The Main Street Lending Program: Who Borrowed and How Have They Benefited? (J. Christina Wang, No. 22-24)
Summary
Federal Reserve Bank of Boston research department working paper No. 22-24, The Main Street Lending Program: Who Borrowed and How Have They Benefited?, by J. Christina Wang, version dated September 2022. The abstract states that the Main Street Lending Program (MSLP) was established by the Federal Reserve to supply credit to small and especially midsize businesses, and that the study uses Dun & Bradstreet data to examine the characteristics of MSLP borrowers and their performance after receiving a loan. It states that a firm was more likely to borrow if it was larger, more active, hit harder by the pandemic or located in a county with fewer community banks, among other factors. It also states that receiving an MSLP loan improved firms' financial condition but did not lead to significant increases in employment over the following year.
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
- The Main Street Lending Program Who Borrowed And How Have They Benefited
- Case
- The Main Street Lending Program Who Borrowed And How Have They Benefited
Cite this document
“The Main Street Lending Program: Who Borrowed and How Have They Benefited? (J. Christina Wang, No. 22-24).”
Original at doi.org. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/the-main-street-lending-program-who-borrowed-and-how-have-they-benefited/
, SHA-256 dea15adb46abc906316fb505a225a068e5249876f2343c467b5caa1b9e1735c8.Full text
No. 22-24
The Main Street Lending Program: Who Borrowed and
How Have They Benefited?
J. Christina Wang
Abstract:
The Main Street Lending Program (MSLP) was established by the Federal Reserve to supply
credit to small and, especially, midsize businesses so they could weather COVID-19–induced
disruptions. This study uses Dun & Bradstreet (D&B) data on the financial condition and overall
viability of firms to examine the characteristics of MSLP borrowers and their performance after
receiving a loan relative to the performance of their peers. Estimates show that, even when
differences in firms' industries and geographic regions are taken into account, a firm was more
likely to borrow from the MSLP if it was larger, more active, had a good but not excellent risk
score, was hit harder by the pandemic, had received a Paycheck Protection Program (PPP) loan
early but was located in a county with a longer delay in PPP lending, operated in a nonessential
industry, or was located in a county with fewer community banks. A nontrivial fraction of MSLP
borrowers also received second-draw PPP loans in 2021, indicating that they were, in fact, in
need of funding. Receiving an MSLP loan improved firms' financial condition progressively and
significantly on average, even though it did not lead to significant increases in employment over
the year following loan receipt.
JEL Classifications: H81, G28, J21
Keywords: COVID-19, Main Street Lending Program, SMEs, credit frictions
J. Christina Wang is a senior economist and policy advisor in the research department of the Federal Reserve Bank of
Boston. Her email address is Christina.Wang@bos.frb.org.
The author thanks Falk Bräuning, José Fillat, María Luengo-Prado, and Joe Peek for helpful discussions and
comments. She gives special thanks to Gustavo Joaquim for deep discussions and county-level data. The author also
thanks Kelly Jackson, Calvin Zhang, and especially Melanie Qing for exceptional research assistance. And she thanks
David Brown, Mike Corbett, Greg Longfield, and especially Catherine Spozio for assistance with borrower name
matching.
This paper presents preliminary analysis and results intended to stimulate discussion and critical comment. The views
expressed herein are those of the authors and do not indicate concurrence by the Federal Reserve Bank of Boston, the
principals of the Board of Governors, or the Federal Reserve System.
This paper, which may be revised, is available on the website of the Federal Reserve Bank of Boston at
https://www.bostonfed.org/publications/research-department-working-paper.aspx.
This version: September 2022 https://doi.org/10.29412/res.wp.2022.24
1 Introduction
As the COVID-19 pandemic and the public health measures implemented in response wreaked
havoc on the US economy, Congress authorized the Paycheck Protection Program (PPP) to
provide liquidity to small businesses through low-cost loans that essentially turned into
grants. At the same time, the Federal Reserve launched unprecedented corporate bond pur-
chase facilities, which proved to be effective liquidity backstops that stabilized the funding
supply to large corporations. By comparison, midsize firms were initially in danger of falling
through the cracks: They tended to be too small to access the corporate bond market and
thus still bank-dependent, but some were too large to borrow under the PPP. The Federal
Reserve therefore established the Main Street Lending Program (MSLP or Main Street) on
April 9, 2020, to provide targeted credit support for these mid-market businesses as well as
nonprofits and any small businesses that might need additional funding beyond what they
could obtain from the PPP.
The MSLP was by far the most challenging of the Fed credit programs to administer
because it “purchases interests in loans that are, by nature, bespoke agreements often with
complex, borrower-specific terms and conditions,” as Eric Rosengren, then president of the
Federal Reserve Bank of Boston, noted in his congressional testimony on August 7, 2020.
This inherent complexity makes it important to understand why certain firms became bor-
rowers whereas other observationally comparable firms, such as those of similar size and
operating in the same industry and the same geographic region, did not. In particular, were
the borrowers more risky than other firms (controlling for size, industry, and location)? Did
their financial situation deteriorate more over the few months after the COVID-19 pandemic
hit and before they received an MSLP loan? As importantly, to what extent did the MSLP
loans help these borrowers’ performance, such as showing relatively greater improvement in
financial health?
Such an analysis is difficult because virtually all of the firms that were eligible for the
MSLP, along with their peers, are private companies about which limited information is
available. We therefore turn to the Dun & Bradstreet (D&B) database, which offers the
most comprehensive coverage of private US firms, especially those firms that have only a few
employees.
The MSLP operated by purchasing participations in bank loans made to eligible borrow-
ers, so whether a firm became an MSLP borrower was jointly determined by the firm’s need
for credit and a bank’s decision to extend credit but only under the MSLP. For this outcome
to be the bank’s optimal decision, intuition suggests that the firm would more likely than
not have been sufficiently viable to benefit from the injection of this funding but too risky to
2
qualify for a private loan of comparable terms. Fully characterizing this joint optimization
problem is clearly challenging and beyond the scope of this paper.
Instead, this study focuses on two aspects of this problem. It (1) examines which firm
attributes, along with local economic and public health conditions, were associated with a
higher likelihood that a firm borrowed from the MSLP, and (2) estimates how much the
MSLP aided the borrowers, especially in shoring up their financial condition after they
received a loan measured relative to the condition of their peer companies that had the
same observable characteristics (and thus were also eligible for the MSLP based on the scale
of their operation) but did not borrow. The answer to the first question can help evaluate
whether the program reached the subset of businesses it was designed to target. Perhaps more
importantly, it can provide useful clues about which program parameters might be altered
to make it more suitable should a similar program be called for in the future. Answers to the
second question serve as an evaluation of the efficacy of the MSLP in providing liquidity to
alleviate financial strains on firms, which should subsequently help to support the recovery
of real economic activity.
First, we find that, as intended by the design of the program, being midsize (that is,
employing 250 to somewhat more than 500 workers) was the strongest predictor of a firm
becoming an MSLP borrower. Firms with fewer than 10 employees were the least likely
to borrow by a notable margin, consistent with the feedback that many firms considered
the certification requirements too onerous. Interestingly, being an early recipient of a 2020
PPP loan increased the odds that a firm would borrow from the MSLP by a smaller but
comparable magnitude, indicating the importance of preexisting banking relationships. It
also suggests that these firms needed more funding than provisioned under the PPP. This
conjecture is supported by the fact that a substantial subset of MSLP borrowers also received
second-draw PPP loans in 2021, which further indicates that these firms suffered relatively
large disruptions due to the pandemic.
By comparison, the influence of risk scores on the probability of borrowing was much more
modest. Consistent with intuition, firms with good risk scores were the most likely to become
borrowers, while those with the best scores were the least likely. Having experienced a larger
improvement in risk scores over the year before the pandemic but a greater deterioration in
risk scores after its onset was also associated with a stronger propensity to borrow. This
indicates that borrowers tended to be firms that had better prospects but were more adversely
affected by the pandemic. Local economic conditions mildly influenced the likelihood of
MSLP borrowing. Greater economic vibrancy in a firm’s county (proxied by a higher mobility
index) and a larger presence of community banks in that county were associated with a lower
likelihood of borrowing, as was operating in an essential industry. A longer delay in PPP
3
lending in a firm’s county raised the likelihood of a firm becoming an MSLP borrower. These
relationships are largely consistent with intuition.
Our subsequent estimates of the MSLP’s treatment effects show that the financial health
of recipient firms improved progressively and significantly over the year following loan receipt.
The reduction in credit risk reached nearly 18 percent four quarters after loan receipt when
estimated using a staggered treatment effects estimator à la Callaway and Sant’Anna (2021).
This suggests that the MSLP did achieve its objective of shoring up the liquidity position
of firms that had promising prospects but were hit hard by the pandemic. By comparison,
MSLP borrowers experienced slightly better recovery in employment over the year following
loan receipt relative to their peers, but the difference was insignificant. This may be due in
part to the imprecision of the employment data. Moreover, full recovery may require more
time. To the extent that firms with more ample liquidity are better able to realize their
real-growth goals, these borrowers should be in a good position to expand operations as the
economy fully normalizes over the coming quarters.
Our analysis complements several existing studies of the MSLP. Arseneau et al. (2021)
offer a comprehensive review of the MSLP—its goals and design, the characteristics of its
borrowers and lenders, etc.1 Bräuning and Paligorova (2021) find that the amount of credit
extended by the MSLP reached about 60 percent of the volume of loans made by large banks
to borrowers of comparable size and leverage. Bräuning et al. (2021) present evidence that
the MSLP accomplished its key goal of directing more funds where and when they were most
needed. Minoiu, Zarutskie, and Zlate (2022) find a positive spillover effect of the MSLP on
participating banks’ commercial and industrial (C&I) lending more generally, which, they
argue, is because the MSLP served as a backstop. Unlike these previous studies, this paper
is able to use data on MSLP borrowers and their peers to explore the treatment effects at
the individual firm level.2
2 The Main Street Lending Program
Given the extensive reviews of the MSLP by studies such as Arseneau et al. (2021), we
discuss only its key features pertaining to borrowers that are relevant for our analysis. First,
only for-profit businesses that obtained MSLP loans are analyzed in this study.3 This study
1
Morgan and Clampitt (2021) provide an earlier review of the MSLP’s main features and ultimate volume.
2
It is closely related to Wang, Ballance, and Qing (2021), which also uses D&B data to understand the
characteristics (particularly the financial condition) of MSLP borrowers relative to those of their peers before
and after the COVID-19 outbreak in 2020.
3
The Federal Reserve also established facilities within the MSLP to support nonprofit organizations.
Those facilities, however, received only a few submissions and thus are omitted from this study.
4
compares borrowers with size-eligible non-borrowers, which are businesses with as many as
15,000 employees in the D&B database.4
Leverage limits set by the MSLP, which differ across the three loan facilities (New, Pri-
ority, and Expanded), cannot be considered because D&B provides no financial variables.5
Instead, risk scores are used to gauge firms’ riskiness. See Appendix B for more details about
the Main Street program design.
3 Data Description
The primary data source for our study is the Dun & Bradstreet (D&B) database. It is the
most comprehensive commercial database of private businesses in the United States.6 The
full “population” for our analysis includes all the US-domiciled employer establishments in
the D&B database that were active as of January 2019 and employed one to 15,000 workers
(which is the MSLP size limit).7 Establishments with missing state and industry codes are
also excluded, although they account for less than 1 percent of the firms.
D&B is aptly regarded as a credit bureau for businesses. It assigns credit scores to
nearly all the businesses in its database based partly on their payment records. These
scores are constructed to predict a firm’s future credit performance (such as making timely
repayment) and serve a role in business lending analogous to that of consumer credit scores
in consumer lending. D&B scores are generally considered the leading scores used by a
broad range of creditors (lenders and suppliers). For example, the primary credit score
used by Bank of America for approving small-business loans is provided by D&B.8 D&B
updates the risk scores on an ongoing basis, based on the stream of signals it receives from a
network of information providers (including landlords, lenders, utility companies, suppliers,
postal services, secretaries of state, etc.). These risk scores thus enable us to adequately
account for businesses’ pre-pandemic financial health and commercial viability as perceived
by lenders. Given that lenders had to retain 5 percent of the principal on an MSLP loan,
4
According to the Main Street term sheets, businesses with no more than $5 billion in revenue were also
eligible, but revenue is sparsely covered in the D&B data.
5
Leverage is capped at four times adjusted 2019 earnings before interest, tax, and depreciation (EBITDA)
for loans purchased by the New Loan Facility (MSNLF), and six times EBITDA for the Priority and Ex-
panded Loan Facilities (MSPLF and MSELF).
6
For years now, every government contractor has been required to obtain a D&B identifier to be eligible
to bid for contracts.
7
The two-employee (equivalent to one-worker) lower bound is needed to obtain a sample of employer firms
because D&B tends to count the proprietor as an employee (see Barnatchez, Crane, and Decker, 2017). This
threshold also excludes sole proprietorships and establishments with zero reported employees.
8
https://www.bankofamerica.com/smallbusiness/education/business-credit-score/. Many digital finan-
cial services platforms (such as Credit Karma and Brex) highlight the importance of D&B credit scores,
especially for small businesses.
5
a business’ creditworthiness would have been an important factor in a bank’s decision on
whether to issue a loan and whether to seek the Fed’s participation in the loan. Furthermore,
it is likely that the most visible near-term effect of the liquidity injection through an MSLP
loan was to enable the borrowing firm to make more timely payments on its liabilities (owed
to suppliers and creditors).9 This should have, in turn, shored up a firm’s credit score, all
else being equal. The D&B data thus enable us to examine the impact of the MSLP on
businesses’ bill payment records and the evolution of their risk scores. Greater financial
resources should, in principle, provide the wherewithal for businesses to expand their real
activity once demand has recovered.
Given the tight near-term link between credit availability and firms’ ability to pay bills
punctually, we focus our analysis of firm risk on the D&B Commercial Credit Score (CCS).
CCS is modeled to measure the risk of delinquency in the next 12 months.10 Scores range from
101 to 670, with each 40-point increase halving the risk of delinquency. When interpreting
our estimates, we use the following formula to convert a change in the CCS to a change
in the delinquency probability: %∆Risk = exp{−ξ · ∆CCS} − 1, where ξ ≡ ln(2)/40. For
example, ∆CCS = 12 maps to an approximately 18 percent reduction in risk. See Appendix
A for more details about the risk scores and ratings, along with comparable statistics of the
other risk scores.
In analyzing firms’ MSLP uptake decisions as well as estimating the MSLP’s treatment
effects, we also consider the 2019 average D&B Viability Score (VS) bin (which smooths
out monthly fluctuations in a small fraction of observations) as a control variable in some
specifications. By definition, the VS provides a more comprehensive measure of a firm’s
overall commercial viability prior to the onset of the pandemic. VS bins are analogous to
credit rating bins and likely more commonly used than the more granular VS Points by
creditors in making lending decisions.11
9
This effect was partly indirect since, with a few exceptions (see Appendix B for details), borrowers had
to “refrain from repaying the principal balance of, or paying any interest on, any debt” until the MSLP loan
was repaid in full. But the MSLP loan would substitute for internal funds that could then be directed to
other uses. To this end, the deferral of interest (for one year) and principal repayment (for two years) further
helped.
10
Schematically speaking, D&B estimates the CCS using its proprietary model of firm payment outcomes
on past payment behavior and an array of firm attributes (such as industry, location, age, size, etc.). The
CCS thus predicts default probability in the future given data to date and can be regarded as the forward-
looking counterpart to the better known D&B Paydex score, which is simply a summary statistic of a firm’s
past payment behavior.
11
VS Points is the more continuous counterpart to VS, ranging from 101 to 800. The relationship between
VS Points and VS bins is similar to the relationship between consumer credit scores and consumer risk profile
categories, an example of which (according to the Consumer Financial Protection Bureau) consists of five
categories: deep subprime, subprime, near-prime, prime, and super-prime. Lenders tend to treat consumers
who have scores in the range of a given category as equally risky.
6
Although the D&B data are available monthly, analyses here, especially of the MSLP
treatment effects, are mostly conducted at the quarterly frequency, because quarterly aver-
ages smooth out occasional transitory fluctuations from month to month in some risk-score
variables. The event time is defined according to the timing of the MSLP operations: Event
quarter τ = 0 in our analysis corresponds to 2020:Q3, as the MSLP began operations on July
6, 2020, while the majority of loans were made in τ = 1 because of the surge in participation
in early December 2020. τ = −1 corresponds to 2020:Q2, coinciding with the initial acute
phase of the COVID-19 pandemic. Overall, our sample runs from 2019:Q1 (τ = −6) through
2021:Q4 (τ = 6).
Because the volume of the D&B data is impractically large, we conduct all the analysis
at the firm level using a representative random sample drawn by strata of geography (state),
industry (two-digit NAICS code), and size (range of employee count) from the population of
US-domiciled firms in the D&B database that were active as of January 2019 and employed
one to 15,000 workers.12 Firm attributes are measured either as of March 2020, our baseline
pre-pandemic period, or as 2019 averages to minimize idiosyncratic monthly noise for more
volatile variables.13 The final sample comprises 418,995 firms for the uptake propensity score
regressions and from 172,000 to 179,000 matched peers for the MSLP firms for the dynamic
treatment effects analysis.
We use several additional data sources to estimate the factors that influenced MSLP
uptake. First, to measure the de facto level of social distancing and commercial activity
in each county, we use the mobility indexes from Google’s COVID-19 Community Mobility
reports.14 In addition, county-level COVID-19 statistics (deaths per million) are used to
account for pandemic severity.15 We use two indicators to measure the type and degree
of impact from the pandemic by industry. The first indicates whether an industry was
designated as essential, which would have allowed it to operate during a lockdown, according
to the Cybersecurity and Infrastructure Security Agency.16 The other involves classifications
for the six-digit NAICS industries by the Chicago Fed indicating which were subject to
12
More precisely, the unit of analysis is “enterprise,” according to D&B—either standalone or headquarters
(but not branch)—which we refer to as “firm” for short. Some variables, such as risk scores, are measured
only at the firm level.
13
As explained in Wang, Ballance, and Qing (2021), given the typical lag of updates in D&B data, March
2020 was the last pre-pandemic month, although using February 2020 values makes virtually no difference.
Moreover, a firm had to be established by March 13, 2020, to be eligible for a Main Street loan.
14
See Google COVID-19 Community Mobility Reports, https://www.google.com/covid19/mobility/,
obtained from Opportunity Insights (see Chetty et al., 2021). These indexes were compiled using mobile
phone tracking data, and they measure percentage changes in the number of visitors to and lengths of stay
at categorized places relative to the baseline period of January 2 through February 6, 2020.
15
The data were compiled by the New York Times and also obtained from Opportunity Insights.
16
Data downloaded from https://www.cdc.gov/niosh/topics/coding/essentialworkers/.
7
moderate, severe, substantial, or basically no impact from the COVID-19 shock.
Since essentially all lending to (small) private businesses is local, to measure such firms’
access to funding, we use the share of bank branches in a county that belongs to commu-
nity banks. We computed the shares using the FDIC’s 2019 annual Summary of Deposits
(SOD) data, which provide the location of all branches (and deposit balance by branch) of
all depository institutions that were operating in the United States as of June 2019. We
aggregate the branch-level data from the SOD to the county level. Community banks are
identified according to the FDIC’s institution directory. This ratio was also found to be
associated with more rapid access to the PPP by small businesses, as community banks were
more active in PPP lending, especially early on (see Faulkender et al., 2021).
Given the unprecedented fiscal support provided to small businesses and households
following the onset of the pandemic, the amount of public assistance received by each locality
(relative to its pre–COVID-19 economic activity) should be accounted for when evaluating
the effects of the MSLP. We use the volume of PPP loans received by small businesses
within a county as a summary measure of the degree of public-sector assistance delivered
to that county. This amount is normalized by the 2019 average monthly payroll of eligible
firms (that is, those with 500 or fewer employees), which is derived using data from the
US Census Bureau’s County Business Patterns (CBP 2020) combined with data from the
bureau’s Statistics of US Businesses (SUSB 2017).17 The PPP data are from the July 2021
data release by the Small Business Administration. These data are also used to identify
which MSLP borrowers also obtained PPP loans.
Identifying MSLP Borrowers in the D&B Database
D&B’s name-matching algorithm is the primary tool for identifying the MSLP borrowers in
the D&B database by locating their Data Universal Numbering System (DUNS) numbers.
Of the accepted MSLP borrowers recorded as of January 8, 2021, we are able to find with high
confidence the DUNS number at the headquarters level for a total of 1,781 borrower firms.
However, we lose around 270 borrowers from the statistical analysis either because there are
no pre-pandemic records for them in the database or they lack peers (in the same state,
industry, and age and size bins) needed to compute the treatment effects. See Appendix B
for further details.
Defining Matched Peers for MSLP Borrowers
Because the MSLP borrowers operate at different scales and in different industries and
17
We use 2019 payroll expenses for normalization to be consistent with the PPP rule that capped the size
of a loan at 2.5 months of a firm’s payroll or $10 million.
8
localities, it is important to compare these firms with their suitably matched peers to accu-
rately assess the “treatment effects” of the program. Industry, locality, age, and size are the
dimensions considered in most studies. Here, industry is classified by two-digit NAICS code
and locality by state.18 The employment size bins are defined as follows: (1) as many as 10
employees, (2) 11 to 50 employees, (3) 51 to 250 employees, (4) 251 to 500 employees, (5)
500-plus employees. Firm age bins are defined as follows: (1) less than 2 years old in 2020
(that is, in the database by March 2020 and started no earlier than 2019), (2) 2 to 10 years
old (that is, started between 2011 and 2018), and (3) started no later than 2010.
4 Why Did Firms Borrow from the MSLP?
This section examines the factors, including firm characteristics and economic conditions,
that influenced whether a firm obtained an MSLP loan. To set the stage, we first compare the
pre-pandemic characteristics of MSLP borrowers with those of firms that were also eligible
(based on employee count) but did not borrow. One notable difference is that the MSLP
firms were much larger than the other firms (with an average of 102 employees versus 8,
respectively), even when differences in location (state), industry (two-digit NAICS), firm-
age bin, and the small-business indicator (defined as up to 500 employees) are accounted
for, as is evident in the Appendix Table B.2. MSLP firms’ higher average employee count
was driven by the substantially larger size of those in the top fifth percentile. Nevertheless,
the majority of these firms had fewer than 500 employees and thus would have qualified for
the PPP. The MSLP firms were also somewhat more established (that is, older by about
a year and a half) than the other firms. By contrast, the unconditional distribution of
MSLP-borrower risk scores is highly similar to that of non-borrowers.
We next formally test whether MSLP firms had the same risk scores and payment records,
on average, as non-borrower firms in terms of both pre–COVID-19 levels (that is, as of March
2020) and the changes over the months (March through June 2020) after the pandemic’s
onset until the MSLP began operations on July 6, 2020.19 Table 1 reports these comparisons,
equivalent to differences in means, for risk scores, payment records, and employment between
MSLP borrowers and all the non-borrowers (in the random sample). Each set of mean
comparisons is conditioned on several commonly used key firm attributes. For example, the
differences in means reported in column (1) are computed among firms in the same state
18
These two dimensions became particularly relevant for this episode because the pandemic struck certain
industries more severely than others, and social-distancing rules were imposed to differing degrees and at
different times across states. Two-digit NAICS codes are used to preserve the sample size of each peer group.
19
As detailed in Wang, Ballance, and Qing (2021), it makes virtually no difference whether February or
March 2020 values are used to measure firms’ pre–COVID-19 condition.
9
as well as the same industry (two-digit NAICS), whereas the mean differences in column
(4) are conditioned on all the attributes (that is, being in the same state, industry, age bin,
employment-size bin, and 2019 average Viability Score bin, and being a PPP borrower or
not).20
The MSLP firms had significantly higher CCS scores in March 2020 compared with non-
MSLP firms in the same state, industry, age bin, and employment-size bin, with a 17-point
higher score mapping into nearly 30 percent lower delinquency risk over the next 12 months.
This differential becomes insignificant once the MSLP firms are compared with non-borrowers
that are also in the same 2019 average Viability Score bin. The score difference disappears
entirely if the comparison is further conditioned on being a PPP borrower or not. Notably,
even when conditioned along all these dimensions, the MSLP borrowers still exhibited a sig-
nificantly larger decline in the CCS after the onset of the pandemic and before the start of the
MSLP. This is consistent with the finding that MSLP firms’ delinquency records (percentage
of debt 31-plus days past due) in March 2020 were slightly better than those of other non-
borrower firms when conditioned on all the attributes, but MSLP firms clearly fell behind
on paying bills more than non-borrowers over the March–June 2020 period.21 Likewise, as of
March 2020, MSLP firms were much more viable overall, according to the Viability Scores,
compared with non-borrower firms in the same state and industry, but that advantage more
or less disappears once the comparison is further confined to firms in the same age bin,
size bin, etc. MSLP firms showed slightly worse, albeit largely insignificant, deterioration in
Viability Scores from March to June 2020 when compared with non-borrowers in the same
state, industry, age bin, size bin, etc.
By comparison, MSLP firms displayed clearly greater financial distress risk (that is,
a lower average Financial Stress Score (FSS)) before the pandemic compared with non-
borrower firms matched along all six dimensions (same state, industry, etc., in column (4)
of Table 1) but did not experience worse deterioration in their FSS after the COVID-19
outbreak. One likely explanation for MSLP borrowers’ worse FSS is the finding reported
in Arseneau et al. (2021) that these firms had higher pre–COVID-19 leverage than other
comparable borrowers of the Y-14 banks (that is, those banks that have than $100 billion in
total consolidated assets and are subject to the Dodd-Frank Act Stress Tests).
The essentially same overall pattern is confirmed (in Table B.3) if we further restrict the
comparison group for each MSLP borrower to only those non-MSLP firms that are in the
20
Specifically, the mean differences reported in column (1) are computed by first removing the mean value
of each variable by state and then by industry.
21
As defined, the CCS should be most correlated with delinquency records. This is confirmed for the
MSLP firms in Wang, Ballance, and Qing (2021) and more generally for all the firms with delinquency data
in Joaquim and Wang (2022).
10
same state and the same two-digit NAICS industry and have the same small-business status
(that is, employ no more than 500 workers). With these matched peer firms, it is more
natural to also compare the change in risk scores from March 2020 to the month when a
borrower’s loan was submitted to the MSLP; it is conceivable that the financial condition of
MSLP firms that obtained their loans in later months continued to weaken beyond June 2020
relative to the condition of their peers. This conjecture is, in fact, borne out in the data:
MSLP borrowers experienced significantly more severe deterioration in their risk profile and
payment records over the months after COVID-19 broke out and before they obtained their
MSLP loans (see Table B.3).
In sum, these estimates indicate that, on average, MSLP borrowers were in fairly sound
condition on the whole before the pandemic, but they were later more adversely affected
by the COVID-19 shock than peer firms comparable in terms of the attributes considered
in most studies. In particular, the contrast between MSLP firms’ FSS and their other two
risk scores suggests these firms were commercially viable and able to service their liabilities
in a timely manner, but perhaps they were able to finance these expenditures by being
more heavily indebted. The cessation of income due to the pandemic would then be more
detrimental to these firms, which is consistent with the observed greater deterioration in
their risk scores after the COVID-19 outbreak. These findings for nearly all the MSLP firms
confirm the findings presented in Arseneau et al. (2021) for a small subset of MSLP firms
that also borrowed from the Y-14 banks.22 This overall pattern is intuitively consistent with
the outcome of these firms borrowing from the MSLP: Each was deemed adequately viable
to be granted a bank loan but also sufficiently risky that it needed assistance from the MSLP
to obtain better terms and perhaps a larger loan.
4.1 Which Factors Predict Firms’ MSLP Uptake?
We now formally estimate which factors help predict whether a firm became a Main Street
borrower. The patterns uncovered above imply that risky and more adversely affected but
still fundamentally sound firms were more likely to seek credit from the MSLP. Guided by
this conjecture, intuition suggests that the pandemic wreaked more havoc on these firms in
part because of the nature of their operations, as some industries (especially those that rely
more on in-person contact, such as leisure and hospitality) suffered more COVID-19–induced
disruptions. In the initial phase of the pandemic, normal operations at firms outside the
essential industries were relatively more disrupted because of government lockdown orders.
22
Arseneau et al. (2021) find that Main Street borrowers had higher leverage and lower bank internal credit
ratings than other comparable Y-14 borrowers, and their ratings declined more rapidly after their loans were
originated.
11
We thus include indicators that measure whether a firm’s industry was deemed essential
or especially susceptible to COVID-19–induced disruptions. In addition, a firm’s ability to
operate could also be undermined by worse local public health conditions and the resulting
restraint on economic activity from either mandatory containment measures or voluntary
social distancing. We thus also consider indicators of pandemic severity and a broad measure
of mobility, which captures the actual level of commercial activity (relative to the pre-
pandemic baseline in January and February 2020), whether it was the result of a mandate
or voluntary restraint.
Given that many MSLP borrowers were also eligible for the PPP, from which many MSLP
firms obtained loans that were expected to be largely forgivable, we include indicators for
PPP loan recipients depending on when a PPP borrower received that funding. In light of
the analysis by Joaquim and Netto (2021) and Joaquim and Wang (2022), we classify the
PPP borrowers into three groups: early (those that received loans before the initial funding
ran out), middle (those that received loans after newly appropriated funding arrived on April
27 and through May 2, 2020, when the PPP lending volume slowed to a trickle), and late (all
the rest).23 The unprecedented fiscal support provided to small businesses and households
in the wake of the COVID-19 outbreak means that the size of the stimulus received relative
to the pre-pandemic size of the local economy likely also mattered for the overall economic
condition in a locality. We use the total (normalized) cumulative volume of PPP loans
received by July 1, 2020, by small businesses in a county as a summary measure of the
level of public-sector assistance delivered to that county.24 More PPP funding should have
reduced the need for businesses to seek additional financing, controlling for local economic
and public health conditions. To further capture the potentially differential impact of PPP
credit even conditional on the overall volume, we also include the share of PPP funds that
were delayed, as proposed in Doniger and Kay (2021), which relies on the volume of PPP
loans made in the two days right after the exogenous 10-day window of funding delay from
April 17 through April 26, 2020, relative to the two days right before it.25
Moreover, findings by previous studies suggest that the structure of the local banking
market affects private firms’ access to loan financing, particularly the access of small private
firms. We thus include the (deposit-volume-weighted) share of bank branches in each county
belonging to community banks, which enabled PPP funds to be disbursed more rapidly after
23
The indicator for PPP eligibility is omitted because it coincides exactly with the above-500-employee
indicator.
24
This amount is normalized by the average monthly payroll expenses in 2019 because of the PPP rule
that capped the size of a loan at 2.5 months of a firm’s payroll or $10 million.
25
In fact, Joaquim and Wang (2022) show that this delayed share is the only instrumental variable that
retains its explanatory power after firms’ pre–COVID-19 financial health is taken into account.
12
the onset of COVID-19 (as shown, for example, in Faulkender et al., 2021).
In sum, we estimate the following linear probability regression as the main specification:
Yi = α + Xi γ + Si0 β1 + ∆Si β2 + Xj κ + Xc θ + Xppp,i λ + i . (1)
Yi is the binary indicator equal to one if firm i was an MSLP borrower and zero otherwise.
Xi refers to a vector of i’s nonfinancial attributes, including binned values of its employment
and age in March 2020. These and other firm attributes are mostly accounted for by using
binned values to allow for potentially nonlinear effects nonparametrically. Si0 denotes a
vector of dummy variables based on binned values of i’s risk score as of March 2020 or 2019
average. We consider two risk scores in turn: the CCS because of the tight linkage between
a firm’s near-term delinquency risk and credit availability, and the more comprehensive
Viability Score. ∆Si denotes the risk score’s change from March to June 2020, measured by
quartile-based dummy variables. Xj includes industry-based indicators: Essentially equal to
one if i operated in an essential industry, and Impacted Industry equal to one if 75 percent or
more of employment in i’s three-digit NAICS industry was deemed substantially or severely
impacted by COVID.26
Xc denotes the vector of county-level regressors. First is the Google mobility indexes
based on time away from home or time at a workplace (measured as the change relative to
January and February 2020), used to measure the de facto level of commercial activity after
COVID-19 hit but before the MSLP commenced operations.27 Missing county-level data are
proxied with state-level values. Each index’s monthly values from March to June 2020 enter
the regression separately to allow for potentially dynamic effects of changes in mobility in a
locality on firms’ MSLP uptake decisions. Pandemic severity is accounted for by using the
county-level COVID-19 death rates (number of new deaths per 1 million population in each
month). This should be more accurate than case rates, which depend on the extent of testing.
Test rates were low and highly uneven across localities early on in the pandemic. New death
rates in each month from March through June 2020 likewise enter the regression separately.
Xc also includes the 2019 community-bank-branch share in each county (CBshare ).
The linear probability model is estimated using all 1,238 MSLP firms for which there are
data along with the random sample of non-borrowers. Each non-borrower observation is thus
scaled up to represent the full population of slightly more than 25 million firms in the D&B
26
The Impacted Industry is a weighted average indicator that equals one if a six-digit NAICS industry was
classified as subject to severe or substantial impact from the COVID-19 shock, and zero otherwise, summed
up to the three-digit NAICS industry level. The weight is each six-digit NAICS industry’s employment share
within the three-digit NAICS industry.
27
The two indexes at the state level have been found to be the most correlated with MSLP uptake (see
Bräuning et al., 2021). Their values are highly correlated at the county level as well.
13
database in March 2020. Clearly, the mean probability of being a borrower is vanishingly
small, hence the minuscule magnitude of the coefficient estimates. Like the linear model,
Equation (1) is estimated using the rare-event logit regression. See Appendix C for details
and output of the logit estimation. By and large, the logit estimates corroborate the linear
estimates.
Table 2 reports the estimates of the uptake propensity described in Equation (1), with
firms’ financial health measured by their CCS.28 Column (1) considers only the firm financial
indicators, assuming that the lenders cared most about potential borrowers’ credit risk.
Firms with pre–COVID-19 CCS of 300 to 600 (roughly corresponding to the 5t h and 95t h
percentiles, respectively) were more likely (by 3 basis points) to obtain MSLP loans than
firms with lower (the omitted category) or higher CCS; the higher the CCS within this
range, the lower the probability of obtaining an MSLP loan, albeit insignificantly so.29 CCS
changes over the year before the onset of the pandemic are used to gauge whether a firm’s
pre–COVID-19 trajectory influenced MSLP uptake. Firms that had experienced a 20-plus-
point increase or decrease in their CCS (equal to a 25 percent change in delinquency risk) and,
to a lesser degree, those that had seen moderately positive changes showed an additional
propensity to borrow. But firms with moderately negative or zero changes (the omitted
subset) in their CCS did not exhibit such a propensity.
By comparison, large CCS changes (more than 20 points in absolute value) from March to
June 2020 were several times more important in boosting the odds of borrowing, while modest
CCS changes during the pandemic raised the odds by a margin comparable to large pre–
COVID-19 CCS changes. This bimodal pattern of large CCS changes, positive or negative,
being associated with a higher probability of borrowing is likely because large changes are
correlated with the (unobserved) firm property of being more commercially active and thus
receiving larger score updates, as suggested by the correlation between an indicator of large
CCS changes and being a relatively large firm and receiving a PPP loan early in the program’s
operation (see Table B.4 in the Appendix). Overall, however, the CCS-based indicators have
rather small explanatory power.
Column (2) adds employment-bin and PPP-timing indicators, and many of them boosted
the propensity to borrow much more than the CCS-based variables, the coefficients on which
shrink noticeably as a result.30 In particular, having more than 250 employees exerted the
greatest influence, followed by having 51 to 250 employees and receiving a PPP loan early.
28
Table D.1 in Appendix D presents the estimates using the 2019 average Viability Score to measure firm
financial health. The coefficients are highly similar to those from estimates using the CCS.
29
Separate coefficients for more granular bins are omitted for brevity.
30
When the other indicators are added, having an excellent pre–COVID-19 CCS (above 600) lowered the
odds of borrowing, while all lower scores had no effect on borrowing.
14
This demonstrates that the MSLP achieved the goal of targeting firms that were too large
for the PPP or likely to find the PPP’s funding cap binding. At the same time, having access
to a PPP loan early, which depended almost entirely on having an existing relationship with
at least one bank, further enhanced a firm’s chances of receiving a Main Street loan. This
highlights the importance of banking relationships in firms’ access to government credit-
support programs administered through private lenders, given the evidence that early PPP
recipients tended to have more established relationships with banks (see, for example, Bartik
et al., 2020). It is thus not surprising that firms that had 10 or fewer employees (the omitted
category) or received a PPP loan late or not at all (the omitted category) in 2020 were the
least likely to borrow from the MSLP.31 These additional regressors also augment the overall
fit of the model by an order of magnitude, although their collective explanatory power is
still rather modest.
Column (3) adds the two industry indicators (measuring sensitivity to the pandemic),
along with the local mobility index and COVID-19 death rates during the period of April
through June 2020. Operating in an essential industry or a county with greater mobility
(proxying for more active commerce) was associated with a slightly lower likelihood of MSLP
uptake, whereas the industry-level severity of COVID-19’s impact or county death rates were
insignificant determinants of uptake.32 Column (4) shows that a higher community-bank-
deposit share in the county or a lower share of delayed PPP loans also reduced MSLP
uptake.33 Both indicators likely serve as proxies for the importance of early injection of PPP
liquidity for small businesses’ survival.34 Perhaps not surprisingly, columns (3) and (4) also
make clear that despite the relevance of several industry- or county-level indicators, they
add little to the model’s fit beyond the firm-level attributes’ contribution.
Discrete-Time Hazard Model of MSLP Uptake
Main Street loans were granted over a six-month period. To account for the timing of
each loan, we next estimate a discrete-time hazard model that explains the probability of
31
Small firms found it more difficult to access the MSLP because of the certification requirement (see
Arseneau et al., 2021). Few truly small businesses borrowed because the minimum loan size was reduced
from $250,000 to $100,000 in October 2020, leaving little time for such firms to complete the underwriting
process before the program’s deadline.
32
The sum of the coefficients on the mobility index is reported in the bottom portion of Table 2, while the
(insignificant) sum of the coefficients on death rates is omitted to save space.
33
We also added the normalized cumulative PPP loan amount through June 2020 but found that it was
insignificant and therefore omitted it for space.
34
Granja et al. (2020) show that community banks were crucial to PPP lending early on, while Doniger
and Kay (2021) document the impact of delays in obtaining PPP loans on local employment.
15
a firm receiving a Main Street loan in a particular month (conditional on the firm having
not borrowed until then), instead of just the cross-section probability of being an MSLP
borrower, as in Equation (1). Specifically, we estimate the following linear model on suitably
organized panel data of (the random sample of) the size-eligible firms:35
Yit = ηt + Si0 β1 + ∆Si,t−1 β2 + Xj γ + Xc,t−1 θ + Xppp,i λ + it . (2)
Yit = 0 if in month t firm i did not borrow; Yit = 1 otherwise, which is also the last observation
in the time series for a borrower, since it is the last period the firm was “at risk.”36 ηt traces
out a nonparametric baseline hazard with t indexing the periods over which the event may
occur, which in this setting spans July through December 2020. ∆Si,t−1 denotes the (binned)
change in i’s risk score from March 2020 to the month before the MSLP loan submission,
while Xc,t−1 denotes the average economic or public health conditions over the same time
period, again approximated using the mobility index and COVID-19 death rates, respectively.
Si0 , Xj , and Xppp,i are the same set of time-invariant regressors as defined in Equation 1.
Table 3 reports the estimates of the linear hazard model. The coefficients clearly exhibit
patterns similar to those from the static probability model above, with just a few noteworthy
differences. Firms with a pre-pandemic CCS of 300 to 400 (below the 5th percentile) now
show a small but statistically significant higher propensity to borrow. With the COVID-19–
period CCS change measured from March 2020 to the month immediately before the loan
submission, arguably the more relevant period for each borrower, only firms that suffered
a CCS decline of more than 20 points were more likely to borrow. The new definition
also boosts the collective explanatory power of the CCS-based regressors by a fair margin.
Mobility in the county becomes insignificant, its explanatory power likely taken over by the
COVID-19–period CCS change defined over the more relevant months.37
In sum, the probability models paint a consistent picture: Being a midsize firm and
having prior banking relationships were the most important predictors of MSLP uptake.
This signifies the success of Main Street’s design to target firms in danger of falling through
the cracks of the other public-support programs. It also highlights the importance of prior
banking relationships along with the potential drawbacks of implementing a lending program
through banks. Having poor to mediocre pre-pandemic risk scores and suffering a large
35
As Allison (1982) shows, a logit probability regression using properly organized panel data can be
interpreted as a discrete-time hazard model: For every subject, there are as many observations as there are
number of periods when a unit is at risk of the event. A logit model corresponds to assuming a proportional
odds ratio, and it closely approximates proportional hazards if the hazard is low.
36
The probability of borrowing more than one loan is ignored given the exceedingly small fraction of such
borrowers.
37
The industry-level pandemic impact, COVID-19 death rates, and cumulative PPP loan amount through
June 2020 remain insignificant.
16
increase in delinquency risk also contributed to the propensity to borrow. These results
confirm that Main Street also succeeded in reaching firms that were more disrupted by
the pandemic, but the program’s design features, including a uniform interest rate and a
relatively high ceiling on the leverage ratio, attracted more risky firms as well, which is
consistent with the findings in Arseneau et al. (2021).
5 The MSLP Helped Shore Up Firm Financial Health
In this section, we document how receipt of an MSLP loan and the receipt’s timing affected
the borrower’s financial condition. Precisely speaking, receipt of an MSLP loan in 2020 is
a staggered treatment even at the quarterly frequency, since the MSLP operation spanned
the last two quarters of 2020 (from July 6 through December 31). A burgeoning body of
work demonstrates that the standard two-way fixed effects (TWFE) estimations applied
to staggered treatments can potentially lead to nontrivial biases.38 The cause of biases
highlighted in this literature is likely minor in our setting because the never-treated group
accounts for the overwhelming majority of the sample.
Nevertheless, we apply the method of Callaway and Sant’Anna (2021) to derive a more
precise nonparametric estimate of the treatment effects of receiving MSLP loans. This
enables us to not only account for the staggered design, but also provide a consistent ag-
gregation of the coefficients estimated for firms receiving loans in different quarters. We
specifically apply the doubly robust (dr) estimand as follows:39
" pc (X)N ! #
Gc 1−p (X)
AT Tdr (c, t) = E − h c i yf,t − yf,c−1 − mnev
c,t (X) , (3)
E(Gc ) E p c (X)N
1−pc (X)
where yf,t is an outcome y of firm f in quarter t, and yf,c−1 is its outcome in the quarter
immediately before the treatment. Gc is an indicator equal to one if the firm was treated
in quarter c and zero otherwise. The treatment quarter also defines the cohort: Cohort c
38
The basic intuition is that, with staggered treatment, in some periods the comparison group for a newly
treated cohort includes units that have been treated earlier, but this is not accounted for by the TWFE
estimator, leading to biased estimates of treatment effects. See, for instance, Sun and Abraham (2021),
Callaway and Sant’Anna (2021), and Baker et al. (2022). Roth et al. (2022) offer an excellent review of the
literature.
39
It is an adaptation of Equation (2.4) in Callaway and Sant’Anna (2021), which applies to settings with
a never-treated group, under the assumption that there is no anticipation effect. This can be justified by
the argument that firms could not take action in anticipation of receiving MSLP loans because not only
was the approval uncertain, but also they could not pay their bills or employees (which would improve their
CCS) until they received the funds. This doubly robust estimand is robust to misspecifications of either the
propensity score or the conditional expectation of the outcome of the comparison group.
17
refers to all the constituent firms that received MSLP loans in quarter c. N is the indicator
equal to one if the firm did not borrow (that is, never treated). pc (X) is the propensity
score for being treated in quarter c, that is, pc (X) = P (Gc = 1|X, Gc = 1orGi = ∞).
mnev
c,t (X) = E[yf,t − yf,c−1 |X, N = 1] denotes the mean change in outcome for non-borrower
firms from quarter t to the pre-treatment quarter for cohort c. X denotes the vector of all
the control variables, which enter as fixed effects here.
Empirical estimates of AT Tdr (c, t), denoted AT [ T dr (c, t), are obtained in two steps. First,
we estimate pc (X) with a parametric (logit) model and mnev c,t (X) with a linear regression,
40
both of which take X as the covariates. The fully saturated specification includes these
fixed effects in X: firm, state-time, industry-time, age-bin-time, employment-bin-time, time
interacted with the firm’s 2019 average Viability Score (VS), and the binned value of the VS
Point change from March 2020 to the month just before the loan submission. Second, we
plug the fitted values (p̂c (X) and m̂nev c,t (X)) into Equation (3) to derive the sample analog
AT
[ T dr (c, t). Note that for AT
[ T dr (c, t) to identify the treatment effect of Main Street, we need
only the assumption of a parallel trend in the absence of a treatment between the borrowers
and the non-borrowers after the treatment quarter (that is, Assumption 4 in Callaway and
Sant’Anna (2021)). It imposes no condition on the evolution of a variable for borrowers
relative to non-borrowers before the treatment time. Hence, any non-parallel trend before
the onset of the pandemic does not per se invalidate the treatment effect estimates. On the
other hand, the presence of a parallel pre–COVID-19 trend provides supporting evidence for
the assumption of a parallel (non-treatment) trend after the treatment time.
We present the dynamic treatment effects by the relative quarter (typically referred to as
event time) τ with τ = t − c, that is, indexed by the number of quarters after receipt of an
MSLP loan. Figure 1 depicts the estimated AT [ T c,τ for the CCS (panel (a)) and employment
41
(panel (b)). Panel (a) shows that both cohorts of Main Street borrowers exhibited an
essentially parallel pre-pandemic CCS trend relative to non-borrowers, except for a noticeable
decline in CCS in the quarter just before they received loans. Receiving an MSLP loan
was then associated with progressive increases in CCS over the four quarters afterward,
reaching 15 and 10 points higher after four quarters for the 2020:Q3 and 2020:Q4 cohorts of
borrowers, respectively.42 The CCS increases in the third and fourth post-borrowing quarters
are significant for the 2020:Q4 cohort because of the much larger number of borrowers (due to
Pn
40
E(Z) = n−1 i=1 Zi for any variable Z.
41
All the estimates in this section are produced using the DiD package in R created by Brantly Callaway
and Pedro Sant’Anna; see https://cran.r-project.org/web/packages/did/.
42
These gains correspond to 23 and 16 percent reductions in delinquency risk, respectively. In the pre–
COVID-19 period, the average within-firm standard deviation in CCS is 9. Thus, MSLP receipt is associated
with a CCS increase that is 1.1 to 1.7 times the within-firm standard deviation.
18
the surge in December) and hence smaller standard errors, but the increase is only marginally
significant for the 2020:Q3 borrowers in the fourth quarter after loan receipt.43
Likewise, Main Street also appears to have boosted both borrower cohorts’ employment
by noticeable margins in the third and fourth quarters after loan receipt, although none
of the increases is statistically significant (shown in panel (b) of Figure 1). This lack of
significance may stem in part from greater measurement errors in the employment data,
which are exacerbated in changes over a relatively brief period.44 It is also plausible that
more time is needed for the effect of Main Street funding to manifest in employment since the
economy was not yet fully recovered from the pandemic disruptions in 2021, including the
shortfall in labor supply that has constrained employment growth. The labor shortage will
likely take some time to abate. In principle, because the liquidity injection from Main Street
put MSLP borrowers on sounder financial footing, they should be in a more advantageous
position to exploit growth opportunities as the economy continues to normalize.
Estimates of AT[ T c,τ for Viability Scores and FSS are presented in Appendix E. Main
Street funding brought about improvement in these two financial indicators as well, albeit
not as notably as for the CCS, consistent with our earlier conjecture that the liquidity injec-
tion would most likely manifest in better payment performance and hence lower borrowers’
delinquency risk. In Appendix F, we also consider the Sun and Abraham (2021) estimator
for staggered treatments to corroborate the findings using the Callaway-Sant’Anna estimator
and to explore the relative contribution of firms’ pre-MSLP-borrowing financial and nonfi-
nancial attributes to the estimates. The overall finding is that these two estimators yield
fairly similar estimates of Main Street’s effects, although accounting for a firm’s propensity
to borrow is important for tight matching between borrowers and their most comparable
non-borrowers to minimize different pre-trends between borrowers and their peers.
In sum, we find generally robust evidence that Main Street credit helped borrowers
improve their financial health—reducing their delinquency risk in particular—over the four
quarters after they received a loan. By comparison, there is only suggestive evidence that
the program also enabled borrowers to recover their employment over the same period. So
it appears that Main Street, at the very least, achieved its goal of bridging the COVID-19–
induced liquidity gap and thus placed firms on more solid financial footing. All else being
equal, the more adequate liquidity should enable these firms to take advantage of growth
43
The average treatment effect AT
[ T dr (τ ) across cohorts (c) in a given quarter τ is fairly close to that for
the 2020:Q4 cohort because of its dominating sample size.
44
See Appendix A and Wang, Ballance, and Qing (2021) for more details. Despite the inadequacies of
D&B employment data, there is no strong reason to suspect that the data shortcomings should bias the
estimates instead of just compromising the precision. Moreover, assessing the impact of Main Street on
employment is an important element for analyzing the policy. We therefore choose to use the employment
data.
19
opportunities in the coming years as the economy continues to recover from pandemic-
induced dislocations.
6 Concluding Thoughts
Using a comprehensive database of businesses’ financial condition, this study compares firms
that borrowed from the Main Street Lending Program with their peers to understand which
attributes and circumstances made it more likely for firms to seek Main Street credit and how
the program benefited them afterward. The patterns that emerge are broadly consistent with
the targeting and objective of the MSLP. In line with the program’s targeting, midsize firms
(those with 250 to somewhat more than 500 employees) exhibited the highest probability of
becoming borrowers. On the other hand, several of the program’s design elements appear to
have made it notably harder for firms with fewer than 10 employees to borrow. Being an early
recipient of a 2020 PPP loan raised the likelihood that a firm would borrow by a nontrivial
margin, confirming the importance of preexisting banking relationships, as well as these firms’
need for funding to weather the COVID-19 shock to their liquidity. This interpretation is
further supported by the fact that a considerable fraction of MSLP borrowers also received
second-draw PPP loans in 2021.
As conjectured, firms with moderately low risk just prior to the COVID-19 outbreak were
more likely to borrow compared with riskier firms, while the safest firms were the least likely
to borrow. Changes in risk scores reveal a pattern that MSLP borrowers tended to be firms
with better prospects but that had suffered worse disruptions due to the pandemic. Local
economic environments also exerted mild influence over a firm’s probability of borrowing in
ways consistent with intuition. Better county economic conditions and a larger presence of
community banks reduced the odds of borrowing, as did operating in an essential industry.
Having more PPP loans in a county delayed by the funding hiatus made it more likely for
firms in that county to borrow from Main Street.
Based on a doubly robust estimator tailored to staggered treatments, Main Street is
estimated to have led to progressive and significant improvement in recipient firms’ finan-
cial health over the year after they received a loan. In particular, the reduction in their
delinquency risk reached nearly 18 percent four quarters after loan receipt. This consti-
tutes evidence that the MSLP achieved the objective of shoring up the liquidity position
of firms that had promising prospects but were hit hard by the pandemic. By comparison,
employment at MSLP firms recovered slightly better than at their peer firms over the same
period after borrowing, but the difference was barely significant. Apart from the noise in
the employment data, full recovery may require more time, especially considering the labor
20
market tightness in general. To the extent that firms in a better liquidity position can more
readily take advantage of growth opportunities, these borrowers should be able to expand
operations more easily as the economy continues to normalize over the coming quarters.
The general pattern that emerges from the findings is that Main Street borrowers were,
on average, more commercially active and viable before COVID-19 hit, but they were more
adversely affected by the pandemic and thus suffered a worse short-term liquidity shortage.
The combination of these conditions made these firms suitable recipients of Main Street loans.
Their promising long-term prospects made the lending relationship valuable to their banks.
The onset of the pandemic, however, rendered the firms more risky, at least temporarily.
Therefore, the banks turned to the MSLP to fund the loans, as the program likely enabled
them to offer lower rates or lend larger amounts, or both, than they would have had they
used solely their own capital. For the firms that were eligible for the PPP but required
more funding, Main Street helped make up for the shortfall. For those too large to qualify
for the PPP, Main Street filled the gap, as intended. The additional credit or lower rates
should enhance MSLP firms’ odds of survival and a more robust recovery, enhancing the
relationship value to the banks.
The analytical results here support the argument that, despite the various challenges, a
program like the MSLP following major disruptions to the economy can confer substantial
real economic benefits by providing the bridge funding to firms that have good prospects
but are suffering from short-term liquidity shortages. More time is needed to assess the
medium-term efficacy of Main Street, especially in terms of promoting real growth. In addi-
tion, more analysis remains necessary to gain a better understanding of the attributes that
identify firms with greater growth potential and design a program accordingly to encourage
their participation. This will reduce the chances of offering cheap credit to fundamentally
nonviable firms that enables them to linger longer than they would otherwise, which could
discourage the entry of new, more promising firms. Additional analysis of the MSLP can
thus help make a similar lending program in the future more efficient in terms of funding
and more effective in terms of providing an economic stimulus.
References
Allison, P. (1982): “Discrete Time Methods for the Analysis of Event Histories,” Socio-
logical Methodology, Leinhardt (ed), 61–98.
Arseneau, D., J. Fillat, M. Mahar, D. P. Morgan, and S. V. den Heuvel (2021):
21
“COVID Response: The Main Street Lending Program,” Federal Reserve Bank of New
York Staff Reports No. 984.
Baker, A. C., D. F. Larcker, and C. C. Wang (2022): “How Much Should We
Trust Staggered Difference-in-Differences Estimates?” Journal of Financial Economics,
144, 370–395.
Barnatchez, K., L. D. Crane, and R. Decker (2017): “An Assessment of the National
Establishment Eime Series (NETS) Database,” FEDS working paper.
Bartik, A. W., Z. E. Cullen, E. L. Glaeser, M. Luca, C. T. Stanton, and
A. Sunderam (2020): “The Targeting and Impact of Paycheck Protection Program Loans
to Small Businesses,” Working Paper 27623, National Bureau of Economic Research.
Bräuning, F., J. L. Fillat, F. Lin, and J. C. Wang (2021): “A Helping Hand to Main
Street Where and When It Was Needed?” Current Policy Perspectives, Federal Reserve
Bank of Boston.
Bräuning, F. and T. Paligorova (2021): “Uptake of the Main Street Lending Program,”
Current Policy Perspectives, Federal Reserve Bank of Boston.
Callaway, B. and P. H. Sant’Anna (2021): “Difference-in-Differences with Multiple
Time Periods,” Journal of Econometrics, 225, 200–230.
Chetty, R., J. Friedman, N. Hendren, M. Stepner, and T. O. I. Team (2021):
“The Economic Impacts of COVID-19: Evidence from a New Public Database Built Using
Private Sector Data,” Working paper.
Doniger, C. L. and B. Kay (2021): “Ten Days Late and Billions of Dollars Short: The
Employment Effects of Delays in Paycheck Protection Program Financing,” Finance and
Economics Discussion Series 2021-003, Board of Governors of the Federal Reserve System.
Faulkender, M. W., R. Jackman, and S. Miran (2021): “The Job Preservation Effects
of Paycheck Protection Program Loans,” SSRN Electronic Journal.
Granja, J., C. Makridis, C. Yannelis, and E. Zwick (2020): “Did the Paycheck
Protection Program Hit the Target?” Working Paper 27095, National Bureau of Economic
Research.
Imbens, G. W. (1992): “An Efficient Method of Moments Estimator for Discrete Choice
Models with Choice-Based Sampling,” Econometrica, 60, 1187–1214.
22
Joaquim, G. and F. Netto (2021): “Bank Incentives and the Effect of the Paycheck
Protection Program,” Working Papers 21-15, Federal Reserve Bank of Boston.
Joaquim, G. and J. C. Wang (2022): “What Do 25 Million Records of Small Businesses
Say about the Effects of the PPP?” Working Paper, Federal Reserve Bank of Boston.
King, G. and L. Zeng (2001): “Logistic Regression in Rare Events Data,” Political
Analysis, 9, 137–163.
Minoiu, C., R. Zarutskie, and A. Zlate (2022): “Motivating Banks to Lend? Credit
Spillover Effects of the Main Street Lending Program,” SSRN Working Paper.
Morgan, D. P. and S. Clampitt (2021): “Up on Main Street,” Federal Reserve Bank
of New York, Liberty Street Economics (blog), February 5, 2021.
Roth, J., P. H. Sant’Anna, A. Bilinski, and J. Poe (2022): “What’s Trending
in Difference-in-Differences? A Synthesis of the Recent Econometrics Literature,” arXiv
preprint arXiv:2201.01194.
Sun, L. and S. Abraham (2021): “Estimating Dynamic Treatment Effects in Event Stud-
ies with Heterogeneous Treatment Effects,” Journal of Econometrics, 225, 175–199.
Wang, J. C., J. Ballance, and M. Qing (2021): “How Did the MSLP Borrowers Fare
before and during COVID-19?” Current Policy Perspectives, Federal Reserve Bank of
Boston.
23
Table 1: Comparison of Pre-COVID Characteristics: MSLP Borrowers versus All Non-
Borrowers
Attribute (1) (2) (3) (4) Min. Obs.
CCS Points (2020M3) 17.033 8.258 4.309 -0.198 485,378
(0.001) (0.031) (0.153) (0.944)
CCS Change (2020M3–2020M6) -5.557 -4.710 -4.545 -4.760 480,890
(0.024) (0.041) (0.06) (0.042)
FSS Points (2020M3) -0.918 -9.711 -11.629 -15.088 485,303
(0.802) (0.001) (0) (0)
FSS Change (2020M3–2020M6) -0.606 -0.159 -0.356 -0.911 480,829
(0.536) (0.859) (0.715) (0.327)
Viability Points (2020M3) 20.057 2.473 3.267 -0.798 489,971
(0) (0.329) (0.011) (0.465)
Viability Change (2020M3–2020M6) -0.833 -0.571 -0.488 -1.236 485,529
(0.038) (0.176) (0.462) (0.057)
% Debt 31+ Days Past Due (2020M3) -0.003 -0.009 -0.023 -0.014 117,622
(0.67) (0.172) (0.001) (0.029)
% Change Past Due (2020M3–2020M6) 0.038 0.036 0.034 0.036 107,792
(0) (0) (0) (0)
Conditioning Variables – – – –
State Yes Yes Yes Yes
Industry Yes Yes Yes Yes
Firm Age Bins No Yes Yes Yes
Employement Bins No Yes Yes Yes
Viability Score Bins (2019) No No Yes Yes
PPP Indicator No No No Yes
Note: This table compares the average value of firm attributes among MSLP borrowers versus all the other
firms that were eligible (based on employee count) but did not borrow from the MSLP. The conditional
mean differences are reported, along with the p values of the equal-mean test in parentheses. The mean
difference for each attribute is computed by regressing that attribute on a set of binary indicators for the
conditioning variables, along with the MSLP borrower indicator, whose coefficient then measures the mean
difference between MSLP borrowers and the other firms while controlling for the conditioning variables. The
conditioning variables for each set of mean comparisons (organized by column) are reported in the bottom
block of the table. “Min. Obs.” lists the minimum number of observations among the four sets of mean
differences reported in each row, universally corresponding to column (4). There are generally around 5%
more observations underlying results in columns (1) and (2). Source: D&B and author’s calculations.
24
Table 2: Probability of Firm MSLP Uptake: Financial Health Based on CCS
(1) (2) (3) (4)
CCS Range (300,600] 0.0003∗∗∗ −0.00001 0.00000 0.00001
(0.0001) (0.00005) (0.00005) (0.0001)
CCS >600 0.0001 −0.0004∗∗∗ −0.0004∗∗∗ −0.0004∗∗∗
(0.0001) (0.0001) (0.0001) (0.0001)
CCS Change Pre <-20 0.0002∗∗∗ 0.00004 0.00005∗ 0.00005∗
(0.00005) (0.00003) (0.00003) (0.00003)
CCS Change Pre [-20,0) −0.00000 0.00001 0.00002∗ 0.00002∗∗
(0.00001) (0.00001) (0.00001) (0.00001)
CCS Change Pre (0,20] 0.0001∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗
(0.00002) (0.00002) (0.00002) (0.00002)
CCS Change Pre >20 0.0003∗∗∗ 0.0002∗∗∗ 0.0002∗∗∗ 0.0002∗∗∗
(0.0001) (0.00004) (0.00004) (0.00004)
CCS Change Covid <-20 0.001∗∗∗ 0.0004∗∗∗ 0.0004∗∗∗ 0.0004∗∗∗
(0.0001) (0.0001) (0.0001) (0.0001)
CCS Change Covid [-20,0) 0.0001∗∗ 0.00003 0.00003 0.00003
(0.00003) (0.00002) (0.00002) (0.00002)
CCS Change Covid (0,20] 0.0002∗∗∗ 0.00005∗∗∗ 0.00004∗∗∗ 0.00004∗∗
(0.00003) (0.00002) (0.00001) (0.00001)
CCS Change Covid >20 0.001∗∗∗ 0.0002∗∗∗ 0.0002∗∗∗ 0.0002∗∗∗
(0.0001) (0.0001) (0.0001) (0.0001)
PPP Early 0.002∗∗∗ 0.002∗∗∗ 0.002∗∗∗
(0.0005) (0.0005) (0.0005)
PPP Mid 0.0004∗∗∗ 0.0004∗∗∗ 0.0004∗∗∗
(0.0001) (0.0001) (0.0001)
PPP Late 0.0001 0.0001 0.0001
(0.0001) (0.0001) (0.0001)
Emp. [11, 50] 0.001∗∗∗ 0.001∗∗∗ 0.001∗∗∗
(0.0002) (0.0002) (0.0002)
Emp. [51,250] 0.003∗∗∗ 0.003∗∗∗ 0.003∗∗∗
(0.001) (0.001) (0.001)
Emp. [251,500] 0.006∗∗∗ 0.006∗∗∗ 0.006∗∗∗
(0.002) (0.002) (0.002)
Emp. >500 0.005∗∗∗ 0.005∗∗∗ 0.005∗∗∗
(0.002) (0.002) (0.002)
Essential Industry −0.0001∗ −0.0001∗
(0.0001) (0.0001)
COVID Impacted 0.0001 0.0001
(0.0001) (0.0001)
Community Bank Share −0.0002∗
(0.0001)
PPP Loans Share Delayed 0.0005∗∗
(0.0002)
Sum mobility coefs (p val) – – -0.002 (0.091) -0.002 (0.081)
Observations 418,995 418,995 418,995 418,995
Adjusted R2 0.0003 0.003 0.003 0.003
Note: This table reports coefficients from Equation (1). CCS Range and Employment (Emp.) bins: as of
2020:M3. CCS Change Pre: binned values of CCS change 2019:M3–2020:M3. CCS Change Covid: binned
values of CCS change 2020:M3–2020:M6. PPP Early, Mid, Late: Indicator if a firm received PPP loans
before April 16, April 27 to May 2, and after May 2, 2020, respectively. See text for the other variables’
definitions. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively. Source: D&B
and author’s calculations.
25
Table 3: Discrete-Time Hazard Model of Firm MSLP Uptake
(1) (2) (3) (4)
M3 CCS Range (300,400] 0.0004∗∗∗ 0.0001∗∗ 0.0001∗∗ 0.0001∗∗
(0.0001) (0.0001) (0.0001) (0.0001)
M3 CCS Range (400,600] 0.0002∗∗∗ 0.0001 0.0001 0.0001
(0.0001) (0.0001) (0.0001) (0.0001)
M3 CCS >600 0.0001 −0.0004∗∗∗ −0.0004∗∗∗ −0.0004∗∗∗
(0.0001) (0.0001) (0.0001) (0.0001)
CCS Change Pre <-20 0.0002∗∗∗ 0.0001∗ 0.0001∗∗ 0.0001∗∗
(0.0001) (0.00003) (0.00004) (0.00004)
CCS Change Pre [-20,0) 0.00000 0.00002∗ 0.00002∗∗ 0.00002∗∗
(0.00001) (0.00001) (0.00001) (0.00001)
CCS Change Pre (0,20] 0.0001∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗
(0.00002) (0.00002) (0.00002) (0.00002)
CCS Change Pre >20 0.0003∗∗∗ 0.0002∗∗∗ 0.0002∗∗∗ 0.0002∗∗∗
(0.0001) (0.00004) (0.00004) (0.00004)
CCS Change Covid <-20 0.001∗∗∗ 0.0003∗∗∗ 0.0003∗∗∗ 0.0003∗∗∗
(0.0001) (0.0001) (0.0001) (0.0001)
CCS Change Covid [-20,0) −0.00001 −0.00002 −0.00002 −0.00002
(0.00001) (0.00001) (0.00002) (0.00001)
CCS Change Covid (0,20] 0.0001∗∗∗ −0.00003 −0.00003 −0.00004
(0.00003) (0.00003) (0.00003) (0.00003)
CCS Change Covid >20 0.0004∗∗∗ 0.0001 0.0001∗ 0.0001
(0.0001) (0.00003) (0.00003) (0.00003)
PPP Early 0.002∗∗∗ 0.002∗∗∗ 0.002∗∗∗
(0.0005) (0.0005) (0.0005)
PPP Mid 0.0004∗∗∗ 0.0004∗∗∗ 0.0004∗∗∗
(0.0001) (0.0001) (0.0001)
PPP Late 0.0001 0.0001 0.0001
(0.0001) (0.0001) (0.0001)
Emp. [11,50] 0.001∗∗∗ 0.001∗∗∗ 0.001∗∗∗
(0.0002) (0.0002) (0.0002)
Emp. [51,250] 0.003∗∗∗ 0.003∗∗∗ 0.003∗∗∗
(0.001) (0.001) (0.001)
Emp. [251,500] 0.005∗∗∗ 0.005∗∗∗ 0.005∗∗∗
(0.002) (0.002) (0.002)
Emp. >500 0.005∗∗∗ 0.005∗∗∗ 0.005∗∗∗
(0.002) (0.002) (0.002)
Essential Industry −0.0001∗ −0.0001∗
(0.00005) (0.00005)
Community Bank Share −0.0004∗∗
(0.0002)
PPP Loans Share Delayed 0.001∗∗
(0.0002)
Observations 2,491,824 2,491,824 2,491,824 2,491,824
Adjusted R2 0.001 0.003 0.003 0.003
Note: This table reports coefficients from Equation (2). M3 CCS Range and Employment (Emp.) bins: as
of 2020:M3. CCS Change Pre: binned values of CCS change 2019:M3–2020:M3. CCS Change Covid: binned
values of CCS change 2020:M3–MSLP submission month. PPP Early, Mid, Late: Indicator if a firm received
PPP loans before April 16, April 27 to May 2, and after May 2, 2020, respectively. See text for the other
variables’ definitions. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively. Source:
D&B and author’s calculations. 26
Figure 1: Effects of the MSLP on Firm Commercial Credit Score and Employment
(a) MSLP Effects on CCS
(b) MSLP Effects on Employment
Note: This figure depicts the dynamic effects of the staggered treatment of MSLP loans on borrowers’
Commercial Credit Score (top panel) and employment (bottom panel), estimated using the doubly robust
DiD estimator à la Callaway and Sant’Anna (2021). Source: D&B and author’s calculations.
27
Part I
Online Appendix — Not For
Publication
Table of Contents
A Dun & Bradstreet Database A-2
B Main Street Lending Program and Identifying Borrow-
ers in the D&B Database A-5
C Rare-Event Logit Model of MSLP Uptake Decision A-11
D Additional Estimates of Linear Model of MSLP Uptake
A-14
E Additional Callaway-Sant’Anna Estimates of MSLP’s
Effects A-16
F MSLP Treatment Effects Using Sun-Abraham Estima-
tor A-17
A-1
A Dun & Bradstreet Database
Employment
D&B offers two types of data on firm employment: (1) actual employee count for a firm,
either directly reported by the firm or gleaned from other sources, and (2) modeled (imputed)
employment, which is estimated using D&Bs proprietary model along with other data items
D&B collects on the firm. Around 25 percent of the firms report actual employment.1 Note,
however, that the actual employment data are updated only with a lag and often at different
times for different firms, or even any given firm, depending on the arrival of new information.
This means that the actual employment may be out of date to varying degrees across firms.
The general pattern is that the smallest businesses (especially those with fewer than 10
employees) and young businesses (those established in 2017 or after) are more likely to have
only estimated or modeled values for employment (as documented in Wang, Ballance, and
Qing, 2021). To maximize the sample size, we thus use all the firms with any type of
employment data in our baseline analysis. In unreported results, we compare estimates for
firms with actual versus modeled employment and find that in general the former are more
creditworthy but do not exhibit significantly different dynamics since COVID-19 hit.
As documented in Wang, Ballance, and Qing (2021), more than a third of the firms were
ranked in the highest decile by employee count in March 2020 relative to other firms in the
same state and three-digit NAICS industry. Also, compared with such peers as of March
2020, nearly a third of them were ranked in the lowest decile and another 20 percent in the
second-lowest decile. Overall, nearly 40 percent of the firms have been established within
the last 10 years. In the D&B population of firms, firm size and age are broadly positively
correlated. So, it seems that, controlling for size just before the pandemic hit, a fraction of
MSLP firms had experienced faster growth relative to their peers.
Risk Scores
We consider the three major risk scores compiled by D&B, all of which are modeled
to be predictive indicators. The granular raw points for every score is structured as it is
for consumer credit scores, so a higher value signifies lower risk. Among these, the most
comprehensive are Viability Points, which range from 101 to 800. They assess a business’s
overall likelihood of going out of business (which includes becoming inactive or filing for
bankruptcy) over the next 12 months. We use the Viability Points discretized counterpart—
Viability Score—in most analyses because this is more widely used by creditors.2 The Com-
1
D&B makes no distinction between full- and part-time employees, and the data are best regarded as
roughly full-time-equivalent employees.
2
This score also contains descriptive components pertaining to the quantity of predictive data available
A-2
mercial Credit Score (CCS, also known as the Delinquency Score), ranging from 101 to 670,
predicts the likelihood of a business having a severely delinquent account (91-plus days past
due) over the next 12 months.3 The Financial Stress Score (FSS, also known as the Failure
Score), ranging from 1,001 to 1,875, predicts the likelihood over the next 12 months that a
business will incur financial distress (such as ceasing operations, leaving unpaid obligations
to creditors, or moving into receivership) or file for bankruptcy. As an example, Figure
A.1 depicts the range of Viability Scores (from the 10th to the 90th percentiles) for MSLP
borrowers from January 2020 onward. The figure shows that the scores generally peaked
around March 2020 and then declined visibly among the top percentiles.4
Payment Records
The last set of variables record the degree of a business’s payment delinquency, if any,
over the past 3 to 24 months (depending on the variable). We focus on the more timely
records over the past three months. We examine the portion of the total amount owed over
the most recent three months that is 31-plus, 61-plus, or 91-plus days past due.5 As shown
in Wang, Ballance, and Qing (2021), the share of firms with bills 31-plus days past due rose
after the March–April 2020 period and mostly peaked in July 2020, followed by a modest
improvement (that is, a decline) in July among firms with visibly higher than average shares
(whereas those with the most serious delinquency level had their portion of bills that were
31-plus days past due stuck at 1).6
As a sufficient statistic of these indicators for how punctually a business has been paying
its bills, D&B compiles the Paydex score. The most commonly used version of Paydex is
based on payment behavior over the past 24 months. A value of 80 or above means a firm
pays on time, 70 equals 15 days beyond terms, 60 equals 22 days beyond terms, 50 equals
to make a reliable risk assessment, as well as the market segment for each firm based on its age, size, and
quantity of data signals. These are not used because we apply our own grouping criteria.
3
The marginal odds of being a good risk doubles for each 40-point increase. This also applies to the
Financial Stress Score described next.
4
The distributional patterns of the time series of these risk scores from January 2020 onward for all the
firms in our internal D&B data are broadly similar. In fact, this holds for basically all the variables analyzed
here. A few minor exceptions are the past due variables, where the highest percentiles of the distribution
tend to take the value of one throughout the sample months.
5
To supplement these near-term records that are based on the amount owed on bills that are due, in
unreported results we also looked at a measure based on the number of accounts (the highest ratio of the
cumulative number of accounts reaching 31-plus days past due divided by the cumulative total number of
active accounts for the most recent month and three months prior) and how persistent a firm’s trouble in
paying bills had been (the portion of months over the last 12 months with reported data that contain a
balance that is 31-plus or 61-plus days past due).
6
The pattern is broadly similar for the portion of bills that were 61-plus and 91-plus days past due,
although the share reporting zero is higher.
A-3
30 days beyond terms, 40 equals 60 days beyond terms, 30 equals 90 days beyond terms, 20
equals 120 days beyond terms, and 0 equals 180-plus days beyond terms. If a Paydex score
is missing, it is due to an insufficient number of qualified trade experiences. This score thus
is available for many more firms than the number of firms that have data for recent (over
the past three months) payment records, even though it is less available than the three risk
scores.7 The 24-month Paydex is available for 90 percent or so of the firms since January
2020 in our downloaded data set, whereas payment records over the last three months are
available for only about one-third of the firms. Unlike the forward-looking risk scores, the
commonly referenced Paydex score is backward-looking, based on a firm’s past payment
records.
Figure A.1: Viability Score Points, MSLP Borrowers, 01/2019 to 12/2020
Count (left axis) avg p99 p95 p90 p75 p50 p25 p10 p05 p01
1600
650
1580
1560 600
1540
Overall Viability Score
550
Obs. Count
1520
1500
500
1480
450
1460
1440
400
1420
Jan 2019 Apr 2019 Jul 2019 Oct 2019 Jan 2020 Apr 2020 Jul 2020 Oct 2020
Note: Number of MSLP firms in the sample (light gray shading) on the left scale; monthly value at selected
percentiles along with the mean value on the right scale. Source: D&B and author’s calculations.
7
The Paydex score is available only for firms with three or more payment experiences from at least two
trade providers, and some firms do not satisfy this condition even over a 24-month period.
A-4
B Main Street Lending Program and Identifying Borrowers in
the D&B Database
Main Street Lending Program
Several facilities were established under the MSLP to offer different parameters and
loan features that would meet the diverse needs of small and midsize businesses. Secured or
unsecured “new loans” with principal as high as $35 million were available for borrowers with
an adjusted debt-to-EBITDA ratio (inclusive of the Main Street loan) as high as 4. “Priority
loans” or “expanded loans” were available for borrowers with an adjusted debt-to-EBITDA
ratio as high as 6 in amounts of as much as $50 million or $300 million, respectively, but these
loan types came with more stringent security and bankruptcy priority requirements. Except
for loans made under the Priority Loan Facility—and only at the time of origination—no
MSLP loans could be used to refinance existing debt owed to lenders other than the MSLP
lender. Moreover, borrowers had to “refrain from repaying the principal balance of, or
paying any interest on, any debt” until the MSLP loan was repaid in full, “unless the debt
or interest payment is mandatory and due.” These covenants would not prohibit borrowers
from repaying a line of credit, taking on and paying additional debt obligation in the normal
course of business, or refinancing debt that was maturing in no more than 90 days.8
All the facilities featured an interest rate of LIBOR plus 300 basis points (bps) and a
maturity of five years, as well as the deferral of principal payments and interest payments for
two years and one year, respectively. The minimum loan size was initially set at $250,000 for
all the facilities, but it was later lowered to $100,000 in October 2020 to better accommodate
very small businesses. Moreover, all MSLP loans required the lender, which collected an
origination fee and loan servicing fees, to retain 5 percent of the loan balance, while the
Main Street Special Purpose Vehicle (SPV) purchased the remaining 95 percent at par.9
When the MSLP ended on January 8, 2021, it was the largest debt-purchase facility
and the fastest growing in the second half of 2020 among all Federal Reserve debt-purchase
facilities established in response to the COVID-19 pandemic (see Morgan and Clampitt,
2021). The program extended a total of $17.5 billion in credit to nearly 2,500 borrowers
across virtually all the states. The volume of loans purchased by the MSLP grew more or
less steadily over time through the end of November 2020, followed by a surge in the first
half of December, after the US Treasury Secretary announced that the program would not
8
See FAQs posted on the Federal Reserve Bank of Boston MSLP site https://www.bostonfed.org/-
/media/Documents/special-lending-facilities/mslp/legal/frequently-asked-questions-faqs-post-
termination.pdf.
9
All vintages of the Main Street term sheets, which detail the parameters for each facility and their
evolution over time, can be found on this official program page:
https://www.federalreserve.gov/monetarypolicy/mainstreetlending.htm
A-5
Table B.1: Summary Statistics of Total Main Street Lending Activity
Principal Loan Amount ($ Millions)
Total Volume ($ Millions) N. Loans N. Borrowers N. Lenders Mean Min p10 p25 p50 p75 p90 Max
All Facilities 17,459 1,830 1,815 319 9.5 0.1 0.7 1.5 4.0 10.6 25.0 300.0
MSELF 1,805 26 26 16 69.4 10.0 11.0 22.0 40.5 90.0 148.0 300.0
MSNLF 2,695 616 608 149 4.4 0.1 0.4 0.8 2.0 4.5 10.0 35.0
MSPLF 12,917 1,173 1,166 258 11.0 0.1 1.1 2.4 6.0 14.8 30.0 50.0
NONLF 42 15 15 13 2.8 0.2 0.4 0.6 2.5 5.0 5.0 8.5
Note: MSELF stands for Main Street Extended Loan Facility; MSNLF stands for Main Street New Loan
Facility; MSPLF stands for Main Street Priority Loan Facility; and NONLF stands for Nonprofit Organi-
zation Loan Facilities (pooling loans in the New Loan and Extended Loan Facilities). Source: Main Street
public-release data.
be renewed beyond December 31, 2020.10
Overall, the MSLP approved the participation requests for 1,830 loans totaling $17.5
billion. By far the most popular was the Priority Loan Facility—accounting for about 74
percent of the uptake (1,173 loans totaling $12.9 billion)—followed by the New Loan Facility
(616 loans totaling $2.7 billion), and the Extended Loan Facility (26 loans totaling $1.8
billion). Loan sizes were generally concentrated in amounts substantially below the program’s
maximum loan-size limits, with a median loan size of $40.5 million in the Extended Loan
Facility, $6 million in the Priority Loan Facility, and $2 million in the New Loan Facility.
The loan-size distribution reveals that the vast majority of loans were much larger than the
minimum requirement of $100,000. This is partly because the minimum requirement was
lowered late in the program, leaving limited time for the smallest loans to be underwritten.
This result is likely also an outcome of the MSLP’s initial design goal of making the program
appeal more to midsize firms.
The geographic coverage of the Main Street program was broad: Participation requests
were submitted by borrowers from 49 states (all but Maine), the District of Columbia, and
two territories (Puerto Rico and the US Virgin Islands).11 Banks’ participation was some-
what limited. About 600 banks registered for the program, but only about half originated
loans that were subsequently sold to the Main Street Special Purpose Vehicle.
Identifying MSLP Borrowers in the D&B Database
We rely mostly on the D&B databases name-matching algorithm (Integration Manager
or IM), supplemented with manual online searches, to identify the DUNS IDs for MSLP
10
In fact, the program was closed to new submissions on December 15, 2020, but the volume of submissions
in December was so large that it was not until January 8, 2021, that all the submissions (some of which had
to be resubmitted due to missing documents) were processed.
11
On the other hand, most of the uptake was concentrated in firms from Texas ($3.1 billion), Florida ($2.1
billion), and California ($2.1 billion). When the volume is scaled relative to the size of the state economy,
the top three states by uptake per dollar of gross state product were Oklahoma, Arkansas, and Missouri.
A-6
borrowers.
Among the 1,813 accepted primary borrowers from the Main Street program recorded as
of January 8, 2021, we are able to find DUNS matches for 1,796 directly through the IM based
on the MSLP firm name and address.12 Among these, we consider only the 1,516 borrowers
matched by the IM with a confidence score of 8 or higher (which is deemed sufficiently
precise).
To these, we are able to add 266 borrowers that initially matched with low confidence
scores by improving their scores to 8 or higher with additional information obtained via man-
ual searches. Likewise, we were also able to find matches for 13 of the 17 initially unmatched
borrowers. On the other hand, 10 borrowers were incorrectly matched by the algorithm
and were removed from our sample after manual inspection. Another four borrowers were
dropped after we aggregated branches to their headquarters DUNS because several variables
were not separately available at the branch level in the D&B database. In sum, our sample
of matched borrowers totals 1,781, all corresponding to headquarters-level DUNS.
Subsequently, several borrowers were removed from our sample either because (1) they
do not exist in the D&B database in the relevant period, (2) they cannot be assigned a peer
group due to missing employment data, or (3) their status code changed (that is, a change
to/from branch/headquarters) over the sample period.13 In March 2020, for instance, 106 of
these firms did not exist in the D&B database, another 49 had missing values for employment,
and five had changed their status during the period used to compute changes in the variables
of interest, from March 2020 to a firm’s MSLP loan submission month. Thus there are 1,605
firms left for our analysis.
12
For multi-borrower loans, which the MSLP started accepting in late September 2020, we consider only
the primary borrower. All the counts reported here are based on the number of distinct DUNS unless
otherwise noted. 1,813 DUNS correspond to 1,804 distinct tax IDs as reported by MSLP borrowers, some
of which are mapped into more than one DUNS number.
13
When a firm’s status changes over a period, changes in many variables’ values are likely to be subject
to composition bias, and we therefore omit those firms.
A-7
Table B.2: Comparison of the Random Sample versus Full D&B Data Set versus MSLP Borrower
Variable Data N Mean Mean Diff (%) 5th Pctl. 25th Pctl. 50th Pctl. 75th Pctl. 95th Pctl.
CCS Points Sample 179,452 503.02 0.02 395 473 496 551 606
Full D&B 6,400,714 502.92 394 473 496 551 606
MSLP Firms 1,506 506.08 416 474 500 554 603
FSS Points Sample 179,452 1469.23 0.029 1394 1428 1471 1510 1550
Full D&B 6,400,714 1468.81 1394 1425 1468 1510 1550
MSLP Firms 1,506 1467.99 1387 1424 1470 1508 1550
Viability Points Sample 179,452 495.77 -0.026 437 465 492 523 567
Full D&B 6,400,714 495.90 437 465 492 523 569
MSLP Firms 1,506 497.45 437 466 496 524 570
Viability Score Sample 179,452 4.61 0.074 2 3 5 6 6
A-8 Full D&B 6,400,714 4.61 2 3 5 6 6
MSLP Firms 1,506 4.56 2 3 5 6 6
Firm Age Sample 179,452 14.68 -0.745 2 6 10 17 42
Full D&B 6,400,714 14.79 2 6 10 17 42
MSLP Firms 1,506 16.30 3 6 10 19 47
Employment Sample 179,452 7.18 -7.324 1 2 2 4 18
Full D&B 6,400,714 7.72 1 2 2 4 18
MSLP Firms 1,506 102.42 1 2 2 6 75
Note: This table presents the summary statistics of key characteristics of i) the MSLP borrowers, ii) the random sample of non-borrower peer firms
matched on state, industry (3-digit NAICS), firm age bin, and small business indicator (defined as up to 500 employees), and iii) all non-borrower
peer firms defined likewise in the full D&B dataset of active firms as of March 2020. The column “Mean Diff (%)” reports the percentage difference
between mean value for the random sample versus the full D&B dataset. Source: D&B and author’s calculations.
Table B.3: Comparison of Pre-COVID Characteristics: MSLP Borrowers versus Peers
Attribute (1) (2) (3) (4) Min. Obs.
CCS Points (2020M3) 11.983 7.807 3.893 -1.245 179,578
(0.01) (0.046) (0.198) (0.622)
CCS Change (2020M3–2020M6) -5.207 -4.679 -4.551 -5.182 177,965
(0.034) (0.041) (0.059) (0.024)
CCS Change (2020M3–MSLP Month) -4.388 -3.136 -3.126 -6.475 176,398
(0.125) (0.184) (0.197) (0.009)
FSS Points (2020M3) -8.204 -8.602 -10.563 -13.263 179,554
(0.035) (0.002) (0) (0)
FSS Change (2020M3–2020M6) 0.491 0.180 -0.097 -0.677 177,946
(0.648) (0.845) (0.918) (0.449)
FSS Change (2020M3–MSLP Month) -1.002 -1.314 -1.756 -3.195 176,367
(0.404) (0.129) (0.084) (0.005)
Viability Points (2020M3) 11.821 3.081 3.489 -0.520 180,946
(0.002) (0.148) (0.001) (0.505)
Viability Change (2020M3–2020M6) -0.147 -0.305 -0.220 -0.778 179,353
(0.712) (0.446) (0.73) (0.174)
Viability Change (2020M3–MSLP Month) -0.503 0.409 0.129 -1.569 177,784
(0.535) (0.517) (0.855) (0.023)
% Debt 31+ Days Past Due (2020M3) 0.000 -0.007 -0.022 -0.015 46,993
(0.987) (0.298) (0.001) (0.012)
% Change Past Due (2020M3–2020M6) 0.038 0.033 0.031 0.033 43,364
(0) (0) (0.001) (0)
% Change Past Due (2020M3–MSLP Month) 0.013 0.012 0.011 0.016 41,984
(0.136) (0.136) (0.191) (0.036)
Conditioning Variables – – – –
State Yes Yes Yes Yes
Industry Yes Yes Yes Yes
Firm Age Bins No Yes Yes Yes
Employement Bins No Yes Yes Yes
Viability Score Bins (2019) No No Yes Yes
PPP Indicator No No No Yes
Note: This table compares the average value of firm attributes among the MSLP borrowers versus their
peer firms that are matched by state, 2-digit NAICS industry codes and whether a firm has no more than
500 employees. The conditional mean differences are reported, along with the p values of the equal-mean
test in parentheses. The mean difference for each variable is computed by regressing it on a set of binary
indicators for the conditioning variables, along with the MSLP borrower indicator, whose coefficient then
measures the mean difference between MSLP borrowers and the other firms controlling for the conditioning
variables. The conditioning variables for each set of mean comparison (organized by column) are reported in
the bottom block of the table. “MSLP Month” refers to the month when a borrower’s loan was submitted
to the MSLP facilities. “Min. Obs.” lists the minimum number of observations among the four sets of mean
differences reported in each row, universally corresponding to column (4). There are generally around 5%
more observations underlying results in columns (1) and (2). Source: D&B and author’s calculations.
A-9
Table B.4: Score Change as An Active-Firm Indicator
Abs. CCS Change Covid >5 Abs. Via Change Covid >5
(1) (2)
Emp. [11,50] 0.189∗∗∗ 0.062∗∗∗
(0.010) (0.006)
Emp. [51,250] 0.227∗∗∗ 0.106∗∗∗
(0.012) (0.009)
Emp. [251,500] 0.335∗∗∗ 0.134∗∗∗
(0.026) (0.016)
Emp. >500 0.401∗∗∗ 0.108∗∗∗
(0.018) (0.016)
PPP Early 0.232∗∗∗ 0.091∗∗∗
(0.008) (0.005)
PPP Mid 0.219∗∗∗ 0.093∗∗∗
(0.008) (0.005)
PPP Late 0.189∗∗∗ 0.080∗∗∗
(0.008) (0.006)
Observations 2,491,824 2,491,824
Adjusted R2 0.056 0.008
Note: This table reports correlation between having a change in CCS or Viability Score of at least 5 points
and larger firm size and being more likely to receive PPP loans earlier. The dependent variable are indicators
if the absolute value of a firm’s CCS Change (column (1)) or Viability Score (column (2)) over 2020:M3–
2020:M6 is 5 points or more. VS Change Covid: binned values of Via. Score change 2020:M3–2020:M6.
Employment (Emp.) bins: as of 2020:M3. PPP Early, Mid, Late: Indicator if a firm received PPP loans
before April 16, April 27 to May 2, and after May 2, 2020, respectively. See text for the other variable
definitions. Source: D&B and author’s calculations.
A-10
C Rare-Event Logit Model of MSLP Uptake Decision
Because MSLP firms constitute a vanishingly small fraction of all active firms in the D&B
database as of January 2019, standard logistic regressions can sharply underestimate the
probability of a rare event. We thus apply choice-based sampling to undersample non-
borrower firms and arrive at a more balanced estimation sample (see Imbens, 1992). Specif-
ically, we draw a small random sample of non-borrower firms that is approximately twice as
large as the number of MSLP firms.14 To also avoid selection on exogenous variables, this
sample is drawn by a stratum of employment bins because MSLP firms are substantially
larger than the other firms but have nearly the same distribution of risk scores and other
attributes conditioning on size bins (see Table B.2). We then apply re-weighting in the esti-
mation to account for the different sample versus population weights of MSLP firms.15 We
further adjust the coefficients to correct for the bias due to the extreme imbalance in the
dependent variable.16
We estimate a rare-event logit model with the same set of regressors as those underlying
Table 2. The coefficients are reported in Table C.1. Despite the much smaller sample (with
non-borrowers massively undersampled), the same general pattern emerges. Having more
than 250 employees and having received a PPP loan early were still the most significant
predictors of a firm becoming a Main Street borrower. Firms with the best pre–COVID-19
CCS scores showed the least propensity to borrow, much lower than the firms with the worst
CCS (the omitted category).17 Firms that experienced a CCS increase in the year before
the onset of the pandemic but a CCS decline afterward were more likely to borrow. Having
a greater presence of community banks in the county or a lower share of PPP loans delayed
were associated with a lower likelihood of borrowing from Main Street. In sum, the rare-
14
According to Imbens (1992), equal-shares sampling is close to optimal in a large number of cases. But
more data up to a point generally improve the efficiency of estimates; for example, King and Zeng (2001)
advocate a number of observations of yi = 0 that is two to five times that of yi = 1. In unreported results,
we verify that altering the ratio of non-borrowers to borrowers between 2 and 5 makes little difference to the
coefficient estimates.
15
Specifically, a weighted log likelihood function is estimated, with πi = P rob(yi = 1), while w1 = τ /ȳ and
w0 = (1 − τ )/(1 − ȳ) denote the weights for yi = 1 and yi = 0, respectively, where τ and ȳ are the fraction of
units for which the dependent variable equal to one (which corresponds to being a borrower in this case) in
the population and the sample, respectively. For more details, see King and Zeng (2001). In our application,
τ is known.
16
The bias adjustment formula is (X 0 W X)( − 1)X 0 W ξ. X is the regressor matrix, W is a n × n diagonal
weight matrix with each element wi = π̂i (1 − pi ˆ i )wi0 , where π̂i denotes the fitted probability before the bias
0
correction, and wi is the ratio of the population weight to the sample weight of observation i’s class (either
0 or 1). ξ is a n × 1 vector, with each element ξi = 0.5Qii [(1 + w1 )π̂i − w1 ], with Qii being the diagonal
elements of Q = X(X 0 W X)( − 1)X 0 (for details, see King and Zeng, 2001).
17
Coefficients on CCS in the range of 300 to 600 become insignificant once all the regressors are included
and are thus omitted to save space.
A-11
event logit estimates confirm the findings from the linear probability estimate regarding the
key factors that determined the odds of a firm becoming an MSLP borrower.
A-12
Table C.1: Logit Estimates of Firm MSLP Uptake Decision
(1) (2) (3) (4)
CCS >600 0.386 −0.780∗∗ −0.846∗∗ −0.752∗
(0.425) (0.389) (0.420) (0.403)
CCS Change Pre <-20 0.816∗∗∗ 0.580∗∗∗ 0.590∗∗∗ 0.556∗∗∗
(0.103) (0.194) (0.180) (0.186)
CCS Change Pre [-20,0) 0.077 0.362∗∗ 0.421∗∗ 0.394∗
(0.085) (0.177) (0.214) (0.227)
CCS Change Pre (0,20] 0.662∗∗∗ 0.594∗∗ 0.730∗∗∗ 0.803∗∗∗
(0.105) (0.246) (0.256) (0.259)
CCS Change Pre >20 1.067∗∗∗ 1.084∗∗∗ 1.321∗∗∗ 1.372∗∗∗
(0.060) (0.219) (0.223) (0.165)
CCS Change Covid <-20 1.547∗∗∗ 0.960∗∗∗ 0.883∗∗∗ 0.843∗∗∗
(0.184) (0.214) (0.260) (0.264)
CCS Change Covid [-20,0) 0.440∗∗∗ 0.466∗∗ 0.326 0.311
(0.145) (0.231) (0.252) (0.282)
CCS Change Covid (0,20] 0.895∗∗∗ 0.195 0.176 0.142
(0.097) (0.152) (0.160) (0.163)
CCS Change Covid >20 1.377∗∗∗ 0.762∗∗∗ 0.726∗∗ 0.754∗∗∗
(0.148) (0.285) (0.288) (0.285)
PPP Early 2.534∗∗∗ 2.696∗∗∗ 2.772∗∗∗
(0.160) (0.165) (0.187)
PPP Mid 1.122∗∗∗ 1.172∗∗∗ 1.133∗∗∗
(0.197) (0.179) (0.181)
PPP Late −0.801 −1.055 −0.939
(0.610) (0.704) (0.653)
Emp. [11, 50] 1.305∗∗∗ 1.369∗∗∗ 1.374∗∗∗
(0.232) (0.277) (0.273)
Emp. [51,250] 2.356∗∗∗ 2.409∗∗∗ 2.528∗∗∗
(0.292) (0.350) (0.310)
Emp. [251,500] 3.276∗∗∗ 3.311∗∗∗ 3.429∗∗∗
(0.487) (0.496) (0.407)
Emp. >500 4.118∗∗∗ 3.775∗∗∗ 3.933∗∗∗
(0.412) (0.469) (0.438)
Community Bank Share −1.538∗∗
(0.733)
PPP Loans Share Delayed 2.245∗∗
(1.119)
Observations 7,428 7,428 7,428 7,428
Log Likelihood −1.722 −1.716 −1.712 −1.709
Note: This table reports coefficients from a rare-event logit model. CCS Range and Employment (Emp.)
bins: as of 2020:M3. CCS Change Pre: binned values of CCS change 2019:M3–2020:M3. CCS Change
Covid: binned values of CCS change 2020:M3–2020:M6. PPP Early, Mid, Late: Indicator if a firm received
PPP loans before April 16, April 27 to May 2, and after May 2, 2020, respectively. See text for the other
variables’ definitions. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively. Source:
D&B and author’s calculations.
A-13
D Additional Estimates of Linear Model of MSLP Uptake
This section reports additional linear probability estimates of Main Street uptake. Instead
of the CCS, these estimates use Viability Score bins to measure firms’ pre–COVID-19 sound-
ness. All the other explanatory variables remain the same as in the regressions underlying
Table 2, which use the CCS as the financial indicator. These estimates reveal that the coef-
ficients on the other regressors remain quite similar regardless of the financial variable used.
We discuss the estimates from the regression with all the regressors, presented in column
(4). Being a midsize firm and having prior banking relationships remain the most important
predictors of MSLP uptake. Being in an industry less severely impacted by COVID or a
county with greater mobility becomes associated with a slightly lower likelihood of uptake,
whereas operating in an essential industry or county COVID-19 death rates each become
an insignificant determinant.18 In addition, a higher community bank deposit share in the
county or a lower share of PPP loans delayed lower MSLP uptake.
At the same time, the coefficients on the VS bins confirm the pattern found using the
CCS—firms with poor or mediocre risk scores were more likely to seek Main Street funding.
Firms with an average 2019 Viability Score in the 7 to 8 range were most likely to borrow,
followed by those with a VS in the 2 to 6 range, more so than those with the best VS (equal
to 1, the omitted category). At the same time, firms that saw their VS Points increase over
the year until March 2020, especially if the increase was more than 20 points, showed greater
propensity to borrow. On the other hand, firms that suffered a decline in their VS Points
after COVID-19 hit (March through June 2020), especially if the loss exceeded 20 points,
were more likely to borrow.
18
The sum of the coefficients on the mobility index is reported in the bottom portion of Table D.1, while
the (insignificant) sum of the coefficients on death rates is omitted to save space.
A-14
Table D.1: Probability of MSLP Uptake: Financial Health Based on Viability Score
(1) (2) (3) (4)
Via. Score Range (1,6] −0.0003∗ 0.001∗∗∗ 0.001∗∗∗ 0.001∗∗∗
(0.0002) (0.0004) (0.0004) (0.0004)
Via. Score Range (6,8] −0.0002 0.002∗∗∗ 0.002∗∗∗ 0.002∗∗∗
(0.0002) (0.0004) (0.0004) (0.0004)
VS Change Pre <-20 0.0001∗∗∗ 0.00002 0.00003 0.00003
(0.00004) (0.00002) (0.00002) (0.00002)
VS Change Pre [-20,0) 0.0002∗∗∗ −0.00000 0.00001 0.00001
(0.00004) (0.00003) (0.00003) (0.00003)
VS Change Pre (0,20] 0.0002∗∗∗ 0.0001∗∗ 0.0001∗∗ 0.0001∗∗
(0.00005) (0.00002) (0.00003) (0.00003)
VS Change Pre >20 0.0003∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗
(0.0001) (0.00003) (0.00004) (0.00004)
VS Change Covid <-20 0.0002∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗
(0.0001) (0.00004) (0.00004) (0.00004)
VS Change Covid [-20,0) 0.0002∗∗∗ 0.0001∗∗ 0.0001∗∗ 0.0001∗∗
(0.0001) (0.00003) (0.00003) (0.00003)
VS Change Covid (0,20] 0.0003∗∗∗ 0.0001∗∗∗ 0.00005∗∗ 0.00004∗
(0.0001) (0.00002) (0.00002) (0.00003)
VS Change Covid >20 0.0001∗∗∗ 0.00004 0.00003 0.00002
(0.00003) (0.00003) (0.00002) (0.00002)
PPP Early 0.002∗∗∗ 0.002∗∗∗ 0.002∗∗∗
(0.0005) (0.0005) (0.0005)
PPP Mid 0.0004∗∗∗ 0.0004∗∗∗ 0.0004∗∗∗
(0.0001) (0.0001) (0.0001)
PPP Late 0.0001 0.0001 0.0001
(0.0001) (0.0001) (0.0001)
Emp. [11,50] 0.001∗∗∗ 0.001∗∗∗ 0.001∗∗∗
(0.0002) (0.0002) (0.0002)
Emp. [51,250] 0.004∗∗∗ 0.004∗∗∗ 0.004∗∗∗
(0.001) (0.001) (0.001)
Emp. [251,500] 0.006∗∗∗ 0.006∗∗∗ 0.006∗∗∗
(0.002) (0.002) (0.002)
Emp. >500 0.005∗∗∗ 0.005∗∗∗ 0.005∗∗∗
(0.002) (0.002) (0.002)
Essential Industry −0.0001 −0.0001
(0.0001) (0.0001)
COVID Impacted 0.0001∗ 0.0001∗
(0.0001) (0.0001)
Community Bank Share −0.0002∗∗
(0.0001)
PPP Loans Share Delayed 0.0005∗∗
(0.0002)
Sum mobility coefs (p val) – – -0.002 (0.065) -0.002 (0.069)
Observations 418,995 418,995 418,995 418,995
Adjusted R2 0.0001 0.002 0.003 0.003
Note: This table reports coefficients from Equation (1). Via. Score Range and Employment (Emp.) bins:
as of 2020:M3. VS Change Pre: binned values of Via. Score change 2019:M3–2020:M3. VS Change Covid:
binned values of Via. Score change 2020:M3–2020:M6. PPP Early, Mid, Late: Indicator if a firm received
a PPP loan before April 16, April 27 to May 2, and after May 2, 2020, respectively. See text for the other
variable definitions. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively. Source:
D&B and author’s calculations.
A-15
E Additional Callaway-Sant’Anna Estimates of MSLP’s Effects
Figure E.1: Effects of Main Street on Firm Viability and Financial Stress Score
(a) MSLP Effects on Viability Score Points
(b) MSLP Effects on Financial Stress Score (FSS)
Note: This figure depicts the dynamic effects of the staggered treatment of MSLP loans on borrowers’
Viability Score (top panel) and Financial Stress Score (bottom panel), estimated using the doubly robust
DiD estimator à la Callaway and Sant’Anna (2021). Source: D&B and author’s calculations.
A-16
F MSLP Treatment Effects Using Sun-Abraham Estimator
This section applies the estimator developed by Sun and Abraham (2021) (SA) to derive an
alternative estimate of the treatment effects of receiving MSLP loans. Like the estimator
by Callaway and Sant’Anna (2021), the SA estimator can also account for the staggered
design and provide a consistent aggregation of the coefficients for firms receiving loans in
different quarters. However, not being a doubly robust estimator by incorporating inverse-
probability weighting, the SA estimator is susceptible to misspecification of the outcome
regression. As will be shown, even with matching between the MSLP borrowers and their
peers along as many dimensions as feasible given the data, the borrowers still display a
different pre-trend compared with their non-borrower peers. On the other hand, the SA
regression can be estimated rapidly using high-dimension fixed-effects estimators optimized
for parallel computing. It is therefore suitable for estimating multiple specifications with
different sets of regressors.19 Hence, the results in this section are meant to explore the
relative contribution of subsets of regressors, and to corroborate the estimates à la Callaway
and Sant’Anna (2021).
Specifically, we estimate the following regression:
XX
yf,t = ψc,τ + αf + γt + ζs,t + ηn,t + θa,t + κe,t + µt × VSf,2019 + εf,t , (F.1)
τ 6=−1 c
where yf,t is an outcome y of firm f in quarter t, and ψc,τ are the coefficients on the in-
teraction terms between indicators for cohort (c) and event time—defined as time relative
to treatment—(τ ). Each cohort is defined by the quarter in which the constituent firms
received MSLP loans. The omitted category is firms that did not receive MSLP loans. ψc,τ
is estimated while controlling for several sets of fixed effects. The fully saturated specifica-
tion includes these fixed effects: firm (αf ), time (γt ) (if the following time-interacted terms
are not included), state-time (ζs,t ), industry-time (ηn,t ), age-bin-time (θa,t ), employment-bin-
time (κe,t ), and time interacted with the firm’s 2019 average Viability Score bin, VSf,2019 .
Estimates of δc,τ for the CCS and employment with fixed effects of only firms’ nonfinan-
cial attributes—state, industry (two-digit NAICS), age bin, employment bin, and the PPP
indicators, all interacted with quarter indicators—are plotted in Figure F.1. Corresponding
estimates of δc,τ with fixed effects of all firm attributes (that is, adding indicators of pre–
COVID-19 Viability Score bins and COVID-19–period change in Viability Score bins, again
both interacted with quarter indicators) are depicted in Figure F.2.
19
All the estimates in this section are produced using the fixest package in R (see https://cran.r-
project.org/web/packages/fixest/).
A-17
Panel (a) of Figure F.2 reveals that, when compared with only non-borrower peers in
the same state, industry, and small- or large-business category and controlling for all the
firm attributes, the MSLP borrowers display a parallel trend of CCS before the pandemic,
albeit at a higher level (on average better by 10 points, mapping into about 16 percent lower
delinquency risk). Their CCS fell to the level of the peer firms after the pandemic hit, in
the quarter before receiving the MSLP loan and the quarter of the loan.20 The negative
CCS CCS
gap between δc,−1 (normalized to zero here) and δc,τ |τ < −1 is qualitatively similar to the
pattern observed in panel (a) of 1 (where the treatment effect for τ = −1 is negative and
not normalized to zero). This is corroborative evidence that Main Street borrowers likely
suffered a more adverse effect on their cash flow due to the pandemic and thus had greater
difficulty paying their bills. By comparison, MSLP borrowers display a slightly upward trend
in employment (especially for the cohort of MSLP firms that borrowed in 2020:Q4) before
the pandemic relative to their peers (panel (b) of F.2). This is suggestive evidence that they
were likely regarded as having better growth prospects before the pandemic but were then
more severely disrupted by it.
Borrowers experienced a relative recovery in their CCS scores, which grew progressively
in magnitude over the four quarters after they received their loan and almost fully regained
the ground they had lost to the CCS scores of borrowers’ peers. The size of the estimated
treatment effects for the 2020:Q4 borrower cohort is fairly similar to one found when using the
Callaway-Sant’Anna estimator, although the SA estimates appear more significant because
of the smaller standard errors.21 Borrowers’ employment also grew faster than at the peer
firms over the post-MSLP year, especially in the last two quarters. These estimates are of
similar magnitude as those from the CS estimator but appear more significant (especially for
the 2020:Q4 borrowers) again due to the narrower confidence bands. All things considered,
it is reassuring that the two somewhat different estimators for staggered treatments yield
basically the same point estimates.22 Using the more conservative simultaneous confidence
band, we conclude that Main Street helped shore up borrowers’ financial position by reducing
delinquency risk (by 18 percent), but it did not boost employment significantly over the year
after the loan disbursement.
Comparing the estimates in Figures F.1 and F.2 reveals that accounting for the VS-based
financial variables hardly alters the estimates of Main Street’s effects on the borrowers’ CCS
20
The horizontal axis depicts the event quarter—relative to the loan submission quarter—and the pre-event
quarter (τ = −1) is omitted (thus equivalent to being normalized to zero).
21
The confidence bands from the SA estimator are generally narrower than those from the CS estimator
because the former are point-wise while the latter are simultaneous bands for all the time periods used in
the estimation and thus robust to multiple testing.
22
The treatment effects estimates for the Viability Score and the FSS are also similar to those from the
CS estimator, which are omitted here for brevity but are available upon request.
A-18
or employment. This is likely because the regression sample is already restricted to only
those non-borrowers that belonged to the same broad pre–COVID-19 Viability Score bins
as the borrowers.
Figure F.3 plots the estimated effects of Main Street on borrowers’ CCS for each cohort
of borrowers grouped by the loan submission month. The broad picture is highly similar to
the one in panel (a) of Figure F.2, although it also reveals some interesting patterns for firms
that borrowed in different months. In particular, July 2020 borrowers exhibited the largest
gain in CCS after receiving a loan, consistent with the interpretation that they tapped into
the program early because their liquidity squeeze was the most severe and thus they also
benefited the most. Not surprisingly, the effects on the December 2020 cohort are the most
precisely estimated because of the much larger sample size.
A-19
Figure F.1: Effects of Main Street on Firm Commercial Credit Score and Employment (Fixed
Effects of Only Nonfinancial Attributes)
(a) MSLP Effects on Commercial Credit Score
(b) MSLP Effects on Employment
Note: This figure depicts the dynamic effects of the staggered treatment of MSLP loans on borrowers’
Viability Score (top panel) and Financial Stress Score (bottom panel), estimated using the Sun-Abraham
estimator as in Equation (F.1). Source: D&B and author’s calculations.
A-20
Figure F.2: Effects of Main Street on Firm Commercial Credit Score and Employment (Fixed
Effects of All Attributes)
(a) MSLP Effects on Commercial Credit Score
(b) MSLP Effects on Employment
Note: This figure depicts the dynamic effects of the staggered treatment of MSLP loans on borrowers’
Viability Score (top panel) and Financial Stress Score (bottom panel), estimated using the Sun-Abraham
estimator as in Equation (F.1). Source: D&B and author’s calculations.
A-21
Figure F.3: Effects of Main Street on Firm Commercial Credit Score: Monthly Estimate
A-22
Note: This figure depicts the dynamic effects by month of the staggered treatment of MSLP loans on borrowers’ Commercial Credit Score, estimated
using the Sun-Abraham estimator as in Equation (F.1). Source: D&B and author’s calculations.
A-23
File and source
- File
- the-main-street-lending-program-who-borrowed-and-how-have-they-benefited.pdf
- Size
- 1,636,462 bytes
- SHA-256
- dea15adb46abc906316fb505a225a068e5249876f2343c467b5caa1b9e1735c8
- Original
- doi.org