Measuring Small Business Dynamics and Employment with Private-Sector Real-Time Data — IZA DP No. 15515

Summary

IZA Discussion Paper No. 15515, Measuring Small Business Dynamics and Employment with Private-Sector Real-Time Data, by André Kurmann, Etienne Lalé and Lien Ta, dated AUGUST 2022. The abstract states that the paper proposes new methods to assess the representativeness of private-sector datasets and to distinguish business openings and closings from sample churn, and applies them to Homebase. It states that the authors match Homebase establishment records with information from Safegraph, Google and Facebook, and use the data to study whether small service-sector businesses were hit harder by the pandemic than larger firms and the extent to which the Paycheck Protection Program (PPP) helped small businesses keep their workforce employed. It states that the real-time estimates are remarkably representative of administrative data.

Summary drafted by a model from the document's text below and checked by script against that text before publication. It is a navigation aid, not a reading of what the document proves. Where AI is used

Issuer: Source documents
Document type: Measuring Small Business Dynamics And Employment With Private Sector R
Case: Measuring Small Business Dynamics And Employment With Private Sector R

Cite this document

“Measuring Small Business Dynamics and Employment with Private-Sector Real-Time Data — IZA DP No. 15515.” Original at www.iza.org. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/measuring-small-business-dynamics-and-employment-with-private-sector-r/, SHA-256 c66cea1c956e979422e916ebc7ee0904adc277df7a5b0ab067f0684f8c5b2567.

Full text

DISCUSSION PAPER SERIES

IZA DP No. 15515

Measuring Small Business Dynamics
and Employment with Private-Sector
Real-Time Data

André Kurmann
Etienne Lalé
Lien Ta

AUGUST 2022
DISCUSSION PAPER SERIES

IZA DP No. 15515

Measuring Small Business Dynamics
and Employment with Private-Sector
Real-Time Data

André Kurmann
Drexel University
Etienne Lalé
Université du Québec à Montréal and IZA
Lien Ta
Drexel University

AUGUST 2022

Any opinions expressed in this paper are those of the author(s) and not those of IZA. Research published in this series may
include views on policy, but IZA takes no institutional policy positions. The IZA research network is committed to the IZA
Guiding Principles of Research Integrity.
The IZA Institute of Labor Economics is an independent economic research institute that conducts research in labor economics
and offers evidence-based policy advice on labor market issues. Supported by the Deutsche Post Foundation, IZA runs the
world’s largest network of economists, whose research aims to provide answers to the global labor market challenges of our
time. Our key objective is to build bridges between academic research, policymakers and society.
IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of such a paper
should account for its provisional character. A revised version may be available directly from the author.

ISSN: 2365-9793

IZA – Institute of Labor Economics
Schaumburg-Lippe-Straße 5–9 Phone: +49-228-3894-0
53113 Bonn, Germany Email: publications@iza.org www.iza.org
IZA DP No. 15515 AUGUST 2022

ABSTRACT
Measuring Small Business Dynamics
and Employment with Private-Sector
Real-Time Data*
The COVID-19 pandemic has led to an explosion of research using private-sector datasets to measure
business dynamics and employment in real-time. Yet questions remain about the representativeness
of these datasets and how to distinguish business openings and closings from sample churn –
i.e., sample entry of already operating businesses and sample exits of businesses that continue
operating. This paper proposes new methods to address these issues and applies them to the case
of Homebase, a real-time dataset of mostly small service-sector sector businesses that has been
used extensively in the literature to study the effects of the pandemic. We match the Homebase
establishment records with information on business activity from Safegraph, Google, and Facebook
to assess the representativeness of the data and to estimate the probability of business closings and
openings among sample exits and entries. We then exploit the high frequency / geographic detail
of the data to study whether small service-sector businesses have been hit harder by the pandemic
than larger firms, and the extent to which the Paycheck Protection Program (PPP) helped small
businesses keep their workforce employed. We find that our real-time estimates of small business
dynamics and employment during the pandemic are remarkably representative and closely fit
population counterparts from administrative data that have recently become available. Distinguishing
business closings and openings from sample churn is critical for these results. We also find that while
employment by small businesses contracted more severely in the beginning of the pandemic than
employment of larger businesses, it also recovered more strongly thereafter. In turn, our estimates
suggests that the rapid rollout of PPP loans significantly mitigated the negative employment effects
of the pandemic. Business closings and openings are a key driver for both results, thus underlining
the importance of properly correcting for sample churn.

JEL Classification: E01, E24, E32, E60
Keywords: small business activity, sample turnover versus business
openings/closings, matching records, COVID-19, Paycheck
Protection Program

Corresponding author:
Etienne Lalé
Université du Québec à Montréal
Department of Economics
C.P. 8888, Succ. centre ville
Montréal (QC), H3C 3P8
Canada
E-mail: lale.etienne@uqam.ca

* First version: April 28, 2020. This draft replaces earlier versions entitled “The Impact of COVID-19 on Small Business Dynamics and
Employment: Real-time Estimates with Homebase Data”. We thank Homebase, Safegraph and CrowdTangle for generously making
their data available. We also thank John Coglianese, Ryan Decker, Cynthia Doniger, John Friedman, Chad Higdon-Topaz, Bart Hobjin,
Erik Hurst and Jay Stewart as well as seminar participants at Arizona State University, the Deutsche Bundesbank, Drexel, McGill,
Placekey, the St. Louis Fed, UQA`M, Wharton, the IZA workshop on labor market dynamics, the North American Meetings of the
Econometric Society, the NBER Summer Institute, and FESAC for comments. All errors are our own.
1 Introduction

The COVID-19 pandemic unfolded with tremendous speed and affected labor markets in unprecedented
ways. This led to an explosion of research with new private-sector establishment-level datasets to measure
the state of the economy. The primary advantage of these datasets is that they provide information in real
or near-real time and at considerably greater detail than what is typically available from official statistics.1
At the same time, questions remain about the representativeness of these datasets and how to distinguish
business openings and closings from sample churn – i.e., sample entries of already operating businesses and
sample exits of businesses that continue to operate. These issues are of particular importance because,
contrary to previous recessions, the disruptions brought about by COVID-19 were concentrated in service
sectors where establishments with fewer than 50 employees account for about half of all jobs, and it is
exactly among these small businesses where opening and closing rates are high even under the best of
circumstances. Hence, the quality of real-time estimates of business dynamics and employment during
the pandemic depends crucially on whether the dataset in question contains small businesses that are
representative of the underlying population and on whether the estimates accurately incorporate the
effects of small business openings and closings.
In this paper, we propose new methods to distinguish business openings and closings from sample
churn, correct for sample selection issues, and assess the representativeness of the resulting estimates.
We apply our methods to the case of Homebase, a scheduling and time clock software provider used by
about 100,000 businesses in the U.S. that shares daily data on hours worked and hourly pay for each
employee of their clients. The Homebase data is particularly well-suited for two reasons. First, the
majority of the businesses in the dataset are small and operate in the service sectors most exposed to the
pandemic. Second, Homebase has experienced strong growth and client turnover both before and during
the pandemic. It therefore provides an excellent case to assess the efficacy of our methods.
The main methodological contribution of the paper relative to the many other studies using Homebase
or similar establishment-level datasets is that we match the Homebase establishment records by name
and address with information on business activity from Safegraph, Google, and Facebook. We use the
match with Safegraph, a database of over 7 million places of interest including many small service-sector
businesses, to assess the representativeness of the Homebase sample by comparing visit patterns from
anonymized cell-phone records to businesses in Homebase with visit patterns to the universe of corre-
1
As described in detail below, statistical agencies in the U.S. publish monthly survey estimates for the labor market,
but these estimates become available only with a lag of several weeks and contain limited breakdowns by industry and
establishment size. Administrative data on the population of establishments and employment are available only at quarterly
or annual frequency and are released several months to several years later. The situation in other countries is similar.

1
sponding small businesses in Safegraph. The match with Google / Facebook, in turn, allows us to estimate
the probability of business closings and openings among sample exits and entries. We then adjust for
sample selection by benchmarking the quarterly birth and death rates implied by these estimates against
pre-pandemic population counterparts from the Bureau of Labor Statistics’ (BLS) Business Employment
Dynamics (BED) and the U.S. Census Bureau’s Business Dynamics Statistics (BDS). Finally, we use the
adjusted estimates to construct weekly time series of small business employment that directly incorporate
gains from business openings and losses from business closings.
To assess the quality of our methods, we compare our estimates during the pandemic with population
counterparts from the BED and the Quarterly Census of Employment and Wages (QCEW) that have
recently become available. We find that our estimates provide a close fit. In addition, given the sparse
frequency of these official data sources, we show that average weekly visits to our Homebase establish-
ments move more or less in lockstep with average weekly visits to all corresponding small businesses in
Safegraph. Together, the results provide in our view compelling evidence that our Homebase estimates
are representative of small business dynamics during the pandemic.
We show that properly distinguishing closings and openings from sample churn is crucial for these
results. Counterfactuals that include either none of the entries and exits into Homebase or that treat
all entries and exits as openings and closings would produce wildly different estimates. As such, the
paper offers a cautionary tale about the increasing use of private-sector datasets to construct estimates
of employment and business activity. These datasets are opportunity samples subject to client turnover
that can result in large, time-varying gaps between sample entry and exit and true business openings and
closings. While Homebase may represent an extreme example in this regard, the issue is also important
for other establishment-level datasets, including the Current Employment Statistics (CES) that forms
the basis of the BLS’ monthly payroll estimate.2
In the second part of the paper, we exploit the high frequency / geographic detail of the Homebase data
to study two important questions: have small service-sector businesses been hit harder by the pandemic
than larger firms; and to what extent did the Paycheck Protection Program (PPP), a major provision
of the 2020 CARES Act that provided small businesses with financial support, help small business keep
their workforce employed.
The analysis yields four key insights. First, small business employment in Retail Trade, Education &

2
As discussed in more detail below, the employment estimates from the CES historically did not take into account
the effects of openings and closings directly and instead estimated employment changes from business birth/death with an
econometric model applied to past data. Faced with extraordinary numbers of business closings in the beginning of the
pandemic, the BLS modified this procedure in April 2020.

2
Health, Leisure & Hospitality and Other Services – the four service sectors hit hardest by the pandemic
– contracted by an estimated 14 million between mid-February and mid-April 2020, a staggering 45%
decline, and then regained about 9 million by mid-June 2020. Between mid-June 2020 and late-November
2021 (the end of the sample under consideration), small business employment gradually recovered most
of the losses, and for Retail Trade and Leisure & Hospitality even exceeded their pre-pandemic levels.
The decline and subsequent rebound in the beginning of the pandemic are both larger than the CES
estimates for total employment in the four sectors. This implies, consistent with Cajner et al. (2020)
and Dalton et al. (2020), that small businesses initially contracted more severely but then also recovered
more strongly than larger businesses. Our estimates show that the faster recovery of small business
employment continued all through 2021, which provides interesting new evidence for the ongoing debate
on whether small businesses are more sensitive to economic shocks than larger businesses (e.g., Moscarini
and Postel-Vinay, 2012; Fort et al., 2013; or Haltiwanger et al., 2018).
Second, we show that temporary closings account for 70% of the initial decline in small business
employment, with total closings spiking to 40% in mid-April 2020. In the months thereafter, about two
thirds of closed businesses reopened, resulting in a cumulative closing rate in the four sectors of 17% one
year after the start of the pandemic. This is only about two percentage points higher than the cumulative
closing rate over the same time period one year prior, which implies, perhaps surprisingly but consistent
with concurrent analysis by Crane et al. (2022) based on alternative measures of business closures as well
as Decker and Haltiwanger (2022) and Fairlie et al. (2022) based on more recently released administrative
data, that the pandemic has not led to a substantially higher rate of permanent shutdowns.3
Third, new businesses openings in the four sectors considered have added more than 1.5 million new
jobs since the pandemic started, constituting an important driver of the recovery from mid-June 2020
onward. Compared to 2019, the rate of new business openings remains lower, however. This differs
from recent evidence by Fazio et al. (2021) and Haltiwanger (2021) who report record numbers of new
business applications since the pandemic started. The difference is likely due to the fact that it takes
several quarters from business application to employment of workers, and that new applications are
disproportionally for businesses without a physical store location (e.g. nonstore retailers) which are
excluded from our data. Our estimates may therefore represent a lower bound of the recovery in small
business employment.
Fourth, we estimate that counties with a larger share of delayed PPP loans experience a persistently

3
Other studies that document a large but temporary spike in small business closures in the beginning of the pandemic
are Chetty et al. (2020) based on Womply revenue data, Yelp (2020) based on Yelp data, and Vaan et al. (2021) based on
Safegraph visit data.

3
lower recovery of small business employment. The estimates are based on a county-by-week panel built
from our Homebase sample that, following the research design proposed by Doniger and Kay (2021),
allows us to exploit plausibly exogenous local differences in the delay of obtaining a PPP loan due to
the temporary exhaustion of PPP funding in mid-April 2020. The negative employment effect, which is
conditional on a rich set of controls including establishment fixed effects and is not driven by pretrends,
is in large part due to higher rates of business closings.4
The result underlines the importance of properly taking into account closing and opening margins,
and suggests that the temporary exhaustion of PPP in mid-April 2020 occurred at a critical moment
when many small business owners, faced with an unprecedented negative shock, had to decide whether to
continue operating and hope for government support or cut their losses and close shop. Quantitatively,
our estimates imply that without delays in PPP loans, small business employment would have been about
10% higher in mid-July 2020 and about 6% higher in January 2021. Despite the many valid critiques
about PPP in terms of inadequate targeting and potential for fraud, this suggests that the rapid and
extensive deployment of economic support in response to the pandemic was an important reason for the
relatively strong recovery in small business activity.
Related literature. The paper contributes to an extensive literature using non-traditional establishment-
or household-level data sources to measure the economic impact of the COVID-19 pandemic; e.g., for the
U.S., Bartik et al. (2020), Bick and Blandin (2020), Cajner et al. (2020), Chetty et al. (2020), Coibon
et al. (2020), Crane et al. (2022), Kahn et al. (2020), Lewis et al. (2021), or the Household Pulse Survey
by the U.S. Census Bureau, among many others. A related set of studies focuses explicitly on employment
changes of smaller businesses; e.g. Bartik et al. (2020), Dalton et al. (2020), Fairlie (2020), Fairlie et al.
(2022), or the Small Business Pulse Survey by the U.S. Census Bureau.
Aside from the high-frequency estimates of the effects of the pandemic on small business activity
that we provide, the novelty of the paper is that we systematically match establishment records with
alternative information on business activity to assess the representativeness of the data, estimate business
closings and openings, and integrate the resulting effects into a consistent measure of small business
employment. This is critical to accurately measure small business employment throughout the pandemic
and to quantitatively evaluate the role played by business closings and openings. While the Homebase
data we use has its limitations in terms of coverage, its focus on small businesses in service sectors hit
hardest by the pandemic and the considerable growth and turnover of its client base provide us with an

4
The regressions incorporate controls for changes in county-level COVID infection and death rates, non-pharmaceutical
interventions (NPIs), school closings and weather, as well as differential week fixed effects by pre-pandemic county average
household income.

4
excellent stress test of our methods. In this sense, the paper can also be seen as a proof-of-concept.
Our paper also contributes to the vigorous debate on the effects of PPP. Autor et al. (2020), Chetty
et al. (2020), and Hubbard and Strain (2020) exploit the 500 employee threshold for PPP loan eligibility
and find only limited effects. Concurrent studies by Bartik et al. (2021), Bartlett and Morse (2020),
Doniger and Kay (2021), and Granja et al. (2020) based on alternative identification techniques generally
find larger effects, especially for smaller businesses. Consistent with Chodorow-Reich et al. (2020), this
suggests that small businesses are generally more financially constrained and were therefore more depen-
dent on PPP loan support. Our estimates provide further support for this conclusion and are broadly in
line with the main results of Doniger and Kay (2021), whose research design we adopt. Importantly, our
analysis provides an explanation for why the temporary exhaustion of PPP loans had such long-lasting
effects: it is primarily due to the adverse effects that these delays had on business closings. This finding is
confirmed in more recent studies by Autor et al. (2022) and Dalton (2021) who compare employment and
closing probabilities of businesses who received a PPP loan earlier with those of businesses who received
a PPP loan later.
The rest of the paper proceeds as follows. Section 2 introduces our employment estimator; Section
3 describes the data; Section 4 discusses how we distinguish business openings and closings from sample
churn and provides first evidence about the representativeness of the Homebase data; Section 5 analyzes
small business dynamics during the pandemic and further assesses the representativeness of our results;
Section 6 estimates the effects of PPP loan delays; and Section 7 concludes.

2 Estimating small business dynamics and employment

This section provides a general description of our approach to construct an estimate of small business em-
ployment from a sample of establishment-level data that directly incorporates the effects of establishment
openings and closings. For each sector (e.g. Leisure & Hospitality), we start with reference employment
!0 , taken from the QCEW, and estimate employment in week t as
level E

" # $
A O
i ωi êi,ti,t + êi,ti,t
!t = E
E !t−1 × # $, (1)
" Ai,t Ci,t
ω
i i ê i,t−1 + ê i,t−1

where ωi denotes the sampling weight for cell i (in our case at the industry-size-region level), constructed
as the ratio of establishment population counts to sample counts in that cell in the reference period;
A
êi,ti,t denotes employment of the set of establishments Ai,t that are active in the sample in both week t

5
O
and t − 1; êi,ti,t denotes employment gains from the set of establishments Oi,t that are newly opening or
C
i,t
reopening in week t; and êi,t−1 denotes employment losses from the set of establishments Ci,t that are
closing temporarily or permanently in week t.
The key challenge in constructing this estimate is to distinguish business openings and closings from
sample churn; i.e. entry of businesses that already operated previously, and exit of businesses that
O
continue to operate thereafter. Including already operating entrants as part of Oi,t would mean that êi,ti,t
overestimates employment gains from openings. Vice versa, including exiters that continue to operate in
C
i,t
Ci,t would mean that êi,t−1 overestimates employment losses from closings.
Sample churn is important for many private-sector establishment-level datasets since the underlying
data providers commonly experience client turnover and/or growth in their customer base over time. As
shown in Sections 4 and 5, this means that simple strategies such as treating all entries as openings,
respectively treating all exits as closings will result in spurious estimates of establishment openings,
O
closings, and employment growth. Similarly, ignoring all entries and exits (i.e. setting êi,ti,t = 0 and
C i,t
êi,t−1 = 0) and effectively estimating total employment growth from the set of continuing establishments
Ai,t will be misleading during large business cycle swings such as the pandemic when there are important
changes in the relative rate of establishment openings and closings.
O C
We address this challenge by constructing a direct estimate of ei,ti,t and ei,t−1
i,t
that exploits information
on individual business activity from alternative sources. Conceptually, we estimate employment gains
from establishment openings as

O
%
O
êi,ti,t = p̂(O!,t |entry!,t ) × e!,t × θ!,t , (2)
!∈i

where p̂(O!,t |entry!,t ) denotes the estimated probability of establishment # opening conditional on enter-
O is a possibly
ing the sample in week t; e!,t measures employment of establishment # at entry; and θ!,t
establishment- and time-specific adjustment factor such that the resulting birth rate is consistent with
pre-pandemic population counterparts from official statistics. Similarly, we estimate employment losses
from establishment closings as

C
%
i,t C
êi,t−1 = p̂(C!,t |exit!,t ) × e!,t−1 × θ!,t , (3)
!∈i

where p̂(C!,t |exit!,t ) denotes the estimated probability of establishment # closing conditional on exiting the
C is an adjustment
sample in week t; e!,t−1 measures employment of establishment # prior to exit; and θ!,t

6
factor such that the resulting death rate is consistent with pre-pandemic population counterparts from
official statistics. As explained in detail in Section 4, these adjustment factors correct for potential sample
selection issues.
It is instructive to compare our approach with the CES employment estimator that the BLS uses for
its monthly Employment Situation.5 While our estimator in (1) is conceptually similar to the “weighted
link-relative technique” of the CES estimator, our estimator directly incorporates employment gains
from openings and employment losses from closings. The CES estimator, in contrast, includes only a
portion of the establishments that report zero employment in month t and establishments that return
to positive employment in month t, respectively, and then adjusts separately for employment effects
from new openings and other closings with a econometric “net birth/death” model based on current and
past data.6 As such, the CES estimator may not reflect the net effect of changes in business openings
and closings that occur in response to large business cycle shocks such as the pandemic whereas in our
O C
approach, p̂(O!,t |entry!,t ) and p̂(C!,t |exit!,t ) and therefore ei,ti,t and ei,t−1
i,t
adjust with real-time information
about business activity. In addition, since our approach is formulated explicitly in terms of opening and
closing probabilities, it is straightforward to benchmark the resulting birth and death rates to population
counterparts from official statistics.

3 Data

The data we use to apply our estimation approach comes from Homebase (HB), a scheduling and payroll
administration provider used primarily by small, independently owned businesses employing fewer than
50 workers. The majority of the businesses operate in service sectors with a large propensity for in-person
interaction that were most exposed to the disruptions and stay-at-home orders in the beginning of the
pandemic. As such, the HB data is particularly well-suited to assess our approach of estimating small
businesses dynamics in response to large shocks.
The HB data consists of anonymized daily records of individual hours worked and wages of employees,
linked longitudinally to the establishment where they work and the firm that owns the establishment
(almost all businesses are single-establishment firms). The data is recorded in real-time through HB’s

5
See https://www.bls.gov/web/empsit/cestn.htm for details on the CES and estimation.
6
Historically, the CES estimator only included establishments that reported positive employment in both t and t − 1 and
adjusted for the net effect of births versus deaths with an econometric model based on past data. By doing so, employment
growth of non-respondents was effectively assumed to equal employment growth of respondents. In light of the large labor
market disruptions caused by the COVID-19 pandemic, the BLS changed its methodology from April 2020 onward by directly
including a portion of reported zeros in the estimator and by adding current period employment growth in the econometric
model to predict the net birth/death effect. See https://www.bls.gov/web/empsit/cesbd.htm for details.

7
proprietary software and is used by many of the businesses for payroll processing. HB provides free data
access to researchers and updates the data frequently with the latest observations.
In addition to the publicly available data, HB shares with us counts of owners and managers that use
the HB software, scheduled hours of employees when available, and name and address information for
each establishment. The information on owners and managers and scheduled hours allows us to construct
a broader measure of employment. In turn, we use the name and address information to match HB
establishments to data from Safegraph, Google, and Facebook. This is a key advance over other studies
using the HB data that allows us to attribute a consistent industry classifier to each establishment;
assess the representativeness of the HB sample; and, most importantly, distinguish business openings and
closings from sample churn. The procedure involves extensive data cleaning and standardization before
matching the records sequentially by exact merges and then fuzzy name match and substring match
algorithms. The Appendix provides details on the different steps as well as match statistics. We only
retain HB establishment records that match exactly or with a high match rate.7

3.1 Employment and business activity

For each establishment, we construct weekly employment as the sum of individuals with tracked hours
(actual or scheduled) during that week plus owners and managers that show activity in the HB software
but do not have tracked hours.8 Including owners and managers broadens the employment concept
beyond hourly paid workers, which attenuates somewhat the decline in estimated employment at the
onset of the pandemic when owners and managers with untracked hours had a higher propensity to
remain active than employees with tracked hours. But otherwise, the results are not affected materially
by their inclusion.
For an establishment to be retained in our sample it must show up at least once for a spell of
three consecutive weeks with at least 40 weekly tracked hours across its employees. We thus exclude
establishments that use HB only for a short trial period. For an establishment in the sample to be
active in a given week, it must have employees with tracked hours in that week. Establishment activity is
therefore independent of owners and managers logging in to the HB software (e.g. for reporting purposes).

7
We compare our match algorithm to Safegraph’s Placekey matching tool and find that our algorithm results in higher
match rates, primarily thanks to extensive pre-cleaning of establishment names and deduplication of establishment records.
Details are available upon request.
8
For establishments that report both scheduled and actual hours, we compare the two measures and find them to be
very close to each other. We are therefore confident that scheduled hours are an accurate measure of actual hours worked.

8
3.2 Industry classification

The historical HB data comes with an industry category for each establishment, but the available cate-
gories do not line up with standard industry classification, and for about one third of the records, industry
category is missing altogether. This is an important limitation for the purpose of constructing estimates
that can be compared to official statistics. We address this issue by using the above-described match of
HB establishments with Safegraph’s Core Places data, which contains consistent NAICS-6 industry cod-
ing for over 7 million Places of Interest (POIs) in the U.S., including a large fraction of all private-sector
establishments.9

3.3 Sample characteristics and representativeness

The sample we consider covers January 2019 through November 2021. The beginning of the sample is
imposed by the availability of sizable HB data.10 The end of the sample is due to upload limitations
to relate HB establishments to Facebook postings.11 The raw HB data contains about 300,000 distinct
establishments; however, many of them do not use HB regularly and therefore do not satisfy our retention
criterion. The sectors with the largest coverage are Leisure & Hospitality (NAICS 71 and 72), Retail
Trade (NAICS 44-45), Education & Health Services (NAICS 61-62), and Other Services (NAICS 81).12
To ensure good coverage, we focus our analysis on these four sectors and exclude all establishments with
50 employees or more.13 According to official statistics, establishments with fewer than 50 employees
accounted for about half of all jobs across the four sectors and for about 23% of all private-sector jobs
prior to the pandemic. Hence, the segment covered by our analysis represents a sizable share of to-
tal employment and, as highlighted above, consists primarily of businesses with a significant in-person
component that were among the hardest hit by the pandemic.

9
See the Appendix for details on the Safegraph data and summary statistics about the NAICS industry codes for the
matched HB establishments. In December 2020, HB independently started attributing NAICS industry codes for each
establishment in their dataset. This classification is available only for establishments active from that month onward. Since
many establishments that were active in 2019 and 2020 exited the HB sample before December 2020, the HB NAICS codes
are not directly useful for our analysis. However, for establishments active in December 2020 and beyond, we find a high
level of overlap between our NAICS codes and the HB NAICS codes.
10
HB data is available starting in January 2018, but the sample size is relatively small until 2019.
11
CrowdTangle, Facebook’s research database, imposes limits on how many records (establishments) can be uploaded
and matched. We have periodically renegotiated additional upload credits and plan to do so again in the future.
12
See the Appendix for details. Other Services includes “Repair & Maintenance” (NAICS 811) and “Personal & Laundry
Services” (NAICS 812), which contains many of the HB establishments categorized under “home and repair”, “beauty and
personal care”, and “health care and fitness”. Aside from these four sectors, the HB data also contains several hundred
establishments each in “Utilities” (NAICS 22), “Construction” (NAICS 23), “Food, Textile & Apparel Manufacturing”
(NAICS 31) and “Real Estate, Rental & Leasing” (NAICS 53).
13
We also exclude “Non-store Retail” (NAICS 454) and “Private Households” (NAICS 814) because HB contains only
very few establishments in these industries, and the QCEW that we use to benchmark does not contain these industries.

9
Table 1 reports the number of establishments that satisfy our retention criterion and that we can
match with a high degree of confidence to Safegraph. Across the entire January 2019 to November
2021 period, the sample contains about 100,000 distinct establishments. The mid-February 2019 base
sample consists of 38,193 establishments that show activity between the beginning of January 2019 and
the second week of February 2019 of which 34,757 are active in the second week of February.14 For the
mid-February 2020 base sample, the corresponding establishment counts are 49,268 and 45,454. From
mid-February 2019 to mid-February 2020, there are 13,289 exits without return and 25,149 new entrants,
and from mid-February 2020 to late-November 2021, there are 28,256 exits and 37,777 new entrants.
Foreshadowing the discussion below, these entry and exit rates are much larger than birth and death
rates in the official statistics, implying that the HB data is subject to considerable sample churn.

Table 1: Establishment counts of retained Homebase sample

Feb. 2019 - Feb. 20 Feb. 2020 - Nov. 21
Mid-February base sample 38,193 (100%) 49,268 (100%)
- active in mid-February 34,757 (91.0%) 45,454 (92.3%)
- temporarily inactive in mid-February 3,436 (9.0%) 3,814 (7.7%)
Exits without return 13,289 (34.8%) 28,256 (57.4%)
New entrants 25,149 (65.8%) 37,777 (76.6%)

Notes: The first column shows counts of HB establishments from mid-February 2019 to mid-February 2020, and the second column
shows counts of HB establishments from mid-February 2020 to the end of November 2021 that (i) were successfully matched to Safegraph;
(ii) belong to either Retail Trade, Education & Health Services, Leisure & Hospitality, or Other Services; and (iii) have fewer than 50
workers when active in mid-February or when entering Homebase.

It is instructive to compare the size of our HB sample to the number of small businesses in the four
sectors sampled by the CES. While the BLS does not publish breakdowns by industry and size, we know
that in 2021 the CES sampled about 45,000 businesses in the four sectors considered, and the share of
businesses with fewer than 50 employees sampled across all sectors was 60%.15 Assuming that this share
is similar across industries, the CES therefore sampled about 27,000 small businesses in the four sectors
considered. By this calculation, our HB sample is considerably larger.
A potentially important issue with opportunity samples such as HB concerns the extent to which the
establishments covered are representative of the population, which in our case consists of the universe
of small businesses in the four sectors considered. First, note that the estimator in (1) corrects for
distributional differences across industry-size-region cells i by weighing each cell by the ratio of QCEW
14
The remaining 3,436 establishments are temporarily inactive; i.e active prior to mid-February and then active again at
some point after mid-February. This is consistent with administrative data from the BED that also reports a substantial
rate of temporary closings prior to the pandemic.
15
The CES survey includes about 130,000 businesses (UI accounts), which cover approximately 670,000 individual work-
sites or establishments. Hence, many of the businesses sampled are larger, multi-establishments firms.

10
to HB establishment counts in the mid-February 2020 base sample so that the relative importance of
each cell in the estimator is the same as in the QCEW.16 Hence, for representativeness to be an issue,
it would need to be the case that HB establishments are systematically different from their population
counterparts within the different cells.
To provide a first assessment of this concern, we compare the pre-pandemic average number of employ-
ees per establishment in each industry-size cell with the QCEW counterparts. As shown in the Appendix,
for all the cells, the HB and QCEW numbers are very close to each other, implying that at least prior to
the pandemic, the HB establishments look highly representative within industry-size cells.
In Sections 4 and 5, we will further assess the representativeness of the HB data by comparing annual
growth rates of small business employment implied by our estimates with population counterparts from
the QCEW. In addition, given the sparse frequency of the QCEW data, we will pursue a novel approach
that exploits our match of HB establishments with Safegraph POIs, which allows us to compare weekly
visit patterns to the establishments in our sample with visit patterns to the entire Safegraph sample of
small businesses in the four sectors considered. For both of these comparisons, we find a very close fit.

4 Distinguishing business closings and openings from sample churn

This section describes the main methodological contribution of the paper, which consists of (i) estimating
the probability p̂(C!,t |exit!,t ) that an establishment # exiting the HB sample in week t is a closing, re-
spectively the probability p̂(O!,t |entry!,t ) that an establishment # entering the HB sample in week t is an
opening; and (ii) adjusting these estimates by benchmarking them against pre-pandemic birth and death
rates from official statistics. We then assess the quality of our methodology by comparing the resulting
small business employment estimates for the pre-pandemic period with population counterparts.

4.1 Identifying business closings and openings

To identify business closings in week t, we take all exiters defined as establishments that are active in
HB in week t − 1 and become inactive in week t and proceed in four steps. First, we check whether an
exiter returns to activity in HB by the end of the sample. If so, we classify the exiter as a temporary
closing and assign probability p̂(C!,t |exit!,t ) = 1. Second, for exiters that do not return to activity by the

16
The Appendix provides details on distributional differences. While the HB sample has generally good coverage across
all industry-size class cells, it under-represents the smallest establishment size class (1-4 employees) and over-represents the
other size classes (5-9, 10-19, and 20-49 employees). Furthermore, the HB sample over-represents NAICS 72 (Accommodation
& Food Services) at the expense of NAICS 62 (Health Care & Social Assistance) and NAICS 81 (Other Services), and the
geographic distribution skews slightly towards Florida and Texas at the expense of California.

11
end of the sample, we match them to Google Places using Google’s API service and assign probability
p̂(C!,t |exit!,t ) = 1 if Google tags them as “temporarily closed” or “permanently closed”.17 These tags are
reported by business owners and customers but cover only a subset of all closed establishments. Hence,
as a third step, we match the remaining exiters to Facebook using CrowdTangle, Facebook’s research
database, and check whether establishments with regular posting histories while being active in HB stop
posting regularly after exiting HB. If so, we assign probability p̂(C!,t |exit!,t ) = 1. Fourth and finally, for
exiters that we cannot match to either Google or Facebook or that do not post regularly on Facebook while
being active in HB, we assign probability p̂(C!,t |exit!,t ) equal to the industry-size cell average probability
obtained in step three.
To identify business openings, we adopt a similar method except that Google Places does not contain
a “new opening” tag. We take all entrants defined as establishments that are inactive in week t − 1 and
become active in week t and proceed in three steps. First, we check whether an entrant was active in
HB at some point before week t − 1. If so, we classify the entrant as a reopening and assign probability
p̂(O!,t |entry!,t ) = 1. Second, for entrants that become active in HB for the first time in week t, we
match them to Facebook and check whether they start posting regularly only after entering HB. If so,
we assign probability p̂(O!,t |entry!,t ) = 1. Third, for the entrants that we cannot match to Facebook or
that do not post regularly after becoming active in HB, we assign probability p̂(O!,t |entry!,t ) equal to the
industry-size cell average probability obtained in the previous step.
As a basic quality check of our identification approach, we exploit the match with Safegraph and
compare visit patterns for the different types of establishments in our HB sample. First, we verify
that exits identified as business closings exhibit a large drop off in average weekly visits relative to
establishments that remain active in HB. Second, we verify that entries identified as new openings appear
in Safegraph only after entering HB. Third, we verify that exits and entries identified as sample churners
show average visit patterns similar to those of establishments that remain active in HB. These results,
which are available in the Appendix, provide independent support for our approach to identify business
closings and new openings.
It is worth mentioning that in principle, the Safegraph visits data could be used to identify individual
establishment openings and closings from their visit patterns. For instance, Crane et al. (2022) define a
Safegraph POI as closed if year-over-year visits decline by more than a certain threshold. However, we
found after extensive analysis that the Safegraph visits data can be very noisy at the individual estab-

17
We include both Google tags “temporarily closed” and “permanently closed” because “temporarily closed” is sometimes
updated to “permanently closed”, and because our first step only identifies temporary closings that have returned by the
end of the sample.

12
lishment level, especially for POIs in buildings with other occupants (e.g. malls, multi-story buildings)
or POIs that conduct a lot of their business by delivery. This makes reliable identification of business
closings extremely challenging. Furthermore, Safegraph regularly backfills historical visit data back to
2017, independent of whether a POI operated during the entire period or newly opened at some point
after 2017.18 This means that the visits data cannot be used to identify business openings. We therefore
use information from Google and Facebook to identify openings and closings and exploit the Safegraph
visits data only to assess the representativeness of aggregates for which measurement issues are likely to
average out.

4.2 Benchmarking against administrative data
C and θ O
The second step of our methodological contribution consists of calculating adjustment factors θ!,t !,t

such that the birth and death rates implied by our estimates are consistent with pre-pandemic population
counterparts from official statistics. These adjustment factors correct for possible selection issues that
arise if the establishments exiting and entering the HB sample have a different propensity to be closings
and openings than in the population.19 Note that this adjustment procedure applies only to new openings
and permanent closings, but not to temporary closings and reopenings. The reason for this is two-
fold. First and as already mentioned above, Safegraph visits data indicate that the HB sample of both
continuously open and temporary closed and reopening establishments is highly representative of the
larger population of small businesses (see above). Second, there is no good publicly available official data
of temporary closings and reopenings at the industry-size cell level.
We use two administrative data sources covering the quasi-population of establishments to calculate
the adjustment factors. The first is the BED, which consists of all longitudinally linked establishments
from the QCEW and reports birth and death rates by sector but not by establishment size class at
quarterly frequency with a delay of about six months. The second source is the BDS, which consists
of all longitudinally linked employer establishments from U.S. tax records and reports entry and exit
rates defined similarly as BED birth and death rates by industry and size class, but only at an annual
frequency and with the latest available data for 2019.20 Since very small establishments have substantially

18
Further information on these issues with Safegraph visits data is available upon request.
19
The adjustment factors also correct for measurement error that could arise if, for example, Google closed tags are
systematically inaccurate or if businesses posts regularly on Facebook even though they are closed. We assess the potential
importance of such measurement issues by looking at Google closed tags and Facebook posting histories for businesses
identified as temporarily closed in step one and find few discrepancies. Nevertheless, measurement error remains a possibility.
20
See the Appendix for details on the definitions of birth and death rates in the BED, respectively entry and exit rates
in the BDS, and a comparison of these rates. We note that while in Retail Trade and Leisure & Hospitality, pre-pandemic
BED birth and death rates line up closely with BDS entry and exit rates, in Education & Health and Other Services, BED

13
higher birth and death rates even in normal times, and since we want to compare our estimates to
official counterparts during the pandemic, we combine the BED with the more detailed cross-sectional
information from the BDS to obtain birth and death rates by sector and size class at a quarterly frequency
that extends through the first year of the pandemic. Specifically, for each sector, we compute ratios of
entry and exit rates by size class relative to the sectoral average from the BDS and multiply these ratios
with the sectoral birth and death rates from the BED. We label the resulting BED/BDS birth and death
rates for each sector-size class i and quarter q as p(birthi,q ) and p(deathi,q ).
C = θ C for each establishment # ∈ i that permanently
For closings, we calculate an adjustment factor θ!,t i,q

exits the sample in week t ∈ q of 2019 such that the death rate for pre-pandemic quarter q in sector-size
cell i implied by the thus adjusted HB closing probabilities equals the BED/BDS counterpart p(deathi,q ).
In practice, since BED/BDS death rates vary little over 2019 and are similar for all but the smallest size
class (1-4 employees), we average over all quarters of 2019 and across sector-size cells for which the HB
sample contains relatively few permanently exiting establishments. We use these adjustment factors when
estimating the employment losses from closings during the pandemic. Hence, closing probabilities and
therefore death rates are allowed to vary in real-time during the pandemic, but the adjustment factors
that correct for selection remain fixed.
For openings, we proceed similarly by calculating adjustment factors for each establishment # ∈ i that
newly enters the sample in week t ∈ q of 2019 such that the average birth rate for 2019 in sector-size cell
i implied by the thus adjusted HB opening probabilities equals the BED/BDS counterpart p(birthi,q ).
Different from above, however, we allow these adjustment factors to be time-varying as a function of
entry A
i,t /ni,t
n
O = θO × (
the number of new entrants relative to total sample size; i.e., θ!,t )−1 , which is akin
i n
entry
/nA
i,0 i,0
to inverse probability weighting. The reason for this difference is that in contrast to exits, which are
naturally bounded by total sample size, entry can vary independently of sample size.21 Hence, opening
probabilities and adjustment factors are both allowed to vary in real-time during the pandemic.
The Appendix provides further details on the methodology and reports the different industry-size
adjustment factors. Generally, adjustment factors for the smallest size class (1-4 employees) are larger
than one, while adjustment factors for the larger size classes (5-9, 10-19, 20-49 employees) are smaller

birth and death rates are notably higher than BDS entry and exit rates. The reason for this difference is unclear and can
only be investigated with access to the underlying micro-data. This illustrates that even in official data, measuring birth
and death rates is far from trivial and depends crucially on the definition of (employer) establishments.
21
As a simple example, suppose that there are no selection issues; i.e., p̂(O!,t |entry!,t ) = p(birthi,q ) for all ! ∈ i and t ∈ q.
Now consider a large increase in sample entry (e.g. as a result of increased customer acquisition efforts by the data provider)
such that the number of new openings doubles. Then with constant adjustment factors, employment gains from openings
would spuriously double, thereby overstating estimated employment growth. By making the adjustment factor time-varying
entry
as a function of ni,t /nA
i,t , we control for this issue.

14
than one. This implies that that the HB sample of entries and exits, respectively the subset of these
establishments that we observe in Google / Facebook to estimate closing and opening probabilities, are
indeed subject to selection. The adjustment factors correct for this issue. In Section 5, we evaluate the
quality of these adjustments when we compare the birth and death rates implied by our methodology
during the pandemic with BED/BDS counterparts that have recently become available.

4.3 Pre-pandemic business dynamics and employment estimates

Figure 1 reports average quarterly rates of all new entries and permanent exits for 2019 in our HB sample,
average quarterly birth and death rates implied by our adjusted estimates of new openings and closings,
and the corresponding average quarterly birth and death rates from the BED/BDS benchmark.

Figure 1: Benchmarking with BED/BDS birth and death rates
(a) Birth rates (b) Death rates

Retail trade Education and Health Services Retail trade Education and Health Services
25 25

20 20

Quarterly rates (in %), average for 2019 Quarterly rates (in %), average for 2019
15 15

10 10

5 5

0 0

10−19 20−99 10−19 20−99 10−19 20−99 10−19 20−99
1−4 5−9 1−4 5−9 1−4 5−9 1−4 5−9

Leisure and hospitality Other services Leisure and hospitality Other services
25 25

20 20

15 15

10 10

5 5

0 0

10−19 20−99 10−19 20−99 10−19 20−99 10−19 20−99
1−4 5−9 1−4 5−9 1−4 5−9 1−4 5−9

BED birth rate HB birth rate HB total entry rate BED death rate HB death rate HB total exit rate

Notes: Quarterly birth and death rates by sector and establishment size class from BED industry data, combined with annual BDS
industry-size ratios; corresponding quarterly birth and death rates from HB; and quarterly entry and exit rates from HB. See text for
details on the computation.

By construction, the birth and death rates from our HB sample fit the BED/BDS counterparts very
closely (the fit is not perfect because, as described above, we pool over some of the sector-size classes).
Total new entry and permanent exit rates, in comparison, are much larger, as high as 25% per quarter
in the Education and Health sector. This confirms that the HB data is subject to important sample
churn: many establishments already operated prior to entry into HB, and many establishments continue
to operate after exiting HB. Finally, the figure illustrates the large differences in birth and death rates
between the smallest size class and the other size classes included in our sample. Taking into account
these differences turns out to be important for the estimation of small business dynamics and employment

15
during the pandemic.
Next, as a quality check of our methodology and to further assess the representativeness of our
HB sample, we take the adjusted opening and closing probabilities to estimate employment losses from
establishment closings and employment gains from establishment openings, equations (2) and (3), and
plug these estimates into equation (1) to compute weekly small business employment for each of the four
sectors.

Figure 2: Small business employment change compared to CES all business estimates for 2019

Retail trade Education and health services
15

Employment change (in % relative to mid−February)
10

ref. ref.
0

−5

−10
Leisure and hospitality Other services
15

ref. ref.
0

−5

−10

Feb 17 Apr 14 Sep 29 Oct 27 Apr 14 Sep 29 Oct 27
Feb 16
Mar 17 May 12
Aug 4 Sep 1 Nov 24 Dec 22 Feb 17 Mar 17 May 12
Aug 4 Sep 1 Nov 24 Dec 22 Feb 16
Jul 7 Jul 7
Jan 20 Jun 9 Jan 19 Jan 20
Jun 9 Jan 19

CES employment in all businesses QCEW small business employment
Homebase small business estimate

Notes: Employment change by small businesses with less than 50 employees and all businesses in percent of respective employment
level during the week of Feb 10 - Feb 16, 2019 for Retail Trade (NAICS 44-45), Education & Health Services (NAICS 61-62), Leisure
& Hospitality (NAICS 71-72), and Other Services (NAICS 81). None of the estimates are seasonally adjusted. The estimates for the
weeks of Thanksgiving, Christmas, and New Year are smoothed by using the estimates of adjacent weeks.

Figure 2 reports the resulting time series from mid-February 2019 to mid-February 2020 together
with the corresponding year-on-year growth rate for small business employment from the QCEW.22
22
We use mid-February 2019 as the reference week because, as described above, the QCEW publishes employment and
establishment counts by industry and establishment size class only for the first quarter of each year, with the numbers
pertaining to the month of February. We use these numbers to construct the jump-off point E0 and the cell weights ωi for
our employment estimator in (1).

16
For comparison, the figure also shows the monthly CES estimates of total sectoral employment (i.e.
employment by businesses of all size classes as opposed to just small businesses).23 Here and below, we
refrain from seasonally adjusting employment estimates since usual adjustment procedures would not be
appropriate for the type of large changes that the four sectors experienced during the pandemic. See
Rinz (2020) for a discussion.
Our HB estimates provide an excellent fit of the year-on-year employment growth rate from the QCEW
benchmark. Aside from the selection issues with regards to openings and closings for which we correct
with the adjustment factors, the HB sample is therefore representative of small business employment
growth during the pre-pandemic period.
The weekly HB estimates also comove with the monthly CES estimates, as should be expected,
although there are some seasonal differences. These differences should not come as a surprise nor do they
invalidate our approach since our HB estimates pertain to establishments with fewer than 50 employees
whereas the CES estimates pertain to total employment by establishments of all size classes. Instead,
the relevant comparison is with the small business employment benchmark from the QCEW and in this
respect, the fit is close.

5 Small business dynamics and employment during the pandemic

As a first application, we study whether small service-sector businesses have been hit harder by the
pandemic than larger businesses. The exercise also allows us to further evaluate the representativeness
of our estimates of small business dynamics and employment during the pandemic. Finally, we exploit
the richness of the HB data to compute results on average hours worked as well as gross job flows.

5.1 Employment

We start in Table 2 by reporting estimates of the change in small business employment in each of the
four sectors considered from mid-February 2020 to mid-April 2020, mid-April 2020 to mid-June 2020,
and mid-June 2020 to mid-February 2021. Across the four sectors, small business employment declined
dramatically between mid-February and mid-April 2020 as states imposed business closures and stay-at-
home orders, combining for a job loss of 14 million or 45% of the 30 million jobs prior to the pandemic.
Between mid-April and mid-June, small business employment in all four sectors then regained about 9

23
The employment estimates from the CES are publicly available only by sector and selected industries, but not by size
class.

17
million or more than two thirds of the initial job loss. Between mid-June 2020 and mid-February 2021,
small business employment further recovered although at a considerably lower pace.

Table 2: Small business employment loss and recovery during the pandemic

Retail Education Leisure & Other
Total
Trade & Health Hospitality Services
Employment in mid-February 2020 7,205 8,539 9,714 4,394 29,852

Mid-Feb to mid-April 2020 -3,019 -3,097 -5,235 -2,209 -13,558
in % relative to mid-Feb 2020 -42% -36% -54% -50% -45%
Mid-April to end-June 2020 2,365 1,752 3,560 1,401 9,078
in % relative to mid-Feb 2020 +33% +21% +37% +32% +30%
Mid-June 2020 to mid-Feb 2021 362 1,059 201 524 2,146
in % relative to mid-Feb 2020 +5% +12% +2% +12% +7%
Notes: Employment is expressed in 1,000s of jobs and pertains to establishments with fewer than 50 employees in Retail Trade (NAICS
44-45), Education & Health (NAICS 61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS 81). None of the
estimates are seasonally adjusted. Employment in mid-February 2020 is constructed as the employment estimate for all businesses from
the CES times by the ratio of employment in businesses with fewer than 50 workers to employment in all businesses from the QCEW.
The other estimates are computed with HB data using the estimator in equation (1).

Figure 3 traces out the week-by-week evolution of these estimates from the mid-February 2020 refer-
ence week to late-November 2021 (the end of the current sample) and compares them with the correspond-
ing monthly CES total employment estimates (as described above, the CES does not publish estimates
for different size classes). For reference, the figure also shows the mid-February 2020 to mid-February
2021 year-on-year growth rate of small business employment from the QCEW.
The figure visualizes the swift decline in small business employment from mid-March to mid-April
2020 as well as the ensuing rebound by mid-June 2020. Thereafter, the recovery continues at a more
gradual pace, with a temporary interruption during Winter 2020-21. By late Spring of 2021, small
business employment had recovered almost all of the losses in Education & Health and Other Services,
and exceeded its pre-pandemic level in Retail Trade and Leisure & Hospitality, although part of this
increase appears to be driven by seasonal factors (compare with Figure 2 above).24
In comparison, the initial decline and subsequent rebound in total employment from the CES is two to
three times smaller than for small business employment, with the exception of Leisure & Hospitality where
the relative magnitudes are almost as large.25 From mid-June 2020 onward, small business employment
24
Interestingly, this rise in small business employment in Retail Trade and Leisure & Hospitality accords well with the
increase in (inflation-adjusted) output above their pre-pandemic levels observed in national income accounts data. See
https://www.census.gov/econ/currentdata/.
25
Digging deeper, we find that even in retail subsectors considered essential such as Building Material Dealers (NAICS
444), Food and Beverage Stores (NAICS 445), Gasoline Stations (NAICS 447), or General Merchandise Stores (NAICS 452)
where the CES estimates show almost no job loss across all businesses, our HB estimates show large declines in small business
employment between mid-February and mid-April, followed by a large rebound. See the online Appendix for details.

18
is estimated to have regained somewhat more of its pre-pandemic level than total employment, especially
in Retail Trade and Leisure & Hospitality where the small business estimate is consistently above the
CES counterpart, ending up about 10% higher by the end of the sample in mid-November 2021.

Figure 3: Small business employment change compared to CES all business estimates

Retail trade Education and health services
20

Employment change (in % relative to mid−February)
ref. ref.
0

−10

−20

−30

−40

−50

−60
Leisure and hospitality Other services
20

10
ref. ref.
0

−10

−20

−30

−40

−50

−60

Feb 23 Aug 16 Sep 20 Oct 25 Nov 29 Feb 7 Apr 18 Aug 1 Oct 10 Feb 23 Aug 16 Sep 20 Oct 25 Feb 7 Apr 18 Aug 1 Oct 10
Nov 14
Jan 19 Mar 29 May 3 Jul 12 Mar 14 May 23 Jun 27 Mar 29 May 3 Jul 12 Nov 29 Mar 14 May 23 Jun 27 Nov 14
Jun 7 Jan 3 Sep 5 Jun 7 Jan 3 Sep 5
Jan 19

CES employment in all businesses QCEW small business employment
Homebase employment in small businesses

Notes: Employment change by small businesses with less than 50 employees and all businesses in percent of respective employment
level during the week of Feb 9 - Feb 15, 2020 for Retail Trade (NAICS 44-45), Education & Health Services (NAICS 61-62), Leisure
& Hospitality (NAICS 71-72), and Other Services (NAICS 81). None of the estimates are seasonally adjusted. The estimates for the
weeks of Thanksgiving, Christmas, and New Year are smoothed by using the estimates of adjacent weeks.

We highlight that these weekly estimates of small business employment were available in near real-
time.26 In the official data, there are no comparable numbers that are as timely and at such a high
frequency. In fact, the only publicly available official product to which we can compare our estimates are
the annual small business employment numbers from the QCEW, which are released with a lag of about
nine months. As the figure shows, our estimates align closely with the year-on-year changes implied

26
The estimation of p̂(C!,t |exit!,t ) and p̂(O!,t |entry!,t ) to calculate job loss from closings and job gains from openings
requires four weeks of Facebook data, although preliminary estimates based on extrapolated values of p̂(C!,t |exit!,t ) and
p̂(O!,t |entry!,t ) could be obtained within only a few days.

19
by these QCEW numbers, except in Leisure & Hospitality where the QCEW year-on-year change is
almost 10% above our estimate. Closer inspection of the QCEW suggests that this discrepancy is due
to the fact that in each year the QCEW reports a new cross-section of employment by establishment
size class. Since many establishments in Leisure & Hospitality were still substantially below their pre-
pandemic employment level by mid-February 2021, this means that the count of establishments in the
QCEW classified as having fewer than 50 employees increased substantially from mid-February 2020 to
mid-February 2021, thereby inflating the year-on-year growth rate of small business employment in that
sector. Our HB estimates, in contrast, follows establishments of a given size class over time and therefore
does not face this compositional issue.
The results in Figure 3 imply that aside from the initial phase, small business employment in the
four service sectors has not been hit harder by the pandemic than employment of larger businesses. The
results offer a potentially interesting new perspective for the ongoing debate on whether small businesses
are more sensitive to economic shocks than larger businesses. While Moscarini and Postel-Vinay (2012)
and Haltiwanger et al. (2018) find that this is generally not the case, Fort et al. (2013) report that
small/young business contracted more during the 2008-09 Great Recession. Our high-frequency estimates
for the pandemic suggest that this difference in results may in part be about the timing of when in a
recession employment is measured. Initially, small businesses may be affected more severely than larger
businesses, perhaps due to more severe financial constraints, but smaller businesses may also recover more
quickly, perhaps due to their ability of adapting better to a new economic environment.
At the same time, we note that the total 14 million loss in small business employment between mid-
March and mid-April 2020 and the rebound of about 10 million between mid-April and mid-June 2020
that we estimate across the four sectors are larger than the corresponding CES all business estimates
(13.5 million and about 6 million, respectively).27 Unless employment in businesses with 50 employees
or more increased from mid-March to mid-April and then declined from mid-April to mid-June – an
implausible scenario – this means that either our HB estimates or the CES estimates did not adequately
capture small business employment at the onset of the pandemic.
As we have already documented, our HB estimates closely align with year-on-year sectoral growth
rates for small business employment from the QCEW. Below, we further show that our methodology to
distinguish closings and openings from sample churn implies birth and death rates during the pandemic
that closely fit the BED/BDS population counterparts. While these are encouraging indicators of the

27
For comparison, the headline CES employment estimate for all private sectors declined by 19 million from mid-February
to mid-April on a seasonally unadjusted basis.

20
representativeness of our estimates, it could nevertheless be the case that the HB sample skews towards
small businesses that were disproportionally affected during the first few months of the pandemic. Given
that there is no publicly available administrative data at sufficiently high frequency to assess this pos-
sibility, we pursue a novel approach that exploits the match of our HB establishments with Safegraph
POIs to show that this is unlikely to be the case.

Figure 4: Homebase weekly visits compared to Safegraph POIs with less than 50 employees

Retail trade Education and health services
40

Visits change (in % relative to mid−February 2020)
ref. ref.
0

−20

−40

−60

−80
Leisure and hospitality Other services
40

ref. ref.
0

−20

−40

−60

−80
Feb 9 Mar 22 May 3 Jun 14 Jul 26 Sep 6 Oct 18 Nov 29 Jan 10 Feb 21
Apr 4
May 16 Jun 27 Aug 8 Sep 19 Oct 31 Feb 9 Mar 22 May 3 Jun 14 Jul 26 Sep 6 Oct 18 Nov 29 Jan 10 Feb 21
Apr 4
May 16 Jun 27 Aug 8 Sep 19 Oct 31

Homebase small businesses Safegraph small POIs

Notes: Weekly visits change by small businesses with less than 50 employees in Homebase (green circles) vs. Safegraph POIs with
less than 50 employees (red triangles) in percent of respective employment level during the week of Feb 9 - Feb 15, 2020 for Retail
Trade (NAICS 44-45), Education & Health Services (NAICS 61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS
81). None of the estimates are seasonally adjusted. Information on the number of employees at Safegraph POIs come from NetWise
employment data.

For many of its POIs, Safegraph reports information on the number of weekly visits and dwell time
per visit. In addition, the Safegraph data can be combined with information from NetWise, a company
collecting and analyzing vast amounts of data on U.S. businesses, to obtain an annual estimate of employ-
ees for each Safegraph POI. We use this information first to cross-check the average number of employees
for each matched HB establishment to the corresponding NetWise estimate and find a close correspon-

21
dence. Second and more importantly, we compare weekly visits of the matched HB establishments to all
Safegraph POIs in the four sectors with fewer than 50 employees.28 Figure 4 shows the results.
The fit is remarkable. Throughout the entire pandemic, weekly visits to matched HB establishments
evolve, with a few exceptions, in lockstep with weekly visits to all corresponding Safegraph POIs. As
shown in the Appendix, a similarly close fit obtains for median dwell time, share of visits lasting longer
than four hours, and weekly visits per visitor. The results indicate that the establishments in our HB
sample have on average very similar visits characteristics than the much larger Safegraph sample.29
Together with the close fit relative to QCEW year-on-year employment changes documented above, this
provides in our view compelling evidence that our HB sample is representative of small businesses in the
four service sectors considered.
This leaves two likely explanations for the larger drop in small business employment in the beginning
of the pandemic according to our HB estimates compared to total employment according to the CES
estimates. First, differences in how employment is measured; and second, differences in how business
closings and openings are identified. Employment in the CES is measured by the number of workers
receiving pay for any part of the pay period that includes the 12th of the month, independent of whether
they actually worked, while employment in HB is measured by the number of workers with positive
tracked hours plus all untracked workers who used the HB software otherwise in a given week (e.g.
owners, managers). So, if some workers who were temporarily furloughed in mid-April still received pay
even though they were no longer working, then they were counted in the CES but not in the HB data,
which would imply that the CES overestimated employment in mid-April.30 Second and as discussed in
Section 2, the CES estimator includes only a portion of the employment changes from establishments
reporting zero employment and does not directly incorporate non-responding establishments. Our HB
estimator, in contrast, directly includes the employment effects of business closings and openings. Since,
as shown below, almost 40% of all small businesses in our sample are estimated to be closed in mid-
April 2020, with about two thirds returning to activity by mid-June 2020, it is conceivable that the CES
estimator missed part of the employment effects from this large increase in temporary closings in the
beginning of the pandemic.
28
The resulting Safegraph sample of POIs in the four sectors with fewer than 50 employees and visits data contains almost
1,000,000 unique observations, or about about 22% of the corresponding universe of establishments in the QCEW.
29
In Figure 4, the Safegraph data refers to POIs that are not associated with a chains of commercial POIs (McDonald’s,
Starbucks, etc.). Results are almost identical when we include POIs that are associated with a brand.
30
One could be concerned that our HB employment measure leaves out many non-hourly workers and that these workers
were less likely to be furloughed in the beginning of the pandemic. However, when we check in the Current Population
Survey (CPS) whether employment of salaried workers declined by more than employment of hourly-paid workers, we find
only small differences. So, even if we do not capture all non-hourly workers, it seems unlikely that this would explain the
large difference with the CES employment estimates.

22
Finally, we note that our estimates of small business employment for the first few months of the
pandemic are broadly consistent with estimates in Cajner et al. (2020) based on data from ADP, the
largest payroll processing company in the U.S., and Dalton et al. (2020) based on CES microdata. As
in the CES, ADP’s employment concept is pay-based, but different from the CES, Cajner et al. (2020)
count all sample exits as closings and abstract from new openings. Similarly, Dalton et al. (2020) count
all establishments reporting zero employment as closings and imputes employment for non-reporting
establishment from employment changes in observed establishments. Since sample churn and new entry is
quantitatively small in the beginning of the pandemic, this shortcut provided a good initial approximation
although it would likely be much less accurate later in the pandemic (see below). Focusing on the results
by Cajner et al. (2020), their estimates imply that employment of all businesses in the four sectors that
we consider declined by 20.2 million between mid-February and late April. Furthermore, they report that
employment in establishments with less than 50 employees across all private sectors of the U.S. economy
declined by about twice as much between March and April 2020 as employment for larger establishments,
but by the end of June 2020 had recovered as much as larger establishments. Considering that businesses
with fewer than 50 employees accounted for almost half of employment in the four sectors prior to the
pandemic and that ADP employment is pay-based, these numbers appear quite close to our estimated
decline in small business employment of 14 million and subsequent rebound of about 10 million during
the same time period.

5.2 The crucial role of small business closings and openings

The preceding discussion highlights the importance of accurately taking into account small business
openings and closings. To assess the quality of our methodology during the pandemic, Figure 5 compares
average quarterly birth and death rates by sector and size class for 2020 as implied by our adjusted
estimates of new openings and closings to the corresponding BED/BDS birth rates that have recently
become available.31 To illustrate the importance of sample churn, the figure also shows average quarterly
rates of all new entries and permanent exits.

31
For the first quarter of 2020, the BED shows a large increase in births in Education & Health that is likely spurious,
due to a technical issue with the annual revision. Also see Decker and Haltiwanger (2022). For this sector, we therefore do
not include the first quarter birth rate for the average BED/BDS birth rate calculation.

23
Figure 5: Comparison to BED/BDS birth and death rates
(a) Birth rates (b) Death rates

Retail trade Education and Health Services Retail trade Education and Health Services
20 20

16 16

Quarterly rates (in %), average for 2020 Quarterly rates (in %), average for 2020
12 12

8 8

4 4

0 0

10−19 20−99 10−19 20−99 10−19 20−99 10−19 20−99
1−4 5−9 1−4 5−9 1−4 5−9 1−4 5−9

Leisure and hospitality Other services Leisure and hospitality Other services
20 20

16 16

12 12

8 8

4 4

0 0

10−19 20−99 10−19 20−99 10−19 20−99 10−19 20−99
1−4 5−9 1−4 5−9 1−4 5−9 1−4 5−9

BED birth rate HB birth rate HB total entry rate BED death rate HB death rate HB total exit rate

The figure shows a close match between HB birth and death rates and the BED/BDS counterparts,
indicating that our methodology of using information from Google / Facebook to identify new openings
and permanent closings in real-time and adjusting for selection with pre-pandemic information works
well.32 Given the large unexpected shock that the pandemic represents, we think this result is quite
remarkable. The figure also shows that, similar to 2019, total entry and exit rates are several times larger
than new openings and permanent closings. Correcting for sample churn is therefore as important during
the pandemic as prior to the pandemic.
To further study small business dynamics during the pandemic, we compute weekly rates of new
closings, reopenings, total closings, and new openings since the beginning of the pandemic and compare
them to the same time period one year earlier. Specifically, we compute

Ii,t "
i ωi n̂i,t
rate(It ) = " # $, (4)
Ai,1 Ci,1
i ωi n̂i,0 + n̂i,0

I
where n̂i,ti,t denotes the count of establishments in industry-size-region cell i that closed either temporarily
or permanently in week t (Ii,t = Ci,t ), reopened in week t (Ii,t = Ri,t ), or newly opened in week t
A C
(Ii,t = Bi,t ), with Oi,t = Ri,t ∪Bi,t by definition; and n̂i,0i,1 +n̂i,0i,1 denotes the count of active establishments
32
It should also be noted that the BED/BDS benchmarks are computed under the assumption that the relative birth and
death rates by size class taken from the 2019 BDS data remain constant during the pandemic. This assumption is unlikely
to hold exactly and so, the BED/BDS benchmarks are themselves only an approximation.

24
in the reference week.33

Figure 6: Small business closings, reopenings, and new openings
(a) Weekly rate of business closings (b) Weekly rate of business reopenings
18 18

16 16

Weekly closing rates Weekly reopening rates
14 14

12 12

10 10

(as % of all businesses in mid−February) (as % of all businesses in mid−February)
8 8

6 6

4 4

2 2

0 0
−2 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 −2 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52
Weeks (0 = mid−February reference week) Weeks (0 = mid−February reference week)

2020 2019 2020 2019

40 40

Share of total closed businesses Share of total new businesses
35 35

30 30

25 25

(as % of all businesses in mid−February) (as % of all businesses in mid−February)
20 20

15 15

10 10

5 5

2020 2019 2020 2019

Notes: Rates of closings, reopenings, total closings, and total new openings of small businesses with less than 50 employees in Retail
Trade (NAICS 44-45), Education & Health Services (NAICS 61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS
81). All rates are computed as a % of the total count of active businesses in mid-February. Week 0 denotes the mid-February reference
week.

As shown in panels (a) and (b) of Figure 6, the weekly rate of small business closings across the
four sectors considered spikes to 16% in the week of March 22-28, 2020 (week 6 after the mid-February
reference week) and then sharply declines to about 2% by mid-April (week 10) before further declining to

33
We define these rates relative to the count of active establishments in the reference week as opposed to the count of
active establishments around week t because the count of active establishments varies dramatically in the beginning of the
pandemic.

25
just above the pre-pandemic average of about 1% per week.34 Concurrent with the decline in the rate of
business closings in April of 2020, reopenings start to increase, reaching about 5% per week in early May
before gradually declining back to the 1.5-2% range between July and September and then the 1-1.5%
range thereafter, just slightly above the pre-pandemic rate.
Panel (c) displays the cumulative effect of these closings and reopenings on the rate of total closed
businesses relative to active businesses in the mid-February reference week. Note first that the rate of
total closed businesses averages about 6% in both mid-February 2019 and mid-February 2020, indicating
that a substantial fraction of businesses are temporarily closed at any point in time (also see Table 1).
From mid-March 2020 onward, total closings rise steeply and peak at 39% in mid-April. Thereafter, the
cumulative closing rate declines, steeply initially as reopenings rise and then more gradually to a low of
about 14% by November before rising to about 16% by mid-February 2021.35 This suggests that only
about one third of all closings in mid-March are permanent.36 Moreover, and perhaps surprisingly, the
cumulative rate of closings one year after the start of the pandemic is only about 2 percentage points
higher than the cumulative closing rate from mid-February 2019 to mid-February 2020. This implies,
perhaps surprisingly but consistent with concurrent analysis by Crane et al. (2022) based on alternative
measures of business closures as well as Decker and Haltiwanger (2022) and Fairlie et al. (2022) based
on more recently released administrative data, that the pandemic did not lead to more permanent small
business closings.
Panel (d), finally, shows total new business openings relative to total active businesses in mid-February.
This rate rises gradually throughout the year, even during the worst of the pandemic in March and April.
Compared to 2019, the pace of new openings is clearly lower during the Spring and Summer of 2020 but
then picks up somewhat in Fall and Winter, finishing at about 8% in mid-February 2021. This is only
about 2 percentage points lower than the rate of new business openings a year earlier, implying that the
pandemic only exerted a modest negative effect on new business openings.
At first sight, the lower rate of new business openings contrasts with recent evidence from the U.S.
Census Bureau that shows record rates of new applications for likely employer business relative to pre-
pandemic levels, in particular in Nonstore Retail (NAICS 454) and Personal & Laundry Services (NAICS
812), but also in Food Services & Drinking Places (NAICS 722) (see Haltiwanger, 2021; also see Fazio
et al., 2021). However, there are a number of potential explanations for this difference. First, based

34
The temporary upticks in closing rates in weeks 41 and 46 capture the weeks of Thanksgiving and Christmas.
35
These cumulative closing rates are less than half of what Crane et al. (2022) report based on HB data. The reason for
this difference is that their study treats all exits as closings.
36
The majority of establishments that closed for more than 10 weeks remain closed.

26
on historical patterns, it typically takes four to eight quarters from business registration to employment
of workers. Second, as we show below, while we see substantial employment gains from new business
openings in the Leisure & Hospitality sector (of which Food Services and Drinking Places is a large
part), new business openings play a more modest role for Retail Trade and Other Services. Since we
exclude Nonstore Retail and many of the new small businesses in Personal & Laundry Services may not
have a physical store location (and are therefore not part of our HB sample since they do not appear in
Safegraph), this implies that our estimates may represent a lower bound of the recovery in small business
employment in Retail Trade and Other Services.
To quantify the role of closings and openings for small business employment, we decompose the
employment change for each sector into the contributions from businesses that continue to operate from
mid-February until at least week t (and possibly longer), businesses that closed at some point after the
mid-February reference week but reopened by week t, businesses that operated in mid-February but are
closed in week t (temporarily or permanently), and businesses that newly opened between mid-February
and week t.37
As Figure 7 shows, business closings account for 70% or more of the initial employment decline
from mid-March to mid-April across the four sectors (red bars) while job losses by continuing businesses
account for rest (blue bars). Reopenings of closed businesses drive most of the rebound in employment
between mid-April and mid-June (smaller red bars), even though the reopened businesses operate at
lower employment than in mid-February (green bars). Finally, the recovery from mid-June onward is
driven primarily by the combination of job gains by continuing businesses (smaller negative and positive
blue bars) and new businesses (yellow bars). Overall, these new openings add almost 1.5 million new
jobs by the end of the sample.
The decomposition also reveals interesting differences across sectors. Large job losses from closings
persist through the end of the sample in Education & Health but reduce to almost zero in Retail Trade.
At the same time, Retail Trade experiences substantial job gains by continuing businesses. Finally, job
gains from new openings are more important in Leisure & Hospitality and Education & Health. These
differences suggest that the pandemic led to varying degrees of restructuring within the different sectors.

37
The Appendix provides details on the decomposition. The employment losses from closed business nets out gains
from establishments that were active in HB prior to the mid-February reference week, temporarily closed in the reference
week, and then reopened at some point thereafter (e.g. seasonal businesses; see Table 1). By netting out these gains, the
contribution from closings represents the employment losses over and above the usual employment losses from business that
temporarily close. See below for further discussion.

27
Figure 7: Contribution of closings, reopenings, and new openings to small business employment change

Retail trade Education and health services
20

−10

−20

Employment change
−30

−40

−50

−60
Nov 15 Dec 27
Nov 27
Feb 16 Mar 8 Mar 29 Apr 19 May 10 May 31 Jun 21 Jul 12 Aug 23 Sep 13 Oct 25 Dec 6 Jan 17 Feb 28 Mar 21 Apr 11 May 23 Jun 13 Jul 25 Aug 15 Sep 26 Oct 17 Nov 7 Mar 8 Mar 29 Apr 19 May 10 May 31 Jun 21 Jul 12 Aug 23 Sep 13 Oct 25 Nov 15 Dec 6 Dec 27 Jan 17 Feb 28 Mar 21 Apr 11 May 23 Jun 13 Jul 25 Aug 15 Sep 26 Oct 17 Nov 7 Nov 27
Aug 2 Oct 4 Feb 7 May 2 Jul 4 Sep 5 Aug 2 Oct 4 Feb 7 May 2 Jul 4 Sep 5
Feb 16

Leisure and hospitality Other services

(in % relative to mid−February)
20

−10

−20

−30

−40

−50

Contribution from: continuing businesses new businesses
re−opened businesses closed businesses

Notes: Contribution to percent employment change relative to mid-February 2020 in Retail Trade (NAICS 44-45), Education & Health
Services (NAICS 61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS 81) by businesses that continued operating
from mid-February until at least week t (blue bars), businesses that closed at some point after mid-February 2020 and but reopened
by week t (green bars), employment changes from businesses that operated in mid-February 2020 but are closed in week t (red bars),
and employment changes from new businesses that opened between mid-February 2020 and week t (orange bars). The estimates for
the weeks of Thanksgiving, Christmas, and New Year are smoothed by using the estimates of adjacent weeks.

5.3 The importance of controlling for sample churn revisited

To further illustrate the importance of distinguishing closings and openings from sample churn, Figure 8
reports different counterfactual employment estimates for the pandemic (similar counterfactual estimates
for the pre-pandemic period are reported in the Appendix) and compares them with our baseline small
business estimates (green circled line).
The brown short-dashed line shows the HB estimate if we abstracted completely from entry and exit
and used only the set of establishments that are continuously active in HB from the beginning through
! Ai,t
!t−1 × ! i ωi êAi,t , where Ai,t denotes continuously active establishments).
!t = E
the end of the sample (i.e., E i,t
i ωi êi,t−1
This estimate would miss much of the large decline and subsequent rebound of small business employment

28
in the initial phase of the pandemic.

Figure 8: Comparison with counterfactual employment estimators

Retail trade Education and health services
200

Employment change (in % relative to mid−February)
160
120
80
40
ref. ref.
0
−40
−80
Leisure and hospitality Other services
200
160
120
80
40
ref. ref.
0
−40
−80
Jan 19 Feb 16 Mar 15 Apr 12 May 10 Aug 30 Sep 27 Oct 25 Nov 22 Dec 20 Jan 17 Feb 14 Mar 14 Apr 11 Aug 29 Sep 26 Oct 24 Nov 21 Jan 19 Feb 16 Mar 15 Apr 12 May 10 Aug 30 Sep 27 Oct 25 Nov 22 Dec 20 Jan 17 Feb 14 Mar 14 Apr 11 Aug 29 Sep 26 Oct 24 Nov 21
Jun 7 Jul 5 Aug 2 May 9 Jun 6 Jul 4 Aug 1 Jun 7 Jul 5 Aug 2 May 9 Jun 6 Jul 4 Aug 1

Our Homebase small business estimate
Homebase always active businesses
Homebase always active + all exiting + re−opening businesses
Homebase always active + all exiting + re−opening + all entering businesses

Notes: Estimated employment change in % relative to mid-February 2020 of small businesses with less than 50 employees in Retail
Trade (NAICS 44-45), Education & Health Services (NAICS 61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS
81) according to different estimation methods (see text). The estimates are constructed based on February 2020 CES employment
estimates (week of Feb 9 – Feb 15) and QCEW shares of small business employment for the first quarter of 2020. The estimates for
the weeks of Thanksgiving, Christmas, and New Year are smoothed by using the estimates of adjacent weeks.

The red dashed-dotted line reports what
"
happens
#
if we treated all exits as either temporary or per-
! Ai,t R
ω ê +êi,ti,t
!t−1 × ! i "i Ai,t
!t = E
manent closings (i.e., E exit
# , where exiti,t denotes the set of all exiting establish-
i,t i,t
i ωi êi,t−1 +êi,t−1
ments in week t and Ri,t the set of all returning establishments in week t). Since a substantial fraction of
establishments that exit HB do not close, this estimate declines even more than our baseline estimate in
the beginning of the pandemic and recovers much less thereafter. In fact, from 2021 onward, this estimate
declines gradually as cumulative sample churn becomes increasingly important.
The orange dashed !t = E
line, finally, adds all entries and treats them as new openings (i.e., E !t−1 ×
$ %
! A entry
i ωi êi,ti,T +êi,t i,t
% , where entry
i,t denotes the set of all entering establishments in week t, including the
$
! Ai,T exiti,t
i ω i ê i,t−1 +ê i,t−1

returning establishments). The resulting estimate shows a dramatic increase in small business employment

29
through the end of the sample, far outweighing the negative effect of treating all exits as closings. This
reflects the fact that even during the pandemic, HB managed to substantially expand its client base.
The different counterfactuals offer an interesting perspective relative to other research estimating the
impact of the pandemic on small business employment with HB and other private-sector establishment
level datasets (e.g. Bartik et al., 2020; Cajner et al., 2020; Chetty et al., 2020; or Dalton et al., 2020 among
others). These studies do not systematically distinguish business closings and new openings from sample
churn and instead used either the set of continuously active businesses or the set of continuously active
businesses plus all exiting and reopening businesses or some combination thereof to estimate employment.
While the consequences of doing so with other datasets may not be as extreme as with the HB data,
the results are nevertheless affected by this choice. As such, our results offer a cautionary tale about the
use of opportunity samples to estimate small business employment series without careful incorporation
of employment gains from establishment openings, respectively employment losses from closings.

5.4 Average weekly hours

To provide further insights on the impact of the pandemic on workers employed by small businesses, we
use the HB data to report changes in average weekly hours (AWH) worked. To do so, we start with the
&H 0 , and then use our HB data to estimate
CES estimate from February 2020, AW
#" $ "
'
i ωi whi,t /( i ωi êi,t )
&H t = AW
AW &H t−1 × # $ , (5)
" ' "
i ωi whi,t−1 / ( i ωi êi,t−1 )

' i,t denotes estimated total weekly hours worked and êit denotes estimated employment at es-
where wh
tablishments in industry-size-region cell i in week t. This estimation of AWH is different from the “link-
and-taper technique” used to construct AWH in the CES, which adjusts the current estimate towards the
previous estimate so as to keep it close to the overall sample average over time. The CES estimate may
therefore not capture large changes in actual AWH that occur in times of economic disruptions, whereas
our estimate does because it is based on current information only.38
Figure 9 shows two estimates of AWH, one for all workers employed in week t, and one for all job
stayers who remained employed continuously in establishments that are active throughout the sample.

" #
38
The link-and-taper estimate used in the CES can be expressed as AW !H t = 0.9 AW!H t−1 − ! awht−1 + !awht , where
$% & $% &
!H t is the official estimate and !
AW awht = ! !
i ωi whi,t / i ωi ei,t . If AW H t−1 > awht−1 in the previous month, then the
current month official estimate will be raised relative to actual data, and vice versa if AW ! t−1 . The CES makes
!H t−1 < awh
a slight adjustment to this estimator to account for atypical reports although it is unclear what makes a report atypical.

30
While the former measure is affected by compositional change, the latter is not since it consists by
definition of a balanced panel of workers.

Figure 9: Average Weekly Hours of Small Business Employees

Retail trade Education and health services
3.0
2.0
1.0
0.0
−1.0
−1.17

Change in average weekly hours
−2.0 −1.82
−2.50 −2.37
−3.0
−4.0
−5.0
−6.0
Leisure and hospitality Other services
3.0

(relative to mid−February)
2.0
1.0
0.0
−1.0
−2.0
−3.0
−3.07−3.13
−4.0
−5.0 −4.59 −4.73
−6.0
Jan 19 Feb 23 Mar 29 Jul 12 Aug 16 Sep 20 Oct 25 Nov 29 Mar 14 Apr 18 May 23 Jun 27 Oct 10 Nov 14 Jan 19 Feb 23 Mar 29 Jul 12 Aug 16 Sep 20 Oct 25 Nov 29 Mar 14 Apr 18 May 23 Jun 27 Oct 10 Nov 14
May 3 Jun 7 Jan 3 Feb 7 Aug 1 Sep 5 May 3 Jun 7 Jan 3 Feb 7 Aug 1 Sep 5

All workers Stayers in always−active businesses

Notes: Changes in average weekly hours of employees in small businesses with less than 50 employees in Retail Trade (NAICS 44-45),
Education & Health Services (NAICS 61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS 81) relative to the
February 2020 CES estimate (week of Feb 9 – Feb 15). The solid red line shows the change in average weekly hours of all workers
employed in all small businesses. The blue shows the change in average weekly hours of job stayers in businesses that remain active
throughout the sample.

Across all four sectors, AWH declines sharply in March 2020 but then recovers quickly and exceeds the
pre-pandemic level somewhat through the end of the sample. The sharp decline in AWH in the beginning
of the pandemic is further evidence of the sudden impact of the pandemic on service-sector jobs. The
larger drop as well as the smaller increase in AWH of job stayers relative to the all workers measure is
due to the compositional change: the workers laid off or furloughed in March worked on average fewer
hours than job stayers. As employment recovered a large part of the loss, this gap then mostly closed.
Remarkably, AWH of job stayers recovers fully within a few months of the start of the pandemic.
This suggests that the labor market during the pandemic has not been characterized by a large increase

31
in involuntary part-time employment as has been the case during previous recessions (e.g. Borowczyk-
Martins and Lalé, 2019). This may be due to the particular nature of the pandemic and its outsize effect
on in-person service sector jobs where part-time is unlikely to be as feasible as in, say, manufacturing that
suffered more heavily during previous recessions. Alternatively, the health risks implied by the pandemic
and the large extensions of unemployment insurance in response may have reduced the incentives for
part-time work. Examining these questions is an interesting topic for future research.

5.5 Job separations, new hires, recalls, and excess turnover

The linked worker-establishment structure of the HB data also allows us to provide a detailed account of
gross job flows. We decompose weekly employment changes into job separations, new hires, and recalls.
As is common in the literature, recalls are defined as workers who are employed with the establishment at
some point in the past, disappear for at least one time period (one week in our case), and then reappear
as employees in the same establishment.
Figure 10 reports the different weekly gross flows as a rate of average employment in the same week
and the preceding week (as for the previous results, all weighted by industry-size-region cells). As panels
(a) and (b) show, the job separation rate spikes the week of March 22-28 (week 6), the same week as
business closures spike, while the new hire rate declines. Both rates then return to their pre-pandemic
average by mid-June and remain essentially the same as one year earlier.
As shown in panel (c), the recall rate of workers previously employed in the same establishment
increases substantially in the weeks following the initial spike in separations, peaking the week of May
3-9 (week 12). The recall rate then declines steadily through the week of June 28 - July 4 (week 20)
and thereafter remains slightly elevated through the end of summer before essentially returning to the
corresponding 2019 value (the spikes around week 42 and 47 are due to the Thanksgiving and Christmas
holidays).39
Panel (d), finally, shows the excess turnover rate, which is computed as the difference between the
sum of separations rate, new hiring rate, and recall rate minus (the absolute value of) net employment
growth. The excess turnover rate drops briefly in the beginning of the pandemic as new hiring and recalls
decline and then jumps up as recalls jump up while some businesses still show excess job separations.

39
It is interesting to compare these recall numbers to recent results on recalls in the literature. In particular Fujita
and Moscarini (2017) document based on monthly household survey data that on average about 40% of workers return to
their previous employment after a jobless spell. Our estimates imply that the corresponding recall rate, measured as recalls
divided by the total of recalls and new hires, averages about 55% for 2019 and rises as high as 85% in mid-April 2020. The
higher average for 2019 is primarily due to time aggregation in monthly data (we observe non-trivial non-employment spells
lasting less than one month with subsequent recall).

32
After mid-June, excess turnover averages about the same rate as one year earlier.

Figure 10: Job separations, new hires, recalls, and excess turnover in small businesses
(a) Weekly rate of job separation (b) Weekly rate of new hiring
35 18

30
15

Weekly separation rates Weekly rates of new hires
12

(in %) (in %)
15

6
10

3
5

2020 2019 2020 2019

15 25

Weekly worker turnover rates (in %)
Weekly rates of recalling workers
12 20

9 15

(in %)
6 10

3 5

2020 2019 2020 2019

Notes: Weekly rates of job separation, new hires, recalls, and excess turnover for small businesses with less than 50 employees in Retail
Trade (NAICS 44-45), Education & Health Services (NAICS 61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS
81). All rates are computed as a percent of average employment in the same week and the preceding week.

The results indicate that the rebound in small business employment following the sharp decline in the
beginning of the pandemic is driven primarily by recalls of temporarily furloughed workers as opposed to
new hires, which is in line with other estimates (e.g. Ganong et al., 2021). This is quite different from
previous downturns (e.g. the Great Recession) where a larger share of separations was permanent and
the recovery was more sluggish due to persistently lower new hiring rates. Furthermore, the quick return
in the excess turnover rate to its 2019 average suggests that, at least within the four in-person service

33
sectors considered, the pandemic has so far not led to major reallocations of labor.

6 Effects of the Paycheck Protection Program

As a second application, we use the HB data to investigate the extent to which small business dynamics
and employment was affected by the Paycheck Protection Program (PPP), which provided loans to small
businesses during the first months of pandemic. The program has been the subject of much controversy
and intense research. The novelty of our investigation is that we exploit the high-frequency and detailed
geographic dimension of the HB data to differentiate the effects of variations in timely access to PPP
loans from the many other changes that occurred in the first months of the pandemic. As before, properly
distinguishing the effects of business closings and openings from sample churn turns out to be central for
the results.

6.1 Delayed access to PPP loans

The 2020 CARES Act that was signed into law on March 27, 2020 appropriated $349 billion in PPP loans
to support firms with fewer than 500 employees prior to the pandemic.40 To allow broad access, many
of the usual eligibility criteria to access government loan programs were waived and the loans came with
very favorable terms: qualifying businesses could apply for 2.5 times the average total monthly payroll
for each employee up to a maximum of $10 million, and the loans had a duration of two years at a 1%
annual interest rate but were forgivable if the business spent at least 75% on payroll within 8 weeks of
loan disbursement.41
While the Small Business Administration (SBA) was responsible for oversight, firms applied for the
loans through local lenders and the first loans were approved on April 3. The demand for loans was so
overwhelming that by April 16, the appropriated funds were depleted. In response and after considerable
uncertainty, Congress voted on an additional $321 billion in PPP funding that the President signed into
law on April 24. Banks started issuing new loans on April 27 and demand spiked immediately, with 60%
of the additional funds allocated within two weeks of reopening of the program. Thereafter, loan demand
declined substantially and PPP stopped taking new applications on August 8, with almost $150 billion

40
For multi-establishment firms in Accommodation & Food Services (NAICS 72), the 500 employee threshold applied to
establishments within certain limits.
41
Businesses also had to maintain or restore employee counts and pay for loans to be forgivable. On June 5, 2020
Congress relaxed the conditions for loan forgiveness, lowering the threshold on PPP funds used for payroll from 75% to 60%
and increasing the number of weeks to use the funds from 8 to 24. See https://www.sba.gov/funding-programs/loans/covid-
19-relief-options/paycheck-protection-program for details.

34
in unused funds remaining.42
As documented in detail by Bartik et al. (2021), Doniger and Kay (2021), and Granja et al. (2020)
among others, the first round of PPP was subject to large geographic disparities in loan allocations,
likely reflecting differences in the ability and willingness of local banks to process and approve the large
initial influx of loan applications. Funds did not necessarily flow to areas of the country where the initial
economic effects from the pandemic were largest but were instead driven by the local presence of the
different lenders. In addition, the first loans were unusually large, made to relatively larger businesses.
Hence, many of the smallest businesses – the ones that are the focus of our study – were subject to
delayed access to PPP loans during the beginning of the pandemic, and the extent of this delay depended
in large part on the quasi-random presence of different banks across localities.

6.2 Research design

We exploit the geographic variation in initial loan access to evaluate the effects of PPP for small business
activity. Similar to Doniger and Kay (2021) we measure delayed access to PPP loans by the share of loans
issued between April 26 and May 2 (the week when additional PPP funding became available) relative to
the total amount of loans issued between April 12 and May 2 (the week when initial PPP funding ran out to
(loans April 26-May 2)
the week when additional PPP funding became available); i.e. share PPP delayedc = (loans April 12-May 2)c
c

, where c denotes the county of the businesses receiving the loans.43 As Doniger and Kay (2021) argue,
focusing on a relatively narrow window around the temporary exhaustion is important to avoid selection
issues associated with the first few weeks of the program.
We construct share PPP delayedc using data on all PPP loans from the SBA. The loans made during
these weeks account for about one third of all loans and for an even larger share of loans made to
businesses with fewer than 50 employees. As shown in the Appendix, the variation in share PPP delayedc
across counties is wide, with a median of 40% and a 10-to-90-percentile range of [26%, 60%].44

42
In December 2020, Congress voted for and the President signed into law a third round of PPP consisting of an additional
$285 billion in funding and new eligibility rules. Loans started in mid-January 2021 and the program ran through the end
of May 2021. This third round is not the focus of our investigation.
43
The weeks in our estimation run from Sunday to Saturday. April 12, 19 and 26 are Sundays. Doniger and Kay use
a narrower 2-day window around the temporary exhaustion of PPP loans to measure the share of delayed PPP loans, and
they compute the measure at the broader CBSA geographic level. Our estimates are robust to using their narrower time
window and the broader CBSA level.
44
In the regression, we use a subset of 1,956 counties for which we have reliable HB data (out of 3,143 counties for which
we have PPP data). The distribution of share PPP delayedc for this subset of counties is almost identical to the distribution
for the full set of counties.

35
We use the share PPP delayedc measure to estimate the following county-level regression

57
%
yc,t = αt ( {week = t} × share PPP delayedc ) + X#c,t γ + φt + µc + εc,t (6)
t=0

where yc,t is either the percent deviation of employment across establishments in county c in week t
relative to its employment in the first week of 2020 (t = 0); the fraction of establishment in county
c being closed in week t; or the fraction of establishments in county c being newly opened in week t.
The vector Xc,t contains a vector of county-specific controls measuring weekly COVID cases and deaths,
non-pharmaceutical interventions (NPIs), school closures, weather, as well as week fixed effects interacted
with average county household income prior to the pandemic.45 Finally, φt is a week fixed effect capturing
time variations in average yc,t ; µc is a county fixed effect controlling for unobserved average differences
across counties; and εc,t is the error term. All regressions are weighted by county level employment prior
to the pandemic in the four sectors considered, and standard errors are clustered at the county level.46
The αt are the main coefficients of interest and measure the effect in week t of the share of delayed PPP
loans in county c. The identifying assumption for estimates of these coefficients to have economic meaning
is that conditional on controls, share PPP delayedc reflects the relative difficulty for small businesses
located in county c to obtain a PPP loan and is independent of other factors affecting small business
activity during the initial phase of the pandemic.
An obvious concern with this assumption is that share PPP delayedc may reflect at least partly system-
atic variations in PPP loan demand across counties. If, for instance, small businesses in counties affected
more severely by the pandemic were also more likely to apply for a PPP loan earlier, then this would
bias our estimates of αt away from zero. As Doniger and Kay (2021) show, however, there is no clear
geographic concentration in loan issuance within the narrow window around the temporary exhaustion of

45
Weekly county-level COVID cases and deaths per capita are obtained from Covid Act Now. The NPIs included are
weekly county-level indicators of industry closures, stay-at-home orders, and gathering bans from the Data Library on
Industry Closures and Reopening by Atalay et al. (2020) and the Center for Disease Control and Prevention. In addition, we
include a weekly containment index containing seven indicators of state-level policies from the Oxford Covid-19 Government
Response Tracker (see Hale et al., 2021). We allow for differential impact of these controls over subperiods of the sample.
Following Bravata et al. (2021), we proxy weekly school closures by the log change in county average school visits relative
to one year prior using data on individual schools from Safegraph. For weather, we use the maximum county temperature
in a given week from the Climatology Lab. We also controlled for other measures of weather and found very similar
results. Finally, county level average household income are 2016-2019 estimates from the American Community Survey. The
predictive value of these controls for small business activity is interesting in its own right and extends earlier results by
Bartik et al. (2020), Chetty et al. (2020), and Goolsbee and Syverson (2020), among others. We discuss these results as well
as further details on the different controls in the Appendix.
46
County level employment prior to the pandemic is computed from the Quarterly Workforce Indicators (QWI). Results
are robust to using county population as weights or estimating the regression at the establishment level that implicitly weighs
counties by the count of HB establishments.

36
PPP considered here, and share PPP delayedc varies substantially between adjacent counties in the same
state. Furthermore, our regressions control for a host of county-specific time-varying factors that may
have affected small business activity throughout the pandemic as well as week fixed effects interacted with
a county’s pre-pandemic average household income that absorb local demand effects related to a county’s
affluence.47 Finally, as shown below, our estimations do not show significant effects on employment and
revenue prior to the start of PPP, which provides direct evidence against this particular demand-side
story.
Another threat to identification could be that there are systematic differences in the distribution
of productivity of businesses across counties and that more productive businesses both grow faster and
applied for a PPP loan earlier. Since average differences are absorbed by county fixed effects, this
would bias estimates of αt towards zero only insofar as these growth effects were time-varying during
the pandemic. To address this possibility, we also estimate regression (6) at the establishment level and
control for establishment fixed effects that differentiate out firm-specific factors including productivity
differences. As shown in the Appendix, all the results are robust to these establishment-level regressions
and are even somewhat stronger.48

6.3 Results

Figure 11 reports the point estimates for αt together with 95% confidence bands. Panel (a) shows that
counties with a higher share of delayed loans experienced lower employment growth starting the week
after the exhaustion of the first round of PPP loans. This negative effect becomes stronger from May
through August and remains significantly negative through the end of the sample, long after PPP ended.
Prior to the temporary exhaustion of PPP in mid-April, the estimates are close to zero and insignificant,
indicating that the negative effect is not driven by pretrends. As a further robustness check for pretrends,
we estimate the same regression for small business revenue from Womply, made available at the county-
week level by Chetty et al. (2020). We find no evidence of pretrends in this regression, either.

47
As Chetty et al. (2020) document, more affluent localities suffered relatievely larger declines in spending on in-person
services and employment in the beginning of the pandemic, presumably due to local demand declining by more in these
localities.
48
The establishment-level regressions also measure share PPP delayedc separately for each of the four service sector con-
sidered, control for week fixed effects interacted with average household income at the zip-code level, and cluster standard
errors at the establishment level.

37
Figure 11: Effect of delayed PPP loans on small business activity
(a) County employment (b) Employment of always active businesses

Notes: Coefficient estimates of share PPP delayedc interacted with weekly fixed effects. Shaded areas show 95% confidence bands.
All regressions are estimated over all weeks between January 5-11, 2020 and January 31 - February 6, 2021. share PPP delayedc is
constructed as the amount of PPP loans issued in county c during the week of April 26 relative to the total amount of PPP loans issued
per county during the weeks of April 12, April 19, and April 26. County employment in Panel (a) is the percent deviation relative to
mid-February 2020 employment for all county-weeks for which the HB sample contains positive employment observations. Employment
of always active businesses in Panel (b) is the percent deviation relative to mid-February 2020 employment for all establishments in a
county that are continuously active throughout the entire sample. Business closings in Panel (c) is the percent ratio of the total count
of establishments closed in a county in week t to the count of businesses in the reference week. Newly opened businesses in Panel (d)
is the percent ratio of the total count of new establishments in a county in week t relative to the count of businesses in the reference
week. All regressions control for county-specific time-varying controls as described in the text as well as week- and county fixed effects.
Regressions are weighted by county employment prior to the pandemic in the four service sectors considered, and standard errors are
clustered at the county level.

The fact that the exhaustion of PPP loans lasted for only 10 days and that the additional funding from
PPP approved in late April 2020 was not used up by the time the program stopped taking applications
in early August 2020, begs the question of why the estimated employment effects are so persistent. To
shed light on this question, we run regression (6) separately for employment growth of businesses that
are continuously active throughout the sample, the share of business closings, and the share of new
business openings (computed, as above, using our methodology that distinguishes business closings and

38
new openings from sample churn and then adjusts for selection).
As shown in panel (b), employment growth by always active businesses in counties with a larger
share of delayed PPP loans is barely affected. There is a small decline around the week of the temporary
exhaustion of PPP loans that turns marginally significant the week after PPP loans restart, but thereafter
the effect is close to zero and insignificant (in the establishment-level regressions, this negative effect is
somewhat larger and temporarily significant).
In contrast, as shown in panel (c), counties with a larger share of delayed PPP loans experience a
significantly higher rate of business closings starting the week of the temporary exhaustion of PPP. This
effect peaks in August 2020, then declines gradually through Fall of 2020, and stabilizes at about half of
its peak value by 2021.
Panel (d), finally, shows that the share of delayed PPP loans has only a very small and generally
insignificant effect on the rate of new business openings, which confirms the validity of the design since
new businesses by definition did not qualify for PPP loans.49
To interpret the magnitude of the estimated coefficients, consider the difference in share PPP delayedc
between counties at the 90th and the 10th percentile of the distribution, which is 34% (= 60%−26%). The
point estimate at the end of the sample in mid-February 2021 is about −0.15 percentage points for the
effect on county employment and about −0.1 percentage points for business closings. This implies that
a county at the 90th percentile of delayed PPP loans has about 5.1% lower small business employment
(relative to mid-February 2020) and an about 3.4% higher rate of business closings than a county at the
10th percentile. Given that in mid-February 2021, average small business employment across the four
sectors considered was about 8% below its pre-pandemic level and the average cumulative closing rate
amounted to about 17%, these magnitudes are substantial.
The estimates imply that the negative and persistent employment effect of delays in PPP funding is
in large part driven by higher rates of business closings, some of which appear to be permanent. This
suggests that the exhaustion of PPP in mid-April 2020 occurred at a critical moment when many small
business owners, faced with an unprecedented downturn amid COVID health concerns, stay-at-home
orders, and business restrictions had to decide whether to continue operating and hope for loan relief
from the government or cut their losses and close shop.50

49
As argued by Acemoglu et al. (2018), supporting incumbent businesses could potentially suppress new openings. The
slight negative estimates are consistent with this possibility, although the effect is very small.
50
As an example of these difficulties, see the NPR Planet Money podcast episode 990 “The Big Small Business Rescue”
from April 10, 2020.

39
6.4 Aggregate employment effects and comparison to the literature

To provide a sense of the the aggregate employment effects of PPP loan delay, we follow an approach
similar to Mian and Sufi (2012) and Berger et al. (2020) that is also used in the PPP context by Granja
et al. (2020). For each county c, we compute the difference between actual small business employment
Ec,t and counterfactual employment Ẽc,t under the assumption that the county experienced zero delay in
loans around the temporary exhaustion in PPP (which is the case for a small set of counties); i.e.

α̂t
Ec,t − Ẽc,t = × share PPP delayedc × Ec,0 (7)
100

where α̂t are the regression estimates reported in Figure 11 and Ec,0 is small business employment in
county c in the pre-pandemic reference period. We then aggregate across counties using pre-pandemic
employment weights.51 The approach implicitly assumes that share PPP delayedc is a good measure of the
difficulty of small businesses in obtaining PPP funding during the first round of loans. The approach also
abstracts from possible general equilibrium effects of more timely availability of PPP loans and any other
differences across counties in the difficulty of obtaining loans that are unrelated to PPP. Nevertheless,
the approach is illustrative because it provides a benchmark for the overall effect of PPP and allows us
to compare our estimates to other results in the literature.
Given that small business employment in the four sectors considered was about 30 million prior to
the pandemic, the point estimate of α̂t = −0.25 for the last week of July 2020 implies that without delays
in PPP loans, small business employment in the four sectors would have been about 3 million or 10%
higher. In turn, for the last week of January 2021, the estimate of α̂t = −0.15 implies that without
delays in PPP loans, small business employment in the four sectors would have been about 1.8 million
or 6% higher. Since the delays in PPP loans could have been avoided by appropriating a larger initial
amount for PPP in the CARES Act, the costs of doing so would have been essentially zero. Vice versa,
the estimates suggest that if PPP had not been part of the CARES Act, small business closings would
have been substantially larger and the pandemic would have caused a larger and longer-lasting decline
in service sector jobs.52

51
Specifically, the aggregate employment effect relative to pre-pandemic employment is estimated as EtE−0Ẽt =
α̂t % Ec,0
100 c E0 share PPP delayedc , where Ec,0 and E0 denote pre-pandemic employment in county c and nationwide, respec-
tively, for small businesses in the four sectors considered.
52
We refrain from attempting to infer the overall number of jobs saved by PPP for two reasons. First, our estimates
pertain to small businesses in four of the service sectors affected most by the pandemic. Larger businesses and businesses
in other sectors that received PPP loans may have been less dependent on PPP funding, but the HB data do not allow us
to quantify the extent of this treatment effect heterogeneity. Second, under the counterfactual assumption that the CARES
Act or the subsequent COVID relief bills had not contained any funding for PPP, small businesses as a whole could have

40
Our results complement event-study estimates of the effects of PPP by Bartik et al. (2021), Doniger
and Kay (2021), and Granja et al. (2020) among others. Granja et al. (2020) use Homebase data like
us but apply a different research design that exploits local variations in the presence of banks that
processed PPP loans at varying expediency. They find that over the months of April, May and June
2020, employment in small businesses would have been 4.5% higher if all banks had been equally expedient
in making loans, which implicitly assumes that the initial PPP funding in the CARES Act would have
been larger (i.e. no loan delays). This number is about half of our estimate. However, their estimation
treats all exits from Homebase as business closings, which is likely to impart substantial error in the
cross-regional variation in small business employment. Bartik et al. (2021) use data from a survey of
small businesses owners in late April 2020 during the temporary exhaustion phase of PPP. Leveraging
information on existing banking relationships as instrumental variables, they find that PPP loan approval
led to a 14 to 30% increase in expected survival probability and had a positive but imprecisely estimated
effect on employment. Our estimates provide confirmation of this finding with actual data on employment
and closings. Using the same aggregation approach as for employment above, the estimated aggregate
reduction in permanent closings if there had been no delay in PPP loans is about 5%. This is lower
than the estimates in Bartik et al. (2021), which may be due to the fact that we use actual data as
opposed to expectations formed in the initial phase of the pandemic when uncertainty was likely higher.
Doniger and Kay (2021), whose research design we adopt here, use monthly household survey data from
the CPS and find persistent effects of PPP loan delay on unemployment and non-participation. From
their estimates, they infer that a reduction in the average share of delayed PPP loans by 20% would
have increased private-sector employment in mid-May 2020 by nearly 2.8 million with much of the job
savings concentrated in small businesses. This number is difficult to compare to our estimates since it
pertains to all private sector employment while we focus on the four hardest-hit service sectors. At the
same time, our estimates confirm that PPP loan delay had a substantial negative effect on small business
employment and provide an explanation for why this effect is persistent: it is in largely due to business
closings of which many appear to be permanent.53
This finding is confirmed in more recent studies by Dalton (2021) and Autor et al. (2022) who match
establishment-level data from the QCEW and ADP, respectively, to individual PPP loan information and

reacted very differently from the present context where PPP funds were available but temporarily ran out for a relatively
short period of time. This in turn could have led to important general equilibrium effects that are difficult to quantify.
53
More generally, our results are consistent with a growing literature documenting that limited cash-on-hand and working
capital adversely affects labor demand and makes small businesses more sensitive to negative shocks (e.g. Chodorow-Reich,
2014, Bacchetta et al., 2019, Barrot and Nanda, 2020, or Mehrotra and Sergeyev, 2020 among others). Our results, however,
put increased emphasis on the effects that these financial frictions can have on the extensive margin – i.e. business closings
– which likely has more permanent effects.

41
compare employment and closing probabilities of businesses who received a PPP loan earlier with those
who recevied a loan later. Like us, they find sizable effects on employment that are concentrated among
businesses with fewer than 50 employees, and that a large part of these employment effects are due to
business closures.
Finally, Autor et al. (2020), Chetty et al. (2020), and Hubbard and Strain (2020) exploit the 500
employee threshold for PPP loan eligibility to estimate the overall impact of PPP. These studies find
more modest employment effects, suggesting that businesses around the 500 employee threshold have
been less dependent on PPP loan support, which is consistent with results by Chodorow-Reich et al.
(2020). This suggests that the effectiveness of PPP could have been enhanced if the program had been
restricted at least initially to the smallest businesses.

7 Conclusion

In this paper, we use establishment-level data from Homebase, a scheduling and time clock software
provider, to construct weekly estimates of the effect of the COVID-19 pandemic on small business dy-
namics and employment in four of the hardest hit service sectors. The main methodological contribution
relative to the many other studies using Homebase and other high-frequency establishment-level datasets
is that we match the Homebase establishment records with independent data to assess the representa-
tiveness of the establishments in the Homebase sample and to distinguish business openings and closings
from sample churn.
Our implementation uses data from Safegraph, Google, Facebook, but other datasets measuring
business activity could be used as well. Similarly, while the Homebase data provides an excellent case to
test the quality of our methods, other establishment-level datasets could be used. As such, we consider our
paper as a proof of concept on how to combine different data sources to construct employment estimates
that directly incorporate the effects of business openings and closings in almost real-time and that can
be benchmarked to official statistics and used to measure the impact of rapidly disseminating shocks and
economic policies.

42
References

Acemoglu, D., U. Akcigit, N. Bloom, and W. R. Kerr (2018). Innovation, Reallocation and Growth.
American Economic Review 108, 3450–3491. 49

Atalay, E., S. Fujita, S. Mahadevan, R. Michaels, and T. Roded (2020). Reopening the Economy: What
are the risks, and what have States done? Research brief, Federal Reserve Bank of Philadelphia. 45

Autor, D., D. Cho, L. D. Crane, M. Goldar, B. Lutz, J. Montes, W. B. Peterman, D. Ratner, D. V.
Vallenas, and A. Yildirmaz (2020). An Evaluation of the Paycheck Protection Program Using Admin-
istrative Payroll Microdata. Working paper, MIT. 1, 6.4

Autor, D., D. Cho, L. D. Crane, M. Goldar, B. Lutz, J. Montes, W. B. Peterman, D. Rat-
ner, D. V. Vallenas, and A. Yildirmaz (2022). The 800BillionP aycheckP rotectionP rogram :
W hereDidtheM oneyGoandW hyDidItGoT here?W orkingP aper29669, N ationalBureauof EconomicResearch.1, 6

Bacchetta, P., B. Kenza, and C. Poilly (2019). Corporate cash and employment. American Economic
Journal: Macroeconomics 11 (3): 30–66 . 53

Barrot, J.-N. and R. Nanda (2020). The Employment Effects of Faster Payment: Evidence from the
Federal Quickpay Reform. The Journal of Finance 75 (6):3139–3173 . 53

Bartik, A. W., M. Bertrand, Z. B. Cullen, E. L. Glaeser, M. Luca, and C. T. Stanton (2020). The Impact
of COVID-19 on Small Business Outcomes and Expectations. Proceedings of the National Academy of
Sciences 117 (30), 17656–17666. 1

Bartik, A. W., M. Bertrand, F. Lin, J. Rothstein, and M. Unrath (2020). Measuring the labor market at
the onset of the COVID-19 crisis. Brookings Papers on Economic Activity forthcoming. 1, 5.3, 45

Bartik, A. W., Z. Cullen, E. L. Glaeser, M. Luca, C. Stanton, and A. Sunderam (2021). The Targeting and
Impact of Paycheck Protection Program Loans to Small Businesses. Working Paper 21-021, Harvard
Business School. 1, 6.1, 6.4

Bartlett, R. P. and A. Morse (2020). Small Business Survival Capabilities and Policy Effectiveness:
Evidence from Oakland. Working paper, National Bureau of Economic Research. 1

Berger, D., N. Turner, and E. Zwick (2020). Stimulating Housing Markets. Journal of Finance 75 (1),
277–321. 6.4

43
Bick, A. and A. Blandin (2020, April). Real Time Labor Market Estimates During the 2020 Coronavirus
Outbreak. Working paper, Arizona State University. 1

Borowczyk-Martins, D. and E. Lalé (2019). Employment Adjustment and Part-time Work: Lessons from
the United States and the United Kingdom. American Economic Journal: Macroeconomics 11 (1),
389–435. 5.4

Bravata, D., J. H. Cantor, N. Sood, and C. M. Whaley (2021). Back to School: The Effect of School Visits
During Covid-19 on Covid-19 Transmission. Working paper, National Bureau of Economic Research.
45

Cajner, T., L. D. Crane, R. A. Decker, J. Grigsby, A. Hamins-Puertolas, E. Hurst, C. Kurz, and
A. Yildirmaz (2020, May). The U.S. Labor Market during the Beginning of the Pandemic Reces-
sion. 1, 5.1, 5.3

Chetty, R., J. N. Friedman, N. Hendren, M. Stepner, and the Opportunity Insights Team (2020). The
Economic Impacts of COVID-19: Evidence from a New Public Database Built from Private Sector
Data. Working paper 26463, National Bureau of Economic Research. 3, 1, 5.3, 45, 6.3, 47, 6.4

Chodorow-Reich, G. (2014). The employment effects of credit market disruptions:Firm-level evidence
from the 2008–9 financial crisis. The Quarterly Journal of Economics129 (1): 1–59 . 53

Chodorow-Reich, G., O. Darmouni, S. Luck, and M. Plosser (2020). Bank Liquidity Provision Across the
Firm Size Distribution. Journal of Financial Economics forthcoming. 1, 6.4

Coibon, O., Y. Gorodnichenko, and M. Weber (2020, April). Labor Markets During the COVID-19 Crisis:
A Preliminary View. Working Paper 27017, National Bureau of Economic Research. 1

Crane, L. D., R. A. Decker, A. Flaeen, A. Hamins-Puertolas, and C. Kurz (2022). Business Exit During the
COVID-19 Pandemic: Non-Traditional Measures in Historical Context. Journal of Macroeconomics 72,
103419. 1, 4.1, 5.2, 35

Dalton, M. (2021, November). Putting the Paycheck Protection Program into Perspective: An Analysis
Using Administrative and Survey Data. Working Paper 542, Bureau of Labor Statistics. 1, 6.4

Dalton, M., E. W. Handwerker, and M. A. Loewenstein (2020, October). Employment changes by
employer size during the COVID-19 pandemic: a look at the Current Employment Statistics survey
microdata. Monthly Labor Review , 1–17. 1, 5.1, 5.3

44
Decker, R. A. and J. C. Haltiwanger (2022). Business Entry and Exit in the COVID-19 Pandemic: A
Preliminary Look at Official Data. Feds notes, Board of Governors of the Federal Reserve System. 1,
31, 5.2

Doniger, C. L. and B. Kay (2021). Ten Days Late and Billions of Dollars Short: The Employment
Effects of Delays in Paycheck Protection Program Financing. Finance and economics discussion series
2021-003, Board of Governors of the Federal Reserve System. 1, 6.1, 6.2, 6.2, 6.4

Fairlie, R. (2020). The impact of COVID-19 on small business owners: Evidence from the first three
months of widespread social-distancing restrictions. Journal of Economics Management Strategy 29,
727–740. 1

Fairlie, R. W., F. M. Fossen, R. L. Johnsen, and G. Droboniku (2022). Were Small Businesses More
Likely to Permanently Close in the Pandemic. Working paper 30285, National Bureau of Economic
Research. 1, 5.2

Fazio, C. E., J. Guzman, Y. Liu, and S. Stern (2021). How is COVID Changin the Geography of
Entrepreneurship? Evidence from the Startup Cartography Project. Working paper 28787, National
Bureau of Economic Research. 1, 5.2

Fort, T. C., J. C. Haltiwanger, R. S. Jarmin, and J. Miranda (2013). How Firms Respond to Business
Cycles: The Roel of Firm Age and Firm Size. IMF Economic Review 61 (3), 520–559. 1, 5.1

Fujita, S. and G. Moscarini (2017). Recall and unemployment. American Economic Review 107 (12),
3875–3916. 39

Ganong, P., F. Greig, M. Liebeskind, P. Noel, D. M. Sullivan, and J. Vavra (2021). Spending and
Job Search Impacts of Expanded Unemployment Benefits: Evidence from Administrative Micro Data.
Working paper, Becker Friedman Institute. 5.5

Goolsbee, A. and C. Syverson (2020). Fear, Lockdown, and Diversion: Comparing Drivers of Pandemic
Economic Decline 2020. Working paper, Becker Friedman Institute. 45

Granja, J., C. Makridis, C. Yannelis, and E. Zwick (2020). Did the Paycheck Protection Program Hit
the Target? Technical report. 1, 6.1, 6.4, 6.4

Hale, T., N. Angrist, R. Goldszmidt, B. Kira, A. Petherick, T. Phillips, S. Webster, E. Cameron-Blake,

45
L. Hallas, S. Majumdar, and H. Tatlow (2021). A global panel database of pandemic policies (Oxford
COVID-19 Government Response Tracker. Nature Human Behaviour 5 (529-538). 45

Haltiwanger, J. C. (2021). Entrepreneurship During the Covid-19 Pandemic: Evidence from the Business
Formation Statistics. Working paper 28912, National Bureau of Economic Research. 1, 5.2

Haltiwanger, J. C., H. R. Hyatt, L. B. Kahn, and E. McEntarfer (2018). Cyclical Job Ladders by Firm
Size and Firm Wage. American Economic Journal: Macroeconomics 10 (2), 52–85. 1, 5.1

Hubbard, G. and M. R. Strain (2020). Has the Paycheck Protection Program Succeeded? Brookings
Papers on Economic Activity forthcoming. 1, 6.4

Kahn, L. B., F. Lange, and D. G. Wiczer (2020, April). Labor Demand in the Time of COVID-19:
Evidence from Vacancy Postings and UI Claims. Working Paper 27061, National Bureau of Economic
Research. 1

Lewis, D., K. Mertens, J. H. Stock, and M. Trivedi (2021). Measuring Real Activity Using a Weekly
Economic Index. Journal of Applied Econometrics forthcoming. 1

Mehrotra, N. and D. Sergeyev (2020). Financial shocks, firm credit and the Great Recession. Journal of
Monetary Economics In Press. 53

Mian, A. and A. Sufi (2012). The Effects of Fiscal Stimulus: Evidence from the 2009 Cash for Clunkers
Program. Quarterly Journal of Economics 127 (1107-1142). 6.4

Moscarini, G. and F. Postel-Vinay (2012). The Contribution of Large and Small Employers to Job
Creation in Times of High and Low Unemployment. American Economic Review 102 (6), 2509–39. 1,
5.1

Rinz, K. (2020). Understanding Unemployment Insurance Claims and Other Labor Market Data During
the COVID-19 Pandemic. Working paper. 4.3

Vaan, M. D., S. Mumtaz, A. Nagaraj, and S. B. Srivastava (2021). Social Learning in the COVID-19 Pan-
demic: Community Establishments’ Closure Decisions Follow those of Nearby Chain Establishments.
Management Science 67 (7), 4446–4454. 3

Yelp (2020, September). Local economic impact report. https://www.yelpeconomicaverage.com/business-
closures-update-sep-2020. 3

46
Supplementary material for:
Measuring Small Business Dynamics and Employment
with Private-Sector Real-Time Data

André Kurmann Etienne Lalé Lien Ta
Drexel University Université du Québec Drexel University
à Montréal

Contents

A Homebase data 2
A.1 Description and definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
A.2 Employment, hours worked and active establishments . . . . . . . . . . . . . . . . . . . . 5
A.3 Industry classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
A.4 Base sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
A.5 Benchmarking establishment counts and employment to the QCEW . . . . . . . . . . . . 8

B Safegraph and Facebook data 13
B.1 Safegraph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
B.2 Facebook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

C Matching procedures 16
C.1 Matching with Safegraph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
C.2 Matching with CrowdTangle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

D Closings and openings 19
D.1 Google/Facebook approach to determine business closings and openings . . . . . . . . . . 20
D.1.1 Permanent closings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
D.1.2 New openings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
D.2 Incorporating closings and openings into our estimation . . . . . . . . . . . . . . . . . . . 24
D.2.1 Employment at establishments closing permanently . . . . . . . . . . . . . . . . . 24
D.2.2 Employment at newly opened establishments . . . . . . . . . . . . . . . . . . . . . 26
D.3 Recap of our approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
D.4 Benchmarking to BED establishment births and deaths . . . . . . . . . . . . . . . . . . . 31

1
E Representativeness and additional checks on closings/openings 36
E.1 Using visits data to assess representativeness of the Homebase sample . . . . . . . . . . . 36
E.2 Using visits data to compare closings and openings against sample churn . . . . . . . . . . 41

F Employment decompositions 43
F.1 Decomposition by establishment status . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
F.2 Decomposition into hiring and separation flows . . . . . . . . . . . . . . . . . . . . . . . . 45

G Additional figures and tables 46
G.1 Counterfactual employment estimates for the pre-pandemic period . . . . . . . . . . . . . 46
G.2 Regression variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
G.3 Additional regression results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
G.3.1 Coefficient estimates for different control variables . . . . . . . . . . . . . . . . . . 51
G.3.2 Coefficient estimates for establishment-level regressions . . . . . . . . . . . . . . . 51

A Homebase data

The Homebase (HB) data contains employment and wage records for small businesses primarily engaged in
customer-oriented services. In February 2020, the database covered 938,072 individual workers employed
in 94,203 business locations (establishments) belonging to approximately 80,921 firms.

A.1 Description and definitions

• A HB customer (firm) is identified by a unique persistent company id. Different business locations
(establishments) belonging to the same firm are identified by a unique persistent location id. The
establishment is our main unit of observation.

• Some HB customers create several location id’s for the same establishment in order to track
different departments. These departments either have the same address or no address. We tag them
by creating a common parent location id. In the sample construction, we collapse location id’s
with the same parent location id into a single location id.1

• Each user of HB’s services has a unique persistent user id. Associated with each user id is a level
(employee, manager, general manager) and an owner status (owner, non-owner). The values of
these fields do not change over time.

1
parent location id’s account for about 6% of all unique HB location id’s., and about 15% of location id’s have a
parent location id.

2
• A user id shows up in the database on a given day if the user had some activity; e.g. scheduling
or logging hours, sending messages to another user, etc.2

• Each user id comes with a hours scheduled and a hours worked field. If both fields are empty, we
designate the user as having “untracked hours”. If either the hours scheduled or the hours worked
field contains data, then we designate the user as having “tracked hours”. When a user has data in
both fields, we use hours worked to measure actual hours worked, although we find that the data
in hours scheduled and hours worked correspond closely.3

• In a given week in 2020, user id’s with untracked hours make up one-fifth of all user id’s in the
raw HB data. For those with tracked hours, 30% of user id’s have hours worked and a missing
hours scheduled field. The remaining 70% of user id’s with tracked hours have entries in both
the hours scheduled and hours worked fields. 42% of these user id’s have hours worked that
are exactly equal to hours scheduled, suggesting they use HB for the purpose of scheduling hours
and that the hours worker field has been populated with information from the hours scheduled
field.

• Figure A1 shows the distribution of weekly hours worked on average over the first 10 weeks of 2020.4
Among managers and general managers, there is a mode at 40 hours, and a substantial proportion
of managers with 0 tracked hours (8.7% of managers, 3.3% of general managers). Median weekly
hours worked equal 21.7 for employees, 28.5 for general managers, and 24.3 for managers. Employees
make up 90% of all user id’s in the data with tracked hours in the first 10 weeks of 2020, while
general managers account for 1% of these user id’s.

• Figure A2 shows the proportion of location id’s in 2020 with tracked hours only, untracked hours
only, and both tracked and untracked hours. At the onset of the pandemic (mid-March 2020),
location id’s with untracked hours make up 10% of all location id’s with some activity. The
corresponding figure for location id’s with both untracked and tracked hours is 66%. During the
first weeks of the pandemic, there is a clear shift towards location id’s with untracked hours only.
The increase is mainly driven by an inflow of location id’s that used to have both untracked and
tracked hours. Transition probabilities (not reported here) of switching across the 3 groups return
2
But simply being logged in to the app on a cellular device is not counted as an activity.
3
The difference between hours worked and hours scheduled, conditional on being different from each other, is symmetric,
bell-shaped centered on zero. For 75% of user id’s with different hours worked and hours scheduled, the absolute difference
between the two measurements is less than 1.2 hours.
4
In Figure A1, the sample is restricted to weekly hours worked (strictly) under 60 hours. Less than 1 percent of workers
have weekly hours worked (averaged over this 10-week window of time) above the 60 hours cutoff.

3
Figure A1: Distribution of weekly hours worked in raw HB data

Employee General Manager Manager
10

Percent

0 0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Weekly hours worked

Notes: Weekly hours worked averaged over the first 10 weeks of 2020 for all HB user id’s with tracked (worked or scheduled) hours.

Figure A2: Establishments with untracked and tracked hours
80

Share as a % of all active businesses
60

Mar 8 − Mar 14 Apr 19 − Apr 25 Oct 25 − Oct 31 Mar 21 − Mar 27 Apr 11 − Apr 17 Oct 17 − Oct 23
Mar 29 − Apr 4 Jul 12 − Jul 18 Oct 4 − Oct 10 Jul 25 − Jul 31
Jul 4 − Jul 10
May 10 − May 16 Jun 21 − Jun 27 Jan 17 − Jan 23 May 23 − May 29 Jun 13 − Jun 19
Feb 28 − Mar 6
Aug 23 − Aug 29 Sep 13 − Sep 19 Aug 15 − Aug 21
May 2 − May 8
May 31 − Jun 6 Nov 15 − Nov 21 Dec 6 − Dec 12 Nov 7 − Nov 13 Nov 28 − Dec 4
Aug 2 − Aug 8 Dec 27 − Jan 2 Sep 26 − Oct 2
Feb 16 − Feb 22 Feb 7 − Feb 13 Sep 5 − Sep 11
Jan 26 − Feb 1

Establishments with: Untracked workers only
Hours−tracked workers only
Both untracked and hours−tracked workers

Notes: HB establishments with untracked only (solid line), hours-tracked only (dashed line), and both untracked and hours-tracked
workers (dotted line), as a share (expressed in %) of all establishments with some activity in 2020.

4
to their pre-pandemic levels by mid-April. The distribution of location id’s across the 3 groups
takes longer to recover.

A.2 Employment, hours worked and active establishments

A key operational definition is that of employment. We measure employment of a location as the sum of all
users with positive hours worked or hours scheduled plus active users with untracked hours. Employment
is set to 0 if the establishment has zero tracked hours. In other words, we only count employment
at establishments that have at least one hourly-tracked employee with positive hours worked or hours
scheduled. Total weekly hours and average weekly hours per worker are computed only for user id’s
users with positive hours worked or hours scheduled.
We use these measurements of employment and hours worked to define active establishments. To be
considered as active in a given week we require a location to have positive employment during that week
and have at least 40 total weekly hours. Our goal in adopting this definition is to purge the data from
“try-outs”, i.e. locations that only show up in HB data for a short period of time and without significant
tracked hours.5

A.3 Industry classification

The historical HB data comes with an industry category for each establishment, but the available cate-
gories do not directly line up with standard industry classification, and for more than 10% of the records
the industry category is not usable because it is “Other” or “Unknown”. Instead of using this industry
classification in an ad hoc manner, we match all available HB locations by name and address to Points
of Interest (POI) from Safegraph “Core Places data”, which come with their own NAICS-6 code. The
Safegraph Core Places data consists of more than 8 million POIs. A POI is defined as a location where
individuals spend time or money. Matching the HB locations to Safegraph’s POIs involves extensive
data cleaning and standardization. The data and match procedure are described in detail in Sections
B.1 and C.1. We only retain establishments that match either exactly or with a high match rate.6 Table
A1 reports the mapping of HB industry categories to 2-digit NAICS codes for the HB locations that are

5
The total count of locations provided at the beginning of this section already purges the data from a few “try-out”
businesses. Specifically, we remove locations that either have (i) less than 1 total weekly hours worked or scheduled in at
least 50% of the weeks with tracked hours or (ii) total weekly hours worked or scheduled that are never higher than 5 hours.
6
In December 2020, HB independently started publishing NAICS industry classification. This classification is available
only for locations that are active from December 2020 onward. Since many establishments that were active in 2019 and 2020
are no longer in the HB sample, this NAICS classification is not directly useful for our estimation and benchmarking, which
starts in 2019. However, we compare our industry classifications to the one provided by HB and find a high level of overlap,
especially at the 2-digit level that we use for our estimation.

5
Table A1: Comparison between HB industry codes and 2-digit NAICS codes obtained through Safegraph

NAICS
11-23 31-33 42 44-45 48 51 52-53 54-56 61 62 71 72 81 92 Total

Beauty & personal care 0 2 4 29 0 0 3 1 6 26 5 13 329 0 418
Charities, education & membership 10 6 12 174 6 32 80 11 312 439 165 82 219 30 1,578
Food & drink 35 1,883 57 1,754 38 37 330 49 198 402 435 20,435 262 12 25,927
Health care and fitness 8 26 57 496 8 11 149 237 85 2,001 545 131 615 15 4,384
Home and repair 85 26 25 328 13 0 70 60 18 38 22 34 330 5 1,054
6 Leisure and Entertainment 3 16 3 179 4 49 53 11 35 36 432 212 38 3 1,074
Professional Services 121 200 108 1,481 101 77 518 333 238 632 533 984 896 69 6,291
Retail 17 40 32 349 30 20 337 255 55 217 63 116 330 50 1,911
Transportation 42 210 300 5,692 63 107 369 82 116 187 201 912 519 12 8,812
Other 8 3 5 125 123 5 39 14 9 32 19 42 92 1 517
Unknown 57 439 75 1,404 53 58 350 177 210 702 530 2,039 508 20 6,622

Total 386 2,851 678 12,011 439 396 2,298 1,230 1,282 4,712 2,950 25,000 4,138 217 58,588

Notes: HB data for February 2020. The table compares HB industry codes (in the first column of each row) to NAICS codes (in each column) obtained using Safegraph data (see Section
C). Each cell is the count of establishments (including establishments with 50+ employees) that are active during the reference week of 2020 (Feb 9 – Feb 15).
active in mid-February 2020.7

A.4 Base sample

For both 2019 and 2020, we form a base sample consisting of HB establishments that are active according
to the above definition for each of the three weeks centered around the week containing February 12 (week
0) plus all establishments that are temporarily inactive during this period but had positive employment
in at least three weeks at some point prior to the mid-February reference period and become active for
at least three consecutive weeks at some point thereafter. For active establishments, we determine their
size class (i.e., 1 to 4, 5 to 9, 10 to 19, and 20 to 49 employees) by taking average employment over the
three week window centered on week 0. For temporarily inactive establishments, size class is determined
by the average of employment over all weeks of activity prior to the reference period.

Table A2: Homebase sample counts

2019 2020
All estab. potentially in base sample 48,993 (100%) 64,807 (100%)
- active in mid-February 44,294 (90.4%) 59,207 (91.4%)
- temporarily inactive in mid-February 4,699 (9.6%) 5,600 (8.6%)

In-scope estab. in base sample 38,193 (100%) 49,268 (100%)
- active in mid-February 34,757 (91.0%) 45,454 (92.3%)
- temporarily inactive in mid-February 3,436 (9.0%) 3,814 (7.7%)
Notes: The table shows counts of establishments that (i) are either active or temporarily inactive in mid-February
(top panel) of 2019 or 2020, and (ii) are in scope according to our industry (i.e. they belong to either Retail Trade,
Education & Health Services, Leisure & Hospitality, or Other Services according to the NAICS codes obtained through
Safegraph), size class (fewer than 50 workers), and geographic requirements (bottom panel).

The top panel of Table A2 reports the count of active and temporarily inactive establishments for
both the 2019 and 2020 base periods. Interestingly, in both base samples there are close to 10% of
temporarily inactive establishments, which is consistent with evidence from administrative data on tem-
porarily closed establishments. From this sample, we further drop all establishments for which we do not
have a sufficiently high quality match with Safegraph to confidently attribute a NAICS-6 industry code;
establishments with a 2-digit NAICS code different from the four sectors that we study (44-45, 61-62,
71-72 and 81);8 establishments with 50 employees or more; and establishments based in the U.S. Virgin

7
The total count of establishments in Table A1 is different from the sample count of active establishments in mid-
February in Table A2 (59,207 vs. 58,588 establishments in Table A1) because a few establishments match to a Safegraph
POI with a missing NAICS code (and are thus included in Table A2 but not in Table A1). About 1% of POIs in Safegraph
have a missing NAICS code.
8
We retain a few (about 1,000 in both 2019 and 2020) businesses for which the quality of the match to SG data is low

7
Islands, Puerto Rico, or the island of Guam. As the lower panel of Table A2, this leaves us with an
in-scope base sample of 38,193 establishments for 2019 and 49,268 establishments for 2020.
For our estimations, we weight establishments by 2-digit NAICS category × size class × geographic
area cells. There are six 2-digit NAICS categories (44-45, 61, 62, 71, 72 and 81), four establishment size
classes (1 to 4, 5 to 9, 10 to 19, and 20 to 49 employees) and thirteen geographic areas listed below in
Table A4. Thus, we have a total of 312 cells. We drop from the analysis those cells where there are too
few establishments in HB compared to the QCEW. In the 2020 base sample, for instance, we retain 296
industry-size-region cells i. The average number of HB establishments per cell i is 169, the median is 80,
the 5th percentile is 11 and the 95th percentile is 640. The smallest cell i in the base sample of 2020
is NAICS 61 of size 5 to 9 employees in the Pacific region excluding California (that is to say Alaska,
Hawaii, Oregon, Washington). This cell contains 7 establishments and its QCEW-HB weight ωi is equal
to 156. The largest cell is NAICS 72 of size 10 to 19 employees in the State of California; it contains
1,993 establishments and its QCEW-HB weight ωi is equal to 11 (i.e., HB covers almost 10 percent of all
establishments in this cell).

A.5 Benchmarking establishment counts and employment to the QCEW

Tables A3a and A3b display the count and distribution of establishments for each 2-digit NAICS categories
in our 2020 base sample, and compares them with the corresponding data from the Quarterly Census
of Employment and Wages (QCEW). Two features are noteworthy. First, the HB data is less skewed
towards very small establishments than the QCEW. This is especially true in NAICS 61, 62, and 71. A
large part of this difference is explained by the (very) large establishment counts in size class 1–4 of the
QCEW. For instance, when we compare establishment counts in size class 1–4 of the QCEW with those
from the Business Dynamics Statistics (BDS), we find enormous differences: for NAICS 61, 62 and 71,
the QCEW counts are respectively 165%, 307%, and 145% higher than those from the BDS.9 Second, the
largest industry in the HB data is by far NAICS 72 (Accommodation and Food Services): it makes up
for about half of all the businesses from the 2020 base sample. In the QCEW data, NAICS 72 accounts

but where the NAICS code obtained through the matching procedure is consistent with the industry code provided in HB
data. Namely, we keep businesses with NAICS 44-45 and HB industry code “Retail”; businesses with NAICS 611 and HB
industry code “Charities, Education & Membership”; businesses with NAICS 621, 623 or 624 and HB industry code “Health
Care and Fitness”; businesses with NAICS 71 and HB industry code “Leisure and Entertainment”; businesses with NAICS
722 and HB industry code “Food & Drink”; businesses with NAICS 811 or 812 and HB industry code “Home and Repair”;
and businesses with NAICS 813 and HB industry code “Charities, Education & Membership”.
9
We run this comparison with data from 2019, which is the most recent available data for the BDS. Barnatchez et al.
[2017] provide a detailed analysis of the discrepancies between these establishment counts. For establishment size 1–4 in
NAICS 62 (Health care and social assistance) in 2019, the magnitude of the discrepancy is remarkable: 376,026 establishments
according to the BDS vs. 1,154,994 in the QCEW.

8
Table A3a: Establishment counts and employment in HB and the QCEW

NAICS 44-45 - Retail Trade
HB data QCEW data
Estab. Workers Estab. Workers
% % % %
# % # % # #
all small all small
1–4 3,706 30.4 10,631 10.6 474,656 45.4 48.2 856,574 5.6 11.5
5–9 4,763 39.1 29,809 29.8 245,749 23.5 25.0 1,657,230 10.9 22.3
10–19 2,796 22.9 34,968 35.0 178,365 17.1 18.1 2,377,561 15.6 32.0
20–49 928 7.6 24,504 24.5 85,740 8.2 8.7 2,548,858 16.7 34.3
50–99 0 0.0 0 0.0 31,076 3.0 – 2,159,731 14.2 –
100+ 0 0.0 0 0.0 29,303 2.8 – 5,658,339 37.1 –
Total 12,193 100 99,912 100 1,044,889 100 – 15,258,293 100 –

NAICS 61 - Educational Services
HB data QCEW data
Estab. Workers Estab. Workers
% % % %
# % # % # #
all small all small
1–4 149 14.2 408 3.2 73,977 56.4 60.9 94,839 3.1 11.2
5–9 342 32.6 2,272 18.0 18,558 14.1 15.3 124,591 4.1 14.7
10–19 372 35.5 4,707 37.4 15,674 11.9 12.9 213,644 7.1 25.3
20–49 186 17.7 5,209 41.4 13,351 10.2 11.0 412,494 13.7 48.8
50–99 0 0.0 0 0.0 5,331 4.1 – 368,226 12.2 –
100+ 0 0.0 0 0.0 4,299 3.3 – 1,806,533 59.8 –
Total 1,049 100 12,596 100 131,190 100.0 – 3,020,327 100.0 –

NAICS 62 - Health Care and Social Assistance
HB data QCEW data
Estab. Workers Estab. Workers
% % % %
# % # % # #
all small all small
1–4 700 18.5 2,074 5.4 1,250,472 72.6 75.4 1,499,144 7.3 20.6
5–9 1,578 41.7 10,067 26.0 184,691 10.7 11.1 1,235,153 6.1 17.0
10–19 1,035 27.3 13,400 34.6 132,454 7.7 8.0 1,788,534 8.8 24.6
20–49 473 12.5 13,216 34.1 90,702 5.3 5.5 2,739,494 13.4 37.7
50–99 0 0.0 0 0.0 33,377 1.9 – 2,330,236 11.4 –
100+ 0 0.0 0 0.0 30,056 1.7 – 10,810,686 53.0 –
Total 3,786 100 38,757 100 1,721,752 100.0 – 20,403,247 100.0 –
Notes: HB and QCEW data for February 2020. The columns titled “#” report the number of establishments by class size, and
employment by establishment class size. In the “HB data” panels, the columns titled “%” show the distribution of establish-
ments by class size and distribution of employment by establishment class size. In the “QCEW data” panels, the columns titled
“% all” show the distribution of establishments by class size and distribution of employment by establishment class size, and
the columns titled “% small” show the distribution among small establishments (establishments with fewer than 50 workers).

9
Table A3b: Establishment counts and employment in HB and the QCEW

NAICS 71 - Arts, Entertainment, and Recreation
HB data QCEW data
Estab. Workers Estab. Workers
% % % %
# % # % # #
all small all small
1–4 442 15.0 1,277 3.7 97,592 61.5 65.1 109,419 4.9 11.9
5–9 953 32.4 6,000 17.4 20,835 13.1 13.9 138,878 6.2 15.2
10–19 983 33.4 12,328 35.8 17,875 11.3 11.9 245,345 10.9 26.8
20–49 567 19.3 14,847 43.1 13,666 8.6 9.1 422,463 18.8 46.1
50–99 0 0.0 0 0.0 5,181 3.3 – 353,624 15.8 –
100+ 0 0.0 0 0.0 3,432 2.2 – 971,527 43.3 –
Total 2,945 100 34,453 100 158,581 100.0 – 2,241,256 100.0 –

NAICS 72 - Accommodation and Food Services
HB data QCEW data
Estab. Workers Estab. Workers
% % % %
# % # % # #
all small all small
1–4 3,424 13.3 9,893 3.2 232,321 31.6 34.0 365,553 2.7 4.3
5–9 8,332 32.3 55,156 17.8 124,265 16.9 18.2 848,051 6.3 10.0
10–19 9,608 37.3 125,595 40.5 160,482 21.8 23.5 2,267,987 16.8 26.8
20–49 4,399 17.1 119,681 38.6 166,507 22.7 24.4 4,972,904 36.9 58.8
50–99 0 0.0 0 0.0 39,886 5.4 – 2,629,706 19.5 –
100+ 0 0.0 0 0.0 11,309 1.5 – 2,391,440 17.7 –
Total 25,763 100 310,325 100 734,770 100.0 – 13,475,641 100.0 –

NAICS 81 - Other Services (except Public Administration)
HB data QCEW data
Estab. Workers Estab. Workers
% % % %
# % # % # #
all small all small
1–4 1,016 28.8 2,728 9.3 594,786 72.6 73.4 848,552 19.2 25.6
5–9 1,444 40.9 9,228 31.3 123,871 15.1 15.3 811,239 18.3 24.5
10–19 797 22.6 10,141 34.4 63,202 7.7 7.8 834,213 18.8 25.2
20–49 275 7.8 7,388 25.1 28,144 3.4 3.5 815,662 18.4 24.6
50–99 0 0.0 0 0.0 6,103 0.7 – 414,017 9.3 –
100+ 0 0.0 0 0.0 3,560 0.4 – 706,529 15.9 –
Total 3,532 100 29,485 100 819,666 100.0 – 4,430,212 100.0 –
Notes: HB and QCEW data for February 2020. The columns titled “#” report the number of establishments by class size, and
employment by establishment class size. In the “HB data” panels, the columns titled “%” show the distribution of establish-
ments by class size and distribution of employment by establishment class size. In the “QCEW data” panels, the columns titled
“% all” show the distribution of establishments by class size and distribution of employment by establishment class size, and
the columns titled “% small” show the distribution among small establishments (establishments with fewer than 50 workers).

10
Table A4: Geographic distribution of establishments in the QCEW and HB data

2019 2020
HB QCEW HB QCEW
base sample small estab. base sample small estab.
# % % # % %
Alaska, Hawaii, Oregon, Washington 2,015 5.3 4.9 2,527 5.1 4.9
California 6,218 16.3 21.4 7,942 16.2 21.7
Arizona, Colorado, Idaho, Montana, Nevada,
3,377 8.8 6.1 4,388 8.9 6.2
New Mexico, Utah, Wyoming
Iowa, Kansas, Minnesota, Missouri,
2,352 6.2 6.5 3,092 6.3 6.5
North Dakota, Nebraska, South Dakota
Illinois, Indiana, Michigan, Ohio,
4,241 11.1 11.3 5,729 11.6 11.3
Wisconsin
Texas 3,585 9.4 6.4 4,711 9.6 6.4
Alabama, Arkansas, Kentucky, Louisiana,
3,000 7.9 7.6 3,913 7.9 7.6
Mississippi, Oklahoma, Tennessee
Connecticut, Massachusetts, Maine,
1,202 3.1 5.7 1,553 3.2 5.6
New Hampshire, Rhode Island, Vermont
New York 1,534 4.0 6.2 1,950 4.0 6.0
Pennsylvania, New Jersey, Delaware 2,066 5.4 6.6 2,451 5.0 6.5
District of Columbia, Maryland, Virginia 1,729 4.5 5.4 2,124 4.3 5.3
Georgia, North Carolina, South Carolina 3,533 9.3 6.0 4,573 9.3 6.1
Florida 3,341 8.7 5.9 4,315 8.8 5.9
Total 38,193 100 100 49,268 100 100
Notes: HB and QCEW data for February 2019 and 2020. The columns titled “#” report the number of establishments in HB data.
The columns titled “%” shows the distribution of establishments in HB data and according to QCEW data.

for 15% of establishments with fewer than 50 workers in the sectors of our analysis. Within NAICS 72,
the distribution across size classes in HB is, again, less concentrated towards small businesses compared
to QCEW data. In terms of employment, NAICS 72 accounts for an even larger share of the HB sample,
namely 60% of employment in February 2020.
Table A4 reports the distribution of businesses with respect to U.S. states or groups of states. In
grouping states together, we attempt to strike a balance between having 2-digit NAICS × size × region
cells with sufficiently many HB establishments (as compared with QCEW data), while preserving some
geographic variation. We group together states that are geographically close and, as it turns out, that
have experienced similar employment outcomes during the crisis. As shown in Table A4, the geography
distribution of HB data in both 2019 and 2020 is similar to the distribution of establishments according
to the QCEW.
Figure A3 reports average employment by establishment within each establishment class size in HB
data and in the QCEW in 2020. Establishments tend to be slightly larger (in employment terms) in
the QCEW data, except in the first class size where establishments are always larger in HB data.10 But
10
Again, this difference could be related to the larger establishments counts of the QCEW in size class 1–4. According to

11
Figure A3: Average employment by establishment by establishment size

Retail trade Educational services Health care and social assistance
35
30
25
20
15
10
5
0
1−4 5−9 1−4 5−9 1−4 5−9
10−19 20−49 10−19 20−49 10−19 20−49
Arts, entertainment, and recreation Accommodation and food services Other services (except public admin.)
12 35
30
25
20
15
Employment per establishment

10
5
0
1−4 5−9 1−4 5−9 1−4 5−9
10−19 20−49 10−19 20−49 10−19 20−49

QCEW data HB data

Notes: Average employment by establishment size class for NAICS-2 industries in HB data and in QCEW data for February 2020.
overall, the data match closely.

B Safegraph and Facebook data

B.1 Safegraph

We use Safegraph (SG) data on business locations, called Points of Interest (POIs) to attribute NAICS
industry codes to HB establishments. SG contains information on more than 8 million POIs in the U.S.
A POI is defined as a location where individuals can spend time or money. For a subset of POIs, SG
reports anonymized visits at daily, weekly, and monthly frequency based on information from cell phone
devices. The information in Safegraph is organized in three main datasets.

• Core Places contains basic information for every establishment including name, address, GPS
coordinates (lat/lon), NAICS industry code, brand, etc. This is the main frame based on which
Safegraph builds the other datasets.

• Patterns contains data for a subset of establishments including visit counts, visit duration, and
mapping of visitors to their home Census block group.

• Geometry contains spatial hierarchy information for a subset of establishments. This information
is important to understand and qualify the accuracy of the visits data.

We note that the Core Places files contain separate opened on and closed on fields that are supposed to
carry the opening date of a new POI, respectively the closing date for a permanently closed establishment.
However, as Safegraph acknowledges in its documentation, coverage and accuracy of this information is
far from perfect. We confirm this through the following checks: (i) there is bunching of opening and
closing in particular months, (ii) we find discrepancies between the closed on field and both our HB
data and Google’s permanently closed indicator, (iii) the rates of business openings and closings implied
by SG’s opened on and closed on fields are too low compared to BED data. We therefore do not rely
on SG’s opened on and closed on variables.

Description and definitions. Each POI is identified by a unique persistent safegraph place id.
This is our main unit of observation when working with SG data. In some cases, which are rare in
Barnatchez et al. [2017], one reason why these counts are larger is that QCEW includes some non-employers in establishments
counts, which would bias downward the ratio between total employment and number of establishments. In BDS data, average
employment per establishment in size class 1–4 is 2.4 workers (vs. 1.4 in the QCEW), which is very close to the HB data in
Figure A3.

13
our matched dataset, a POI may appear under multiple safegraph place id’s. This happens when
SG changes details about some time-invariant attribute of the POI, such as the address or the NAICS
code. We identify duplicated safegraph place id’s through a deduplication procedure described in the
paragraph “Deduplication of POIs” below. We keep the time-invariant attributes provided in the most
recent release of the Core Places data for POIs that are duplicate of each other. We combine visits data
to duplicated POIs by adding them up.

Industry codes and visits. Except for a handful of POIs (about 1% of the universe of SG POIs),
each safegraph place id comes with a 6-digit industry NAICS code that SG attributes based on an
algorithm. For details about SG’s methods, see this Documentation.
About 80-85% of safegraph place id’s come with information on visits. Safegraph constructs visits
data by attributing cell-phone pings (visits) to a POI’s polygon.11 In Section E, we use a weekly aggregate
of visits, which is provided in the Weekly Patterns file, to run several checks on selection into Homebase, to
compare the dynamics of Homebase employment with visits, and to contrast the visits data of continuing
establishments vs. entry and exit churners in Homebase. SG weekly visits are also available in a “bucketed
dwell time” format, which yields similar results to the results presented in those sections.

Deduplication of POIs. Our algorithm to identify and deduplicate Safegraph POIs is as follows:

1. Find all POIs that have the same location name (normalized using Step 1a of our matching algorithm
for SG data; see Section C) and same GPS coordinates rounded up to one decimal place. This defines
sets of potential duplicate POIs.

2. Within each set of potential duplicates, perform all pairwise comparisons to identify subsets of POIs
that are duplicates of each other. Specifically:

(a) Given two POIs that belong to the same set, compute the geographic distance between them
using GPS coordinates, and the string distance between their street addresses (normalized
using Step 1b of our SG matching algorithm; see Section C) concatenated with the 5-digit zip
code. We use Levenshtein distance normalized by the length of the longest string to define
string distance. If either the geographic distance is less than 250 meters or the string distance
is under 0.250, we tag the two POIs as duplicates of each other.

11
Safegraph’s documentation provides information about spacial hierarchy for each polygon that is important for visit
attribution: see Places Manual as well as this blog post for details.

14
(b) Form subsets of all POIs that are direct duplicates of each other. Indirect duplicates, as
opposed to direct duplicates, refer to instances such as this one: A is a duplicate of B but
not a duplicate of C, but B is a duplicate of C, making A and C indirect duplicates of each
other. A and B could be included in the same subset while C is left aside. Alternatively, A
could be left aside while B and C are included in the same subset. In such instances (which
are very rare, as we explain below) we break the tie by assigning B to the closest POI (A or
C) as measured by the geographic and Levenshtein distances.

3. For POIs that are duplicates of each other, we construct a dedup safegraph place id to over-
write their time-invariant attributes (address, zip code, and most importantly NAICS codes)
with those of the dedup safegraph place id. We set dedup safegraph place id equal to the
safegraph place id of the POI that (i) was seen most recently in the Core place files and (ii) has
a missing closed on field (since otherwise SG will no longer attribute visits to this POI).

4. In our deduplicated SG data, visits for a POI that has duplicates in the raw SG data are taken to
be the sum of the POI’s own visits and visits at its duplicates POIs.

In Step 1 of our deduplication algorithm, we have 1,105,224 POIs (about 15% of SG Core place POIs)
that belong to a set with more than one POI. 70% of these sets contain exactly two POIs and another
15% have only three POIs. The largest set contains 281 POIs. After Step 2, 476,770 POIs belong to a
subset with more than one POI, meaning that roughly 50% of POIs from Step 1 have been found to have
no duplicate. There are 197,280 subsets with more than one POI, 80 percent of which contain exactly
two POIs and 10 percent of which contain exactly three POIs.

B.2 Facebook

We use information on Facebook (FB) posting activity by establishments in HB data to estimate new
business openings and closings, as described in Section D. The FB data comes from CrowdTangle (https:
//www.crowdtangle.com/), FB’s research tool to analyze social media activity.

Description and definitions. We use Google to search for a Facebook URL address for each HB
establishment. A Facebook URL can be linked to a unique facebookid, which we can then use to retrieve
information on FB posts. However, there are some caveats:

• In some cases, we obtain the same Facebook URL for several establishments. It so happens when
different establishments that belong to the same company use and manage the same Facebook

15
page. We drop these establishments from the analysis of FB posts at a later stage, upon getting
information on posting activity;

• If the Facebook account is deactivated by the user, CrowdTangle no longer tracks the page’s activity.
When it so happens, we cannot attribute a facebookid to the establishment.

We access CrowdTangle information using two main datasets:

• Leaderboard, which contains information for every uploaded account, including a user name, a
unique facebookid, page growth, etc.

• Historical data, which contains posting data for a subset of facebookid from the Leaderboard
dataset.

Historical posts. CrowdTangle tracks public content, including the date when a users posts some
content on Facebook, and from which facebookid the content is posted. We obtain historical posting
data of all establishments with a valid Facebook URL in our sample. We aggregate counts of FB posts to
the weekly frequency. We define a HB establishment as being an active user of FB if its posting history
averages at least one post per week during the weeks when the establishment is also active in HB.

C Matching procedures

We augment HB’s establishment records with information from Google and then match the records with
SG’s POIs based on name and geography. In addition, for the purpose of determining birth and death,
we match the data with information from CrowdTangle. Below we describe both match algorithms and
provide basic match statistics.

C.1 Matching with Safegraph

To match HB locations to Safegraph, we consider the entire catalog of SG POIs that ever appear in the
Core place files between March 2020 and November 2022. The total number of SG POIs is 8,537,035 –
out of which we remove duplicates, as described in Step 1 below. Our algorithm to match HB locations
is as follows:

1. Pre-treat the data by cleaning and standardizing names, and deduplicate SG data:

16
(a) Clean company and location names in HB, and location names in Safegraph and Google
by: (i) removing company titles, such as “inc”, “incorporated”, “corp”, “corporation”, “llc”,
(ii) removing “and” and “the”, and (iii) removing any spaces and keeping only numeric and
alphabetic characters.

(b) Clean addresses in HB, Safegraph and Google using Stata’s stnd address command. Then
standardize addresses and city names by removing any spaces and keeping only numeric and
alphabetic characters.

2. Merge or match HB data:

(a) Merge/Match is performed using 3 possible names for the establishment: HB location name,
HB company name, and Google name retrieved using the Google place identifier.

(b) Try to merge using each of the name (sequentially in this order: HB location name, HB com-
pany name, and Google name) combined with the following information (again, sequentially
in this order): (i) Latitude and longitude (rounded up to 3 decimal places), (ii) Address and
zip code, (iii) Address and city, (iv) Address and State, (v) Address only.
(At each level of the merge, we only keep the unique merges: i.e., a HB establishment gets
linked to a unique Safegraph POI. We discard merges when a HB establishment merges to
more than one Safegraph POI.)

(c) Try to match using each of the names and the geographical information as described in the
previous step. A matching score is assigned to each pair of HB establishment and Safegraph
POI, representing how similar their names are. We consider matching to be successful if the
HB establishment and the Safegraph POI have (1) the same geographical information, and (2)
a matching score of 80 or higher.

(d) Try to merge using each of the name and broader geographical information (sequentially in
this order): (i) Zip code, (ii) City, (iii) State.

(e) Try to match using each of the name and broader geographical information as described in the
previous step.

Table C1 presents the outcomes of the algorithm for the pooled 2019 and 2020-2021 data. 45% of the
sample is made up of locations that we merge to SG based on names and GPS coordinates. Table C1 also
shows that there is some variation by sector, in that the high-level merge/matches (those based on the

17
name combined with either GPS coordinates or street address) are likely to be in Leisure & Hospitality
while the low-quality matches (those based on the name combined with either city or state) are more
likely to be in Retail Trade.

Table C1: Results of matching HB establishment to Safegraph

Sample NAICS sector (%)
# % 44-45 61-62 71-72 81
Merge on name and GPS coordinates 46,925 44.8 22.3 7.8 63.9 6.1
Merge on name and address 3,221 3.1 23.8 13.2 56.9 6.2
Match on name and address 12,772 12.2 21.8 12.0 59.7 6.5
Merge on name and zip code 6,488 6.2 24.9 8.7 59.9 6.6
Merge on name and city 2,828 2.7 28.4 12.6 51.3 7.8
Merge on name and state 5,607 5.4 31.5 14.6 43.2 10.7
Match on name and zip code 2,147 2.1 26.5 17.8 44.5 11.2
Match on name and city 5,044 4.8 31.7 21.6 33.5 13.2
Match on name and state 17,644 16.9 37.2 18.2 31.6 12.9
Others 2,039 2.0 34.4 15.1 45.8 4.7
Total 104,715 100 26.4 11.8 53.8 8.1
Notes: The table reports counts (#) and distribution (%) of HB locations in the 2019 base sample 2020 base sample,
and all new entrants (entering after mid-February 2019 but no later than end of November 2021) across the outcomes
of the algorithm for matching HB locations to Safegraph, and distribution across sectors for the base sample and new
entrants sample combined. The last outcome category (“Others”) refers to HB locations that match to Safegraph with
a low quality but have a SG NAICS code that matches the industry code provided in the raw HB data; see Footnote 8
for details. The four sectors are Retail Trade (NAICS 44-45), Education and Health Services (NAICS 61-62), Leisure &
Hospitality (NAICS 71-72), and Other Services (NAICS 81).

C.2 Matching with CrowdTangle

Our algorithm to match HB establishments to CrowdTangle is as follows:

1. Pre-treat the data by cleaning and standardizing names:

(a) Clean company and location names in HB, and location names in Google by: (i) removing
company titles, such as “inc”, “incorporated”, “corp”, “corporation”, “llc”, (ii) removing “and”
and “the”, and (iii) removing any spaces and keeping only numeric and alphabetic characters.

(b) Clean addresses in HB and Google using Stata’s stnd address command. Then standardize
addresses and city names by removing any spaces and keeping only numeric and alphabetic
characters.

2. Use Google to find a Facebook address for each HB location, and clean the Facebook address by

(a) Removing links that are posted in public pages or groups, such as “event”, “careers”, “places”,
“marketplace”, etc.;

18
(b) Extracting the Facebook page name from the Facebook address, e.g. extract the end part of
http://www.facebook.com/page name.

3. Batch upload locations with a valid Facebook URL into CrowdTangle.

4. Merge or match the resulting CrowdTangle dataset to HB data. Specifically,

(a) From CrowdTangle, we obtain the “Leaderboard” dataset (Section B.2), which contains a
unique user name and facebookid for each Facebook account;

(b) Try to merge the FB page name obtained in Step 2b to CrowdTangle’s user name ;

Table C2 presents the outcomes of matching HB establishments to CrowdTangle for the pooled 2019 and
2020-2021 data. Most HB establishments can be linked to a valid Facebook URL: more than 90% for both
new entrants and permanent exits during this period. We manage to upload almost 30% of new entrants
and permanent exits to CrowdTangle, either by merging or matching on page name. Note, however, that
we only use a small subset of these establishments in our analysis of FB posts (see Tables D1 and D2)
because many establishments are not actively posting on FB.

Table C2: Results of matching HB establishment to CrowdTangle

New entrants Exits without return
All estabs. 62,926 (100%) 41,545 (100%)
Establishments with a Facebook URL 57,491 (91.4%) 37,970 (91.4%)
Establishments uploaded to CrowdTangle 17,698 (28.1%) 10,632 (25.6%)
Notes: The table shows counts of establishments that newly enter HB after mid-February 2019 and no later than
end of November 2021 (new entrants), and establishments that exit HB with no return to activity before the end of
November 2021 (exits without return). For these establishments, the table reports counts of establishments with a
valid Facebook URL and establishments that successfully upload to CrowdTangle.

D Closings and openings

Recall our employment estimator from (F.1)

" # $
A O
i ωi êi,ti,t + êi,ti,t
!t = E
E !t−1 × # $, (D.1)
" Ai,t Ci,t
ω
i i ê i,t−1 + ê i,t−1

19
A
where êi,ti,t denotes week t employment of the set of establishments Ai,t that were active in both week
C
i,t
t − 1 and week t; êi,t−1 denotes week t − 1 employment of the set of establishments Ci,t that are closing in
O
week t; and êi,ti,t denotes week t employment of the set of establishments Oi,t that are opening in week
C
i,t T
i,t D
i,t O R B
t. Using êi,t−1 = êi,t−1 + êi,t−1 and êi,ti,t = êi,ti,t + êi,ti,t , this estimator can be written as

" # $
Ai,t Ri,t " Bi,t
ω ê + ê
! ! i i i,t i,t
! i ωi êi,t
Et = Et−1 × " # $ + Et−1 × " # $, (D.2)
Ai,t Ti,t Di,t Ai,t Ti,t Di,t
i ω i ê i,t−1 + ê i,t−1 + ê i,t−1 i ω i ê i,t−1 + ê i,t−1 + ê i,t−1

Ti,t
where êi,t−1 denotes week t − 1 employment of the set of establishments Ti,t that closed temporarily in
D i,t
week t; êi,t−1 denotes week t−1 employment of the set of establishments Di,t that closed permanently (i.e.
R
deaths) in week t; êi,ti,t denotes week t employment of the set of establishments Ri,t that reopen in week t
B
after being temporarily closed; and êi,ti,t denotes week t employment of the set of establishments Bi,t that
newly open (births) in week t. Notice that establishments that cease to be active in week t and become
active again at some later date t + n belong to the set of temporarily closed establishments Ti,t ⊆ Ci,t in
week t, and later on they are added to the set of re-opening establishments Ri,t+n ⊆ Oi,t+n .12 The key
challenge facing the implementation of (D.1) is that sample churn prevents us from directly observing
Di,t and Bi,t . That is, among the establishments that cease to be active in some week t and are not part
of Ti,t , some continue to operate outside of HB and must therefore not be included in Di,t , Likewise,
among establishments that open during week t and do not belong to Ri,t , those establishments that
were operating before entry into HB in week t must not be included in Bi,t . The next sections present
our methodology to address this challenge. Section D.1 explains how we use Google and Facebook to
disentangle sample churn from (permanent) closings and (new) openings, while Section D.2 explains how
i,t D B
we subsequently incorporate this information for the estimation of êi,t−1 and êi,ti,t .

D.1 Google/Facebook approach to determine business closings and openings

D.1.1 Permanent closings. We use information from Google and Facebook to estimate whether
establishments that exit HB and do not re-enter before the end of the sample are closed or continue to
operate outside of HB. We match HB establishments to Google Places using their API. At the same time,

12
The maintained assumption is that these establishments do not continue to operate while temporarily abstaining from
using the HB service. Notice that as t gets closer to the end of the sample, this approach implies that we miss some temporary
closings. This is not a problem as long as the closing probability (subsection D.1) correctly identifies these establishments as
being closed, and since the employment estimator in Equation (D.1) does not require us to distinguish between temporary
and permanent closings. By the same token, the definition of temporarily closed establishments implies that past real-time
employment estimates are subjected to revisions upon using more recent data to estimate equation (D.1). In practice,
revisions of our estimates have been unimportant.

20
we use Google to search for a Facebook address for each HB establishment based on name and address
and then use CrowdTangle, Facebook’s research utility, to extract the history of posts for each available
Facebook address. We then proceed in 3 steps:

1. For HB establishments that can be matched to Google Places, we identify establishment " that exits
HB in week as closed if Google attributes a “closed” indicator.

2. For HB establishments that either cannot be matched to Google Places or are not flagged as Google-
closed, we retain all establishments with a unique Facebook address that average at least one post
per week during the weeks when they are active in HB. Then, for any establishment " that exits
during week t and satisfies these criteria,

(a) if Facebook posts continue for more than 4 weeks after the establishment exits HB, we identify
it as an establishment that continues to operate outside of HB;

(b) otherwise, we identify establishment " as a business closing.

3. For all other exiting establishments that cannot be matched to either Google Places or Facebook,
we identify them as closed with probability equal to the proportion of closings estimated in Step 2.
We compute this proportion separately by quarter q for each of the four sectors i.

Concretely, the procedure means that for all establishments " belonging to cell i that exit HB permanently
in week t of quarter q, we define:


 = 1 if " is Google-closed, or not Google-closed but matched to FB and closed





 based on FB posts




 = 0 if " is not Google-closed but matched to FB and operating outside of HB
p̂(D!,t |exit!,t )

 based on FB posts





 = cell-i probability (∈ [0, 1]) of closing conditional on exit from HB in the





current quarter q

p̂(" ∈ D!,t |exit!,t ) is a key ingredient of our estimation approach, summarized in Table D3 at the end
of this section. Notice that we rely on the set of establishments that are either Google-closed or matched
to Facebook to compute cell-i’s probability of closing conditional on exit from HB. When compared
to establishments that permanently exit HB and are neither Google-closed nor matched to Facebook,
we find no evidence of systematic differences between the two sets of establishments: they share similar

21
industry-size-region distributions and have similar employment and hours dynamics while active in HB.13
Since we have few permanent exits from HB that are either Google-closed or matched to Facebook in a
given week, we pool these establishments together by quarter to compute cell-i’s probability of closing
conditional on exit from HB.
Table D1: Google / Facebook procedure to determine closings

2019 2020
Exiting estab. that do not reopen 13,289 (100%) 28,256 (100%)
- Google closed 3,011 (22.7%) 4,031 (14.3%)
- Not Google-closed and matched to FB 2,197 (16.5%) 6,826 (24.2%)
- Closed from FB posts 419 (3.2%) 2,000 (7.1%)
Proportion estimated as closed 37.4% 39.0%
Notes: For 2019, exiting establishments that do not reopen refer to the set of establishments included in the base
sample or new entrants sample of 2019 that cease to be active at some point before mid-February 2020 and do not
return to activity in HB by the end of the sample (currently end of November 2021). For 2020, exiting establish-
ments that do not reopen refer to the set of establishments included in the base sample or new entrants sample of
2020 that cease to be active and do not return to activity in HB by the end of the sample. Establishments that
are Google-closed are establishments that can be matched to Google Places and are flagged as either “temporary
closed” or “permanently closed”. Establishments that can be matched to Facebook are establishments that can be
matched to Facebook pages, do not belong to multi-establishment company id’s and post on average at least once
a week while active in HB. Establishments matched to Facebook are flagged as “closed” if their Facebook posts do
not continue for more than 4 weeks after exit from HB. The proportion estimated as closed is the proportion of
Google closed plus the proportion of establishments that are either identified (if they are not Google-closed and are
included in the FB-HB matched sample) or estimated as closed based on the FB estimation.

Table D1 provides statistics on the procedure to deal with permanent closings. Consider for instance
the 13,289 establishments that exit HB in 2019 without return before the end of the sample (currently end
of November 2021). 22.7% of these establishments are estimated as closed because they receive a “closed”
flag when matched to Google Places (Step 1 of the above procedure). The remaining establishments that
can be matched to Facebook and are actively posting while being active in HB account for 16.5% of all
permanent exists in 2019. Our procedure flags 19.1% of these establishments (419/2,197) as “closed”
because their Facebook posts stop within less than 4 weeks after exit from HB. For 60.8% of exiting
establishments in 2019 (100% minus 22.7+16.5%), the Google/Facebook approach is not applicable (Step
3 of the above procedure), and we estimated that 19.1% of these establishments are closed (based on the
proportion of “closed” among establishments matched to FB). Put together, these numbers imply that
the proportion estimated as closed among all 2019 exits from HB is 37.4%.14
13
For the most part, it seems that these establishments cannot be matched to either Google Places or Facebook due to
idiosyncratic details in company names or addresses that our algorithm fails to take into account (see Section C.2), or due
to inexistent or erratic posting behaviors on Facebook that prevent us from relying on this source of information (as we use
Facebook only for establishments that average at least one post per week while in HB).
14
Table D1 in addition shows that we identify 13.4% of exiting establishments in 2019 as sample churn (! ∈
/ Di,t ) because

22
Table D1 shows that in 2020 our Google / Facebook approach yields a similar proportion of closings
conditional on exit from HB, namely 39.0% in 2020 against 37.4% in 2019. There are fewer “temporary
closed” or “permanently closed” based on the information obtained from Google places, and a higher
proportion of establishments matched to Facebook, among which we flag 29.3% (2,000/6,826) of the
establishments as closed based on their Facebook posts.

D.1.2 New openings. For establishments that become active in HB for the first time after the
reference week (i.e. new entries, not reopenings), we proceed similarly as for closings. One difference
is that we only exploit the information coming from Facebook posts since Google does not contain an
indicator for new openings as it does for closings. We use Google to search for a Facebook address for all
newly entering HB establishments based on name and address and then use CrowdTangle to extract the
history of posts for each available Facebook address. We then proceed in 2 steps:

1. We retain all newly entering establishments with a unique Facebook address that average at least
one post per week during the weeks when they are active in HB. Then, for any establishment " that
enters HB in week t and satisfies these criteria,

(a) if Facebook posts start before the base period (mid-February of the respective year), we identify
establishment " as one that operated already prior to entering HB;

(b) otherwise, we identify establishment " as a new opening.

2. For all other newly entering establishments, we identify them as new openings with probability
equal to the proportion of new openings estimated in Step 1.

For all establishments " that become active in cell i for the first time in week t, this procedure yields:


 = 1 if " is matched to FB and new based on FB posts




 = 0 if " is matched to FB and operating prior to entry based on FB posts
p̂(B!,t |entry!,t )

 = cell-i probability (∈ [0, 1]) of birth conditional on entry into HB in the




 current quarter q

As for closings, we find little evidence that those establishments that can be matched to Facebook
and post actively while being active in HB are a selected sample among the set of all new entrants in
HB, in the sense that they have similar industry-size-region distributions as well as similar employment
and hours dynamics while active in HB.
they either do not receive a “closed” indicator from Google, and they post actively in Facebook and continue to do so for
more than 4 weeks after exit from HB.

23
Table D2 provides statistics on the procedure for new openings. Of the 25,149 establishments that
newly entered in 2019, 31.7% can be matched to Facebook and post regularly while being active in HB.
Then, we estimate the proportion of new establishments among newly entering HB establishments to be
equal to 9.1% (727 divided by 7,960 establishments). Notice that the remaining establishments matched
to FB that were posting before the base period (7,233 establishments, or 28.8% of the sample of newly
entrants in 2019) are considered to be part of sample churn. In 2020, the Facebook approach can be
implemented for a lower portion of newly-entering establishments (25.3%), and the proportion that is
estimated as new conditional on entry in HB (7.5%) is very similar to that for 2019.

Table D2: Facebook procedure to determine new openings

2019 2020
Newly entering establishments 25,149 (100%) 37,777 (100%)
- Matched to FB 7,969 (31.7%) 9,566 (25.3%)
- Newly opened from FB posts 726 (2.9%) 718 (1.9%)
Proportion estimated as new 9.1% 7.5%
Notes: For 2019, newly entering establishments refer to the set of establishments that are active in HB for
the first time after mid-February 2019 but no later than mid-February 2020. For 2020, newly entering estab-
lishments refer to the set of establishments that are active in HB for the first time after mid-February 2020
but no later than late-November 2021. Establishments matched to Facebook includes all establishments that
can be matched to unique Facebook pages and post on average at least once a week while active in HB. Estab-
lishments matched to Facebook are flagged as “new” if their Facebook posts do not start before mid-February
of the corresponding week. The proportion estimated as new is the number of newly opened establishments
based on FB posts divided by the number of newly entering, FB-HB matched establishments.

D.2 Incorporating closings and openings into our estimation

Having described the identification of closings and reopenings / new openings, we now explain how this
information is incorporated into our employment estimator. For establishments identified as temporary
i,t T R
closings, respectively reopenings, we can directly measure êi,t−1 and êi,ti,t . To estimate employment
i,t D
losses from establishments closing permanently êi,t−1 (deaths) and employment gains from newly opened
B
establishments êi,ti,t (births), we need to perform several adjustments as described in what follows.

D.2.1 Employment at establishments closing permanently. We estimate employment for per-
manent closings (death) in industry-size cell i in week t as

D
)
i,t D
êi,t−1 = θ!,t × p̂(D!,t |exit!,t ) × ê!,t−1 , (D.3)
!∈i

24
where ê!,t−1 denotes employment of exiting establishments in the week prior to exit from HB; p̂(" ∈
D is an
Di,t |exitit ) denotes the probability estimated from the above Google / Facebook approach; and θ!,t
adjustment factor that corrects for possible selection issues, namely that the survival probability of the
average HB business may differ systematically from the population survival probability of the average
small business and that survival rates conditional on exit as implied by our Google / Facebook approach
may differ systematically from survival rates of the average HB business.
D is calculated to fit the unconditional average death rate for 2019 for cell i
The adjustment factor θ!,t
in the BED/BDS. Specifically, for each quarter q, we compute the unconditional death rate in our HB
data as
D
N̂i,q
p̂HB (deathi,q ) = # $, (D.4)
1 A + N̂ A
2 N̂i,q i,q−1

D is the count of HB establishments that permanently close in quarter q as implied by our Google
where N̂i,q
A and is the count
/ Facebook approach (i.e., establishment " is multiplied by its p̂(D!,t |exit!,t )); and N̂i,q
of HB establishments active in quarter q. We define these counts analogous to how they are defined by
A is the count of
the BLS to construct quarterly BED death rates (see Section D.4). In particular, Ni,q
D is defined as the
all establishments with positive employment in the third month of quarter q; and N̂i,q
count of establishments with positive employment in the third month of quarter q − 1 but not in the third
month of quarter q, and estimated to represent a permanent closing according to Google / Facebook.
Then, we average the resulting quarterly death rates for 2019, and compute the adjustment factors as

1"
q∈2019 p̂BED/BDS (deathi,q )
θi = 4 1 "
D
, (D.5)
4 q∈2019 p̂HB (deathi,q )

where p̂BED/BDS (deathi,q ) denotes the BED/BDS quarterly death rate for industry-size cell i in quarter
q.15 Finally, we set θ!,t D = θ D . Thus, by construction of the θ D , the adjusted HB death rates (i.e.,
i !,t
" D
!∈i θ!,t × p̂(D!,t |exit!,t )) averaged over the four quarters of 2019 matches the average BED/BDS quarterly

death rates over this period.
To allow for sufficient sample size, we compute the adjustment factor θiD for Retail Trade, Education
& Health Services, and Other Services separately for establishment size 1 to 4 and establishment size
5-10, 11-19, and 20-49 pooled together. As shown in the main text, the fit with the BED death rate is
excellent for all size classes despite this pooling. For Leisure & Hospitality, we compute θiD separately for
all size classes, both because our sample is larger and because death rates vary more substantially across
15
BED death and birth rates are available by industry but not by size class. We use information from the BDS to derive
size-adjusted BED death and birth rates; see Section D.4.

25
size classes 5-10, 11-19, and 20-49 than for the other sectors. See the main text for results.

D.2.2 Employment at newly opened establishments. Conceptually, employment gains from new
openings (births) in industry-size cell i in week t can be estimated similarly as employment losses from
permanent closings; i.e.
B
)
B
êi,ti,t = θ!,t × p̂(B!,t |entry!,t ) × ê!,t (D.6)
!∈i

where ê!,t denotes employment of entering establishments during the week of entry into HB, p̂(B!,t |entry!,t )
denotes the estimated probability of new openings (births) conditional on entry into HB obtained from
B is the adjustment factor. However, the computation of θ B is more involved than that
Facebook; and θ!,t !,t
D because entry of establishments in HB may vary in ways that are not necessarily taken into account
of θ!,t
by corresponding changes in our Google / Facebook probability of new openings, p̂(B!,t |entry!,t ).
B , consider the hypothetical case where
To illustrate this issue and motivate our strategy to compute θ!,t
Homebase samples randomly from the population such that p̂(B!,t |entry!,t ) = p(B!,t ); i.e. there are no
B
selection issues. But then, according to Equation (D.6), employment from new openings êi,ti,t in week t
"
would be entirely driven by ! p(B!,t )ê!t , which in turn is driven by the number of new entrants during
that week. This number may vary independently of the size of the sample used to run the estimation of
(D.1) as a result of, e.g., changes in HB’s efforts to attract new customers or changes in HB’s competitive
environment.16 In the present case, we can use an approach akin to inverse-probability weighting to
adjust for this issue. Suppose that multiply p̂(B!,t |entry!,t ) × ê!,t by the inverse of the ratio of new
entry A
i,t /n̂i,t
n̂
entrants to existing HB establishments (normalized by this ratio at t = 0), n̂entry . Then, employment
i,0 /n̂A
i,0
B " n̂
entry A
i,t /n̂i,t
from new openings would be calculated as êi,ti,t = ( !∈i n̂entry/n̂A )−1 × p̂(B!,t |entry!,t )× ê!,t , or equivalently
i,0 i,0
B " B
êi,ti,t = !∈i θiB × p(Bi,t ) × n̂A B n̂entry Thus êi,ti,t would be the estimated number
i,t × !,t/n̂i,t , with θi =
entry
ê / i,0 n̂A
i,0
!
of new births (p(Bi,t ) × n̂A
i,t ) times average employment of newly entering establishments ( / entry
!∈i ê!,t n̂i,t ).
In practice, we also want to adjust for possible selection issues, namely that HB establishments may
not represent a random sample and that conditional birth rates as implied by our Google / Facebook
approach may differ systematically from survival rates of the average HB establishment. We do by setting
entry A
i,t /n̂i,t
n̂
B = θB × (
θ!,t )−1 and calculating the adjustment factor θiB to fit the unconditional average birth
i n̂
entry
/n̂A
i,0 i,0
rate for 2019 for cell i in the BED. We pool class sizes in the same way as for θiD : establishments with 5
to 49 employees are pooled together for Retail Trade, Education & Health Services, and Other Services,
16
Put differently, while exits are naturally bounded by HB sample size (i.e. the total number of establishments in HB),
the upper bound for entry is theoretically the population of establishments not already covered by HB. Hence, differences
in HB’s efforts to attract new customers or changes by competing service providers may lead to larger fluctuations in entry
of new establishments.

26
while for Leisure & Hospitality we keep the four class sizes separate from each other. We compute the
unconditional birth rate of the HB data for each quarter q of 2019 as

B
N̂i,q
p̂HB (birthi,q ) = # $, (D.7)
1 A + N̂ A
2 N̂i,q i,q−1

B is the count of newly entering establishments with positive employment in the third month
where N̂i,q
of quarter q and multiplied by the estimated probability p̂(B!,t |entry!,t ). One difference with θiD is
that we compute θiB by regressing the four quarterly values of the BED/BDS birth rates for 2019,
p̂BED/BDS (birthi,q ), on the p̂HB (birthi,q )’s. We find that this approach does a better job at controlling for
changes in p̂(B!,t |entry!,t ) induced by the large swings in the number of new entrants in HB in the first
and last quarters of the year. While the regression implies that the average of the adjusted HB birth
rates will be different from the average BED quarterly birth rates for 2019, Figure 1 in the main text
shows that they are very close to each other.

D.3 Recap of our approach

Table D3 recaps and summarizes our approach. For permanent closings, which are establishments that
do not return at a later date: Each of these establishments is multiplied by the probability p̂(D!,t |exit!,t )
from our Google / Facebook approach, and multiplied by the industry-size adjustment factor for closing
θiD . The adjustment factor is calibrated to match average BED death rates in 2019. For any subsequent
period, since θiD constant, all the variations in closing rates in our estimates come from exits from HB
data and from the Google / Facebook indicator and probabilities of permanent closings. As shown by
Figure 5 in the main text, the estimates come very close to the BED/BDS death rates for 2020, despite
our assumption of constant θiD ’s. For new openings, we multiply new entrants in HB in a given week
entry A
i,t /n̂i,t
n̂
t by the Google / Facebook probability p̂(B!,t |entry!,t ), weight them by the inverse of n̂entry to take
i,0 /n̂A
i,0
account of changes in new entries in t relative to the sample size of establishments that contributed to the
estimation in week t − 1, and multiply them by the industry-size adjustment factor for openings θiB . θiB is
calibrated to BED/BDS birth rates in 2019. For any subsequent period, the source of variations in new
openings is coming from variations in the Google / Facebook indicator and probabilities and controlling
for variations in total counts of new entrants in HB data, n̂entry A
i,t , relative to n̂i,t−1 .

We now present implementation results on our approach.
For permanent closings, the probability p̂(D!,t |exit!,t ) derived from Google and Facebook is de-

27
" A O! # ! " A Ri,t Bi,t
#
êi,ti,t +êi,ti,t
i ωi
i,t
i ωi êi,t +êi,t +êi,t
!t = E
Table D3: Implementing E !t−1 × ! " A Ci,t
!t−1 × ! " A
# =E Ti,t Di,t
#
i,t i,t
i ωi êi,t−1 +êi,t−1 i ωi êi,t−1 +êi,t−1 +êi,t−1

Approach:
" active in A i,t
" A Ai,t " A
both t and t − 1 êi,t−1 = !∈i ê!,t−1 , êi,t = !∈i ê!,t

Approach:
T i,t
" T
returns in t + n êi,t−1 = !∈i ê!,t−1

" active in identifies sample churn
t − 1 but not in t
among exits from HB
Google-closed or
without return matched to FB D "
i,t D exit
before the êi,t−1 = !∈i θ!,t × p̂(D!,t |exit!,t ) × e!,t−1
end of sample
neither Google-closed
nor matched to FB

Properties: Varying over time t? Specific to estab. "? Specific to cell i?
 Yes Yes Yes



 weekly if " is Google-closed by construction of by construction of "


 or matched to FB matching at the level of " belonging to cell i
p̂(D!,t |exit!,t )

 Yes No Yes



 quarterly if " is not Google- computed at the

closed and not matched to FB sector level
D
θ!,t = θiD No No Yes
matches sector-size
BED 2019 death rates

Approach:
R " R
active in t − n êi,ti,t = !∈i ê!,t

" active in identifies sample churn
t but not in t − 1 among entries in HB
in HB data matched to FB B " B entry
for the first êi,ti,t = !∈i θ!,t × p̂(B!,t |entry!,t ) × ê!,t
time in t not matched to FB

Properties: Varying over time t? Specific to estab. "? Specific to cell i?
 Yes Yes Yes



 weekly if " is by construction of by construction of "


 matched to FB matching at the level of " belonging to cell i
p̂(B!,t |entry!,t )

 Yes No Yes



 quarterly if " is not computed at the

not matched to FB sector level
entry A
i,t /n̂i,t
n̂
B
θ!,t = θiB × ( n̂entry )−1 Yes No Yes
i,0 /n̂A
i,0
entry
through n̂i,t /n̂A
i,t computed matches sector-size
at the weekly frequency BED 2019 birth rates

Notes: “matched to FB” means that establishment ! has a unique Facebook address that averages at least one post per week during
the weeks when they are active in HB.

28
D = θ D that multiply
scribed in Table D1. The left panel of Table D4 presents the adjustment factors θ!,t i

this probability to align the unadjusted HB death rates to BED/BDS death rates in 2019. These fac-
tors show a very consistent pattern: they are larger for small than for larger establishments, meaning
that the unadjusted data would underestimate permanent closings for small and overestimate them for
larger establishments. We note, however, that across size classes the unadjusted data comes close to the
BED/BDS quarterly death rates.
For new openings, the probability p̂(B!,t |entry!,t ) is described in Table D2 and the adjustment
B , averaged over the weeks of 2019, is described in Table D4. Similarly as for closings, the
factors θ!,t
adjustment factors tend to be larger for small than for larger establishments, although not Education &
Health Services, where in addition they are lower than one for all size classes, i.e. the unadjusted birth
rates would always overestimate the BED/BDS birth rates for this sector. Figure D1 complements Table
B ’s for each sector and size class over time. Recall that the time variation is coming
D4 by showing the θ!,t
entry A
i,t /n̂i,t
n̂
from ( n̂entry )−1 . During the pandemic, we observe a significant drop in the counts of new entrants in
i,0 /n̂A
i,0
B . Although not reported
HB across all four sectors, which leads to an increase in the adjustment factors θ!,t
here, this effect is partly offset by a reduction in the Google / Facebook probability, p̂(B!,t |entry!,t ), during
this period. For smaller establishments in Leisure & Hospitality and Other Services, and to a lower extent
B are noticeably higher during 2020 compared to the other
for Retail Trade, the adjustment factors θ!,t
time periods, meaning that birth rates would be substantially underestimated without the adjustment
factors.

29
D and θ B
Table D4: Adjustment factors θ!,t !,t

D (= θ D ) for permanent closings
θ!,t B (average for 2019) for new openings
θ!,t
i
Retail Education Leisure & Other Retail Education Leisure & Other
Class size Trade & Health Hospitality Services Class size Trade & Health Hospitality Services
1–4 1.52 2.57 1.51 1.26 1–4 1.70 0.79 4.16 2.23
5–9 0.40 0.79 0.78 0.33 5–9 0.54 0.92 1.96 0.23
10–19 0.40 0.79 0.58 0.33 10–19 0.54 0.92 1.46 0.23
20–49 0.40 0.79 0.30 0.33 20–49 0.54 0.92 0.52 0.23
D B
Notes: The table reports the adjustment factors for permanent closings, θ!,t , and new openings, θ!,t , calibrated to make the quarterly
HB death and birth rates match the quarterly BED/BDS rates on average for 2019. For permanent closings, the adjustment factor are
constant over time. For new openings since the adjustment factors vary over time, the table report the average value for 2019. Ad-
justment factors are identical for establishments with 5 to 49 employees within Retail Trade, Education & Health Services, and Other
Services, since we pool class sizes together to avoid small cell issues.

B
Figure D1: Adjustment factors for new openings θ!,t

Retail trade Education and health services
25
mid−Feb mid−Feb mid−Feb mid−Feb mid−Feb mid−Feb
of 2019 of 2020 of 2021 of 2019 of 2020 of 2021
20

0
Leisure and hospitality Other services
25
mid−Feb mid−Feb mid−Feb mid−Feb mid−Feb mid−Feb
of 2019 of 2020 of 2021 of 2019 of 2020 of 2021
20

0
Apr 19
Oct 31
Feb 24 May 5 Jul 14 Sep 22 Dec 1 Feb 9 Jun 28 Sep 6 Nov 15 Jan 24 Apr 4 Jun 13 Aug 22 May 5 Jul 14 Sep 22 Dec 1 Feb 9 Apr 19 Jun 28 Sep 6 Nov 15 Jan 24 Apr 4 Jun 13 Aug 22 Oct 31
Feb 24

1−4 5−9 10−19 20−49

B , calibrated to make the quarterly HB birth rates match the quarterly BED/BDS rates on average for
Notes: Adjustment factor for θ!,t
2019. Adjustment factors are identical for establishments with 5 to 49 employees within Retail Trade, Education & Health Services,
and Other Services, since we pool class sizes together to avoid small cell issues.

30
D.4 Benchmarking to BED establishment births and deaths

The adjustments presented in the previous section and benchmarking of our HB data rely on establishment
births and death rates from the Business Employment Dynamics (BED). The BLS generates these rates by
longitudinally linking establishment records of the QCEW. The BED reports quarterly rates of business
closings and openings and business births and deaths by industry as well as employment gains and losses
associated with these events. These rates are computed using the following definitions:17

• BED openings in quarter q are establishments with positive employment in the third month of
quarter q and no employment in the third month of the previous quarter (q − 1);18

• BED closings in quarter q are establishments with zero employment in the third month of quarter
q and positive employment in the third month of the previous quarter (q − 1);

• BED births in quarter q are establishments with positive employment in the third month of quarter
q and no employment in the third month of the preceding four quarters (q − 4, q − 3, q − 2, q − 1);

• BED deaths in quarter q are establishments with positive employment in the third month of quarter
q − 1 and no employment in the third month of the subsequent four quarters (q, q + 1, q + 2, q + 3).

As of this writing, the last BED data release was in January 2022 and includes data up to the second
quarter of 2021. This means that the last available death rates are for the third quarter of 2020.

Adjustment of BED birth and death rates by size class. One important issue with comparing
our HB data with the BED is that, at the industry level, the BED does not report statistics by establish-
ment size class. This issue matters because entry and exit rates of small establishments are substantially
higher than for larger establishments. We resolve this problem by using data from the Business Dynamics
Statistics (BDS) of the U.S. Census Bureau, which contains entry and exit rates by industry and estab-
lishment size class. This data is only available annually up to 2019, and exit and entry rates are computed
somewhat differently than BED entry and exit rates.19 As Figure D2 shows, BDS and BED annual entry
17
See https://www.bls.gov/news.release/cewbd.tn.htm for details on the BED. Rates are computed by dividing flows
in quarter q by the average count of establishments in quarters q and q − 1.
18
No employment means either zero reported employment or no reported employment (e.g. because establishment appears
for the first time in that quarter).
19
BDS establishment entry and exits are defined as rates in a way similar to the BED’s annualized opening and closing
rates, defined as
• annualized openings refer to establishments with positive employment in the third month of quarter q that were not
present (or had zero employment) in the third month of quarter q − 4,
• annualized closings refer to establishments with positive employment in the third month of quarter q − 4 that are no

31
Figure D2: BDS and BED annual entry and exit rates
(a) Retail Trade
Entry rates Exit rates
18 18

15 15

12 12

9 9

6 6

3 3

0 0
1994 1998 2002 2006 2010 2014 2018 1994 1998 2002 2006 2010 2014 2018

BDS entry rate BED entry rate BDS exit rate BED exit rate

(b) Education & Health Services
Entry rates Exit rates
18 18

15 15

12 12

9 9

6 6

3 3

0 0
1994 1998 2002 2006 2010 2014 2018 1994 1998 2002 2006 2010 2014 2018

BDS entry rate BED entry rate BDS exit rate BED exit rate

15 15

12 12

9 9

6 6

3 3

0 0
1994 1998 2002 2006 2010 2014 2018 1994 1998 2002 2006 2010 2014 2018

BDS entry rate BED entry rate BDS exit rate BED exit rate

(d) Other Services
Entry rates Exit rates
18 18

15 15

12 12

9 9

6 6

3 3

0 0
1994 1998 2002 2006 2010 2014 2018 1994 1998 2002 2006 2010 2014 2018

BDS entry rate BED entry rate BDS exit rate BED exit rate

Notes: BDS (solid lines) and BED (dashed lines) annual entry and exit rates, 1994-2019.

32
Table D5: BDS conversion factors

Entry Exit
Retail Education Leisure & Other Retail Education Leisure & Other
Class size Trade & Health Hospitality Services Class size Trade & Health Hospitality Services
1–4 2.01 1.89 2.16 1.51 1–4 2.05 1.89 2.34 1.51
5–9 0.36 0.41 0.94 0.24 5–9 0.32 0.40 0.88 0.22
10–19 0.24 0.27 0.57 0.16 10–19 0.19 0.28 0.45 0.16
20–99 0.19 0.23 0.23 0.09 20–99 0.20 0.25 0.17 0.11
Notes: BDS data for 2015-2019, conversion factors by class size for entry and exit derived for Retail Trade (NAICS 44-45), Education and
Health Services (NAICS 61-62), Leisure & Hospitality (NAICS 71-72) and Other Services (NAICS 81).

and exit rates line up closely for Retail Trade and Leisure & Hospitality, but less so for Education &
Health Services (NAICS 61-62) and Other Services (NAICS 81) where BED rates are several percent
above BDS rates. While some of these differences are due to data source (Business Register for the BDS
and QCEW for the BED), the main reason for the larger BED rates in NAICS 61-62 and NAICS 81 are
industry definitions / reclassifications.20 Investigating the details behind these differences is beyond the
scope of this paper. Besides, they should be innocuous for our estimates since we adjust both entries
and exits to match the BED rates, and the discrepancy between the BED and the BDS in respectively
NAICS 61-62 and NAICS 81 seems to affect both rates by the same order of magnitude.
We use the BDS rates to construct conversion factors by industry and size class that we then apply
to BED industry rates to benchmark our HB data. Figure D3 shows that differences in entry and exit
rates across class sizes within an industry are very stable over time. We take the average over the past
5 years of available data to construct the conversion factors reported in Table D5. As the table shows,
entry and exit rates decrease with establishment size.

BED death rates for 2020. For Figure 5 of the paper comparing the (untargeted) BED birth and
deaths rates for 2020 with their HB counterparts, we need BED death rates for the four quarters of
2020. However, as of this writing, the BED death rates are available only up to 2020Q3. We compute
death rates for the fourth quarter of 2020 using an approximation proposed by Ryan Decker: deathsi,q =
closuresi,q − repoeningsi,q+1 , where repoeningsi,q = openingsi,q − birthsi,q . We can compute business
reopenings for 2020Q4 because the available BED data covers openings and births for 2021Q1. We verify

longer present (or have zero employment) in the third month of quarter q
We use these annual BED openings and closing rates to compare to the BDS annual entry and exit rates in Figure D2.
20
For NAICS 61-62, the rates do not line up as closely before 2013, primarily because the definition of this sector
in the BED was different from the definition of the (new) BDS. In particular, in 2013:Q1 the QCEW program reviewed
establishments that provide non-medical, home-based services for the elderly and persons with disabilities and classified these
establishments into services for the elderly and persons with disabilities (NAICS 624120). Many of these establishments
were previously classified in the private households industry. (BDS industry rates are typically volatile around Economic
Census years, which occur in 1997, 2002, 2007, 2012, when most of the reclassification of businesses occurs).

33
Figure D3: BDS annual entry and exit rates by size classes
(a) Retail Trade
Entry rates Exit rates
25 25

20 20

15 15

10 10

5 5

0 0
1991 1995 1999 2003 2007 2011 2015 2019 1991 1995 1999 2003 2007 2011 2015 2019

All sizes 1−4 5−9 10−19 20−99 All sizes 1−4 5−9 10−19 20−99

(b) Education & Health Services
Entry rates Exit rates
25 25

20 20

15 15

10 10

5 5

0 0
1991 1995 1999 2003 2007 2011 2015 2019 1991 1995 1999 2003 2007 2011 2015 2019

All sizes 1−4 5−9 10−19 20−99 All sizes 1−4 5−9 10−19 20−99

20 20

15 15

10 10

5 5

0 0
1991 1995 1999 2003 2007 2011 2015 2019 1991 1995 1999 2003 2007 2011 2015 2019

All sizes 1−4 5−9 10−19 20−99 All sizes 1−4 5−9 10−19 20−99

(d) Other Services
Entry rates Exit rates
25 25

20 20

15 15

10 10

5 5

0 0
1991 1995 1999 2003 2007 2011 2015 2019 1991 1995 1999 2003 2007 2011 2015 2019

All sizes 1−4 5−9 10−19 20−99 All sizes 1−4 5−9 10−19 20−99

Notes: BDS annual entry and exit rates, 1991-2019. The solid line in each plot reports the BDS rates for all class sizes, which include
establishments with more than 100 employees. The other lines report BDS rates for establishment sizes that match our HB data.

34
that Decker’s formula works well at the quarterly frequency for the four sectors of our analysis, even
during the pandemic crisis where it predicts death rates for the second and third quarters of 2020 that
come close to the actual rates.

Comparison between BED and HB rates. To compare BED rates to HB rates, we aggregate
weekly tracked hours for each HB establishment " in the base sample to the monthly level and then
define establishment " as having positive employment in quarter q if its tracked hours in the third month
of that quarter are positive. We then define

• HB total entry for quarter q as HB establishments with positive employment in the third month of
quarter q and no employment in the third month of the preceding four quarters (q − 4, q − 3, q −
2, q − 1);21

• HB total exit for quarter q as HB establishments with positive employment in the third month of
quarter q − 1 and no employment in the third month of the subsequent four quarters (q, q + 1, q +
2, q + 3).

Because of sample churn, the HB total entry and exit rates are higher than the BED counterparts, as
shown by Figure 1 in the paper. On the other hand, the adjusted HB birth and death rates (also shown
in Figure 1) come very close to their BED counterparts. The HB birth rate in quarter q is the sum of
p̂ (birthi,t ) over the weeks of quarter q. The HB death rate in quarter q is the subset of HB establishments
with positive employment in the third month of quarter q − 1, no employment as of quarter q and that
do not continue to operate according to our Google/Facebook approach. Since the Google/Facebook
approach is intended to identify permanent closings, we effectively estimate that the establishment has
no employment in the third month of the subsequent quarters.22,23

21
As in the BED, no employment means either zero reported employment or no reported employment because the HB
establishment appears for the first time in that quarter.
22
If we strictly followed BED’s definition of death, we would include HB locations that return to activity in q + 4 or later.
We exclude them here since these may simply be businesses that stopped using HB for a while and later on return as clients.
In a previous version, we analyzed HB hazard rates defined as the ratio between the number of establishments that return
to activity after t weeks of inactivity to the total number of establishments returning to activity. These hazard rates suggest
returns to activity in q + 4 is an extremely unlikely event.
23
One caveat is that although Google Places distinguishes “temporary closed” and “permanently closed”, we treat both
indicators the same and tag the establishment as closed. Google’s “permanently closed” accounts for more than 80% of all
the exiting establishments from HB that are found to be Google-closed. Moreover, an establishment flagged as “temporary
closed” at a given point in time might later on be flagged as “permanently closed” in Google Places.

35
E Representativeness and additional checks on closings/openings

E.1 Using visits data to assess representativeness of the Homebase sample

We use Safegraph’s Points of Interest (POI) visits data to check for issues of representativeness and
potential selection of small businesses into usage of the Homebase software. The basic idea is to take
advantage of the much larger sample size of the Safegraph data to compare it to the Homebase data.
The result of this exercise, which we explain in detail in this section, is readily summarized. Along
the dimensions that can be compared, we find that Homebase establishments are not different in any
significant manner from establishments from the larger Safegraph sample that covers over 20 percent
of the universe of small businesses in the four sectors of our analysis. Thus, there is little evidence of
sample selection into Homebase, at least among small businesses in service sectors that require in-person
interaction.

Data preliminaries. We extract a sample of Safegraph POIs that meet two requirements: (1) they
can be characterized as small establishments, and (2) have weekly foot traffic data available. (1) is
challenging because the Safegraph data does not directly include a measure of establishment size. To
address this issue, we take advantage of an extra data product called the NetWise dataset. NetWise
is a data company that specializes in identifying business persons associated to company datapoints
for sales, advertising and marketing purposes, using a variety of data aggregation techniques of online
information; see https://www.netwisedata.com/our-data/ for details. In September 2021, Safegraph
released a cross-sectional NetWise dataset of counts of business persons that can be linked to the universe
of POIs that were tracked by Safegraph at this point. POIs in this dataset are identified by a placekey
identifier, which in turn allows us to match them to (a subset of) the Safegraph Core Places data (where
NAICS codes can be obtained) and Safegraph Weekly Pattern (containing visits data that are relevant
to characterize local economic activity).24 The second data requirement slightly reduces the sample size
because not all Safegraph POIs have visits (see Section B.1 for details on Safegraph)
Table E1 describes the samples of the analysis presented in the next paragraphs. After using the
NetWise employment information to extract the set of POIs with fewer than 50 workers, we restrict the

24
Between March 2020 and September 2020, POIs in the Safegraph Core Places data were identified using a
safegraph place id (see Section B.1 for details). From the November 2020 release of the Core Places data to June
2021, they are identified using both a safegraph place id and placekey identifier. As of July 2021, the Core Places
data only use the placekey identifier. We use the November 2020 through June 2021 release to create a crosswalk between
safegraph place id and placekey identifiers. For the Safegraph Weekly Pattern, we rely on the latest release of these data,
which, by construction of generating visits data based on the recent Core Places data, use placekey to identify POIs.

36
Table E1: Sample sizes of Safegraph and Homebase small establishments with available visits data

Retail Education Leisure & Other
Total
Trade & Health Hospitality Services
Safegraph sample of small estabs:
All 324,561 200,910 291,088 151,830 968,389
Without brand 188,622 187,023 167,980 146,139 689,764

Homebase sample 14,092 5,448 35,683 2,285 57,508

Notes: The table reports sample sizes for Safegraph small POIs in Retail Trade (NAICS 44-45), Education & Health (NAICS
61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS 81) with available visits data from Safegraph weekly
patterns. The last row of the table reports sample size for the Homebase establishments from either the mid-February 2020 base
sample or 2020-2021 new entrants sample with available Safegraph visits data.

sample to those in the four sectors of interest and with available visits data. The overall sample contains
almost 1 million establishments (upper panel). Compared to the 4.4 million of small businesses for the four
sectors according to QCEW establishment counts, this means the Safegraph-NetWise extract covers 22
percent of its universe. In Table E1, we further distinguish between POIs that are associated with a brand
and those that are not. A Safegraph brand corresponds to a chains of commercial POIs (McDonald’s,
Starbucks, etc.); see https://docs.safegraph.com/docs/core-places#section-brands. The table
shows that brands are pervasive in Retail Trade and Leisure & Hospitality, and much less so in Education
and Health and in Other Services. As will be shown below, there are differences in visits data between
POIs that have no brand associated (which typically are single commercial locations) and branded POIs.
However, the patterns of changes in visits over time are similar across the two sets of POIs.
The last row of Table E1 describes the sample of Homebase establishments that we use to run the
comparative analysis of visits data. The overall sample size is about 58,000 establishments. This cor-
responds to the mid-February 2020 base sample and the 2020-2021 new entrants sample put together,
and for which we have available visits data from Safegraph weekly patterns. Note that we also use our
2020 samples to check the accuracy of the NetWise employment information. Not all the establishments
from the Homebase samples can be linked to NetWise, due to differences in identifiers across databases,25
in addition to issues such as business closing between February 2020 and September 2021. Meanwhile,
among the roughly one third of establishments from the 2020 base and new entrants samples that can be
linked across datasets, we find almost 90 percent of them have fewer than 50 workers according to the
NetWise employment dataset.26
25
In the NetWise dataset (which covers data for September 2021), establishments are identified using the placekey
identifier. In our Homebase data, we link establishments to a safegraph place id, and not all safegraph place id’s that
have existed since the March 2020 Core Places of Safegraph can be linked to a placekey identifier of September 2021.
26
Across all four sectors, NetWise attributes fewer than 50 workers to 88 percent of the Homebase establishments from

37
Comparisons of visits data. Having described the two data samples, we now compare them with
respect to the various dimensions of foot traffic data measured by Safegraph. We analyze four such
dimensions: weekly counts of visits, median dwell time and share of long visits (defined as daily visits
longer than 240 minutes) among all visits, and weekly visits per unique visitor. For each samples, we
construct time series covering the period from mid-February 2020 through the end of the sample period,
and compare the Homebase sample denoted by green crosses with the Safegraph ones denoted in orange
and red marks. Figures E1–E4 present the results.
First, Figure E1 reports average weekly visits counts. Typically, establishments in Retail Trade and
Leisure & Hospitality receive more visits than those in Education & Health Services and Other Services.
Across all four sectors, we observe a large drop in visits at the beginning of the pandemic. In relative
terms, the drop is larger in Education and health services (visits fall by almost 80 percent) and in Leisure
& Hospitality (the decrease is by 60 percent in mid-April 2020). More importantly for our purposes, we
observe a very similar behavior over time of the time series that correspond to the different samples. In
Retail Trade, Homebase establishments are more similar (based on average weekly visits) to the sample
of all Safegraph small POIs, while in Leisure & Hospitality it resembles that of Safegraph small POIs
that have no brand associated. No matter these differences, in relative terms compared to mid-February
2020, the Homebase establishments do not behave differently from those of the larger Safegraph sample.
In Figure E2 and E3, we turn our attention to the duration of visits, by looking at median dwell
times and the share of visits lasting longer than 4 hours among all visits. In Retail Trade and Leisure
& Hospitality, visits are significantly shorter than in the other two sectors, with visits duration for the
Homebase establishments between those of Safegraph establishments hat have no brand associated and
those branded. In Other Services, we see no difference between the different samples for median dwell
times and the share of long visits. In Education & Health Services, we observe an unclear pattern in
what concerns long visits, which become relatively more important during the pandemic in the Homebase
sample compared to the Safegraph samples. This discrepancy might be driven by outliers and leaves
almost no discernible difference in median dwell times between the different samples in Education &
Health Services.
Last, in Figure E4 we consider the number of visits per unique weekly visitor. Again, Retail Trade
and Leisure & Hospitality differ from the other two sectors in that they have a lower share of returning

the 2020 base sample, and fewer than 100 workers to 95 percent of them. At the same time, there are differences across
sectors. 92 percent of the Homebase establishments in Leisure & Hospitality have fewer than 50 workers according to Netwise
employment, while the corresponding number is only 74 percent for Education & Health Services. In the latter sector, the
distribution shifts towards small businesses if we exclude establishments in NAICS 611 (“Educational services”) and 622
(“Hospitals”): Netwise employment then classifies 82 percent of the Homebases businesses as having fewer than 50 workers.

38
Figure E1: Safegraph small vs. Homebase establishments: Weekly counts of visits

Retail trade Education and health services
2,500

2,000

1,500

1,000

Weekly visits counts
500

0
Leisure and hospitality Other services
2,500

2,000

1,500

1,000

500

0
Feb 9 Mar 22 May 3 Jun 14 Jul 26 Sep 6 Oct 18 Nov 29 Jan 10 Feb 21
Apr 4
May 16 Jun 27 Aug 8 Sep 19 Oct 31 Feb 9 Mar 22 May 3 Jun 14 Jul 26 Sep 6 Oct 18 Nov 29 Jan 10 Feb 21
Apr 4
May 16 Jun 27 Aug 8 Sep 19 Oct 31

Homebase estab.
Safegraph small POIs Safegraph small POIs, not associated to a brand

Notes: Safegraph visits data. The orange circles and red triangles denote Safegraph POIs with fewer than 50 workers according to
NetWise employment data. The green crosses denote Homebase establishments matched to Safegraph visits data.

Figure E2: Safegraph small vs. Homebase establishments: Median dwell time

Retail trade Education and health services
150

120

Median dwell time (in minutes)
60

0
Leisure and hospitality Other services
150

120

Homebase estab.
Safegraph small POIs Safegraph small POIs, not associated to a brand

39
Figure E3: Safegraph small vs. Homebase establishments: Share of long visits

Retail trade Education and health services
40

Share of long visits (240+ minutes) among all visits
30

0
Leisure and hospitality Other services
40

Homebase estab.
Safegraph small POIs Safegraph small POIs, not associated to a brand

Figure E4: Safegraph small vs. Homebase establishments: Weekly visits per visitor

Retail trade Education and health services
2.0
1.9
1.8
1.7
1.6
1.5
1.4

Weekly visits per visitor
1.3
1.2
1.1
1.0
Leisure and hospitality Other services
2.0
1.9
1.8
1.7
1.6
1.5
1.4
1.3
1.2
1.1
1.0
Feb 9 Mar 22 May 3 Jun 14 Jul 26 Sep 6 Oct 18 Nov 29 Jan 10 Feb 21
Apr 4
May 16 Jun 27 Aug 8 Sep 19 Oct 31 Feb 9 Mar 22 May 3 Jun 14 Jul 26 Sep 6 Oct 18 Nov 29 Jan 10 Feb 21
Apr 4
May 16 Jun 27 Aug 8 Sep 19 Oct 31

Homebase estab.
Safegraph small POIs Safegraph small POIs, not associated to a brand

40
visitors within the week. What it is also remarkable in this figure is that we see little changes over time in
the number of visits per visitor, including in Retail Trade and Leisure & Hospitality. The picture shown
in Figure E4 is similar to the other figures, indicating that the Homebase establishments are not different
in any significant manner from establishments from the larger Safegraph samples.

E.2 Using visits data to compare closings and openings against sample churn

We use the Safegraph visits to run additional checks on our procedure to distinguish new openings and
closings from sample churn (see Section D.1). Recall that our Google/Facebook approach identifies
establishments that were already operating before entry into HB and establishments that continue to
operate outside of HB after disappearing from the data. In principle, these “churn” establishments
should behave similarly to those that are included in the base sample of our analysis. This is the basic
test that we perform in Figure E5.
The upper panel of Figure E5 compares changes in visits among entrant churners (they enter HB
after the mid-February base period and we identify that they were already operating before) with the
continuously active establishments from the base sample. The latter is the relevant comparison group as it
consists of the most stable establishments among those already active in the base period. Changes in visits
among the two sets of establishments are extremely similar in Retail Trade and Leisure & Hospitality. In
Education & Health Services and Other Services, there is a discrepancy coming from the larger decline
in visits at the beginning of the pandemic among churners, but the subsequent dynamics is similar. In
the lower panel of Figure E5, we compare exit churners to a larger set of establishments that includes
returning establishments in addition to the continuing establishments from the base sample. The reason
we include returning establishments is that they are the most relevant counterpart to establishments that
exit but continue to operate outside of HB: if exit churners were to use HB software again, we would
classify them as returning establishments. The plots show a great deal of overlap between the different
series. In Retail Trade and Education & Health Services, they are virtually identical to each other. In
Leisure & Hospitality, visits at exit churners seem to recover faster than among continuing and returning
establishments, but by the end of the sample period the series are almost on top of each other. The
opposite happens in Other Services. Overall, entry and exit churners, as identified by our Google /
Facebook approach, seem to behave in a very similar manner to their counterparts from the base sample.

41
Figure E5: Safegraph visits among Homebase “churn” establishments
(a) Entrant churners

Retail trade Education and health services
30

Visits change (in % relative to mid−February 2020)
0

−15

−30

−45

−60

−75
Leisure and hospitality Other services
30

−15

−30

−45

−60

−75
Feb 9 Mar 22 May 3 Jun 14 Jul 26 Sep 6 Oct 18 Nov 29 Jan 10 Feb 21
Apr 4
May 16 Jun 27 Aug 8 Sep 19 Oct 31 Feb 9 Mar 22 May 3 Jun 14 Jul 26 Sep 6 Oct 18 Nov 29 Jan 10 Feb 21
Apr 4
May 16 Jun 27 Aug 8 Sep 19 Oct 31

SG visits of continuing establishments SG visits of entrants churners

(b) Exit churners

Retail trade Education and health services
30

Visits change (in % relative to mid−February 2020)
0

−15

−30

−45

−60

−75
Leisure and hospitality Other services
30

−15

−30

−45

−60

SG visits of continuing and returning establishments SG visits of exit churners

Notes: Visit changes by continuing and returning businesses (dotted lines) vs businesses that we identify as churners among HB entries
and exits (crossed lines) in percent of respective visits level during the week of Feb 9 - Feb 15, 2020 for Retail Trade (NAICS 44-45),
Education and Health Services (NAICS 61-62), Leisure & Hospitality (NAICS 71-72), and Other Services (NAICS 81).

42
F Employment decompositions

F.1 Decomposition by establishment status

As described in the main text, our employment estimator is

" # $
Ai,t Oi,t
i ω i ê i,t + ê i,t
!t = E
E !t−1 × # $ (F.1)
" Ai,t Ci,t
ω
i i ê i,t−1 + ê i,t−1

where ωi denotes the sampling weight for industry-size-region cell i, constructed as the ratio of QCEW
A
establishment counts in 2020:Q1 to HB establishment counts in that industry-size-region cell; êi,ti,t denotes
O
week t employment of the set of establishments Ai,t that are active in HB in both week t and t − 1; êi,ti,t
denotes week t employment of the set of establishments Oi,t that are either newly opening or reopening
Ci,t
in week t; and êi,t−1 denotes week t − 1 employment of the set of establishments Ci,t that are closing
either temporarily or permanently in week t.
To motivate our decomposition, suppose that there is no sample churn; i.e. all exits from the HB
sample are temporary or permanent closings and all entrants in the HB sample are new openings or
reopenings. Under this scenario, the following equality holds

A i,t−1 O i,t−1 A i,t C i,t
êi,t−1 = êi,t−1 + êi,t−1 = êi,t−1 + êi,t−1 (F.2)

for every industry-size-region cell i. Intuitively, Ai,t−1 is the set of establishments active in week t − 1 and
t − 2 and Oi,t−1 the set of establishments active in week t − 1 but not t − 2. Together they account for the
set of all establishments active in week t − 1. Without sample churn, this set is the same as Ai,t , the set
of establishments active in week t − 1 that continue to be active in t plus Ci,t , the set of establishments
active in week t − 1 but not in t. With sample churn, this equality would not hold since Ai,t−1 would
also contain establishments that exit HB in t but continue to operate (i.e. not closings) and Ai,t would
also contain establishments that enter HB in t − 1 but operate already beforehand (i.e. not openings).
Given (F.2), we can iterate Equation (F.1) backward to week 0 and obtain

" # $ " # $
Ai,t Oi,t Ai,t Oi,t
ω i ê + ê ω
i i ê + ê
!t = E0 × i # i,t
E
i,t
$ = E0 × "
i,t i,t
. (F.3)
" Ai,1 Ci,1 ωi êi,0
i ω i ê i,0 + ê i,0 i

where E0 is CES employment in reference week 0, and êi,0 is HB employment of all establishments

43
belonging to cell i in reference week 0. Subtracting E0 from both sides, we can therefore express the
change in employment relative to the reference week 0 as

" # $
Ai,t Oi,t
i ω i ê i,t + ê i,t − ê i,0
!t − E0 = E0 ×
E " . (F.4)
ω ê
i i i,0

O B R
Now, we split êi,ti,t into employment êi,ti,t from new openings (births) in week t and êi,ti,t employment
from establishments that were active in reference week 0, temporarily closed at some point, and reopen
in week t. Hence,
A O A R B
êi,ti,t + êi,ti,t = êi,ti,t + êi,ti,t + êi,ti,t . (F.5)

Next and still supposing no sample churn, week t employment of establishments active in week t − 1 and
A
t, êi,ti,t , can be decomposed as
(2)
A A R B
êi,ti,t = êi,ti,t + êi,ti,t−1 + êi,ti,t−1 , (F.6)
(2)
A
where, with some abuse of notation, êi,ti,t denotes week-t employment of establishments continuously
R
open from week t − 2 to t, êi,ti,t−1 denotes week-t employment of establishments reopening in t − 1, and
B
êi,ti,t−1 denotes week-t employment of new establishments opening in t − 1. Combining Equations (F.5)
and (F.6) and iterating back to reference week 0, we obtain

(t) t
) t
)
A O A R B
êi,ti,t + êi,ti,t = êi,ti,t + êi,ti,s + êi,ti,s , (F.7)
s=1 s=1

(t)
A
where êi,ti,t denotes week-t employment of establishments that stayed continuously open from reference
" R
week 0 to t, ts=1 êi,ti,s denotes the sum of week t employment of all establishments that temporarily
" B
closed at some point after reference week 0 and reopened by week t, and ts=1 êi,ti,s denotes the sum of
week t employment of all establishments that newly opened after reference week 0. Similarly, we can
decompose week-0 employment of all establishments belonging to cell i in reference week 0 as

(t) t
)
A R C
êi,0 = êi,0i,t + êi,0i,s + êi,0i,t (F.8)
s=1

C " R
where êi,0i,t is added by definition of ts=1 êi,0i,s not including establishment that are still closed in week
" B
t, and ts=1 êi,0i,s is missing by definition of week-0 employment of establishments newly opening in week
s = 1, ..., t being zero.

44
Plugging (F.7) and (F.8) into Equation (F.4), we finally obtain
* (t) +
" "t # Ri,s $
(t)
Ai,t Ai,t "
ω ê − ê Ri,s
i i i,t i,0 ω
i i s=1 ê i,t − ê i,0
!t − E0 = E0 ×
E " + E0 × "
ωi êi,0 ωi êi,0
, -.i / , -.i /
Change from continously active estabs Change from reopenings
" "t Bi,s " Ci,t
i ωi s=1 êi,t i ωi êi,0
+ E0 × " − E0 × " (F.9)
ωi êi,0 ωi êi,0
, -.i / , -.i /
Change from new openings Change from closings

This decomposition holds exactly under no sample churn. With sample churn, the decomposition holds
approximately under the assumption that while being active in HB, employment growth of entering
establishments that operated prior to entry and exiting establishments that continue to operate after exit
is about equal to employment growth of continuously active establishments. We verify that this indeed
the case by comparing the change from continuously active establishments (the first term on the right-
hand side above) with the residual obtained from subtracting the change from reopenings, the change
from new openings and the change from closings (the last three terms on the right-hand side above) from
!t − E0 (computed with Equation (F.1)).
E

F.2 Decomposition into hiring and separation flows

For a given establishment ", we can decompose employment growth into hiring and separations

ê!,t − ê!,t−1 = ĥ!,t − ŝ!,t (F.10)

where ĥ!,t are all the employees in establishment " who work in week t but not in t − 1, and ŝ!,t are all the
employees in establishment " who work in week t − 1 but not in t (for firms with several establishments,
we define hiring and separations at the firm level; i.e. if an employee works at one establishment in one
week but another establishment of the same firm in another week, we do not count it as a separation /
hire). Hence,

" " " "
!t − E
!t−1 = E
!t−1 × "i ω i !∈i ĥ !,t !t−1 × "i ω i !∈i ŝ!,t
E −E . (F.11)
i ωi êi,t−1 i ωi êi,t−1
, -. / , -. /
Change from hiring Change from job separation

We let " " " "
i ωi !∈i ĥ!,t i ωi !∈i ŝ!,t
hiring ratet = " and separation ratet = "
ω ê
i i i,t−1 ω ê
i i i,t−1

45
denote, respectively, the hiring rate and separation rate in week t. hiring ratet and separation ratet are
plotted in Figure 8 of the paper, where separation ratet is split into the hiring rate of new workers and
that of recalled employees. The turnover rate (also in Figure 8 of the paper) is defined as
"
| ωi êi,t − êi,t−1 |
i"
turnover ratet = (hiring ratet + separation ratet ) − .
i ωi êi,t−1

In Section 5 of the paper, we also discuss recall rates in our data. The recall rate in week t is defined as
the ratio between recalled employees and the sum of recalled employees and new hires in week t.

G Additional figures and tables

G.1 Counterfactual employment estimates for the pre-pandemic period

Figure G1 is the counterpart of Figure 8 in the main text. The figure reports different counterfactual
employment estimates: the brown short-dashed line uses only the set of establishments that are continu-
ously active in HB; the red dashed-dotted line treats all exits as either temporary or permanent closings;
the orange dashed line, finally, adds all entries and treats them as new openings. The green circled line
corresponds to our baseline small business estimates. The latter shows that employment in mid-February
2020 was roughly similar to employment in mid-February 2019 in the four sectors considered. This is
in line with the QCEW year-on-year employment growth rates reported in Figure 2 of the paper. The
counterfactual employment estimates, on the other hand, would predict very large changes in employment
between mid-February 2019 and 2020. For example, the orange dashed line yields year-on-year employ-
ment growth rates between 50 and 75 percent, depending on the sector considered. In sum, Figure G1
demonstrates that distinguishing closings and openings from sample churn is important even during the
pre-pandemic period.

46
Figure G1: Comparison with counterfactual employment estimators

Retail trade Education and health services
75

Employment change (in % relative to mid−February)
50

ref. ref.
0

−25
Leisure and hospitality Other services
75

ref. ref.
0

−25
Jan 20 Feb 17 Mar 17 Apr 14 May 12 Sep 29 Oct 27 Nov 24 Dec 22 Jan 19 Feb 16 Jan 20 Feb 17 Mar 17 Apr 14 May 12 Sep 29 Oct 27 Nov 24 Dec 22 Jan 19 Feb 16
Jun 9 Jul 7 Aug 4 Sep 1 Jun 9 Jul 7 Aug 4 Sep 1

G.2 Regression variables

Table G1 describes the main variables used in the regressions reported in the paper. Table G2 provides
descriptive statistics for the main variables. Figure G2 complements the description of time-varying
control variables by showing distributions for the weeks of April 12 – April 18 (when small business
employment is at its lowest) and October 4 – October 10 (when large parts of the economy had reopened).
In Figure G2, the plots are based on data for the counties used in the regressions (1,957 counties), but the
distributions are virtually identical if we use data for all counties. Figure G3 compares the distribution of
delays in PPP loan among counties included in the regressions vs. all counties. Again, the distributions
are very similar when comparing counties included in our regressions and all counties.

47
Table G1: Description of the control variables used in Section 6

Variable Description Source

COVID cases Number of COVID cases per 1,000 county inhabitants COVID Act Now
COVID deaths Number of COVID deaths per 1,000 county inhabitants https://covidactnow.org/

containment index Weighted measure of restrictions for all containment Oxford COVID-19 Government
measures (school closure, workplace closure, Response Tracker
cancel public events, gathering restriction, public
transportation restriction, stay-at-home order,
internal movement control) at the state level

NPI1 0: No restriction Atalay et al. [2020]
1: 50% or more of all industries within a county https://reopeningdata.github.io/
having a closure restriction
NPI1 sector 0: No restriction
1: businesses in a specific 2-digit NAICS within a
county is required to be closed
NPI2 0: No restriction/Advisory Centers for Diseases Control
1: Mandate stay-at-home policy for high-risk people and Prevention
2: Mandate stay-at-home restriction for all people
NPI3 0: No restriction Centers for Diseases Control
1: Ban gathering above certain sizes and Prevention
2: Ban gathering of all sizes

SG school visits Safegraph school visits in log difference relative to Safegraph Weekly Patterns
same week of the previous year visits to places associated with
NAICS code 611110 (“Elemen-
tary and Secondary Schools”)

weather max temp Weekly average of maximum daily temperature (in °F) Climatology Lab GRIDMET
climatologylab.org/gridmet.html

hh income County-level household income 2016-2019 American Community Survey
5-year Estimates (5-year ACS)

PPPdelay PPP loans (at the county level) received during Small Business Administration
the week of Apr 26 - May 2 divided by the sum and Doniger and Kay [2021]
of PPP loans during theweeks of Apr 12
- Apr 18, Apr 19 - Apr 25, and Apr 26 - May 2
PPP loans (county × 2-digit NAICS level) received
PPPdelay sector during the week of Apr 26 - May 2 divided by
the sum of PPP loans during theweeks of Apr 12
- Apr 18, Apr 19 - Apr 25, and Apr 26 - May 2

48
Figure G2: Distribution of control variables in Apr 12 – Apr 18 vs. Oct 4 – Oct 10
(a) COVID cases
0.40

0.30

Fraction
0.20

0.10

0.00
0.0 10.0 20.0 30.0
COVID cases