Screen Now, Save Later? The Trade-Off between Administrative Ordeals and Fraud (August 18, 2022)
Summary
A paper, Screen Now, Save Later? The Trade-Off between Administrative Ordeals and Fraud, by Shan Aman-Rana, Daniel Gingerich and Sandip Sukhtankar of the University of Virginia, dated August 18, 2022. The abstract states that the authors examine trade-offs of screening requirements in the Paycheck Protection Program (PPP), using a database that includes nearly 11.5 million PPP loans and variation in the documentation standards applied to loan applications of different values. It states that screening significantly reduced measures of loan irregularities indicative of fraud and that a subset of borrowers with a checkered history strategically reduced their loan application amounts to avoid screening. It also states that the estimates imply that screening led to a $737 million reduction in losses due to fraud.
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
- Screen Now Save Later The Trade Off Between Administrative Ordeals And Fraud
- Case
- Screen Now Save Later The Trade Off Between Administrative Ordeals And Fraud
Cite this document
“Screen Now, Save Later? The Trade-Off between Administrative Ordeals and Fraud (August 18, 2022).”
Original at doi.org. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/screen-now-save-later-the-trade-off-between-administrative-ordeals-and-fraud/
, SHA-256 e6ea638dd7eed420a0bb11fd35976df88e2ff6e7d540a8a286fec6fb2fae55ee.Full text
Screen Now, Save Later? The Trade-Off between
Administrative Ordeals and Fraud∗
Shan Aman-Rana† Daniel Gingerich‡ Sandip Sukhtankar§
University of Virginia University of Virginia University of Virginia
August 18, 2022
Abstract
Screening requirements are common features of fraud and corruption mitigation efforts around
the world. Yet imposing these requirements involves trade-offs between higher administrative
costs, delayed benefits, and exclusion of genuine beneficiaries on one hand and lower fraud on
the other. We examine these trade-offs in one of the largest economic relief programs in US
history: The Paycheck Protection Program (PPP). Employing a database that includes nearly
11.5 million PPP loans, we assess the impact of screening by exploiting temporal variation in
the documentation standards applied to loan applications for loans of different values. We find
that screening significantly reduced the incidence and magnitude of various measures of loan
irregularities that are indicative of fraud. Moreover, our analysis reveals that a subset of borrowers
with a checkered history strategically reduced their loan application amounts in order to avoid
being subjected to screening. Borrowers without a checkered history engaged in this behavior at a
much lower rate, implying that the documentation requirement reduced fraud without imposing
an undue administrative burden on legitimate firms. All told, our estimates imply that screening
led to a $737 million reduction in losses due to fraud.
JEL codes: D22, D73, G38, H25, H32, H81
Keywords: Fraud, screening, monitoring, ex-ante, firms, CARES act, COVID relief, business loans
∗ We thank Jenny Le, Ashwin Nair, Rui Cao, and Sasha Ruby for excellent research assistance. We are grateful
to Prabhat Barnwal, Kerem Cosar, Leora Friedberg, James Harrigan, Tarun Jain, Justin Kirkland, Lee Lockwood, Luis
Martinez, John Mclaren, Karthik Muralidharan, Yusuf Neggers, Paul Niehaus, Sam Norris, Rachel Potter, Sheetal Sekhri,
Sebastian Tello-Trillo, Laura Zimmerman and seminar participants at multiple universities, conferences, and workshops
for comments and suggestions. The authors are grateful for financial support from the CLEAR Lab (Democracy
Initiative) at UVA. All mistakes are our own.
† Department of Economics, University of Virginia. sa8ey@virginia.edu.
‡ Department of Politics, University of Virginia. dwg4c@virginia.edu.
§ Department of Economics, University of Virginia. sandip.sukhtankar@virginia.edu.
1 Introduction
Corruption and fraud plague public programs around the world (Glaeser and Saks, 2006; Olken
and Pande, 2012; Finan et al., 2017).1 Public benefits may be captured by ineligible beneficiaries,
or beneficiaries may obtain benefits greater than those to which they are entitled (Becker et al.,
2005; Olken, 2006; Niehaus and Sukhtankar, 2013; Fang and Gong, 2017). To prevent this leakage,
governments can impose screening requirements ex-ante, potentially including small “ordeals” in
order to induce ineligible applicants to select out of the program before injury to the public fisc
has taken place (Nichols and Zeckhauser, 1982; Besley and Coate, 1992).2
However, these requirements may reduce the timeliness of delivery, impose added costs for
both the government and beneficiaries, and could lead to the exclusion of legitimate beneficia-
ries (Currie, 2006; Kleven and Kopczuk, 2011; Herd and Moynihan, 2018). Whether the costs of
higher administrative expenses, delayed benefit transfers, and increased exclusion (Type I error)
are higher than the savings from lower inclusion (Type II) errors is ultimately an empirical ques-
tion. Prior evidence is mixed. Alatas et al. (2016) find that small ordeals improved the targeting
of benefits by dissuading ineligible beneficiaries from applying for cash transfers in Indonesia.
Yet in the US, Deshpande and Li (2019) find that application costs reduce targeting efficiency by
dissuading poorer and more severely disabled beneficiaries from applying for disability benefits,
and Gray (2019) and Homonoff and Somerville (2021) similarly show that reporting and interview
requirements undermine retention among legally eligible households in the Supplemental Nutri-
tion Assistance Program (SNAP). Understanding and identifying when costs of the administrative
requirements are low relative to their value as a tool for sanctioning fraud is thus important in
maintaining the integrity and viability of public programs.
We examine these issues in one of the largest economic relief programs in US history: The Pay-
check Protection Program (PPP), a $814-billion small business stimulus program adopted as part
of the Coronavirus Aid, Relief, and Economic Security (CARES) Act signed into law on March 27,
2020. The PPP permitted small businesses, nonprofits and certain other entities, typically with 500
or fewer employees, to apply for federally-backed loans administered by banks and other private
lenders on behalf of the Small Business Administration (SBA). These loans in essence operated as
grants, since program rules stipulated that repayment would not be required of borrowers as long
as the funds were used for purposes prescribed by the federal government.3 Moreover, there were
weak incentives for due-diligence by lenders, since loans were 100% guaranteed by the SBA if the
borrower defaulted for “legal purposes”.4 The primary oversight mechanism was ex-post auditing
1Starting with the seminal work by Becker and Stigler (1974), a number of scholars have put forth conceptual
frameworks to assist our understanding of corruption, including but by no means limited to Shleifer and Vishny (1993);
Banerjee (1997); Banerjee et al. (2012).
2An alternative to ex-ante screening is to deliver benefits first and audit ex-post. Auditing serves a punitive, and
ipso facto, deterrent function. While a large literature demonstrates that rigorous ex-post auditing reduces corruption
(Di Tella and Schargrodsky, 2003; Olken, 2007; Bobonis et al., 2016; Avis et al., 2018; Zamboni and Litschig, 2018), our
focus in this paper is on the marginal value of screening in addition to performing audits.
3Loan forgiveness for eligible businesses was built into the program through Section 1106, CARES Act, 2020.
4Section 1102, CARES Act, 2020. Lenders were held harmless if they acted in good faith and all documents were
1
by the SBA. The PPP was divided into two distinct phases: phase 1 (April 2020 to August 2020)
and phase 2 (January 2021 to May 2021).
In order to study the impact of screening requirements on loan irregularities, we exploit a sudden
change in the rule regarding proof of loan eligibility announced in phase 2 of the PPP.5 The new
rule stipulated that all firms who had previously received a PPP loan and were requesting loans
greater than $150K were required to submit with their PPP application documentation proving
that they had experienced a reduction in gross receipts in excess of 25% in 2020 relative to 2019.
Firms requesting loans of ≤$150K were also only legally eligible to receive funds from the program
if they experienced a 25% or more reduction in gross receipts. Yet they were not required to
provide up front to the lender any documentation proving this in order to have their applications
processed. In phase 1, there were no differences in the documentation that firms were required to
provide up front to lenders based upon the size of the loan for which they were applying.6
Given variation in screening by program phase and loan value, we employ a difference-in-
differences estimation strategy to compute the impact of screening on serious irregularities in the
receipt of loans from the PPP. Since our data consists of the full corpus of 11.5 million PPP loans
allocated over the course of the two phases of the program, this allows us to use the universe of
borrowers in phase 1 to determine the set of firms that were subjected to the advance documentation
requirement in their second loan applications. We define our treated firms as those that had
received a loan greater than $150k in phase 1, i.e., before the announcement of the rule change, and
reapplied in phase 2. The control firms are those firms that received a loan ≤$150k in phase 1 and
reapplied in phase 2.
We utilize as primary outcomes three measures of irregularities that are indicative of (though not
dispositive of) fraudulent behavior on the part of the borrower: the receipt of a loan that exceeded
the maximal permissible payment given firm characteristics presented in the loan application, the
amount of overpayment expressed as a fraction of the maximal permissible payment (overpayment
rate), and the receipt of multiple loans from the PPP during a single phase of the program. For each
measure, we find an economically meaningful and statistically significant reduction in irregular
loans attributable to the screening requirement.
To investigate the mechanisms underlying these results, and to better understand the trade-off
between the administrative burden of complying with upfront documentation requirements and
fraud prevention, we present a simple conceptual framework to demonstrate the expected impact of
the rule change. In our model, legitimate and fraudulent firms both face some administrative costs
of complying with either upfront or ex-post documentation requirements; the costs for fraudulent
firms increase discretely with an upfront requirement given the prospect of legal sanction upon
discovery of their fraud. We show that firms will respond to the documentation requirements
complete.
5The SBA made an announcement on January 6, 2021; the change was effective for loans made after January 14 2021
(13 CFR Parts 120 and 121).
6Other requirements also more or less remained the same across phases 1 and 2, as well as for loans above and below
$150K.
2
by strategically requesting less than $150K in phase 2, leading to “bunching” of the frequency
distribution under $150K; however, if the costs of the administrative requirements are low relative
to their value as a tool for sanctioning fraud, the bunchers will be composed disproportionately of
fraudulent firms.
Consistent with our framework, we find that borrowers with loan irregularities in phase 1 who
would have been subject to screening in phase 2 responded to the new upfront documentation
requirement by reducing their loan requests (as opposed to exiting the program). We explore
this intensive margin response further and find strong evidence of the strategic evasion of the
documentation requirements associated with screening. In phase 2 of the program, we observe
bunching in the number of borrowers receiving loans right below and just at the $150K threshold
that determined the use of screening. In phase 1, by contrast, there was no discontinuity at this
threshold. This suggests that many borrowers strategically set their loan requests just below the
threshold to avoid submitting documentation.
The shift in the locus of fraud to lower loan values resulted in considerable savings to the PPP: We
estimate that it reduced overpayment by $737 million, or 88% of the total reduction in overpayment
between phases 1 and 2.7 Moreover, even firms that appear to have strategically evaded screening
nonetheless engaged in less fraudulent activity once it had been introduced. The strong behavioral
reaction to screening that we document, concentrated as it was among borrowers with previous
loan irregularities, implies that screening was perceived of as a genuine risk for borrowers who
may have been defrauding the program. Simply applying the same screening requirement in
phase 1 would have resulted in overall savings of $1.5 billion, ceteris paribus. An across-the-board
advance documentation requirement - for all loan values - would likely have had an even more
significant impact in reducing losses due to fraud.
Our results indicate that substantial savings may be generated by forms of screening that impose
minimal bureaucratic hurdles. The administrative burden of providing proof of eligibility in
phase 2 of the PPP was quite low. Any legally functioning enterprise that pays taxes must of
necessity have documentation that could be used to show recent changes in revenues. Thus,
simply requiring program participants to present documents which they must have in order
to satisfy their other legal obligations–a minimal ordeal–can be highly effective in reducing the
participation of fraudulent actors without imposing an undue burden on legitimate participants.
These findings have implications that extend beyond the context of the US federal government’s
Covid-19 relief efforts. This includes programs for which the timeliness of relief is essential, the
potential participants are large in number, and the capacity to detect fraudulent intent is limited.
For instance, emergency relief programs have these features.8 Another example of a critical
7Amounts calculated based upon the estimates presented in Table A4.
8The incidence and scope of such relief programs is only likely to grow in the future, not just due the prospect of
new variants of Covid or future pandemics, but due to the realities of climate change. The warming climate has led to
more extreme natural disasters, with concomitantly greater economic costs (Coronese et al., 2019; Estrada et al., 2015).
Evidence suggests that it has also reduced economic growth (Dell et al., 2012; Kotz et al., 2021). Given this environmental
and economic reality, the organization of emergency relief is poised to become an increasingly salient responsibility of
government.
3
government task that shares these features is tax collection.9
Our paper contributes to the literature on screening requirements for public benefits. The
theoretical literature (Nichols and Zeckhauser, 1982; Besley and Coate, 1992; Kleven and Kopczuk,
2011) clearly lays out the trade-offs involved. The empirical literature finds mixed evidence. In
addition to the papers cited above, Finkelstein and Notowidigdo (2019) show that an intervention
designed to reduce the cost of applying to the SNAP in the US mostly benefited richer beneficiaries,
although they suggest that this reflected the poor targeting of the intervention rather than the
general properties of screening requirements. Our results suggest that in general, in settings
where program scope or timing are such that rigorous ex-post auditing systems are costly and/or
infeasible, screening can be a particularly valuable tool in mitigating fraud. This is especially the
case if the documentation requirements that characterize screening are easy for legitimate program
participants to satisfy at a reasonable cost.
Our paper also relates to the broader literature on regulation in law and economics.10 In
particular, our work has a natural link to studies that compare ex-ante regulation with ex-post
enforcement of harmful behavior, both in the private and public sectors. There is a long-standing
and large theoretical literature that describes the conditions under which screening/regulation and
auditing/enforcement act as complements or substitutes.11 In contrast, the empirical literature is
extremely limited. To the best of our knowledge the only study that empirically investigates
this issue is Behrer et al., 2021, who show that water quality improved when ex-post oversight
mechanisms were replaced by ex-ante regulation by the 1972 Clean Water Act (CWA). In a related
vein, the distinction between ex-ante and ex-post monitoring features prominently in the political
science literature on legislative oversight of the executive branch. The central concern of this
literature is understanding how legislators can mitigate opportunism by bureaucrats in light of
informational asymmetries and the prospect of ex-post monitoring by constituents and interest
groups (McCubbins and Schwartz, 1984; Lupia and McCubbins, 1994). Our contribution to these
related bodies of work lies in empirically establishing an approach for assessing the trade-offs
entailed by greater ex-ante regulation, which we apply to a government program of unprecedented
9In the five year period from 2015-2019, the audit rate of individual tax returns was only 0.6% (TIGTA, 2021),
suggesting that the risk of a formal audit may be unlikely in and of itself to dissuade many actors from engaging in tax
evasion. Yet the payment of taxes requires ample ex-ante documentation (e.g., W2 forms for wage earners), facilitating
the use of automated systems to detect discrepancies or irregularities. A large literature examines the role of this type of
third-party document reporting in tax compliance; see for example Kleven et al. (2011); Pomeranz (2015). These upfront
documentation requirements introduce additional hurdles and risk for those who would seek to evade their taxes, and
surely contribute to broad compliance in spite of a low incidence of formal auditing.
10A large theoretical body of work investigates optimal regulation. Laffont, 1994 and Estache and Wren-Lewis, 2009
present an excellent review of the key ideas. Some studies have focused on the possibility of collusion between the
regulator and the regulated (Stigler, 1971; Posner, 1974; Burgess et al., 2012; Jia and Nie, 2017). Along these lines, several
papers study the problem of regulation from the lens of incentives of regulators (Glaeser et al., 2001; Duflo et al., 2013,
2018), while others investigate how the selection of the regulator affects social welfare (Besley and Coate, 2003).
11Theoretically, whether ex-ante regulation is better than ex-post control depends on, for example, the relative costs
of enforcement (De Chiara and Livio, 2017; Strausz, 2005), in particular transaction costs (Coase, 1960) and whether
there are fixed cost of lawsuits (Posner, 1998); heterogeneity in offense severity and limits on the violator’s ability to
pay (Shavell, 1984a,b); the degree of uncertainty in potential harm and uncertainty in whether and to what extent the
legal system will penalize the violator (Kolstad et al., 1990; Mookherjee and Png, 1992); and the possibility of ex-post
subversion of justice by the potential violator (Glaeser and Shleifer, 2003).
4
scope.
Finally, our paper also contributes to a small but growing literature on COVID relief funds,
in particular the PPP. Much of this literature examines its impact on employment and business
survival, with some evidence that it boosted both outcomes but debate over magnitudes (Hubbard
and Strain, 2020; Autor et al., 2020; Granja et al., 2020; Chetty et al., 2020); as well as on appropriate
targeting, with evidence that larger firms were better able to access the program (Bartik et al., 2020;
Humphries et al., 2020; Balyuk et al., 2021). Our paper is most closely related to two studies that
examine fraud in the PPP, with Griffin et al. (2021) suggesting that FinTech lenders were responsible
for much of this fraud, while Beggs and Harvison (2021) find that 6% of PPP funds that went to
investment management firms likely consisted of overallocations. Our paper corroborates much of
these two papers’ findings about systemic fraud, while highlighting the importance of institutional
design.
2 Background
A reaction to the economic disruptions created by Covid-19, the PPP was designed to provide
small businesses with large influxes of money in a very short period of time. The first phase of the
program was established under the CARES Act and lasted from April 2020 until August 2020. The
Economic Aid Act established the second phase of the PPP program that lasted from January 2021
until May 2021. The second phase of the program operated under the same terms and conditions
as phase 1 with a few important exceptions that we outline below.
Although nominally structured as loans, loan forgiveness for eligible businesses was built into
the program and widely advertised across both phases of the program.12 The monies disbursed
under the program did not need to be repaid if used for certain purposes (such as payroll costs,
payments on covered mortgage obligations, payments on covered lease obligations, or covered
utility payments).13 The program was formally managed by the Small Business Administration
(SBA), an independent federal agency. Private-sector financial institutions (henceforth lenders)
played a central role as intermediaries in the program.
The program had several eligibility criteria for borrowers under the PPP. Since the program was
oriented towards small businesses, there were ceilings on the size of firms: Eligible firms had
to employ five hundred employees or less in phase 1 and three hundred or less in phase 2.14 In
terms of economic criteria, firms applying in phase 1 of the program had to certify that “current
economic uncertainty makes this loan request necessary to support the ongoing operations of the
Applicant.” To receive a PPP loan for the second time, eligibility was limited to firms that had
experienced a reduction in gross receipts in excess of 25% in 2020 relative to 2019, but, as explained
above, only firms requesting loans above $150K in value were required to provide documentation
12Section 1106, CARES Act 2020.
13There was a penalty on firms if they retained fewer workers or reduced their total wages in excess of 25%.
14The ceiling was set at three hundred employees in phase 1 for housing cooperatives, member-based professional
organizations, and tourist boards.
5
of this as part of the application.15 These documents could include relevant tax forms, including
annual tax forms, or, if relevant tax forms are not available, quarterly financial statements or
bank statements could be used.16 Others had to retain these documents and could be asked for
these later by the SBA or if the borrower applied for loan forgiveness. The SBA announced the
change in the documentation requirement for loans greater than 150K on January 6, 2021 via the
release of a rule change; it became effective for second time loans made after January 14, 2021.17
This announcement came five months after the conclusion of the first phase of the PPP program
(August 8, 2020).
Figure 1 describes the timeline of the Paycheck Protection Program with key events and attempts
to mitigate fraud.
4/3/2020 8/8/2020 1/11/2021 5/31/2021
Start of Phase 1 End of Phase 1 Start of Phase 2 End of Phase 2
6/20/2020 1/6/2021
SBA announced it would SBA announced the document
release names of borrowers of requirement for second-draw
loans over 150K USD loans over 150K USD
4/28/2020 12/27/2020 1/14/2021
U.S Treasury Secretary The announcement The document
announced audits for PPP of Phase 2 requirement took effect
loans of at least 2 million USD
Figure 1: Timeline of the Paycheck Protection Program.
In addition to eligibility criteria, PPP rules specified the maximum loan amount that firms
applying to the program could obtain. The maximum amount a firm could receive in phase 1 of
the program was equal to the average employee compensation (salary and benefits) during the
previous twelve months multiplied by 2.5. For the purposes of this calculation, employee salaries
were capped at $100K. In this phase, no firm was permitted to receive more than $10 million
from the program. The caps stayed much the same in phase 2, with the following exceptions: 1)
restaurants and other firms within the accommodation and food services sectors could receive 3.5
times the average monthly compensation of their employees; 2) no firm could receive more than
$2 million. Rules also stipulated that borrowers could not receive multiple loans for the same
purpose.18
Given the emphasis on injecting capital in the private sector as quickly as possible, adherence to
15In phase 1 of the program documentation requirements were not linked to loan amounts.
16SBA’s Interim Final Rule 13 CFR Parts 120 and 121.
17SBA’s Interim Final Rule 13 CFR Parts 120 and 121.
18Section 1102, CARES Act 2020.
6
the eligibility criteria outlined above and even compliance with loan maximums appears to have
followed an honor system. There were weak incentives for due diligence by lenders. Lenders
provided the loans to recipients using money fully backed by the federal government. They were
charged with processing the loan applications and verifying that the proper attestations were
submitted by applicants. In exchange, the lenders received a fee from the SBA for each loan that
they administered, with the size of the fee expressed as a percentage of the loan.19 Crucially,
the CARES Act contained a “hold harmless” clause stipulating that lenders that had received
attestations from borrowers that the loans were used for authorized purposes could neither be
subject to enforcement actions nor penalties related to said loans. Moreover, since the loans were
backed by the federal government, lenders would not be on the hook in instances where borrowers
needed to repay but failed to do so. In this way, the SBA allocated significant fees to lenders for
managing loans for which they incurred zero risk.20 In total, 5,460 lenders participated in the PPP.
The default oversight mechanism was ex-post auditing by the SBA. The SBA Administrator
could review borrower eligibility for loans and loan forgiveness, loan amounts, and whether a
loan was used for the permitted purposes at any time. PPP loans that are either taken for a first
or a second time by firms could be subject to the review. To allow for ex-post auditing, borrowers
were required to retain their application and all related documents for four to six years after the
disbursement of the loan. Yet the frequency of auditing and the degree of scrutiny it entailed were
never publicly disclosed. Only loans above $2 million were guaranteed to be audited.
Evidence of widespread fraud in the program is now abundant.21 On March 26, 2021 the Justice
Department announced that it had charged nearly five hundred defendants with engaging in
fraud related to the PPP and other pandemic relief programs. Cases recently concluded with
convictions illustrate the weak financial controls instituted in the PPP and give a sense of how
the schemes operated. Dinesh Sah was sentenced in July 2021 to more than eleven years in
prison for fraudulently obtaining over $17 million in funds from the PPP, using the money to
purchase multiple homes and luxury automobiles, as well as to send millions overseas through
wire transfers. By his own admission, he had filed fifteen fraudulent applications to eight different
lenders, claiming employees and payroll expenses in his businesses that were vastly at odds with
the true figures. To give a sense of the ease of fraud, Sah registered a fake company on May 18, 2020
(well after the stipulated February 15 cutoff date) then filed a PPP application for the company
that same day. One day later, he received $2 million in his account for the non-existent firm.
Another notable case was that of former NFL wide receiver Joshua Bellamy, who was sentenced
19Until February 2021, lenders received as a fee 5% of the loan amount for loans of $350K or less, 3% for loans between
$350K and $2 million, and 1% for loans greater than $2 million. After February 2021, lenders received either a fee of
50% of the loan amount or a fixed payment of $2,500, whichever was smaller, for loans of $50K or less. For other loans,
the fee schedule remained unchanged. For loans ineligible for forgiveness, lenders also receive a 1% interest rate.
20The fees do not represent pure profits, however, due to the administrative costs of managing the loan applications.
Early reporting suggests there was significant variation across lenders in the profitability of PPP loan administration.
See Cowley (2020).
21A recent article from the New York Times bemoans the fact that prosecutors are unable to process the “tidal wave
of pandemic fraud.” https://www.nytimes.com/2022/08/16/business/economy/covid-pandemic-fraud.html, ac-
cessed August 16, 2022.
7
in December 2021 to three years in prison for fraudulently obtaining a $1.2 million loan for his
company, Drip entertainment, much of which was spent on jewelry, hotels, and other personal
expenses. The federal complaint alleges that Bellamy was part of a group of conspirators who
worked with an intermediary who prepared the loan applications and falsified paperwork in
exchange for kickbacks on the PPP loan proceeds. FinTech lenders appear to have approved a
disproportionate number of fraudulent loans identified by the Justice Department.22
In addition to fraud related to overstatements of job figures and employee compensation, there
have been reports of companies receiving more than one PPP disbursement in the same round.
On March 15, 2021, the Office of Inspector General (OIG) released a report in which it found 4,260
borrowers were approved for more than one loan as of August 31, 2020, which cost the program
approximately 692 million U.S dollars.
3 Data and variable coding
3.1 Data sources
PPP loans. Our primary data is the universe of PPP loans approved across the two phases of
the program. This data is made available by the Small Business Administration (SBA).23 The total
number of approved loans is 11,475,004 (5,136,454 in phase 1 and 6,338,537 in phase 2). The data
includes details on the names and addresses of the borrowers, the loan approval date, whether
the loan is a first or a second time loan, the borrower’s industry (NAICS codes), the number of
employees reported by the borrower, the loan amount, the status of the loan (whether paid in full
or charged off), loan maturity, whether SBA guaranteed the loan, the purpose for which the loan
is sought, the business type, the congressional district of the borrower and the names and address
(only headquarter) of the lenders. It also includes other information on the borrowers such as
race, ethnicity, gender, veteran status, whether the firm is located in a rural or urban area, and
whether the firm is a non-profit. Table A1 shows descriptive statistics of firms that received PPP
loans across the two phases. The majority of firms that received loans were urban and were either
corporations or limited liability companies. Over seventy percent of firms had 10 employees or
less.24
Dun & Bradstreet (DNB) data on number of employees. Since loan amounts were based on
the number of employees there was an incentive for firms to misreport their employee numbers
on the PPP application. To observe whether there are any such discrepancies, we use the firm
employee figures compiled by DNB. DNB maintains two sets of data on firms and their number
22For information on cases brought by the Department of Justice for PPP fraud, see the website of the COVID-19 Fraud
Enforcement Task Force, www.justice.gov/criminal-fraud/cares-act-fraud. On FinTechs and the incidence of fraud, see
Griffin et al. (2021).
23We retrieved this data from the SBA’s website on the 24th of November 2021.
24Demographic questions had a low response rate on PPP applications. However, those who did report their race
were mainly white. Similarly, more men than women received PPP loans.
8
of employees: one is verified and based on actual figures from the firms and the other is imputed
by DNB. We use the verified part of their data on 1,691,849 million PPP loan applicants.25 This
dataset was verified and updated in July 2021. For the firms in this dataset, we have a snapshot of
each firm’s number of employees both in phase 1 of the PPP program as well as in phase 2.26
3.2 Key variables
Overpayment on PPP Loans. We classify the approval of a loan that exceeded the maximum
permissible payment as an overpayment on a PPP loan. To observe such overpayments, we combine
PPP loan level data with SBA’s rule on disbursement of funds. Using information presented in
the loan applications about the number of employees and the industry in which the firm was
operating, we first compute the maximum payment for which the firm was eligible as per SBA’s
rules.
In Phase 1, the following maximum payment method was applied to every firm with employees:
𝑀 𝑎𝑥𝑖𝑚𝑢𝑚 𝑝𝑎 𝑦𝑚𝑒𝑛𝑡 = 𝑁 𝑜. 𝑜 𝑓 𝑒𝑚𝑝𝑙𝑜 𝑦𝑒 𝑒 𝑠 × (2.5 × $100, 000/12 + $9, 166),
where $100,000 is the maximum annual salary for each employee that firms can report on their
PPP applications. For self-employed workers without employees, the maximum payment was
calculated without the $9,166 that is the average benefit spending on each employee.27
In Phase 2, the maximum payment is calculated similarly for firms except those from the Ac-
commodation and Food Services industry that took out loans for a second time. For these firms,
SBA set a higher threshold. Therefore, we use the following method was instead:
𝑀 𝑎𝑥𝑖𝑚𝑢𝑚 𝑝𝑎 𝑦𝑚𝑒𝑛𝑡 = 𝑁 𝑜. 𝑜 𝑓 𝑒𝑚𝑝𝑙𝑜 𝑦𝑒 𝑒 𝑠 × (3.5 × $100, 000/12 + $9, 166)
We then compare the maximum payment due with the actual approved amount and define
overpayment on PPP loans in three different ways. The first is an indicator variable, overpayment
dummy, equal to 1 if any of a firm’s approved loans in a phase is above the maximum amount
due, 0 otherwise. The second is the variable overpayment rate, equal to the amount of overpayment
expressed as a fraction of the maximum payment that was due to borrowers. If a firm had multiple
loans with overpayments in a phase, we used the maximum overpayment rate across the set of
loans in any given phase. The third is the variable overpayment amount, equal to the total dollar
amount by which all the loans in any phase exceeds the maximum. Figure 2 plots the distribution
25The DNB dataset was matched to the PPP dataset by DNB: 3.3 million observations were matched in total, out of
which 1.7 million were verified and not imputed.
26While the data we use was updated overall in 2020 and 2021 by DNB, there is heterogeneity in the date that any
particular firm’s record was updated. Therefore, we place less weight on this data in our analysis and caution readers
that some of the discrepancies between jobs reported in PPP applications and those in the DNB data may be due to
other factors besides fraud.
27The $9,166 benefit spending amount was derived from the SBA’s method of calculating maximum loan payment
as presented in their January 2021 report (https://www.sba.gov/sites/default/files/2021-01/SBA%20OIG%20Report-
21-07.pdf).
9
of overpaid amounts conditional on a firm reapplying in Phase 2 and having overpaid loans in
phase 1. The figure shows that the distribution of overpaid amounts in Phase 2 shifted to the left
of those in Phase 1 of the program. Further descriptive statistics for the overpayment indicators
are shown in Table A2. The approved amount per loan was on average $101,589 USD in phase 1.
The amount approved per loan fell in phase 2 to $42,748 USD. Both the number of overpaid loans
as well as the amount overpaid fell in phase 2. The share of loans that had overpayments was 0.01
in phase 1, while it was 0.003 in phase 2. Since the total number of loans was 5,136,454 in phase
1, this suggests that almost 50,000 loans that were approved had payments above the maximum
stipulated by law. This number fell to more than 19,000 overpaid loans in phase 2. Similarly, the
overpaid amount per loan was $725 USD in phase 1, while it was much lower ($91 USD) in phase
2.
Multiple loans to the same borrower. The SBA does not provide a numerical firm identifier in
the version of data for public access. Therefore, to determine whether two or more loans were
disbursed to the same company, we use string matching on business name and address to assign
a unique firm identification number to each group of loans associated with the same business
name and address (the appendix for online publication has more details on our string matching
algorithm). A firm could have received multiple loans in violation of the rules in any of the
following ways:
• if a firm that participated solely in Phase 1 of the program received more than one loan in
phase 1 (5,290 firms).
• if a firm that participated in both Phase 1 and Phase 2 of the program received more than one
loan in any given phase (3,320 firms).
• if a firm borrowed for the first time in Phase 2 and received more than two loans (7,099 firms).
The average number of loans issued to a firm with multiple loans is 2 in Phase 1 and 3 in Phase 2.
Nevertheless, the incidence of duplicate loans is low when compared to overpayment. Borrowers
with duplicate loans make up only 0.15 percent and 0.18 percent of all participating firms in Phase
1 and Phase 2, respectively (Table A2). Since we were interested in looking at changes in firm
behavior across phases in response to the screening requirement, we define multiple loans as a
dummy variable equal to 1 when a firm that participated in both Phase 1 and Phase 2 of the
program received more than one loan in any given phase.
Gap in jobs reported. Firms report their number of employees on the PPP applications and
these figures are used to compute the loan amount to any firm. We combine this data with DNB
data on employee figures and define the gap in jobs reported in two ways. First, we employ the
simple difference in employee numbers across the two data sources (job gap). Second, we utilize
the percentage by which PPP employee figures (positively) deviate from the DNB figures, and on
10
this basis create a series of dummy variables that capture instances of extreme deviations of PPP
figures from those contained in the DNB database (>300%, 400% or 500%). While evidence from
both outcomes can be interpreted as suggestive of fraud, in our interpretation of the results we
place more weight on the latter measure than the former one.28
Note that the above measures by no means represent a comprehensive account of fraud in the
PPP. Our variables would miss any more sophisticated types of fraud - for example registering
the same business with distinct names - and they do not account for what individuals actually
did with the money - for example buying luxury consumption goods rather than spending it on
employee wages. One should thus consider our measures as capturing some - relatively crude -
aspects of fraud.29
4 Did screening affect fraud in PPP?
4.1 Identification strategy
The advance documentation requirement in phase 2 stipulated that firms with loan applications
greater than $150k must submit documentation showing a reduction in gross receipts of more
than 25% in 2020 relative to 2019. Those with loan requests of less than $150k were not required
to submit such documentation with their loan applications, but they were required to retain said
documents should the SBA later request it.
We use a differences-in-differences approach to estimate the effect of the screening requirement
on fraud in the PPP. We define our treatment group as the set of firms that had applied for a loan
greater than $150k in phase 1. Presuming that the firm fundamentals determining loan need and
eligibility remained constant across the two phases, these are the firms that were affected by the
changes in the documentation requirement. Our control group is the set of firms that had loan
amounts less than $150k in phase 1.
The identification assumption motivating the differences-in-differences estimation strategy is
parallel trends, i.e. firms whose loan amounts were greater than $150k in phase 1 would have
experienced, on average, the same changes in fraudulent behavior across phases as those firms
whose loan amounts were ≤$150k in phase 1, were it not for the fact that the documentation
requirement was imposed on the former (treatment) group. As suggested in the literature, we
assess the evidence in support of this assumption using an event-study plot prior to conducting
the main analysis.
28This is the case since the DNB data on any particular firm is updated with variable frequency and there is little
clarity on whether any update refers to an update of the number of employees that work in a given firm or some other
characteristic of that firm.
29We attempted to incorporate other data to measure other types of fraud, but were not successful in these attempts.
For example, businesses were meant to have been in existence prior to February 15, 2020, and we attempted to find firms
who registered after this date yet received loans. However, the OpenCorporates business registry database we used
does not contain data on a number of large states. In addition, we tried to match loan recipients to firms listed on the
Federal “Do Not Pay” list, a list of firms and individuals previously found to be fraudulent by the federal government,
but could only find about 100 matches.
11
We estimate the relationship between treatment group status and our outcomes in the months
prior to and after the imposition of screening. Specifically, we estimate the following equation,
𝜌 𝑔 𝑇𝑖0 × [1(𝑔 = 𝑚𝑡)] + 𝑒 𝑖𝑚𝑡
Õ
𝑌𝑖𝑚𝑡 = 𝜏𝑖 + 𝛾𝑚𝑡 + (1)
𝑔≠𝐴𝑢 𝑔20
where 𝑌𝑖𝑚𝑡 is one of our three loan irregularity overpayment measures (overpayment dummy, over-
payment rate, and multiple loans dummy), which corresponds to a particular firm (i) that receives
a loan in a given month (m) during a given year (t). 𝑇𝑖0 is an indicator variable equal to 1 for
treatment firms, i.e. those that had loan amounts greater than $150k in phase 1 of the program,
0 otherwise. 𝜏𝑖 is a firm fixed effect and 𝛾𝑚𝑡 is a month-year of approval fixed effect. Standard
errors (𝑒 𝑖𝑚𝑡 ) are clustered at the firm level. The coefficient 𝜌 𝑔 estimates the effect of belonging to
the treatment group for each month-year from April 2020 until May 2021. The reference category
is the last month in phase 1 of the PPP program (August 2020). If 𝜌 is statistically insignificant for
all months of phase 1, then this lends support to the validity of the parallel trends assumption.
Figure 3 through Figure 5 plot 𝜌 and 95% confidence intervals for each month-year of the PPP
program. Figure 3 shows the effect on the overpayment rate, Figure 4 presents the result for the
overpayment dummy, and Figure 5 plots the findings for multiple loans. In all the figures one
can see that before the advanced documentation requirement was introduced in phase 2 (January
2021) for loans >150k, there are no statistically significant differences in outcomes across treatment
(𝑇𝑖0 ) and control groups, lending support to the parallel trends assumption. At the end of this
section we discuss other checks on the validity of our identifying assumptions.
4.2 Estimation
For firms (𝑖) with a PPP loan approved at date (𝑡) we estimate the following equation:
𝑌𝑖𝑡 = 𝛼 𝑖 + 𝜔𝑃 ℎ𝑎𝑠𝑒2𝑡 + 𝛾𝑇𝑖0 × 𝑃 ℎ𝑎𝑠𝑒2𝑡 + 𝑒 𝑖𝑡 (2)
where 𝑌𝑖𝑡 stands for irregularities in PPP loans defined in the ways described earlier. In addition
to our treatment indicator (𝑇𝑖0 ), we include the indicator variable Phase2𝑡 , equal to 1 for loans in
phase 2 of the program, 0 otherwise. The parameter of interest is 𝛾, which captures the impact of
the screening requirement for firms whose previous loan amount indicates that they were subject
to it. We consider specifications of our model with fixed effects (𝛼 𝑖 ), which we specify at the firm
level. Standard errors (𝑒 𝑖𝑡 ) are clustered at the firm level.
4.3 Results
Figure 6 through Figure 8 graphically present the raw mean differences in differences in the data.
There are several takeaways from these figures. First, for the overpayment measures one finds that
both the volume and frequency of overpayment were concentrated among the treatment firms.
This was true across both phases. Second, average outcomes for these measures declined across
12
phases for both treatment and control firms; however, the fall was much greater for the former
than the latter. Finally, the incidence of multiple loans was slightly higher among the control firms
than treatment firms in phase 1. Yet the incidence of multiple loans declined sharply in phase 2
for the treatment firms while it increased for the control firms.
Table 1 moves beyond raw mean differences in the data and presents the results of estimating
equation 2. Columns (1)-(2) present results for the overpayment dummy, Columns (3)-(4) for the
overpayment rate, while Columns (5)-(6) present results for multiple loans as a dummy variable.
Columns (1), (3) and (5) present results without firm fixed effects while Columns (2), (4) and (6)
include them.
Results in Column (2) show that firms that were subject to the upfront documentation require-
ment, i.e. firms that had loans greater than a $150k in phase 1 and reapplied in phase 2, had a
5% reduction in the probability of overpayment relative to firms that were not subjected to the
screening requirement. The results are statistically and economically significant. Given the relative
infrequency of overpayments, this is a large effect, equal to 63% of the value of the control group
mean.
Findings for the rate of overpayment were even more pronounced. As shown in column (4),
the upfront documentation requirement led to a statistically significant reduction in the rate of
overpayment of 1.7% (more than 5 times the control mean) for treated firms in phase 2 of the
program.
The results for multiple loans point in a similar direction. Column (6) shows that the upfront
documentation requirement reduced the probability of receiving multiple loans by 0.2% (an effect
equal to the value of the control mean). This effect is both statistically and economically significant.
Robustness using the “honest approach.” Following recent developments in differences-in-
differences estimation (see Roth et al., 2022 for a review), we use the "honest approach" to parallel
trends suggested by Rambachan and Roth, 2019. We use this to investigate the robustness of our
results to alternative assumptions about different outcome trends for treated (firms that had loan
amounts greater than $150k in phase 1 of the program) and control firms (firms that had loan
amounts ≤$150k in phase 1 of the program).
Rambachan and Roth, 2019 suggest that restrictions on the possible violations of parallel trends
must be specified by the researcher, and the choice should depend on the economic context. In our
case, one might be concerned about violations of parallel trends due to secular trends that evolve
smoothly over time. We, therefore, bound the change in slope of the differential trend between
treated and control firms using the following formula:
Δ𝑆𝐷 := {𝛿 : |(𝛿 𝑡+1 − 𝛿 𝑡 ) − (𝛿 𝑡 − 𝛿 𝑡−1 )| ≤ 𝑀, ∀𝑡}
where 𝛿 𝑡 refers to the difference in trends between the treated and control firms at time t. M is the
maximum possible error of the linear extrapolation of the pre-trend. If M=0 the difference in trends
between treated and control firms would be exactly linear, while M > 0 relaxes the assumption of
13
exact linearity.
The estimates for probability of overpayment and overpayment rates are summarized in Figure 9
and Figure 10, respectively, while those for multiple loans are in Figure 11. The data used for these
plots are at the loan - day level, while the treatment is defined at a firm level. These figures show
that for the overpayment dummy and the overpayment rate the results remain robust for at least an
M=0.0025 and M=0.005, respectively. In the case of multiple loans dummy this value is M=0.0025.
As discussed by Rambachan and Roth, 2019, nothing in the data itself can place an upper bound
on the parameter M. For us to interpret these values of M we use data from phase 1 of the program.
For each of the outcomes, we use only this data to create a linear trend that is extrapolated to phase
2 of the PPP program. We calculate the median of the absolute deviations of the coefficient 𝜌 (see
Equation 1) from this linear trend in phase 1 of the PPP program.30 For the overpayment dummy
this method leads to a value of M=0.0015, while for the overpayment rate and the multiple loans
dummy the values are M=0.0026 and M=0.001, respectively. These are lower than the M at which
our results are robust according to Figure 9 and Figure 10.
Other checks. In this section we present further checks that aim to support a causal interpretation
of our estimates. Appendix Table A3 presents the results defining the treatment firms (𝑇 0 ) as those
with a loan amount between $151K-$200K and the control firms as those with loan amounts
between $100K - $150K in phase 1.31 Firms that are on either side of the $150k threshold in a sample
of those firms whose approved loan amounts were between $100-200k in phase 1 are likely to be
similar to each other in their fundamentals and the baseline probability of fraudulent behaviour.
The results are in the same direction as those in Table 1, albeit of smaller magnitude as the sample
is restricted to firms whose approved loan amounts were between $100-200k in phase 1.
Next we check whether there is attrition of firms in Phase 2 of the program and investigate the
direction of the bias. Since we can observe the universe of firms that received PPP loans in phase
1, i.e. before the announcement of the advanced documentation requirement in January 2021, if
any firm does not reapply in phase 2 of the PPP program due to the advanced documentation
requirement we can observe this in the data. Results show that those with loan amounts greater
than $150k and with multiple or overpaid loans in phase 1 were more likely to reapply in phase 2
of the program (see Table 2a and 2b Column (1) below). This suggests that we have a conservative
estimate of the true effect of screening.
In this context one might also be worried about anticipation effects of an upfront documentation
requirement based on loan size. However, the documentation requirement based on the $150k
threshold was not announced or even discussed by policy makers in phase 1 and phase 1 is when
treatment firms are defined (see timeline in Figure 1). The announcement of the documentation
requirement was on January 6, 2022, right before the start of phase 2 of the program.32
30For a graphical representation see Figure 3 to Figure 5. These figures plot the coefficient 𝜌 from Equation 1 for each
month in phase 1 and 2 of the PPP program. The black line is a linear trend using only phase 1 data.
31If a firm had been approved for multiple loans, 𝑇 0 would take on a value of one if the firm had at least one loan
amount greater than $150K in Phase 1.
32This announcement was via the release of a rule change and advanced documentation was effective for second time
14
5 A Model of Fraud with a Value-Based Documentation Requirement
The findings presented thus far establish that there were benefits from the PPP’s documentation
requirement in terms of fraud reduction. Yet such requirements typically impose administrative
burdens on all program participants in addition to dissuading fraud. Thus, it is important to verify
that the documentation requirement adopted in the PPP was effective in reducing losses due to
the participation of fraudulent firms without imposing undue burdens on firms with a legitimate
right to participate in the program. We develop here a theoretical framework that provides precise
empirical implications about when this will be the case, which we subsequently evaluate using
our loan data.
Consider a relief program that distributes a highly valued good (e.g., money) among firms in
the economy. A given firm 𝑖 participates in the program by submitting an application for the good
in the amount 𝑔𝑖 ∈ [0, 𝑔], where 𝑔 is the maximum level of support for any firm permitted by the
program. The economy contains two types of firms, legitimate firms and fraudulent firms, with the
former equal to a proportion 𝜁 ∈ (0, 1) of all firms. One can conceptualize fraudulent firms as shell
corporations that engage in no legitimate economic activity, or, alternatively, as firms that engage in
economic activity but are ineligible to participate in the program based on the criteria stipulated by
the program. Crucially, due to the need to rapidly provide relief in order to mitigate the emergency
which gave rise to the program, firm type cannot be discerned by program administrators prior
to allocating the good. Yet the status of firms may become apparent during an ex-post review
of the program after the emergency has abated, with penalties potentially assigned to fraudulent
firms that received the good. At the time of application, each firm is privy to its eligibility status
𝑠 𝑖 ∈ {0, 1}, where 𝑠 𝑖 = 1 indicates that firm 𝑖 is legally eligible to participate (legitimate firm) and
𝑠 𝑖 = 0 indicates that it is not eligible (fraudulent firm).
Utility from participation in the program varies by firm type. For legitimate firms, participation
in the program entails no risk of punishment, so demand for the good is mediated only by
idiosyncratic tastes for asking the government for support and the fixed cost of submitting an
application. For fraudulent firms, who are officially barred from the program, the prospect of
punishment at some point after the program has concluded is a distinct possibility, so this fact will
shape demand for the good.
In line with the institutional structure of the PPP, we consider a documentation requirement that
is based on the amount of the good being solicited from the program. Specifically, we consider
a documentation requirement imposed on all loan applications greater than the amount 𝑔
e. This
requirement has two consequences. First, all firms seeking values of the good above this amount
will incur a cost 𝜙 > 0, which represents the administrative burden associated with satisfying the
documentation requirement. Second, the likelihood that a fraudulent firm will be identified as such
upon post-program review increases discretely with the documentation requirement, implying
that for fraudulent firms the cost of fraud shifts upward at this point. This discrete increase in the
PPP loans made after January 14, 2021. https://www.sba.gov/sites/default/files/2021-01
15
cost of fraud at 𝑔
e can be conceptualized as reflecting firms’ beliefs about the likelihood that fraud
will be discernable based on the information contained in the document itself, or, alternatively,
firms’ beliefs about the resolve of program administrators to more stringently audit firms that
receive amounts of the good in excess of 𝑔
e.
These considerations lend themselves to the following characterization of firm utility:
(
𝑣(𝑔𝑖 ) − 𝜂 𝑖 𝑔𝑖 − 𝜙 I(𝑔𝑖 > 𝑔
e) if 𝑠 𝑖 = 1
𝑢𝑖 = (3)
𝑣(𝑔𝑖 ) − 𝑐(𝑔𝑖 ) − 𝜂 𝑖 𝑔𝑖 − 𝜙 I(𝑔𝑖 > 𝑔
e) if 𝑠 𝑖 = 0
where 𝑣(0) = 0, 𝑣 ′ > 0, 𝑣 ′′ < 0, 𝑣 ′(0) = +∞, and 𝜂 𝑖 captures idiosyncratic tastes for procuring
government support. We assume 𝜂 𝑖 is distributed according to a continuous density 𝐹1 with
support [𝜂, 𝜂] for legitimate firms, and according to a continuous density 𝐹0 with identical support
for fraudulent firms. I(𝑥) is an indicator function equal to 1 if the expression 𝑥 is true (0 otherwise).
For a fraudulent firm, 𝑐(𝑔𝑖 ) represents the cost of soliciting the good in amount 𝑔𝑖 given that the
firm is not entitled to participate in the program. We characterize the cost function as follows:
𝑐(𝑔𝑖 ) = 𝜋(𝑔𝑖 )[1 + 𝜏I(𝑔𝑖 > e
𝑔 )], (4)
where 𝜋 satisfies 𝜋(0) = 0, 𝜋′ > 0, 𝜋′′ > 0. The parameter 𝜏 > 0 captures the discrete jump in the
cost of fraud at 𝑔
e. The cost function reflects a setting in which in which the punishment for fraud
is a smooth and convex increasing function of the level fraud but the likelihood of detection jumps
discretely upwards for all levels of fraud greater than the amount 𝑔
e.
To fix ideas, we start with the scenario in which there is no discrete change in documentation
requirements at 𝑔
e, i.e. 𝜏 = 𝜙 = 0. The optimally selected value of the good for a legitimate firm is
equal to: (
b𝑖 ≡ 𝑔𝑖 |𝑣 ′(𝑔𝑖 ) = 𝜂 𝑖
𝑔 b𝑖 ≤ 𝑔
if 𝑔
𝑔𝑖1∗ (𝜂 𝑖 ) = , (5)
𝑔 if 𝑔
b𝑖 > 𝑔
which implies that 𝑔𝑖1∗ is (weakly) decreasing in 𝜂 𝑖 .
In contrast, the optimal request of a fraudulent firm is equal to:
(
𝑔𝑖† ≡ 𝑔𝑖 |𝑣 ′(𝑔𝑖 ) − 𝜋′(𝑔𝑖 ) = 𝜂 𝑖 if 𝑔𝑖† ≤ 𝑔
𝑔𝑖0∗ (𝜂 𝑖 , 𝜏 = 0) = , (6)
𝑔 if 𝑔𝑖† > 𝑔
where again 𝑔𝑖0∗ is decreasing in 𝜂 𝑖 .
Now define the threshold points 𝜂𝑈
1
, 𝜂𝑈
0
as follows:
𝜂𝑈
1
≡ 𝜂 𝑖 | 𝑔𝑖1∗ = 𝑔
e (7)
𝜂𝑈
0
≡ 𝜂 𝑖 | 𝑔𝑖0∗ = 𝑔
e.
All legitimate firms with a taste parameter above 𝜂𝑈
1
optimally request amounts of the good below
16
𝑔
e, so they are not affected by the documentation requirement. Similarly, all fraudulent firms with
a taste parameter above 𝜂𝑈
0
optimally request amounts of the good below 𝑔
e.
For all firms for which the taste parameter lies below the relevant threshold, on the other hand,
the documentation requirement may affect their requests. Any such firm can set 𝑔𝑖 = 𝑔
e, thereby
ensuring a utility disbursement equal to:
(
e) − 𝜂 𝑖 𝑔
𝑣( 𝑔 if 𝑠 𝑖 = 1
𝑢𝑖 ( 𝑔
e
e) = (8)
𝑣( 𝑔 e) − 𝜂 𝑖 𝑔
e) − 𝜋( 𝑔 e if 𝑠 𝑖 = 0
For a legitimate firm with 𝜂 𝑖 ≤ 𝜂𝑈
1
that sets 𝑔𝑖 ≠ 𝑔
e, the best such request will be 𝑔𝑖1∗ , defined by
equation (5). Similarly, for a fraudulent firm with 𝜂 𝑖 ≤ 𝜂𝑈
0
that sets 𝑔𝑖 ≠ 𝑔
e, the best such request
will be: ( ‡ ‡
𝑔𝑖 ≡ 𝑔𝑖 |𝑣 ′(𝑔𝑖 ) − (1 + 𝜏)𝜋′(𝑔𝑖 ) = 𝜂 𝑖 if 𝑔𝑖 ≤ 𝑔
𝑔𝑖0∗ (𝜂 𝑖 , 𝜏) = ‡ . (9)
𝑔 if 𝑔𝑖 > 𝑔
Naturally, the question arises as to whether it would be optimal for firms to avoid documentation
by requesting 𝑔
e or to solicit more from the program in spite of the added administrative burden
and/or greater risk of sanction. Define the threshold points 𝜂1𝐿 , 𝜂0𝐿 as follows:
𝜂1𝐿 ≡ 𝜂 𝑖 |𝑢𝑖 (𝑔𝑖1∗ (𝜂 𝑖 )) = 𝑢𝑖 ( 𝑔
e) (10)
𝜂0𝐿 ≡ 𝜂 𝑖 |𝑢𝑖 (𝑔𝑖0∗ (𝜂 𝑖 , 𝜏)) = 𝑢𝑖 ( 𝑔
e)
where the utilities on the LHS of the equalities above incorporate the fixed cost of the documen-
tation requirement (𝜙) and, for fraudulent firms only, also incorporate the parameterized jump in
the cost function (𝜏). Firms with a taste parameter equal to the relevant threshold above will be
indifferent between setting 𝑔𝑖 = 𝑔
e and optimally choosing a higher value of 𝑔𝑖 that is subject to
the documentation requirement. Firms with a taste parameter below the relevant threshold will
choose a value of 𝑔𝑖 above 𝑔
e. Specifically, all legitimate firms with a taste parameter equal to or
above 𝜂1𝐿 but equal to or below 𝜂𝑈
1
, i.e. all 𝑖 such that 𝑠 𝑖 = 1 and 𝜂 𝑖 ∈ [𝜂1𝐿 , 𝜂𝑈
1
], will set 𝑔𝑖 = 𝑔
e.
Moreover, all fraudulent firms with a taste parameter equal to or above 𝜂0𝐿 but equal to or below
𝜂𝑈
0
, i.e. all 𝑖 such that 𝑠 𝑖 = 0 and 𝜂 𝑖 ∈ [𝜂0𝐿 , 𝜂𝑈
0
], will also set 𝑔𝑖 = 𝑔
e. Thus, the existence of the
documentation requirement creates a spike in the mass of requests at 𝑔𝑖 = 𝑔
e equal to:
Δ = 𝜁[𝐹1 (𝜂𝑈
1
) − 𝐹1 (𝜂1𝐿 )] + (1 − 𝜁)[𝐹0 (𝜂𝑈
0
) − 𝐹0 (𝜂0𝐿 )], (11)
resulting in a concomitant reduction in the mass of requests above 𝑔
e. Following the nomenclature
employed by the public finance and labor literatures, we refer to firms that contribute to this spike
as ‘bunchers’ and Δ as the bunching mass.33
33There is a large public economics (Saez, 2010; Chetty et al., 2011; Kleven and Waseem, 2013) and labor (Burtless
and Hausman, 1978; Aaron et al., 1981; Chetty et al., 2011) literature that uses discrete changes in the level and slope
of choice sets as a way to estimate elasticity of behavioral responses like income, wealth and labor supply. See Kleven,
2016 for an excellent survey of this work.
17
Figure 12 depicts the impact of the documentation requirement on the density of all requests
for the good from the program.34 The initial density of requests–without the documentation
requirement–is shown by the solid line. Note that the density is smooth throughout its range.
The density of requests with documentation required for all requests above 𝑔
e is shown by the
dashed line. This density is characterized by a large upward spike at 𝑔
e, reflecting the choice of a
subset of legitimate firms to avoid the administrative costs of the documentation requirement and
the choice of a subset of fraudulent firms to simultaneously evade the risk of detection created by
documentation requirement and to avoid its administrative costs. Relative to the no documentation
density, the density of requests with the documentation requirement has a large excess mass at 𝑔
e
and too little mass for an interval to the right of this point. Otherwise, at points sufficiently far to
the left and right of 𝑔
e, the two densities overlap perfectly.
It is instructive to consider how the parameters representing the administrative burden of the
documentation requirement (𝜙) and its value as a tool for sanctioning fraud (𝜏) affect the size and
composition of the bunching mass. The proposition below summarizes these relationships.
Proposition 1. a) An increase in the administrative burden of the documentation requirement (𝜙) results
in a larger bunching mass due to increased strategic avoidance of the requirement by both legitimate and
fraudulent firms; b) An increase in the fraud sanctioning value of the documentation requirement (𝜏)
results in a larger bunching mass due strictly to an increase in strategic avoidance of the documentation
requirement by fraudulent firms seeking to reduce their risk of sanction. Thus, the bunching mass will be
disproportionately populated by illegitimate firms when 𝜏 is high and 𝜙 is low.
Proof. Note that Δ is decreasing in 𝜂1𝐿 and 𝜂0𝐿 , each of which may depend on 𝜏 and 𝜙. Thus, changes
in parameters that lower both of these points or that lower one but leave the other unaffected will
unambiguously increase Δ. Using the equalities in (10), implicit differentiation and application
of the envelope theorem reveals that an increase in 𝜙 lowers both 𝜂1𝐿 and 𝜂0𝐿 . This implies that
greater 𝜙 leads to a larger bunching mass, and that it does so because greater numbers of both
legitimate and illegitimate firms engage in strategic avoidance. Repeating the same procedure for
𝜏 reveals that 𝜂0𝐿 is decreasing in 𝜏 but 𝜂1𝐿 is unaffected by changes in this parameter. This implies
that greater 𝜏 leads to a larger bunching mass, but that it does so solely due to the fact that greater
numbers of illegitimate firms engage in strategic avoidance. □
The proposition has immediate implications for program design. An effective documentation
requirement is one for which compliance costs are low, but which significantly raises the real or
perceived risk of engaging in fraud. If a documentation requirement is effective in this sense,
then the composition of the bunchers will be quite different from that of the overall population of
firms: bunchers will consist of a disproportionate number of fraudulent firms. If a documentation
requirement is ineffective, on the other hand, then the composition of the bunchers will mirror that
34This graph was created by parameterizing the model as follows: 𝑣(𝑥) = ln(𝑥), 𝜋(𝑥) = 𝑥 2 /2, 𝜌 = 1/2, 𝜏 = 1,
𝜙 = 1/1000, and 𝐹1 , 𝐹0 are truncated normal densities with mean 5 and support [1,9].
18
of the overall population. For this reason, empirical analyses which can detect bunching and elu-
cidate the composition of bunchers may be highly revealing about the efficacy of a documentation
requirement. We investigate these in more detail in the next section.
6 Mechanisms: Extensive or intensive margin effects?
In this section, we examine mechanisms that might explain why the advanced documentation
requirement led to a systematic fall in overpayment and multiple loans, particularly on loans
greater than $150k in phase 2. The introduction of screening may have affected firm behavior along
the extensive margin, by which we refer to the decision of firms to exit the program. Specifically,
those firms whose phase 1 loan indicated that they may have participated in fraud could have
reacted to the prospect of screening in phase 2 by abandoning the program entirely. Alternatively,
as elucidated by our formal model, the introduction of screening may have affected firm behavior
along the intensive margin, by which we refer to the decision of firms to manipulate the details of
their loan applications so as to avoid crossing the $150k threshold. Firms with past indicators of
fraud may have continued using the PPP but simply asked for less out of the program in order to
avoid the documentation requirement. To the extent that fraudulent firms reduced their requests
in this manner at a significantly greater rate than legitimate firms, this constitutes evidence that
the fraud reducing effect of the documentation requirement dominated the administrative burden
it may have generated. Below we provide evidence on both types of mechanisms. We show that
the behavioral reaction of firms to the documentation requirement was overwhelmingly located
on the intensive margin, and that the nature of the behavioral response indicates that the fraud
reduction properties of the requirement dominated its compliance costs.
6.1 Estimation
For this section we use the cross-section of firms (𝑖) in phase 2 of the PPP program and estimate
the following equation:
𝑌𝑖 = 𝜅 + 𝜃𝐹𝑖0 + 𝜇𝑇𝑖0 + 𝜋𝑇𝑖0 × 𝐹𝑖0 + 𝜀𝑖 (12)
where 𝑌𝑖 includes the following outcomes: an indicator variable equal to 1 if firm 𝑖 exited in phase
2 of the PPP program (i.e., did not reapply for PPP loans); an indicator variable equal to 1 if the
firm applied to a loan in phase 2 that is greater than $150k; the dollar loan amount itself; the
difference in the number of jobs reported from phase 1; and the difference in the loan amount per
job. While the first outcome describes changes in firm behavior along the extensive margin, the last
four outcomes present evidence on the intensive margin. To assess whether screening resulted in
behavioral changes for firms that may have participated in fraud, we include the indicator variable
𝐹 0 . The variable 𝐹 0 is equal to 1 for firms that were paid above the maximum permissible amount
under the PPP rules or firms that received multiple loans in phase 1. We interact 𝐹 0 with our
19
treatment (𝑇𝑖0 ) in order to specifically assess how potentially fraudulent firms that were most likely
to be affected by the documentation requirement responded to it. Standard errors (𝜀𝑖 ) are clustered
at the firm level.
6.2 Results
Table 2a and Table 2b present the findings. Table 2a shows the results classifying suspect firms
as those that were paid a PPP loan amount greater than the permissible amount in phase 1 of the
program. Table 2b presents the results classifying such firms as those that received multiple PPP
loans in phase 1 of the program.
Screening does not appear to have prompted bad actors to leave the program (Column (1)). To
the contrary, firms that were subjected to upfront documentation and that had previously obtained
loans in excess of the maximum or received multiple loans were more likely than other firms to
obtain loans in phase 2 of the PPP program (9.6 and 5.9 percentage points of the control mean,
respectively).
However, screening did appear to have important implications for behavior along the intensive
margin, i.e. for the content of the loans firms received in phase 2. There are two key takeaways.
First, these results show that, unlike for bad actors, there was continuity in loan requests for those
firms that did not have any irregularities in phase 1. Firms with loan amounts greater than $150k
in phase 1 but no overpayments or multiple loans, continued to ask for loans greater than $150k
(approximately 90%). The loan amounts they received in phase 2 were also larger in phase 2 than
control firms (Columns (2) - (3)). Second, Column (2) in both Panels (a) and (b), show that firms
that were subjected to screening and that had previously obtained loans in excess of the maximum
or received multiple loans were 11% and 13.7% less likely to receive a loan amount greater than
$150k, respectively. The effects are similar for loan amounts as an outcome (Column (3)).
Table 2a Columns (4) and (5) show that the suspect firms did not request smaller loan amounts
because they reported that they had fewer employees to support. In fact, they reported significantly
greater growth in the number of employees relative to phase 1 than did other firms (Column (4).
What distinguished these firms is that they reduced the loan amounts they requested per employee
much more than did other firms. Table 2a Column (5) shows that the differences in this respect are
substantial. Relative to firms that received a loan amount in excess of $150K in phase 1 but were not
overpaid, overpaid firms in the same category received approximately $18k less per employee in
phase 2 than they did in phase 1. Table 2b Columns (4) and (5) present a similar trend. Thus, given
the specter of screening for high loan requests, firms with past loan irregularities were inclined to
cut their requests down to levels carrying less risk.
6.3 Strategic Evasion of Screening
The fact that the upfront documentation requirement in phase 2 was made conditional on the
value of the loan request introduces the possibility of strategic evasion of screening. Borrowers
20
who sought to maximally exploit the program but wanted to avoid providing evidence of a fall in
their gross receipts could do so by following a simple strategy: Set phase 2 loan requests just below
the $150k cut-off. As in our formal model, we refer to borrowers who pursue such as strategy as
‘bunchers’. If strategic evasion was truly systemic, then the data should reveal the existence of a
large mass of such bunchers. Moreover, if the fraud reduction properties of screening dominated its
administrative burden, then the firms belonging to any such bunching mass should be significantly
more likely to have indicators of past wrongdoing than non-bunching firms.
Due to the fact the we observe loan allocations across the two phases of the program, our data
are exceptionally well suited for detecting the existence of bunchers. Since in phase 1 there was no
documentation requirement activated for loans greater than $150K, the distribution of loan values
from that phase serves as a credible counterfactual distribution, i.e. a distribution capturing what
the loan values in phase 2 would have been had screening not been implemented.35 If bunchers
engaging in strategic evasion do exist, then we should observe specific patterns in the distributions
of loan values in the two phases. First, in phase 2 one should observed a sharp upward spike in
the density of loans immediately below the $150K cutoff, and concomitant reduction in the density
above the cutoff. Second, in phase 1 we should not observe any large spikes in the density around
the $150K cutoff, as the screening requirement was not operative at this time.
Figure 13 presents the density of loan amounts for phase 1 and phase 2 loans. Figure 14 presents
the same information, but breaks down the phase 2 loans into first and second time borrowers.36
The figures provide evidence of systemic strategic evasion. In both figures, one observes that
relative to the counterfactual (phase 1) distribution, there is a marked excess mass just to the left
of the $150K cutoff and a missing mass of borrowers above the cutoff. The spike immediately to
the left of the $150K cutoff is exceptionally stark and dwarfs the magnitude of other bumps in the
distribution attributable to reference effects.37
We complement our visual inspection of loan amounts with a formal test of the continuity of
the distribution of loan amounts around the $150k cut-off in the two phases of the PPP program.
Figure 15 presents the density as well as the p-value from a McCrary density test (McCrary, 2008).
The McCrary density test tests the null hypothesis of the continuity of the density of loans against
the alternative of a jump in the density function at the $150k cut-off.38 There are two key takeaways.
First, we fail to reject the null of a continuous loan distribution in phase 1 around the $150K cut-off
(p-value=0.266). This lends support to our use of the phase 1 loan distribution as a counterfactual
for the behavior of the firms in phase 2. Second, we can reject the null of a continuous loan
35A common challenge in identifying bunchers in the public economics and labor literature is the absence of a
counterfactual distribution reflecting aggregate behavior in the absence of the reform (Blomquist et al., 2021; Jakobsen
et al., 2020; Londoño-Vélez and Avila-Mahecha, 2020)
36Naturally, incentives for firms that were applying for the first time in phase 2 were similar to those for repeat
borrowers. Any new program participant who wished to garner maximal benefit from the PPP but avoid upfront
documentation of a fall in gross receipts could do so by locating just below the $150K cutoff.
37By reference effects, we refer to the concentration of loans in amounts that are easy to remember, typically numbers
that are factors of ten or five.
38The McCrary density test (McCrary, 2008) is the standard test of discontinuities in the conditional density of the
forcing variable in regression discontinuity designs (Imbens and Lemieux, 2008).
21
distribution in phase 2 around the $150K cut-off (p-value=0). This finding is consistent with firms
changing their behavior following the introduction of the advanced documentation requirement
and ‘bunching’ below the cut-off of $150K.
Identifying “bunchers”. While a visual inspection of Figure 13 and Figure 14 can be utilized
to identify the interval of loan values whose frequency was inflated by bunching due to strategic
evasion, we formally test and locate the bunching interval using a Kolmogorov–Smirnov (KS) test.
The KS-test tests the equality of the loan distributions in phase 1 and phase 2 loans. Figure 16
plots p-values from this test on the y-axis with loan amounts on the x-axis. The figure suggests
that the two distributions across phase 1 and phase 2 for second time borrowers are statistically
significantly different in the $136-150K range. We therefore, define ‘bunchers’ as those firms that
had PPP loans of greater than $150K in phase 1, but then chose to get PPP loans between $136-150K
in phase 2.
The behavior of bunchers before-and-after screening. Having established the existence of sys-
temic strategic evasion and having identified the set of firms that appear to have practiced it, two
questions arise. First, do the bunching firms differ systematically from non-bunchers in terms of
their behavior in the first phase of the program? Specifically, are they more likely to have received
irregular loans or to have reported employment figures that appear to be inflated (in order to
qualify for greater loans) in Phase 1 of the program? If so, then this would be consistent with
the notion that bunching firms were disproportionately composed of bad actors, some of whom
may have engaged in strategic evasion to avoid detection for both their past (phase 1) and present
(phase 2) misdeeds. A finding of a disproportionate level of such fraudulent firms among the
bunchers would indicate that the documentation requirement worked well, in the sense that it dis-
suaded participation in the PPP by bad actors to a greater degree than it dissuaded participation
by legitimate firms. A second question that arises is to what extent, if any, did screening lead to
a fall in irregularities among bunching firms? Findings linking screening to a reduction in fraud
among bunchers might indicate that the introduction of screening may have induced extra caution
on the part of bad actors, even though they had set their loan requests at lower level to avoid the
documentation requirement.
To address these questions, we use our data at the firm (𝑖) and phase (𝑡) level and estimate the
following:
𝑌𝑖𝑡 = 𝜋 + 𝜃𝐵𝑢𝑛𝑐 ℎ𝑒𝑟𝑠 𝑖 + 𝜏𝑃 ℎ𝑎𝑠𝑒2𝑡 + 𝛽𝐵𝑢𝑛𝑐 ℎ𝑒𝑟𝑠 𝑖 × 𝑃 ℎ𝑎𝑠𝑒2𝑡 + 𝑣 𝑖𝑡 (13)
where 𝑌𝑖𝑡 includes the following outcomes: overpayment dummy; overpayment rate; the difference
in the number of jobs reported on the PPP loan application as compared to the employees data
compiled by DNB (job gap); a set of indicator variables denoting whether the percent gap in jobs
reported in the PPP loan application as compared to the DNB employees data was greater than
300, 400 or 500%, respectively; and multiple loans. 𝐵𝑢𝑛𝑐 ℎ𝑒𝑟𝑠 𝑖 is an indicator variable equal to 1 for
22
firms that had PPP loans of greater than $150K in phase 1, but who then chose to get PPP loans
between $136-150k in phase 2, and equal to 0 for non-bunchers. Non-bunchers are any firms that
fell outside the bunching interval of $136-150K in phase 2 of the program, irrespective of their loan
amounts in phase 1. Standard errors (𝑣 𝑖𝑡 ) are clustered at the firm level.
The coefficient 𝜃 captures the outcomes for bunching firms in phase 1. If bunching firms were
disproportionately comprised of bad actors, we should expect 𝜃 > 0 for the measures of fraud
in our study. The coefficient 𝛽 allows us to quantify the extent to which the new documentation
requirement was correlated with a reduction in such outcomes among bunching firms in phase 2.
Table 3 presents the results. For nearly all outcomes, we find that bunchers were more likely
than non-bunchers to have been involved with loan irregularities in phase 1. Table 3 Column (1)
and (2) show that relative to non-bunchers, they were significantly more likely to have received an
overpayment on their loans (1.7% or nearly twice the average for the control group) and the rate of
overpayment on their loans was higher (1.6% or 2.7 times the control group’s average). Columns
(3) and (4)-(6) show that the gap between the number of jobs bunchers reported in their PPP
applications and that recorded by DNB was higher (column (3)), and the likelihood that said gap
was extreme–exceeding a percentile difference of 300%, 400%, or 500%–was also higher (columns
(4)-(6)). In all these cases, the differences between bunchers and non-bunchers were statistically
significant by any conventional standard. The one outcome for which there was no statistically
significant difference between bunchers and non-bunchers was the incidence of multiple loans
(although the estimated impact is non-trivial in substantive terms). This one partial exception
notwithstanding, it appears clear that bunchers were disproportionately comprised of firms with
a highly checkered history in the program.
The table also shows that the screening requirement was associated with a significant fall in
irregular activities by bunchers in phase 2. With the introduction of screening in phase 2, relative
to phase 1, bunchers became significantly less likely to receive loans with an overpayment (1.3%
or 1.3 times the control group mean), the rate of overpayment on their loans declined (1.5% or
2.5 times the control group mean), the gap between the number of jobs they reported in their
applications and the DNB figures declined (1.9 jobs or half the control group mean), and the
likelihood of extreme gaps between these sources declined as well (approximately 1% fall or a
fall of 10-23% of the control group mean in the category of percent job gap of >300%, >400% and
>500%, respectively). The probability of multiple loans by bunchers also fell by 0.2% (100 pp of
the control group mean) in phase 2 relative to phase 1.
These results suggest that the advance documentation requirement was able to screen out the
bad actors from loan brackets greater than >$150k and in the process significantly reduced the
irregularities committed by suspect firms. The effects are economically large and statistically
significant. In fact, the declines of bunchers were significantly larger in magnitude than the
declines experienced by non-bunchers. Thus, even though bunchers had taken pains to avoid
screening by setting their loan requests to values below the $150K cutoff, it appears the existence
of the documentation requirement seems to have instilled greater caution, thereby reducing the
23
incidence of fraud.
Conclusion
The findings of our paper establish the utility of screening as a means of reducing fraud in large-
scale public programs. Examining changes over time in the behavior of borrowers in the Paycheck
Protection Program (PPP), we find that screening was effective in reducing irregularities in loan
disbursements that are indicative of fraud. However, due to the peculiarities of the implementation
of screening in the PPP, it was possible for borrowers to strategically evade the documentation
requirements it imposed by setting their loan requests below a threshold value of $150K. A non-
negligible mass of borrowers did precisely that. Even so, the advent of screening nonetheless
marked a reduction in the aggregate level of fraud, including among those borrowers who had
engaged in strategic evasion. In our judgement, this speaks to the merit of screening in programs
like the PPP that have extensive scope and tight timelines.
The PPP is far from unique in this regard. There are varied circumstances when it is simply
impossible to have a robust oversight apparatus up-and-running prior to engaging with the po-
tential beneficiaries from a public program. Emergency relief programs by their very nature face
this challenge, since the welfare improvements they convey are highly contingent upon the time-
liness with which funds are allocated to recipients. And yet such programs cannot possibly be
effective unless they are able to channel resources to legitimate program participants and prevent
debilitating levels of fraud. In precisely these types of contexts, this paper shows that a little bit of
screening can go a long way.
A key trade-off for policymakers is the timeliness of support versus the extent to which said
support is used for its intended purpose. Screening requirements that can be satisfied by program
participants promptly and at low cost, but which are also effective in discriminating between
actors who should or should not have access to the program hit the veritable sweet spot. Actors
who are legitimate program participants should be able to put together the relevant documentation
quickly—meaning that funds can go out the door quickly to those who need them. Conversely, fake
documentation should be difficult to create and/or easy to detect, thereby deterring criminally-
minded actors from attempting to defraud the program.
Of course, it is important to recognize that there is no free lunch in implementing screening. The
introduction of documentation requirements and other verification procedures can easily dissuade
legitimate beneficiaries of a public program from taking the steps necessary to receive the benefits
intended for them by policymakers. The returns to fraud reduction offered by screening should
always be contemplated with this point in mind. Yet our paper demonstrates that there are
contexts in which the trade-off between fraud reduction and uptake by legitimate beneficiaries
can be empirically assessed, thereby informing future changes in program design. In particular,
the approach developed in this paper should be applicable to any program in which screening
requirements are a function of the value of the good solicited by program participants. In these
24
settings, the identification of a bunching mass and an analysis of the characteristics of bunchers can
provide useful information about the fraud-inhibiting effect of screening relative to its compliance
costs.
25
References
Aaron, S. F. H. J., H. J. Aaron, J. A. Pechman, et al. (1981). How taxes affect economic behavior.
Number 14. Washington, DC: Brookings Institution.
Alatas, V., R. Purnamasari, M. Wai-Poi, A. Banerjee, B. A. Olken, and R. Hanna (2016). Self-
targeting: Evidence from a field experiment in indonesia. Journal of Political Economy 124(2),
371–427.
Autor, D., D. Cho, L. D. Crane, M. Goldar, B. Lutz, J. Montes, W. B. Peterman, D. Ratner, D. Villar,
and A. Yildirmaz (2020). An evaluation of the paycheck protection program using administrative
payroll microdata. Technical report, MIT.
Avis, E., C. Ferraz, and F. Finan (2018). Do government audits reduce corruption? estimating the
impacts of exposing corrupt politicians. Journal of Political Economy 126(5), 1912–1964.
Balyuk, T., N. Prabhala, and M. Puri (2021). Small bank financing and funding hesitancy in a crisis:
Evidence from the paycheck protection program. In Conference “How is the Covid-19 experience
changing finance, pp. 7–8.
Banerjee, A., S. Mullainathan, and R. Hanna (2012). Corruption. Technical report, National Bureau
of economic research.
Banerjee, A. V. (1997). A theory of misgovernance. The Quarterly journal of economics 112(4),
1289–1332.
Bartik, A. W., Z. B. 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. Technical report,
National Bureau of Economic Research.
Becker, D., D. Kessler, and M. McClellan (2005). Detecting medicare abuse. Journal of Health
Economics, 189–210.
Becker, G. S. and G. J. Stigler (1974). Law enforcement, malfeasance, and compensation of enforcers.
The Journal of Legal Studies 3(1), 1–18.
Beggs, W. and T. Harvison (2021). Fraud and abuse in the paycheck protection program? evidence
from investment advisory firms. Technical report, University of San Diego.
Behrer, A. P., E. L. Glaeser, G. A. Ponzetto, and A. Shleifer (2021). Securing property rights. Journal
of Political Economy 129(4), 1157–1192.
Besley, T. and S. Coate (1992). Workfare versus welfare: Incentive arguments for work requirements
in poverty-alleviation programs. The American Economic Review 82(1), 249–261.
26
Besley, T. and S. Coate (2003). Elected versus appointed regulators: Theory and evidence. Journal
of the European Economic Association 1(5), 1176–1206.
Blomquist, S., W. K. Newey, A. Kumar, and C.-Y. Liang (2021). On bunching and identification of
the taxable income elasticity. Journal of Political Economy 129(8), 2320–2343.
Bobonis, G., L. R. C. Fuertes, and R. Schwabe (2016). Monitoring corruptible politicians. American
Economic Review 106(8), 2371–2405.
Burgess, R., M. Hansen, B. A. Olken, P. Potapov, and S. Sieber (2012). The political economy of
deforestation in the tropics. The Quarterly journal of economics 127(4), 1707–1754.
Burtless, G. and J. A. Hausman (1978). The effect of taxation on labor supply: Evaluating the gary
negative income tax experiment. Journal of political Economy 86(6), 1103–1130.
Chetty, R., J. N. Friedman, N. Hendren, M. Stepner, et al. (2020). The economic impacts of covid-19:
Evidence from a new public database built using private sector data. Technical report, National
Bureau of Economic Research.
Chetty, R., J. N. Friedman, T. Olsen, and L. Pistaferri (2011). Adjustment costs, firm responses, and
micro vs. macro labor supply elasticities: Evidence from danish tax records. The quarterly journal
of economics 126(2), 749–804.
Coase, R. H. (1960). The problem of social cost. the Journal of Law and Economics 3, 1–44.
Coronese, M., F. Lamperti, K. Keller, F. Chiaromonte, and A. Roventini (2019). Evidence for sharp
increase in the economic damages of extreme natural disasters. Proceedings of the National Academy
of Sciences 116(43), 21450–21455.
Cowley, S. (2020). Despite billions in fees, banks predict meager profits on p.p.p. loans. New York
Times (Oct.2), B–1.
Currie, J. (2006). The take-up of social benefits, pp. 80–148. Russell Sage Foundation.
De Chiara, A. and L. Livio (2017). The threat of corruption and the optimal supervisory task.
Journal of Economic Behavior & Organization 133, 172–186.
Dell, M., B. Jones, and B. Olken (2012). Temperature shocks and economic growth: Evidence from
the last half century. American Economic Journal: Macroeconomics 4(3), 66–95.
Deshpande, M. and Y. Li (2019). Who is screened out? application costs and the targeting of
disability programs. American Economic Journal: Economic Policy 11(4), 213–48.
Di Tella, R. and E. Schargrodsky (2003). The role of wages and auditing during a crackdown on
corruption in the city of buenos aires. Journal of Law and Economics 46(1), 269–292.
27
Duflo, E., M. Greenstone, R. Pande, and N. Ryan (2013). Truth-telling by third-party auditors
and the response of polluting firms: Experimental evidence from india. The Quarterly Journal of
Economics 128(4), 1499–1545.
Duflo, E., M. Greenstone, R. Pande, and N. Ryan (2018). The value of regulatory discretion:
Estimates from environmental inspections in india. Econometrica 86(6), 2123–2160.
Estache, A. and L. Wren-Lewis (2009). Toward a theory of regulation for developing countries:
Following jean-jacques laffont’s lead. Journal of Economic Literature 47(3), 729–70.
Estrada, F., W. Botzen, and R. Tol (2015). Economic losses from us hurricanes consistent with an
influence from climate change. Nature Geoscience 8(11), 880–884.
Fang, H. and Q. Gong (2017). Detecting potential overbilling in medicare reimbursement via hours
worked. American Economic Review, 562–591.
Finan, F., B. A. Olken, and R. Pande (2017). The personnel economics of the developing state. In
Handbook of Economic Field Experiments, Volume 2, pp. 467–514. Elsevier.
Finkelstein, A. and M. J. Notowidigdo (2019). Take-up and targeting: Experimental evidence from
snap. The Quarterly Journal of Economics 134(3), 1505–1556.
Glaeser, E., S. Johnson, and A. Shleifer (2001). Coase versus the coasians. The Quarterly Journal of
Economics 116(3), 853–899.
Glaeser, E. L. and R. E. Saks (2006). Corruption in america. Journal of Public Economics 90(6),
1053–1072.
Glaeser, E. L. and A. Shleifer (2003). The rise of the regulatory state. Journal of economic literature 41(2),
401–425.
Granja, J., C. Makridis, C. Yannelis, and E. Zwick (2020). Did the paycheck protection program hit
the target? Technical report, National Bureau of Economic Research.
Gray, C. (2019). Leaving benefits on the table: Evidence from snap. Journal of Public Economics.
Griffin, J. M., S. Kruger, and P. Mahajan (2021). Did fintech lenders facilitate ppp fraud? Technical
report, Available at SSRN 3906395.
Herd, P. and D. P. Moynihan (2018). Administrative Burden: Policymaking by Other Means. Russell
Sage Foundation.
Homonoff, T. and J. Somerville (2021). Program recertification costs: Evidence from snap. American
Economic Journal: Economic Policy, 271–298.
Hubbard, R. G. and M. R. Strain (2020). Has the paycheck protection program succeeded? Technical
report, National Bureau of Economic Research.
28
Humphries, J. E., C. A. Neilson, and G. Ulyssea (2020). Information frictions and access to the
paycheck protection program. Journal of public economics 190, 104244.
Imbens, G. W. and T. Lemieux (2008). Regression discontinuity designs: A guide to practice.
Journal of Econometrics 142(2), 615–635.
Jakobsen, K., K. Jakobsen, H. Kleven, and G. Zucman (2020). Wealth taxation and wealth ac-
cumulation: Theory and evidence from denmark. The Quarterly Journal of Economics 135(1),
329–388.
Jia, R. and H. Nie (2017). Decentralization, collusion, and coal mine deaths. Review of Economics
and Statistics 99(1), 105–118.
Kleven, H. J. (2016). Bunching. Annual Review of Economics 8, 435–464.
Kleven, H. J., M. B. Knudsen, C. T. Kreiner, S. Pedersen, and E. Saez (2011). Unwilling or unable to
cheat? evidence from a tax audit experiment in denmark. Econometrica 79(3), 651–692.
Kleven, H. J. and W. Kopczuk (2011). Transfer program complexity and the take-up of social
benefits. American Economic Journal: Economic Policy 3(1), 54–90.
Kleven, H. J. and M. Waseem (2013). Using notches to uncover optimization frictions and structural
elasticities: Theory and evidence from pakistan. The Quarterly Journal of Economics 128(2), 669–
723.
Kolstad, C. D., T. S. Ulen, and G. V. Johnson (1990). Ex post liability for harm vs. ex ante safety
regulation: substitutes or complements? The American Economic Review, 888–901.
Kotz, M., L. Wenz, A. Stechemesser, M. Kalkuhl, and A. Levermann (2021). Day-to-day temperature
variability reduces economic growth. Nature Climate Change 11(4), 319–325.
Laffont, J.-J. (1994). The new economics of regulation ten years after. Econometrica: Journal of the
Econometric Society, 507–537.
Londoño-Vélez, J. and J. Avila-Mahecha (2020). Behavioral responses to wealth taxation: evidence
from a developing country. In Annual Congress of the IIPF, Volume 3.
Lupia, A. and M. McCubbins (1994). Learning from oversight: Fire alarms and police patrols
reconstructed. Journal of Law, Economics, and Organization 10(1), 96–125.
McCrary, J. (2008). Manipulation of the running variable in the regression discontinuity design: A
density test. Journal of econometrics 142(2), 698–714.
McCubbins, M. and T. Schwartz (1984). Congressional oversight overlooked: Police patrols versus
fire alarms. American Journal of Political Science 28(1), 165–179.
29
Mookherjee, D. and I. P. Png (1992). Monitoring vis-a-vis investigation in enforcement of law. The
American Economic Review, 556–565.
Nichols, A. L. and R. J. Zeckhauser (1982). Targeting transfers through restrictions on recipients.
The American Economic Review 72(2), 372–377.
Niehaus, P. and S. Sukhtankar (2013). The marginal rate of corruption in public programs: Evidence
from india. Journal of Public Economics 104, 52–64.
Olken, B. (2007). Monitoring corruption: evidence from a field experiment in indonesia. Journal of
Political Economy 115(2), 200–249.
Olken, B. A. (2006). Corruption and the costs of redistribution: Micro evidence from indonesia.
Journal of Public Economics 90(4-5), 853–870.
Olken, B. A. and R. Pande (2012). Corruption in developing countries. Annu. Rev. Econ. 4(1),
479–509.
Pomeranz, D. (2015). No taxation without information: Deterrence and self-enforcement in the
value added tax. American Economic Review 105(8), 2539–69.
Posner, R. (1998). Economic analysis of law, 5th edn (new york: Aspen law & business).
Posner, R. A. (1974). Theories of economic regulation. Technical report, National Bureau of
Economic Research.
Rambachan, A. and J. Roth (2019). An honest approach to parallel trends. Unpublished manuscript,
Harvard University.
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.
Saez, E. (2010). Do taxpayers bunch at kink points? American economic Journal: economic policy 2(3),
180–212.
Shavell, S. (1984a). Liability for harm versus regulation of safety. The Journal of Legal Studies 13(2),
357–374.
Shavell, S. (1984b). A model of the optimal use of liability and safety regulation. The Rand Journal
of Economics 15(2), 271–280.
Shleifer, A. and R. W. Vishny (1993). Corruption. The quarterly journal of economics 108(3), 599–617.
Stigler, G. J. (1971). The theory of economic regulation. The Bell journal of economics and management
science, 3–21.
30
Strausz, R. (2005). Timing of verification procedures: Monitoring versus auditing. Journal of
Economic Behavior & Organization 59(1), 89–107.
TIGTA (2021). Trends in Compliance Activities Through Fiscal Year 2019. Number 2021-30-011. Wash-
ington, DC: Treasury Inspector General for Tax Administration.
Zamboni, Y. and S. Litschig (2018). Audit risk and rent extraction: Evidence from a randomized
evaluation in brazil. Journal of Development Economics 134, 133–149.
31
Tables
Table 1: Did screening affect fraud in PPP loans?
Dependent Variable: Overpayment dummy Overpayment rate Multiple loans dummy
(1) (2) (3) (4) (5) (6)
T0 0.013∗∗∗ 0.019∗∗∗ -0.000
(0.000) (0.001) (0.000)
Phase 2 -0.003∗∗∗ -0.003∗∗∗ -0.002∗∗∗ -0.002∗∗∗ 0.001∗∗∗ 0.001∗∗∗
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
T0 × Phase 2 -0.005∗∗∗ -0.005∗∗∗ -0.017∗∗∗ -0.017∗∗∗ -0.002∗∗∗ -0.002∗∗∗
(0.000) (0.001) (0.001) (0.002) (0.000) (0.000)
Control mean of outcome 0.008 0.008 0.003 0.003 0.002 0.002
Observations 2988104 2988104 2988104 2988104 2988104 2988104
Firm FE No Yes No Yes No Yes
32
Note: The table shows how the advanced documentation requirement affects the main indicators of fraud. The unit of observation is at the firm-phase level. The
sample is restricted to firms that received loans in both Phase 1 and Phase 2. Overpayment dummy is a dummy variable that takes a value of 1 when a firm received
any overpayment in a phase, and 0 otherwise. Overpayment rate on each loan is the overpaid amount divided by the maximum payment that a firm was eligible for.
For firms with multiple overpaid loans, the regression was run with the maximum rate among those loans. Multiple loans dummy is a dummy variable that takes
a value of 1 when a firm is identified to have been approved for more than one loan in any given phase, and 0 otherwise. T0 , the treatment group, consists of firms
with at least one loan greater than $150,000. Standard errors are clustered at the firm level. Significance levels are denoted as: * p<0.1, ** p<0.05, *** p<0.01.
Table 2: Did screening affect fraud through the extensive or intensive margin?
(a) Suspect firms defined based on overpayment in phase 1
Exited Whether loans Average loan amount Δ average no. jobs Δ average loan amount/job
Dependent Variable: from Phase 2 >150k in Phase2 in Phase 2 across phases across phases
(1) (2) (3) (4) (5)
T0 -0.032*** 0.911*** 140850.21*** -4.392*** 10.06
(0.001) (0.001) (2390.79) (0.082) (11.50)
Overpaid in phase 1 0.014*** -0.005*** -7398.27*** 2.384*** -22372.91***
(0.002) (0.001) (309.69) (0.071) (207.21)
T0 × Overpaid in Phase 1 -0.037*** -0.111*** -59343.86*** 20.618*** -18639.64***
(0.005) (0.006) (6160.15) (0.592) (1579.51)
Control mean of outcome 0.713 0.017 40273.74 -0.443 844.78
Observations 5128185 1494052 1494052 1494052 1494052
Fixed effects No No No No No
Phase 1 control No No Yes No No
33 (b) Suspect firms defined based on multiple loans in phase 1
Exited Whether loans Average loan amount Δ average no. jobs Δ average loan amount/job
Dependent Variable: from Phase 2 >150k in Phase2 in Phase 2 across phases across phases
(1) (2) (3) (4) (5)
T0 -0.033*** 0.909*** 140061.00*** -3.928*** -686.10***
(0.001) (0.001) (2387.03) (0.081) (37.53)
Multiple loans in phase 1 -0.042*** -0.005** 2493.60 0.440* 945.95***
(0.006) (0.002) (3115.85) (0.258) (135.49)
T0 × Multiple loans in Phase 1 -0.009 -0.137*** -28115.77* 3.466* -740.19
(0.016) (0.022) (16678.58) (1.871) (493.77)
Control mean of outcome 0.713 0.016 40326.64 -0.425 669.78
Observations 5128185 1494052 1494052 1494052 1494052
Fixed effects No No No No No
Phase 1 control No No Yes No No
Note: The table shows how the advanced documentation requirement affected the behavior of firms with overpayment (Panel (a)) and with multiple loans (Panel (b))
in Phase 1. The unit of observation is at the firm-phase level. The sample is restricted to firms that received loans in both Phase 1 and Phase 2. Exited from Phase 2 is
a dummy variable that takes a value of 1 when a firm showed up only in Phase 1, and 0 otherwise. Whether loans greater than $150K in Phase 2 is a dummy variable
that turns on when at least one of a firm’s loans in Phase 2 exceeds $150,000. The average loan amount is calculated from all the loans issued to a firm in Phase 2.
The average number of jobs is computed using the numbers of employees reported by a firm on its applications. The average loan amount per job is calculated by
dividing the total dollar amount of loan by the total number of employees reported in the same phase. For the last two outcomes, we further subtract Phase 2 values
from Phase 1 values to get an across-phase difference for each firm. T0 , the treatment group, consists of firms with at least one loan greater than $150,000. Standard
errors are clustered at the firm level. Significance levels are denoted as: * p<0.1, ** p<0.05, *** p<0.01.
Table 3: Were outcomes of bunching firms different from others?
Overpayment Overpayment Job gap between Percent job gap Multiple loans
Dependent Variable: dummy rate PPP and DNB PPP - DNB dummy
>300% >400% >500%
(1) (2) (3) (4) (5) (6) (7)
Bunchers 0.017∗∗∗ 0.016∗∗∗ 1.713∗∗∗ 0.066∗∗∗ 0.057∗∗∗ 0.055∗∗∗ 0.001
(0.002) (0.003) (0.354) (0.006) (0.005) (0.005) (0.001)
Phase 2 -0.004∗∗∗ -0.004∗∗∗ -1.083∗∗∗ -0.005∗∗∗ -0.005∗∗∗ -0.005∗∗∗ 0.000∗∗∗
(0.000) (0.000) (0.024) (0.000) (0.000) (0.000) (0.000)
Bunchers × Phase 2 -0.013∗∗∗ -0.015∗∗∗ -1.903∗∗∗ -0.010∗∗∗ -0.009∗∗∗ -0.014∗∗∗ -0.002∗∗∗
(0.002) (0.003) (0.310) (0.003) (0.003) (0.003) (0.001)
Control mean of outcome 0.010 0.006 3.674 0.096 0.073 0.059 0.002
Observations 2988104 2988104 970440 970440 970440 970440 2988104
Fixed effects No No No No No No No
34 The table shows how the outcomes of bunching firms differ from non-bunching firms. The unit of observation is at the firm-phase level. The sample is restricted to
firms that received loans in both Phase 1 and Phase 2. Bunchers are firms with at least one loan greater than $150K in Phase 1, all loans at $150K and below in Phase
2, and at least one loan in Phase 2 falling between $136K and $150K. Overpayment dummy is a dummy variable that takes a value of 1 when a loan is identified
as overpayment, and 0 otherwise. Overpayment rate is the overpaid amount divided by the maximum payment that a firm was eligible for. Job gap between PPP
and DNB is calculated by subtracting the number of employees shown on a firm’s DNB record (as of July 2021) from the maximum number of employees that firm
reported on its PPP applications in a phase. Note that DNB data shows the sum of all employee figures across all locations of a firm, while the number reported on
PPP applications can be from one of the offices. Therefore, to increase the comparability of that the two job numbers, we exclude from the regression observations
with large, negative (< -50) gaps between PPP and DNB figures and observations with DNB job numbers exceeding 500. For each firm, we further estimate the size of
this job gap as a fraction of its DNB employee number, and report dummy outcomes indicating very large, positive differences. Multiple loans dummy is a dummy
variable that takes a value of 1 when a firm is identified to have been approved for more than one loan in any given phase, and 0 otherwise. Standard errors are
clustered at the firm level. Significance levels are denoted as: * p<0.1, ** p<0.05, *** p<0.01.
Figures
.00006
Density of firms with overpaid loans
.00002 .00004
0
0 50000 100000 150000 200000
Overpaid USD
Phase 1 Phase 2
Figure 2: The distribution of overpaid amounts issued to borrowers with overpayments in Phase
1 and who reapplied in Phase 2. Data is at the firm-phase level, restricted to firms that applied
in both phases and had at least one overpayment in Phase 1. Overpaid amount on a loan is the
approved dollar value less the maximum payment due. For each firm that took out more than one
loan in a phase, the overpaid amount is the total overpaid value across all the loans the firm was
approved for in that phase. While the plot only shows up to $200,000 of overpayment, there is a
small number of firms with overpaid amounts exceeding this level.
35
.05 Phase 1 (2020) Phase 2 (2021)
Overpaid USD/max payment due
Estimated coefficient (relative to Aug20)
-.05 0
-.1
Apr May Jun Jul Aug Jan Feb Mar Apr May
Month
Coefficient Pre-treatment trend 95% CI
Figure 3: Event study plot for the overpaid amount (in USD) as a fraction of maximum payment
due per firm. Data is at the loan-date level, restricted to firms that took out loans in both Phase
1 and Phase 2. For each firm with multiple loans, overpayment rate is the maximum rate among
all the loans it was approved for. Each coefficient is obtained from the interaction terms between
treatment and the corresponding month as shown in equation Equation 1. The treatment group
consists of firms with at least one loan greater than $150,000 in Phase 1. The trend line connects
predicted monthly outcomes generated from regressing the pre-treatment coefficients on a linear
term of months.
36
Phase 1 (2020) Phase 2 (2021)
.02
Likelihood of overpayment
Estimated coefficient (relative to Aug20)
-.04 -.02 0
-.06
Apr May Jun Jul Aug Jan Feb Mar Apr May
Month
Coefficient Pre-treatment trend 95% CI
Figure 4: Event study plot for the probability that a firm was overpaid. Data is at the loan-date
level, restricted to firms that took out loans in both Phase 1 and Phase 2. Each coefficient is obtained
from the interaction terms between treatment and the corresponding month as shown in equation
Equation 1. The treatment group consists of firms with at least one loan greater than $150,000
in Phase 1. The trend line connects predicted monthly outcomes generated from regressing the
pre-treatment coefficients on a linear term of months.
37
.02 Phase 1 (2020) Phase 2 (2021)
Likelihood that a loan was a duplicate
Estimated coefficient (relative to Aug20)
-.04 -.02 0
-.06
Apr May Jun Jul Aug Jan Feb Mar Apr May
Month
Coefficient Pre-treatment trend 95% CI
Figure 5: Event study plot for the probability that a loan was a duplicate. A duplicate loan is
defined as any loan that is not the first loan issued to a firm. Data is at the loan-date level,
restricted to firms that took out loans in both Phase 1 and Phase 2. Each coefficient is obtained
from the interaction terms between treatment and the corresponding month as shown in equation
Equation 1. The treatment group consists of firms with at least one loan greater than $150,000
in Phase 1. The trend line connects predicted monthly outcomes generated from regressing the
pre-treatment coefficients on a linear term of months.
38
0.025
0.021
Likelihood of overpayment
0.013
0.010 0.015 0.020 0.008
0.005
0.005
0.000
Phase 1 Phase 2
Control Treatment 95% CI
Figure 6: Probability that a firm was overpaid. Data is at the firm-phase level. The sample is
restricted to firms that took out loans in both Phase 1 and Phase 2.
39
0.025
0.022
Overpaid USD/max payment due
0.005 0.010 0.015 0.020
0.003
0.003
0.001
0.000
Phase 1 Phase 2
Control Treatment 95% CI
Figure 7: Overpaid amount (in USD) as a fraction of maximum payment due as per rules. Data is
at the firm-phase level. For firms with multiple overpaid loans in a phase, the overpayment rate
plotted is the maximum rate among all those loans. The sample is restricted to firms that took out
loans in both Phase 1 and Phase 2.
40
0.0025 0.0024
Likelihood of having multiple loans
0.0018
0.0017
0.0010 0.0015 0.0020
0.0008
0.0005
Phase 1 Phase 2
Control Treatment 95% CI
Figure 8: Probability that a firm had multiple loans. Data is at the firm-phase level. The sample is
restricted to firms that took out loans in both Phase 1 and Phase 2.
41
0.000
−0.025
−0.050
−0.075
0.000 0.005 0.010 0.015
M
FLCI Original
Figure 9: Honest DID plot for the probability that a firm was overpaid.The blue line denotes the
coefficient obtained from the interaction term between treatment and Phase 2 in an event-study
regression similar to Equation 1. The only difference between this regression and Equation 1 is
that in this version, we let the treatment indicator interact with Phase 2 instead with individual
post-treatment months. The treatment group consists of firms with at least one loan greater than
$150,000 in Phase 1.
42
0.025
0.000
−0.025
−0.050
−0.075
−0.100
0.000 0.005 0.010 0.015
M
FLCI Original
Figure 10: Honest DID plot for the overpaid amount (in USD) as a fraction of maximum payment
due per firm. The blue line denotes the coefficient obtained from the interaction term between
treatment and Phase 2 in an event-study regression similar to Equation 1. The only difference
between this regression and Equation 1 is that in this version, we let the treatment indicator interact
with Phase 2 instead with individual post-treatment months. The treatment group consists of firms
with at least one loan greater than $150,000 in Phase 1.
43
0.025
0.000
−0.025
−0.050
0.000 0.005 0.010 0.015
M
FLCI Original
Figure 11: Honest DID plot for the probability that a loan was a duplicate. A duplicate loan is
defined as any loan that is not the first loan issued to a firm in any given phase. Data is at the
loan-date level, restricted to firms that took out loans in both Phase 1 and Phase 2. The blue line
denotes the coefficient obtained from the interaction term between treatment and Phase 2 in an
event-study regression similar to Equation 1. The only difference between this regression and
Equation 1 is that in this version, we let the treatment indicator interact with Phase 2 instead with
individual post-treatment months. The treatment group consists of firms with at least one loan
greater than $150,000 in Phase 1.
44
Figure 12: The theoretical impact of the advanced documentation requirement on the density of
all requests for the good from the program. This graph was created by parameterizing the model
as follows: 𝑣(𝑥) = ln(𝑥), 𝜋(𝑥) = 𝑥 2 /2, 𝜌 = 1/2, 𝜏 = 1, 𝜙 = 1/1000, and 𝐹1 , 𝐹0 are truncated normal
densities with mean 5 and support [1,9].
45
.03
.025
Density of loan amounts
.01 .015 .02
.005
0
100 110 120 130 140 150 160 170 180 190 200
Loan amounts (in '000 USD)
Phase 1 Phase 2 (first and second draws)
Figure 13: Density of loan amounts across the two phases of PPP program. Phase 2 borrowers
include both first and second time borrowers. The vertical red line represents the $150,000 thresh-
old. Bin width is $500.
46
.04
.035
Density of loan amounts
.01 .015 .02 .025 .03
.005
0
100 110 120 130 140 150 160 170 180 190 200
Loan amounts (in '000 USD)
Phase 1 Phase 2 - Draw 1 Phase 2 - Draw 2
Figure 14: Density of loan amounts across the two phases of PPP program. First and second draw
refers to whether a firm took a PPP loan for the first or second time. The vertical red line represents
the $150,000 threshold. Bin width is $500.
47
Phase I (p-value=0.266)
5
4
3
2
1
0
100000 120000 140000 160000 180000 200000
Loan Amount
Phase II (p-value=0.000)
5
4
3
2
1
0
100000 120000 140000 160000 180000 200000
Phase II (First draw) (p-value=0.000)
5
4
3
2
1
0
100000 120000 140000 160000 180000 200000
Phase II (Second draw) (p-value=0.000)
5
4
3
2
1
0
100000 120000 140000 160000 180000 200000
Figure 15: Distribution of approved loan amounts in phase I and phase II with p-value from a
McCrary Density Test (McCrary, 2008) of continuity of densities around the $150k cut-off
48
.8
P-value from Kolmogorov-Smirnov test
.1 .2 .3 .4 .5 .6 .7
0
100 110 120 130 140 150 160 170 180 190 200
Loan amounts (in '000 USD)
Notes: The vertical red line denotes the $150,000 threshold; the horizontal black line marks the point where p-value equals 0.05.
As shown in the plot, the bunching window for second drawers in Phase 2 is between $136,000 and $175,000.
Figure 16: P-values from the Kolmogorov–Smirnov test.
49
Appendix tables
Table A1: Characteristics of borrowers
Firms Firms reapplied
in Phase 1 in Phase 2 Bunchers Non-bunchers
Business profile in Phase 1
Business type
Corporation 0.295 0.349 0.483 0.348
Limited Liability Company 0.283 0.299 0.250 0.300
Subchapter S Corporation 0.134 0.137 0.169 0.137
Sole Proprietor/Self-employed 0.198 0.131 0.014 0.132
Others 0.090 0.083 0.084 0.083
Business size
At most 10 employees 0.775 0.740 0.210 0.743
11-20 employees 0.107 0.128 0.491 0.126
21-50 employees 0.076 0.092 0.250 0.091
More than 50 employees 0.042 0.040 0.048 0.040
Industry
Construction 0.096 0.096 0.129 0.096
Professional, Scientific, and Tech. 0.131 0.126 0.136 0.126
Healthcare and Social Assistance 0.103 0.100 0.154 0.099
Accommodation and Food Services 0.074 0.106 0.039 0.107
Retail trade 0.091 0.079 0.059 0.079
Others 0.480 0.475 0.473 0.475
Unanswered 0.026 0.019 0.011 0.019
Location
Urban 0.802 0.829 0.897 0.829
Rural 0.198 0.171 0.103 0.171
Registration date
Before February 2020 0.997 0.998 0.999 0.998
After February 2020 0.003 0.002 0.001 0.002
Share of loan proceeds in Phase 1
Payroll 0.958 0.956 0.962 0.956
Utilities 0.013 0.014 0.011 0.014
Rent 0.015 0.016 0.016 0.016
Debt interest 0.001 0.001 0.001 0.001
Observations 5,128,815 1,494,052 8,256 1,485,796
The table summarizes the characteristics of all firms that took out loans in Phase 1. The unit of obseration is at the firm
level. Bunchers are firms with at least one loan greater than $150K in Phase 1, all loans at $150K and below in Phase 2,
and at least one loan in Phase 2 falling between $136K and $150K. Business size is the maximum number of employees
a firm reported on its Phase 1 applications. Information about registration dates is from the OC dataset.
50
Table A2: Summary statistics of loan amounts and overpayment indicators
Phase 1 Phase 2
Approved amount per loan 101,589.58 42,748.19
(348,642.24) (141,714.77)
Share of overpaid loans 0.010 0.003
Overpaid USD per loan 725.80 91.80
(31,129.07) (7,858.05)
Overpaid USD per overpaid loan 75,367.62 29,480.93
(308,219.80) (137,714.31)
Overpaid USD per $10k of max payment 66.04 8.57
(2,852.92) (308.07)
Overpaid USD per $10k of max payment for overpaid loans 6,857.39 2,751.02
(28,259.83) (4,789.10)
Share of firms with multiple loans 0.002 0.002
Number of loans 5,136,454 6,338,537
Number of firms 5,128,185 5,745,589
Note: The table shows the mean value and standard deviation (values in parentheses).
51
Table A3: Did the screening requirement affect fraud in PPP loans? (Loans between $100-200K in Phase 1)
Dependent Variable: Overpayment dummy Overpayment rate Multiple loans dummy
(1) (2) (3) (4) (5) (6)
T0 0.003∗∗∗ 0.003∗∗∗ 0.000∗∗
(0.001) (0.001) (0.000)
Phase 2 -0.005∗∗∗ -0.005∗∗∗ -0.008∗∗∗ -0.008∗∗∗ -0.000 -0.000
(0.000) (0.001) (0.000) (0.001) (0.000) (0.000)
T0 × Phase 2 -0.002∗∗∗ -0.002∗ -0.003∗∗∗ -0.003∗∗ -0.000 -0.000
(0.001) (0.001) (0.001) (0.001) (0.000) (0.000)
Control mean of outcome 0.012 0.012 0.010 0.010 0.000 0.000
Observations 322752 322752 322752 322752 322752 322752
Firm FE No Yes No Yes No Yes
Note: The table shows how the screening requirement affects the main indicators of fraud. The unit of observation is at the firm-phase level. The sample is restricted
to firms that received loans in both Phase 1 and Phase 2 with more than half of the loans in Phase 1 falling between $100K and $200K. Overpayment dummy is a
52 dummy variable that takes a value of 1 when a firm received any overpayment in a phase, and 0 otherwise. Overpayment rate on each loan is the overpaid amount
divided by the maximum payment that a firm was eligible for. For firms with multiple overpaid loans, the regression was run with the maximum overpayment rate.
Multiple loans dummy is a dummy variable that takes a value of 1 when a firm is identified to have been approved for more than one loan, and 0 otherwise. T0 , the
treatment group, consists of firms with at least one loan greater than $150,000. Standard errors are clustered at the firm level. Significance levels are denoted as: *
p<0.1, ** p<0.05, *** p<0.01.
Table A4: Did the screening requirement affect overpaid amounts?
Overpayment amount
Dependent Variable: Overpayment amount (robustness check)
(1) (2) (3) (4)
T0 4549.48∗∗∗ 282.75∗∗∗
(174.80) (38.74)
Phase 2 -69.59∗∗∗ -69.59∗∗∗ -316.55∗∗∗ -316.55∗∗∗
(2.21) (3.12) (18.94) (26.79)
T0 × Phase 2 -3551.33∗∗∗ -3551.33∗∗∗ -226.96∗∗∗ -226.96∗∗∗
(175.93) (248.81) (43.49) (61.51)
Control mean of outcome 119.97 119.97 479.07 479.07
Observations 2988104 2988104 322752 322752
Firm FE No Yes No Yes
Note: The table shows how the screening requirement affects the overpaid amounts issued to firms. The unit of
observation is at the firm-phase level. The sample is restricted to firms that received loans in both phases of the program
(columns 1-2), in addition to having more than half of the loans in Phase 1 falling between $100K and $200K (columns
3-4). Overpaid amount on a loan is the approved dollar value less the maximum payment due. For each firm that
took out more than one loan in a phase, the overpaid amount is the total overpaid value across all the loans the firm
was approved for in that phase. T0 , the treatment group, consists of firms with at least one loan greater than $150,000.
Standard errors are clustered at the firm level. Significance levels are denoted as: * p<0.1, ** p<0.05, *** p<0.01.
53
Appendix: For Online Publication
String matching to create a unique firm level identifier and identification of multiple
loans to a firm in a given phase
The advanced documentation requirement was levied on second loans by firms that had already
taken a PPP loan once. Our analysis also required that we identify firms that had taken multiple
loans in a given phase. Since there were no unique firm level identifiers in the PPP data, we carried
out a string matching exercise. This was both a time and computation heavy activity.
Using names and addresses that were submitted by borrowers with their PPP application, the
following steps were used.
1. PPP loan data had an identifier for first or second time loans. We created two separate datasets
using this identifier.
2. To reduce computation burden, we further split the data by states.
3. Within each state and first or second time loans, we then string matched borrowers on names
and then addresses with other borrowers. This helped us identify borrowers with multiple
loans to the same firm.
4. For each state, borrowers were then matched across first and second time loans using the same
matching algorithm as in step 1. This helped us in creating a unique firm level identifier.
Following this algorithm we were able to create a unique firm-level identifier for approximately
70% of all second time loans. To verify the accuracy of these matches, we randomly sampled 1000
firms out of 2078901 total firms.Approximately 20 firms were chosen from each state. To check for
false positives, we manually checked whether the names and addresses of borrowers that were
identified as belonging to the same firm were in fact correct. If the addresses were different we
used Google Maps to check whether the address were very different in terms of distance. Of the
1000 firms, we could only find 2 firms for which there were mismatches in addresses.
54
File and source
- File
- screen-now-save-later-the-trade-off-between-administrative-ordeals-and-fraud.pdf
- Size
- 1,617,911 bytes
- SHA-256
- e6ea638dd7eed420a0bb11fd35976df88e2ff6e7d540a8a286fec6fb2fae55ee
- Original
- doi.org