.070∗∗∗
(0.015) (0.015) (0.014) (0.015) (0.015) (0.015) (0.014) (0.015)
Leverage Ratio 0.000
(.)
ALLL to Total Loans -0.060 -0.049 -0.058 -0.036 -0.067 -0.054 -0.064 -0.044
(0.068) (0.063) (0.066) (0.059) (0.067) (0.062) (0.065) (0.058)
Cases Per 100k 0.044 0.046 0.025 0.029 0.055 0.058 0.036 0.041
(0.121) (0.122) (0.119) (0.119) (0.124) (0.125) (0.121) (0.122)
Constant -5.941∗∗∗ -5.711∗∗∗ -5.310∗∗∗ -4.817∗∗∗ -6.038∗∗∗ -5.755∗∗∗ -5.472∗∗∗ -4.810∗∗∗
(0.802) (0.834) (0.842) (0.907) (0.853) (0.888) (0.885) (0.962)
Observations 3,895 3,895 3,895 3,895 3,426 3,426 3,426 3,426
Loglik -1,180.57 -1,180.20 -1,171.72 -1,170.10 -1,142.41 -1,141.92 -1,134.03 -1,132.16
Pseudo R2 0.12 0.12 0.13 0.13 0.11 0.11 0.11 0.11
Notes: Dependent variable is an indicator for any PPP loans outstanding at quarter end. Sample is 2020:Q2. Regressor balance sheet variables are
measured as four quarter averages from 2019.
Robust standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
96
Table J.6: 2020:Q3 PPP Participation Determinants
All Banks Banks < $1 billion
(1) (2) (3) (4) (5) (6) (7) (8)
CI to assets -0.001 -0.022 0.002 0.000
(0.011) (0.023) (0.012) (0.029)
Small CI to assets -0.008 0.003 -0.010 -0.021
(0.013) (0.026) (0.012) (0.031)
Unused CI Commitments to Assets 0.118∗∗∗ 0.142∗∗∗ 0.117∗∗∗ 0.129∗∗∗
(0.040) (0.045) (0.039) (0.043)
ln Assets 0.718∗∗∗ 0.706∗∗∗ 0.623∗∗∗ 0.610∗∗∗ 0.762∗∗∗ 0.748∗∗∗ 0.676∗∗∗ 0.635∗∗∗
(0.066) (0.067) (0.074) (0.077) (0.069) (0.070) (0.076) (0.082)
ROA 0.268∗∗∗ 0.273∗∗∗ 0.289∗∗∗ 0.291∗∗∗ 0.257∗∗∗ 0.263∗∗∗ 0.275∗∗∗ 0.291∗∗∗
(0.085) (0.085) (0.087) (0.086) (0.085) (0.085) (0.087) (0.088)
Liquid Assets To Assets -0.009∗∗ -0.010∗∗ -0.007 -0.009∗∗ -0.009∗∗ -0.010∗∗ -0.007 -0.009∗∗
(0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004)
Leverage Ratio -0.072∗∗∗ -0.073∗∗∗ -0.069∗∗∗ -0.070∗∗∗ -0.063∗∗∗ -0.064∗∗∗ -0.061∗∗∗ -0.063∗∗∗
(0.017) (0.017) (0.016) (0.016) (0.016) (0.016) (0.015) (0.016)
ALLL to Total Loans -0.005 0.002 -0.007 0.019 -0.023 -0.011 -0.021 0.003
(0.080) (0.073) (0.080) (0.065) (0.077) (0.069) (0.076) (0.062)
Cases Per 100k -0.060 -0.058 -0.062 -0.056 -0.064 -0.061 -0.066 -0.061
(0.051) (0.052) (0.051) (0.051) (0.052) (0.052) (0.051) (0.051)
Constant -5.664∗∗∗ -5.474∗∗∗ -4.920∗∗∗ -4.637∗∗∗ -6.282∗∗∗ -6.016∗∗∗ -5.599∗∗∗ -5.002∗∗∗
(0.849) (0.882) (0.886) (0.926) (0.867) (0.906) (0.900) (0.982)
Observations 3,891 3,891 3,891 3,891 3,427 3,427 3,427 3,427
Loglik -1,180.44 -1,180.15 -1,168.60 -1,166.14 -1,130.70 -1,130.29 -1,119.81 -1,117.86
Pseudo R2 0.11 0.11 0.12 0.12 0.10 0.10 0.10 0.11
Notes: Dependent variable is an indicator for any PPP loans outstanding at quarter end. Sample is 2020:Q3. Regressor balance sheet variables are
measured as four quarter averages from 2019.
Robust standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
97
Table J.7: PPP Participation Intensity Determinants: 2020:Q2
All Banks Banks < $1 billion
(1) (2) (3) (4) (5) (6) (7) (8)
CI to assets 0.344∗∗∗ 0.097∗∗ 0.422∗∗∗ 0.221∗∗∗
(0.028) (0.044) (0.032) (0.052)
Small CI to assets 0.455∗∗∗ 0.187∗∗∗ 0.508∗∗∗ 0.101
(0.059) (0.072) (0.053) (0.075)
Unused CI Commitments to Assets 0.745∗∗∗ 0.570∗∗∗ 0.821∗∗∗ 0.551∗∗∗
(0.044) (0.066) (0.053) (0.076)
ln Assets 0.770∗∗∗ 1.529∗∗∗ 0.210∗ 0.518∗∗∗ 1.502∗∗∗ 2.310∗∗∗ 0.840∗∗∗ 1.139∗∗∗
(0.099) (0.120) (0.109) (0.144) (0.147) (0.165) (0.160) (0.188)
ROA -0.100 -0.267 -0.147 -0.166 -0.259 -0.438 -0.297 -0.279
(0.319) (0.328) (0.343) (0.334) (0.333) (0.345) (0.364) (0.348)
Liquid Assets To Assets 0.107∗∗∗ 0.099∗∗∗ 0.088∗∗∗ 0.108∗∗∗ 0.119∗∗∗ 0.106∗∗∗ 0.093∗∗∗ 0.120∗∗∗
(0.010) (0.010) (0.010) (0.010) (0.011) (0.011) (0.011) (0.011)
Leverage Ratio -0.231∗∗∗ -0.201∗∗∗ -0.243∗∗∗ -0.211∗∗∗ -0.193∗∗∗ -0.166∗∗∗ -0.225∗∗∗ -0.187∗∗∗
(0.045) (0.047) (0.043) (0.044) (0.045) (0.046) (0.044) (0.044)
ALLL to Total Loans 0.340 0.384∗ 0.332 0.307 0.159 0.238 0.203 0.155
(0.227) (0.231) (0.244) (0.233) (0.229) (0.235) (0.253) (0.235)
Cases Per 100k -0.129 -0.032 -0.082 -0.102 0.244 0.294 0.236 0.231
(0.256) (0.256) (0.249) (0.247) (0.285) (0.284) (0.274) (0.273)
Constant -4.019∗∗∗ -13.025∗∗∗ 4.139∗∗∗ -1.786 -13.819∗∗∗ -22.945∗∗∗ -3.652∗ -9.952∗∗∗
(1.513) (2.121) (1.585) (2.296) (2.094) (2.509) (2.191) (2.781)
Observations 3,509 3,509 3,509 3,509 3,047 3,047 3,047 3,047
Adjusted R2 0.148 0.112 0.169 0.194 0.184 0.134 0.183 0.221
Notes: Dependent variable is PPP loans as a share of total loans in 2020:Q2. Regressor balance sheet variables are measured as four quarter averages
from 2019. COVID cases are county level case counts averaged over counties where the bank operates a branch according to the Summary of Deposit
data. Daily county-level COVID case counts are drawn from John Hopkins.
t statistic in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
98
Table J.8: PPP Participation Intensity Determinants: 2020:Q3
All Banks Banks < $1 billion
(1) (2) (3) (4) (5) (6) (7) (8)
CI to assets 0.374∗∗∗ 0.141∗∗ 0.462∗∗∗ 0.284∗∗∗
(0.033) (0.062) (0.038) (0.077)
Small CI to assets 0.479∗∗∗ 0.159∗∗ 0.537∗∗∗ 0.056
(0.059) (0.081) (0.052) (0.090)
Unused CI Commitments to Assets 0.781∗∗∗ 0.567∗∗∗ 0.867∗∗∗ 0.547∗∗∗
(0.048) (0.075) (0.059) (0.085)
ln Assets 0.702∗∗∗ 1.515∗∗∗ 0.115 0.408∗∗∗ 1.424∗∗∗ 2.297∗∗∗ 0.720∗∗∗ 0.993∗∗∗
(0.102) (0.124) (0.113) (0.152) (0.150) (0.172) (0.165) (0.194)
ROA -0.431 -0.628 -0.451 -0.463 -0.598 -0.814∗ -0.607 -0.580
(0.432) (0.452) (0.450) (0.443) (0.447) (0.475) (0.474) (0.457)
Liquid Assets To Assets 0.104∗∗∗ 0.095∗∗∗ 0.082∗∗∗ 0.105∗∗∗ 0.117∗∗∗ 0.103∗∗∗ 0.087∗∗∗ 0.116∗∗∗
(0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011)
Leverage Ratio -0.191∗∗∗ -0.153∗∗∗ -0.210∗∗∗ -0.176∗∗∗ -0.151∗∗∗ -0.116∗∗ -0.191∗∗∗ -0.152∗∗∗
(0.050) (0.054) (0.049) (0.048) (0.051) (0.054) (0.050) (0.048)
ALLL to Total Loans 0.396∗ 0.432∗ 0.388 0.361 0.228 0.300 0.268 0.218
(0.236) (0.240) (0.253) (0.242) (0.237) (0.244) (0.263) (0.244)
Cases Per 100k 0.170 0.222∗∗ 0.324∗∗∗ 0.246∗∗ 0.224∗∗ 0.274∗∗ 0.385∗∗∗ 0.286∗∗∗
(0.107) (0.110) (0.107) (0.103) (0.113) (0.116) (0.112) (0.108)
Constant -3.685∗∗ -13.286∗∗∗ 4.897∗∗∗ -0.926 -13.452∗∗∗ -23.241∗∗∗ -2.622 -8.684∗∗∗
(1.586) (2.223) (1.661) (2.317) (2.211) (2.670) (2.312) (2.783)
Observations 3,513 3,513 3,513 3,513 3,057 3,057 3,057 3,057
Adjusted R2 0.155 0.111 0.168 0.197 0.195 0.133 0.184 0.230
Notes: Dependent variable is PPP loans as a share of total loans in 2020:Q3. Regressor balance sheet variables are measured as four quarter averages
from 2019. COVID cases are county level case counts averaged over counties where the bank operates a branch according to the Summary of Deposit
data. Daily county-level COVID case counts are drawn from John Hopkins.
t statistic in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
99
Table J.9: Bank Outcomes in Q4 2020 IV Regression: Employment share in
COVID-affected industries
(1) (2) (3) (4) (5)
NIM dNIM C&I Gwth Non-PPP C&I Gwth CRE Gwth
PPP Loans to Total Loans 0.109∗∗∗ -0.850 16.292∗∗∗ 1.155 1.053∗
(0.023) (1.449) (1.966) (0.811) (0.630)
ln Assets -0.169∗∗∗ 7.589∗∗∗ -2.956 -2.022∗∗∗ 0.244
(0.023) (1.363) (2.006) (0.713) (0.557)
CI to assets -0.028∗∗∗ -0.637 -7.238∗∗∗ -0.429 -0.087
(0.008) (0.485) (0.705) (0.263) (0.201)
Leverage Ratio -0.005 -0.859∗∗ 1.161∗∗ 0.526∗∗∗ 0.531∗∗∗
(0.006) (0.366) (0.551) (0.199) (0.144)
Liquid Assets To Assets -0.032∗∗∗ -0.579∗∗∗ -0.847∗∗∗ -0.110 -0.061
(0.002) (0.130) (0.198) (0.075) (0.056)
ALLL to Total Loans -0.009 -3.853 -6.932∗∗∗ -3.527∗∗∗ -1.614∗∗
(0.032) (2.389) (2.190) (0.744) (0.730)
ROA 0.152∗∗∗ -18.039∗∗∗ -1.151 -2.009∗∗ -2.230∗∗∗
(0.054) (2.822) (3.229) (0.882) (0.752)
Cases Per 100k 0.018 -0.186 0.907 0.107 0.207
(0.014) (0.914) (1.159) (0.487) (0.376)
Constant 5.653∗∗∗ -80.332∗∗∗ 78.078∗∗∗ 23.369∗∗∗ -2.281
(0.247) (15.149) (20.413) (7.155) (5.467)
Observations 3,518 3,518 3,518 3,518 3,518
Adjusted R2 -0.517 0.114 0.392 -0.055 -0.042
Notes: Instrumental variable is employment share in COVID-affected industries. Sample is 2020:Q2 and 2020:Q3.
Regressor balance sheet variables are measured as four quarter averages from 2019. CRA origination share is calculated
using 2019 origination volumes reported on the banks CRA disclosure. COVID cases are county level case counts
averaged over counties where the bank operates a branch according to the Summary of Deposit data. Daily county-level
COVID case counts are drawn from John Hopkins. COVID-affected employment share is employment in industries that
underwent the largest decline in employment averaged over counties where the bank operates a branch according to the
Summary of Deposit data. County-level employment share in COVID-affected industries is obtained from the QCEW
databse of the Bureau of Labor Statistics.
t statistic in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
100
Table J.10: Bank Outcomes IV Regression: Effect of PPP share on outcomes
Instrumental Variable (1) (2) (3) (4) (5)
NIM ∆NIM C&I Non-PPP CRE
Gwth C&I Gwth Gwth
Small firm employment share -0.005 -1.334∗∗ 17.258∗∗∗ 0.886∗∗∗ 0.466∗∗
(0.007) (0.553) (0.890) (0.279) (0.200)
Core Deposits To Assets 0.063∗∗∗ -8.488∗∗∗ 4.898∗∗∗ -1.457∗∗∗ -0.199
(0.013) (1.270) (1.626) (0.472) (0.320)
Unused CI Commitments To Assets -0.023∗∗∗ -0.887 6.094∗∗∗ -0.942∗∗∗ 0.322∗∗
(0.006) (0.578) (0.768) (0.205) (0.154)
Notes: Sample is 2020:Q2 and 2020:Q3. Regressor balance sheet variables are measured as four quarter averages
from 2019. CRA origination share is calculated using 2019 origination volumes reported on the banks CRA
disclosure. Small firm share is employment share in firms with less than 500 employees per county averaged
over counties where the bank operates a branch according to the Summary of Deposit data. Employment share
in small firms is obtained from the QWI database of the U.S. Census. Employment share in COVID-affected
industries is obtained from the QCEW databse of the Bureau of Labor Statistics.
t statistic in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
101
K 2020:Q1 C&I Loan Draw Effect
This appendix presents results using C&I loan growth and loans from 2020:Q1. Dur-
ing the onset of the pandemic in the first quarter, many banks experienced large
draws on existing lines of credit. We hypothesize that banks experiencing greater
draws would have been more active in the program because firms may have returned
any precautionary draws after receiving the PPP funds. In that way, the PPP helped
to reduce credit risk to the banks by transferring the default risk from their own cap-
ital to the government balance sheet.
Table K.1 shows the impact of these draws on participation, intensity, and the
change in net interest margins. The results for participation and intensity are qual-
itatively similar with the most statistically important effect of loan draws occurring
on the intensity of participation in the program. Moreover, the change in net inter-
est margins was larger compared to banks that experienced less C&I loan growth.
This effect is statistically important for banks that experienced the highest C&I loan
growth impacts.
102
Table K.1: Effect of 2020:Q1 C&I Draws on participation, intensity, and ∆NIM
Participation PPP Intensity ∆NIM (participants) ∆NIM (non-participants)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
CI gwth2020:Q1 0.001 0.009 0.07 0.022
[0, 0] [0.01, 0.01] [0.04, 0.1] [-0.03, 0.08]
CI gwth75th 0.091 1 3.32 1.204
[0.03, 0.16] [0.64, 1.36] [0.58, 6.11] [-3.6, 6.07]
CI U sage2020:Q1 0.005 0.036 0.159 0.103
[0, 0.01] [0.01, 0.06] [-0.05, 0.38] [-0.36, 0.57]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC draws with a burn-in of
5000.
103
Tables K.2 and K.3 report the results on total C&I loan growth and non-PPP
C&I growth. These specifications show that PPP participants that experienced the
largest C&I loan growth in the first quarter had more total C&I loan growth and
more non-PPP loan growth during the subsequent quarters the PPP was active.
104
Table K.2: Effect of 2020:Q1 C&I Draws on participation, intensity, and C&I growth
Participation PPP Intensity CI Gwth (participants) CI Gwth (non-participants)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
2020:Q1
CI gwth 0.002 0.008 0.275 0.027
[0, 0] [0.01, 0.01] [0.24, 0.31] [-0.02, 0.07]
CI gwth75th 0.113 1.108 15.59 4.687
[0.05, 0.17] [0.76, 1.45] [12.14, 19.01] [0.57, 8.79]
2020:Q1
CI U sage 0.005 0.032 0.642 -0.72
[0, 0.01] [0.01, 0.06] [0.35, 0.94] [-1.07, -0.38]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC draws with a burn-in of 5000.
105
Table K.3: Effect of 2020:Q1 C&I Draws on participation, intensity, and non-PPP C&I growth
Participation PPP Intensity Non-PPP CI Gwth (participants) Non-PPP CI Gwth (non-participants)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
CI gwth2020:Q1 0.001 0.008 0.218 0.077
[0, 0] [0.01, 0.01] [0.2, 0.23] [0.04, 0.12]
CI gwth75th 0.103 1.081 14.498 5.154
[0.04, 0.16] [0.74, 1.42] [12.94, 16.03] [1.12, 9.17]
2020:Q1
CI U sage 0.005 0.037 0.383 -0.685
[0, 0.01] [0.01, 0.06] [0.28, 0.49] [-1.04, -0.33]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC draws with a burn-in of
5000.
106
Finally, Table K.4 reports the results for commercial real estate lending (CRE)
as a check on spillover effects. We find mixed evidence that participants in the PPP
program made more CRE loans. In at least one specification, the sign is negative
and statistically unimportant. However, in the other specifications we find positive
and statistically important effects. Notably, for non-participants we find negative
impacts of first quarter C&I loan growth on CRE lending, suggesting that the cap-
ital protection that PPP provided may have encouraged some additional, non-C&I
lending for the most active C&I lenders in the first quarter.
107
Table K.4: Effect of 2020:Q1 C&I Draws on participation, intensity, and non-PPP CRE growth
Participation PPP Intensity CRE Gwth (participants) CRE Gwth (non-participants)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
CI gwth2020:Q1 0.000 0.010 0.013 -0.018
[0, 0] [0.01, 0.01] [0, 0.02] [-0.05, 0.01]
CI gwth75th 0.083 0.998 2.068 -2.803
[0.02, 0.15] [0.65, 1.35] [0.9, 3.24] [-5.84, 0.18]
CI U sage2020:Q1 0.005 0.034 -0.048 -0.028
[0, 0.01] [0.01, 0.06] [-0.13, 0.03] [-0.26, 0.21]
Note: The reported values are posterior means of the parameters, and 95% credibility intervals in brackets. The results are based on 55,000 MCMC draws with a burn-in
of 5000.
108
File and source
- File
- government-loan-guarantees-during-a-crisis-the-effect-of-the-ppp-on-bank-lending-and-profitability.pdf
- Size
- 783,147 bytes
- SHA-256
- 4aa7a4834c81e0368659f861c09752d74c362f68389ee33064c712aea0d26e44
- Original
- doi.org