Putting the Paycheck Protection Program into Perspective — Bureau of Labor Statistics Working Paper 542
Summary
BLS Working Paper 542, Putting the Paycheck Protection Program into Perspective: An Analysis Using Administrative and Survey Data, by Michael Dalton of the U.S. Bureau of Labor Statistics, dated November 5, 2021. The abstract states that after matching over 3 million loans from the $669 billion Paycheck Protection Program to administrative wage records, the author estimates a doubly robust dynamic difference-in-difference event study showing robust, causal impacts of the loans on employment, wages and opening status of establishments 7 months after PPP approval. It states a range of $20,000 to $34,000 of PPP spent per employee-month retained, with about 24% of the PPP money going towards wage retention in the baseline model, and that small and low-wage establishments show the largest impact.
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
About this paper. The author, at the Bureau of Labor Statistics, finds causal effects of the loans on employment, wages and the opening status of establishments seven months after approval. In back-of-the-envelope calculations, the author estimates $20,000 to $34,000 of PPP spent per employee-month retained, with about 24% of PPP money going to wage retention in the baseline model. Another paper, Did the Paycheck Protection Program Hit the Target?, finds that many firms used the loans for non-payroll fixed payments and savings buffers, "which can account for small employment effects." This paper does not measure fraud.
- Issuer
- Source documents
- Document type
- Putting The Ppp Into Perspective Administrative And Survey Data
- Case
- Putting The Ppp Into Perspective Administrative And Survey Data
Cite this document
“Putting the Paycheck Protection Program into Perspective — Bureau of Labor Statistics Working Paper 542.”
Original at www.bls.gov. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/putting-the-ppp-into-perspective-administrative-and-survey-data/
, SHA-256 21965b827bb2cc860e8748125f2a89b8e446f8e0db79932fcc43737836935166.Full text
BLS WORKING PAPERS
U.S. Department of Labor
U.S. Bureau of Labor Statistics
Office of Employment and Unemployment Statistics
Putting the Paycheck Protection Program into Perspective:
An Analysis Using Administrative and Survey Data
Michael Dalton, U.S. Bureau of Labor Statistics
Working Paper 542
November 5, 2021
Putting the Paycheck Protection Program into Perspective:
An Analysis Using Administrative and Survey Data
Michael Dalton
U.S. Bureau of Labor Statistics∗
November 5, 2021
Abstract
After matching over 3 million loans from the $669 billion Paycheck Protection Program to admin-
istrative wage records, I estimate a doubly robust dynamic difference-in-difference event study showing
robust, causal impacts of the loans on employment, wages, and opening status of establishments 7 months
after PPP approval. Doing back-of-the-envelope calculations, I find a range of $20,000 to $34,000 of PPP
spent per employee-month retained, with about 24% of the PPP money going towards wage retention in
the baseline model. Small and low-wage establishments show the largest impact from PPP.
∗ contact: dalton.michael@bls.gov
1 Thank you to Brittany Borg, Trent Thompson, Elizabeth Handwerker, Mark Loewenstein, Anne Polivka, Jeremy Oreper,
C.J. Krizan, Mina Kim, Emily Thomas, David Ratner, and seminar participants at the Small Business Administration and
Bureau of Labor Statistics for helpful comments. The views expressed herein are those of the author and do not necessarily
reflect the views of the United States Bureau of Labor Statistics.
1
1 Introduction
The Coronavirus Aid, Relief, and Economic Security Act (CARES Act) passed in March 2020 established the
Paycheck Protection Program (PPP) administered by the Small Business Administration (SBA). Initially,
$349 billion was allocated to the program, followed by an additional $320 billion authorized in April 2020.
This $669 billion for the PPP program alone amounted to approximately 85% of the initial estimated size of
the entire American Recovery and Reinvestment Act of 2009 making it a remarkable program in both size
and scope. The majority of employers in the United States were eligible for a loan through the program and
the explicit objective of the program was to keep employers from terminating employees by allowing them
to maintain the typical wages of their employees even when their businesses were adversely affected by the
pandemic.
This paper provides a thorough analysis that answers the questions of how many jobs and how much in
wages were protected by the PPP, the extent that businesses were able to remain open after PPP and for
how long, and which employers benefitted the most from the program. Linking the full set of administrative
PPP loan microdata to the Quarterly Census of Employment and Wages (QCEW) allows for observing
monthly employment and quarterly wages before and after PPP approval. This paper is also able to contem-
poraneously observe employers who have not received PPP loans because the QCEW includes all employers
that pay into the Unemployment Insurance system covering 95% of all employment in the United States.
The other benefit to this linkage is that the QCEW is the sampling frame for all other Bureau of Labor
Statistics employer surveys. This allows further linking of unique information for each establishment from
other surveys—monthly hours worked reported in the Current Employment Statistics survey (CES), occu-
pational employment composition in the Occupational Employment and Wage Statistics survey (OEWS),
and information to confirm the quality of record linking using the 2020 Business Response Survey (BRS)
that asked employers questions related to the pandemic. Combined, this offers a rich collection of data on
employers to more fully and granularly understand the impact that PPP had on the labor market.
A number of papers have offered preliminary insight into the question of the impact of the PPP, and
Section 5.2.1 and Table 5 consider these papers in detail. One common theme in these papers has been the use
of the SBA employment cutoff for PPP eligibility,1 and this has resulted in findings of null or smaller effects of
PPP on employment. Another set of results using instrumental variables find larger local average treatment
effect estimates2 . These local average treatment effects can be identical to the average treatment effect if
there is no heterogeneity in the effect of PPP across timing of PPP approval or employer characteristics.
Furthermore, some of these papers have had to rely on geographically aggregating the data because they do
not have access to both outcome data and PPP approval information for individual employers limiting the
heterogeneity analysis that can be done on employer characteristics.
This paper expands on the previous research in some key ways. First, this is the first paper to rely on
administrative wage records that covers more than 95% of employment in the United States. Second, this
paper uses an econometric strategy based on Callaway and Sant’Anna (2020) and Sant’Anna and Zhao (2020)
to estimate the effect of PPP via a doubly robust dynamic difference-in-difference routine, only Autor et al.
(2021) have thus far made use of a similar strategy. This is an improvement because it provides estimates
that are consistent in the presence of heterogeneity in the effect of PPP on employers, allows for an event
study style estimate of parameters, and it allows for controlling for static characteristics of the employer to
make the estimation more robust to selection into PPP approval. Third, this paper is the first to combine
1 Autor et al. (2020), Chetty et al. (2020), Hubbard and Strain (2020)
2 Bartik et al. (2021), Faulkender et al. (2020), Doniger and Kay (2021)
2
estimates of the effect of PPP on employment, wages, and closure status simultaneously to paint a complete
picture of what has happened to employers over the pandemic recession and how PPP has impacted those
employer outcomes. Fourth, this paper translates the PPP program into dollars spent per employee-month
retained as well as dollars-of-wages retained to better understand the impact of PPP. Lastly, due to the
rich information related to employer characteristics and geography contained in the BLS wage records and
related surveys, this paper is able to conduct a thorough heterogeneity analysis that prior research have
not had the opportunity to do.3 This heterogeneity analysis puts the full SBA program into perspective by
identifying characteristics where PPP had the biggest impact.
This paper also contributes to the flood of findings assessing the impact of the COVID-19 pandemic
on the labor market. Most relevant to this paper are findings in Cajner et al. (2020), which shows small
employers and low-wage workers suffered the largest employment losses early in the pandemic; Crane et al.
(2021) find significant closures for small employers, though, by the latter half of 2020, not drastically dif-
ferent than previous years; Dalton et al. (2020) which finds that the smallest employers had a deep dip in
employment at the start of the pandemic but bounced back very quickly; and Dalton et al. (2021) , which
finds disproportionate effects of the pandemic on employment for low-wage workers and employers through
2020 and into 2021. This paper places these findings in the context of the large-scale PPP program to
understand how some of these patterns may have been influenced by it. The results in this paper show that
PPP had a larger effect for smallest, young, and low-wage employers and these findings help explain some
of the patterns found in prior research examining employment over the pandemic.
2 Data
2.1 Paycheck Protection Program Administrative Data
In December 2020, the Small Business Administration (SBA) began publishing data for all approved PPP
loan applications including the loan amount, date that the loan was approved, business name, business
type, address of business, reported industry, and reported number of jobs saved due to the loan. The key
information used here is the date of the loan approval, the name and address of the business, and the loan
amount. This analysis only considers loans approved in 2020 and future work will examine the loans approved
in 2021.
2.2 Quarterly Census of Employment and Wages
The Quarterly Census of Employment and Wages (QCEW) is an administrative collection of monthly em-
ployment and quarterly wages for all establishments that pay into the unemployment insurance system,
covering more than 95% of all employment. The data can be linked over time to follow any single establish-
ment. Employment and wages can be tracked prior to the pandemic until the most recent month available
in the data - March, 2021. In addition to employment levels, the QCEW also allows the tracking of closure
status on a month-to-month basis. Closures are defined as an establishment having zero employment or
returning an inactive code for a particular month, and the QCEW allows monthly tracking of an employer
who, for example, moves from open to closed and back to open over the course of three months.
The QCEW also contains employer name and address information for every establishment, which allows
for linking to the PPP. Lastly, since the QCEW is the sampling frame for all employer surveys conducted
3 Outside researchers can apply for access to the microdata here: https://www.bls.gov/rda/home.htm
3
by the Bureau of Labor Statistics, responses to other surveys that contain valuable additional information
can be also be linked to the PPP data. The QCEW is therefore well-suited for assessing the effects of PPP
on private businesses.
2.3 Business Response Survey
The Business Response Survey (BRS) was an online survey conducted during the summer of 2020 that asked
businesses a series of questions related to the pandemic. This survey collected 162,000 responses from a
nationally representative sample from July through September 2020. One question in the survey asks the
establishment whether it had received a loan or grant from the federal government. Since there were a
number of such programs enacted at the time and the question was not specific to PPP, affirmative answers
are not sufficient for denoting that PPP was received, but replying yes to the question should be a necessary
condition if the establishment had received a PPP loan.4 After linking the PPP data to the QCEW, the
data was then further linked to BRS survey responses to validate that a high proportion of BRS respondents
who received a PPP loan are also reporting having received a loan or grant from the federal government.
2.4 Occupational Employment and Wage Statistics survey
The Occupational Employment and Wage Statistics survey (OEWS) is a semiannual survey of employers
used to construct wage and employment estimates across detailed occupations. This is a unique survey
both in terms of its focus on obtaining detailed occupation information from employers and the size of
the survey–more than 300,000 responses per year. The information about occupational distribution within
an establishment was linked to the PPP data to identify how PPP affected establishments with varying
occupational compositions.
2.5 Current Employment Statistics survey
The Current Employment Statistics survey (CES) is a monthly survey of nearly 700,000 worksites used to
construct monthly hours, wage and employment estimates. This survey is used as a robustness check against
the results from the QCEW analysis and additional analysis on hours worked and pay per hour from CES
add more nuance to the QCEW results.
3 Match Rate
Despite the PPP data including both employer names and addresses, it is still a nontrivial task to link the
two datasets. The QCEW can contain multiple addresses for a particular business, and these addresses
may not be the same as the one provided by the individual filing the PPP loan application with the bank.
Furthermore, the QCEW has both trade and legal names of establishments, which may take a different
format or be an entirely different name than what is used on the PPP loan application. For this reason,
this paper employs a record-linking algorithm to identify the best match of an establishment in the QCEW
to the information contained in the PPP loan application based on address, business name, and industry.
Further details on the match are provided in Appendix Section A.1.
4 Approximately 5% of the sample responded that they did not know whether a loan or grant was received. For the purposes
of this analysis, these respondents are treated as missing and the percentages are relative to the pool of respondents answering
”yes” or ”no”.
4
Table 1 provides match details from linking the loan data to the Fourth Quarter of 2019 QCEW. From
the overall set of PPP loans approved in 2020, 61% of the PPP loans and 85% of the PPP loan money is
successfully matched to the QCEW. However, a number of business types reported on the loan applications
are typically out of scope for the QCEW, such as ”self-employed individuals”, ”independent contractors”,
”sole proprietorship”, and ”non-profits” in the Religious Organization industry. The second row in the table
removes these loans. From the set of the remaining 3.8 million PPP loans, 88% of the dollars distributed are
matched, which is for 76% of the loans. There are half-a-million PPP loans that state that only one job was
saved on the application, potentially suggesting this was a single-employee business, which likely would not
be included in the QCEW because they would not pay taxes into the UI system. After removing these estab-
lishments, the match rate improves to 81% of PPP loans and 89% of the PPP loan dollar amount.Another
600,000 PPP loans state that zero jobs were saved from the receipt of the loan, which could be a clerical
error, or also indicative of the business having no employees besides the owner filling out the application.
Removing these PPP loans further increases the match rate to 83% of PPP loans. Figure 1 shows the match
rate, weighted by loan amount, by the number of jobs reported saved on the PPP application. As seen in
the Figure, for applications reporting some jobs saved, there is a positive relationship between reported jobs
saved and the match rate, moving well above a 90% match rate in PPP loan dollars for the largest categories.
Table 1: Match Rate of PPP to Wage Records
Total Number Total Dollar % of Loan $
% of Loans
Descriptor of Loans Amount Amount
Matched
(millions) ($billions) Matched
All Loans 5.1 511 61.2 84.8
After removing...
Out of Scope Business Types 3.84 483.5 76.3 87.9
Out of Scope + Reporting Only 1 Job 3.3 476.7 80.5 88.5
Out of Scope + Reporting 0 or 1 Job 2.7 412.2 82.5 88.8
Notes: Matching between PPP data and the QCEW using record linking techniques identifying similar addresses and
employer names. Details in Appendix A.1.
There are a number of reasons that a PPP loan may not match. In some cases, correct matches are
removed because they do not meet the necessary threshold for a quality match by text of employer name
and address; in cases where this was a mistake, this would be a standard type 2 (false negative) error. If
the addresses do not match exactly between the PPP loan and the QCEW, and a business name provided
in the PPP loan application is very different from either the legal or trade names in the QCEW, this will be
rejected as a potential match. There may also be cases where the business that the PPP loan is intended for
is simply not in the QCEW database, possibly because the establishment does not pay into the UI system.
Since the unmatched PPP loans tend to skew towards smaller loan amounts and fewer reported jobs saved,
this may imply that smaller establishments were more difficult to match. If smaller employers are more
likely to have incomplete information on their loan application, then they may be less likely to match to
the QCEW. Further, smaller employers may be less likely to pay into the UI system and therefore excluded
from the QCEW.
Considering the overall high match rate in dollars disbursed, and the evidence that the unmatched loans
may likely be out-of-scope for the QCEW, we move forward with the analysis and examine type 1 (false
positive) errors in the matching.
5
Figure 1: Match Rate by Number of Reported Jobs
Proportion of Loan Amount Matched,
Based on Jobs Reported on PPP Application
0
[8% of All $]
1
[3% of All $]
# of Jobs Reported on PPP Application
2-5
[7% of All $]
6-10
[9% of All $]
11-50
[31% of All $]
51-499
[39% of All $]
500
[3% of All $]
0.0 0.2 0.4 0.6 0.8
Proportion
Notes:Percentage of PPP loan amount matched to the QCEW using record-linking techniques described in Appendix A.1.
3.1 Match Quality in the PPP-to-QCEW Linkage
Figure 2 displays the distribution of PPP loans based on match type, weighted by loan amount. The match
type is defined by the geography level where a sufficient establishment match in the QCEW is found for the
PPP loan. The second feature of the match type is whether it was an exact or fuzzy match on employer
name5 . Over 50% of the PPP loan amount falls into the highest quality match category - exact address and
exact name match - and 70% of the $511 billion in PPP loan amount have at least an exact address match.
Figure 3 gives a box plot of the ratio of jobs reported on the PPP application to the average monthly
employment reported in the QCEW for 2019 for each of the match types. The dotted light blue line
represents the expected median ratio of 1. However, the reporting of jobs saved on the PPP application
was an imperfect measure, as it was left up to the interpretation of the applicant. Also, there were a large
percentage of applications that left this question blank and zero was filled in for the applicant by the bank.
Since applications reporting zero jobs made up 8% of the loan money disbursed (see Figure 1), this is enough
to bring down the observed median ratio below the expected ratio of 1. Additionally, the employment may
have changed from 2019 to the time of application submission in 2020. For these reasons, it should be
expected that there is some variance around the median ratio. It is worth noting that as the match quality
decreases, going left to right, the interquartile ranges get wider, which is further evidence of somewhat lower
quality matches. Though, importantly, the box plots for the first three match types are fairly similar, and
make up approximately 97% of the total PPP matches and have medians close to 1.
A measure that is less prone to mismeasurement on the application is the loan amount itself. One
5 ”Fuzzy” match refers to the identification of employer names with similar text. Details are in Appendix A.1
6
Figure 2: Proportion of Loans Matched
Distribution of Loan Amount Matched,
Based on Match Type
Exact Address Match
Address Match
Exact Match,
City
Exact Match,
County
Type of Match
Exact Match,
State
Fuzzy Match,
City
Fuzzy Match,
County
Fuzzy Match,
State
No Match
0.3 0.0
0.4 0.5 0.1 0.2
Proportion
Notes: The record-linking algorithm matching the PPP to the QCEW moves in successive steps in the order displayed in the
graph. If a sufficient employer match is found at one stage of the match type, then that PPP observation is removed.
Figure 3: Match Rate by Number of Reported Jobs
Box Plot of Ratio of Jobs Reported
in PPP to Avg. Ann. Emp. in LDB
3.5
PPP Jobs Saved = 2019 Employment
3.0
2.5
2.0
1.5
1.0
0.5
0.0
Fuzzy Match, Fuzzy Match, Fuzzy Match,
Exact Address Match
Address Match
Exact Match, Exact Match, Exact Match,
City County State City County State
Type of Match
Notes: Dotted blue line represents equal average employment in 2019 and number of jobs reported on PPP application. The
green line is the median and the blue box shows the interquartile range.
7
Figure 4: Match Rate by Number of Reported Jobs
Box Plot of Ratio of Loan Amount
in PPP to Annual Wages in LDB
10 Weeks of Wages
0.5
0.4
0.3
0.2
0.1
0.0
Fuzzy Match, Fuzzy Match, Fuzzy Match,
Exact Address Match
Address Match
Exact Match, Exact Match, Exact Match,
City County State City County State
Type of Match
Notes: The dotted blue line represents 10 weeks of wages relative to a full year’s wages. The green line is the median and the
blue box shows the interquartile range.
restriction on the PPP loan amount was that it could not exceed 10 weeks of wages for establishments prior
to the pandemic. This should give a ratio of approximately 0.19 for the PPP loan amount to the 2019 wages
reported by the establishment. Figure 4 shows the box plots for this measure. The light blue dotted line
represents a ratio of 0.19, or a loan equal to 10 weeks of year’s wages. Each of the match types have a median
near the 19% mark, with interquartile ranges around 13% to 22% at every geography level, which is more
refined than the state level. This provides evidence that many of the matches are of high quality.
As a final test for quality of matches, Table 2 displays a test for false positives using data obtained from the
Business Response Survey (BRS). Each row represents a different match type in the record-linking process,
based on type of employer name match (column 1), geography where the match was identified (column 2),
and the quality of text match between employer names (column 3). Column 4 reports the number of BRS
respondents for each row, and the final column reports the percentage of respondents reporting in the BRS
that they had received a loan or grant from the government as of the time of the survey, which was collected
from end of July 2020 through September 2020. The most relevant comparison is the last column in each row
relative to the bottom row - the percent of establishments not matched to a PPP loan that reported having
received a loan or grant from the government in the BRS. As is clear from the table, having been matched to
a PPP loan correlates very strongly to a much higher percentage reporting having received a loan or grant.
Additionally, the three highest quality match types making up the largest percentage of matches, all have
at least a 94% affirmative response to the BRS question.
Combined, this evidence suggests that the accepted matches are of very high quality, and justifies con-
tinuing forward with identifying the effect of PPP on employment and closures.
8
Table 2: Confirming Successful Data Linkage Using the Business Response Survey (BRS)
Percent Reporting
Fuzzy Match Number of BRS
Match Type Geography Received Loan/Grant
Score Respondents
in BRS of Any Type
Exact Match Address Exact 45714 97.7%
Fuzzy Match Address - 11024 94.6%
Exact Match City Exact 6485 97.0%
Fuzzy Match City High 341 92.8%
Fuzzy Match City Medium 471 91.4%
Fuzzy Match City Low 531 86.4%
Fuzzy Match City Lowest 235 80.0%
Exact Match County Exact 1126 96.4%
Fuzzy Match County High 64 95.3%
Fuzzy Match County Medium 58 91.4%
Exact Match State Exact 997 92.3%
BRS Respondents with no PPP Match 95338 39.0%
Notes: Record linking between PPP administrative data and the QCEW using matching techniques identifying similar
addresses and employer names. Business Response Survey was an online survey of employers fielded between July
and September 2020. Details in Appendix.
4 PPP Take-up Rate
The take-up rate is the set of establishments that received PPP loans divided by the eligible set of estab-
lishments. For the denominator, the number of eligible establishments in the QCEW, where eligibility is
based on industry, establishment employment, and firm-level links between establishments. Typically, the
SBA defines small businesses according to industry-specific employment cutoffs, where employment levels are
based on all affiliated establishments. One caveat to this definition is that firms that are officially designated
as franchises have their employment determined at the establishment level.6 Employers in the QCEW are
matched to the franchise database to identify which EINs (employer identification numbers) fall under the
franchise designation. This is an imperfect measure as there is no geography that allows for identifying
quality matches - only employer name. This means there is both a higher likelihood of false positives, as well
as a higher proportion of franchises that are not matched to the QCEW because of employer names that are
too different to find quality matches.
One significant change that the SBA made to the eligibility criteria for PPP is allowing all employers
with NAICS code 72 (the Accommodation and Food Services sector) to be eligible if the establishment-level
employment is below 500, regardless of the firm-level employment or franchise status.7
The last key point for eligibility determination is that EIN is an imperfect link of establishment affiliates.
The SBA defines entities as affiliated “if one has the power to control the other or a third party has the
power to control both” establishments.8 For a variety of reasons, establishments that are in the same firm are
not guaranteed to have the same EIN. This makes determining firm-level employment, or affiliate-summed
employment, and therefore eligibility an imperfect process. This would most likely lead to incorrectly
assigning eligibility status to an establishment because the EIN employment is undercounting the true
6 https://www.sba.gov/sba-franchise-directory
7 The intention of this eligibility change was to be less restrictive on qualifications for the hardest-hit sector.
8 https://www.sba.gov/sites/default/files/bank_eligibility_questionnaire_0.pdf
9
firm-level employment. This would increase the denominator and result in undershooting the true take-up
rate.
For the numerator, PPP receipt is defined similarly as eligibility: if a PPP loan is matched to an EIN that
is not a franchise or NAICS 72 (the Accommodation or Food Services sector), all establishments with that
same EIN are determined to have received the PPP loan. For franchisees and NAICS 72 establishments, PPP
receipt is determined at the establishment level. One ad hoc imposed additional criteria: if the average loan
amount per employee at the EIN-level is less than $500, then that loan is instead treated as an establishment-
level loan. This somewhat arbitrary cutoff is an attempt to identify loans that may have been misidentified
as applying to an entire firm when they were more likely applied for by the single establishment. 9% of those
matched to an EIN-level PPP loan are instead treated as an establishment-specific PPP loan.
Overall, the take-up rate among eligible establishments is 45%, and 54% when weighted by employment.
As 24% of in-scope loans are not matched, this take-up rate is an underestimate. Though, as described in
the previous section, a number of these unmatched loans are likely out of scope for the QCEW. Specifically,
there are about 1.2 million loans that report having an employment of 1 on the PPP application, and many
of these PPP applicants are likely not in the QCEW frame.
Another reason why this may be an underestimate of the true take-up rate is the imperfect, under-
counted measure of affiliate-summed employment, which increases the denominator by identifying ineligible
establishments as eligible. However, when the sample is focused only on NAICS 72 establishments where
PPP eligibility is determined by employment at the establishment level thus sidestepping the affiliate-linked
issue, a 48% take-up rate is observed, only a slight increase above the overall take-up rate.
These take-up rate estimates are noticeably lower than other estimates. One example is the Census
Pulse Small Business Survey, which consistently found that approximately 72% of eligible establishments
reported having received PPP.9 One notable difference is that the Census Pulse survey targeted single-unit
establishments with employment less than 500, and a number of industries are out-of-scope for the Census
Pulse. When the QCEW sample is restricted to a comparable definition, the take-up rate is 49% (and 66%
when employment weighted). Adding to the numerator the 640,000 loans that are a) in scope, based on
reported business type, b) reported jobs saved greater than 1 or equal to 0 on the PPP application, and c)
remain unmatched to the QCEW, this would increase the take-up rate to 61% for the subset of establishments
that meet the Census Pulse sample definition.
Even conditional on an underestimate of the true take-up rate, it is still of interest to examine how the
take-up rate varies across different employer characteristics. Figure 5 displays eligibility and PPP approval
by sector, where the blue bar is the overall proportion of employment among eligible establishments and the
orange line is the proportion of employment among establishments receiving PPP. In cases where the orange
bar exceeds the blue bar, this shows that the group has a higher take-up rate than average. Retail trade has
a higher take-up rate than average, and leisure and hospitality has a lower take-up rate than average. The
other sectors all manage to stay approximately near the average.
Figure 6 breaks establishments down by their average wage in 2019. The very lowest wage establishments
had a slightly lower-than-average take-up rate, and the 2nd lowest wage establishments (where the average
employee makes between $20,000 and $40,000 a year) had a somewhat higher-than-average take-up rate.
Table 7 examines take-up by establishment size, determined as the average monthly employment in 2019.
The majority of eligible establishments are very small - slightly more than 50% had an average of fewer than
3 employees in 2019. However, there is also a big gap in the take-up rate, as only about one-third of PPP
9 Based on percentages from September through November 2020. https://portal.census.gov/pulse/data/
10
Figure 5: Take-up Rate by Sector
Employment Distribution Within
Same-Colored Bars
Eligible Establishments
PPP Recipients
0.20
0.15
Proportion
0.10
0.05
0.00
Financial / Insurance
Professional Serv.
Transportation and Warehousing
Educational Serv. Wholesale Trade
Other services
Management Serv.
Leisure / Hospitality
Manufacturing
Agriculture Construction Health Care Information Retail Trade
Utilities
Real Estate
Sector
Figure 6: Takeup Rate by Avg. Wage at Establishment in 2019
Employment Distribution Within
Same-Colored Bars
Eligible Establishments
PPP Recipients
0.35
0.30
0.25
Proportion
0.20
0.15
0.10
0.05
0.00
20k-40k 40k-60k 60k-80k
¡20k 80k+
Avg. Wage per Employee
11
Figure 7: Takeup Rate by Establishment Size in 2019
Establishment Distribution Within
Same-Colored Bars
Eligible Establishments
PPP Recipients
0.5
0.4
0.3
Proportion
0.2
0.1
0.0
500+
0-2 3-5 6-9
10-49 50-249 250-499
Establishment Size
receiving establishments are in the smallest size class. This can be interpreted in two competing ways: either
it is evidence of poorer linking between the PPP and QCEW of the smallest establishments (demonstrated
in Figure 1), or it may simply be that the smallest UI-paying employers had the lowest take-up rate. All
other size classes show a larger-than-average take-up rate.
Table 8 shows take-up by occupation employment in establishments. This sample is conditional on also
being surveyed in the 2017-2019 OEWS. Establishments employing sales occupations had the highest relative
take-up rate, though food preparation occupations also had a larger-than-average take-up rate. Production
and construction occupations seemed to be underrepresented in PPP take-up relative to their share of
employment in eligible establishments.
Overall, take-up rates are fairly similar across a variety of employer characteristics, but the lower-than-
average take-up rate among the smallest employers is notable, which may be an artifact of higher false
negative matches to PPP loans among these employers. This homogeneity in take-up likely speaks to the
expansive scope of the PPP, which allowed a variety of employers access to the program.
5 The Effect of PPP on Employer Outcomes
5.1 Estimation Strategy
One reason this paper is unique is because of its ability to rely on an analytical sample that represents the
entire universe of UI-paying establishments through the QCEW, which covers more than 95% of employment
in the United States. This is an improvement over other datasets that rely on subsamples of establishments in
particular industries or size classes, or where a selection bias into those samples may confound any estimates.
Another improvement is relying on a dynamic difference-in-difference (DDID) estimation strategy based
on Callaway and Sant’Anna (2020), referred to as CS from this point forward. This estimation routine has
12
Figure 8: Takeup Rate by Occupational Employment
Employment Distribution Within
Same-Colored Bars, OES Sample Weights
Transportation and Material Moving
Production
Installation, Maintenance, and Repair
Construction and Extraction
Farming, Fishing, and Forestry
Office and Administrative Support
Sales and Related
Personal Care and Service
Building and Grounds Cleaning and Maintenance
Food Preparation and Serving Related
2-Digit SOC
Protective Service
Healthcare Support
Healthcare Practitioners and Technical
Arts, Design, Entertainment, Sports, and Media
Education, Training, and Library
Legal
Community and Social Services
Life, Physical, and Social Science
Architecture and Engineering
Computer and Mathematical
Business and Financial Operations
Management Eligible Establishments
PPP Recipients
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14
Proportion
Notes: Distribution based on 2019 employment in QCEW, where occupational composition for the establishments comes from
the 2017-2019 OEWS.
many useful features that are valuable to this context. First, it allows for estimating an average treatment
on the treated effect (ATT). The treatment effect of interest is the impact of receiving a PPP loan on
establishment outcomes, such as employment, the probability of closure, and wages. The ATT is the causal
parameter of primary interest with regards to the impact of PPP.
Second, DDID allows for estimating dynamic treatment effects. This is important for constructing an
event study analysis that follows the path of employer outcomes of an establishment from prior to the start
of the pandemic through March 2021. The nature of the pandemic shifted throughout 2020, on both a local
and global scale, and the PPP has the ability to influence employer behavior over time. The PPP loans
could be large enough to cover 10 weeks of wages, so the effects of PPP may reasonably be observed for three
months, or more. Additionally, businesses may have used the PPP loan as a supplement for maintaining
wages as businesses were only partially closed. This may extend the impact of PPP to many months beyond
10 weeks. Lastly, the PPP can have longer-term impact if it explicitly allowed a business to remain open.
All of these impacts should be studied in a dynamic framework, which DDID allows for.
Estimates of ATT using two-way fixed effects (TWFE) are very common in the scientific literature.
Recently, researchers have identified issues with TWFE that may give misleading ATT estimates10 . When
there are heterogeneous treatment effects between the treatment groups, TWFE can give coefficient estimates
that are different from the true ATT. In the PPP context, the month of PPP receipt defines discrete treatment
groups.11 Given the timing of the first round of PPP loans, there is one treatment group for each month
from April 2020 through August 2020. Since the labor market was very fluid through 2020 as the pandemic
10 See Goodman-Bacon (2021).
11 The reported monthly employment level in the QCEW is for the pay period including the 12th of the month. The
establishment is determined to have received a PPP loan in month t if the date of the PPP approval was before or on the 12th
of the month. If it comes after the 12th, then the treatment group will be t + 1.
13
spread across the country, there is reason to think receiving a PPP loan in August may have had a different
impact on employment than receiving the loan in the first tranche of loans in April. Estimating separate
dynamic treatment effects allows for an event study that considers ATT effects that differ based on when
the loan was approved.
The following notation closely follows the one used by CS. The key benefit of this strategy is that it allows
for different treatment timing, heterogeneous treatment effects, and controlling for time invariant covariates
to estimate an uncontaminated ATT from the regression.
AT Tp,t = E[Yt (1) − Yt (0)|X, P P Pp = 1] (1)
where the AT Tp,t is the ATT effect of PPP on employment Yt in month t for establishments receiving
PPP in month p. There is a unique AT Tp,t for each calendar month t from February 2020 through March
2021 and for each treatment group (defined by month of PPP approval) p from April 2020 through August
2020.12
One implication is the assumption that AT Tp,t = 0 for all t < p − δ, conditional on the time invariant
covariates X and δ is the number of periods of anticipation of treatment. This simply states that there
should be no effect of PPP on employment in months before the establishment anticipates receiving PPP.
Yt (1) is the outcome when PPP is received, compared to Yt (0), the counterfactual outcome when PPP is not
received. In this context13 , contemporaneous counterfactual outcomes remain unobserved. As an alternative,
assuming that the time trend for untreated establishments is comparable to what the counterfactual time
trend would have been for treated establishments, then untreated establishments are a sufficient control
group. Testing for AT Tp,t = 0 where t < p − δ becomes an informative test of the conditional parallel trends
assumption needed for unbiased estimates of AT Tp,t where t ≥ p.
E[Yt (0) − Yt−1 (0)|X, P P Pp = 1] = E[Yt (0) − Yt−1 (0)|X, C = 1] ∀ t > p − δ (2)
Equation 2 represents the conditional parallel trends assumption, which says that conditional on a set
of time invariant covariates X, the employment change from month t − 1 to t for establishments receiving
PPP in month p in the counterfactual world where they never received PPP equals the same change in
employment for the control establishments who have not received PPP, C = 1. Note that the control group
can both be establishments that never receive PPP and the group of establishments that eventually receive
0
PPP but have not as of time t. In order words, establishments treated at month p > t. Equation 2 is the
key assumption necessary for estimating unbiased ATT.14
The initial rush to obtain PPP by businesses led to a quick depletion of the initial funding. There was a
gap of 11 days in April where no loans were approved, as the second round of funding was approved for the
program. Businesses that submitted applications during the first tranche but were not approved until the
second round of funding was approved, may have made payroll decisions in anticipation of receiving PPP a
number of days before PPP approval. Furthermore, the dates that are used for this analysis are the date of
PPP approval - which may come a number of days after application submission. In both scenarios, business
12 To avoid collinearity, January 2020 is the reference group in the estimates.
13 Time-travel machines are not publicly available.
14 Additional assumptions are i.i.d. data and common support among the control variables and receipt of PPP.
14
behavior may change prior to actually getting approval. Lastly, because QCEW employment is reported as
a pay period including the 12th of the month, it is possible for PPP approval to occur after the 12th but
employment in that pay period is still directly impacted by PPP approval.15 For these reasons, anticipation,
represented by δ, needs to be considered in this analysis. I assume one month of anticipation, or δ = 1.
CS show that the AT Tp,t for each period can be semi-parametrically estimated from the product of a
propensity score matching weight predicting the probability an establishment is in treatment group p and
the difference-in-difference estimate for each period t and each PPP receipt month p:
Inverse Probability Weight
z }| {
P robp,t+δ (X)(1−Dt+δ )
Outcome Regression
P P Pp
z }| {
1−P robp,t+δ (X)
AT Tp,t = E − (Yt − Yp−δ−1 − cp,t,δ (X))
, (3)
E(P P Pp ) E[ P robp,t+δ (X)(1−Dt+δ ) ]
1−P robp,t+δ (X)
where cp,t,δ (X) = E[Yt − Yp−δ−1 |X, Dt+δ + P P Pp = 0]
P robp,t+δ (X) = E(P P Pp |X, P P Pp + (1 − Dt+δ ) = 1)
The third line of Equation 3 is the propensity score prediction for receiving PPP in month p, predicted
on the group of establishments receiving PPP in month p (or, P P Pp = 1) and those that either never receive
PPP or receive PPP no earlier than period t + δ + 1 (or, Dt+δ = 0). This is also referred to as the ”inverse
probability weight” (see Abadie (2005)). The second line is the observed change from p − δ − 1 to t in
outcome Yt for the group never receiving PPP or receiving PPP no earlier than period t + δ + 1, conditional
on covariates X. The change in outcome Yt over the same period for the establishments receiving PPP
in month p minus the second line, and times the inverse proportion of establishment-months observed for
PPP
establishments receiving PPP in month p (this is E(P P Ppp ) ) would be the ”outcome regression” (Heckman
et al. (1998)). Combining the inverse probability weighting with the outcome regression gives an estimate
that is ”doubly robust” (Sant’Anna and Zhao (2020) and Sun and Abraham (2020)) in the sense that only
either the propensity score function needs to be properly specified *or* the outcome regression for the control
observations needs to be properly specified for the AT Tp,t estimate to be valid (Sant’Anna and Zhao (2020)).
Once each AT Tg,t is estimated, a weighted average of each AT Tp,t is constructed to create a set of results
reflecting an event study - AT Te - identifying the effect of PPP for each period e relative to the period of
PPP approval p, providing a set of parameters for:
T̃
X
AT Te = AT Tp,p+e E(P P Pp |p + e ≤ T, C 6= 1) (4)
e=0
where E(P P Pp |p + e ≤ T, C 6= 1) is the fraction of all PPP-approved establishments-months that were
15 Mapping the timing of PPP approval to reported employment in the QCEW and the CES can be tricky. The QCEW
employment measure is for the pay period containing the 12th of the month - a pay period could be a week, 2 weeks (ending in
the week of the 12th or the week after) or even a monthly pay period. For this reason, it is possible that the month classification
of assigning establishment i to treatment group p = t + 1 if the PPP loan was approved after the 12th of month t may not
reflect reported employment prior to PPP approval in month t. For this reason, I allow for one month of anticipation of PPP
receipt. Returning to Equation 3, this sets δ = 1, or one period of anticipation. This may also reflect the reality that there is
a lag between when the PPP application is submitted and when it is actually approved. If the establishment has taken action
with regards to employment decisions in anticipation of an approval, then there may be a treatment effect ”seeping” into the
period prior. For the purposes of estimation, all this does is change the comparison pre-treatment period to be p − 2 instead of
p − 1 when calculating each AT Tp,t .
15
approved for a PPP loan in month p, and T̃ is 7 months, the last date for which data is available for each
month of PPP receipt.16 Although balancing the analytical sample removes establishment-observations from
the estimation (for instance, all months after November 2020 are dropped for establishments receiving PPP
in April 2020), there is an upside: when comparing AT Te of different time periods, there is no concern
of a compositional effect. For example, for the estimate AT Te=11 , the only set of establishments where
there are observations 11 months post-PPP approval are those loans approved in April 2020. Thus, the
sample composition used to estimate AT Te=11 would be different than the sample composition for estimating
AT Te=1 . For this reason, the panel is balanced such that AT Te=7 is the farthest out estimate from PPP
approval that will be estimated on the full sample. Lastly, likely due to reporting errors, there are a fraction
of observations reporting extreme increases in employment. For this reason, .01% of the observations are
removed for having employment or wage changes exceeding 100 times the baseline.17
The R package publicly provided by CS is used for all displayed estimates18 . These estimates result from
using the doubly robust method (which relies on the propensity score match estimates for weighting), with
bootstrapped standard errors that are clustered at both establishment level and county level. All reported
confidence intervals are the 95% simultaneous confidence band, which are more robust to multiple hypothesis
testing. Lastly, to reduce computational burden, a 5% random sample of all establishments with 15 months
of non-missing reporting in the QCEW make up the analytical sample.
5.2 Effects of PPP on Employment
Table 3 presents the ATT effects along with 95% simultaneous confidence bands for the full event study
going from five months prior to PPP approval to 7 months after. The dependent variable is establishment
employment in month t relative to average employment in the same calendar month from 2017-2019 for that
establishment, so that the estimates can be viewed as a percentage change in employment.19 Using the same
calendar month as the baseline helps avoid issues with seasonality. Controlling for seasonality is important
in this context because Accommodation and Food Services establishments, which were targeted for PPP, are
more likely to be affected by seasonality. Additionally, bonuses are often paid at the end of the year, which
are also important to control for.20
The first column of Table 3 displays the DDID estimates with no control variables and the second column
includes the control variables.21 Comparing the columns without controls to the columns with controls,
there is an attenuation towards zero for each month where e ≥ 0 in the event study for almost every month.
Additionally, the change in estimates after adding in the controls gets larger the further from PPP approval,
potentially suggesting that long-term effects observed may be due to some selection that is captured by the
time invariant controls. For this reason, the estimates with controls are the preferred specification.
Focusing on estimates with controls, within the first month of being approved for a PPP loan, there is
an increase of between 6.7 to 8 percent in employment absent PPP loan approval. The effect slowly tapers
16 This would be November 2020 for PPP approval in April 2020, December 2020 for PPP approval in May 2020,...,March
2021 for PPP approval in August 2020.
17 Cutoffs of 5 and 10 times the baseline yield nearly identical results.
18 https://www.rdocumentation.org/packages/did/versions/2.0.0
19 Only years where the establishment exists are included in the average. For months in establishments that have an average
of zero employment in prior years, the employment percentage is imputed as 100. The number of observations where this is
imputed is 2%. The results are unaffected whether these observations are included or excluded.
20 Results using average employment for the establishment across all 2019 calendar months as the baseline are slightly larger
but follow the same patterns and statistical significance as the results shown here.
21 Appendix A.2 gives detail about the control variables chosen. Since the sample is sufficiently large, the choice of control
variables is not particularly parsimonious, which is another benefit of using the QCEW.
16
Figure 9: Average Treatment on the Treated (ATT) of PPP Loan Approval on Employment Outcomes for
Establishments near Eligibility Cutoff
Effects of Receiving a PPP Loan on
Employment Relative to Pre-Pandemic Employment Baseline
8.0% 95% Simultaneous Confidence Band
ATT of PPP Loan
Effect on Employment %
6.0%
Relative to Pre-Pandemic Employment Baseline
4.0%
2.0%
[Average Treatment on Treated]
0.0%
-2.0%
0 -5 -4
1 -3 2 -2 3 -1
4 5 6 7
Months Until/Since PPP Approval
Notes: Based on a random sample of 5% of establishments in the QCEW. Visualization of Column 2 of Table 3.
as the months progress, though have statistically significant and positive effects through seven months after
PPP loan approval. This is a particularly interesting result as it provides some evidence of the longer-term
impact of being approved for a PPP loan. As observed in Figure 4 and according to the rules governing
maximum PPP loan amount, the loan amounts were intended to be up to 10 weeks of wages. Ten weeks of
wages implies an employment effect that should completely dwindle after three months in the event study if
all wages are being covered using PPP. Many businesses simply had reduced demand without fully closing,
suggesting that PPP may cover a portion of wages over a longer time period than just 10 weeks. Also, the
effects may be longer-lasting because one of the stipulations for converting the loan to a grant are that the
employment and compensation levels pre-pandemic must be maintained for 24 weeks following PPP receipt,
through the end of the calendar year of 2020.
Figure 9 visualizes the ATT effects displayed in Column 2 of Table 3. The months prior to being
approved for a PPP loan all have estimates close to zero, and mostly statistically indistinguishable from
zero. A key test of the primary assumption of parallel trends displayed in Equation 2 is that the ATT in
months prior to being approved for PPP should be zero. In all months prior to e = 0, the estimated ATT
is indistinguishable from zero. Overall, there is no evidence of pre-treatment selection after controlling for
time invariant regressors.
Given evidence shown previously that may indicate a low match rate for PPP applications with 1 em-
ployee, there may be concern that the poor match rate for these employers is affecting the results in some
way. Column 3 in Table 3 shows the results removing all employers with 2019 monthly employment less
than 2. Going from Column 2 to Column 3, the coefficients attenuate by a few tenths of a percent but do
not meaningfully impact the results.
Overall, the results show a strong, persistent effect on employment of PPP receipt. As an additional
17
Table 3: Estimates of the Effect of PPP Approval on Employment
Dependent Variable
% of Employment Baseline
Months Until /
Removing Small
After PPP Estimate Without Controls With Controls
Employers
Loan Approval
(1) (2) (3)
ATT -0.75 -0.24 0.18
-5
[95% C.I.] [-2.9, 1.4] [-2.4, 1.9] [-1.1, 1.5]
ATT -2.48 -1.2 -1.16
-4
[95% C.I.] [-4.3, -0.6] [-3.0, 0.6] [-2.3, -0.0]
ATT -0.73 0.02 -0.03
-3
[95% C.I.] [-1.4, -0.1] [-0.7, 0.7] [-0.4, 0.3]
ATT -1.04 -0.17 0.12
-2
[95% C.I.] [-1.5, -0.6] [-0.7, 0.3] [-0.2, 0.4]
ATT -2.82 -0.12 -0.05
-1
[95% C.I.] [-3.4, -2.2] [-0.8, 0.6] [-0.6, 0.5]
ATT 4.42 6.69 5.93
0
[95% C.I.] [3.8, 5.0] [6.0, 7.4] [5.4, 6.5]
ATT 5.87 8.0 7.52
1
[95% C.I.] [5.2, 6.5] [7.2, 8.8] [6.9, 8.1]
ATT 4.37 6.16 5.79
2
[95% C.I.] [3.7, 5.0] [5.3, 7.0] [5.1, 6.5]
ATT 3.39 4.98 4.87
3
[95% C.I.] [2.7, 4.1] [3.9, 6.0] [4.1, 5.6]
ATT 3.22 4.44 4.28
4
[95% C.I.] [2.5, 3.9] [3.5, 5.4] [3.6, 5.0]
ATT 3.35 4.54 4.13
5
[95% C.I.] [2.6, 4.1] [3.5, 5.6] [3.4, 4.8]
ATT 2.9 4.1 4.14
6
[95% C.I.] [2.1, 3.7] [3.1, 5.1] [3.4, 4.9]
ATT 2.71 4.04 4.09
7
[95% C.I.] [1.9, 3.6] [3.0, 5.0] [3.3, 4.8]
Notes: These are estimates from a dynamic difference-in-difference semi-parametric estimation, based on Callaway
and Sant’anna (2020). Results displayed are aggregate group-time average treatment on the treated (ATT) effects,
showing simultaneous 95% confidence bands via bootstrapping over 1000 iterations. The standard errors are clustered
at the establishment level and the county level. Controls included are described in Appendix A.2. The unit of observa-
tion is an establishment month in the QCEW. This was estimated on 5,548,635 observations, or a 5% random sample
of private establishments in the QCEW with positive employment in 2019. The pre-pandemic employment baseline is
the average employment for that establishment in the same calendar month from the years 2017-2019.
robustness check that offers additional nuance to the results, Table 4 shows results from an analogous
analysis using monthly microdata from the CES. The CES has a panel structure that allows the use of survey
responses from prior months to be the baseline outcome to compare the month t outcome. Furthermore,
CES contains information about hours worked that can help understand the effect of PPP on employment.
The sample is restricted to respondents that give a valid response for the entire 2020 calendar year. To
avoid any compositional bias in the estimates, the ATT estimates are produced only for 4 months after PPP
approval - the last value of e where all groups p can be estimated.
For Table 4, the dependent variables are month t values relative to the reported January 2020 values.
18
Column 1 shows results very similar to Column 2 in Table 3. Each month’s 95% confidence bands overlap
despite being statistically different from zero, suggesting that the CES - an entirely different source of
establishment employment information - corroborates the results found using the QCEW. Interestingly, the
hours estimates in column 2 suggest that employers were not simply paying employees while they did not
work. To the extent that PPP may have been used to keep businesses afloat while employees stayed home
without working during the pandemic, this does not seem to be common enough to show in the average
estimates. The effect of PPP on hours is again comparable to the effect on employment. The last column
looks at the effect on the ratio of hours per employee each month relative to the reported value in January
2020. One month after PPP approval shows a statistically significant 1.3% positive effect, but that dissipates
quickly int he following months. Compared to the employment estimates, this suggests that the effect of
PPP seem to only hinge on the extensive margin of keeping employees on payrolls, and not the intensive
margin of how many hours those workers were paid for.
One thing to note about the CES results is that they are subject to nonresponse bias in a way that
the QCEW results are not. In particular, the CES is more subject to a nonresponse being the result of a
closure22 . For this reason, it may bias the results from the CES downward as some closed establishments
are not included in the analytical sample.
One final robustness check is to do a placebo test, doing the exact same analysis but for years prior to
the pandemic. Specifically, looking at establishments in existence in 2018 and estimating the employment
trajectory through March, 2019 for establishments approved for PPP in 2020. The sample is restricted to
establishments that also continue to exist in 2020, so as to not bias the estimates on PPP approval upward.
Month of PPP approval is the same calendar month as approval the establishment received in 2020.
Figure 10 shows the same estimates for a 5% sample of establishment-months from January 2018-March
2019, where the employment baseline is the same calendar month employment average from 2015-2017. The
results show small, positive coefficients with point estimates less than 1 and indistinguishable through zero
up to five months after the calendar month that they received PPP. Based on the estimates from the placebo,
any potential bias is small relative to estimates presented above, as all confidence bands exist outside of the
confidence bands shown in Table 3. This is strong evidence that the baseline results are not an artifact of
seasonality or some other patterns of recent growth correlated to PPP application and approval.
In the next section, these estimates are put into context compared to estimates from other research
projects.
5.2.1 Comparison to Previous Results
Table 5 shows estimates from previous research estimating the effects of PPP on employment and business
outcomes. The results presented in previous research so far have been varied, both in terms of results and
methodology. Nearly all of the papers have found at least some positive impact of PPP on employment,
though a number of them amount to an economically small effect. Chetty et al. (2020), Hubbard and Strain
(2020), and Granja et al. (2020) find effects on employment that do not exceed 2%. Autor et al. (2020) finds
a central effect of a little more than 3%. These results are all notably smaller than the effects identified
in the baseline results presented thus far. Faulkender et al. (2020) find a large 12% effect on employment
of increased PPP access, though that number is not an ATT effect and therefore not an apples-to-apples
comparison to the results presented.
22 Documented in Dalton et al. (2021).
19
Table 4: Estimates of the Effect of PPP Approval on Employment in the CES
Dependent Variable
Months Until / % of Hours per
% of January 2020 % of January 2020
After PPP Estimate Employee Relative
Employment Hours
Loan Approval to January 2020
(1) (2) (3)
ATT -2.7 -2.43 -1.1
-5
[95% C.I.] [-7.3, 1.9] [-13.2, 8.3] [-7.9, 5.7]
ATT -3.21 -5.54 -0.52
-4
[95% C.I.] [-6.8, 0.4] [-13.0, 1.9] [-4.6, 3.6]
ATT -2.14 -1.06 0.35
-3
[95% C.I.] [-4.8, 0.5] [-6.4, 4.3] [-2.2, 2.9]
ATT 0.73 0.17 0.04
-2
[95% C.I.] [-0.2, 1.7] [-1.2, 1.6] [-0.9, 1.0]
ATT 0.84 1.41 0.11
-1
[95% C.I.] [-0.3, 2.0] [-0.3, 3.1] [-0.9, 1.1]
ATT 5.87 6.98 0.88
0
[95% C.I.] [4.7, 7.1] [5.1, 8.9] [-0.1, 1.9]
ATT 8.11 9.85 1.31
1
[95% C.I.] [6.7, 9.5] [7.6, 12.1] [0.1, 2.5]
ATT 6.34 6.97 0.19
2
[95% C.I.] [4.9, 7.8] [4.6, 9.3] [-1.5, 1.9]
ATT 5.2 5.55 -0.24
3
[95% C.I.] [3.7, 6.7] [3.2, 7.9] [-1.7, 1.2]
ATT 4.52 5.14 0.41
4
[95% C.I.] [3.1, 5.9] [2.7, 7.6] [-1.0, 1.8]
Notes: These are estimates from a dynamic difference-in-difference regression, based on Callaway and Sant’anna
(2020). Results displayed are aggregate group-time average treatment on the treated (ATT) effects, showing simulta-
neous 95% confidence bands via bootstrapping over 1000 iterations. The standard errors are clustered at the estab-
lishment level. Controls included are described in Appendix A.2. The unit of observation is an establishment month
in the Current Employment Statistics survey (CES). The dependent variable is the establishment’s reported value in
the month relative to their report in January 2020.
Bartik et al. (2021) also find a large increase on employment, where their focus is on the delay in funding
caused in the few weeks between the first tranche of PPP in April 2020 running out due to demand and
the second tranche opening up again. The effect that they estimate is limited to a specific subset of the
establishments—those that were unable to access PPP loans in the first tranche because of the inability of
nearby banks to deal with an influx of PPP applications.
Doniger and Kay (2021) find positive effects of PPP going through September - up to five months after
the first PPP loans were disbursed, consistent with this paper’s findings on the seven post-PPP approval
months of effects as shown in Table 3.
In a recent paper, Autor et al. (2021) employ a similar methodology to this paper estimating cohort-
specific average treatment effects but using payroll processor data. The baseline estimates between the
papers are comparable: they find an overall employment effect of 6% across the size distribution, which is
very similar to the overall effect estimated here of 5.3% averaged across months-since-PPP-approval. More
comparisons will be made to this paper in later sections.
While it is difficult to reconcile all of the different results given their unique methodological approaches,
20
Figure 10: Average Treatment on the Treated (ATT) of PPP Loan Approval on Employment Outcomes in
Years Prior to Pandemic (Placebo Effect)
Effects of Receiving a PPP Loan on
Employment Relative to Pre-2018 Employment Baseline
Pre-Pandemic Placebo Test
95% Simultaneous Confidence Band
8.0% ATT of PPP Loan
Effect on Employment %
6.0%
Relative to Pre-Pandemic Employment Baseline
4.0%
2.0%
[Average Treatment on Treated] 0.0%
-2.0%
0 -5 -4
1 -3 2-2 -1
3 4 5 6 7
Months Until/Since PPP Approval
Notes: Based on a random sample of 5% of establishments from the year 2018. Same estimation as Table 3.
there are some common elements that can be examined alongside the methodology in this paper to help
understand the varying results. First, Chetty et al. (2020), Autor et al. (2020), Hubbard and Strain (2020),
and Granja et al. (2020) all rely on the variation for the largest establishments, or largest loans (which is a
proxy for establishment size), when analyzing the microdata. For Chetty et al. (2020) and Autor et al. (2020),
they explicitly rely on a discontinuity design based around the employment cutoff SBA sets for determining
whether an employer fits the definition of a ”small business” for the purposes of PPP. The benefit of this
restriction is that the eligibility criteria is set exogenously, but this necessarily means estimates are derived
from the group of establishments getting PPP just below the employment threshold compared to the group
of establishments immediately above the employment threshold. To the extent that the effect of PPP is
the same for larger employers (those near a size of 500, for instance) as it is for small employers, then this
strategy is valid. Similarly, Hubbard and Strain (2020) and Granja et al. (2020) rely on loans greater than
$150,000, which is a small subset of all of the PPP loans more likely to have been received by larger employers
as it is for small employers, then these strategies are valid.
However, if the effects vary for small and large employers, then applying the results from the large employ-
ers to all eligible establishments may be a misleading interpretation of the impact of PPP on employment.
To better understand these results, Figure 11 shows the AT Te estimates after reducing the sample to
only employers where the employment is between .5 and 1.5 times the SBA industry-specific employment
cutoff. This sample is close in spirit to the samples used for estimates in Chetty et al. (2020), Autor et al.
(2020), Hubbard and Strain (2020), and Granja et al. (2020). Furthermore, these papers tend to use a
specific month in 2020 as the employment baseline. For that reason, the employment baseline used here
is January 2020. When the sample is limited in this way, the effects on employment as seen in Figure 11
become notably similar to the results in those papers - 2% increase in employment in the third month after
21
Table 5: Estimates of the Effect of PPP on Employment from Prior Research
Effect of PPP Identification
Paper Limitations
on Employment Strategy
Chetty et al. Diff-in-Diff Focuses only on employers with 100+
2%
(2020) around cutoff employees; Specific Industries
Autor et al. Diff-in-Diff Focuses only on employers with 100+
3.25%
(2020) around cutoff employees
Autor et al. Robust No wage analysis and Limited
6%
(2021) Diff-in-Diff heterogeneity analysis with ADP data
Hubbard and Only identifies establishments with
.9% Diff-in-Diff
Strain (2020) PPP loan of $150k+
Granja et al. 1-2% decline in Instrumental Aggregated local effects; only uses
(2020) closure Variables $150k+ loans; Relies on Homebase
Uses first tranche of loans from April;
Bartik et al. Instrumental
16-35% small sample size so harder to identify
(2021) Variables
heterogeneity
+ Effects on
Instrument
Doniger and Employment Aggregated to location; no
using delay in
Kay (2021) through heterogeneity analysis besides size
first tranche
September
Faulkender Instrumental
12% Not an ATT effect
et al. (2020) Variables
4.7% increase in Control for
Bartlett and One metropolitan area; small sample;
survival for size application
Morse (2020) not using observed closures
3 success
PPP approval though the estimates have large standard errors and are not statistically different from zero.
This is in the same ballpark as the 2% estimate in Chetty et al. (2020), and smaller than the 3.3% estimate
in Autor et al. (2020) and larger than the .9% estimate in Hubbard and Strain (2020).
Overall, the sample differences help explain the inconsistent results among a number of papers in assessing
PPP. Understanding these limitations emphasizes the importance of doing a more detailed heterogeneity
analysis
5.3 Closures and Wages
Table 6 presents results examining the effect of PPP on establishment closures and wages. A closure in this
context is when an establishment reports 0 employment for the month or receives an inactive status in the
QCEW.
The following formula translates reported quarterly wages in the QCEW into monthly wages:
22
Figure 11: Average Treatment on the Treated (ATT) of PPP Loan Approval on Employment Outcomes for
Establishments near Eligibility Cutoff
Effects of Receiving a PPP Loan on
Employment Relative to 2019 Employment
Restricting Sample to Employers Near SBA Cutoff
95% Simultaneous Confidence Band
8.0% ATT of PPP Loan
Effect on Employment %
6.0%
Relative to Pre-Pandemic Employment Baseline
4.0%
2.0%
[Average Treatment on Treated] 0.0%
-2.0%
0 -5 -41 -3 2 -2 -1
3 4 5 6 7
Months Until/Since PPP Approval
Notes: Based on a random sample of 25% of establishments. Same estimation as Table 3 but with a sample reduced to the
establishments between .5 and 1.5 times the SBA industry-specific employment cutoff for PPP eligibility.
empyi,t
wageyi,t = wageyi,q ∗ P y
t0 ∈q empi,t0
where wageyi,q are the wages reported in the quarter q in year y for establishment i, proportioned to each
month in that quarter, weighted by the proportion of total employment across each of the three months of
that quarter in month t.
Similar patterns seen in Table 3, which examines the effects of PPP loans on employment, emerge for
PPP loans’ effects on closures as well—a post-PPP ATT effect that maintains statistical significance in
reducing the likelihood of closure through 7 months post-PPP. Within the first month of PPP approval, an
establishment is about 5.8% less likely to close relative to an establishment that did not receive approval
for a PPP loan. One interesting pattern is that the employment effect seems to dissipate faster than the
reduction in closures—as the employment effect goes from 8% in month 1 to 4% in month 7, closures go from
a 5.8% drop in month 1 to a 3.5% drop in month 7. These results suggest that PPP loans had a long-term
impact in keeping businesses from closing, too.
The ATT effect on wages also shows a statistically significant jump at PPP approval of 11.5% - a few
percentage points higher than the employment effect. Furthermore, the effect on wages also had a quicker
decline as the months progress, maintaining a 3.8% increase of wages 7 months after PPP approval. There
are a few possible interpretations of these results.
One key caveat to this analysis on wages is that by using reported wages for the entire establishment, this
does not control for compositional changes to employment. In other words, the translation from quarterly
23
Figure 12: Average Treatment on the Treated (ATT) of PPP Loan Approval on Business Operating Status
and Wages
Effects of Receiving a PPP Loan on
2.0%
Closure Status in Month
1.0%
95% Simultaneous Confidence Band
ATT of PPP Loan
0.0%
Effect on Probability of Closing
-1.0%
-2.0%
[Average Treatment on Treated]
-3.0%
-4.0%
-5.0%
-6.0%
-5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
Months Until/Since PPP Approval
Effects of Receiving a PPP Loan on
12.0%
Monthly Wages Relative to Pre-Pandemic Wage Baseline
10.0%
95% Simultaneous Confidence Band
ATT of PPP Loan
8.0%
Effect on Wages % 6.0%
Relative to Pre-Pandemic Wage Baseline
4.0%
[Average Treatment on Treated]
2.0%
0.0%
-2.0%
-4.0%
0 -5 1-4 -3 2 -2 3 -1 4 5 6 7
Months Until/Since PPP Approval
Notes: Visualization of estimates and 95% confidence bands shown in Table 6.
24
Table 6: Estimates of the Effect of PPP Approval on Business Operating Status and Wages
Dependent Variable
Months Until /
After PPP Estimate Y/N: Closure Status % of Wage Baseline
Loan Approval
(1) (2)
ATT -0.05 -1.95
-5
[95% C.I.] [-0.7, 0.6] [-7.1, 3.2]
ATT 1.07 -2.32
-4
[95% C.I.] [0.5, 1.7] [-4.8, 0.1]
ATT 0.01 0.12
-3
[95% C.I.] [-0.2, 0.2] [-1.0, 1.2]
ATT 0.03 -0.26
-2
[95% C.I.] [-0.1, 0.2] [-1.1, 0.5]
ATT -1.42 1.48
-1
[95% C.I.] [-1.7, -1.1] [0.3, 2.7]
ATT -5.44 9.5
0
[95% C.I.] [-5.7, -5.2] [8.2, 10.8]
ATT -5.76 11.48
1
[95% C.I.] [-6.1, -5.5] [10.0, 13.0]
ATT -4.88 7.86
2
[95% C.I.] [-5.2, -4.6] [6.3, 9.4]
ATT -4.4 5.8
3
[95% C.I.] [-4.8, -4.0] [4.0, 7.6]
ATT -4.01 5.5
4
[95% C.I.] [-4.4, -3.7] [3.8, 7.1]
ATT -3.71 4.46
5
[95% C.I.] [-4.1, -3.3] [2.8, 6.2]
ATT -3.62 3.91
6
[95% C.I.] [-4.0, -3.3] [2.2, 5.6]
ATT -3.48 3.8
7
[95% C.I.] [-3.9, -3.1] [1.9, 5.7]
Notes: These are estimates from a dynamic difference-in-difference semi-parametric estimation, based on Callaway
and Sant’anna (2020). Results displayed are aggregate group-time average treatment on the treated (ATT) effects,
showing simultaneous 95% confidence bands via bootstrapping over 1000 iterations. The standard errors are clustered
at the establishment level and the county level. Controls included are described in Appendix A.2. The unit of ob-
servation is an establishment month in the QCEW. This was estimated on 6,164,370 observations, or a 5% random
sample of private establishments in the QCEW with positive employment in 2019. The pre-pandemic wage baseline
is the average monthly wage for that establishment in the same calendar month from the years 2017-2019.
wages to monthly wages is assuming each employee-month is earning the same salary in that quarter. For
instance, if an establishment temporarily reduces staffing of low-wage employees, and then the following
month in the same quarter they are all hired back, the estimated wages for the last month may be an
overestimate because the below-average employees are returning, though they are treated as ”average wage”
employees within the establishment. The CES, which is a monthly report of wages, as opposed to the QCEW,
which requires extrapolation from a quarterly report to a monthly report, is not subject to this specific form
of error. Table 7 shows the effects of PPP on closures and wages using the CES monthly reported data
on wages paid. Column 1 in Table 7 shows similar patterns to column 2 in Table 6 and also each of the
post-PPP AT Te estimates have overlapping 95% confidence bands, suggesting no statistically significant
25
difference. This helps rule out the theory that mismeasurement of monthly wages in the QCEW is driving
the results.
Table 7: Estimates of the Effect of PPP Approval on Business Operating Status and Wages
Dependent Variable
Months Until /
% of Pay per Hour Relative to
After PPP Estimate % of January 2020 Pay
January 2020
Loan Approval
(1) (2)
ATT -5.57 0.76
-5
[95% C.I.] [-15.8, 4.7] [-3.0, 4.5]
ATT -5.4 -0.89
-4
[95% C.I.] [-12.5, 1.7] [-4.4, 2.6]
ATT -1.14 1.01
-3
[95% C.I.] [-6.9, 4.7] [-1.0, 3.0]
ATT 0.03 -0.64
-2
[95% C.I.] [-1.5, 1.5] [-1.4, 0.1]
ATT 1.28 0.25
-1
[95% C.I.] [-0.8, 3.4] [-0.8, 1.3]
ATT 7.87 1.35
0
[95% C.I.] [5.7, 10.0] [0.4, 2.3]
ATT 11.47 1.48
1
[95% C.I.] [8.9, 14.1] [0.5, 2.5]
ATT 7.98 0.96
2
[95% C.I.] [5.5, 10.5] [-0.2, 2.1]
ATT 6.02 0.41
3
[95% C.I.] [3.5, 8.6] [-0.7, 1.5]
ATT 5.55 0.8
4
[95% C.I.] [2.9, 8.2] [-0.1, 1.7]
Notes: These are estimates from a dynamic difference-in-difference regression, based on Callaway and Sant’anna
(2020). Results displayed are aggregate group-time average treatment on the treated (ATT) effects, showing simulta-
neous 95% confidence bands via bootstrapping over 1000 iterations. The standard errors are clustered at the estab-
lishment level. Controls included are described in Appendix A.2. The unit of observation is an establishment month
in the Current Employment Statistics survey (CES). The dependent variable is the establishment’s reported value in
the month relative to their report in January 2020.
After checking for the measurement issue, I considered whether this pattern might actually be a result of
how employers receiving PPP loans avoided cutting wages for their employees. The 2020 BRS showed that
11% of employers reported decreasing wages for at least some employees in response to the pandemic, which
highlights the importance of considering the PPP’s effects on wages in addition to employment.23 Column
2 of Table 7 shows the effect on pay per hour worked. All post-PPP estimates are positive, though of small
size, and with only statistically significant effects within the first month after PPP approval. A 1.5% increase
in pay per hour is observed for employers in the month following PPP approval, which suggests that there
is modest evidence that employers receiving PPP loans actually paid higher wages for retained workers.
However, it is important to note that these findings could be a result of employers explicitly retaining
or hiring back higher-wage employees and not hiring back lower-wage employees. If true, this would give
the appearance of an increase in wages even though workers are simply receiving the same wage they were
23 https://www.bls.gov/brs/, Table 6
26
receiving before. There is no information contained in any employer survey about which employees were
hired or retained. However, one of the stipulations for being eligible for converting the PPP loan to a grant
is for all employees to be hired back over the 8 or 24-week period following receipt of the PPP loan.24
There is a caveat in the forgiveness rules that states that if an employee quits or refuses to return after an
offer is made by the employer, then the employee count to determine eligibility for loan forgiveness will not
include that worker. If low-wage workers are more likely to turn down an offer from their employer to return
to work, then this may explain the pattern of a slight increase in wages per hour.
One way to test this last hypothesis is to do a heterogeneity analysis where the employers are broken
into two groups - states with high median replacement unemployment insurance rates and low median
replacement unemployment insurance states. The replacement rate varies from state-to-state in normal
times, and, even with the federally implemented Pandemic Unemployment Compensation (PUC), there
remains some variation across states in terms of the midpoint of replacement rates due to unemployment
insurance among the claimants. Ganong et al. (2020) provide estimates for the median replacement rate
across states. Figure 13 shows estimates from the same baseline estimation but separately estimated for
the 15 states with the highest replacement rates and then again for the 15 lowest rates. The test in this
case is to see if high replacement states saw a statistically significant lower impact of PPP on employment
and wages than in low replacement states. If unemployment insurance is keeping employees from accepting
offers to return to their previous jobs, there should be a divergence between the ATT effect on employment
compared to wages for the high replacement states versus the low replacement states. Although the point
estimates for the high UI replacement rate states (the second column of graphs) are consistently smaller
than the low UI replacement rate states (the first column), the confidence bands all overlap, suggesting no
statistically different estimate of effects of PPP based on state-specific UI replacement rates. This result is
consistent with results showing only a small impact of expanded UI on job finding rates of workers (Greig
et al. (2021)).
6 Estimating Employment and Wages Retained per PPP Dollars
Spent
This analysis can also be expanded to include matched loan amounts to each establishment in order to
identify how much PPP money went to each of the outcomes estimated. Table 8 provides a back-of-the-
envelope calculation for the amount of PPP money spent per job saved. Columns (1) and (2) are the AT Te
estimates displayed in Tables 3 and 6. Column (3) takes the number of employees at establishments that
matched to a PPP loan (47.6 million) and multiplies that number by the estimates in column (1). Column
(4) does the same but for total wages of the establishments matched to a PPP loan ($196 billion) times
column (2). Columns (5) and (6) are then the sum of columns 3 and 4, respectively - for a total of $20
million employee-months and $103 billion in monthly wages retained seven months after PPP approval.
Lastly, columns 7 and 8 are the amount of PPP money in the matched sample ($426 billion) divided by
columns 5 and 6, respectively. This gives an overall estimate of $20,737 of PPP loans per employee-month
retained and $4.13 of PPP loans per dollar-wage retained. The $4.13 estimate can also be inverted to say
how much of the PPP money went to wage retention: 24%. This estimate may be a lower-bound because it
does not include employee or wage retention after March 2021. However, it is close to the bounds shown by
24 The PPP loan forgiveness terms were altered which extended loan forgiveness timeline to 24 weeks from 8 weeks.
27
Figure 13: Average Treatment on the Treated (ATT) of PPP Loan Approval Employment and Wages, by
State UI Replacement Rates
Effects of Receiving a PPP Loan on Effects of Receiving a PPP Loan on
Employment Relative to Pre-Pandemic Employment Baseline Employment Relative to Pre-Pandemic Employment Baseline
10.0%
Restricting Sample to Low UI Replacement Rate States 10.0%
Restricting Sample to High UI Replacement Rate States
8.0% 95% Simultaneous Confidence Band 8.0% 95% Simultaneous Confidence Band
Effect on Employment % Effect on Employment %
ATT of PPP Loan ATT of PPP Loan
6.0% 6.0%
Relative to Pre-Pandemic Employment Baseline Relative to Pre-Pandemic Employment Baseline
4.0% 4.0%
[Average Treatment on Treated] [Average Treatment on Treated]
2.0% 2.0%
0.0% 0.0%
-2.0% -2.0%
-3 -2 -1 0 1 2 3 4 5 6 7 -3 -2 -1 0 1 2 3 4 5 6 7
Months Until/Since PPP Approval Months Until/Since PPP Approval
Effects of Receiving a PPP Loan on Effects of Receiving a PPP Loan on
Wages Relative to Pre-Pandemic Wage Baseline Wages Relative to Pre-Pandemic Wage Baseline
14.0%
Restricting Sample to Low UI Replacement Rate States 14.0%
Restricting Sample to High UI Replacement Rate States
12.0% 12.0%
95% Simultaneous Confidence Band 95% Simultaneous Confidence Band
10.0% ATT of PPP Loan 10.0% ATT of PPP Loan
Effect on Wages % Effect on Wages %
8.0% 8.0%
Relative to Pre-Pandemic Wage Baseline Relative to Pre-Pandemic Wage Baseline
6.0% 6.0%
4.0% 4.0%
[Average Treatment on Treated] [Average Treatment on Treated]
2.0% 2.0%
0.0% 0.0%
-2.0% -2.0%
-4.0% -4.0%
-3 -2 -1 0 1 2 3 4 5 6 7 -3 -2 -1 0 1 2 3 4 5 6 7
Months Until/Since PPP Approval Months Until/Since PPP Approval
Notes: Visualization of estimates and 95% confidence bands from samples restricted to UI generosity, based on reported
numbers in Ganong et al. (2020).
Autor et al. (2021), which are that 25-40% of PPP dollars went to workers’ wages.
To put the 24% into a broader economic context, it is first worth noting that, in order to be eligible
for forgiveness, only 60% of the loan amount needs to be spent on payroll costs; the rest can be spent on
recurring business expenses. Another aspect to consider is the amount that a firm saves by not dealing with
the churn associated with letting a worker go and then hiring a new worker to fill their place. Boushey
and Glynn (2012) estimate that the cost to an employer of turnover is about 20% of the salary of the
lost employee. This 20% is due to costs of recruitment, training, and loss of production from sub-optimal
employment. To the extent that this estimate holds in the context of the pandemic, this would mean $20
billion of the $102 billion in retained wages 7 months after PPP approval would have otherwise been a cost
to the employer. Another point is the financial cost of employees losing their jobs during times of high
unemployment. Davis and von Wachter (2011) find workers lose an average of $78,000 in lifetime earnings
as a result of unemployment during a mass layoff, and that amount increases with high unemployment, like
seen during 2020. Also, there are other benefits that are harder to put a dollar amount on - the benefit
to communities from keeping businesses from permanently closing, the societal costs avoided by keeping
workers from potentially experiencing long-term unemployment, and the government-cost savings from the
employees not moving to unemployment insurance.
Lastly, since there is an anticipation effect estimated for establishments prior to PPP approval, observed
in the AT T−1 estimate in Column 1 of Table 6, there are potentially some retained wages not included in
this analysis, suggesting an even smaller number in column 8 of Table 8 if these retained wages, plus those
unobserved in the data occurring after March 2021, were added to the aggregated retention effects. For that
28
Table 8: Value of PPP Loans per Employee Retained
AT Te Retained due to PPP
e Months after Employee Monthly Wages
Employment % Wages %
PPP Approval Months ($)
From Column From Column
(3) (4)
(2) in Table 3 (2) in Table 6
AT T0 6.69 9.5 3,181,668 18,631,381,561
AT T1 8.0 11.48 3,806,132 22,519,287,722
AT T2 6.16 7.86 2,929,664 15,412,910,998
AT T3 4.98 5.8 2,370,402 11,389,037,993
AT T4 4.44 5.5 2,111,974 10,785,133,311
AT T5 4.54 4.46 2,161,518 8,756,225,479
AT T6 4.1 3.91 1,950,064 7,678,496,949
AT T7 4.04 3.8 1,925,030 7,457,182,953
Total Retained due to PPP $ of PPP Loans per...
Employee-
Employee Monthly Wages Dollar-Wage
Month
Months ($) Retained
Retained
(5) (6) (7) (8)
20,436,452 102,629,656,965 $20,737 $4.13
Notes: AT Te come from Tables 3 and 6. The employee months saved is the product of the number of employees as of
2019 in establishments that received PPP approval (47.6 million) times the AT Te for employment %. The monthly
wages retained is the product of the total monthly wages in 2019 at establishments that received PPP approval ($196
billion) times the the AT Te for wages %.
reason, the benefits per PPP loan may be even higher.
Prior research has estimated the following— Bartik et al. (2021) estimates between $32,000 and $67,000
per job retained, Doniger and Kay (2021) estimates $100,000 per job retained, Autor et al. (2020) estimates
$224,000 per job retained, and Chetty et al. (2020) estimates $377,000 per job retained. Autor et al. (2021)
use a similar methodology to estimate a cost between $170,000 and $258,000 per job-year saved, but our
estimate of $20,700 per employee-month saved or $249,000 per job-year saved is below their upper bound.
Table 8 aggregates multiple months so that these are not specifically jobs retained but employee-months
retained, which will be larger as it double counts jobs.
6.1 Heterogeneous Effects of PPP
Using the QCEW as the basis of analysis allows us to look at a variety of employer characteristics like wage
class, establishment size, and age. The following tables are based on the identical analysis used in Table
3 but estimating each set of parameters separately for each subset of establishments, based on different
characteristics. Then analogous results to Columns 5 through 8 in Table 8 are provided for each subset of
the sample. The full set of AT Te for each grouping can be provided upon request.
One caveat with the following analysis is that the numbers reported are average effects and not marginal
effects. While using the average as an approximation for marginal effects is still useful to understand
heterogeneity across this variety of employer characteristics, it is important to keep that distinction in mind
when interpreting these results.
29
6.1.1 Wages
One documented fact over the pandemic is that lower-wage workers experienced a larger employment loss
and a slower recovery compared to high-wage workers (Chetty et al. (2020), Crane et al. (2021), and Dalton
et al. (2021)). Table 9 reports estimates that breakdown the PPP impact by wage class, where wage class is
based on the establishment-wide average wage paid in 2019. On an employee-month retention basis, there
were $6,700 in PPP loans per employee-month retained for the lowest wage class, compared to $55,000
for the highest wage class. This shows a strong positive gradient moving from lowest wage to highest
wage establishments. The lowest wage establishments also had the lowest dollar-wage retained estimate
at $4.41, with the $60-80,000 wage category showing the largest estimate of $7.37. Totaling the sum of
employee-months and wage-dollars retained using each group’s AT Te estimates, gives an estimate of an
average PPP-dollars per employee-month retained of $21,000 and per dollar-wage retained of $5.41.
Table 9: Value of PPP Loans per Employee Retained
$ of PPP Loans per...
Percent of
Employee-Month Dollar-Wage Percent of PPP Employment in
Wage Class
Retained Retained Going to Group Group Receiving
PPP
(1) (2) (3) (4)
< 20k $6,714 $4.41 7.23% 40%
20-40k $13,689 $4.61 23% 42%
40-60k $27,200 $4.94 28% 42%
60-80k $42,385 $7.37 18% 35%
80+k $55,076 $6.38 22% 24%
$ of PPP Loans per...
Employee-Month Retained Dollar-Wage Retained
(5) (6)
$21,434 $5.41
Notes: AT Te come from estimates analogous to Tables 3 and 6 but on subsets of the sample based on wage char-
acteristics. The employee months saved is the product of the number of employees as of 2019 in establishments that
received PPP approval for each group times the AT Te for employment %. The monthly wages retained is the product
of the total monthly wages in 2019 at establishments that received PPP approval for each group times the the AT Te
for wages %. This analysis is based on a 10% random sample of private establishments in the QCEW with positive
employment in 2019.
6.1.2 Establishment Size
Another notable pattern that emerged over the pandemic is that small establishments struggled more than
larger establishments through 2020 (Cajner et al. (2020) and Dalton et al. (2020)). An important caveat to
this is that the smallest establishments (fewer than 10 employees) had a steep drop in employment in April
2020 but then bounced back rather quickly compared to the other size classes.25 One potential explanation
is that PPP was particularly effective for the smallest employers. In fact, that is exactly what we see in Table
10, which shows a low ratio of PPP dollars spent on the smallest establishments relative to the number of
25 See Dalton et al. (2020).
30
employee-months retained by employers. This also holds in the dollar-wage retained measure, too. Figures
14 and 15 display the AT Te estimates for each size category. This offers one explanation for why the smallest
size class rebounded so much quicker than the other size classifications - PPP had a drastic and immediate
impact for small businesses, with an 13.3% increase in wages within a month of PPP approval. This is in
contrast to the much smaller effects on wages for the 100+ size group. These results help better understand
some of the unique patterns observed early in the pandemic: small businesses suffered but PPP had an
important impact for the smallest employers that allowed them to maintain employment and bounce back
quicker. Combining the estimates gives an estimate of about $34,000 in PPP loans per employee-month
retained, a higher estimate than the baseline estimate in Table 9.
Table 10: Value of PPP Loans per Employee Retained
$ of PPP Loans per...
Percent of
Establishment Size Employee-Month Dollar-Wage Percent of PPP $ Employment in
Class Employment Retained Retained Going to Group Group Receiving
PPP
(1) (2) (3) (4)
1-10 Employment $21,474 $4.2 23% 48%
10-50 Employment $27,014 $5.86 40% 53%
50-100 Employment $57,534 $12.97 16% 44%
100+ Employment $290,822 $697 19% 19%
$ of PPP Loans per...
Employee-Month Retained Dollar-Wage Retained
(5) (6)
$33,847 $7.22
Notes: AT Te come from estimates analogous to Tables 3 and 6 but on subsets of the sample based on establishment
size characteristics. The employee months saved is the product of the number of employees as of 2019 in establishments
that received PPP approval for each group times the AT Te for employment %. The monthly wages retained is the
product of the total monthly wages in 2019 at establishments that received PPP approval for each group times the the
AT Te for wages %. This analysis is based on a 25% random sample of private establishments in the QCEW with
positive employment in 2019.
6.1.3 Age
Haltiwanger et al. (2013) emphasize that firm age, more than firm size, is the key determinant for firm
growth. In light of that, Table 11 presents results by establishment age. The gradient is not as obvious as
it is by establishment size, but younger establishments are associated with fewer PPP dollars per employee-
month retained compared to the oldest establishments, though higer PPP dollars per dollar-wage retained.
35% of PPP money went to the oldest set of establishments, despite composing only 18% of establishments
and 30% of employment. Even after accounting for age heterogeneity, PPP-dollars per employee-month and
dollar-wage retained are similar than the baseline estimates observed in Table 3.
31
Figure 14: Average Treatment on the Treated (ATT) of PPP Loan Approval on Employment, by Establish-
ment Size
Effects of Receiving a PPP Loan on Effects of Receiving a PPP Loan on
Employment Relative to Pre-Pandemic Employment Baseline Employment Relative to Pre-Pandemic Employment Baseline
by Establishment Size Class : 1-10 Employment by Establishment Size Class : 10-50 Employment
10.0% 10.0%
8.0% 95% Simultaneous Confidence Band 8.0% 95% Simultaneous Confidence Band
Effect on Employment % Effect on Employment %
ATT of PPP Loan ATT of PPP Loan
6.0% 6.0%
4.0% 4.0%
Relative to Pre-Pandemic Employment Baseline Relative to Pre-Pandemic Employment Baseline
2.0% 2.0%
[Average Treatment on Treated] [Average Treatment on Treated]
0.0% 0.0%
-2.0% -2.0%
-4.0% -4.0%
-6.0% -6.0%
-3 -2 -1 0 1 2 3 4 5 6 7 -3 -2 -1 0 1 2 3 4 5 6 7
Months Until/Since PPP Approval Months Until/Since PPP Approval
Effects of Receiving a PPP Loan on Effects of Receiving a PPP Loan on
Employment Relative to Pre-Pandemic Employment Baseline Employment Relative to Pre-Pandemic Employment Baseline
by Establishment Size Class : 50-100 Employment by Establishment Size Class : 100 + Employment
10.0% 10.0%
8.0% 95% Simultaneous Confidence Band 8.0% 95% Simultaneous Confidence Band
Effect on Employment % Effect on Employment %
ATT of PPP Loan ATT of PPP Loan
6.0% 6.0%
4.0% 4.0%
Relative to Pre-Pandemic Employment Baseline Relative to Pre-Pandemic Employment Baseline
2.0% 2.0%
[Average Treatment on Treated] [Average Treatment on Treated]
0.0% 0.0%
-2.0% -2.0%
-4.0% -4.0%
-6.0% -6.0%
-3 -2 -1 0 1 2 3 4 5 6 7 -3 -2 -1 0 1 2 3 4 5 6 7
Months Until/Since PPP Approval Months Until/Since PPP Approval
Notes: These are the AT Te estimates used in Table 10. They are analogous to the DDID estimates constructed for Table 3
but estimated separately for each subset of the sample.
Figure 15: Average Treatment on the Treated (ATT) of PPP Loan Approval on Wages, by Establishment
Size
Effects of Receiving a PPP Loan on Effects of Receiving a PPP Loan on
Wages Relative to Pre-Pandemic Wage Baseline Wages Relative to Pre-Pandemic Wage Baseline
15.0%
by Establishment Size Class : 1-10 Employment 15.0%
by Establishment Size Class : 10-50 Employment
12.5% 12.5%
95% Simultaneous Confidence Band 95% Simultaneous Confidence Band
ATT of PPP Loan ATT of PPP Loan
Effect on Wages % Effect on Wages %
10.0% 10.0%
7.5% 7.5%
Relative to Pre-Pandemic Wage Baseline 5.0% Relative to Pre-Pandemic Wage Baseline 5.0%
[Average Treatment on Treated] [Average Treatment on Treated]
2.5% 2.5%
0.0% 0.0%
-2.5% -2.5%
-3 -2 -1 0 1 2 3 4 5 6 7 -3 -2 -1 0 1 2 3 4 5 6 7
Months Until/Since PPP Approval Months Until/Since PPP Approval
Effects of Receiving a PPP Loan on Effects of Receiving a PPP Loan on
Wages Relative to Pre-Pandemic Wage Baseline Wages Relative to Pre-Pandemic Wage Baseline
15.0%
by Establishment Size Class : 50-100 Employment 15.0%
by Establishment Size Class : 100 + Employment
12.5% 12.5%
95% Simultaneous Confidence Band 95% Simultaneous Confidence Band
ATT of PPP Loan ATT of PPP Loan
Effect on Wages % Effect on Wages %
10.0% 10.0%
7.5% 7.5%
Relative to Pre-Pandemic Wage Baseline 5.0% Relative to Pre-Pandemic Wage Baseline 5.0%
[Average Treatment on Treated] [Average Treatment on Treated]
2.5% 2.5%
0.0% 0.0%
-2.5% -2.5%
-3 -2 -1 0 1 2 3 4 5 6 7 -3 -2 -1 0 1 2 3 4 5 6 7
Months Until/Since PPP Approval Months Until/Since PPP Approval
Notes: These are the AT Te estimates used in Table 10. They are analogous to the DDID estimates constructed for Table 3
but estimated separately for each subset of the sample.
32
Table 11: Value of PPP Loans per Employee Retained
$ of PPP Loans per...
Percent of
Age of Employee-Month Dollar-Wage Percent of PPP $ Employment in
Establishment Retained Retained Going to Group Group Receiving
PPP
(1) (2) (3) (4)
< 6 Years old $18,894 $5.43 22% 34%
6-11 years old $19,514 $3.83 16% 38%
11-21 years old $18,063 $3.32 26% 37%
21+ years old $25,062 $4.48 35% 39%
$ of PPP Loans per...
Employee-Month Retained Dollar-Wage Retained
(5) (6)
$20,546 $4.14
Notes: AT Te come from estimates analogous to Tables 3 and 6 but on subsets of the sample based on establishment
age characteristics. The employee months saved is the product of the number of employees as of 2019 in establishments
that received PPP approval for each group times the AT Te for employment %. The monthly wages retained is the
product of the total monthly wages in 2019 at establishments that received PPP approval for each group times the
the AT Te for wages %. This analysis is based on a 5% random sample of private establishments in the QCEW with
positive employment in 2019.
6.1.4 Poverty
Table 12 looks at establishments based on their geography. In this case, every establishment is matched to
the poverty rate from 2015-2019 in its county to see if PPP had differential impacts26 . On wages retained,
establishments in the highest poverty areas had higher estimates of AT Te compared to the lower poverty
counties, which explains the higher PPP dollars per dollar-wage retained for the highest poverty counties.
The combined PPP-dollar per employee-month and dollar-wage retained estimates are very similar to the
baseline estimates.
6.1.5 Telework
Lastly, Table 13 matches the subsample of establishments that are in the 2017-2019 OEWS to get occupation
composition within the establishment. Specifically, every detailed occupation is matched to the binary
telework classification from Dingel and Neiman (2020). From there, the percentage of employment within
the establishment is determined to be ”teleworkable” or not. The results show an interesting pattern where
establishments with the least employment in teleworkable occupations have about 2/3rds the cost of PPP
per employee-month retained compared to the most employment in teleworkable occupations, but also has
the highest PPP cost per dollar-wage retained. These conflicting results are likely related to the somewhat
conflicting results seen in wage class estimates, where the highest-wage workers had a lower PPP dollars per
dollar-wage retained compared to the other two highest wage classes. This is relevant here since high-wage
occupations tend to be more likely to be teleworkable.
26 Based on the 5-Year American Community Survey.
33
Table 12: Value of PPP Loans per Employee Retained
$ of PPP Loans per...
Percent of
Poverty Rate in Employee-Month Dollar-Wage Percent of PPP $ Employment in
County Retained Retained Going to Group Group Receiving
PPP
(1) (2) (3) (4)
< 10% Poverty $19,611 $4.53 27% 38%
10-15% Poverty $20,813 $4.12 42% 37%
> 15% Poverty $18,793 $3.43 27% 37%
$ of PPP Loans per...
Employee-Month Retained Dollar-Wage Retained
(5) (6)
$19,867 $4.00
Notes: AT Te come from estimates analogous to Tables 3 and 6 but on subsets of the sample based on poverty in the
county of the establishment, determined from the 2015-2019 American Community Survey. The employee months
saved is the product of the number of employees as of 2019 in establishments that received PPP approval for each
group times the AT Te for employment %. The monthly wages retained is the product of the total monthly wages in
2019 at establishments that received PPP approval for each group times the the AT Te for wages %. This analysis is
based on a 5% random sample of private establishments in the QCEW with positive employment in 2019.
Table 13: Value of PPP Loans per Employee Retained
$ of PPP Loans per...
Percent of
Employee- Percent of PPP
% of Establishment in Dollar-Wage Employment in
Month $ Going to
Teleworkable Occupations Retained Group
Retained Group
Receiving PPP
(1) (2) (3) (4)
< 25% of Estab. in
$24,055 $5.51 51% 28%
Telework Occupations
25-75% of Estab. in
$37,830 $5.81 29% 30%
Telework Occupations
75 + % of Estab. in
$33,442 $4.45 19% 25%
Telework Occupations
$ of PPP Loans per...
Employee-Month Retained Dollar-Wage Retained
(5) (6)
$28,676 $5.35
Notes: AT Te come from estimates analogous to Tables 3 and 6 but on subsets of the sample based on occupation
composition information merged from the 2017-2019 OEWS microdata. The employee months saved is the product of
the number of employees as of 2019 in establishments that received PPP approval for each group times the AT Te for
employment %. The monthly wages retained is the product of the total monthly wages in 2019 at establishments that
received PPP approval for each group times the the AT Te for wages %.
34
7 Conclusion
Estimating impacts of PPP on employers via a dynamic difference-in-difference regression using adminis-
trative and survey data from the Bureau of Labor Statistics linking microdata directly to over 3 million
PPP loans, this paper is able to more thoroughly establish the positive effects of the PPP on employment,
closures, and wages while reconciling the results from prior research. I show an 8% increase in employment,
11.5% increase in wages, and a 5.8% decline in the likelihood of closure within one month of PPP approval.
7 months after PPP approval, I find a 4% impact on employment, a 3.8% effect on wages, and a 3.5% decline
in the probability of closure. Doing back-of-the-envelope calculations, I find a range of $20,000 to $34,000 of
PPP spent per employee-month retained, with about 24% of the PPP money going towards wage retention
in the baseline model. Seven months after PPP approval, establishments are estimated to have retained $102
billion in wages as of March, 2021.
I am also able to separate out the effects for larger employers relative to the full sample, which explains a
number of the conflicting results shown in the literature. Using the QCEW allows for analysis of the effects
of PPP across a variety of employer characteristics like establishment size and wage. I find that low-paying
employers had both lower PPP spent per dollar-wage retained and a smaller percentage of PPP dollars
received relative to high-wage employers. The smallest establishments and establishments in geographies
with higher poverty also had the most employee-months and dollars of wages retained per PPP dollar spent.
Further study will look at longer-term impacts of the PPP both for employers and for the communities
that received PPP. Additionally, study of the establishments that had their loans converted to grants should
be considered in future work.
35
References
Abadie, Alberto (2005), “Semiparametric difference-in-differences estimators.” The Review of Economic
Studies, 72, 1–19.
Autor, D., D. Cho, L. Crane, M. Goldar, B. Lutz, J. Montes, W. Peterman, D. Ratner, D. Villar, and
A. Yildirmaz (2020), “An evaluation of the paycheck protection program using administrative payroll
microdata.” MIT Working Paper.
Autor, D., D. Cho, L. Crane, M. Goldar, B. Lutz, J. Montes, W. Peterman, D. Ratner, D. Villar, and
A. Yildirmaz (2021), “The $800 billion paycheck protection program: Where did the money go and why
did it go there?” MIT Working Paper.
Bartik, A. W., Z. B. Cullen, E. L. Glaeser, M. Luca, C. T. Stanton, and A. Sunderam (2021), “The targeting
and impact of paycheck protection program loans to small businesses (no. w27623).” National Bureau of
Economic Research.
Bartlett, R. P. and A. Morse (2020), “Small business survival capabilities and policy effectiveness: Evidence
from oakland.” National Bureau of Economic Research, no. w27629.
Boushey, H and SJ Glynn (2012), “There are significant business costs to replacing employees. washington,
dc: Center for american progress.”
Cajner, Tomaz, Leland D. Crane, Ryan A. Decker, John Grigsby, Adrian Hamins-Puertolas, Erik Hurst,
Christopher Kurz, and Ahu Yildirmaz (2020), “The U.s. labor market during the beginning of the pandemic
recession.” Working paper, June 14.
Callaway, B. and P. H. Sant’Anna (2020), “Difference-in-differences with multiple time periods.” Journal of
Econometrics.
Chetty, Raj, John N Friedman, Nathaniel Hendren, Michael Stepner, and The Opportunity Insights Team
(2020), “How did covid-19 and stabilization policies affect spending and employment? a new real-time
economic tracker based on private sector data.” National Bureau of Economic Research Cambridge, MA.
Crane, L., R. Decker, A. Flaaen, A. Hamins-Puertolas, and C. J. Kurz (2021), “Business exit during the
covid-19 pandemic: Non-traditional measures in historical context (no. 2020-089r1).” Board of Governors
of the Federal Reserve System (US.
Dalton, M., J. A. Groen, M. A. Loewenstein, D. S. Piccone, and A. E. Polivka (2021), “The k-shaped
recovery: Examining the diverging fortunes of workers in the recovery from the covid-19 pandemic using
business and household survey microdata.” The Journal of Economic Inequality, 1–24.
Dalton, Michael, Elizabeth Weber Handwerker, and Mark A. Loewenstein (2020), “An update on employment
changes by employer size during the covid19 pandemic: A look at the current employment statistics
microdata.” BLS Working Papers (532).
Davis, Steven J. and Till von Wachter (2011), “Recessions and the costs of job loss.” Brookings Papers on
Economic Activity :, 1–73.
Dingel, Jonathan I. and Brent Neiman (2020), “How many jobs can be done at home?” Journal of Public
Economics, 189.
Doniger, C. and B. Kay (2021), “Ten days late and billions of dollars short: The employment effects of delays
in paycheck protection program financing.” Federal Reserve Working Paper.
Faulkender, R., R. Jackman, and S. Miran (2020), “The job preservation effects of paycheck protection
36
program loans.” US Department of the Treasury, Office of Economic Policy.
Ganong, Peter, Pascal Noel, and Joseph Vavra (2020), “Us unemployment insurance replacement rates
during the pandemic.” Working paper, May 15.
Goodman-Bacon, Andrew (2021), “Difference-in-differences with variation in treatment timing.” Journal of
Econometrics.
Granja, J., C. Makridis, C. Yannelis, and E. Zwick (2020), “Did the paycheck protection program hit the
target?” National Bureau of Economic Research, (No. w27095).
Greig, Fiona, Daniel M Sullivan, Peter Ganong, Pascal Noel, and Joseph Vavra (2021), “When unemployment
insurance benefits are rolled back: Impacts on job finding and the recipients of the pandemic unemployment
assistance program.” Available at SSRN 3896667.
Haltiwanger, John, Ron S. Jarmin, and Javier Miranda (2013), “Who creates jobs? small versus large versus
young.” Review of Economics and Statistics, 95, 347–361.
Heckman, J. J., H. Ichimura, J. Smith, and P. Todd (1998), “Characterizing selection bias using experimental
data.” Econometrica, 66, 1017–1098.
Hubbard, R. G. and M. R. Strain (2020), “Has the paycheck protection program succeeded? (no. w28032).”
National Bureau of Economic Research.
Sant’Anna, P. H. and J. Zhao (2020), “Doubly robust difference-in-differences estimators.” Journal of Econo-
metrics, 219, 101–122.
Sun, L. and S. Abraham (2020), “Estimating dynamic treatment effects in event studies with heterogeneous
treatment effects.” Journal of Econometrics.
A Appendix
A.1 Record Linking
Linking the QCEW to other surveys is trivial, since the QCEW is the sample frame for BLS establishment
surveys. However, there is no direct way to link with other data sources like the PPP loan microdata.
However, in the absence of unique identifiers or other administrative codes, common data between the two
databases can be used for matching. The PPP Data includes employer name, address, reported NAICS code.
All of the PPP information comes directly from the loan application. For a variety of reasons, there
may be inconsistencies between what is input on the loan application for an establishment and what would
show up in the QCEW data. For instance, the person filling out the application may use an abbreviation for
the employer that is not used in the trade or legal name in the QCEW. The applicant may provide a local
contact address instead of the physical location of the establishment, which may be more common during the
pandemic as people were less likely to come directly into the office. For these reasons, it is not immediately
straight-forward to match the PPP application data to the QCEW. For this, more advanced record-linking
techniques are necessary.
All addresses in the QCEW and PPP are standardized using the Postal Service’s geocoding tool that takes
address inputs and returns the nearest address in the full index of current addresses. Next, the employer
name provided on the PPP application and both the legal and trade names are standardized by removing
punctuation and capitalization, and replacing all references to states with the appropriate state abbreviation.
37
Also uninformative words are removed, such as ”INC” or ”CORP” and city names repeated in the business
name.
After these standardization rules are applied, a fuzzy text matching between QCEW and PPP business
names are applied. The TF-IDF, or Term Frequency (TF) – Inverse Document Frequency (IDF), text
matching method is used on business name matching between PPP and the QCEW. TF-IDF is both faster
than other common methods, which is especially important in a Big Data matching space, and more flexible.
To apply TF-IDF, business names were broken up into n-grams, or n-letter strings, where n=3. These n-
grams are also referred to as tokens and are useful in breaking up text strings into component pieces that
can be compared between records. TF-IDF calculates how common or unique each of these n-grams are.
Then TF-IDF scores are computed such that term frequency (TF) is the fraction of n-grams in employer
name that are in specific 3-letter string, IDF (inverse document frequency) is the natural log of the inverse
fraction of employer names containing the specific string, and the final score is TF time IDF.
The TF-IDF score is for each n-gram, and thus creates a vector of scores across the universe of n-grams
in the name space. Every employer name has its own vector, and can be compared against the vector of
another employer name by calculating the cosine of the angle between the two vectors. This number is the
cosine similarity index and allows for a simple pair-wise comparison across all potential employer names. A
cosine similarity index of 1 means an exact match, with 0 being the lowest score.
To minimize Type I errors and also reduce the state space in order to make the matching process more
efficient computationally, the match is done in 3 stages. The three stages follow concentric circles of geography
- city, county, and then state. At each stage, the state space is only employers that are a) within the same
geography and b) have not previously been matched to a PPP loan. When a sufficient match is found, both
the PPP loan and the QCEW establishment are removed from the state space. As the linking proceeds
to larger geographies, the minimum threshold for a sufficient match increases. For the match at the state
level, only exact matches are accepted. One other piece is do the match for matching sector - if the PPP
establishment and the QCEW establishment have the same sector, a small ”bonus” match is added to the
cosine similarity score to indicate a higher quality match.
A.2 Time invariant Control Variables
The doubly robust difference-in-difference estimates allow for controlling time invariant characteristics that
may predict an establishment choosing to participate in the PPP. These controls are sets of dummy variables
for sector, state, urban classification, whether the establishment is part of a franchise, bins for distance from
nearest bank, bins of age of establishment, bins for average wage of establishment, bins for EIN size, bins
for establishment size, bins for local industry employment concentration, whether the establishment was
eligible for PPP, whether the establishment grew or declined between 2018 and 2019, 1 dummy for each
quarter of 2019 determining whether the wage’s for the quarter were lager than the full-year average for
the establishment, 1 dummy for each month of 2019 for if the establishment was closed, and 1 dummy each
for if the establishment was matched to an Economic Injury Disaster Loan (EIDL) or grant from the EIDL
program. Lastly, a continuous measure of 2018 to 2019 growth for the establishment is included as a control
as well.
38
File and source
- File
- putting-the-ppp-into-perspective-administrative-and-survey-data.pdf
- Size
- 842,590 bytes
- SHA-256
- 21965b827bb2cc860e8748125f2a89b8e446f8e0db79932fcc43737836935166
- Original
- www.bls.gov