The Unequal Effects Of The Health Economy Trade Off During The Covid 19 Pandemic
Summary
A research article, The unequal effects of the health-economy trade-off during the COVID-19 pandemic, published in Nature Human Behaviour, Volume 8, February 2024, with Marco Pangallo as corresponding author. The authors introduce an agent-based model simulating epidemic and economic outcomes across industries, occupations and income levels for a synthetic population of 416,442 individuals representing the New York-Newark-Jersey City metro area. The model is calibrated to the first wave of COVID-19 there, with fear of infection reducing consumption. The article reports a similar epidemic-economic trade-off from both behavior change and mandated non-pharmaceutical interventions, strongest for low-income workers in in-person, customer-facing jobs. It is published under a Creative Commons Attribution 4.0 International License.
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
“The Unequal Effects Of The Health Economy Trade Off During The Covid 19 Pandemic.”
Original at doi.org. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/the-unequal-effects-of-the-health-economy-trade-off-during-the-covid-1-95a7d72630/
, SHA-256 71634f34b49efdd1c759c67a7cfc6ea83657fc12350c74c9dc11b18da2363c21.Full text
nature human behaviour
Article https://doi.org/10.1038/s41562-023-01747-x
The unequal effects of the health–economy
trade-off during the COVID-19 pandemic
Received: 8 December 2022 Marco Pangallo 1 , Alberto Aleta 2, R. Maria del Rio-Chanona 3,
Anton Pichler 3, David Martín-Corral 4, Matteo Chinazzi 5,6,
Accepted: 5 October 2023
François Lafond 7, Marco Ajelli 8, Esteban Moro 4,9, Yamir Moreno ,
1,2,3
Published online: 16 November 2023 Alessandro Vespignani 5 & J. Doyne Farmer7,10
Check for updates
Despite the global impact of the coronavirus disease 2019 pandemic,
the question of whether mandated interventions have similar economic
and public health effects as spontaneous behavioural change remains
unresolved. Addressing this question, and understanding differential
effects across socioeconomic groups, requires building quantitative and
fine-grained mechanistic models. Here we introduce a data-driven, granular,
agent-based model that simulates epidemic and economic outcomes
across industries, occupations and income levels. We validate the model
by reproducing key outcomes of the first wave of coronavirus disease
2019 in the New York metropolitan area. The key mechanism coupling
the epidemic and economic modules is the reduction in consumption
due to fear of infection. In counterfactual experiments, we show that a
similar trade-off between epidemic and economic outcomes exists both
when individuals change their behaviour due to fear of infection and when
non-pharmaceutical interventions are imposed. Low-income workers, who
perform in-person occupations in customer-facing industries, face the
strongest trade-off.
From the inception of the coronavirus disease 2019 (COVID-19) pan- economic effects of mandated NPIs were industry and occupation
demic, global efforts have focused mostly on curbing the spread of specific; for instance, lower-income workers who are primarily engaged
severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) through in customer-facing industries and in-person occupations were more
the implementation of mandated non-pharmaceutical interventions at risk of layoffs during industry shutdowns2,3.
(NPIs)1. These strategies, encompassing the partial or complete closure In parallel with NPIs, the COVID-19 pandemic also triggered behav-
of non-essential, customer-facing economic activities such as entertain- ioural adaptations, with individuals voluntarily minimizing their con-
ment and dining, and the enforcement of remote work policies, have tacts and reducing their use of customer-facing services due to fear
impacted people differently across socioeconomic groups. In particu- of the disease. However, the effect of these self-imposed behavioural
lar, employees in non-essential industries or able to work remotely were changes, as opposed to NPIs, remains contentious4–6. Moreover, it is
less likely to be exposed to the virus, while workers engaged in essential, an open question whether these behavioural changes result in uneven
in-person tasks experienced a higher risk of exposure. Similarly, the consequences across socioeconomic groups, similarly to NPIs.
1
CENTAI Institute, Turin, Italy. 2Institute for Biocomputation and Physics of Complex Systems and Department of Theoretical Physics, University of Zaragoza,
Zaragoza, Spain. 3Complexity Science Hub, Vienna, Austria. 4Department of Mathematics and GISC, Universidad Carlos III de Madrid, Leganes, Spain.
5
MOBS Lab, Northeastern University, Boston, MA, USA. 6The Roux Institute, Northeastern University, Portland, ME, USA. 7Institute for New Economic Thinking at
the Oxford Martin School, and Smith School of Enterprise and the Environment, University of Oxford, Oxford, UK. 8Laboratory for Computational Epidemiology
and Public Health, Department of Epidemiology and Biostatistics, Indiana University School of Public Health, Bloomington, IN, USA. 9Connection Science,
Institute for Data Science and Society, MIT, Cambridge, MA, USA. 10Santa Fe Institute, Santa Fe, NM, USA. e-mail: marco.pangallo@centai.eu
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 264
Article https://doi.org/10.1038/s41562-023-01747-x
Addressing the effects of mandated NPIs and behavioural changes, Economic module
both at the aggregate and granular level, requires building theoreti-
cal, mechanistic models that jointly simulate epidemic and economic
dynamics at a fine-grained level. While several models have been pro-
posed7–11, they tend to provide aggregate perspectives on either the
epidemic or economic dimension, and thus fall short in characterizing
heterogeneous outcomes across diverse socioeconomic groups. The
few agent-based models (ABMs) that simulate epidemic spreading and
economic decisions at the level of individual, heterogeneous agents12–14
are primarily designed for qualitative assessments of different policies,
using a basic parameter calibration that considers only a few aggregate
data points. Fear of infection Hiring and firing
In this Article, we introduce an ABM developed to simulate the Workplace
epidemic and economic impacts on a large synthetic population contacts
representative of the New York–Newark–Jersey City, NY–NJ–PA Household
contacts
metro area. The model incorporates detailed socioeconomic attrib-
utes of agents, along with their consumption and contact patterns
Age 40 years
derived from comprehensive census, survey and mobility data. The School
Income $45,000
contacts
structure of the economy is initialized from input–output tables and Community
Employed? Yes
contacts
national and regional accounts. This joint epidemic–economic ABM Industry Manufacturing
largely extends our former epidemic15,16 and economic2,17 COVID-19 Occupation Technician
Work from No
models. home?
Epidemic module
Results Fig. 1 | Schematic representation of the joint epidemic–economic model.
We build a data-driven, granular ABM of the New York–Newark– The economic module depicts the flow of goods and services between industries
Jersey City, NY–NJ–PA metro area. The main agents of the model are and from industries to final consumers (input–output network), while the
the 416,442 individuals of a synthetic population that is representative epidemic module tracks pathogen exposure at workplaces, community/
of the real population across multiple socioeconomic characteristics, consumption venues, schools and households (contact network). Agents display
including household composition, age, income, occupation and pos- high heterogeneity across various socioeconomic characteristics (see box).
sibility to work from home (WFH) (for a schematic representation, see The economic and epidemic modules are closely linked: the epidemic module
Fig. 1, and for a detailed description of the model, see Methods and impacts economic outcomes by reducing consumption due to infection fear,
Supplementary Information). while the economic model influences epidemic spread by altering workplace and
The epidemic module of the ABM is built on the contact network community contacts through employment changes in different industries.
that connects synthetic individuals. This network has multiple layers,
where each layer captures interactions occurring (1) in the household,
(2) in school, (3) in the workplace and (4) in the community (during Modelling the first wave of COVID-19 in New York
on-site consumption, such as in shops, restaurants or movie theatres). We calibrate the model’s key parameters, including the parameter
Epidemic propagation occurs on these networks, built via anonymized, regulating the behavioural changes, dubbed ‘fear of infection’, to fit
privacy-enhanced mobility data from opted-in users, which inform crucial epidemic and economic statistics from the first wave in the NY
workplace and community interactions. These data, collected through metro area (Supplementary Section 4). Epidemiological parameters
a General Data Protection Regulation (GDPR)-compliant framework are adjusted to fit ancestral SARS-CoV-2 lineages (Supplementary
by Cuebiq, provide daily workplace visitation patterns and estimates Table 5). Simulations start on 12 February 2020, protective measures
for colocation probabilities in community spaces, based on a Four- are imposed on 16 March and relaxed on 15 May 15, and simulations
square dataset. The ABM employs a stochastic, discrete-time disease end on 30 June. As protective measures, we close schools, mandate
transmission model on the contact network and synthetic population, WFH and shut down all non-essential economic activities, such as
with individuals transitioning between epidemic states based on key entertainment and most of the accommodation–food industry, but
time-to-event intervals (for example, incubation period, generation also large parts of manufacturing and construction. We use the offi-
time and so on) derived from SARS-CoV-2 transmission data. cial NY regulations to estimate the degree to which a given industry is
From an economic point of view, individuals play a role both as essential (Supplementary Section 3.2.1) and assume that workers who
workers and consumers. They work in one of multiple industries, pro- can WFH are not directly affected by these closures2. We name this set
ducing goods and services that are either sold to other industries as of assumptions the empirical scenario.
intermediate products or sold to final consumers as consumption
products. The economic module specifically emphasizes employment Economic validation
and consumption. In particular, hiring and firing decisions are driven by Our model accurately matches the six official economic statistics we
industry workforce requirements, closures of economic activities and calibrated it on (Fig. 2a). It correctly reproduces the fact that employ-
possibility of remote work. Consumption patterns vary among agents ment declined more strongly than gross domestic product (GDP) (this is
based on age and income, and dynamically adjust in response to the because industries most affected by shutdown orders produce less out-
evolving state of the pandemic. Specifically, households tend to curtail put per worker). It also correctly reproduces the fact that consumption
their demand for services from customer-facing industries because of goods and services produced by customer-facing industries declined
of the fear of the disease (customer-facing industries are entertain- more strongly than consumption of goods and services produced by
ment, accommodation–food, other services, retail, transportation, industries that are not customer facing, ranging from manufacturing
health and education; Supplementary Section 3.1.3). The model also products to utilities and financial services (Supplementary Fig. 15).
considers the input–output network of intermediates that industries Our model can also be validated against empirical properties that
use to produce final goods and services18, leading to the propagation were not directly targeted in the parameter calibration procedure
of COVID-19 shocks to the entire economy. (Fig. 2c,d). First, thanks to our estimate of pandemic shocks2 and shock
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 265
Article https://doi.org/10.1038/s41562-023-01747-x
Economic statistics Epidemic statistics
a b
Model Data
Model Data
Employment (NY)
GDP (NY) March 2020
GDP (US)
April 2020
Investment + net exports (US)
May 2020
Consumption customer facing
June 2020
Consumption not customer facing
0 0.5 1.0 1.5
−20 −15 −10 −5 Deaths per 1,000 individuals
2020Q2 versus 2019Q4 (% change)
c d Low income e
100
High income 1.00
Ratio community contacts
Real estate
Wholesale Finance
Employment
Employment 2020
Health
Transports Information
Retail (model) 0.75
80 Health
Other services Professional
services
Education Employment
0.50
after/before 16 March (model)
Manufacturing (data) Transports
April/February (model)
Administrative
60 Construction Retail
services
Consumption Other services
(model) 0.25
Accommodation−food
Consumption Education Accommodation−Food
40 (data) 0 Entertainment
Entertainment
0 0.25 0.50 0.75 1.00
−20 −10 0
40 60 80 100 Ratio community contacts
2020Q2 versus after/before 16 March (data)
Employment 2020 April/February (data)
2019Q4 (% change)
Fig. 2 | The first wave of COVID-19 in New York: empirical scenario. across the main two-digit NAICS industries, in the model and in the data.
a,b, Statistics that were directly targeted in the parameter calibration. a, The The circle size is proportional to employment in February 2020. d, Employment
percentage change from October–December 2019 (2019Q4) to April–May and consumption, in the model and in the data3, among low-income and
2020 (2020Q2) across six official economic statistics, in the model and in the high-income households (low income <$27,000 and high income >$60,000;
data (Supplementary Section 4). Here, and throughout the paper, we report these bands are chosen for comparison to real data; Supplementary Section
the mean and error bars (2.5–97.5 percentile range) across simulation runs that 5.1.1). e, The ratio between community contacts with infectious individuals
differ by stochastic factors (Methods). b, A comparison between model and data (Supplementary Section 5.2) before and after the imposition of protective
for the number of COVID-19 deaths, which is the key epidemic statistic that we measures, in the model and in the data, for the seven customer-facing industries
targeted. c–e, Validation results for statistics that were not directly targeted. (Supplementary Section 3.1.3). The circle size is proportional to the share of
c, Employment in April 2020 as a percentage of employment in February 2020, pre-pandemic contacts.
propagation model17, we are able to recover industry-specific changes the changes in contact patterns that occurred after protective measures
in employment induced by the pandemic (Fig. 2c), with a Pearson were implemented, although these data were not used for parameter
correlation coefficient of 0.82 (P value 2 × 10−4) between model and calibration (Fig. 2e). Both in the model and in the data, community
data. This accuracy is in large part due to our model’s ability to take contacts substantially reduced (Pearson 0.75 and P value 0.05), more
industry-specific estimates as inputs, but also to the propagation in mostly non-essential industries such as entertainment and restau-
mechanisms embedded in the model. Indeed, the estimates by supply rants than in mostly essential industries such as retail and health. We
shocks alone have only a Pearson correlation of 0.69 (P value 4 × 10−3, also accurately estimate the reduction in workplace contacts across
Supplementary Fig. 14). Second, thanks to our granular and data-driven industries (Supplementary Fig. 20; Pearson 0.88 and P value 5 × 10−6),
characterization of employment and consumption patterns (Supple- the temporal profile of reduction in contacts (Supplementary Fig. 19)
mentary Figs. 8 and 12), we reproduce a key fact: low-income individuals and the increase in prevalence over time (Supplementary Fig. 18).
were more likely to become unemployed but reduced consumption Finally, the model makes a number of estimates about how many infec-
less than high-income individuals3,19 (Fig. 2d). This happens because tions happen across each layer and industry over time, as well as which
low-income individuals are more likely to work in the occupations most occupation, income and age groups are most affected (Supplementary
affected by closures, such as ‘food preparation and serving’, ‘building Figs. 21–23). While we are not able to find data to quantitatively evaluate
and grounds cleaning’ and ‘personal care and service’ (Supplementary these estimates, our literature review provides some support to these
Fig. 16), but they spend a larger share of their income on essential goods findings (Supplementary Section 5.2.1).
and services such as housing and utilities (we do not consider here the
effect of the stimulus programme, which would further increase the Counterfactual scenarios
spending of low-income individuals). In our analysis, we quantitatively explore the effects of three key factors
that shaped the behavioural and policy response to the first COVID-19
Epidemic validation wave. In the following, we use ‘baseline’ to refer to estimated parameters
On the epidemic side, our model correctly matches the death count calibrated with empirical data.
data on which it has been calibrated, correctly replicating the spike in As a first set of counterfactual scenarios, we explore the effect of
the number of reported deaths in April 2020 and the strong reduction adjusting the magnitude of the fear of infection parameter that regu-
in June (Fig. 2d and Supplementary Fig. 18). It also correctly estimates lates behaviour change. Generally, we treat this parameter as uniform
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 266
Article https://doi.org/10.1038/s41562-023-01747-x
a b c
Deaths per
0.06
Total deaths per 1,000 individuals Total deaths per 1,000 individuals
Closures Start measures
5 5 0.04
All open Low fear Late
1,000 individuals
4 Customer facing 4 Baseline 0.02
Low fear
Non-essential Early
0
3 3
20
Baseline fear
20
Baseline fear 2
M 020
ay
2 2 ch Ju 02
r il2
ne 0
ar Ap 2020
M
1 High fear 1 50
Unemployment
Empirical scenario Empirical scenario High fear
40
0 0 30
rate (%)
5 10 15 20 25 5 10 15 20 25 20
Mean unemployment rate (%) Mean unemployment rate (%) 10
0
Ap 02 2
ril 0
2
M 020
ay
ch Ju 20
ne 0 2
ar 2020
M
Fig. 3 | Aggregate results on counterfactuals. a,b, Deaths and unemployment specific closure of economic activities (all non-essential industries, as occurred
across scenarios. For each scenario, we show the aggregate unemployment rate empirically, only customer-facing industries and no closures), keeping the start of
and the cumulative number of deaths, as averaged throughout the simulation protective measures fixed at the baseline, empirically observed date. b, Scenarios
period and the simulation runs (Supplementary Fig. 26 shows the variability are further distinguished by the start of protective measures (baseline: 16 March
across simulation runs and discusses its interpretation). The empirical scenario is 2020, as empirically; early: 17 February 2020 and late: 30 March 2020), keeping
highlighted to serve as a benchmark. Scenarios are distinguished by the strength closures fixed at all non-essential industries. c, For the specific combination of high
of behaviour change, as exemplified by the fear of infection parameter (square: fear of infection and three different starts of protective measures, a time series of
low; circle: baseline; triangle: high). a, Scenarios are further distinguished by the unemployment and deaths corresponding to the three scenarios is shown.
across individuals, per survey evidence20. However, we also consider an Although the total death count and average unemployment can vary
age-specific fear counterfactual. Our baseline calibration yields a fear substantially across simulation runs, the relative impacts of different
of infection parameter distribution (Supplementary Fig. 13), implying policies remain robust (Supplementary Fig. 26).
a 14% consumption demand reduction in customer-facing industries Both higher fear of infection and stricter closures lead to saving
due to infection fear at the epidemic peak. This calibrated value, which lives at the expense of jobs, for low and high income workers alike
merges NPI effects with behaviour change, cannot be causally inter- (Fig. 4a). However, for low income workers, higher fear of infection
preted. Absence of NPIs would necessitate a steeper consumption drop or stricter closures have a larger effect, leading to more lives saved
for the model to explain observed behavioural changes, thus estimating and more jobs lost, compared with high-income workers. As we will
higher infection fear. But the lack of real-world data without NPIs hampers show later, outside the household setting, most infections occur in
this estimate. Instead, we explore two counterfactuals: ‘low’ (0.1 times customer-facing industries, where most low-income workers are con-
baseline) and ‘high’ (10 times baseline), translating to a 1% and 77% con- centrated. Thus, mandated closure or spontaneous avoidance of these
sumption reduction due to fear at the peak, respectively. These scenarios industries leads to both more unemployment and fewer workplace
facilitate comparing stronger fear of infection effects with stricter NPIs. infections among low-income workers.
Next, we vary two policy-related factors. First, we experiment with The unequal economic outcomes of the empirical scenario also
different economic activity closures. Besides the baseline scenario with lead to geographical disparities. Figure 4b shows two maps of unem-
all non-essential industries closed, we consider two milder closure sce- ployment in Manhattan in the empirical scenario (asterisk) and in a
narios: (1) only non-essential customer-facing industries are closed and counterfactual with low fear and no closures (hash). We see that in
(2) no closures, with all economic activities open. Second, we simulate the counterfactual, the unemployment rate is very evenly spatially
protective measures starting either 4 weeks earlier (17 February 2020) distributed, while in the empirical scenario, low-income areas such
or 2 weeks later (30 March 2020). Additional counterfactuals, including as the Queens and the Bronx have a high unemployment rate of more
partial closure of customer-facing industries and no WFH or school than 20%, compared with high-income areas such as Manhattan, with
closures, are explored in Supplementary Information (Supplementary unemployment rates around 15%.
Figs. 24 and 25). Overall, these results contribute to the ongoing debate on the rela-
tive effectiveness of behavioural change versus NPIs in preserving both
Behaviour change versus NPIs public health and the economy. While it is intuitive to expect stricter
Aggregate economic and epidemic results are shown in Fig. 3, while mandated NPIs to increase unemployment and decrease COVID-19
results disaggregated by income, geography and industry are shown deaths, it is less apparent that heightened behavioural adaptation
in Fig. 4 (Supplementary Figs. 27–33). Figure 3a conveys our first main would yield similar results (Discussion). Our findings highlight a quali-
result: stricter closure of economic activities and higher fear of infec- tative parallel between substantial behavioural change and stringent
tion both lead to increased unemployment and fewer COVID-19 deaths. economic activity closures. Spontaneous avoidance of services offered
To illustrate this, consider a scenario with baseline fear of infection and by customer-facing industries, akin to their mandated closure, results
all economic activities open, represented by the light-coloured circle. in increased unemployment but fewer fatalities. This trend is particu-
If we maintain the fear of infection at the baseline level, but close all larly pronounced among low-income individuals.
non-essential economic activities (as in the baseline scenario), unem-
ployment surges by 64%, while the number of deaths drops by 35%. Industry-specific closures
Likewise, if we instead keep the closure level at the empirical baseline Our model also evaluates the efficacy of closing all non-essential
but increase fear of infection (represented by the dark triangle), we see economic activities, including large segments of manufacturing and
a 40% rise in unemployment and a 50% decrease in deaths relative to construction, compared with exclusively closing customer-facing
the empirical scenario. Similar trends are observed in other scenarios. industries. We find that the mandated closure of all non-essential
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 267
Article https://doi.org/10.1038/s41562-023-01747-x
a b c 30
(*)
Unemployment rate (%)
25
Workplace infections Infections per
Empirical
160
20 scenario 20
140 15
1,000 individuals
(*) Accommodation
per 1,000 individuals
–food
120 10 Entertainment
10
(*) Professional
5 services
100 Manufacturing
All Customer Non- All Customer Non-
Construction
open facing essential open facing essential
Closures Closures All open Customer Non-essential
(#)
facing
Consumption reduction compared
Unemployment rate (%)
Workplace infections
25
30
160 (#) 20
140 15 20
per 1,000 individuals
with pre-pandemic level (%)
120 10
5 (#) 10
100
Low Baseline High Low Baseline High
Fear of infection Fear of infection Unemployment 0
Income class Low income High income rate (%)<5 10 14 19 23 >28
All open Customer Non-essential
facing
Closures
Fig. 4 | Results on counterfactuals disaggregated by income, geography and baseline date. b, Maps of unemployment across census tracts in New York City,
industry. a, Workplace infections and unemployment across income classes. corresponding to two scenarios in a, including the empirical scenario (asterisk)
Top: we vary the level of closures, keeping fear of infection and the start of and the counterfactual with no closures and low fear (hash). c, Infections and
protective measures to their baseline values (so the case with all non-essential reduction in consumption across five selected industries and three levels of
activities closed is the empirical scenario). Bottom: we vary fear of infection closures, for baseline fear and start of protective measures.
keeping all economic activities open and starting protective measures on the
activities only marginally decreases deaths compared with solely clos- too. Our preliminary investigation (Supplementary Section 6.4) also
ing customer-facing industries, but it drastically increases unemploy- shows that with an early start of protective measures, it is possible to
ment. In comparison with the baseline scenario, a counterfactual that avoid an excessive burden on the healthcare system, as measured by a
only closes customer-facing industries results in a slightly higher death usage of more than 50% of the nominal capacity of intensive care units.
rate (4% higher), but substantially mitigates unemployment, reducing
it by 36%. To explain these results, consider Fig. 4c. Most infections Age-specific fear of infection
occur in customer-facing sectors such as ‘entertainment’ and ‘accom- In the empirical scenario, the parameterization of the ‘fear of infec-
modation–food’. Their closure curbs infections considerably but also tion’ is uniformly applied across individuals, as suggested by survey
consumption and employment. Conversely, closing ‘manufacturing’ evidence20. However, we also examined a counterfactual in which young
or ‘construction’ marginally impacts infections but drastically reduces individuals adopt less behavioural changes (low fear of the disease) than
consumption. ‘Professional services’ remains largely unaffected also older individuals (high fear of the disease), considering this might be
because of WFH adaptability. a more optimal situation for pandemic control through behavioural
Methodologically, these industry-specific results were obtained change. Here, at-risk older individuals would internalize the infection
because we associated each consumption venue from mobility data to risk more, while younger individuals, less likely to suffer severe conse-
an economic activity, allowing the quantification of industry-specific quences, could maintain higher consumption and contribute to herd
contacts. This granular, data-driven approach provides insights that immunity. We explore how this scenario plays out quantitatively in the
more aggregate, qualitative models might overlook. data-driven, granular agent-based model.
To explore the effects of heterogeneous behavioural changes,
Timing of interventions we group all households into three classes based on the age of their
Another counterfactual exploration concerns the effectiveness of household head (0–34, 35–64 and 65+ years). We assume that fear of
starting epidemic mitigation and control earlier (4 weeks before) infection in each class is proportional to the risk of death in that class
or later (2 weeks later) than in the empirical scenario. As we show in (Supplementary Section 6.5). We also normalize the fear of infection
Fig. 3b, delaying these measures marginally reduces unemployment by parameter across age bracket so that the mean takes the same baseline
2% but causes a notable 50% rise in deaths. In a high fear-of-infection value that we considered in the empirical scenario. This normaliza-
scenario, late measures result in both a 46% increase in deaths and a 12% tion ensures that results are driven by a different distribution of fear
rise in unemployment. The mechanism for these results is suggested across age groups, rather than by changes in overall fear. At the end
in Fig. 3c, where we show time series across the three counterfactu- of this procedure, the fear level among households aged 0–34 years
als with high fear of infection. We see that an early start of mitigation is 0.02 times the baseline, households aged 35–64 years have a fear
measures prevents an epidemic wave, leading to no further increase in level of 0.48 times the baseline and households aged 65+ years have a
unemployment due to fear of infection. Conversely, with a baseline or considerably higher fear level of 4.91 times the baseline. We compare
late start, substantial behaviour change leads to reduced consumption, the age-specific scenario with the scenario in which all agents have
and this, in turn, leads industries to fire their employees, increasing uniform baseline fear (as in the empirical scenario).
unemployment. Thus, starting mitigation measures early is crucial We report aggregate results in Fig. 5a, which reproduces the
to improve epidemic outcomes, and possibly economic outcomes same scenarios as Fig. 3 for uniform fear, next to the new results for
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 268
Article https://doi.org/10.1038/s41562-023-01747-x
a b
3.00 Age 0–34 years 65+ years Fear of infection Age specific Uniform
Closures
Consumer demand reduced
All open
Total deaths per 1,000 individuals
1.00 1.00
Contacts reduced
Customer
2.75 facing
0.75 0.75
Non−essential
0.50 0.50
due to fear of infection
due to fear of infection
2.50
0.25 0.25
0 0
2.25 Uniform fear 20 02 20 20 20 20 02 20 20 20
20 0 20 20 20 20 0 20 20 20
ril 2 ril 2
M May Ju
e
Ju
y M May Ju
e
Ju
y
arch Ap n l arch Ap n l
2.00 Age-specific fear c
Infections per 1,000 individuals Infections per 1,000 individuals
Outside households 1.00 Customer facing
Actual consumption demand
8
8 10 12 14
Mean unemployment rate (%) 6 0.95
4 4
0.90
Start measures
2 Accommodation
Late 0.85 –food
Total deaths per 1,000 individuals
Health
3 Baseline 0
Early Not customer facing
Within households
Actual consumption demand
8 1.06
Uniform fear
2
6
1.04
4
Finance
Age-specific fear 1.02
1 Real estate
2
0 1.00
13 14 15 16 20 20 20 20 20 20 20 20 20 20
20 20 20 20 20 20 20 20 20 20
Mean unemployment rate (%) ch ril ay ne
Ju
ly ch ril ay ne
Ju
ly
ar Ap M Ju ar Ap M Ju
M M
Fig. 5 | Results on age-specific fear of infection. a, Aggregate deaths and These time series show how fear of infection reduces contacts and consumption
unemployment across scenarios. The general interpretation is the same as Fig. 3; demand. c, For the same scenario as b, the time series of infections per 1,000
here, uniform fear is represented by circles and age-specific fear is represented individuals disaggregated by age groups and by whether they occurred
by plus symbols. b, For the scenario ‘all open–early start’, time series of the level outside or within households, and actual consumption demand (relative to the
of workplace and community contacts and consumption demand of customer- pre-pandemic level) disaggregated by industry and by whether industries are
facing industries, disaggregated by type of fear (solid lines, age-specific fear customer facing or not.
and dashed lines, uniform fear) and by age groups with heterogeneous fear.
age-specific fear. Adjusting for age-specific fear, while keeping other comparable in both scenarios. In contrast, we see an opposite trend
factors constant, marginally reduces both unemployment and deaths among young households, where a large number of infections hap-
compared with uniform fear. In the ‘all open’ scenario, where the effect pen later due to their very low fear of infection. Within households,
is most pronounced, age-specific fear reduces deaths by 6% and unem- the differences between age-specific and uniform scenarios are less
ployment by 5%. For comparison, closing customer-facing industries pronounced.
cuts deaths by 28% but increases unemployment by 22%. The relatively small decrease in deaths with age-specific fear can
In Fig. 5b,c, we examine industry- and age-specific effects, focusing be explained in two ways. On the one hand, the time series of reduc-
on ages 0–34 years and 65+ years. First, in Fig. 5b, we show how fear of tions in contacts and infections show that older individuals drastically
infection reduces contacts and consumption demand. As expected, reduce their contacts only after the epidemic peak. This delay results
uniform fear leads to equal reductions across all ages by construction, from the lag between infection and death reporting; behaviour change
while age-specific fear instead leads to the least reduction in young intensifies when individuals become aware of the number of COVID-19
agents and the most in older agents. Total consumption decreases less deaths. On the other hand, older individuals cannot avoid infections
than contacts as it may not require direct contact, such as ordering within their own households.
takeaway food (Methods). As we can also see in Fig. 5c (right plots), age-dependent fear of
In Fig. 5c, we first consider infections (left plots). We distinguish infection alters consumption demand across industries. Consump-
infections occurring inside households from those happening outside tion demand decreases more in health, a customer-facing industry on
(community or workplace contacts). Outside the household, infections which old agents spend a disproportionate amount of income, and less
among older agents decrease with age-specific fear compared with in accommodation–food, on which young agents spend a higher share
uniform fear, especially around the epidemic peak where they are 30% of their income. At the same time, consumption demand increases
lower. However, in the waning phase of the epidemic, infections are towards industries that are not customer facing because households
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 269
Article https://doi.org/10.1038/s41562-023-01747-x
reallocate part of their consumption budget to those industries. As our key results to hold, and we view our model as mostly applicable
individuals in older age groups decrease consumption more and thus to the short-term management of emerging/re-emerging infectious
have more budget to reallocate because of higher fear of infection, this diseases. Another important limitation is that the matching between
results in higher consumption demand towards industries that have synthetic individuals and mobility traces is probabilistic, as we do not
high consumption share among old individuals, such as finance. By have socioeconomic information about specific Cuebiq users. None-
contrast, because younger individuals spend a large fraction of their theless, our privacy-preserving matching algorithm based on census
income on real estate and, with low fear, they do not reallocate much, tracts is likely to be accurate given the strong socioeconomic dispari-
the increase in demand for real estate services is lower than if fear of ties in different parts of the New York metro area. From the epidemio-
infection was uniform across ages. logical standpoint, we assume the same per-contact risk of infection
In summary, these findings show that even when individuals in different occupational settings. If empirical epidemiological data
adjust their behaviours in response to their personal risk levels during were collected about the contribution of these settings to SARS-CoV-2
a pandemic, it only modestly affects health and economic outcomes. transmission, our estimates could be further refined. Moreover, we did
Moreover, our results quantify the complex ripple effects across vari- not consider differential risk of severe disease and death for individu-
ous sociodemographic groups. als with different socioeconomic status. The inclusion of this factor
into the model could further exacerbate the highlighted heterogene-
Discussion ity in the health and economic impact of the pandemic and adopted
Addressing the health impacts of the COVID-19 pandemic required policies on different segments of the population. From the economic
important societal and economic disruptions, sparking intense debates. standpoint, we consider industries located over the entire metro area,
On the one hand, stringent restrictions and government-enforced rather than heterogeneous firms at specific geographical locations.
measures were critical to suppress the virus spread. On the other hand, Although this is a limitation of our analysis, we believe that this is the
some contend that individual behavioural adaptations could have right level of aggregation for the questions considered here, but we
served as a more effective tool in managing the epidemic’s trajectory. acknowledge that a more detailed representation of the production
They suggest that allowing individuals to spontaneously lower their sector may be needed to address questions such as the effectiveness
exposure risk according to the epidemic trajectory would lead to the of spatially targeted lockdowns. Finally, the infection transmission
most favourable balance of health and economic outcomes. and economic models are combined through a ‘fear’ mechanism that
Determining whether behavioural change or NPIs are more effec- was modelled in a simple manner (that is, as a function of the number
tive in minimizing the pandemic’s health and economic impacts is of reported deaths on the previous day). For example, individuals may
complex. Each operates differently; NPIs function by curtailing labour not retrieve information on a daily basis and media may amplify some
supply, creating a supply shock, while behavioural changes act as a information (for example, a spike in the number of deaths) at certain
demand shock with time-varying effects. Additionally, behavioural times, thus altering the perception of the population21; different seg-
changes typically occur only when reported deaths rise considerably, ments of the population may have a different risk perception20. This
which usually lags about 3 weeks behind infection transmission. The highlights the importance of conducting future studies to better char-
effectiveness of behavioural changes versus NPIs depends on quanti- acterize the relation between risk perception and human behaviour
tative details such as how long they take to reduce virus circulation to during epidemic outbreaks.
very low levels, enabling a prompt rebound in consumption. From a policy standpoint, we focus on strategies actually imple-
Our findings indicate a parallel between behavioural responses mented in the real world. We also performed preliminary explorations
and economic activity shutdowns: both substantial behavioural of more sophisticated policy options, including the activation of pro-
changes and stringent closures lead to similar patterns of rising unem- tective measures when infections go past a certain threshold, and the
ployment and fewer infections. This impact is particularly heightened deactivation when they go below another threshold (Supplementary
among low-income workers compared with high-income workers. Section 6.6). Exploring these strategies, we find interesting results,
Furthermore, this trend persists even when older individuals dem- such as the possibility of quasi-steady states with intermittent closures,
onstrate stronger behavioural responses than younger individuals. but these results do not change our main conclusions. While our model
Indeed, even if individuals change their behaviour proportionally to can support policymakers in exploring these scenarios, assessing their
their own age-specific death risk, it only slightly enhances epidemic practical feasibility is essential and requires case-by-case examination,
and economic outcomes compared with a situation where behav- based on resources, logistics and the objective functions to optimize,
ioural change is uniformly distributed across the population. Our which is beyond this paper’s scope. Our findings, however, under-
results also show that the trade-off between health and the economy score the importance of some targeted policies. For instance, clos-
strongly depends on which economic activities are closed. The closure ing customer-facing industries, particularly if done early, effectively
of non-customer-facing industries, such as manufacturing and con- reduces viral transmission. This allows income-support schemes to
struction, results in a substantial spike in unemployment with only a specifically aid certain occupations such as food preparation, serving
marginal decrease in fatalities. Additionally, implementing protective or personal care services, rather than a broader worker base such as
measures late in high fear-of-infection scenarios leads to a dual blow those in construction or manufacturing. Enhanced surveillance and
of increased deaths and unemployment. In other cases, a delayed start contact tracing in industries employing these low-income workers
of protective measures substantially escalates fatalities while only could yield both health and economic benefits. Importantly, these
slightly reducing unemployment. These results underscore a crucial policies are crucial not only during government-mandated closures
distinction between behavioural changes and NPIs: while behavioural but also when spontaneous behaviour changes reduce consumption
changes are a result of self-organization, NPIs can be implemented as in these industries. Our findings can thus guide the development of
soon as needed for highest effectiveness. policies to mitigate the health and economic impacts of pandemics,
Our results have the usual limitations pertaining to modelling while also safeguarding low-income populations to reduce inequalities.
studies (for more details on the model, see Methods and Fig. 6). In The model presented in this paper has potential impact on both
this paper, we exclusively focus on the first wave of COVID-19 in one epidemiological and economic impact analysis. From an epidemiologi-
specific metropolitan area. It could also be important to consider cal perspective, we have incorporated industries, occupations and the
other aspects of the COVID-19 pandemic that became relevant after feasibility of remote work into a granular transmission model, dem-
the first wave of infections such as masks, test, trace and quarantining, onstrating the value of integrating economic, social and behavioural
variants, vaccination and waning of immunity. However, we expect dimensions into epidemic spreading models22,23. Economic impact
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 270
Article https://doi.org/10.1038/s41562-023-01747-x
a c
Inputs Model Outputs
Locations ε–γ
Foursquare
Mobility Epidemic Infections p IA R
µ
Cuebiq module recoveries
deaths S L
Synthetic
population 1–p PS IS R
Fear infection Hiring/firing γ µ
ACS, BLS
rβ β βS
Closures δ
Economic Employment ε D
NY
module consumption
government IA IS PS
production
BLS, BEA
Input–output
BEA
b d Supply shockst Epidemicst–1
ID Age Income Industry Occupation WFH Employed
a 54 $30k Manufacturing Production No Yes Employmentt–1
b 39 $100k Manufacturing Management Yes Yes Employmentt Outputt
c 39 $50k Retail trade Sales-related No No Demandt–1
d 75 $20k / Retired / /
Consumption
e 73 $40k / Retired / / Employmentt–1 Employmentt
demandt
f 9 / / / / /
… … … … … … … Outputt–1 Other demandt Total demandt
Fig. 6 | A schematic of the methodology. a, Inputs and outputs of the epidemic symbols. d, A causal diagram of the economic module. Blue rectangles show
and economic modules. For each group of input data, we list the data source. The processes exogenous to the economic module, black rectangles represent
feedback from the epidemic to the economic module is through fear of infection, industry-level endogenous variables and red diamonds are individual-level
while a reverse feedback occurs through hiring and firing. b, Socioeconomic endogenous variables (in this case, the employment status of each individual).
characteristics of six example synthetic individuals, including age, income, A link between variables indicates that the linking variable influences the linked
industry, occupation, possibility to WFH and employment status. c, A schematic variable. We only show the key variables. On the left, we show lagged values of
of the epidemic module. The different compartments and transition times from these variables at time t − 1, while on the right we show the relations between
one compartment to another are shown. See the text for the meaning of the these variables at time t.
studies of disasters frequently limit their focus to industries24,25, or Section 4.1), and so it is a valid approximation to only consider the NY
employ an aggregate representative household17. The detailed syn- MSA. The economic module models in detail the NY MSA area and also
thetic population used in our model, mirroring the real-world popu- features a simplified model of the rest of the United States. The NY MSA
lation, enables us to explore a new array of questions centred around economy is deeply integrated with that of the United States, and this
understanding the disparate impacts across socioeconomic groups. In integration must be taken into account throughout the pandemic to
doing so, this paper has potential impact in future research aimed at for- properly estimate economic impacts.
mulating more targeted, effective public health policies and strategies.
Agents
Methods The main agents of the epidemic–economic ABM are the 416,442 indi-
This section provides an overview of the economic and epidemic viduals of a synthetic population that is representative of the NY MSA
modules, how they are coupled, which data we use to initialize them (Supplementary Section 3.3.5). The population size is determined by
and what outputs they produce (Fig. 6a). A longer description that the availability of mobility data (see below). The agents are hetero-
gives all details and justifies all the assumptions can be found in the geneous in several of their socioeconomic characteristics (Fig. 6b),
Supplementary Information. including age, income, employment status, occupation, possibility
to WFH and the census tract where they live. Individuals are grouped
Geography into 153,547 households, whose composition is consistent with cen-
The epidemic–economic model focuses on the New York–Newark– sus microdata. We derive socioeconomic characteristics of synthetic
Jersey City, NY–NJ–PA metropolitan statistical area (federal informa- individuals from tables provided by the American Community Survey
tion processing standard code C3562) (NY MSA). The NY MSA includes (ACS) and the Bureau of Labor Statistics (BLS), trying to get as many
the highly urbanized area of New York City (that is, the five boroughs joint distributions of variables as we can. For instance, to study hetero-
of Manhattan, Bronx, Queens, Brooklyn and Staten Island), but also geneous outcomes across socioeconomic groups, it is important that
some rural and industrial areas. The total population, as of 2019, is synthetic individuals have the correct joint distribution of income,
19,216,182 individuals, making up about 6% of the US population. Total occupation and industry; this is important to replicate the empirical
GDP, as of 2019, is around US$1.5 trillion, a bit more than 7% of the fact that managers working in the finance industry earn high income,
US GDP. A map of the NY MSA is shown in Supplementary Fig. 5. The while food preparation workers in the restaurants’ industry earn low
epidemic module exclusively considers the NY MSA. This is because, income. To achieve this objective, we combine United States-level BLS
while case importation from other parts of the United States and from tables reporting incomes by industry–occupation pair with spatially
abroad is an important factor in the early phase of a pandemic, it has detailed ACS tables giving incomes down to the census tract level. More
little impact once the local incidence grows26. Our model is initialized details on the synthetic population building algorithm and validation
in a period when local incidence was already relevant (Supplementary tests can be found in Supplementary Section 3.3.
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 271
Article https://doi.org/10.1038/s41562-023-01747-x
In the economic module, we also treat industries as agents, con- the incubation period, to account for pre-symptomatic transmission.
sidering a single representative firm per industry. We use the two-digit Whether an individual becomes symptomatic or not depends on the
North American Industry Classification System (NAICS) level of aggre- age-specific symptomatic probability, p. Lastly, the individual will be
gation, giving 20 different industries. Industries are mainly dependent removed (R) from the infectious pool according to an exponential pro-
on one another through the input–output network of consumption cess with rate μ−1, where μ is the average length of the infectious period
of intermediate goods. Since no official data for the NY MSA exist, in days. Note that the removed compartment does not imply recovery,
we downloaded national data from the Bureau of Economic Analysis only that the individual is no longer able to infect. After δ days, removed
(BEA) and then used a regionalization method known as Flegg Location individuals might transition to the death compartment according to the
Quotient27 to obtain an input–output table that distinguishes between empirical age-dependent infection fatality ratio. A new death is reported
the NY MSA and the rest of the United States. The main idea behind this Tn days after the actual event to account for notification delays.
method is that a region that is more specialized in some good or service
(such as information or finance in NY) can just rely on itself to source Economic module
that good or service, but if it is less specialized (as for manufacturing We introduce a dynamic macro-economic model that is specifically
in NY) it is more likely to import that good or service from the other(s) suited to study the economic effects of COVID-19, both at the macro-level
region. More details on our reconstruction of the input–output table of industries and at the micro-level of individuals and households.
can be found in Supplementary Section 3.4. Figure 6d shows the causal relations between the variables of the eco-
nomic module. It distinguishes between variables that are exogenous to
Epidemic module the economic module (blue rectangles), such as the epidemic trajectory
The epidemic module is a standard epidemiological model that runs on or the shelter-in-place policies that lead to supply shocks, and endog-
top of a contact network extracted from Global Positioning System (GPS) enous variables. These are further distinguished into industry-level
location data. Except for the integration with the economic module, endogenous variables (black rectangles) and individual-level endog-
the basic model is the same as the one described in ref. 16. In this model, enous variables (red diamonds). Examples of industry-level variables
individuals interact through a contact network composed of four layers: include employment, output, consumption and total demand. The
(1) the community layer captures occasional interactions between indi- only individual-level variable that we consider is employment status,
viduals, for instance occurring in consumption venues; (2) the workplace although we do consider other agent attributes (such as age) that are
layer captures interactions between workers; (3) the household layer fixed within our simulation period. At every time step t:
captures interactions between household members and (4) the school 1. Industries decide the size of the workforce they need based on
layer captures interactions between children attending the same school. their past employment, past demand and the current levels of
To initialize contacts in the community and workplace layers to the restrictions (we do not consider labour shortages due to illness
pre-pandemic situation, we use privacy-preserving location intelligence or quarantine, as they would be difficult to model and prelimi-
data provided by Cuebiq, merging information about visits to points of nary simulations showed that they were a second-order effect).
Interest with a large database by Foursquare that characterizes points Conditional on the restrictions and on the previous employ-
of Interest. We devised a privacy-preserving algorithm to match Cuebiq ment status of individuals, industries decide which specific
users to synthetic individuals, mainly based on the census tract where workers they hire or fire uniformly at random
they live (Supplementary Section 3.3.6). Our approach to reconstruct- 2. Individuals, grouped into households, decide their consump-
ing contacts is probabilistic: because we cannot observe colocation tion demand based both on fixed attributes such as age or
of individuals reliably in the data, we use mobility data to estimate the income and on variable outcomes such as the situation of the
probability that any pair of individuals are in contact on a given day and epidemic, and on their employment status (workers who lose
in a certain venue (Supplementary Section 1). The contact networks are their job may cut back on spending). Aggregating over agents
initialized using data on pre-pandemic mobility, and modified over time produces a total consumption demand for each industry
due to exogenous interventions, feedback from the economic module 3. Total final demand for each industry is obtained by summing up
and fear of infection (as explained in the sections below). consumption demand, orders of intermediate goods from oth-
We use a stochastic, discrete-time infection transmission model er industries and other components of final demand (including
coupled to the contact network (Fig. 6c) that extends the classical sus- government expenditures, investments, imports and exports)
ceptible, latent, infected, removed model. A susceptible individual (S) 4. Finally, industries produce goods and services. Industries aim
may become infected upon contacting an infectious individual, mov- to produce as much as demanded, but their production can be
ing to the latent compartment (L). Three states describe individuals limited by labour shortages (for simplicity, in this paper we do
who are potentially infectious, each of them with their corresponding not consider intermediate inputs shortages, which were not a
transmission rate: pre-symptomatic (PS), with transmission rate βS; first-order effect in the first few months of the pandemic17). In
infectious symptomatic, with rate β; and infectious asymptomatic (IA), the case of shortages, demand is rationed on a pro-rata basis
with rate rβ. Contacts between infectious and susceptible individuals across intermediate and final consumers
depend on the contact network estimated for each day. Therefore, the
probability that a susceptible node i gets infected by an infectious node Following the literature on dynamic input–output models that
j in infectious compartment type and place p is: estimate the economic impact of natural disasters24,25, we assume
constant prices. We think that this is a good assumption at least in the
P(Si + Ij → Li + Ij ) = 1 − e−βtype wi,j,p (t)Δt (1) short-term management of a pandemic. For instance, empirically, in the
United States, prices decreased by 1.3% on an annual rate in Q2 2020,
where Δt = 1 day and wi,j,p is a weight that modulates the effectiveness which was the smallest price change in the last three years (National
of contacts in a given setting in terms of spreading. We assume that all Income and Products Account table 1.1.7, ref. 28, accessed 23 April
locations within the same layer (schools, workplaces, households and 2023). Thus, for simplicity, we do not consider prices in our model.
community) have the same weight, except for indoor/outdoor spaces
in the community layer (Supplementary Section 4.1). Coupling the epidemic and economic modules
Once infected, the individual will enter the incubation compart- The epidemic and economic modules are strongly coupled, in the sense
ment (L) for ϵ days, during which they will be infected but not infectious that at time t, both modules take as input the output that the other mod-
yet. A latent individual will become infectious γ days before the end of ule generated at t − 1. More specifically, the economic module takes the
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 272
Article https://doi.org/10.1038/s41562-023-01747-x
number of deaths reported in the epidemic module, Dt−1, as an input to the fear of the disease, several individuals reduced their in-person
compute reduction in consumption demand, as explained in the ‘Fear consumption and contacts in the community and workplace as the
of infection’ section below. At the same time, the epidemic module takes epidemic situation worsened. We follow the large literature in behav-
the employment status of each individual in the synthetic population ioural epidemiology30–35, and in particular the approach introduced in
as input from the economic module. This information is used in the ref. 36, letting behaviour change follow the functional form
epidemic module as previously employed individuals who get fired can
Λt (ϕ, Dt−1 ) = 1 − exp (−ϕDt−1 ) , (2)
no longer get infected in the workplace, while previously unemployed
individuals who are hired can get infected. From a technical point of
view, the epidemic module is written in C, while the economic module where Dt−1 is the number of daily reported deaths in the NY MSA on day
is written in Python. To implement coupling between the two models, t − 1, and ϕ is a sensitivity parameter that we name fear of infection.
we use a Python–C application programming interface (API)29 that When Dt−1 = 0, Λt = 0, so there is no behaviour change. In contrast, when
makes it possible to initialize a Python interpreter from within a C run. Dt−1 grows large, there is substantial behaviour change, as Λt → 1. Since
Dt−1 ≥ 0, ∀ t, behaviour change increases with fear of infection. The expo-
Timeline nential functional form in equation (2) corresponds to a non-linearity
A time step in the epidemic–economic model corresponds to one in behavioural response that gives larger weight to an initial increase in
calendar day. Time effectively begins on 12 February 2020. On this Dt−1, with saturation afterwards36, in line with the behavioural response
date, the epidemic module starts running in calendar time (Supple- to the first wave (for a comparison between our approach and other
mentary Section 4.1) and producing epidemic outcomes on a daily modelling efforts for fear of infection during the COVID-19 pandemic,
scale. The economic module starts in a steady state that represents the see Supplementary Section 2.4). For parsimony, we assume that fear
economic situation at the beginning of 2020. All our simulations finish of infection is constant over time. This implies that the dynamics of
at the end of June 2020, giving a total of 140 time steps. In the empiri- the behaviour change are driven by the evolution of the death rate.
cal scenario, where we use our model to reproduce what happened While behaviour change in both the epidemic and economic
during the first wave of COVID-19 in the NY MSA, we impose a number modules is based on the functional form in equation (2), there are
of exogenous interventions (see below) to both the epidemic and eco- slight differences in how behaviour change affects consumption and
nomic modules on 16 March 2020. We choose that date for simplicity, workplace and community contacts.
as several interventions were imposed at different times by the states of • Reduction in consumption (effect in the economic module).
New York and New Jersey from 9 to 23 March, and we see an abrupt During the COVID-19 pandemic, individuals reduced ‘risky’
change in the social dynamics in our data around that date16. We remove consumption of customer-facing services such as restaurants,
the economic interventions on 15 May 2020, again as an approximation cinemas, hairdressers and so on. However, they did not reduce
to the actual relaxation of protective measures that took place in the consumption of financial and real estate services due to fear of
NY MSA during spring 2020. Other measures, such as the closure of infections. Most people kept paying rent, and even increased
schools and working from home, are kept in place until the end of the consumption of some manufacturing goods such as houseware.
simulation. Similarly, for the modelling of counterfactuals, we start Therefore, we assume that fear of infection only decreased
protective measures at various times, ranging from 17 February to 30 consumption in customer-facing industries, using the following
March, but still relax them on 15 May 2020. functional form
Exogenous interventions ΛECO
t,k
= Λ (ϕECO , Dt−1 ) τk , (3)
During the COVID-19 pandemic, governments and local authorities
imposed a number of NPIs with the goal of reducing the epidemic where τk is an indicator that takes value τk = 1 if industry k is customer
spreading. In this paper, we consider three types of interventions: facing, and τk = 0 if it is not. Supplementary Section 3.1.3 lists which
1. Closure of economic activities. We assume that a certain frac- industries are customer facing, and explains our classification. The
tion sk,t of industry k at time t can be exogenously shut down. parameter ϕECO is a fear of infection parameter specific to the economic
This implies that a fraction sk,t of the in-person workers of indus- module, as we detail below
try k cannot work at t, reducing economic output of industry k. • Reduction in community contacts (effect in the epidemic
In the epidemic module, we assume the same reduction sk,t of module). As individuals reduce consumption of customer-facing
contacts in the industry k. In the empirical scenario, we repro- services, they also decrease their contacts in the community,
duce the closures that occurred in the NY MSA in spring 2020 which has an effect on the epidemic module. However, the
(ref. 2), and we name this set of closures ‘non-essential’. We reduction in consumption is not identical to the reduction in
consider alternative closing strategies when studying counter- community contacts. For instance, individuals may order takea-
factuals, such as the closure of customer-facing industries only way meals from restaurants, thus reducing contacts but not
2. Imposition of WFH. All workers who can WFH must do so. We consumption. We assume that reductions in consumption and
assume that this has no impact on the economic module—a community contacts are proportional, letting fear of infection
worker who can WFH is as productive at home as in the work- in the epidemic module be given by
place—but it reduces contacts and so infections in the work- ϕEPI = ϕECO /ϕ,̃ (4)
place in the epidemic module
3. School closures. All contacts between children going to school • with ϕ̃ as a parameter giving the proportion between the two
are removed, effectively cutting off schools from disease trans- fear of infection parameters (for how we calibrate these
mission and thus reducing overall infections. We do not consider parameters, see Supplementary Section 4). The reduction in
the impact of this policy on economic outcomes, due to the community contacts is then given by
difficulty of calibrating the productivity loss related to childcare
ΛEPI
t,k
= Λ (ϕEPI , Dt−1 ) τk . (5)
Fear of infection • In practice, we implement this reduction in community contacts
Further to government interventions, a distinctive hallmark of the by multiplying the weights wi,j,p in equation (1) by 1 − ΛEPI
t,k
, which
COVID-19 pandemic has been individuals’ behaviour change: Due to is the new level of contacts as modified by fear of infection
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 273
Article https://doi.org/10.1038/s41562-023-01747-x
• Reduction in workplace contacts (effect in the epidemic module). 5. Farboodi, M., Jarosch, G. & Shimer, R. Internal and external
If WFH is not imposed by the government or local authority, indi- effects of social distancing in a pandemic. J. Econ. Theory 196,
viduals may nevertheless decide to WFH if they can. We assume 105293 (2021).
that this is uniform across industries, so that the reduction in 6. Krueger, D., Uhlig, H. & Xie, T. Macroeconomic dynamics and
workplace contacts among individuals who can WFH is given by reallocation in an epidemic: evaluating the ‘Swedish solution’.
Econ. Policy 37, 341-398 (2022).
ΛEPI,work
t,k
= Λ (ϕEPI , Dt−1 ) . (6) 7. Ash, T., Bento, A. M., Kaffine, D., Rao, A. & Bento, A. I. Disease-
economy trade-offs under alternative epidemic control
• As above, we implement this reduction in workplace contacts by strategies. Nat. Commun. https://doi.org/10.1038/s41467-022-
multiplying the weights by 1 − ΛEPI,work
t,k
. We assume no reduction 30642-8 (2022).
in productivity from working from home—a worker who can 8. Eichenbaum, M. S., Rebelo, S. & Trabandt, M. The macroeconomics
WFH is as productive at home as in the workplace—so this has no of epidemics. Rev. Financ. Stud. 34, 5149–5187 (2021).
effect on the economic module 9. Kaplan, G., Moll, B. & Violante, G. L. The great lockdown and the
big stimulus: tracing the pandemic possibility frontier for the us.
Calibration and stochasticity NBER https://doi.org/10.3386/w27794 (2020).
For both the epidemic and economic results, uncertainty comes from 10. Alvarez, F., Argente, D. & Lippi, F. A simple planning problem
(1) stochasticity in the simulation runs, namely inherent stochasticity for COVID-19 lockdown, testing, and tracing. Am. Econ. Rev. 3,
of transmission in the susceptible, latent, infected, removed model 367–382 (2021).
and inherent stochasticity in the hiring/firing process of the economic 11. Baqaee, D., Farhi, E., Mina, M. & Stock, J. H. Policies for a second
module; (2) uncertainty over parameter values, as obtained from the wave. Brook. Pap. Econ. Act. 2020, 385–443 (2020).
Approximate Bayesian Computation calibration algorithm that we 12. Delli Gatti, D. & Reissl, S. Agent-based COVID economics (ABC):
use to calibrate the seven parameters that we cannot pin down inde- assessing non-pharmaceutical interventions and macro-
pendently (Supplementary Section 4). This means that, in line with a stabilization policies. Ind. Corp. Change 31, 410–447 (2022).
Bayesian approach, we run simulations sampling from all parameter 13. Basurto, A., Dawid, H., Harting, P., Hepp, J. & Kohlweyer, D.
values accepted by the Approximate Bayesian Computation. Economic and epidemic implications of virus containment
policies: insights from agent-based simulations. Bielefeld Working
Reporting summary Papers in Economics and Management https://doi.org/10.2139/
Further information on research design is available in the Nature ssrn.3635329 (2020).
Portfolio Reporting Summary linked to this article. 14. Mellacher, P. COVID-town: an integrated economic–
epidemiological agent-based model. GSC Discussion Paper
Data availability Series, 23 (2020).
All the data used in the economic module are publicly available and 15. Aleta, A. et al. Modelling the impact of testing, contact tracing
can be obtained from the American Community Survey, Bureau of and household quarantine on second waves of COVID-19.
Economic Analysis, Bureau of Labor Statistics and other sources, as Nat. Hum. Behav. 4, 964–971 (2020).
described in Supplementary Information. The mobility data used in 16. Aleta, A. et al. Quantifying the importance and location of
the epidemic module are available from Cuebiq, available upon request SARS-CoV-2 transmission events in large metropolitan areas.
submitted to https://www.cuebiq.com/about/data-for-good/. The Proc. Natl Acad. Sci. USA 119, e2112182119 (2022).
places data are obtainable from the Foursquare API (https://foursquare. 17. Pichler, A., Pangallo, M., del Rio-Chanona, R. M., Lafond, F. &
com/products/places, accessed 16 February 2021). Farmer, J. D. Forecasting the propagation of pandemic shocks
with a dynamic input-output model. J. Econ. Dyn. Control 144,
Code availability 104527 (2022).
The code that has been used to obtain all the quantitative results in 18. Miller, R. E. & Blair, P. D. Input–Output Analysis: Foundations
the paper is publicly available on Zenodo (https://doi.org/10.5281/ and Extensions (Cambridge Univ. Press, 2009); https://doi.org/
zenodo.7946867). Given that we cannot share the mobility data, as 10.1017/CBO9780511626982
explained in the data availability statement above, we created syn- 19. Cox, N. et al. Initial impacts of the pandemic on consumer
thetic random mobility networks. Thus, we cannot expect the code behavior: evidence from linked income, spending, and savings
to quantitatively reproduce the results in the paper. Despite this, we data. Brook. Pap. Econ. Act. 2020, 35–82 (2020).
provide a demo with a sanity check that shows that the model behaves 20. Hodbod, A., Hommes, C., Huber, S. J. & Salle, I. The
as expected in obtaining aggregate unemployment and infections as a COVID-19 consumption game changer: evidence from
consequence of different closure policies. a large-scale multi-country survey. Eur. Econ. Rev. 140,
103953 (2021).
References 21. Poletti, P., Ajelli, M. & Merler, S. The effect of risk perception
1. Flaxman, S. et al. Estimating the effects of non-pharmaceutical on the 2009 h1n1 pandemic influenza dynamics. PLoS ONE 6,
interventions on COVID-19 in europe. Nature 584, 257–261 (2020). e16460 (2011).
2. del Rio-Chanona, R. M., Mealy, P., Pichler, A., Lafond, F. & 22. Bedson, J. et al. A review and agenda for integrated disease
Farmer, J. D. Supply and demand shocks in the COVID-19 models including social and behavioural factors. Nat. Hum. Behav.
pandemic: an industry and occupation perspective. Oxf. Rev. 5, 834–846 (2021).
Econ. Policy 36, 94–137 (2020). 23. Tizzoni, M. et al. Addressing the socioeconomic divide in
3. Chetty, R., Friedman, J. N., Stepner, M. & the Opportunity Insights computational modeling for infectious diseases. Nat. Commun.
Team. The economic impacts of COVID-19: evidence from https://doi.org/10.1038/s41467-022-30688-8 (2022).
a new public database built from private sector data. Q. J. Econ. 24. Hallegatte, S. An adaptive regional input–output model and its
https://doi.org/10.1093/qje/qjad048 (2023). application to the assessment of the economic cost of Katrina.
4. Goolsbee, A. & Syverson, C. Fear, lockdown, and diversion: Risk Anal. 28, 779–799 (2008).
comparing drivers of pandemic economic decline 2020. J. Public 25. Guan, D. et al. Global supply-chain effects of COVID-19 control
Econ. 193, 104311 (2021). measures. Nat. Hum. Behav. 4, 577–587 (2020).
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 274
Article https://doi.org/10.1038/s41562-023-01747-x
26. Russell, T. W. et al. Effect of internationally imported cases on acknowledge the use of the computational resources of the
internal spread of COVID-19: a mathematical modelling study. Complex Systems & Networks (COSNET) Lab at the Institute for
Lancet Public Health 6, e12–e20 (2021). Biocomputation and Physics of Complex Systems (BIFI), funded by
27. Flegg, A. T. & Webber, C. Regional size, regional specialization and Banco Santander through grant Santander-UZ 2020/0274 and by the
the FLQ formula. Reg. Stud. 34, 563–569 (2000). Government of Aragón (FONDO-COVID19-UZ-164255). The funders
28. Bureau of Economic Analysis. Table 1.1.7. Percent change from had no role in study design, data collection and analysis, decision to
preceding period in prices for gross domestic product (2020); publish or preparation of the manuscript.
https://apps.bea.gov/iTable/?reqid=19&step=3&isuri=1&1921=surv
ey&1903=11 Author contributions
29. Python/C API Reference Manual. Python (2020) https://docs. M.P. proposed the collaboration. All authors designed the research.
python.org/3/c-api/ M.P. and A.A. performed the research, with contributions from
30. Liu, W.-m, Hethcote, H. W. & Levin, S. A. Dynamical behavior of R.M.d.R.C., A.P. and D.M.-C. M.P. and A.A. wrote the first draft. All
epidemiological models with nonlinear incidence rates. J. Math. authors contributed to the interpretation of the results, writing and
Biol. 25, 359–380 (1987). editing of the paper.
31. Funk, S., Gilad, E., Watkins, C. & Jansen, V. A. The spread of
awareness and its impact on epidemic outbreaks. Proc. Natl Acad. Competing interests
Sci. USA 106, 6872–6877 (2009). The authors declare no competing interests.
32. Fenichel, E. P. et al. Adaptive human behavior in epidemiological
models. Proc. Natl Acad. Sci. USA 108, 6306–6311 (2011). Additional information
33. Funk, S., Salathé, M. & Jansen, V. A. Modelling the influence of Supplementary information The online version
human behaviour on the spread of infectious diseases: a review. contains supplementary material available at
J. R. Soc. Interface 7, 1247–1256 (2010). https://doi.org/10.1038/s41562-023-01747-x.
34. Verelst, F., Willem, L. & Beutels, P. Behavioural change models for
infectious disease transmission: a systematic review (2010–2015). Correspondence and requests for materials should be addressed
J. R. Soc. Interface 13, 20160820 (2016). to Marco Pangallo.
35. Manfredi, P. & D’Onofrio, A. Modeling the Interplay between
Human Behavior and the Spread of Infectious Diseases Peer review information Nature Human Behaviour thanks the
(Springer Science & Business Media, 2013); https://doi.org/ anonymous reviewers for their contribution to the peer review of
10.1007/978-1-4614-5474-8 this work. Peer reviewer reports are available.
36. Perra, N., Balcan, D., Gonçalves, B. & Vespignani, A. Towards
a characterization of behavior–disease models. PLoS ONE 6, Reprints and permissions information is available at
e23084 (2011). www.nature.com/reprints.
Acknowledgements Publisher’s note Springer Nature remains neutral with regard to
M.P., R.M.d.R.C. and A.P. acknowledge funding from the James jurisdictional claims in published maps and institutional affiliations.
S. McDonnell Foundation Postdoctoral Fellowship Award. F.L. and
J.D.F. acknowledge funding from Baillie Gifford and the Institute Open Access This article is licensed under a Creative Commons
for New Economic Thinking at the Oxford Martin School. A.A. Attribution 4.0 International License, which permits use, sharing,
acknowledges support through the grant RYC2021-033226-I adaptation, distribution and reproduction in any medium or format,
funded by MCIN/AEI/10.13039/501100011033 and the European as long as you give appropriate credit to the original author(s) and the
Union ‘NextGenerationEU/PRTR’. Y.M was partially supported by the source, provide a link to the Creative Commons license, and indicate
Government of Aragon, Spain and ‘ERDF A way of making Europe’ if changes were made. The images or other third party material in this
through grant E36-23R: Física Estadística y No Lineal (FENOL), article are included in the article’s Creative Commons license, unless
and by Ministerio de Ciencia e Innovación, Agencia Española de indicated otherwise in a credit line to the material. If material is not
Investigación (MCIN/AEI/10.13039/501100011033) grant no. PID2020- included in the article’s Creative Commons license and your intended
115800GB-I00. M.A., M.C. and A.V. acknowledge the support of the use is not permitted by statutory regulation or exceeds the permitted
from the HHS/CDC 6U01IP001137 and HHS/CDC 5U01IP0001137. use, you will need to obtain permission directly from the copyright
E.M. acknowledges acknowledges support by Ministerio de holder. To view a copy of this license, visit http://creativecommons.
Ciencia e Innovación, Agencia Española de Investigación (MCIN/ org/licenses/by/4.0/.
AEI/10.13039/501100011033) grant PID2019-106811GB-C32 and
the National Science Foundation grant no. 2218748. All authors © The Author(s) 2023
Nature Human Behaviour | Volume 8 | February 2024 | 264–275 275
File and source
- File
- the-unequal-effects-of-the-health-economy-trade-off-during-the-covid-19-pandemic.pdf
- Size
- 5,744,724 bytes
- SHA-256
- 71634f34b49efdd1c759c67a7cfc6ea83657fc12350c74c9dc11b18da2363c21
- Original
- doi.org