Casado et al., The Aggregate Effects of Fiscal Stimulus: Evidence from the COVID-19 Unemployment Supplement (NBER w27576 (rev. Apr 2021), 2020-08)

Summary

NBER Working Paper 27576, dated August 2020 and revised April 2021, by Miguel Garza Casado and co-authors. The paper uses administrative records on unemployment claimants and transaction level data on economic activity to build a framework for evaluating real-time effects of fiscal policy on local economic activity. It uses the changing size of the Federal Pandemic Unemployment Compensation (FPUC) supplement, ($600, then $300) between March and September 2020, across counties and over time. The abstract says higher earnings replacement rates lead to significantly more consumer spending, even with increases in the unemployment claimant rate.

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

Cite this document

“Casado et al., The Aggregate Effects of Fiscal Stimulus: Evidence from the COVID-19 Unemployment Supplement (NBER w27576 (rev. Apr 2021), 2020-08).” Original at www.nber.org. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/casado-et-al-the-aggregate-effects-of-fiscal-stimulus-evidence-from-th-8757aa7f86/, SHA-256 381f4b70bebc069314ce1bb2dfe84975849dd3e87090a12fea40b4922ead5fcd.

Full text

NBER WORKING PAPER SERIES

THE AGGREGATE EFFECTS OF FISCAL STIMULUS:
EVIDENCE FROM THE COVID-19 UNEMPLOYMENT SUPPLEMENT

Miguel Garza Casado
Britta Glennon
Julia Lane
David McQuown
Daniel Rich
Bruce A. Weinberg

Working Paper 27576
http://www.nber.org/papers/w27576

NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue
Cambridge, MA 02138
August 2020, Revised April 2021

Previously circulated as "The Effect of Fiscal Stimulus: Evidence from COVID-19." This work
was funded by Schmidt Futures, Amazon Web Services, the Overdeck Family Foundation the
Bill and Melinda Gates Foundation and NIH grants UL1 TR002733 and R24 HD058484. George
Putnam, Daniela Hochfellner, Nathan Caplan and Frauke Kreuter provided extremely useful input
and guidance. We are grateful to Illinois Department of Employment Security for providing data
access. The views expressed herein are those of the authors and do not necessarily reflect the
views of the National Bureau of Economic Research.

At least one co-author has disclosed additional relationships of potential relevance for this
research. Further information is available online at http://www.nber.org/papers/w27576.ack

NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.

© 2020 by Miguel Garza Casado, Britta Glennon, Julia Lane, David McQuown, Daniel Rich, and
Bruce A. Weinberg. All rights reserved. Short sections of text, not to exceed two paragraphs, may
be quoted without explicit permission provided that full credit, including © notice, is given to the
source.
The Aggregate Effects of Fiscal Stimulus: Evidence from the COVID-19 Unemployment
Supplement
Miguel Garza Casado, Britta Glennon, Julia Lane, David McQuown, Daniel Rich, and Bruce
A. Weinberg
NBER Working Paper No. 27576
August 2020, Revised April 2021
JEL No. E21,E62,H3,J65

ABSTRACT

The current economic crisis has highlighted the need for data that are both timely and local so
that the effects of fiscal policy options on local economies can be evaluated more immediately.
This paper highlights the potential value of using two new sources of near real-time data to
inform decisions about the appropriate stimulus approach to implement. The first data source is
administrative records that provide universal, weekly, information on unemployment claimants.
The second data source is transaction level data on economic activity that are available on a daily
basis. We make use of discrete changes in stimulus payments to construct a framework for
evaluating real-time effects of fiscal policy on local economic activity. In particular, we leverage
cross-county and over-time variation in the relative size of the Federal Pandemic Unemployment
Compensation (FPUC) COVID-19 supplement to Unemployment Insurance – from $0 to $600 to
$300 between March and September 2020 - to estimate the local economic impact of
unemployment, earnings replacement, and the interaction between the two. We find that higher
earnings replacement rates lead to significantly more consumer spending, even with increases in
the unemployment claimant rate, which is consistent with the goal of the fiscal stimulus.

Miguel Garza Casado David McQuown
University of Pennsylvania Chapin Hall at the University of Chicago
3620 Locust Walk 1313 East 60th Street
Philadelphia, PA 19104 Chicago, IL 60637
mgarzaca@sas.upenn.edu dmcquown@chapinhall.org

Britta Glennon Daniel Rich
The Wharton School Illinois State University
University of Pennsylvania Campus Box 4200
3620 Locust Walk Normal, IL 61790
Philadelphia, PA 19104 dprich@ilstu.edu
and NBER
bglennon@wharton.upenn.edu Bruce A. Weinberg
The Ohio State University
Julia Lane Department of Economics
Wagner School of Public Service 410 Arps Hall
New York University 1945 North High Street
295 Lafayette Street Columbus, OH 43210
New York, NY 10012-9604 and NBER
and Coleridge Initiative weinberg.27@osu.edu
julia.lane@nyu.edu
1. Introduction

The COVID-19 pandemic led to a series of unprecedented and sharp fiscal interventions designed to
attenuate the sharp economic downturn. The most immediate action was a federal supplement to existing
Unemployment Insurance (UI) benefits of an additional $600 a week, beginning in the week ending
March 21, 2020 (Karpman and Acs 2020). Those benefits ended at the end of July. The second action
was the replacement of the $600 a week benefit by a supplement of $300 a week that lasted to the
beginning of September. The structure of those benefits resulted in dramatic, high-frequency fluctuations
in the amount of economic stimulus to states and counties, depending on the number of local claimants
who received the benefits and the amount to which their lost wages were replaced. In this paper we make
use of the universe of weekly benefits data in the fifth largest state in the economy – Illinois – to estimate
the links between those two stimuli and the aggregate spending in each county. We also argue that the
source data used in this paper can be used for similar and extended analyses in other states.

The literature on the effects of fiscal stimuli on consumption spending provide mixed evidence on the
effectiveness of such interventions. The early arguments of Friedman and Modigliani were that temporary
increases in income would be saved, rather than spent, resulting in little fiscal impact(Friedman 1957;
Merton 1987). Later work has argued that many consumers – even the wealthy – operate on a “hand-to-
mouth” basis, and spend large fractions of government transfers(Kaplan, Violante, and Weidner 2014).
Empirical work consistently shows much greater marginal propensities to consume than predicted by
basic models; more recent work in heterogeneous agent macroeconomics has tried to rectify the disparity
between theory and evidence(Kaplan and Violante 2014; Carroll et al. 2020).

Given the theoretical ambiguity, the need for rigorous empirical work and local, real-time data to inform
policy interventions is especially pressing. Key decisions are being made with considerable uncertainty
about how state and local individuals and businesses “will respond to recent fiscal and monetary policy
actions taken by the federal government”(Congressional Budget Office [CBO] July 2020; Swagel 2020).
This paper shows how the data toolkit for decision-makers can be expanded to evaluate fiscal impact in
near-real time. Existing data can be combined to do so: in particularly, state administrative records on
unemployment claims can aggregated at the county level and combined with county level credit- and
debit- card transaction data. The first dataset, unemployment claims records, provides universal and
weekly information on the claims, benefits, and previous earnings of claimants, and makes it possible to
calculate individual-level replacement rates and unemployment claimant rates that can be aggregated to
any level of geography, demographic group and industry. The second data source, transaction-level data
on credit and debit card purchases can proxy economic activity at a daily basis at the county level (Chetty
et al. 2020). Unlike much prior work which has focused on estimating the fiscal impact on a micro
marginal propensity to consume, these detailed individual- and county-level data allow us to estimate a
macro spending impact of a change in UI benefits that encompasses multiplier effects.

The analytical contribution is made possible by the differential impact of COVID-19 – and of the
subsequent enhanced unemployment benefits – across industries, regions, and time to obtain plausibly
exogenous variations in unemployment and replacement rates to estimate their effects on economic
activity. Our estimates are relatively unique in that they give aggregate effects that are relevant for policy
makers - namely estimates that include local spillovers from all individuals in a local area to local
businesses - rather than the effects on individual consumption, or effects across geographic areas which
are more typical in the literature (Chodorow-Reich 2019). Importantly, this means that we estimate a
macro spending impact of a cash payout that accounts for any multiplier effects, rather than a micro
marginal propensity to consume. The data and the approach can scale to all states in the nation, can
incorporate spillovers across counties, and can inform decision-making in a timely manner, not just for
this downturn, but for future recessions.

The data are drawn from the state of Illinois, which like every other state in the country, saw a massive
shock in March 2020. Unemployment claims soared from an average of 140,000 a month in 2019 to
almost 800,000 a month for the subsequent six months. The stimulus injected about $475 million a week
into the Illinois economy over and above the unemployment benefits that would otherwise have been paid
out. For many individuals, the payout of UI benefits plus $600 exceeded their prior earnings – a wage
replacement rate greater than one. The analysis outlined below estimates that replacing the Federal
Pandemic Unemployment Compensation (FPUC) supplement to unemployment insurance of $600 at the
end of July 2020 and replacing it with $300 supplement reduced consumption by 5%. Thus, the effects of
large reductions in benefits on consumer spending are quite substantial.

2. Background
As indicated, the effects of fiscal stimulus on consumption have been debated at least since Friedman and
Modigliani in the 1950’s, and been reinvigorated by recent empirical work suggesting that temporary
stimulus is more effective than was previous believed, albeit with heterogeneous effects(Kaplan,
Violante, and Weidner 2014). In an empirical analysis of the effects of the 2001 and 2008 tax cuts, Misra
and Surico use consumer expenditure survey data to document substantial heterogeneity in the spending
impact depending on the income level of individual families (Misra and Surico 2014). Parker et al. also
use survey data, with targeted questions, to show that spending increased by 12-30% as a result of the
stimulus tax payments of about 100 billion dollars sent to 130 million families (Parker et al. 2013). The
analysis of the effects of the stimulus during the 2007-2009 Great Recession also support the notion that
temporary measures are effective. Elmendorf and Furman provide a good overview of lessons learned and
argue that interventions should be timely, targeted and temporary (Elmendorf and Furman 2008). In other
words, they mean that a stimulus should be implemented only if a decline occurs over a specific period,
should be targeted at those who spend the most (the most vulnerable families), and should be temporary
in nature. While both fiscal and monetary policy should be implemented, fiscal policy has the advantage
of being able to be implemented quickly, particularly if, as now, the federal funds rate is close to zero.
They argue that increasing spending on unemployment insurance benefits (or food stamps) or targeted tax
cuts, is to be preferred to infrastructure investments, temporary tax incentives or more general tax cuts.

A large stream of empirical work has emerged examining the implications of unemployment insurance
benefits for labor markets, household spending, and economic recovery, given the focus on increasing
unemployment insurance benefits as a policy response to the 2020 crisis8. Of particular relevance to this
paper and to the hand-to-mouth literature, is Ganong et al. (2020), which combined public use data from
the Current Population Survey with information about each state’s UI system under the CARES Act to
calculate wage replacement rates. Their sample, which consisted of 444 unemployed workers, provided

8
See, for example, the compendium produced by the National Bureau of Economic Research
https://nber.org/wp_covid19_07202020.html
the basis for examining the effect of different policies on the distribution of benefits and labor supply
incentives – and emphasizes the heterogeneity of those effects even in such a small sample (Ganong,
Noel, and Vavra 2020).

The third strand of work has focused on estimating the impact of epidemics on spending patterns. Baker
et al. use bank account data transaction-level data from a non-profit to study changes in the spending
patterns of 4,735 customers (Baker et al. 2020). A study of the spending patterns of about 5 million Chase
Bank customers suggests that spending declines were directly related to the pandemic rather than labor
market effects (Cox et al. 2020). They find that there are substantial differences between customers with
different income levels. On the theoretical side, Carroll et al. have built a sophisticated consumer
spending model with micro economic foundations that shows the link between layoffs, benefits and
consumption spending (Carroll, Slacalek, and White 2020). It distinguishes between those individuals
who are “deeply” unemployed versus those who are “normally” unemployed and allows policy-makers to
explore the effects of different assumptions about who gets unemployed, and how long it takes them to
return to employment.

While these papers help understand individual responses to fiscal stimulus, they do not provide effects on
aggregate activity, which are particularly useful for policy. Our work suggests that including local
multiplier effects can result in sizeable estimates of the economic impact of fiscal policies such as
increased unemployment benefits. It is worth noting that the UI claims data studied here are very similar
in structure in every state across the country. State Departments of Labor have been learning how to
convert claims data to data on claimants, and then to longitudinal data on claimant cohorts so the
approach can be scaled up and used to evaluate fiscal policy impact on local economies in near real-time
for future crises.9 In addition, since the data described in this paper have detailed information on
education, prior earnings, race and ethnicity for well over a million individuals, they could be used in
future research to provide more precise estimates about the distributional consequences of different
policies; state Departments of Labor have also been learning how to do so.

3. Data

We combine two primary data sources in order to generate a unique dataset that permits the analysis of
the combined effect of increased unemployment claimant rates and stimuli-enhanced earnings
replacement rates on economic activity10. The first dataset provides information at the individual level
about weekly unemployment claims, benefits received, and previous earnings, which are then used to
construct unemployment claimant and earnings replacement rates at the county and industry level. The
second dataset provides information on daily changes in debit and credit spending at the county level.

3.1 Unemployment Claimant and Earning Replacement Data

9
Indeed, the Department of Labor’s Employment and Training Administration recently sponsored a training
program for 30 states to make similar use of the claims data for analysis; 8 states have already done so.
https://performancereporting.workforcegps.org/announcements/2020/10/14/18/46/Coleridge-training-course.
10
The source data are deidentified and hosted in a secure facility – the Administrative Data Research Facility at the
Coleridge Initiative – which is FedRAMP moderate certified and has Authorization to Operate from both the Census
Bureau and USDA. Only authorized individuals can access the data, only for approved and statistical purposes, and
only summary analytical statistics are released.
The data that we use to construct measures of unemployment claimant rates and earnings replacement
rates come from the administrative records of the Illinois Unemployment Insurance system, which are
themselves broken into two main data sets.
The first of these is the Unemployment Insurance (UI) certified claims data generated weekly for the
Program for Measuring Insured Unemployed Statistics (PROMIS) system. The dataset contains records of
UI claimants’ weekly certifications for past weeks of benefits and provides a complete universe of
certified claims with a two-week lag. These claims data are transformed into claimant data, and linked
longitudinally to create cohort level data11. There are many key advantages for policymakers when
compared to survey data. The data are extremely timely, since they are reported every week, so that
policies can be developed in response to immediate needs. They are highly geographically granular, so
that heterogeneity in spatial effects can be taken into consideration. Finally, it is possible to examine
demographic heterogeneity since, for each individual claimant, the dataset contains not only benefit
details such as total amount paid, but also claimant details such as age, race, gender, educational
attainment, and pre-separation occupation and industry.
The second is the Quarterly UI Wage Records12, which contain quarterly wages for all UI-covered jobs in
Illinois. These records are filed by employers and include roughly 96% of private non-farm wage and
salary employment(Stevens 2007). Total wages (sometimes called wages or gross wages) for a quarter are
the total amount of wages paid or payable (depending on the wording of the State law) to covered workers
for services performed during the quarter. Bonuses paid are included in the payroll figures. Also included,
when furnished with the job, is the cash value of such items as meals, lodging, tips and other gratuities, to
the extent that State laws and regulations provide. Total wages include both taxable and nontaxable
wages.
Unemployment claimant rate
The construction of the unemployment claimant rate measure is conceptually different from the BLS
unemployment measure, which directly asks individuals whether they are “actively looking for work in
the survey week”. The unemployment claimant rate measure is directly based on the count of individuals
who certified and received unemployment benefits for the weeks between that ending January 25, 2020
and that ending September 5, 2020. The BLS measure has the advantage of familiarity, but has been
criticized for being atheoretical (Card 2011) and for not providing a measure that is useful for
policymakers(Brandolini 2018; Brandolini and Viviano 2016). The sample size is a major limitation:
Ganong’s analyses of pandemic unemployment studied responses from fewer than 500 individuals. The
claimant measure has the advantage of timeliness, and granularity, but has the disadvantage of not
covering specific sectors of the labor market.(Stevens 2007). The sample size is generous: over 13
million datapoints on over a million individuals in the time-period studied. In order to minimize
definitional confusion, we refer to this new measure as the “unemployment claimant rate” throughout.
We construct two versions of the unemployment claimant rate: one at the county-industry-week level, and
one at the county-week level. The former is necessary for constructing the shift-share instrumental
variables, the latter is used for the core specification. The numerator comes from aggregating the weekly
number of certified UI claims to a county-industry level, or to a county-industry-week level respectively.
The denominator for the unemployment claimant rate at the county-week level is derived from the 2019
Local Area Unemployment Statistics (LAUS) workforce data. The denominator for the unemployment
claimant rate at the county-industry-week level is derived from the 2018 Illinois American Community

11
https://ada.coleridgeinitiative.org/eta-class-home (password adrf).
12
UI wage records and their coverage are extensively described elsewhere. (Stevens 2007)
Survey (5 year) data on the workforce. While the ACS data have the advantage of industry-specific
workforce estimates, one shortcoming is that they only include county-industry level data for 15 of the 47
counties with consumption data. The regression analysis that uses a shift-share instrumental variable
approach is therefore limited to those 15 counties13. This constraint does not apply to our OLS
specifications.
Earnings Replacement rates
We calculate the earnings replacement rates by creating a ratio of the total amount paid in UI benefits14
for each certified claimant - for each week from the week ending January 25, 2020 to the week ending
September 5, 2020 - to their 2019 average weekly wage.
The 2019 average weekly wage is calculated for each individual from their earnings as reported in the
quarterly Illinois UI Wage Record file. These records are used to construct measures on each claimant’s
work and earnings history in the year prior to the COVID-19 pandemic. We then divide the aggregate
2019 wages by 52 to get a 2019 average weekly wage for each certified claimant.
Weekly county-level earnings replacement rates are calculated by taking the mean of the county
residents’ weekly UI benefit paid amount and dividing it by the mean of their weekly 2019 wages. Only
individuals with nonzero wages and paid benefit amounts are included in this calculation.
3.2 Economic Activity Measures
Our second primary source of data comes from a public database15 that tracks a variety of timely
measures of local economic activity (Chetty et al. 2020). For the purposes of this paper, we utilize their
spending data from Affinity Solutions. These are aggregated and anonymized purchase data from
consumer credit and debit card spending. The data are seasonally adjusted and indexed relative to January
2020 and are presented as a 7-day moving average at the county level.

3.3 Final Dataset

The final dataset is at the county-week level, from January 25 to September 5, 2020, Illinois is the sixth
largest state in the United States, with a population of about 13 million16, a workforce of about 6 million17
and a GDP of about $900 billion18. In the week ending July 4, nearly 650,000 workers claimed
unemployment insurance. Total UI payments received in that week, including the $600 FPUC
supplement, were $445 million.

Appendix Table 1 provides summary descriptive statistics, but three sets of pictures illustrate the variation
in each of the three key measures that we leverage in this paper: the variation over time, the variation
across counties, and the heterogeneity in the relationship between the earnings replacement rate and
spending.

13
The unemployment claimant rates in the maps, in contrast, use the 2019 Local Area Unemployment Statistics
program data as a workforce measure, which encompasses all 102 counties in Illinois, to provide a broader depiction
of unemployment claimant rates across the state.
14
The total UI payment received field includes both regular benefits, dependent allowances, and the Federal
Pandemic Unemployment Compensation addition and subtracts tax withholding and benefit discounts due to wages
earned.
15
https://tracktherecovery.org/
16
https://www.census.gov/quickfacts/IL
17
https://www.bls.gov/eag/eag.il.htm
18
https://www.bea.gov/sites/default/files/2020-07/qgdpstate0720_0_0.pdf
First, the effect of the pandemic was sharp and strong in Illinois. Figure 1 shows the trends in
unemployment claimant rates, replacement rates and spending over time for Illinois. The shock of the
pandemic is apparent: we see a massive drop in consumer spending and a large spike in unemployment in
March. The immediate effect of the $600 FPUC supplement in the CARES Act fiscal stimulus is also
apparent in both the sharp increase in earnings replacement (in excess of 100%), as well as in the sharp –
and immediate - increase in consumer spending, despite continued high levels of unemployment. The
shock of the change in unemployment benefits at the end of July is also apparent, with the average
replacement rate dropping from 1.2 to just under 1 and with the flattening of consumer spending.

Figure 1: The Pandemic – and CARES Act - Effect

Figure 2 shows spending across different sets of counties. Specifically, we split counties based on
values at the beginning of April 2020. Panel A shows the replacement rate in counties with high and low
replacement rates. Panel B shows spending in each of these counties. The replacement rates and in
spending (relative to the beginning of the year) are remarkably similar through early March, at which
point spending declines essentially equally in both sets of counties. After the $600 FPUC supplement,
replacement rates jump up more (by construction) by about 20% in the counties with higher replacement
rates. Spending also rebounds by a bit under 10% more and much more quickly in the counties with high
replacement rates than in those with low replacement rates. Thus, the differences in replacement rates
correlate in an intuitive way with differences in spending.
Panel C reports the rates of UI receipt for counties with high and low UI receipt in early April.
The counties with high UI rates in early April initially had UI rates of roughly 2% (compared to roughly
1% for the counties with low UI rates in early April). The high UI counties experience a much larger
increase in UI rates (to 11%) compared to 7% for those with lower UI rates in early April, before closing
the gap as both sets of counties trend downward. Panel D shows spending by unemployment claimant rate
in April. Spending drops by roughly the same amount in high and low unemployment counties, but
rebounds more and more quickly in low unemployment counties than in high unemployment counties
after the introduction of the FPUC19

Panel A Panel B

Panel C Panel D
Figure 2. Spending by Counties.

4. Conceptual Framework and Empirical Approach
In this framework, we use counties by week as the unit of analysis to estimate the impact of changes in
unemployment levels and earnings replacement rates on local economic activity. Our approach has the
advantage of both describing local-level heterogeneity in economic pain and enabling local policy makers
to target additional specific interventions. Of course, as noted above, since similar data are available for
each state, a similar approach could be used for every state.

19
An parallel analysis that stratified counties by 2019 average per capita income shows that spending falls more in
high income counties, which have the lowest replacement rates, than in low income counties.
The main outcome measure of interest is the level of economic activity, as described above: an index of
credit/debit spending at the county level compared to January 2020. The independent measures include
the unemployment claimant rate and the earnings replacement rate.

More formally, the baseline model can be written as:

𝐸𝑐𝑜𝑛𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦 𝛽 𝛽 𝑈𝑛𝑒𝑚𝑝 𝛽 𝑅𝑒𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡𝑅𝑎𝑡𝑒 𝑓 𝑡, 𝑐, 𝜃 𝜖 (1)
where c indexes counties and t indexes weeks. 𝑈𝑛𝑒𝑚𝑝 denotes our measure of the unemployment
claimant rate based on UI receipt rates and 𝑅𝑒𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡𝑅𝑎𝑡𝑒 gives the replacement rate, both of
which vary by county and week. Our main specifications include two-way county and time fixed effects,
but we also explore a specification with cubics in time fully interacted with county fixed effects, both
captured flexibly by 𝑓 𝑡, 𝑐, 𝜃 . One might naturally expect the effects of the replacement rate to vary
with the share of people receiving benefits (or, alternatively, the effects of unemployment to vary with the
replacement rate). We also estimate a model that includes interactions between 𝑈𝑛𝑒𝑚𝑝 and the
𝑅𝑒𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡𝑅𝑎𝑡𝑒 . Formally,
𝐸𝑐𝑜𝑛𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦 𝛽 𝛽 𝑈𝑛𝑒𝑚𝑝 𝛽 𝑅𝑒𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡𝑅𝑎𝑡𝑒
(2) 𝛽 𝑈𝑛𝑒𝑚𝑝 𝑅𝑒𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡𝑅𝑎𝑡𝑒 𝑓 𝑡, 𝑐, 𝜃 𝜖
Of course, as we show below, such a model with three correlated variables, a moderate sized sample, and
two-way fixed effects is challenging to estimate, but yields similar estimates to those from model (1).
Instrumental Variables
There are potential concerns with endogeneity in our analysis. We might worry about reverse causality;
local spending may well affect local unemployment. Furthermore, we might worry about omitted
variables; for example, the counties with the highest incomes both experienced the largest declines in
spending and have the lowest replacement rates. To address these concerns, we develop a series of shift-
share instruments for the endogenous independent variables, 𝑈𝑛𝑒𝑚𝑝 and 𝑅𝑒𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡𝑅𝑎𝑡𝑒 .
In the case of our instrument for the unemployment claimant rate, we take the share of workers in each
industry i who are unemployed in week t, 𝑈𝑛𝑒𝑚𝑝 , and weight them by the share of employment in each
county that is in that industry, 𝑆ℎ𝑎𝑟𝑒 , , in a base period, which is 2018. We calculate the industry
unemployment claimant rate by taking the number of certified UI claims for each industry in each week
and dividing that by the workforce in that industry in 2018. We then sum the weighted product across
industries to construct a shift-share instrument for unemployment, as follows:

𝐼𝑉 𝑆ℎ𝑎𝑟𝑒 , ∗ 𝑈𝑛𝑒𝑚𝑝

Our instrument for replacement rates is developed analogously, using the baseline industry employment
shares to weight industry-specific UI replacement rates. We estimate the industry-specific UI replacement
rate using the mean total benefits received in each week for the workers previously employed in each
industry divided by the mean weekly UI earnings in 2019 in that industry. Our UI benefit level is the
actual amount of money that people receive including the $600 FPUC supplement (as relevant) after
deductions like childcare. Formally, we estimate,
𝐼𝑉 𝑆ℎ𝑎𝑟𝑒 , ∗ 𝑅𝑅

5. Results
Table 1 reports summary statistics and OLS estimates, introducing the explanatory variables one at a
time. Column (1) reports the relationship between the unemployment claimant rate and spending. We find
that each 1% increase in unemployment claims leads to a decline in spending of 1.6%. If the only
regressor is the replacement rate (as in column (2)), then a 1% increase in the replacement rate increase
spending by roughly .2%. When both the unemployment claimant rate and replacement rate are included
together (in column (3)), the estimates on each are slightly lower than when they are included separately,
but both remain economically and statistically significant.20 The estimates indicate that a 1% increase in
unemployment claimants reduces spending by 1.491% and a 1% increase in the replacement rate
increases spending by .167%. Reducing the $600 FPUC unemployment insurance supplement to $300
reduced the replacement rate by roughly .3, which would be associated with a 5% reduction in spending.
To provide context, the $300 reduction in unemployment insurance benefits corresponds to a decline in
income of roughly 25% for 7% of the population or a reduction in income of 1.75% if the unemployed
and employed have similar incomes.
Column (4) adds an interaction between the replacement rate and the unemployment claimant rate.
Intuitively, the effect of the replacement rate should depend on the number of people who are receiving
benefits and vice versa. Unfortunately, since this specification includes three correlated variables and
two-way (county and time) fixed effects, the estimates become noisy.21 That said, when evaluated at the
mean unemployment claimant rate, the implied effect of an increase in the replacement rate is .143, which
is quite similar to the coefficient in column (3). Similarly, at the mean replacement rate, the implied effect
of an increase in the unemployment claimant rate is -1.77, which is also similar to the coefficient in
column (3). Given that the interaction in column (4) introduces noise, we instead focus on the estimates in
column (3) and use these as our baseline results.
Table 2 shows that the estimates are remarkably robust to a series of alternative specifications. Column
(1) repeats our baseline model for convenience. Column (2) uses a different replacement rate measure
based on the gross benefit as opposed to the amount of the payment (net of deductions). Column (3)
addresses the considerable uncertainty about benefit levels after the expiration of the $600 FPUC.
Ultimately $300 Lost Wage Assistance (LWA) was paid per week retroactively after a delay of several
weeks. Our baseline assumption is that people anticipated the $300 LWA supplement, but column (3)
calculates benefits assuming that people did not build the $300 LWA supplement into their consumption.
The estimates are remarkably similar to the baseline estimates. Our model obviously leverages variation
over time as well as variation across counties, but the correct timing assumptions are not clear. Columns
(4) and (5) address concerns about timing by lagging unemployment and replacement rates by one and
two weeks respectively. While these estimates are broadly similar, the one-week lag models fit the data
slightly better. This finding suggests that the effects of unemployment and replacement rates may not be
immediate, which seems plausible. Perhaps more importantly, these estimates suggest that most

20
While the UI rate and the replacement rate are strongly positively correlated overall (ρ=.68, weighted), indicating
that unemployment was higher in lower income counties (i.e. those with lower replacement rates), after eliminating
county and week variation, as in our regressions, the two are actually slightly negatively correlated (ρ=-.13,
weighted). Thus, including both variables simultaneously weakens both.
21
The correlation between the Unemployment claimant rate and the interaction is .87 and the correlation between
the replacement rate and the interaction is .28.
identification is off the broad rather than the precise timing. Column (6) provides unweighted estimates,
in contrast with prior estimates, which weight by population. These are broadly similar, but yield a
smaller estimate of the replacement rate. Column (7) excludes Cook County, which accounts for roughly
40% of the population and economic activity in the state of Illinois. In this specification, the
unemployment claimant rate becomes much smaller and insignificant, while the replacement rate
becomes slightly larger. Column (8) explores an alternative specification of our time and county effects,
including a cubic in the week interacted with county. The coefficient on the unemployment claimant rate
in this model is considerably higher than that in our baseline model while the coefficient on the
replacement rate is slightly higher. At the same time, the fit of this model is considerably worse than the
previous models.
As indicated, we are concerned that the OLS estimates may be biased because changes in consumption
would affect local employment and because the counties with the highest incomes experienced the largest
declines in consumption while having the lowest replacement rates. Table 3 reports instrumental variables
estimates, with the second stage estimates in the top panel and the first stage estimates in the bottom
panel. (As indicated, one disadvantage of the IV estimates is that they can only be computed for 15
counties instead of the 47 used elsewhere because of the further need to use county-level data on industry
composition from the American Community Survey.) Column (1) reports estimates with the
unemployment claimant rate alone. The first stage estimates are significant and the first stage F-statistic
on the excluded instrument is just above 10. The second stage estimate is quite similar to the
corresponding estimate in column (1) of Table 1. Column (2) reports estimates for the replacement rate
alone. Here the instrument is a strong predictor and first stage F-statistic on the excluded instrument is 40.
The second stage estimate is larger than the corresponding estimates in column (2) of Table 1, but it is
imprecise. Columns (3) and (4) report first stage estimates for the model with both the unemployment
claimant rate and replacement rate, with instruments for those two variables. Column (3) shows that the
instrument for the replacement rate is essentially unrelated to the unemployment claimant rate. In column
(4), the instrument for the unemployment claimant rate is a strong predictor of the replacement rate, but
the instrument for the replacement rate itself remains quite similar to that in column (2). In both columns
(3) and (4) the first stage F-statistics on the excluded instruments fall (because there are now 2
instruments). The second stage estimates for the unemployment claimant rate and replacement rate are
both larger (in magnitude) than the corresponding estimates in Table 1, although neither difference is
statistically significant.
Appendix Table 2 reports two stage least squares estimates for the same robustness checks that are
reported in Table 2. The estimates are higher than those in Table 2, but the differences are typically
smaller than those for our baseline specification and the estimates are not statistically significantly
different from those in Table 1. We conclude that whatever bias introduced by endogeneity that is not
controlled using the two-way fixed effects is secondary to the large shocks that were experienced because
of the pandemic itself.
Conclusion
This paper quantifies the large aggregate effect of the COVID-19 fiscal stimulus on spending and
economic activity. In Illinois, proposed cuts to unemployment insurance benefits are expected to have
large effects on consumer spending, with a $300 change in benefits resulting in a 5% decrease on
spending. The results are robust across a wide range of specifications and the use of instrumental
variables.
In addition to quantifying the aggregate effects of fiscal stimulus, this paper contributes in another way.
The weekly data used here are timely and universal, can be used to estimate the impact of stimulus effects
for all states in the country, and inform decision-making at all levels of government – federal, state and
local. More precise analysis is also possible, because the dataset is large enough to permit estimates to be
generated by different demographic, geographic, and income groups. As a result, further research could
examine the size of the consumption effect on both the “hand-to-mouth” segment and the low income
segment of the population. The individual data can be structured as cohorts, so that the impact of the
composition of unemployment by whether claimants are “deeply unemployed” or “normally
unemployed” can be examined. Policy makers can be informed about the regional differences in fiscal
stimulus, to better inform the allocation of resources to training providers and economic development
agencies.
Table 1. Baseline Fixed Effects Estimates of the Determinants of County-Level Spending

(1) (2) (3) (4)
Mean Spending relative Spending relative Spending relative Spending relative
(SD) to Jan2020 to Jan2020 to Jan2020 to Jan2020
Unemployment 0.0704 -1.595*** -1.491*** -2.809*
(0.0406) (0.418) (0.382) (1.053)

Replacement Rate 0.815 0.204* 0.167* 0.0538
(0.352) (0.0782) (0.0812) (0.105)

Unemployment*RR 0.0690 1.270
(0.0495) (0.949)
N 1598 1598 1598 1598
F 14.55 6.813 12.12 8.253
R-squared 0.872 0.864 0.875 0.876
County FEs YES YES YES YES
Week FEs YES YES YES YES
Standard errors in parentheses
Standard errors clustered at the county level
The units of observation are county-week pairs.
All are weighted by workforce.
The unemployment claimant rate is measured using the share of people claiming unemployment insurance.
The mean of the dependent variable is -0.150 (SD=0.134).
* p < 0.05, ** p < 0.01, *** p < 0.001
Table 2. Alternative Fixed Effects Estimates of the Determinants County-Level Spending.
(1) (2) (3) (4) (5) (6) (7) (8)
Baseline Alternate RR No LWA 1 Week Lag 2 Week Lag Unweighted Without Cook Time
County Specification
Unemployment -1.491*** -1.377** -1.554*** -1.505*** -1.420*** -1.395* -0.258 -4.647***
(0.382) (0.395) (0.387) (0.346) (0.336) (0.571) (0.738) (0.505)

Replacement Rate 0.167* 0.160* 0.143* 0.184* 0.185* 0.0594 0.201* 0.243***
(0.0812) (0.0596) (0.0572) (0.0721) (0.0746) (0.0715) (0.0849) (0.0527)

N 1598 1316 1598 1598 1598 1598 1564 1598
F 12.12 16.35 14.48 18.34 16.43 4.502 4.441 141.5
R-squared 0.875 0.886 0.875 0.877 0.876 0.685 0.807 0.752
County FEs YES YES YES YES YES YES YES
Week FEs YES YES YES YES YES YES YES
County * Cubic in Week YES
Standard errors in parentheses
Standard errors clustered at the county level
The units of observation are county-week pairs.
All are weighted by workforce unless otherwise specified
The unemployment claimant rate is measured using the share of people claiming unemployment insurance.
The alternative measure of the replacement rate is the gross benefit level as opposed to the net payment received.
* p < 0.05, ** p < 0.01, *** p < 0.001
Table 3. Baseline Instrumental Variables Estimates of the Determinants of County-Level Spending.
(1) (2) (3) (4)
Second Stage: Spending relative to Spending relative to Spending relative to Jan2020
Jan2020 Jan2020
Unemployment -1.778** -2.249***
(0.760) (0.732)
Replacement Rate 0.311 0.313**
(0.189) (0.143)
First Stage: Unemp. Rep. Rate Unemp. Rep. Rate

Unemployment IV 6.439** 6.439** 9.401*
(1.942) (1.941) (3.706)

Replacement Rate IV 5.160*** 0.00182 5.153***
(0.815) (0.261) (0.812)

N 510 510 510 510
F-Statistic (Excluded IVs) 10.99 40.08 5.753 23.52
County FEs YES YES YES YES
Week FEs YES YES YES YES
Standard errors in parentheses
Standard errors clustered at the county level
The units of observation are county-week pairs.
All are weighted by workforce.
The unemployment claimant rate is measured using the share of people claiming unemployment insurance.
* p < 0.05, ** p < 0.01, *** p < 0.001
Appendix Table 1: Summary Descriptive Statistics

Mean SD
Credit/Debit Spending relative to January 2020 -0.150 0.134
Share of Labor Force claiming Unemployment 0.0704 0.0406
Replacement Rate 0.815 0.352
Share of Labor Force claiming Unemployment interacted with the Replacement Rate 0.0690 0.0495
Appendix Table 2. Alternative 2SLS Estimates of the Determinants County-Level Spending.
(1) (2) (3) (4) (5) (6) (7) (8)
Baseline Alternate RR No LWA 1 Week Lag 2 Week Lag Unweighted Without Cook Time
County Specification
Unemployment -2.249** -2.189* -2.313** -1.916* -1.741* -3.239 -74.95 -6.014***
(0.732) (0.940) (0.749) (0.755) (0.773) (3.983) (2886.3) (0.299)

Replacement Rate 0.313* 0.238* 0.237* 0.264 0.245 0.182 -0.660 0.347***
(0.143) (0.105) (0.100) (0.134) (0.130) (0.168) (38.01) (0.0231)

N 510 510 420 510 510 510 510 510
First Stage F- 5.753 4.094 5.776 5.468 5.093 1.602 0.250 719.99
Statistics 23.53 18.39 22.02 23.34 23.09 9.854 19.48 874.16
Week FEs YES YES YES YES YES YES YES
County * Cubic in YES
Week
Standard errors in parentheses
Standard errors clustered at the county level
The units of observation are county-week pairs.
All are weighted by workforce unless otherwise specified
The alternative measure of the replacement rate is the gross benefit level as opposed to the net payment received.
The unemployment claimant rate is measured using the share of people claiming unemployment insurance.
*
p < 0.05, ** p < 0.01, *** p < 0.001
References
Baker, Scott R, Robert A Farrokhnia, Steffen Meyer, Michaela Pagel, and Constantine Yannelis. 2020.
'How does household spending respond to an epidemic? consumption during the 2020 covid-19
pandemic', NBER Working Paper w26949, w26949.
Brandolini, Andrea Viviano. 2018. 'Measuring employment and unemployment', IZA World of Labor.
Brandolini, Andrea, and Eliana Viviano. 2016. 'Behind and beyond the (head count) employment rate',
Journal of the Royal Statistical Society: Series A (Statistics in Society), 179: 657-81.
Card, David. 2011. 'Origins of the unemployment rate: the lasting legacy of measurement without theory',
American Economic Review, 101: 552-57.
Carroll, Christopher D, Edmund Crawley, Jiri Slacalek, and Matthew N White. 2020. 'Modeling the
Consumption Response to the CARES Act', International Journal of Central Banking.
Carroll, Christopher D, Jiri Slacalek, and Matthew N White. 2020. 'Modeling the consumption response
to the CARES Act'.
Chetty, Raj, John N Friedman, Nathaniel Hendren, and Michael Stepner. 2020. 'Real-time economics: A
new platform to track the impacts of COVID-19 on people, businesses, and communities using
private sector data', NBER Working Paper, 27431.
Chodorow-Reich, Gabriel. 2019. 'Geographic cross-sectional fiscal spending multipliers: What have we
learned?', American Economic Journal: Economic Policy, 11: 1-34.
Congressional Budget Office [CBO]. July 2020. "An Update to the Economic Outlook: 2020 to 2030." In.
Washington DC.
Cox, Natalie, Peter Ganong, Pascal Noel, Joseph Vavra, Arlene Wong, Diana Farrell, and Fiona Greig.
2020. 'Initial impacts of the pandemic on consumer behavior: Evidence from linked income,
spending, and savings data', University of Chicago, Becker Friedman Institute for Economics
Working Paper.
Elmendorf, Douglas W, and Jason Furman. 2008. "If, when, how: A primer on fiscal stimulus." In.:
Brookings Institution.
Friedman, Milton. 1957. 'The permanent income hypothesis.' in, A theory of the consumption function
(Princeton University Press).
Ganong, Peter, Pascal J Noel, and Joseph S Vavra. 2020. "US Unemployment Insurance Replacement
Rates During the Pandemic." In.: National Bureau of Economic Research.
Kaplan, Greg, G Violante, and Justin Weidner. 2014. 'The Wealthy Hand-to-Mouth', Brookings Papers on
Economic Activity, Spring.
Kaplan, Greg, and Giovanni L. Violante. 2014. 'A Model of the Consumption Response to Fiscal
Stimulus Payments', Econometrica, 82: 1199-239.
Karpman, Michael, and Gregory Acs. 2020. 'Unemployment Insurance and Economic Impact Payments
Associated with Reduced Hardship Following CARES Act', Washington, DC: Urban Institute.
Merton, Robert C. 1987. 'In Honor of Nobel Laureate, Franco Modigliani', Journal of Economic
Perspectives, 1: 145-55.
Misra, Kanishka, and Paolo Surico. 2014. 'Consumption, income changes, and heterogeneity: Evidence
from two fiscal stimulus programs', American Economic Journal: Macroeconomics, 6: 84-106.
Parker, Jonathan A, Nicholas S Souleles, David S Johnson, and Robert McClelland. 2013. 'Consumer
spending and the economic stimulus payments of 2008', American Economic Review, 103: 2530-
53.
Stevens, David W. 2007. "Employment that is not covered by state unemployment insurance laws." In.:
Center for Economic Studies, US Census Bureau.
Swagel, Phill. 2020. 'Updating CBO’s Economic Forecast to Account for the Pandemic', Blog post,
Congressional Budget Office.

File and source

File: nber-w27576.pdf
Size: 356,444 bytes
SHA-256: 381f4b70bebc069314ce1bb2dfe84975849dd3e87090a12fea40b4922ead5fcd
Our copy: nber-w27576.pdf
Original: www.nber.org