The Impact of Non-Pharmaceutical Interventions on ... · The Impact of Non-Pharmaceutical...

The Impact of Non-Pharmaceutical Interventions onUnemployment During a Pandemic∗

Edward KongHarvard University and Harvard Medical School

Daniel PrinzHarvard University

April 20, 2020

Abstract

We use high-frequency Google search data, combined with data on the announce-ment dates of non-pharmaceutical interventions (NPIs) during the COVID-19 pan-demic in U.S. states, to isolate the impact of NPIs on unemployment in an event-studyframework. Exploiting the differential timing of the introduction of NPIs (focusing onrestaurant closures) and geographic variation in the share of individuals employed infood service, we analyze how Google searches for claiming unemployment insurance(UI) varied from day to day and across locations. We translate these estimates into es-timates of UI claiming and quantify how much of the overall growth in unemploymentduring the COVID-19 pandemic was due to NPIs. We find that the announcement ofrestaurant and bar limitations is associated with an average 15 unit (43%) increase inGoogle search volume on the day of announcement as well as the following two days.We then estimate that 6.2% of the UI claims between March 14 and March 28, orabout 631,000 claims, were caused by states’ implementation of bar and restaurantlimitations. Thus, we find that 29% of the 2,177,000 claims filed in the Food andAccommodation industry were caused by bar and restaurant limitations.

∗Kong: [email protected]. Prinz: [email protected]. We thank Sam Burn, Monica Farid,Ed Glaeser, Larry Katz, Jim Stock, seminar participants at the Harvard Seminar in the Economics ofCOVID-19, and especially Tim Layton for useful comments.

1

Electronic copy available at: https://ssrn.com/abstract=3581254

mailto:[email protected]

mailto:[email protected]

1 Introduction

During a pandemic, governments may implement non-pharmaceutical interventions (NPIs)

to slow the spread of disease. Possible NPIs include shutting down businesses where social

interactions take place. For example, during the COVID-19 pandemic, most U.S. states and

cities chose to shut down bars and restaurants (in most cases only allowing take-out and

delivery), while some states shut down all businesses deemed non-essential.

Measuring the impact of the implementation and the later relaxation of NPIs on economic

outcomes such as employment is not straightforward. Pandemics may impact the economy

through a number of channels. They have direct effects on consumer demand for particular

goods and services, as individuals avoid public places like malls. They may directly decrease

labor supply if managers reduce worker density to avoid outbreaks at their firms. Such

direct impacts operate at the same time NPIs are implemented. This makes it difficult to

distinguish between economic impacts caused by NPIs and economic impacts caused by the

“direct pandemic effects” described above. A further empirical challenge is that data on

employment and unemployment is not readily available at the frequency at which policies

change during pandemics. For example, data on U.S. unemployment insurance (UI) claims

are released weekly, but during the COVID-19 pandemic, information on new cases and the

implementation of new NPIs changed daily.

In this paper, we use high-frequency Google search data, combined with data on the exact

dates of the announcement and implementation of NPIs during the COVID-19 pandemic

in U.S. states to isolate the impact of NPIs on UI claims in an event study framework.

Exploiting the differential timing of the introduction of NPIs (focusing on restaurant closures)

and geographic variation in locations’ shares of employment in the affected industries, we

analyze how Google searches for claiming unemployment varied from day to day and across

states. We then translate these estimates into estimates of UI claiming and quantify how

much of the overall growth in unemployment during the COVID-19 was due to NPIs.

We find that the announcement of restaurant and bar limitations is associated with an

2


average 15 unit (43%) increase in Google search volume on the day of announcement as well

as the following two days. We then estimate that 6.2% of the UI claims between March 14

and March 28, or about 631,000 claims, were caused by states’ implementation of bar and

restaurant limitations. Thus, we find that 29% of the 2,177,000 claims filed in the Food and

Accommodation industry were caused by bar and restaurant limitations.

Our work contributes to the emerging literature studying the impact of NPIs adopted

during the COVID-19 pandemic and other pandemics on unemployment. Most closely re-

lated to our work, Ugarov (2020) presents a calculation of lost productivity related to NPIs

and Correia, Luck and Verner (2020) estimate the economic impact of public health inter-

ventions during the 1918 Flu epidemic. To our knowledge, we are the first to present quasi-

experimental evidence on the employment consequences of NPIs adopted during COVID-19

pandemic.

More broadly, we contribute to the literature that examines the economic consequences

of the COVID-19 pandemic. Much of this literature (Atkeson, 2020; Bethune and Korinek,

2020; Eichenbaum, Rebelo and Trabandt, 2020; Jorda, Singh and Taylor, 2020; Guerrieri,

Lorenzoni, Straub and Werning, 2020; Ludvigson, Ma and Ng, 2020) presents macroeco-

nomic simulations of the consequences of the pandemic and evaluates different pandemic

scenarios and the consequences of potential policies using macroeconomic models. Focusing

on labor markets, Alon, Doepke, Olmstead-Rumsey and Tertilt (2020) study the potential

consequences of the COVID-19 pandemic on gender equality and Dingel and Neiman (2020)

provide estimates of what share of jobs can be done from home. Using the Nielsen Homescan

panel, Coibion, Gorodnichenko and Weber (2020) find that job loss during the COVID-19

pandemic has been higher than implied by new UI claims, but many individuals who lost

their jobs are not actively looking for new jobs. This suggests that the unemployment rate

is likely to rise less than expected but there is likely to be a large drop in labor force par-

ticipation. Other papers study short-term aggregate impacts on economic activity (Lewis,

Mertens and Stock, 2020), consumption impacts (Baker et al., 2020), heterogeneity across

3


firms (Bartik et al., 2020; Hassan, Hollander, van Lent and Tahoun, 2020), as well as the

economic uncertainty (Baker, Bloom, Davis and Terry, 2020) induced by the pandemic.

We build on work in labor economics and beyond that has used Google search data to

study questions that are difficult to study with more traditional survey and administrative

datasets. Within labor economics, our work is most closely related to Goldsmith-Pinkham

and Sojourner (2020) who use Google search volumes to predict unemployment insurance

claims during the COVID-19 pandemic and earlier work by Baker and Fradkin (2017) who

estimate measures of job search intensity based on Google search and other data and study

the consequences of unemployment insurance policy changes for these measures of search.

The remainder of this paper proceeds as follows. We provide background information on

the COVID-19 pandemic and NPI responses to the pandemic in Section 2. We then describe

our data in Section 3 and methods in Section 4. We present our results in Section 5. In

Section 6, we conclude with brief discussion.

2 Background

2.1 The COVID-19 Pandemic in the U.S.

In January 2020, coronavirus disease 2019 (COVID-19), an infectious disease caused by by-

severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) spread to the United States.

Within three months, COVID-19 had spread to all states and caused between 30,000 and

40,000 deaths (Centers for Diseases Control and Prevention, 2020; Johns Hopkins University,

2020).

COVID-19 is a highly infectious disease: most studies suggest that its basic reproduction

number (R0) is 2.2-2.7. Its most common symptoms include fever, dry cough, fatigue,

sputum production, and shortness of breath. Other symptoms may include loss of smell,

muscle or joint pain, sore throat, headache, chills, nausea or vomiting, nasal congestion and

diarrhoea. In a share of cases, COVID-19 progresses to pneumonia, multi-organ failure, and

4


death.

Currently no vaccine or specific treatment exists for COVID-19. U.S. states and cities,

like most countries dealing with COVID-19, have adopted quarantine and social distancing

measures, including non-pharmaceutical interventions that include shutting down certain

types of businesses or severely limiting their operations.

2.2 Non-Pharmaceutical Interventions

Common non-pharmaceutical interventions adopted by U.S. states to mitigate the spread of

COVID-19 include stay-at-home orders, mandatory quarantines for travelers, non-essential

business closures, large gatherings bans, state-mandated school closures, and bar and restau-

rant limits. By April 20, 2020, all U.S. states with the exceptions of Arkansas, Iowa, Ne-

braska, North Dakota, South Dakota, and Wyoming have issued some form of a stay at home

order. By the same time, all states with the exceptions of Arkansas, Minnesota, Nebraska,

South Dakota, Texas, Utah, and Wyoming have implemented some form of non-essential

business closures. Strict bar and restaurants limits had been imposed in all states with the

exception of South Dakota. All other states have closed restaurants and bars except for

takeout and delivery, with the exceptions of Kansas and New Mexico which allow limited

on-site service and Oklahoma where restaurants and bars are closed except for takeout and

delivery only in affected counties (The Henry J. Kaiser Family Foundation, 2020).

Importantly for our analysis, while almost all states eventually implemented these NPIs,

initial implementation was staggered. For example, restaurants and bars were limited to

takeout and delivery in 35 states by March 18, while 4 states still had restaurants and bar

operating a week later. Likewise, 7 states closed all non-essential businesses as early as

March 20th, whereas 16 states had not closed non-essential businesses by April 1st.

5


3 Data

We combine data on internet searches from Google Trends, data on NPI implementation

dates from state announcements, as well as state economic data (e.g., industry composition)

and data on the spread of COVID-19.

3.1 Search Data

We use internet search data released through Google Trends for the February 1, 2020 to

March 31, 2020 period.1 Google Trends releases data on relative search intensity for different

terms by day and geographic location (including state, metro area, and city). Because

Google only releases relative search volumes, throughout our analysis, we will normalize

search volumes setting the highest search volume during our time period in California as

100.2

3.2 NPI timing data

We identify states that have implemented NPIs from The Henry J. Kaiser Family Foundation

(2020). For each state and NPI, we identify the precise date on which the NPI was first

announced. In cases where multiple announcements pertained to the same NPI, we use the

first recorded announcement. In the current version of this paper, we report results for the

NPI of restaurant and bar limitations, which restricts restaurants and bars to take-out and

delivery operations. Forbidding dine-in services is intended to reduce close-quarters, person-

to-person contacts and thus slow the spread of the virus, but also reduces demand for the

affected establishments, which may need to lay off workers in response.

1Approximately 90% of internet searches in the U.S. happen on Google.2We use data originally collected and publicly released by Goldsmith-Pinkham and Sojourner (2020).

6


3.3 Other data

We identify the share of each state’s employed individuals working in each sector using the

American Community Survey 5-year extract for 2013-2017 (U.S. Census Bureau, 2020). We

use confirmed COVID-19 cases and deaths from Dong, Du and Gardner (2020) and Johns

Hopkins University (2020). We use data on industry-level unemployment growth from March

14-28 in Massachusetts from the Massachusetts Executive Office of Labor and Workforce

Development (2020). Total initial UI claims filed at a national level for weeks ending March

21 and March 28 are derived from weekly news releases from the U.S. Department of Labor

(2020).

4 Empirical Framework

Measuring the causal impact of NPIs on employment is not straightforward. Pandemics may

impact employment through a number of channels. They have effects on consumer demand

for particular goods and services, as individuals avoid public places like malls. They may

decrease labor supply if managers reduce worker density to avoid outbreaks at their firms.

These responses, which we call “direct pandemic effects,” occur at the same time NPIs are

implemented, confounding estimates of the impact of NPIs. A further empirical challenge

is that data on employment and unemployment is not readily available at the frequency at

which policies change during pandemics. Data on U.S. unemployment insurance (UI) claims

are released weekly, but during the COVID-19 pandemic, information on new cases and the

implementation of new NPIs changed daily.

We use high-frequency Google search data, combined with data on the exact dates of the

announcement and implementation of non-pharmaceutical interventions (NPIs) during the

COVID-19 pandemic in U.S. states and cities to isolate the impact of NPIs on UI claims in

an event study framework. Exploiting the differential timing of NPI introductions (focusing

on restaurant and bar limits) and geographic variation in locations’ shares of employment in

7


the affected industries, we analyze how Google searches for claiming unemployment varied

from day to day and across cities and states. We then translate these estimates into estimates

of UI claiming and quantify how much of the overall growth in unemployment during the

COVID-19 was due to NPIs.

4.1 Estimation

To quantify the impact of NPIs on unemployment exploiting the differential timing of the

introduction of policies in different states, we estimate an event study regression of the form:

Sit =6∑

τ=−7

γτ × 1 {r = τ}+ αi + αt + εit (1)

where Sit is google search volume in state i and date t, r denotes the days relative to NPI

announcement, and αi, αt are state and date fixed effects. The coefficients of interest γτ

estimate the search volume differential for each day τ relative to the NPI announcement.

We normalize γτ=−1 = 0 and cluster standard errors at the state level.

4.2 Robustness

To assess the robustness of our results, we estimate several alternative specifications. First,

we re-estimate our main event-study specification, excluding California and Washington

which had a large number of cases earlier than other states and New York that was partic-

ularly strongly hit by the pandemic.

Second, we re-estimate our main event-study specification weighting each state by its

total employment. This ensures that our results are not driven by small states that may be

on different trajectories from other areas.

Third, we estimate a difference-in-differences event study specification where we compare

“early adopters” and “late or never adopters”. We label states as “early adopters” if they

implemented an NPI within a week of the first state (March 14-21). We label states as “late

8


adopters” if they implemented an NPI on or after March 22, or not at all. We then run the

following regression:

Sit =March21∑τ=March7

δτ × 1 {Early Adopter}+ β × 1 {Early Adopter}+ ξt + µit, (2)

where the δτ coefficients describe the differential evolution of search volume in “early adopters”

relative to “late adopters.” We normalize δτ=March13 = 0, so β captures the average differ-

ence in Sit between early and late adopters on March 13th. The ξt denote date fixed effects

which control for the time trend in search behavior. We limit to the period of March 7 to

March 21, which allows for 7 days where no states have implemented bar and restaurant

limits, followed by 7 days where the early adopters began implementing limits but the late

adopters did not. The late adopters are thus never treated during our estimation window.

Fourth, we re-estimate our main event-study specification with additional controls for

case growth and deaths, both interacted with state dummies. Case growth is defined as the

additional cases relative to the previous day and is intended to control for the possibility

that individuals change their Google search behavior in response to news reports of new

cases. Controlling for cumulative deaths follows a similar logic. Interacting both variables

with state dummies allows the effects of case growth and deaths to vary by state. This

specification ensures that our results are not driven by differential case growth and deaths.

Finally, to further demonstrate that the NPI-timing variation we utilize is not not driven

by the different epidemiological experiences of each state, we separate states into those that

registered their first COVID-19 death early in the epidemic (on or before March 19) and

those that registered their first COVID-19 death later (after March 19). We use March 19

as the cutoff date because it is the average (and median) date of the first COVID-19 death

across states. We then run our main specification (Equation 1) separately for both sets of

states and assess whether we recover the same qualitative pattern as in our main pooled

specification.

9


4.3 Identification

We are interested in identifying the causal impact of introducing an NPI on unemployment

claiming in a given state or city, where the counterfactual is the trajectory of UI claims in

a world where the NPI was not enacted. Our identification strategy aims to separate this

“causal NPI effect” from the “direct pandemic effects” described above. By separating these

two effects, we can determine the number of UI claims caused by the NPI as a share of the

total increase in UI claims.

Our empirical strategy relies on several assumptions about how individuals respond to

NPI announcements. Our primary assumption is that workers fully internalize the effects

of the NPI on their probability of being unemployed in the near future. Workers who face

an increased unemployment probability will be more likely to search for the term “file for

unemployment,” which has been shown to predict actual growth in UI claims (Goldsmith-

Pinkham and Sojourner (2020)). To the extent that workers under-react (over-react) to the

NPI announcement, our estimate of the causal NPI effect will be biased downward (upward).

Because we adopt an event-study approach to estimation, we require the usual identifi-

cation assumption that the observed untreated outcomes of units that are not (yet) treated

(the late adopters) can approximate the unobserved potential outcomes for units that have

already been treated (the early adopters), conditional on observables. This is similar to the

parallel-trends assumption underlying difference-in-differences (DID) estimators.

Lastly, we assume that individuals do not anticipate the implementation of NPIs. For

states that were early to adopt restaurant and bar limitations, this seems like a reasonable

assumption, as Google search volume was essentially zero before the NPIs were announced.

This assumption is more problematic for states that adopted restaurant and bar limits later

(or not at all). Workers in these states might reasonably have anticipated that such limits

were impending, and this may have driven increases in search behavior. We assume that

searches that occurred prior to NPI announcements are due to direct pandemic effects and

not due to anticipation of future NPIs. To the extent that firms and workers anticipated NPI

10


announcements in late-adopting states and this led to higher Google searches, our estimates

of the causal NPI effect will be biased toward zero. However, our results would still be policy-

relevant: they estimate the effect of an NPI policy taking as given firms’ expectations. The

policy-relevant treatment effect of the intervention accounts for the possibility that the NPI

results in a smaller unemployment increase because firms have already laid off workers in

anticipation of the NPI.

If the above assumptions hold, then the relative-time effects (given by the γτ coefficients)

estimate the causal impact of the NPI announcement on Google searches, and the estimated

nonlinear time trend (given by the αt coefficients) estimates the growth in search volume

that is unrelated to the NPI announcement.

4.4 Quantifying the Impact of NPIs on UI claims

We use a method proposed by Goldsmith-Pinkham and Sojourner (2020) to translate our

estimates of the impact of NPIs on internet searches into estimates of the impact of NPIs

on actual UI claims.

Our key assumption is that the number of UI claims in a given period is proportional to

the area under the curve defined by the search intensity over the same period.3 With this

assumption, the integral under the estimated NPI effect (γτ ) is proportional to the number

of UI claims caused by the NPI. By comparing this integral to the integral under the time

trend (αt), we can dis-aggregate the overall increase in UI claims into the causal NPI effect

and the direct pandemic effect. Formally, let INPI denote the integral under the event-study

coefficients γτ for τ >= 0 and Iα,t1,t2 denote the integral under the date fixed effects αt

between t1 and t2 (which estimate the direct pandemic effect). The share of the UI claims

3In general, the coefficient of proportionality is difficult to interpret, since over any requested time window,the Google search data are always normalized so that the maximum search intensity equals 100.

11


between t1 and t2 that was caused by the NPI can be estimated as:

Share of UI claims caused by NPI =INPI

INPI + Iα,t1,t2. (3)

Because NPIs can have industry-specific impacts, another quantity of interest is the share

of UI claims in a given industry that was caused by the NPI. If the NPI targets industry s

and ρs ∈ [0, 1] is the industry s share of the overall increase in UI claims, then the share of

UI claims for s that was caused by the NPI can be estimated as:

Share of UI claims in s caused by NPI =INPI

ρs × (INPI + Iα,t1,t2). (4)

Defining the appropriate time window [t1, t2] is challenging and will affect estimation of

the shares defined above. We estimate the above shares for bar and restaurant limitations

using a window of t1 = March 14, when the first state adopted limits, up through t2 =

March 28, approximately when the last states adopted limits. This also allows us to simply

use two periods worth of the weekly UI claims data, avoiding interpolation issues.

5 Results

5.1 Event-study Estimates

Figure 2 shows our main event study estimates for restaurant and bar closures. We find that

the announcement of these policies is on average associated with an approximate 15 unit

increase in internet search volume on the day of announcement as well as the following two

days. This represents a 43% increase in search volume relative to a mean (normalized) search

volume of 35.1. This transitory effect diminishes after three days and internet search volume

falls close to its pre-announcement level. This transitory response may reflect an “impulse

response” effect of announcements: workers affected by the restaurant and bar closures may

search online intensively but then search less after they locate the appropriate resources for

12


filing a UI claim.

Figure 3 shows the same estimates separately for states with high (above-median) and

low (below-median) food service employment shares. The effect of a restaurant and bar

closure announcement is larger for states with a high share of their residents employed in

food service.

5.2 Robustness

Figure 5 shows alternative versions of our main event study estimate in Figure 2. Panel

(a) repeats our main estimate in Figure 2. Panel (b) re-estimates the same event study,

excluding California, Washington, and New York. Panel (c) re-estimates the same event

study, weighted by total employment in the state. Panel (d) re-estimates the same event

study, excluding California, Washington, and New York and weighted by total employment

in the state. We find that our results are largely similar under these alternative specifications.

Figure 6 shows an alternative specification comparing “early adopters” (March 14-21)

with “late or never adopters” (after March 22 or never). Panel (a) presents baseline results.

Trends for early and late adopters are identical until the first NPI announcement, at which

point the early adopter states experience a jump in search volume that is sustained through

additional announcements by early adopter states. Panel (b) shows the analysis with Wash-

ington, California, and New York excluded. Panels (c) and (d) repeat the analyses in Panels

(a) and (b), with states weighted by their total employment. We find qualitatively similar

results in each of the specifications.

Figure 7 shows an alternative specification that includes controls for case growth and

deaths, both interacted with state dummies to allow their effects to vary by state. The

estimated coefficients on these controls are frequently significant (not shown), but the inclu-

sion of these controls has a very small effect on our results. This suggests that our results

are not driven by correlation between NPI timing and state trajectories of the epidemic (as

measured by case growth and deaths).

13


Figure 8 shows an alternative specification where we separately estimate the baseline

event study (Equation 1) for states that had their first death early on in the epidemic (on

or before March 19) or later (after March 19) (see Figure A2 for the distribution of these

first-death dates). The qualitative pattern of point estimates is very similar across the two

sub-samples. Again, this suggests that our estimated NPI effect is not confounded by omitted

variables governing the progression or “first signs” of COVID-19 outbreaks by state.

5.3 Quantitative impact of NPIs on UI claims

We use the method outlined in Section 4.4 to compute the number and share of UI claims

caused by the introduction of bar and restaurant limits. The results from our main speci-

fication (Figure 2) suggest that the effects of the NPI on searches was limited to the first

four days after and including the announcement date. Accordingly, we compute INPI as the

integral from 0 to 3 and obtain INPI ≈ 47.4. We compute Iα,t1,t2 ≈ 712.7 by integrating

the estimated time trend from the event study from t1 = March14 through t2 = March28.

Figure 4 visualizes these integrals. Finally, we impute the share of UI claims between March

14 and March 28 that is due to the Food and Accommodation industry using Massachusetts

data4, which indicates a share of 21.4%. Lastly, we use a total number of UI claims (sea-

sonally adjusted) between March 14 and March 28 of 10,174,000 (U.S. Department of Labor

(2020)).

Combining these data points, we compute that 6.2% of the UI claims between March

14 and March 28, or about 631,000 claims, was due to states’ implementation of bar and

restaurant limitations. We infer that about 2,177,000 claims are filed in the Food and

Accommodation industry and that 29% of these were caused by limitations on bars and

restaurants.

4We could not locate the most recent data on industry-specific UI claims at the national level, so we usethe share from Massachusetts as a proxy measure (see Figure A1)

14


6 Discussion

The effects of introducing an NPI and the effect of relaxing the same NPI may not be equal

and opposite. In the case of unemployment, if jobs have been lost due to an NPI, rescinding

the same NPI may not recover all of these lost jobs if employer-worker relationships were

severed. This is itself a prominent concern: if the costs of rebuilding these linkages is high,

then the economy may not bounce back even if NPIs are lifted and direct pandemic effects

mollified (for example, due to an effective therapeutic or vaccine).

Furthermore, the effect of relaxing an NPI depends on firms’ beliefs about how long the

NPI would have lasted. If firms had already anticipated an early relaxation of an NPI (for

example, because of statements made by politicians), then actually relaxing the NPI may

have little additional impact5.

In general, the effect of rescinding an NPI will depend on the costs of re-hiring workers

as well as agents’ prior beliefs on when the original NPI would have been rescinded (when

NPIs are announced, agents likely do not expect them to last forever).

Thus, our estimates can be interpreted as an upper bound on the number of jobs that

could be recovered by relaxing a given NPI; the actual number of recovered jobs may be

much fewer. Even if this upper bound is close to the true number, our estimates suggest

that only a modest fraction (6.2%) of the increase in unemployment can be attributed to a

prominent NPI (bar and restaurant limits) and that this NPI accounts for only 29% of the

growth in unemployment for the Accommodation and Food Service industry.

References

Alon, Titan M, Matthias Doepke, Jane Olmstead-Rumsey, and Michele Tertilt.

2020. “The Impact of COVID-19 on Gender Equality.” National Bureau of Economic

5Conversely, this also has implications for how one interprets the original NPI announcement. If firmsanticipated that the NPI would be in effect for only a short period, then any unemployment effects would beattributed to the short shutdown period, which in turn scales up estimates of the “per week” unemploymenteffects of the NPI.

15


Research Working Paper 26947.

Atkeson, Andrew. 2020. “What Will Be the Economic Impact of COVID-19 in the US?

Rough Estimates of Disease Scenarios.” National Bureau of Economic Research Working

Paper 26867.

Baker, Scott R., and Andrey Fradkin. 2017. “The Impact of Unemployment Insurance

on Job Search: Evidence from Google Search Data.” Review of Economics and Statistics,

99(5): 756–768.

Baker, Scott R, Nicholas Bloom, Steven J Davis, and Stephen J Terry. 2020.

“COVID-Induced Economic Uncertainty.” National Bureau of Economic Research Work-

ing Paper 26983.

Baker, Scott R, R.A. Farrokhnia, Steffen Meyer, Michaela Pagel, and Constan-

tine Yannelis. 2020. “How Does Household Spending Respond to an Epidemic? Con-

sumption During the 2020 COVID-19 Pandemic.” National Bureau of Economic Research

Working Paper 26949.

Bartik, Alexander W, Marianne Bertrand, Zoe B Cullen, Edward L Glaeser,

Michael Luca, and Christopher T Stanton. 2020. “How Are Small Businesses Ad-

justing to COVID-19? Early Evidence from a Survey.” National Bureau of Economic


Bethune, Zachary A, and Anton Korinek. 2020. “Covid-19 Infection Externalities:

Trading Off Lives vs. Livelihoods.” National Bureau of Economic Research Working Paper

27009.

Centers for Diseases Control and Prevention. 2020. “Cases of Coronavirus Dis-

ease (COVID-19) in the U.S.” https: // www. cdc. gov/ coronavirus/ 2019-ncov/

cases-updates/ cases-in-us. html .

16


https://www.cdc.gov/coronavirus/2019-ncov/cases-updates/cases-in-us.html

https://www.cdc.gov/coronavirus/2019-ncov/cases-updates/cases-in-us.html

Coibion, Olivier, Yuriy Gorodnichenko, and Michael Weber. 2020. “Labor Mar-

kets During the COVID-19 Crisis: A Preliminary View.” National Bureau of Economic


Correia, Sergio, Stephan Luck, and Emil Verner. 2020. “Pandemics Depress the

Economy, Public Health Interventions Do Not: Evidence from the 1918 Flu.” https:

// ssrn. com/ abstract= 3561560 .

Dingel, Jonathan I, and Brent Neiman. 2020. “How Many Jobs Can be Done at Home?”

National Bureau of Economic Research Working Paper 26948.

Dong, Ensheng, Hongru Du, and Lauren Gardner. 2020. “An interactive web-based

dashboard to track COVID-19 in real time.” The Lancet Infectious Diseases.

Eichenbaum, Martin S, Sergio Rebelo, and Mathias Trabandt. 2020. “The Macroe-

conomics of Epidemics.” National Bureau of Economic Research Working Paper 26882.

Goldsmith-Pinkham, Paul, and Aaron Sojourner. 2020. “Predicting Initial Un-

employment Insurance Claims Using Google Trends.” https: // paulgp. github. io/

GoogleTrendsUINowcast/ google_ trends_ UI. html .

Guerrieri, Veronica, Guido Lorenzoni, Ludwig Straub, and Ivan Werning. 2020.

“Macroeconomic Implications of COVID-19: Can Negative Supply Shocks Cause Demand

Shortages?” National Bureau of Economic Research Working Paper 26918.

Hassan, Tarek Alexander, Stephan Hollander, Laurence van Lent, and Ahmed

Tahoun. 2020. “Firm-level Exposure to Epidemic Diseases: Covid-19, SARS, and H1N1.”

National Bureau of Economic Research Working Paper 26971.

Johns Hopkins University. 2020. “COVID-19 Dashboard by the Cen-

ter for Systems Science and Engineering (CSSE) at Johns Hopkins Univer-

17


https://ssrn.com/abstract=3561560

https://ssrn.com/abstract=3561560

https://paulgp.github.io/GoogleTrendsUINowcast/google_trends_UI.html

https://paulgp.github.io/GoogleTrendsUINowcast/google_trends_UI.html

sity (JHU).” https: // www. arcgis. com/ apps/ opsdashboard/ index. html#

/bda7594740fd40299423467b48e9ecf6 .

Jorda, Oscar, Sanjay R Singh, and Alan M Taylor. 2020. “Longer-run Economic Con-

sequences of Pandemics.” National Bureau of Economic Research Working Paper 26934.

Lewis, Daniel, Karel Mertens, and James H Stock. 2020. “U.S. Economic Activity

During the Early Weeks of the SARS-Cov-2 Outbreak.” National Bureau of Economic


Ludvigson, Sydney C, Sai Ma, and Serena Ng. 2020. “Covid19 and the Macroeconomic

Effects of Costly Disasters.” National Bureau of Economic Research Working Paper 26987.

Massachusetts Executive Office of Labor and Workforce

Development. 2020. “Massachusetts Weekly Initial Unemploy-

ment Claimant Data (04-02-20).” https: // www. mass. gov/ news/

massachusetts-weekly-initial-unemployment-claimant-data-04-02-20 .

The Henry J. Kaiser Family Foundation. 2020. “State Data and Policy Ac-

tions to Address Coronavirus.” https: // www. kff. org/ health-costs/ issue-brief/

state-data-and-policy-actions-to-address-coronavirus/ .

Ugarov, Alexander. 2020. “Inclusive Costs of NPI Measures for COVID-19 Pan-

demic: Three Approaches.” https: // www. medrxiv. org/ content/ early/ 2020/ 03/

30/ 2020. 03. 26. 20044552 .

U.S. Census Bureau. 2020. “American Community Survey 2013-2017 5-year Public-Use

Microdata Sample (PUMS).” https: // www. census. gov/ programs-surveys/ acs/

data/ pums. html .

U.S. Department of Labor. 2020. “Office of Unemployment Insurance Weekly Claims

Report.” https: // oui. doleta. gov/ unemploy/ claims_ arch. asp .

18


https://www.arcgis.com/apps/opsdashboard/index.html#/bda7594740fd40299423467b48e9ecf6

https://www.arcgis.com/apps/opsdashboard/index.html#/bda7594740fd40299423467b48e9ecf6

https://www.mass.gov/news/massachusetts-weekly-initial-unemployment-claimant-data-04-02-20

https://www.mass.gov/news/massachusetts-weekly-initial-unemployment-claimant-data-04-02-20

https://www.kff.org/health-costs/issue-brief/state-data-and-policy-actions-to-address-coronavirus/

https://www.kff.org/health-costs/issue-brief/state-data-and-policy-actions-to-address-coronavirus/

https://www.medrxiv.org/content/early/2020/03/30/2020.03.26.20044552

https://www.medrxiv.org/content/early/2020/03/30/2020.03.26.20044552

https://www.census.gov/programs-surveys/acs/data/pums.html

https://www.census.gov/programs-surveys/acs/data/pums.html

https://oui.doleta.gov/unemploy/claims_arch.asp

Figure 1: Timing of bar and restaurant limits

05

1015

Freq

uenc

y

3/14/2

020

3/15/2

020

3/16/2

020

3/17/2

020

3/18/2

020

3/19/2

020

3/20/2

020

3/21/2

020

3/22/2

020

3/23/2

020

3/24/2

020

3/25/2

020

3/30/2

020

Missing

Note: Figure shows the number of states announcing restaurant and bar limitations by date.The underlying data were originally collected by The Henry J. Kaiser Family Foundation(2020). Note that the X-axis omits dates with no announcements and that “Missing” indi-cates states that did not announce restaurant and bar limitations as of April 5th.

19


Figure 2: Event study of the response of search volume to NPI announcement

-10

010

2030

40R

elat

ive

sear

ch v

olum

e

-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6Days relative to announcement

Note: Figure shows event study estimates of the impact of restaurant and bar closure an-nouncements on internet search volume. These estimates are based on Equation (2). Theday prior to the announcement is normalized to zero and standard errors are clustered atthe state level. For more details, see Section 4.1.

20


Figure 3: Response of search volume to NPI announcement by share employed in food service

-20

020

4060

Rel

ativ

e se

arch

vol

ume


Low Food Services Share High Food Services Share

Note: Figure shows event study estimates of the impact of restaurant and bar closure an-nouncements on internet search volume, separately for states with low (below-median) andhigh (above-median) food service employment shares. These estimates are based on Equa-tion (2). The day prior to the announcement is normalized to zero and standard errors areclustered at the state level. For more details, see Section 4.1.

21


Figure 4: Dis-aggregating Unemployment Effects by Policy and Pandemic Causes

(a) Effect of bar and restaurant limits-1

00

1020

3040

Rel

ativ

e se

arch

vol

ume


(b) Time trend

050

100

Rel

ativ

e se

arch

vol

ume

01mar2020 07mar2020 14mar2020 20mar2020 27mar2020Date

Note: Figure shows how NPI effects and time trends from event studies are used to decomposeUI claims growth into NPI and direct pandemic effects. The left panel shows the impulseresponse in searches following NPI announcements. The right panel shows the overall timetrend of searches during the same period. The areas under the curves represent the shares ofunemployment growth associated with the NPI and the direct pandemic effects, respectively.For more details, see Section 4.4.

22


Figure 5: Event study of the response of search volume to NPI announcement

(a) Main Specification-1

00

1020

3040

Rel

ativ

e se

arch

vol

ume


(b) Excluding California, Washington, and NewYork

-10

010

2030

40R

elat

ive

sear

ch v

olum

e


(c) Weighted by Total Employment

-10

010

2030

Rel

ativ

e se

arch

vol

ume


(d) Excluding California, Washington, and NewYork and Weighted by Total Employment

-10

010

2030

40R

elat

ive

sear

ch v

olum

e


Note: Figure shows alternative versions of our main event study estimate, based on Equa-tion (1). Panel (a) repeats our main event study estimate from Figure 2. Panel (b) re-estimates the same event study, excluding California, Washington, and New York. Panel(c) re-estimates the same event study, weighted by total employment in the state. Panel(d) re-estimates the same event study, excluding California, Washington, and New York andweighted by total employment in the state. For more details, see Section 4.1.

23


Figure 6: Comparison of early and late adopters

(a) Baseline

-20

020

4060

07mar2020 10mar2020 14mar2020 17mar2020 21mar2020x_dt

(b) Excluding California, Washington, and NewYork

-20

020

4060


(c) Weighted by Total Employment

-10

010

2030


(d) Excluding California, Washington, and NewYork and Weighted by Total Employment

-10

010

2030

40


Note: Figure shows difference-in-differences estimates of the impact of restaurant and barclosure announcements on internet search volume. These estimates are based on Equation (2)and compare “early adopters” (March 14-21) with “late or never adopters” (after March 22or never). Panel (a) presents baseline results. Panel (b) shows the analysis with Washington,California, and New York excluded. Panels (c) and (d) repeat the analyses in Panels (a)and (b), with states weighted by their total employment. The vertical lines represent thefirst NPI announcement, the median announcement date among early adopters, and the lastannouncement date among early adopters. Standard errors are clustered at the state level.For more details, see Section 4.2.

24


Figure 7: Event study controlling for case growth and deaths

(a) Main Specification

-10

010

2030

40R

elat

ive

sear

ch v

olum

e


(b) With controls, states weighted equally

-20

020

40R

elat

ive

sear

ch v

olum

e


(c) With controls, weighting states by total employ-ment

-10

010

2030

Rel

ativ

e se

arch

vol

ume


Note: Figure shows alternative versions of our main event study estimate, based on Equation(1). Panel (a) repeats our main event study estimate from Figure 2. Panel (b) repeats ourmain event study estimate from Figure 2, but includes controls for case growth and numberof deaths, both interacted with state dummies to allow the effect of case growth and deathsto vary by state. Case growth is defined as the increase in cases relative to the prior day.Panel (c) re-estimates the same event study, but weights states by their total number ofemployed individuals. For more details, see Section 4.2.

25


Figure 8: Event study for states with early and later signs of an outbreak

(a) Main Specification

-10

010

2030

40R

elat

ive

sear

ch v

olum

e


(b) States with an early first COVID-19 death

-20

020

40R

elat

ive

sear

ch v

olum

e


(c) States with a later first COVID-19 death

-20

020

4060

Rel

ativ

e se

arch

vol

ume


Note: Figure shows alternative versions of our main event study estimate, based on Equation(1). Panel (a) repeats our main event study estimate from Figure 2. Panel (b) repeats ourmain event study estimate from Figure 2, but limits to the 26 states that registered theirfirst COVID-19 death on or before March 19. The average date of the first COVID-19 deathwas March 14 for these states. Panel (c) shows the same specification, but limited to statesthat registered their first COVID-19 death later on in the pandemic, after March 19. Theaverage date of the first COVID-19 death among these states was March 24. See Figure A2for the distribution of when states registered their first COVID-19 death. For more details,see Section 4.2.

26


A Additional Figures and Tables

Appendix Figure A1: Unemployment Insurance claims by sector in Massachusetts

0 5 10 15 20% share of claims

Accommodation and food servicesHealth care and social assistance

Retail tradeInformation Not Available

ConstructionOther services, except public administration

Professional and technical servicesManufacturing

Administrative and waste servicesTransportation and warehousing

Arts, entertainment, and recreationWholesale trade

Educational servicesReal estate and rental and leasing

Public AdministrationInformation

Management of companies and enterprisesFinance and insurance

Agriculture, forestry, fishing and huntingUtilities

Mining, quarrying, and oil and gas extraction

Note: figure shows new UI claims for Massachusetts between March 14 and 28, broken downby industry. These data are collected from Massachusetts Executive Office of Labor andWorkforce Development (2020).

1


Appendix Figure A2: Histogram of when states registered their first COVID-19 death

050

100

150

Freq

uenc

y

01mar2020 07mar2020 14mar2020 20mar2020 27mar2020Date of first death

Note: Figure shows the timing of when states registered their first COVID-19 death. Dataon COVID-19 deaths by state was obtained from Johns Hopkins University (2020).

2


The Impact of Non-Pharmaceutical Interventions on ... · The Impact of Non-Pharmaceutical...

Documents

Transcript of The Impact of Non-Pharmaceutical Interventions on ... · The Impact of Non-Pharmaceutical...