Macroeconomic consequences of the COVID-19 pandemic

Abstract

We estimate the economic impacts of COVID-19 in the U.S. using a disaster economic consequence analysis framework implemented by a dynamic computable general equilibrium (CGE) model. This facilitates identification of relative influences of several causal factors as “shocks” to the model, including mandatory business closures, disease spread trajectories, behavioral responses, resilience, pent-up demand, and government stimulus packages. The analysis is grounded in primary data on avoidance behavior and healthcare parameters. The decomposition of the influence of various causal factors will help policymakers offset the negative influences and reinforce the positive ones during the remainder of this pandemic and future ones.

1. Introduction and overview

1.1. Background

COVID-19 has had major consequences for the economy of the United States. Several studies have estimated its total impacts on GDP in the trillions of dollars, even before the Delta and Omicron variants ran their course (del Rio-Chanona et al., 2020; Dixon et al., 2020; Ludvigson et al., 2020; Thunström et al., 2020; Walmsley et al., 2021b). The pandemic's total economic impacts are estimated to be twice as great as those of the Great Recession, 20 times greater than the 2001 World Trade Center attacks, and 40 times greater than any natural disaster that befell the country in this century (Rose, 2021). The more than 1,000,000 COVID-19 deaths in the U.S. through December 2022 (CDC, 2022), are greater than the U.S. death toll over the past four decades of the HIV/AIDS epidemic of approximately 700,000 people (KFF, 2021) and U.S. deaths from the Spanish Flu a century ago of approximately 675,000 people (CDC, 2020h). U.S. deaths from COVID-19 are more than nine times the country's death toll from the influenza pandemic of 1957–1958 (CDC, 2020h) and the Hong Kong flu in 1968 (CDC, 2020h). COVID-19 is caused by the SARS-CoV-2 virus, first recorded in the U.S. in January 2020.

Initial studies of the economics of COVID-19 have primarily used macroeconomic and financial models, likely because the most recent great disaster to hit the U.S. was a financial meltdown. Moreover, given the nature of these models in dealing with aggregates and the lack of data at the early stages, the modeling approaches have primarily been top-down. Thus, they were only able to include a limited number of causal factors unique to the pandemic and its impact on the overall economy.1

An alternative is a bottom-up approach that focuses on individual causal factors prevalent in disasters, including economy-wide impacts, though not including all financial considerations (likely to be minor in comparison). Research on the impacts of disasters has led to the development of a comprehensive economic consequence analysis (ECA) framework that has applicability to COVID-19. This framework extends ordinary impact analysis covering standard direct and indirect effects to include resilience and behavioral responses that have played an especially prominent role in a range of disasters including earthquakes, hurricanes, ocean oil spills, terrorist attacks among others (Rose et al., 2014; Rose et al., 2017; Prager et al., 2017; Botzen et al., 2020). We will demonstrate below how the framework is applicable to the COVID pandemic as well. For example, resilience is exemplified by telework and spending out of pent-up demand. Behavioral responses are exemplified by avoidance of public gathering places, mass transit, workplaces, and educational institutions.

1.2. Purpose

The purpose of this paper is to estimate the total economic impacts of COVID-19 in the U.S. We apply the CREATE ECA framework in the process, which enables us to identify the relative influence of several causal factors. We utilize a state-of-the-art dynamic, multi-country CGE model to estimate the time path of the impacts of these factors. CGE analysis essentially models the economy as a set of interconnected supply chains, which enables us to calculate indirect effects as well. Moreover, our analysis is based on the most up-to-date historical data on key causal factors and primary data relating to avoidance behavior and health outcomes and the effects of alternative intervention scenarios. Avoidance is the main behavioral response, and in this case pertains to people engaging in 10 types of activities like retail shopping, going to work, sending children to school, using public transportation, and going to entertainment venues. Moreover, we base our estimates of health outcomes on interviews with medical professionals. We decompose our analysis by performing a set of sequential dynamic simulations beginning with mandatory closures, gradual reopenings, avoidance behavior, and shifts in level and composition of healthcare. We include aspects of resilience, such as the increase in telework and the injection of pent-up demand at later stages of the recovery. Finally, we include the impact of various stimulus packages. All of these effects are disaggregated according to the 65 sectors of the dynamic CGE model to improve the accuracy of the estimation since general equilibrium (supply-chain) effects differ by sector. Given the many uncertainties around several causal factors and the trajectory of infection/hospitalization/deaths due to COVID, we consider alternative scenarios embodying combinations of key assumptions, and including several sensitivity tests.

This study builds on prior work of the research team on the economic consequences of COVID-19. Walmsley et al. (2021a) estimated the consequences of mandatory closures not only in the U.S. but around the globe. Walmsley et al. (2021b) used more up-to-date data to examine the effect of reopenings through the summer of 2020. Rose et al. (2021) analyzed the transmission of these consequences through international trade between the U.S., China, and the rest of the world. All of these studies included major causal factors analyzed in the current study, though they omitted the effects of fiscal stimulus packages. Moreover, all of the studies were based on secondary data and used a static CGE model, while the current study utilizes primary data and performs the analysis with a dynamic CGE model.

The decomposition of the influence of various causal factors is important for several reasons. First, it improves analytical insight into their workings and interrelationships. Second, it enables policymakers to identify the most prominent causal factors, which is the first step toward making adjustments to offset the negative influences and reinforce the positive ones. For example, our analysis indicates that the mandatory closures by far had the largest negative effect on the results and that the slow reopenings, due to such factors as supply-chain bottlenecks and mixed messages and political debate about COVID contagion, continued some of that influence. Avoidance effects and telework were also prominent influences. In the case of the former, improved risk communication, a relatively inexpensive way to quell fear, has great potential to reduce economic losses. The same is true of public and private decision-makers placing a greater emphasis on providing remote work options.

1.3. Overview

This paper is divided into 11 sections. Section 2 presents a summary of the ECA framework. Section 3 provides the basis of the survey of avoidance behavior and its results. Section 4 presents the modeling results of health outcomes under alternative pandemic scenarios in terms of the efficacy of interventions informed by healthcare interviews. Section 5 presents the analysis of data on mandatory closures, reopening, telework, and pent-up demand, as well as the refinement of the avoidance and healthcare data for inclusion in the CGE model. This section also presents an assessment of the six federal stimulus packages that have been enacted to date and how they are translated for use in the CGE model. Section 6 presents the workings of the CGE model. Section 7 provides a presentation and the interpretation of the aggregate impacts, as well as a decomposition of the results. Section 8 provides a comparison of the results with other recent findings and discusses the limitations of our analysis. Section 8 provides a summary and suggestions for future research.2

2. A disaster economics framework

2.1. Overview

A major example of a disaster economics approach is the CREATE3 Economic Consequence Analysis (ECA) Framework (Rose et al., 2009; 2017; Rose, 2015), which has been applied to numerous actual and simulated disasters (see, e.g., Rose et al., 2009; Sue Wing et al., 2016; Chen et al., 2017), including a moderate-sized pandemic (Prager et al., 2017). It represents an advance over conventional economic impact analysis that has been applied for decades to more ordinary situations such as the closing of an automobile plant.

It begins with the specification of direct economic impacts and then typically applies a state-of-the-art economy-wide model, such as a computable general equilibrium (CGE) model, to estimate total economic impacts. However, its major contribution is the inclusion of two additional considerations unique to disasters.

The first is resilience, which pertains to various actions taken by producers, consumers, and governments to recover from the shock and mute the negative impacts on economic activity by using remaining resources efficiently (Rose and Liao, 2005; Rose et al., 2009; Rose et al., 2017; Martin and Sunley, 2014). For example, Rose et al. (2009) estimated that the rapid relocation of tenants of the World Trade Center following the 9/11 attacks reduced the potential GDP impacts by 72%. Several other studies have found various other resilience tactics (e.g., conservation, input substitution, inventories, excess capacity) to be a powerful strategy to reduce business interruption (BI) losses from disasters in the range of 20%–80% of the potential losses based on surveys and simulation analyses (see, e.g., Kajitani and Tatano, 2009; Wei et al., 2020; Dormady et al., 2022).

The ECA framework also includes behavioral responses, which are typically motivated by fear, often far beyond the site of the disaster, and exacerbate the losses (e.g., von Winterfeldt and O’Sullivan, 2006; Gordon et al., 2007; Rose, 2022). The 9/11 Study also found that more than 80% of the GDP impacts were due to the almost two-year reduction in airline travel and related tourism. Other studies have found that behavioral responses to disasters can increase BI losses by one or two orders of magnitude (see, e.g., Giesecke et al., 2012; Rose et al., 2017; Gertz and Davies, 2019).

2.2. Causal factors affecting the economic consequences of COVID-19

The CREATE ECA Framework helps us distinguish several different categories of factors affecting the consequences of disasters. These factors can be summarized in general terms as:

• •Ordinary direct impacts relating to property damage, economic activity, and health
• •Remediation, repair, and reconstruction
• •Resilience
• •Behavioral responses
• •Indirect economic impacts stemming from the above four categories.

In this paper we populate these categories with specific causal factors related to COVID-19, which become our model “shocks”:

• 1.Mandatory closures [Behavioral Linkage]
• 2.Reopenings [Behavioral Linkage]
• 3.Workforce declines due to health issues [Direct Effects]
• 4.Telework [Resilience]
• 5.Consumption and workforce declines due to avoidance [Behavioral Linkage]
• 6.Changes in net demand for health care services [Direct Effects of Pandemic and Behavioral Linkage in part]
• 7.Pent-up demand [Resilience]
• 8.Stimulus packages [Resilience at the macroeconomic level]
• 9.Indirect effects of all of the above

Nearly every one of these factors involves some aspects of resilience and behavioral linkages. This helps illustrate that conventional top-down macroeconomic analysis is likely to miss important considerations in estimating the impacts of the pandemic. This modeling approach serves as a valuable organizing framework for the various factors and is especially adept at estimating indirect effects (see Dixon et al., 2020; Walmsley et al., 2021a, 2021b). Moreover, the dynamic CGE formulation enables us to model the time-path of the impacts of the various factors, including lags in their effects. In the next two sections we will indicate how two of these causal factors are incorporated into the modeling in terms of shocks.

3. Avoidance behavior survey

3.1. Overview

Avoidance behavior has the potential to significantly affect the bottom-line economic impacts of COVID-19. The present study advances prior methods developed for survey research by the authors on the public's response to flu pandemics (Rosoff et al., 2012) and urban biological terror attacks (Rosoff et al., 2013). The survey estimates changes in behavior due to the COVID-19 pandemic in 10 different domains:

• 1)Staying home from work
• 2)Keeping children home from school
• 3)Canceling or postponing medical and dental appointments
• 4)Canceling or postponing professional grooming and spa treatments
• 5)Canceling or postponing domestic and international air travel
• 6)Avoiding public transportation, e.g., taking the bus, rideshare
• 7)Avoiding local leisure activities, e.g., dining out, bars
• 8)Avoiding shopping, e.g., grocery, miscellaneous shopping
• 9)Avoiding recreational activities, e.g., golf, tennis, swimming
• 10)Avoiding large crowds, e.g., sports events, concerts, shows

For each domain, a representative sample of U.S. adult participants was asked to consider their behavior over the six months from November 2020 to April 2021 compared to their behavior before the pandemic (pre-March 2020). Participants indicated whether specific activities had increased, decreased, or stayed the same in the previous six months compared to the same activities before the pandemic. Participants who reported a change in the activity were asked to indicate the amount of increase or decrease and the reason for the change.

3.2. Survey content

A series of specific activities were identified within each of the 10 domains, and participants were asked to reflect on their engagement in each activity during the two distinct periods. Participants who indicated a change in engagement were also asked to estimate absolute frequencies or percentages of increase or decrease (depending on the activity). To estimate changes due specifically to pandemic avoidance, participants were also asked to indicate whether the change was due primarily to one of three reasons: 1) Unavailability of the activity due to COVID, 2) Avoidance of the activity by the respondent due to COVID, or 3) Other.4 A complete copy of the entire survey can be found in Rose et al. (2021).

3.3. Participants and sampling weights

Data were collected in two separate waves – April 23, 2021, and May 3, 2021. All data were collected using the crowdsourcing platform, Prolific, a widely accepted source of online participants for behavioral research (Palan and Schitter, 2018; Peer et al., 2017). A sample of U.S. adults of just over 660 adult respondents was obtained in each wave, for a total sample of N = 1328.

3.4. Estimating decrease in activities due to COVID-19 avoidance

The sample of respondents who participated in paid work reported a mean decrease of 1.25 days/week working outside the home during the pandemic due to all reasons, including avoiding exposure to COVID-19. The same sample of working respondents also reported a mean increase of 0.86 days/week worked at home during the pandemic due to all reasons. The sample of respondents with school-age children reported a mean decrease of 2.02 days/week of in-person school attendance during the pandemic due to all reasons, including parents' decision to avoid children's exposure to COVID-19.

We estimated the impact of COVID-19 avoidance behavior on health/medical and professional grooming visits in terms of the average number of canceled or postponed appointments from November 2020 through April 2021. We isolated the effects of behavioral avoidance of COVID-19 from other effects, such as mandated closures. Respondents indicated whether they had canceled or postponed three separate types of medical-care visits and whether they had substituted telemedicine visits for in-person visits during the pandemic. Respondents also indicated the approximate number of medical visits avoided (or substituted) and the primary reason for avoiding (or substituting) during this period. We focus on medical visits avoided due to behavioral avoidance of COVID, as opposed to other reasons, such as no insurance due to loss of job or unavailability of in-person medical visits for whatever reason.

In order to calculate the mean number of health care visits avoided per person due to COVID for those who utilized health care, we used the following equation:

$${\mathrm{M}}_{\mathrm{VA}}={\mathrm{M}}_{\mathrm{VA}|\mathrm{AE}}*\mathrm{p}\left(\mathrm{A}\mathrm{E}|\mathrm{D}\mathrm{V}\right)*\mathrm{p}\left(\mathrm{D}\mathrm{V}|\mathrm{U}\mathrm{C}\right)$$

where,

M_VA = Mean number of health care visits canceled or not scheduled to avoid COVID,

M_VA|AE = Conditional mean number of health care visits canceled or not scheduled to avoid COVID for those who reported avoiding COVID exposure,

p(AE|DV) = Proportion of those reporting decreased medical visits who reported avoiding COVID exposure,

p(DV|UC) = Proportion of those utilizing health care who reported decreased medical visits.

Avoidance of COVID-19 exposure during the pandemic (November 2020 through April 2021) was responsible for 0.73 fewer routine medical visits, 0.26 fewer medical visits for non-life-threatening acute problems, and 0.10 fewer visits for medical (diagnostic and treatment) procedures. Respondents’ desire to avoid exposure to COVID-19 accounted for about 0.39 additional telemedicine visits substituted for in-person visits over six months of the pandemic.

The impacts of COVID-19 avoidance behavior across the six remaining activity categories were isolated and estimated in terms of percentage decrease during the pandemic using the same approach as described for medical/health appointments. Estimates are based on respondents who are engaged in each activity and are summarized in Fig. 1 , which plots the percentage of decreased activity. These percentages isolate the effects of behavioral avoidance of COVID-19 from other effects, such as mandatory business closures and stay-at-home mandates. The bars represent the percentage decrease in activity during six months of the pandemic (November 2020 through April 2021) compared to pre-pandemic (before March 2020) activity. Pure behavioral avoidance effects, over and above business closures and other government mandates, resulted in a 40%–70% reduction in activities during the pandemic (November 2020 through April 2021) compared to pre-pandemic activities. The only exceptions to this were two activities that did not require sustained presence in an indoor facility, namely to-go/delivery dining and outdoor recreational activities.

Fig. 1.

Summary of percentage decreases in activities across six domains during the pandemic compared to pre-pandemic attributable to behavioral avoidance of covid.

Increases in shopping and dining substitution activities attributable to COVID avoidance, over and above convenience considerations, or mandated business closures, were also isolated and estimated in terms of percentage increases in substitution activities. These substitution effects are substantial, ranging from 23% to 30% for shopping activities to over 36% for the use of to-go and food-delivery dining alternatives, making up for some of the shopping and dining activity reductions attributable to behavioral avoidance of COVID-19 during the pandemic.

4. Trajectory of the pandemic under select intervention scenarios

The U.S. started to screen for COVID-19 at select airports on January 20, 2020, and confirmed the first domestic case of the infection on January 21, 2020. A public health emergency was declared on February 3, 2020, followed by a declaration by the WHO on March 11, 2020, of a pandemic (American Journal of Managed Care, 2021).

4.1. Scenarios

We develop four COVID-19 pandemic scenarios in the U.S. between January 2020 and June 2022, with a benchmark scenario calibrated using reported COVID-19 deaths, and three scenarios with alternative assumptions on the efficacy of interventions. Our estimates of SARS-CoV-2 transmission use a traditional Susceptible-Exposed-Infectious-Recovered (SEIR) compartmental model. We draw on estimates of disease characteristics and mathematical models estimating the effectiveness of select interventions and virus seasonality, referenced below. Where empirical estimates were not available, we leveraged expert interviews to develop the most likely and realistic epidemiological scenarios. Our model is described in detail in Hlávka et al. (forthcoming). The key results of this model are reported in Tables A3, A4 and Figure A4 in Appendix A.

Table 3.

Magnitude of Avoidance Behaviors and CGE Modeling Method.

Avoidance Behavior	GTAP CGE Modeling Linkage	Impacts on the GTAP Sector(s)
Staying home from work	Labor force productivity impact	−4.58%
Keeping children from school (education sector impact)	Reduced demand in Education sector	−1.59%
Reduction of in-person school attendance (caregiver impact)	Labor force impact	−2.71%
Avoiding medical professionals	Demand reduction in Human Health and Social Work sector	−29.16%
Reducing shopping	Demand reduction in Wholesale and Retail Trade sector	−5.85%
Avoiding local leisure activities	Demand reduction in Recreation & Other Services sector	−21.17%
Avoiding dining out	Demand reduction in Accommodation, Food and Service Activities sector	−26.50%
Avoiding public transportation	Demand reduction in Land Transport and Transport via Pipelines sector	−4.21%
Canceling air travel	Demand reduction in Air Transport sector	−57.48%
	Demand reduction in Accommodation, Food and Service Activities sector	−9.67%

4.1.1. Scenario 1: real-world scenario

In approximating real-world estimates, our model draws on data published by the CDC until July 2021 (we calibrate it to approximate mortality reported for the age bracket under-65-years and the age bracket 65-years and older in the U.S., using baseline modeling parameters in Table 1 and transmission reduction ratios that result as mortality estimates aligned with weekly CDC reporting) (CDC, 2021e). The model makes projections through June 2022. Our calibration was conducted at weekly increments in 2020 and monthly increments thereafter. We assume that the reduction in the effective reproductive value attributed to behavioral interventions in July 2021 remains consistent through June 2022, but we do not explicitly model new viral strains that may increase compliance or decrease the effectiveness of behavioral and pharmaceutical interventions.5

Table 1.

Modeling Parameters and Assumptions.

Parameter	Value	Source(s)
Basic reproductive number (R0)	3.32	Alimohamadi et al. (2020)
Initial susceptible population	274,165,495 (<65) 54,074,028 (65+)	U.S. Census Bureau (2019)
Initial infections on January 1, 2020	0	Assumed
Daily infected international arrivals, January and February 2020	30 (January 2020) 45 (February 2020)	Calibrated based on real-world data
Daily infected international arrivals, March 2020 through June 2022	between 0 and 947	Calculated based on I-94 Visitor Arrival Program; Bureau of Transportation Statistics 2019–2020 (Bureau of Transportation Statistics, 2021)
Latent period (pre-infectious)	2.5 days	Vardavas et al. (2021); Rhee et al. (2021)
Infectious period (asymptomatic, pre-symptomatic and symptomatic)	7.5 days	CDC (2021b); Rhee et al. (2021)
Hospitalization rate (h)	2.90% (<65) 21.21% (65+)	CDC (2021c)
Age-specific infection fatality rate (IFR)	0.0969% (<65) 3.3049% (65+)	O'Driscoll et al. (2021)
Vaccine effectiveness (VE)	90%	CDC (2021a)
Proportion of population willing to vaccinate	73.54% (<65) 84.96% (65+)	Kapteyn and Gutsche (2021) weighted by American Community Survey (U.S. Census Bureau, 2019)
IFR Multiplier during peak hospitalization periods	Under 75,000 hospitalizations: 1.00 Over 75,000 hospitalizationsa: 1.20 Over 125,000 hospitalizationsb: 1.30	Based on Rossman et al. (2021)
Seasonality adjustment coefficient	0.75–1.27	The effective reproductive rate is a factor of the basic reproductive number and the seasonality coefficient. Based on Gavenčiak et al. (2021), see Appendix 5.A.6

^aBased on the onset of highest hospitalization counts in Fall (2020) and Summer 2021, which resulted in local shortages of intensive care beds and personnel.

^bBased on the reported peak of hospitalizations when many states were running out of intensive care beds and personnel to provide the highest-quality care.

4.1.2. Scenario 2: high-efficacy scenario

We model a hypothetical high-efficacy scenario to indicate a combination of successful policies that would reduce the adjusted reproduction number to under 1.0, which indicates suppression of transmission over time. We consider effective disease management policies that other countries achieved through measures such as mask mandates, remote work, closure of non-essential businesses, contact tracing, and quarantines (Brauner et al., 2021; Flaxman et al., 2020).

4.1.3. Scenario 3a: very low-efficacy scenario

The very low-efficacy scenario assumes limited willingness of policy-makers to adopt public health interventions affecting social contact, such as business closures. As a result, compliant behavioral response leads to a maximum behavior-attributed reduction in transmission of 32% (based on Sharma et al., 2021). This reduction reduces the effective reproductive number from 3.32 to 2.26, on average. Behavioral fatigue is also taken into account, and, over time, effectiveness is expected to decrease (see Figure A-3c in Appendix A).

4.1.4. Scenario 3b: low-efficacy scenario

In this scenario, we use the estimated effectiveness of mask use from Sharma et al. (2021) in conjunction with the restriction of gatherings. Scenario 3b uses 50% as the maximum reduction (in the initial months of intervention) before declining due to behavioral fatigue (see Appendix Figure A-3d). Vaccination was incorporated into both the real-world scenario and the high-efficacy scenarios. Data for vaccination was drawn from the CDC (CDC, 2021f). In our modeling, 90% of individuals who were vaccinated were removed from the susceptible population permanently and moved to the vaccinated population (see Section B3 of Appendix B).

4.2. Inputs

4.2.1. Expert interviews

We conducted 12 interviews with experts from variable backgrounds, including scientists in the field of public health and epidemiology from the United States, Canada, and the United Kingdom, state- and federal-level government institutions and clinicians. IRB approval was obtained, and honoraria were offered. These interviews covered multiple areas of interest to the research team, including specification of interventions (social distancing, mask-wearing, and others), uncertainty in key parameters of interest (notably the infection fatality rate (IFR), seasonality, hospitalization rates, vaccination effectiveness, and others), and effects of pandemic outcomes on the behavioral response. Further details about expert input are provided in Section B1 of Appendix B.

4.2.2. Data and model inputs

Drawing on published literature as well as expert input, our parameters and assumptions for SEIR and health outcome modeling are listed in Table 1.6

4.3. Scenario specification

We present the number of cases of COVID-19 by age, period, and scenario in Table 2 . Other outcomes of interest are reported in Section B6 of Appendix B.

Table 2.

Age-Stratified COVID-19 Cases in Each 6 Months (in thousands).

	Cases Age under 65				Cases Age 65 or Over
Scenario:	1	2	3a	3b	1	2	3a	3b
H1/2020	30,531	6564	209,178	134,521	3430	617	36,930	18,575
H2/2020	53,606	2143	15,611	47,298	6896	237	6571	17,200
H1/2021	43,923	3900	354	1505	4178	298	40	622
H2/2021	4433	1699	266	376	1834	92	27	36
H1/2022	3347	5455	91	198	7228	139	10	18
Total	135,840	19,761	225,500	183,897	23,565	1383	43,578	36,451

Sums of semi-annual periods may not equal total due to rounding.

4.3.1. Scenario 1: real-world scenario

In our real-world scenario, we estimate 142.6 million cases of COVID-19 by June 2021. Over the full 30-month period, 159.4 million infections are expected (23.6 million among 65+, 135.8 million among <65). These cumulative infection numbers represent 43.5% and 49.5%, respectively, of the initial susceptible populations in each age group. These infections result in 910,000 deaths between January 2020 and June 2022 (see Appendix A for more detail).

Notably, even with a relatively high number of cases and deaths, and vaccination willingness at 73.54% for the <65 population and 84.96% for the 65+ population, our model predicts another wave of COVID-19 deaths in Fall (2021)/Spring 2022 if behavioral non-pharmaceutical interventions continue unchanged from July 2021 onwards.

4.3.2. Scenario 2: high-efficacy scenario

For this scenario, we estimate 21.1 million cases of COVID-19 by June 2022. These cumulative infection numbers represent 2.6% and 7.2%, respectively, of the initial susceptible populations in each age group. These infections result in 65,000 deaths between January 2020 and June 2022.

4.3.3. Scenario 3a: very low-efficacy scenario

Here we estimate 269.1 million cases of COVID-19 by June 2022. These cumulative infection numbers represent 80.6% and 82.2%, respectively, of the initial susceptible populations in each age group. These infections result in 1,993,000 deaths between January 2020 and June 2022.

4.3.4. Scenario 3b: low-efficacy scenario

Here we estimate 220.3 million cases of COVID-19 by June 2022. These cumulative infection numbers represent 67.4% and 67.1%, respectively, of the initial susceptible populations in each age group. These infections result in 1,504,000 deaths between January 2020 and June 2022.

Achieving the most optimistic outcomes would require a coordinated national response from early on in the pandemic (in our model, we assumed most interventions would be in place by April 2020, consisting of measures such as broad testing, quarantines, isolation, social distancing, and remote work) so that any further fluctuations in incidence would not result in exponential growth, which was observed in the real world. Given insights from the pandemic responses of other countries, it is reasonable to contend that better management of the pandemic could have saved hundreds of thousands of lives in the U.S., particularly among the elderly (Brauner et al., 2021; Sharma et al., 2021).

Our analysis has several limitations. Notably, SEIR modeling assumes homogenous mixing of the population with no specific modeling of local outbreaks (Tolles and Luong, 2020). We aim to address the different dynamics for younger lower-risk individuals, and older higher-risk individuals by utilizing two different susceptible population pools. Moreover, we adjust for temporal variability by introducing a seasonality coefficient, but the seasonal estimates we draw on are not specific to the United States. We also do not model other health outcomes – it is yet to be determined what the short- and long-term consequences of deferred health care will be. Another limitation includes the dependence on fixed parameters in the initial model design. We conducted expert interviews to collect independent input to key model design choices. Finally, we draw on external estimates, including the expected effectiveness of social distancing, which also contains uncertainty and variability over time, and we finalize our disease modeling inputs in June 2021, before data about the more infectious Delta (B.1.617.2) and Omicron (B.1.1.529) variant were documented. As a result, our results should be interpreted as approximations of possible disease dynamics rather than accurate estimates for each scenario we study.

5. Data on major drivers of economic consequences and fiscal stimulus

We collected data for some of the basic drivers of the economic consequences of the COVID-19 pandemic from various secondary sources. These include federal and state government agency websites on the status of mandatory closures/reopenings, CDC reports and other health publications, U.S. Bureau of Labor Statistics (BLS) survey results on telework, and foot traffic and credit card spending data for pent-up demand estimation. In this section, we summarize the collection of these data and their refinement and translation into CGE model inputs.

5.1. Mandatory shutdown and reopening process

During the initial outbreak of the coronavirus disease in the U.S. in 2020, 45 states plus D.C. implemented “stay-at-home” orders that also required the shut-down of non-essential businesses. In most states, the orders were issued in March and lifted by the end of May or early June, with an average length of 45 days (Wu et al., 2020; New York Times, 2021). To determine the impacts of the mandatory closures on various economic sectors, we defined three categories of sectors (Entirely Non-Essential; Partially Non-Essential; and Essential) based on the DHS Cybersecurity and Infrastructure Security Agency list of Essential Critical Infrastructure Sectors during COVID-19 (CISA, 2020).

To estimate the reopening process by sector between June and September 2020, detailed data on reopening stages and timelines were collected from the reopening plans issued in five major states: California, Texas, New York, Illinois, and Florida (Ting and Duree, 2020; Cowan and Arora, 2021; Florida Department of Health, 2021; Office of the Texas Governor, 2021; State of California, 2021; State of Illinois, 2021; State of New York, 2021). We also collected data on the reopening status for each state at three additional points in time: September and December 2020, and March 2021.7

Telework has been one of the most important resilience tactics during the mandatory closures and the phased-in reopening process. The data on telework potential (i.e., the percentage of workers in a specific sector that have teleworked or worked from home) by sector and by month post-COVID-19 were collected from U.S. Bureau of Labor Statistics (2021) survey results. The weighted average telework potential across economic sectors was over 40% at the beginning of the pandemic and reduced to about 30% as the economy gradually reopened. Telework potentials ranged from about 8% for Leisure and Hospitality sectors to over 70% for the Financial and Education sectors during mandatory closures (BLS, 2021). Appendix Table B-1 presents the estimated direct percentage reduction in U.S. annual GDP (after telework adjustments) by sector due to mandatory closures and the phased-in reopening process by 6-month period. The estimates for other countries are also discussed in Appendix B.

5.2. COVID-19 health expenditures and workday losses

The health outcomes of COVID-19 in terms of the number of outpatients, hospitalizations, and deaths are translated into total COVID health expenditures by using the average per-patient cost by health outcome category and by age group (see Appendix Table B-2) (Bartsch et al., 2020; Fusco et al., 2021). For the base case scenario, the total estimated COVID-related health expenditures are estimated to be $214 billion over the entire study period, or about 8.35% of the total annual gross output of the Health Care and Social Service sector. The impact is 0.78%, 14.23%, and 11.70% for Scenarios 2, 3a, and 3b, respectively. The detailed results are presented in Appendix Table B-3.

We also estimate lost workdays due to COVID illness and deaths and caring for sick family members. For hospitalized patients, this is calculated by adding the average length of illness onset to hospitalization, hospital stay by age group, and time between hospital discharge and returning to work/school (Fusco et al., 2021; Chen et al., 2020; Zhao et al., 2020; Walmsley et al., 2021). For outpatients, productivity losses are assumed to be 1.5, 1.9, and 5.3 days, respectively, for the three age groups (Prager et al., 2017). However, we assume that patients of all age groups spend 10 days “sick at home” given the CDC isolation guidelines, while adjusting for telework potential (23.7% economy-wide) for the 20–64 age group to estimate workday losses. Appendix Table B4 summarizes per patient loss of productivity measured in days by age and health outcome category.

Lost productivity due to caring for sick family members includes caring for sick children in the 0–18 age group, a sick spouse in the 18–64 age group, and sick elderly family members in the 65+ age group. We made similar assumptions as in Prager et al. (2017) and Dixon et al. (2010) in calculating the productivity day losses for this category. For the Base Case scenario, the total estimated workday losses are over 600 million days. Appendix Table B-5 presents total lost productivity for each scenario.

Table 5.

Percent Changes in Semi-annual U.S. GDP by Causal Effect (cumulative percent differences from baseline).

	2020_1	2020_2	2021_1	2021_2	2022_1	2022_2	2023_1	2023_2
Mandatory Closures	−26.3	−22.9	−21.7	−15.9	−12.3	−10.3	−8.9	−7.8
Avoidance	−12.2	−11.3	−10.2	−5.4	−4.6	−4.1	−3.8	−3.6
Deaths and Illness	0.0	−0.1	−0.3	−0.5	−0.6	−0.7	−0.8	−0.8
Health Care	0.8	0.9	0.8	0.7	0.6	0.5	0.5	0.4
Pent-up Demand	3.1	2.9	5.0	5.4	6.1	5.8	6.2	6.9
Fiscal Policy
Rounds 1–4	7.1	8.9	2.1	1.9	1.2	1.4	1.4	1.6
Round 5	0.0	0.0	1.7	1.4	0.2	0.2	0.1	0.1
Round 6	0.0	0.0	0.3	0.0	−4.3	−3.6	−3.5	−3.3
Totals	−27.6	−21.6	−22.3	−12.4	−13.7	−10.8	−8.8	−6.5

For other countries, workday losses were assumed to occur at the same rate as in the U.S. Similarly, for health care expenditures, an adjustment was made to the cost of health care based on differences in per capita health care expenditure in each country.

5.3. Avoidance behaviors

Our survey results on various types of avoidance behavior are translated into direct production impacts in the relevant GTAP sectors. In most cases, we multiply the percentage net decrease of a specific type of activity because of avoidance behavior by the percentage that the activity represents of the total output in the relevant GTAP sector (see Walmsley et al., 2021 for more discussion on the translation of survey results to CGE model inputs). Table 3 first presents the linkage that is used to analyze the impacts of various types of avoidance behavior in the CGE model and then summarizes the direct impact inputs to the corresponding GTAP sectors..

5.4. Pent-up demand

Pent-up demand represents another important source of economic resilience that helps reduce the negative impacts from shutdowns and individual avoidance behaviors during the pandemic. To estimate the changes in consumer demand for key types of goods and services over time because of COVID-19, we first estimate the “Lowest Point” level in consumption since the onset of COVID-19. For most consumption categories, this “Lowest Point” took place in early or mid-April 2020. Next, the percentage increase from the “Lowest Point” level for each category is estimated at five points in time with three-month intervals. The estimates are presented in Appendix Table B-6, calculated based on micro-level data from three distinct sources that track consumer credit card-spending information or foot traffic data at retail locations across different industries (Opportunity Insights, 2021; Unacast, 2021; SafeGraph, 2021).

Table 6.

Percent Changes in Semi-annual U.S. Employment by Causal Effect (cumulative percent differences from baseline).

	2020_1	2020_2	2021_1	2021_2	2022_1	2022_2	2023_1	2023_2
Mandatory Closures	−27.9	−23.5	−21.2	−15.6	−11.2	−8.8	−7.0	−5.7
Avoidance	−11.7	−8.9	−6.8	−4.7	−3.4	−2.3	−1.7	−1.3
Deaths and illness	0.1	0.7	1.1	0.7	0.5	0.9	1.1	1.1
Health Care	0.1	0.3	0.0	0.0	0.0	0.0	0.1	0.1
Pent-up Demand	2.1	0.4	6.1	9.4	5.4	3.9	4.1	5.0
Fiscal Policy
Rounds 1-4	12.0	12.0	1.0	0.7	0.1	0.1	0.1	0.2
Round 5	0.0	0.0	2.2	1.5	0.2	0.1	−0.1	−0.2
Round 6	0.0	0.0	0.4	0.3	−4.7	−3.5	−3.3	−3.0
Totals	−25.4	−19.0	−17.2	−7.7	−13.1	−9.6	−6.9	−3.9

Factors influencing consumer expenditures during the pandemic include, but are not limited to, closures and reopenings of businesses, government stimulus paychecks, consumer avoidance behavior, and pent-up demand. It is difficult to collect data to separate the effects of these individual factors. Therefore, we first simulate the actual changes in consumption using the data presented in Table B-6. Next we compare the model output with the results of simulations in which several other key factors are incorporated (including closures/reopenings, avoidance, and fiscal policy), and then attribute any positive differences to the effects of pent-up demand. Appendix D presents more details of the method used. Table D-6 in Appendix D shows the actual shocks given to private household demand.

5.5. U.S. Federal government fiscal stimulus legislation

The U.S. federal government enacted six stimulus bills related to COVID-19 from March 2020 to March 2021. Altogether, the six bills amount to an estimated $5.13 trillion in assistance, forgivable and subsidized loans, funding support for state and local government, tax credits, and government expenditures (CBO, 2021a). We break down each bill into several provisions. The cost of each provision was based on the text of the bill or projections by the Congressional Budget Office (CBO). Cash payments to individuals and enhanced unemployment insurance benefits were assumed to benefit households. Other provisions were matched to the most appropriate GTAP sector. For most stimulus provisions, we assume the federal government spends the stimulus funds over time rather than at once. Appendix C details how fiscal stimulus provisions are incorporated into the model. For further detailed analysis of the direct and general equilibrium impacts of the fiscal stimulus packages, see Walmsley et al. (2022b).

6. The CGE model

We refine a dynamic global supply-chain computable general equilibrium model to examine the impact of COVID-19 on the U.S., China, and the rest of the world.8 The model is based on the publicly available Dynamic GTAP model (Ianchovichina and Walmsley, 2012), which is a dynamic adaptation of the widely used GTAP model (Hertel et al., 1997; Corong et al., 2017).9 The model is further adapted to incorporate supply chains based on Walmsley and Minor (2016a, 2016b).10 Current and earlier versions of the model have been used in numerous studies (Hertel et al., 2014; Walmsley et al., 2014, 2021a, 2021b; Walmsley and Minor, 2020a, 2020b, 2021; Rose et al., 2021). In the description below and in Appendix D, we focus on a few key differences between this model and the GTAP model.

In the model, in each region, personal income is received by households supplying factors of production, and business and personal tax revenues are collected by a regional household. This income is then allocated to private $C$ and government $G$ consumption, and to savings $S$ which fund investment, using a Cobb-Douglas demand function:

$$U=C^{\alpha }.G^{\beta }.S^{\gamma }$$ (1)

where: U is the total utility for the region; and.

$\alpha ,\beta ,\gamma$ are distribution (shares) parameters.

Changes in savings rates caused by the pandemic and to government spending due to the fiscal stimulus are modeled through the changes in these distribution parameters ($\alpha ,\beta ,\gamma$)11

Private consumption is then allocated across commodities using constant difference elasticity (CDE)12

$$1=\sum _i{B_i.U_P}^{{{\rm Y}}_i.R_i}.{\left(\frac{P_i}{C}\right)}^{{{\rm Y}}_i}$$ (2)

Where: $U_P$ is the utility associated with private expenditure (C);

$P_i$ is the price of commodity i;

$B_i$ are the distribution parameters of commodity i; and

${{\rm Y}}_i$ and $R_i$ are the substitution and expansion parameters of commodity i.

As in the case of savings and government spending, commodity-specific distribution parameters can be used to incorporate changes in private consumption of commodity i. This is done by incorporating the distribution parameter variables as changes in relative willingness to pay (i.e., like prices) giving the following demand function (as it appears in the model code).13

qp(i,r) - pop(r) = sum(k, TRAD_COMM, EP(i,k,r) * pp(k,r)) + EY(i,r) * [yp(r) - pop(r)]

- sum(k, TRAD_COMM, EP(i,k,r) * [dpp_agg(MAPTC2TCA(k),r) - dppall_agg(r)])

- sum(k, TRAD_COMM, EP(i,k,r) * [dpp_stop(k,r) - dppall_stop(r)])

where: qp(i,r) is private consumption of commodity i in region r;

pop(r) is population;

EP(i,k,r) are the cross-price elasticities between commodities i and k in region r;

pp(k,r) is the composite price of commodity i in region r, for private household consumption;

EY(i,r) is the income elasticity for commodity i, in region r;

yp(r) is private consumption expenditure in region r;

dpp_agg(a,r) is the distribution parameter for essentials (/non-essentials) in region r;

dppall_agg(r) is the average of dpp_agg(a,r);

dpp_stop(k,r) is the distribution parameter for commodity i, in region r; and

dppall_stop(r) is the average of dpp_stop(i,r).

We allow for two changes in willingness to pay, the first is related to our assumption that demand for essentials is fixed (dpp_agg(MAPTC2TCA(k),r)). The second is commodity specific (dpp_stop(k,r)) and is used for avoidance and later health care costs and pent-up demand. For more details on incorporating willingness to pay (distribution parameters) into a CGE model (see Walmsley and Minor (2020b) and Dixon and Rimmer (2002)). Private consumption of commodity i is then allocated across domestic and foreign sources (Equation (3)) and foreign sources (Equation (4)) using a series of nested Armington Constant Elasticity of Substitution (CES) functions that characterize imports as imperfect substitutes for domestic goods. At the first level:

$$Q_i={\left({D_i}^{-\rho }+{M_i}^{-\rho }\right)}^{-\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\rho $}\right.}$$ (3)

Where: $Q_i$ is demand for commodity i by private final consumers;

$D_i$ is demand for the domestic commodity i by private final consumers;

$M_i$ is demand for the imported commodity i by private final consumers; and

$\rho$ is the CES substitution parameter (which also differs by commodity i).

At the second level:

$$M_i={\left(\sum _r{M_{i,r}}^{-{\rho }_M}\right)}^{-\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{${\rho }_M$}\right.}$$ (4)

Where: $M_{i,r}$ is demand for the commodity i imported from region r by final consumers14 ; and

${\rho }_M$ is the CES substitution parameter (which also differs by commodity i).

Given the demand for their goods by domestic and foreign agents, firms then produce these goods by combining intermediate inputs and value-added according to a Leontief (fixed proportion) relationship.15 Intermediate demand for commodity i is then sourced domestically or imported, and then allocated across foreign sources using a series of CES functions similar to equations (3), (4). We also add distribution parameters to intermediate demand and government consumption to incorporate avoidance, increased demand for health care, pent-up demand and even fiscal policy, using the techniques outlined in Dixon and Rimmer (2002) and Walmsley and Minor (2020b).

Demand for factors of production is determined using a CES production relationship of value-added.

$${VA}_i={\left(\sum _f{F_{e,i}}^{-{\rho }_E}\right)}^{-\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{${\rho }_E$}\right.}$$ (5)

Where: ${VA}_i$ is demand for value-added in production of commodity i;

$F_{e,i}$ is demand for factor e in production of commodity i; and

${\rho }_E$ is the CES substitution parameter (which also differs by commodity i).

With the mandatory closure of non-essential businesses, production declines. This is achieved through an endogenous “phantom tax,“16 which raises prices, leading consumers and essential businesses to either reduce demand for these goods or purchase these goods from other (foreign) sources, where mandatory closures are less severe (equations (3), (4))). One of the main benefits of a general equilibrium model is its focus on the allocation of scarce factors of production across sectors. First, demand for factors (determined by Equation (5)) must equal factor supply (${FS}_e$)17

$${FS}_e=\sum _i{F}_{e,i}$$ (6)

The supply of factors (${FS}_e$) depends on exogenous changes in the supply of factors (${FX}_e$), such as deaths and illness caused by the pandemic, and the rate of unemployment (${URate}_e$):

$${FS}_e={FX}_e.\left(1-{URate}_e\right)$$ (7)

The unemployment rate is determined endogenously through an inequality. If changes in demand and supply of labor cause wages to rise, then the equilibrium wage $W_e$ rises, and unemployment is permitted to gradually fall back towards the natural rate; otherwise, wages remain unchanged (${\hat{W}}_e=0$, where ${\hat{W}}_e$ is the proportionate change in the equilibrium wage $W_e$) and unemployment increases endogenously.18 This approach has been added to the model and is based on an approach documented in Walmsley and Strutt (2011) and also similar to that used in Dixon and Rimmer (2011). This means that when businesses close during the pandemic and demand for labor falls, wages are assumed to be sticky downwards, and hence any decline in demand for labor causes increased unemployment. In later periods, when businesses reopen and demand for labor increases again, wages rise and unemployment falls. We also allow for the unemployment/under-utilization of other resources, such as capital and land. As production falls, capital and land lay idle until businesses reopen and these resources are re-deployed.

The pandemic has had a profound impact on both the supply and demand of factors of production. Deaths, illness, and the avoidance of work have led to an exogenous reduction in the supply of workers (${FSX}_e$), while mandatory closures and avoidance of goods and services, as well as the general decline in production, have reduced demand for all factors of production ($F_{e,i}$). The general equilibrium model is therefore an ideal tool for analyzing the economic impact of the pandemic.

The model is also dynamic, allowing us to examine the impact of COVID over eight semi-annual periods (2020_1, 2020_2, …, 2023_2). The dynamic model also incorporates investment behavior through adaptive expectations that allow for the gradual equalization of global rates of return over time, and additional accounting relations to keep track of foreign ownership of capital based on the GDyn model (Ianchovichina and Walmsley, 2012), although many of these features are not important since we are examining the short run impacts of the pandemic.

Since the model is dynamic, a baseline scenario of how the world economy was expected to change without the pandemic must be established. To build this baseline scenario, we use historical and pre-COVID forecasts from the Bureau of Economic Analysis for the U.S. and the International Monetary Fund's World Economic Outlook (IMF, 2019) for real gross domestic product (GDP), investment, and labor; United Nations (UN, 2017) data on population growth to update the 2014 GTAP data to the beginning of 2020 and create a baseline for 2020_1 to 2023_2 (in semi-annual periods). The baseline scenario between 2020_1 and 2023_2 reflects what the global economy would have looked like if the pandemic had not occurred.

See Appendix D for explanations of how the model was applied to each of the causal factors.

7. Results

The model described in Section 6 estimates the macroeconomic impacts of the COVID-19 pandemic on the U.S. economy based on the “bottom-up” economic consequence analysis framework presented in Section 1. This enables us to isolate the effect of individual causal factors (“drivers”) of these consequences.

Fig. 2 presents the time-path of the estimated U.S. GDP over the eight semi-annual periods between 2020 and 2023 (i.e., cumulative GDP growth in six-month periods) in the baseline and with the pandemic.19 The corresponding estimates in billions of dollars of GDP in the baseline during the pandemic, and the cumulative difference between the baseline and pandemic (in both percentage and dollars terms) are given in Table 4. In the first six months of the pandemic, mandatory closures, avoidance, and deaths and illness caused a significant decline in the semi-annual real GDP of approximately 39%. With the introduction of the first four rounds of fiscal stimulus, increased healthcare expenditures, and increased (pent-up demand) consumer spending as businesses were allowed to re-open, this decline was reduced to the total decline of 27.6% shown in Fig. 2. Growth rose in late 2020 as businesses continued to reopen, but dipped again slightly in early 2021 with the re-introduction of some mandatory business closures and further avoidance in response to another wave. Over time, the reopening of businesses, the decline of avoidance behaviors, and pent-up demand have resulted in a gradual crawl back toward baseline growth. The second dip in the first half of 2022 corresponds to the later stages of Round 6 of the fiscal stimulus, which is found to crowd out investment and exports, leading to a decline in growth. But in the second half of 2022, the gap starts to diminish again, and by the end of 2023, the semi-annual gap between economic growth under the pandemic and baseline growth is just under $800 billion.

Fig. 2.

Percent Changes in Semi-annual GDP by Causal Factor. (cumulative percent change over time).

Table 4.

Impacts on Semi-annual U.S.GDP (billions of dollars, unless otherwise specified).

	2020_1	2020_2	2021_1	2021_2	2022_1	2022_2	2023_1	2023_2
Difference (% change)	−27.5	−21.6	−22.3	−12.4	−13.7	−10.8	−8.8	−6.5
Baseline	10,909	11,104	11,292	11,480	11,645	11,810	11,965	12,120
Policy	7898	8705	8774	10,057	10,050	10,534	10,912	11,332
Difference in Semi-annual GDP	−3011	−2398	−2518	−1424	−1595	−1275	−1053	−788

Table 5 presents the decomposition of the gap between the pandemic outcome and the baseline into the various contributing factors. The largest contributing factor is mandatory business closures, followed by avoidance. The mandatory business closures were most prevalent in the first half of 2020, with restrictions partly lifted in the second half. Additional closures in the first half of 2021 caused a further slowing of the recovery. In addition to mandatory closures, the avoidance of in-person shopping for goods and services and of the workplace also contributed significantly to the decline in GDP. The decline in labor due to illness and deaths of workers was relatively small due to the low percentage of workers impacted and the availability of other workers due to the high levels of unemployment. The small decline in real GDP here is due to the decline in population and demand rather than the decline in the labor force.

The remaining factors tended to alleviate the losses in economic growth. Increased demand on the health care sector to treat COVID patients had positive implications for other sectors, albeit small. Pent-up demand was found to be an important factor in the recovery, raising real GDP in the first half of 2020 by 3% and by a further 2% in the first half of 2021, which then continued to boost growth due to a rise in investment. Pent-up demand in the second half of 2020 was found to be relatively small but increased again in early 2021. It is worth noting that we only capture pent-up demand in 2020 and the first part of 2021, although it is likely that there is further pent-up demand in the latter part of 2021 and perhaps even into 2022. Hence, this is likely to be an underestimate of the longer-run impact of all pent-up demand.

The fiscal stimulus was also found to positively impact the economy, particularly when first implemented. Rounds 1–4 had the greatest impact, alleviating the decline in growth at its most severe stages. Round 5 provided further relief in 2021, while round 6 was found not to be beneficial for economic growth, in large part because it was becoming harder to fund the fiscal stimulus using domestic savings.20

With the decline in production due to mandatory business closures and avoidance, employment of all factors declined, with the largest declines due to the mandatory closure of businesses (Table 6 ). The decline in labor is particularly large, with unemployment rising slightly more than GDP in percentage terms. This is because non-essential businesses, and services, in particular, tend to be more labor-intensive and hence have a greater impact on labor than other factors of production, such as capital.

With deaths and illness, the supply of labor declines. Those left unemployed as a result of the mandatory closures and avoidance fill these vacant positions, causing unemployment to fall slightly. Similarly, with avoidance, there is also a decline in labor supply as some employees avoid going to work for fear of unsafe working conditions, or the need to stay home to take care of children or household members in need of care. Overall, however, unemployment still rises with avoidance as the decline in production and demand for labor resulting from the avoidance of non-essential goods and services is greater than the fall in the supply of labor due to avoidance of work. The fall in labor supply due to deaths, illnesses, and avoidance of work also leads to a shift towards capital that has implications for investment, discussed below. Increased health care expenditures, pent-up demand, and Rounds 1–5 of the fiscal stimulus also raise the demand for labor.

The impact on investment (Table 7 ) is generally consistent with the impacts on real GDP. There are a few notable exceptions: avoidance and fiscal stimulus (round 6). Avoidance by workers raises investment because the decrease in the supply of labor raises the use of capital (e.g., increased automation). Increased demand for capital increases returns to capital and hence investment. Investment then adds to capital stocks over time, and so increases in investment increase growth in real GDP in the following period. Similarly, declines in investment result in a decline in growth in real GDP in the following period.

Table 7.

Percent Changes in Semi-annual U.S. Investment by Causal Effect (cumulative percent differences from baseline).

	2020_1	2020_2	2021_1	2021_2	2022_1	2022_2	2023_1	2023_2
Mandatory Closures	−22.4	−26.7	−25.9	−16.2	−9.1	−6.5	−4.4	−2.9
Avoidance	1.3	0.1	−6.5	−7.7	−5.4	−4.6	−4.0	−3.5
Deaths and Illnesses	0.1	2.4	3.9	3.5	1.7	2.7	3.1	3.2
Health Care	−0.1	−0.6	−1.5	−1.0	−0.7	−0.6	−0.4	−0.3
Pent-up Demand	−0.7	0.8	2.4	1.2	6.7	10.1	13.9	18.3
Fiscal Policy
Rounds 1-4	5.1	5.0	6.0	8.5	12.4	12.9	13.4	13.4
Round 5	0.0	0.0	1.4	1.1	1.6	0.9	0.6	0.2
Round 6	0.0	0.0	−0.1	−6.9	−1.1	−1.4	−1.4	−1.5
Totals	−16.8	−19.1	−20.3	−17.6	6.2	13.6	20.7	26.8

The different impacts of the various rounds of the fiscal stimulus on investment reinforce the results in real GDP from the various fiscal stimulus rounds. Rounds 1–4 cause investment to increase, while in Round 6 investment decreases, resulting in declining growth in real GDP in 2022 and beyond. This decrease in investment in Round 6 is caused by government spending crowding it out. This crowding out is exacerbated in Round 6 because much of its spending is implemented in late 2021, just as savings rates are returning to normal. In 2020, as Rounds 1–4 of the fiscal stimulus are implemented, savings rates rise as people could not spend their money due to business closures and avoidance. This rise in savings provided the government with sufficient funds to stimulate the economy without significant crowding out of exports and investment. By late 2021, however, businesses are projected to be open, and savings rates have fallen; savings are therefore scarcer, and foreign savings must be used to fund Round 6 of the fiscal stimulus, which has implications for the trade balance and trade (see Walmsley et al., 2021).

Appendix Figure E−1 shows the impact of the various causal factors on sectoral production in the first half of 2020. As expected, mandatory closure of businesses causes most of the declines in production, followed by avoidance. These declines are most significant for the non-essential sectors, although there are clear indirect implications for essential sectors too, as incomes fall and non-essential businesses reduce their demand for intermediate inputs (e.g., pharmaceuticals used in health care). The impact of avoidance behavior in accommodation and food services, recreational services, transportation, and health care on production is also clear. We also see a reflection of some pent-up demand in the first half of 2020, as people eager to enjoy restaurants increase demand when businesses resume operation. Rounds 1–4 of the fiscal stimulus (Rounds 5 and 6 are not implemented until 2021) have a clear, positive impact on most sectors.

8. Discussion

8.1. Impact of alternative scenarios

In Table 8 , we show how changes in our assumptions about the number of deaths and illnesses, and the resulting impact on the health sector, impact the economy. These are compared with Scenario 1. The results show that increased efficacy of interventions (Scenario 2) reduces the impact of the deaths and illnesses and health care expenses on the economy substantially. In Scenarios 3a and 3b, lower efficacy means that the impacts are larger in absolute terms, with deaths and illnesses causing further declines in real GDP for Scenarios 1 and 2 and the increase in health care expenditures raising real GDP sizably. The extent of the changes depends on the level of social distancing assumed in Scenarios 3a and 3b – with more social distancing in the latter resulting in fewer cases and hence a smaller absolute economic impact.21

Table 8.

Percent Changes in Semi-annual Real GDP Due to Deaths and Illnesses and Health Care Expenses under the Alternative Scenarios, (cumulative percent differences from baseline).

Scenario	2020_1	2020_2	2021_1	2021_2	2022_1	2022_2	2023_1	2023_2
Scenario 1
Deaths and Illnesses	0.0	−0.1	−0.3	−0.5	−0.6	−0.7	−0.8	−0.8
Health care expenses	0.8	0.9	0.8	0.7	0.6	0.5	0.5	0.4
Scenario 2
Deaths and Illnesses	0.0	0.0	−0.1	−0.1	−0.2	−0.2	−0.2	−0.2
Health care expenses	0.2	0.1	0.2	0.2	0.2	0.2	0.2	0.2
Scenario 3a
Deaths and Illnesses	0.0	−0.3	−0.8	−1.0	−1.2	−1.4	−1.5	−1.6
Health care expenses	7.4	5.3	2.9	2.1	1.9	1.9	2.0	2.0
Scenario 3b
Deaths and Illnesses	0.0	−0.2	−0.7	−0.9	−1.0	−1.1	−1.2	−1.3
Health care expenses	5.8	4.9	3.1	2.5	2.2	2.0	2.0	1.9

8.2. Comparison with results of other studies

Few studies have performed a decomposition of the economy-wide impacts of COVID-19 according to several causal factors, such as those provided in this report. One of the reasons is that many of the estimates utilize macro-econometric or time-series models, which are typically “top-down” approaches and hence less amenable to decomposition in comparison to “bottom-up” models such as ours. However, there are a few studies that lend themselves to comparison. The first set of studies, including one by some of the authors of this paper, has focused on mandatory closures and generally found the impacts on GDP to be in the range of 20–25% for closure scenarios similar to those that took place (see del Rio-Chanona et al., 2020; Minor et al., 2015; Office of the Texas Governor, 2021; Walmsley et al., 2021a).22 Dixon et al. (2020), using a quarterly dynamic CGE model, explored the macroeconomic impacts of COVID-19 concerning several drivers over a two-year time horizon. They estimated a 19% reduction in GDP at the trough of the economic downturn at the end of the first quarter of 2020 and a 12% decline by the end of the second quarter. They incorporated our estimates of telework and also included government expenditures on health care and some countervailing fiscal policies, such as unemployment compensation and tax relief, all of which dampen the negative impacts. CBO (2021b) estimates that real GDP would decline by 11% in the second quarter of 2020, resulting in the number of people employed being almost 26 million lower than the number in the fourth quarter of 2019. If this rate were to continue, the decline in U.S. real GDP for the year was projected to be up to 38% on an annual basis; however, given the reopenings, the overall annual decline was projected to drop to 5.4%.

8.3. Limitations

This is a short-run study of the impact of the pandemic over the period of 2020 to 2023 using data collected or estimated for that period. While we partially examine the impact of higher and lower efficacy of intervention scenarios on the labor force and health care, we do not consider the potential ramifications of these or other alternative scenarios on mandatory closures and avoidance. In addition, we do not take into account the impact of a long-term decline in labor due to longer-term illness, or “long COVID,” or the potential for behavior changes, such as teleworking and avoidance of certain activities, to become more permanent. Nor do we include a measure of the total value of lives lost.

Our results should be considered upper-bound estimates for several reasons. First, we have assumed that reductions in business output are accompanied by reductions in wages and salaries paid as people become unemployed, though some businesses continued paying their employees. Second, we have omitted some sources of business resilience, such as the use of inventories, relocation (e.g., haircuts in parking lots), and the internet (e.g., education) to continue to help produce goods and services (Rose, 2017; Dormady et al., 2019). We also do not consider the increase in demand for communications due to the increased use of the internet and remote working. Finally, we only include avoidance and pent-up demand in 2020 and 2021. The extent to which avoidance may continue is uncertain, but further increases in pent-up demand spending that are likely to occur in the latter part of 2021 and into 2022 are not considered.

We also acknowledge the limitations of our model and its application, as well as of our assessment on how they bear on the results. As is the case in most GTAP-based analyses, we assume that, except for factor markets, all other markets clear and firms are perfectly competitive. Factors of production are also assumed to be immobile across sectors within six months, consistent with this paper focusing on the very short-run impacts of the pandemic. Our results also rely on the estimated elasticities taken from the literature and used in the GTAP database. Although we calculated the expected changes in government savings (deficit) and private household savings separately for our decomposition, for modeling purposes these changes had to be aggregated into a single change in domestic savings, as the model does not separately identify government savings (deficit) and private household savings.

9. Conclusion

We refined a dynamic CGE model to estimate the economic consequences of COVID-19 over eight semi-annual periods during 2020–2023. A baseline scenario in which the COVID pandemic is assumed not to occur was first established based on pre-pandemic growth forecasts. The COVID pandemic was then incorporated into the model and compared to baseline growth. The various causal factors of the pandemic were added sequentially to decompose their contributions to the overall impact of COVID on the U.S. economy. These causal factors include the mandatory closure of businesses and gradual reopenings, the avoidance of workplace and various activities (such as restaurant dining), the impact of deaths and illness on the labor force, the increase in hospitalizations and health care expenses, the fiscal stimulus implemented in 2020 and 2021, and the increase in pent-up demand once businesses were allowed to reopen. Sensitivity analysis was performed on disease spread with various interventions relating to vaccine availability, efficacy, and take-up. The decomposition of the influence of various causal factors will help policymakers make adjustments to offset the negative influences and reinforce the positive ones during the remainder of this pandemic and in anticipation of future ones.

The analysis of the decomposition revealed that the largest losses from COVID were associated with the mandatory closure of businesses and the slow reopening process, followed by the avoidance of workplace and other activities by households. While deaths and illnesses resulted in a minimal decline in real GDP, primarily due to the decline in demand caused by the declining population, the increase in demand for health care led to a rise in real GDP. Pent-up demand is a significant factor in the recovery process, raising growth ever closer to the original baseline growth. Early rounds (1–4) of fiscal policy were also very helpful in alleviating some of the losses in economic growth due to mandatory business closures, avoidance, and other causal factors. The benefits of the last round of fiscal policy are considerably lower, and even negative, compared to earlier rounds due to crowding out of private investment and the need for businesses to repay loans.

Several areas of future research present themselves. A prime example would be developing future scenarios involving new variants, more effective vaccines, and more effective anti-viral treatments. Another would be conducting surveys to ascertain the long-term impact of avoidance and changes in the way people work. Survey research can also be extended to improve the accuracy of estimates of pent-up demand by major consumption categories. An additional topic area would be to examine the macroeconomic consequences of supply-chain bottlenecks caused by the pandemic.

Handling editor: Angus Chu

Appendix A. Health Considerations – Appendix to Section 3

A1. Summary of Expert Interviews

We conducted 12 interviews with experts from variable backgrounds including scientists in the field of public health and epidemiology from the United States, Canada, and the United Kingdom; health officials from state- and federal-level government institutions in the United States; and clinicians in the United States. The convenience sample was assembled from experts who work in academic, public, and clinical settings, and who have published work on COVID-19. Interviews focused on specific issues of interest where literature (at the time of modeling) was limited, including the effectiveness of social distancing, the likely behavioral response of the public, and the ability of the health care system to handle excessive numbers of COVID-19 patients during pandemic peaks.

Due to the evolving nature of the COVID-19 pandemic and the limited availability of strong data regarding intervention effectiveness, experts were consulted both on the parameters used in our model and on possible additions that should be accounted for. Two parameter changes we made on expert advice are of note.

Multiple interviewees suggested that we account for the effects of seasonality on the effective reproductive number. To account for this variable, we implemented monthly adjustments to the reproductive number based on Gavenčiak et al. (2021).

It was noted that some interventions, including stay-at-home orders, experienced differing compliance between age groups (CDC, 2020a). While more granular modeling is possible, the dichotomous distinction between the population under 65 and the population 65 and older has been sufficient for the outcome of interest in the employed model. Parameters including hospitalization rate, infection fatality rate, and vaccine hesitancy were age-stratified.

A2. Estimating the Number of Infected International Travelers

COVID-19 was initially introduced to the U.S. through international travel (CDC, 2020c). Our model estimates the number of international travelers entering the United States on any given day from 6 regions – the Americas, Africa, Europe, Middle East, Western Pacific, and South and Southeast Asia. Our model utilizes historical estimates for the percentage of travelers from each region and predicts each traveler's likelihood of carrying COVID-19 to determine the likely number of COVID-19 cases entering the country.

International travel trends for the 2.5 years included in the model are adjusted based on domestic travel reduction trends in 2019 and 2020 (Bureau of Transportation Statistics, 2021; International Trade Administration, 2021), referred to as the coefficient reduction due to travel avoidance (we abstract from international travel restrictions, which affected the numbers of international travelers unevenly). The incoming travelers are then split into two age groups: travelers under 65 and travelers over 65 based on historical estimates of traveler demographics (it is, however, possible that older travelers reduced international travel to a higher degree than younger ones).

We draw on I-94 statistics (International Trade Administration, 2021) and develop a model of infection risk as well as an adjustment for lower travel volumes due to avoidance behavior. We utilize domestic traveler data from 2019 to 2020 from the Bureau of Transportation Statistics to determine how many fewer international travelers would have come to the United States, had there not been travel restrictions in other countries (their effects were not modeled explicitly due to their time-varying nature and limited effect given the magnitude of domestic infection counts). Using these data, we determine a coefficient reduction (as a fraction of 2020 travel relative to 2019). For 2021 and 2022, we estimate a stepwise growth in international travel based on estimates by the U.S. Travel Association (2021). The calculation of coefficient reductions for each month between January 2020–June 2022 is shown in Table A1.

Table A1.

Estimating International Traveler Volumes due to the COVID-19 Pandemic

Month	2019 International Travelers (total)	2019 Domestic Travelers (total)	2020 Domestic Travelers (total)	Coefficient Reduction (19 vs 20)	Coefficient Reduction (21 vs 19)	Coefficient Reduction (22 vs 19)
Jan	5,836,439	58,033,637	61,610,519	106.2%	46.5%	55.5%
Feb	5,091,485	55,679,100	59,849,890	107.5%	47.0%	56.0%
Mar	6,258,503	70,234,011	34,410,743	49.0%	47.5%	56.6%
Apr	7,044,707	66,938,177	2,877,134	4.3%	48.0%	57.2%
May	6,703,974	71,365,029	8,238,474	11.5%	48.5%	57.8%
Jun	6,326,970	72,789,897	16,207,341	22.3%	48.9%	58.4%
Jul	7,702,922	75,281,255	22,878,828	30.4%	49.4%
Aug	8,114,715	72,715,907	23,910,297	32.9%	49.9%
Sep	6,703,451	63,979,337	23,854,435	37.3%	50.4%
Oct	6,687,063	69,922,300	28,026,462	40.1%	50.9%
Nov	6,095,118	64,816,897	26,293,043	40.6%	51.4%
Dec	6,876,248	69,718,719	27,263,496	39.1%	51.9%
Total	79,441,595	811,474,266	335,420,662	41.3%	49.2%	56.9%

Next, we divide countries of travel origin based on the WHO (2021) and U.N. World Tourism Organization's classification: Americas, Africa, East Asia and Pacific, Europe, Middle East, and South and Southeast Asia. Using data from the (WHO, 2021), the percentage of individuals infected with COVID-19 within each region per month was estimated, which allowed us to calculate the likelihood of a traveler from that region being infectious upon entry to the U.S.

As Figure A-1 shows, Europe had the largest risk of infection among travelers to the United States for most of 2020 while Africa had the least. For periods when data were not yet available (starting in April 2021 at the time of modeling, shaded in the figure), we estimated 10% monthly decrements of a moving average from the previous 6-month period, resulting in a near-complete elimination of the risk of infection by June 2022 (0.03% for a traveler from the Americas, 0.02% from South (East) Asia, and 0.01% or less for travelers from the Middle East, Europe, and East Asia/Pacific.

Fig. A 1.

Probability of Infection by Region, by Month.

Following this, we combine these estimates to determine the likely number of infected travelers by region entering the United States on any given day. Initially, the most infected travelers came from Europe, similar to conclusions by Gonzalez-Reiche et al. (2020), who suggest that most index cases of the March 2020 New York State outbreak originated in Europe. Figure A-2 shows the estimated numbers of daily infectious travelers entering the United States by region of origin.

Figure A-2.

Estimates of Infected International Travelers Entering the U.S.(Daily Average by Month).

A3. Vaccination projections

Vaccination projections for the period starting at the beginning of the model (January 1, 2020) and ending on August 12, 2021 were drawn from CDC vaccination data and based on a delay period after vaccination, vaccine effectiveness, and an assumption on the number of vaccinated individuals who had already been conferred immunity through previous infection.

Vaccine effectiveness $\left(VE\right)$ was assumed as 90% based on a CDC real-world analysis of the Pfizer and Moderna mRNA vaccines (CDC, 2021d); the effectiveness of the Johnson & Johnson vaccine was not incorporated due to a lack of data on the exact breakdown of administered vaccines by day and the assumption that the vaccine was not prevalently used (Follow et al., 2021). We accounted for individuals being vaccinated who have already conferred immunity from a previous infection by weighting the number of vaccinations $\left(v\right)$ given by the proportion of the population susceptible to disease (S) out of all initially susceptible (S₀), as shown in the equation below.

$$S\to Effectivelyvaccinated=v*VE*\frac{S}{S_0}$$

From August 13, 2021 until the end of the model (July 31, 2022), we halted data scraping from the CDC and stipulated that the number of individuals transitioning from susceptible to vaccinated on a given day was equal to the average of model days August 6, 2021, through August 12, 2021 (a 7-day average) until the population that had undergone vaccination (including those determined to have had an “ineffective” vaccination regimen) was equal to the population expected to be vaccinated (Kapteyn and Gutsche, 2021).

A4. Reduction in R₀ in each scenario

In Figures B-3a/b/c/d, we show the percentage reduction in R₀ for each week of 2020 and months of 2021/2022 modeled in the four scenarios.

Fig. A 3a.

Scenario 1 - Percent Reductions in R₀ for 2020–22.

Fig. A 3b.

Scenario 2 - Percent Reductions in R₀ for 2020–22.

Percentage reduction estimates in scenario 2 were calibrated to approximate deaths due to influenza in the United States over a comparable period. We use CDC data (CDC, 2020b) to calculate the mean number of influenza deaths over 2.5 years:

Table A2.

Deaths Caused by Influenza from 2015 to 2020

Year	Deaths <65	Deaths 65+	Total
2015–2016	5248	17,458	22,706
2016–2017	5396	32,833	38,229
2017–2018	10,197	50,903	61,100
2018–2019	8603	25,555	34,158
2019–2020	8236	13,673	21,909
Means (per year)	7536	28,084	35,620
Mean estimate over 2.5 years	18,840	70,211	89,051

Fig. A3c.

Scenario 3a - Percent Reductions in R₀ for 2020–22.

Fig. A 3d.

Scenario 3b - Percent Reductions in R₀ for 2020–22. .

A5. Viral Seasonality

Liu et al. (2021) showed that the transmission rate of COVID-19 was time-dependent and cyclical, with reductions during the warm season in the Northern Hemisphere of 46.38 ± 29.10%. We draw on seasonality adjustments estimated by Gavenčiak et al. (2021), calculated for 143 European regions, and assume the reductions apply similarly in the United States.

A6. Estimates of Hospitalizations and Deaths by Scenario, Period, and Age

Table A3.

Age-stratified Estimated Hospitalizations due to COVID-19 in each 6 Months (in millions)

	Hospitalizations <65				Hospitalizations 65+
Scenario:	1	2	3a	3b	1	2	3a	3b
H1/2020	863.8	187.9	5963.7	3746.8	714.0	128.9	7546.6	3671.5
H2/2020	1493.3	64.0	561.6	1528.8	1399.3	51.5	1679.9	3906.4
H1/2021	1352.7	112.5	10.3	45.9	958.6	63.2	8.6	141.7
H2/2021	133.9	49.2	7.8	11.0	364.9	19.5	5.7	7.7
H2/2022	99.1	158.3	2.8	6.0	1559.3	29.9	2.2	3.9
Total	3942.8	571.9	6546.1	5338.5	4996.1	293.1	9243.0	7731.2

Note: Sums of semi-annual periods may not equal total due to rounding. These are model estimates.

Table A4.

Age-stratified Estimated Deaths from COVID-19 in each 6 Months (In millions)

	Deaths <65				Deaths 65+
Scenario:	1	2	3a	3b	1	2	3a	3b
H1/2020	28.3	6.2	228.4	128.3	109.7	19.8	1397.8	599.4
H2/2020	48.2	2.2	24.1	56.4	209.8	8.2	339.5	692.6
H1/2021	47.0	3.7	0.3	1.6	158.5	9.9	1.4	23.7
H2/2021	4.7	1.7	0.3	0.4	54.4	3.1	0.9	1.2
H1/2022	3.4	5.3	1.0	.204	245.6	4.7	0.4	0.6
Total	131.6	19.1	253.2	186.8	778.0	45.6	1739.9	1317.5

Note: Sums of semi-annual periods may not equal total due to rounding.

We indicate how cumulative deaths differ by scenario and between the age-stratified groups in Figure A-4.

Fig. A4.

Graphs of Cumulative Predicted Deaths for the 4 Scenarios.

Appendix B. Economic Data – Appendix to Section 4

Appendix Table B-1 presents the estimated direct percentage reduction in U.S. annual GDP by sector due to mandatory closures and the phased-in reopening process by 6-month period. We defined three categories of sectors based on the DHS Cybersecurity and Infrastructure Security Agency list of Essential Critical Infrastructure Sectors during COVID-19 (CISA, 2020): Category 1 includes sectors that fall entirely under the non-essential category and thus were shut down under the mandatory closures (such as Recreation & Entertainment); Category 2 includes sectors within which only some of their subsectors are non-essential (such as Retail Trade and Business Services); Category 3 includes sectors that are essential and were, therefore, able to maintain operation to the extent possible.

Table B1.

Percentage Reduction of Output by Sector due to Mandatory Closure and Reopening Process (with Telecommuting)

#	Sector	Mandatory Closure Categorya	% Reduction in U.S. annual GDP by sector due to mandatory closures after factoring in telework		% Reduction in U.S. annual GDP by sector due to phased-in reopening after factoring in telework
			2020_1	2020_2	2020_1	2020_2	2021_1
1–25	Agriculture, Fishing, and Forestry; Mining; Food Processing	3	0.00%	0.00%	0.00%	0.00%	0.00%
26	Beverages and Tobacco products	2	2.90%	0.00%	1.30%	0.00%	0.00%
27	Manufacture of textiles	1	10.60%	0.00%	5.70%	0.01%	0.00%
28	Manufacture of wearing apparel	1	14.10%	0.00%	5.60%	0.02%	0.00%
29	Manufacture of leather and related products	1	14.10%	0.00%	5.60%	0.02%	0.00%
30	Lumber	3	0.00%	0.00%	0.00%	0.00%	0.00%
31–34	Paper and Paper Products; Petroleum and Coke Products; Manufacture of chemicals and products; Manufacture of pharmaceuticals, medicinal chemical, and botanical products	3	0.00%	0.00%	0.00%	0.00%	0.00%
35	Manufacture of rubber and plastic products	1	11.50%	0.02%	5.50%	0.02%	0.00%
36	Manufacture of other non-metallic mineral products	1	12.00%	0.01%	5.50%	0.02%	0.00%
37	Iron & Steel: basic production and casting	3	0.00%	0.00%	0.00%	0.00%	0.00%
38, 39	Non-Ferrous Metals; Manufacture of fabricated metal products	3	0.00%	0.00%	0.00%	0.00%	0.00%
40	Manufacture of computer, electronic and optical products	1	13.80%	0.01%	5.70%	0.01%	0.00%
41–44	Manufacture of electrical equipment, machinery and equipment, and transport equipment	3	0.00%	0.00%	0.00%	0.00%	0.00%
45–47	Other Manufacturing: includes furniture	1	12.70%	0.01%	5.50%	0.03%	0.00%
46–48	Electricity; Gas manufacture, distribution; Water supply; sewerage, waste management and remediation activities	3	0.00%	0.00%	0.00%	0.00%	0.00%
49	Construction	2	2.90%	0.00%	1.70%	0.01%	0.00%
50	Wholesale and retail trade; repair of motor vehicles and motorcycles	2	5.40%	0.01%	2.50%	0.01%	0.00%
51	Accommodation, Food and service activities	2	9.30%	0.02%	5.10%	7.44%	4.52%
52	Land transport and transport via pipelines	2	2.20%	0.00%	1.00%	0.15%	0.00%
53	Water transport	2	4.70%	0.00%	2.60%	0.28%	0.00%
54	Air transport	2	8.30%	0.01%	4.50%	0.51%	0.00%
55	Warehousing and support activities	3	0.00%	0.00%	0.00%	0.00%	0.00%
56	Information and communication	2	0.50%	0.00%	0.50%	0.00%	0.00%
57	Other Financial Intermediation: auxiliary activities but not insurance and pensions	2	1.40%	0.00%	1.40%	0.00%	0.00%
58	Insurance	3	0.00%	0.00%	0.00%	0.00%	0.00%
59	Real estate activities	1	7.50%	0.01%	7.70%	0.59%	0.00%
60	Other Business Services not elsewhere classified	2	5.10%	0.01%	5.10%	0.40%	0.00%
61	Recreation & Other Services	1	10.30%	0.01%	14.80%	23.32%	15.00%
62	Other Services (Government)	2	3.60%	0.01%	1.90%	0.13%	0.00%
63	Education	1	4.10%	0.01%	8.10%	10.06%	6.02%
64	Human health and social work	3	0.00%	0.00%	0.00%	0.00%	0.00%
65	Dwellings: imputed rents of owner-occupied dwellings	3	0.00%	0.00%	0.00%	0.00%	0.00%

^aMandatory Closure Categories: 1. Sector is entirely non-essential and thus is completely shut down; 2. Sector for which only some subsectors are non-essential (see notes in the last column); 3. Sector that is essential and thus still able to operate in its usual manner to the extent possible.

For the rest of the world, data were collected on the timings of mandatory closures in each country (Wikipedia, 2020), and the U.S. categorization of essential/non-essential sectors are applied. Where closures were considered partial (e.g., city- or region-wide only), we applied a 50% closure rate. Each country's production data were then used to determine the overall share of the sector closed in the rest of the world, assuming the same essential sector shares as in the U.S. case. The phased-in schedule of U.S. reopenings was also used to determine the timing of reopenings for the rest of the world, adjusted for differences in the mandatory closure periods between each country and the U.S. Although our focus is on the U.S. economy, it is important to capture the simultaneous slowdown of China and the rest of the world that occurred over this period, particularly in a global model. For this reason, we compared the macroeconomic results for China and the rest of the world with actual data to ensure they were consistent. However, we did not collect the same level of detailed information about China and the rest of the world, as our focus was on the U.S. and it is unlikely that these additional details would have had a significant impact on our results.

Appendix Table B-2 presents the average per-patient health expenses by health outcome category and by age group. The total estimated COVID-related health expenditures for each scenario are calculated by multiplying the number of outpatients, hospitalizations, and deaths by the corresponding per-patient cost, and are presented in Appendix Table B-3.

Table B2.

COVID per Patient Health Expenses by Treatment Category and by Age Group

Age Group	Hospital Cost		Outpatients (with mild symptoms)
	Non-ICU	ICU
0–19	$15,620	$55,542	$91
20–64	$15,430	$54,866	$57
65+	$16,758	$59,587	$96
Weighted Average	$16,131	$56,677	$67

Source: Calculated based on Bartsch et al. (2020) and Fusco et al. (2021).

Table B3.

COVID Total Health Expenditures (in millions of 2020 dollars)

Scenario	Outpatient	Hospitalizations - Non-ICU	Hospitalizations - ICU	COVID Health Expenses	% of Total Annual Output of GTAP Health Sector
Scenario 1	6211	125,952	82,265	214,427	8.35%
Scenario 2	789	11,312	7903	20,004	0.78%
Scenario 3a	10,559	212,972	142,086	365,617	14.23%
Scenario 3b	8662	176,103	115,678	300,443	11.70%

Appendix Tables B-4 and B-5 present the per-patient productivity losses (measured in days) and the total losses in workdays for each health scenario modeled in Section 3.

Table B4.

COVID per Patient Lost Productivity (in days)

Age Group	Hospital LOS		Productivity Days Loss		Outpatient
	Non-ICU	ICU	Non-ICU	ICU	Productivity Days Loss	Additional Days in Isolation
0–19	4.0	9.4	11.5	18.9	1.5	8.5
20–64	6.0	14.1	13.5	23.6	1.9	8.1
65+	7.0	16.6	14.5	26.1	5.3	4.7

Sources: Calculations based on Prager et al. (2017), Chen et al. (2020), Zhao et al. (2020), Fusco et al. (2021), and Walmsley et al. (2021).

Table B5.

Lost Productivity Due to Own COVID Illness and Caring for Sick Family Members from COVID (in thousands of days)

Scenario	Outpatient Medical Treatment	Hospitalizations Non-ICU	Hospitalizations ICU	Fatalities	Total	Percentage Change in Labor Force
Scenario 1	392,153	78,665	25,350	113,700	609,868	−0.59%
Scenario 2	48,507	7384	2514	8083	66,488	−0.07%
Scenario 3a	669,611	138,956	45,412	249,136	1,103,115	−1.03%
Scenario 3b	549,885	115,291	37,348	188,049	890,572	−0.83%

Appendix Table B-6 presents the percentage increase of consumption from the “Lowest Point” level for major categories of goods and services at five points in time with three-month intervals after the beginning of the reopening. The linkages to the CGE model sectors and the share of the sectoral output affected are also presented in the table.

Table B6.

Percentage Consumption Changes Compared to the “Lowest Level” after COVID by Commodity/Service Type and Mapping to CGE Sectors

Good/Service	2020.6	2020.9	2020.12	2021.3	2021.6	GTAP Sector	Share of GTAP Sector
Automobiles	79%	65%	20%	200%	200%	50 Wholesale and retail trade; repair of motor vehicles	5.5%
Real Estate	50%	60%	55%	50%		59 Real estate activities	100%
Air Travel	32%	37%	25%	72%	56%	54 Air transport	94.1%
Restaurant Dining	56%	55%	43%	119%	120%	51 Accommodation, Food and service activities	71.7%
Live Experiences (sporting events, concerts, etc.)	39%	35%	30%	65%	85%	61 Recreation & Other Services	16.5%
Apparel	72%	56%	66%	143%	148%	50 Wholesale and Retail Trade; Repair of Motor Vehicles	4.1%
General Merchandise	67%	22%	19%	158%	142%	50 Wholesale and Retail Trade; Repair of Motor Vehicles	6.8%
Hotels and other Hospitality	45%	49%	54%	79%	81%	51 Accommodation, Food and Service Activities	16.8%
Wellness and Fitness	54%	38%	12%	106%	96%	61 Recreation and Other Services	3.0%

Appendix C. Physical Stimulus Bills – Appendix to Section 4

C1. Summary of COVID-19 Stimulus Bills

Table C-1 summarizes the main provisions of each of the six stimulus bills the U.S. Congress enacted.

Table C1.

Summary of COVID-19 Stimulus Bills Enacted by the U.S. Federal Government

Stimulus Bill	Date Enacted	Main Provisions	Estimated Spending (billions, USD)a
Coronavirus Preparedness and Response Supplemental Appropriations Act, 2020	March 6, 2020	Funding for federal, state, and local health agencies, and the purchase of vaccines and treatments	8
Families First Coronavirus Response Act	March 18, 2020	Tax credits for paid medical leave, increased funding for nutritional assistance, and unemployment benefits	192
Coronavirus Aid, Relief, and Economic Security (CARES) Act	March 27, 2020	Cash payments for individuals, extra unemployment benefits, forgivable loans for small businesses, loans for medium and large businesses, and corporate tax relief	1902
Paycheck Protection Program and Health Care Enhancement Act	April 24, 2020	Extension of forgivable loan program for small businesses, reimbursements for health care providers, and funding for testing	342
Coronavirus Response and Relief Supplemental Appropriations Act, 2021 and Additional Coronavirus Response and Relief Act	December 27, 2020	Cash payments for individuals, extra unemployment benefits, and extension of the forgivable loan program for small businesses	864
American Rescue Plan Act of 2021	March 11, 2021	Cash payments for individuals, extra unemployment benefits, support for state and local governments	1825

^aEstimated spending is based on preliminary projections by the Congressional Budget Office (CBO, 2021a) and includes increases in direct spending as well as declines in tax and fee revenue.

The first bill was signed into law in early March 2020 and focused on enhancing the government's preparedness and public health response by providing funds to the Department of Health and Human Services (HHS) and state and local health agencies. Less than two weeks later, the U.S. Congress passed the Families First Coronavirus Response Act (CBO, 2020a). Like the first bill, this Act aimed to improve the public health response to the crisis, but it also sought to address the economic impact of the pandemic by enhancing safety net provisions. A paid sick leave provision, which required certain employers to pay the full or partial salary of employees unable to work due to COVID-19, was estimated to cost $105 billion (more than half the total cost of the bill) in tax credits.

As COVID-19 cases rose in mid-March, Congress enacted the Coronavirus Aid, Relief, and Economic Security (CARES) Act, enacted in late March, which amounted to more than $1.9 trillion in spending and tax credits (CBO, 2020b). The bill included $100 billion for hospitals and $50 billion for other public health provisions. Airline companies received $w50 billion for not laying off employees. The legislation also provided $340 billion for state and local governments, mostly directed to COVID-19 response efforts. Three other provisions amounted to approximately half of the bill's total cost: payments for individuals and dependents, enhanced unemployment insurance, and the Paycheck Protection Program (PPP), a forgivable loan scheme for small businesses. The cash payments provided up to $1200 per person making less than $99,000 a year, with an additional $500 per child. Unemployed individuals received supplemental weekly payments of $600 for 13 weeks. The PPP loans are fully forgivable if businesses use the funds to cover payroll or other eligible expenses and meet certain conditions such as not laying off workers – we assume that they will not be paid back in our simulations.

After the $350 billion allocated by the CARES Act to PPP ran out, Congress funded an extension in April 2020 (Duehren and Omeokwe, 2021). The Paycheck Protection Program and Health Care Enhancement Act appropriated another $310 billion for PPP loans. It also included another $75 billion for hospitals and $25 billion for testing. In our model, we group these initial four bills because they went into effect during a relatively short period, from March to April 2020.

Eight months later, after a second wave of COVID-19 infections in the Fall of 2020, the U.S. Congress approved another comprehensive stimulus bill of nearly $900 billion as part of the Appropriations Act of 2021 (Wall Street Journal, 2020). Many of its provisions were scaled-back extensions of CARES Act provisions: cash payments of up to $600 for individuals earning up to $87,000 a year and additional $600 per dependent; weekly unemployment supplement payments of $300 for 11 weeks; and a third extension of the PPP program.

In March 2021, following a surge of COVID-19 cases and deaths, Congress approved another major stimulus bill. Some of the major provisions extended support provided to households and unemployed individuals under the December 2020 package. These provisions included additional one-time cash payments of up to $1400 per adult and child for eligible households and an extension of boosted unemployment payments through September, additional funding for vaccines and testing, schools and universities, $350 billion to state and local governments, and an increase to the child tax credit (CBO, 2021b).

C2. Stimulus Implementation and Impact

Several issues with the implementation of the CARES Act delayed relief. These issues include confusion among borrowers and lenders about the PPP program and its eligibility criteria, and difficulties in delivering cash payments to nine million, hard-to-reach individuals (GAO, 2020). State governments were also overwhelmed and slow to process enhanced unemployment payments (Iacurci, 2021). Finally, the emergency programs administered by the Federal Reserve only provided $41.1 billion in assistance, a fraction of the announced size of those programs of $4 trillion (CRS, 2021). In the analysis below, we do not model the Federal Reserve loans directly because we assume they will be repaid and because the low uptake suggests the lending terms did not represent a subsidy relative to commercial credit.

Some economists have criticized the cash payments to individuals as not sufficiently targeted to those most in need, arguing that further increasing unemployment benefits or extending their duration would be a more efficient use of resources (Schwartz and Friedman, 2020). Coibion et al. (2020) find that those who received cash payments under the CARES Act on average planned to spend only 40% of their stimulus checks. Baker et al. (2020) similarly estimate that during the first weeks, payment recipients spent 25–40% of the stimulus, but that share was higher for households with lower or declining incomes, highlighting the importance of targeting. In our model, we assumed 7.8% of unemployment benefits and 36% of cash payments and other household payments were saved based on household surveys (Armantier et al., 2020; Coibion et al., 2020; Leer, 2021).

Another criticism of the stimulus provisions is that they had a limited capacity to restore economic activity for the most impacted sectors such as hospitality and entertainment, where consumer spending was constrained by mandatory shutdowns and avoidance (Chetty et al., 2020). We assumed that the stimulus provisions could not raise production in sectors constrained by mandatory closures.

Furthermore, the PPP loans appear to have had a limited effect on preserving and creating jobs, the program's main goal. The short- and medium-term employment effects of the Program were small, with many firms using the loans to make non-payroll payments and build up savings (Granja et al., 2020). Chetty et al. (2020) estimate that the initial round of PPP loans increased employment at small businesses by only 2%, at $377,000 per job saved. For sectors constrained by mandatory closures, we assumed that a portion of the loans—equivalent to the portion of the sector subject to mandatory closures—went directly to the owners or employees as income, allowing us to capture some of this effect.

The aforementioned studies suggest that stimulus provisions are not equally cost-effective, and that policy design is critical to meeting the specific objectives of policymakers. Targeting and timing are important considerations. Provisions that provide aid to the most affected individuals (e.g., enhanced unemployment benefits), and sectors (e.g., forgivable loans to airlines) may be more effective than other measures.

C3. Data Refinement

For most stimulus provisions, we assume the federal government spends the stimulus funds over time rather than at once. The CBO provides estimates of the direct cost of each provision, measured as an increase in spending or decline in revenue, yearly for the next 10 years. In the dynamic CGE model, the effects of the stimulus are estimated every six months. Annual CBO cost estimates are converted to semiannual figures for each provision by assuming an even distribution over a given year after accounting for the date that the related stimulus bill was passed. Some of the CBO cost estimates are implemented after 2023 and hence do not enter our simulations, but 98% of the fiscal stimulus spending is estimated to take place before then. Table C-2 shows how the funds were implemented into the model over time.

Funding for the PPP was assigned to various GTAP sectors based on loan disbursement breakdowns published by the Small Business Administration (SBA, 2020a). The initial $342 billion provided by the CARES Act for PPP was apportioned to corresponding GTAP sectors based on loan approvals through April 16, 2020, when the initial PPP funds had run out, but before the U.S. Congress had approved additional funds.23

The following assumptions are adopted when simulating the impacts of loans to businesses. The method used in the CGE simulations depends on the assumption of whether the sector/firm receiving the forgivable loans is open or closed for business. The distribution of loans is assumed to take place through the following two mechanisms:

• •Option 1: As a direct payment to households when the intention is to give money directly to the workers or owners to survive a period when their business or place of work is closed. This is simulated as increased payments to households in the CGE modeling.
• •Option 2: As a subsidy to businesses when the intention is to help the business stay open and keep producing, despite issues with sick workers and transition to telework. This is simulated as subsidies split between capital and labor using value-added shares for each sector in the CGE modeling or where information on who the loans were intended for is available (e.g., workers or owners), this information was used to allocate the subsidy.24

Table C-2 depicts how the amounts for each provision over time for each of the 6 rounds were incorporated into the model as three rounds (1–4, 5, and 6) and for five major categories (Unemployment assistance; Assistance to households (direct payments and assistance); Government spending (Federal, State, and Local); Assistance to businesses (e.g., loans) and Corporate tax relief). Not all rounds included all categories and increases in foreign aid and unallocated funds were not modeled.

Table C2.

Total Spending Amount by Round and Aggregate Category Modeled over Time

Round	Details	2020	2021	2022	2023	Post-2023 (or not captured)	Total
1–4	Unemployment Assistance	222.0	49.9	0.0	0.0	0.0	272.0
1–4	Direct Payments to Individuals	270.0	23.0	0.0	0.0	0.0	293.0
1–4	Assistance to individuals	216.1	110.1	−3.5	−2.2	−6.5	314.0
1–4	Government spending (Federal, State and Local)	348.4	234.7	82.9	23.1	9.0	698.2
1–4	Assistance to Businesses (e.g., Loans)	630.6	10.0	1.9	1.4	0.0	643.9
1–4	Corporate tax Relief	348.5	258.5	−174.0	−173.0	−56.0	204.0
1–4	Foreign aid					1.2	1.2
5	Unemployment Assistance	0.0	117.0	0.3	0.7	1.2	119.2
5	Direct Payments to Individuals	0.0	169.2	0.1	−0.1	−0.3	168.9
5	Assistance to individuals	0.0	36.8	0.3	0.1	0.2	37.4
5	Government spending (Federal, State and Local)	0.0	96.8	66.1	30.5	23.4	216.7
5	Assistance to Businesses (e.g., Loans)	0.0	315.7	6.8	0.2	−0.5	322.1
6	Unemployment Assistance	0.0	221.3	10.4	0.0	−0.1	231.6
6	Direct Payments to Individuals	0.0	393.7	16.9	0.0	0.0	410.6
6	Assistance to individuals	0.0	68.3	126.9	11.1	−3.8	202.5
6	Government spending (Federal, State and Local)	0.0	398.5	266.2	94.9	107.0	866.7
6	Assistance to Businesses (e.g., Loans)	0.0	72.8	3.3	0.6	−2.6	74.1
6	Corporate tax Relief	0.0	2.7	5.7	−3.1	−26.3	−21.0
6	Foreign aid					10.3	10.3
6	Unallocated					50.3	50.3
	Totals	2035.7	2579.1	410.3	−16.0	106.5	5115.6

Note: The negative numbers for some of the provisions in later years represent the impacts of tax deferral or removal of some tax exemptions.

Appendix D. Modeling the Impact of Causal Factors – Appendix to Section 5

In this appendix we provide further details about the model, closures and shocks used in the simulations.

D1. Model and Time

The dynamic model is run on a semi-annual basis. This is achieved using the variable time, provided in the GDyn model (Ianchovichina and Walmsley, 2012) combined with the approach developed in Dixon et al. (2010):

• 1.Shocking the value of the variable time by 0.5 each period to represent half a year, rather than by 1 for a whole year. The GDyn model is set up so that all of the equations that depend on the length of the period are multiplied by the variable time. By shocking time by a half, only half of a region's investment adds to the capital stocks during the period, and the percent change in capital stocks is half of what it would have done if time had been shocked by 1.
• 2.While the cumulation equations were fixed by the inclusion of the time variable, some of the other convergence parameters needed to be reduced to ensure slower convergence of rates or return and return to full employment.
• 3.Similarly, some of the shocks, primarily those related to avoidance,25 also need to be divided by two in order to obtain annualized 6-monthly shocks. The results from the model were then converted back to semi-annual percent changes by multiplying the result by two.

D2. Baseline

Shocks are given to the following variables:

• -Time (0.5 since the period length is ½)
• -Population forecasts from the UN (pop)
• -Forecasts of factor supply by education (qedusup) based on CEPII forecasts (see Minor et al. (2015) for how the model incorporates labor by education and occupation)
• -Forecasts of GDP (qgdp), investment (qcgds) and government spending (yg) from World Economic Outlook.

Since some of these variables are endogenous in the standard closure of the model, we make the following closure changes provided in Table D-1:

Table D1.

Closures used in Baseline

Variable in model (originally endogenous, becomes exogenous)	Description	Variable in model used for swap (originally exogenous, becomes endogenous)	Description
GDP
qgdp(REG)	Real GDP	afereg(REG) (GDyn variable)	Productivity on factors of production
Investment
sqcgdsreg(“all regions but last one”) (GDyn variable) qcgds(r) = sqcgdsreg(r) + sqcgdsworld	Shift variables in GDyn for fixing investment. When sqcgdsreg(r) fixed, investment is fixed. Sqcgdsworld is fixed in the standard closure.	SDRORT(“all regions but last one”) (GDyn variable)	Risk premium in GDyn (similar to cgdslack in GTAP)
sqcgdsreg(“last region”) (GDyn variable)	Shift variable related to investment for the last region25	dpsavewld (not in GTAP)	Global savings adjustment that alters dpave. dpsave(r) = dpsaveshift(r) + dpsavewld
Government spending
yg(REG)	Government spending	dpgov(REG)	Distribution parameter on government spending

D3. Policy simulations

The policy simulations were added consecutively to the baseline. In the initial mandatory business closures simulation, the difference between the mandatory closure simulation and the baseline are attributed to mandatory closures. Avoidance, deaths and illness, health care, fiscal policy and pent-up demand are then added consecutively to the previous simulation and the differences between policy simulations is attributed to the addition of that policy. In this section, we first examine some of the preliminary closure swaps (Table D-2), before turning to the closures and shocks used to implement each of the causal factors (Table D-3).26

Preliminaries

Table D 2 provides a list of closure swaps incorporated to set the stage for the analysis of COVID.

First, we assume that consumption of essentials remains constant, thereby ensuring that the decline in demand resulting from lower household income falls more heavily on non-essential items, as we might expect with unemployment and the mandatory closure of non-essential businesses.

As in the baseline, we also continue to fix government spending, but rather than allowing both private consumption and savings to adjust as is typical in GTAP because of the regional household, we include an additional swap that ensures that any excess government spending over tax revenues impacts only savings (i.e., the government deficit), and not private consumption. Later in the fiscal stimulus simulations, government spending will be raised, but for the initial simulations, government spending is assumed to be fixed. We also assume that demand for health care services remains unchanged in the initial simulations. This allows us to more easily implement the changes expected in health care due to avoidance, illness and government spending in our subsequent simulations. This also ensures that any changes made to the share of private consumption implemented through the distribution parameter (dppriv) will impact the savings rate.

Finally, we make a small change to the Dynamic GTAP (GDyn) model theory in order to stop the “normal” growth rate from moving with the actual growth rate during COVID. This reflects our belief that COVID is an unusual and temporary event, and while COVID significantly reduces actual economic growth in the short run, its impact on a country's normal, long run, growth rate is uncertain – particularly over the short time horizon we consider here. We therefore chose to fix the normal growth rate in the model. This was done to reduce some of the instability in investment and trade balances caused by excess savings from the mandatory closure of businesses and avoidance, which were eventually used to pay for government spending and pent-up demand.

Table D2.

Closures used for COVID simulations

Variable in model (originally endogenous)	Description	Variable in model used for swap (originally exogenous)	Description
Consumption of Essential goods was fixed
qp_agg(“essentials”,REG) (not in GTAP)	Aggregated private household consumption of all essential and all non-essential commoditiesa	dpp_agg(“essentials”,REG) (not in GTAP)	Distribution parameter applied to private consumption of essentials. Implemented using methods discussed in Walmsley and Minor (2020b)
Fixed Government spending
yg(REG)	Government expenditure	dpgov(REG)	Distribution parameter on government spending
dpav(REG)	Average distribution parameter on Cobb Douglas function of regional householdb	dpsave_shift1(REG) (essentially dpsave in the GTAP model)	Savings distribution parameterc
Demand for Health fixed
qp(“hht",REG) and qg(“hht”,REG)	Consumption of health care by private and government assumed fixed	dpp_stop(“hht",REG) and dpg_stop(“hht”,REG)	Phantom taxes on demand for health care
Turned off a long run adjustment in the dynamic modeld
DKHAT(REG) (GDyn model variable)	Change in normal growth rate from GDyn model	SDKHAT(REG) (GDyn model variable)	Shifter to turn off equation determining the change in normal growth rate

^aEach commodity is categorized as either essential or non-essential (mapping MAPTC2TCA), then this is used to create an aggregate consumption bundle of essential goods that can be held exogenous. In GTAP notation: qp_agg(a,s) = sum(i, TRAD_COMM: MAPTC2TCA(i) = a, SHRVPA(i,s) * qp(i,s)), where SHRVPA(i,s) = VPA(i,s)/sum(k, TRAD_COMM: MAPTC2TCA(k) = i, VPA(k,s)).

^bIf a distribution parameter for consumption, government spending or savings is altered, this causes the other expenditures to adjust and ensures that expenditure equals income. Since a Cobb-Douglas function is used changes in the distribution parameters also represent changes in shares and the shares must always sum to 1.

^cAdjustment required to ensure that savings adjust in response to fixed government spending – i.e., the government deficit rose or fell. The GTAP model uses a regional household and fixing government spending using only the first closure swap, causes both savings (government deficit) and private consumption to adjust to ensure equilibrium. Ordinarily this does not cause any significant issues, but in the case of COVID, where changes in income are significant, the adjustments in savings and private consumption can be large.

^dThe Dynamic GTAP (GDyn) model theory allows for a ‘normal’ growth rate in an economy that gradually converges on the typical, average growth rate of the economy. This gradual convergence occurs as the normal growth rate partially adapts to changes in the actual rate of return, believing that over time changes these changes in the actual return will be incorporated into the long run average. In these simulations we did not allow the normal growth rate to respond to the reduction in actual growth caused by COVID. The assumption is that COVID is an unusual and temporary event that significantly reduces economic growth in the short run, but is unlikely to impact a country's normal, long run rate, particularly during the short period we considered. This reduces the variability in the movement of investment between countries in response to differences in the timing of mandatory closures across countries due to the cyclical nature of COVID variants.

Causal Factors

To model the impact of COVID, we sequentially add each of the following contributing factors.

Mandatory closures

To capture the impact of the mandatory business closures, we reduce the production of the affected sectors using the expedient device of a “phantom tax” to raise prices and lower production through a reduction in final demand. To do this we use a closure swap to exogenise sectoral production and endogenize the corresponding output tax (first section of Table D-3),. The rents, captured by the phantom output tax, accrue to the GTAP regional household.27

Table D3.

Closures and shocks by causal factor

Causal Factors	CGE Model Shocks	Closure swap (if required)	Approach
Mandatory Closures	Production (qo(“selected sectors”,REG))	Production tax (to(“selected sectors”,REG))	“Phantom tax” approach used to raise prices and lower production through a reduction in demand
	Reduction in share of private consumption and rise in savings rate (dppriv(REG))	n.a.
Decline in the workforce due to deaths and illness	Supply of labor (qedusup(REG))	n.a.
	pop(REG)	n.a.
	qp_agg(“essentials”,REG)	dpp_agg(“essentials”,REG)	This closure swap appears in all simulations to fix consumption of essentials
Avoidance	Reductions in the effective supply of labor (qedusup(REG))	n.a.
	Reductions in private demand (qp(“selected sectors”,REG))	dpp_stop(“selected sectors”,REG)	Distribution parameters approach
	Reductions in government demand for health (qg(” hht”,REG))	dpg_stop(“hht”,REG)	Distribution parameters approach
	Reduction in share of private consumption and rise in savings (dppriv(REG))	n.a.
Increased demand for health care	Increased private demand for health care (qp(“hht”,REG))	dpp_stop(“hht”,REG)	Distribution parameters approach
	Increased government demand for health care (qg(“hht “,REG))	dpg_stop(“hht”,REG)	Distribution parameters approach
	Rise in share of private consumption and rise in savings (dppriv(REG))	n.a.
	Rise in government spending (yg(REG))	dpgov(REG)	This closure swap appears in all simulations to stop government spending from falling with income.
Pent-up demand	Increased demand for specific commodities and services (qp(“selected sectors”,REG))	dpp_stop(“selected sectors”,REG)	Distribution parameters approach
	Rise in share of private consumption and rise in savings (dppriv(REG))	n.a.
Fiscal policy	Unemployment and other direct payments to households (dppriv(REG))	n.a.
	Rise in total government spending (yg(REG))	dpgov(REG)	This closure swap appears in all simulations to stop government spending from falling with income.
	Rise in government spending for the public sector (vg(“ogs”, REG))a	dpg_stop(“ogs”, REG)	Distribution parameters approach
	Rise in government spending of other specific commodities (education and health) (vg(“selected commodities”, REG))	dpg_stop(“selected commodities”, REG)	Distribution parameters approach
	Rise in public sector spending of specific commodities (vf(“selected commodities”, “osg”,REG))b	dpf_stop(“selected commodities”, “osg”,REG)	Distribution parameters approach
	Rise in health spending on pharmaceuticals (vf(“bph”, “hht”,REG))	dpf_stop(“bph”, “hht”,REG)	Distribution parameters approach
	Rise in spending on commodities for investment purposes (vf(“selected commodities”, “cgds”,REG))c	dpf_stop(“selected commodities”, “cgds”,REG))	Distribution parameters approach
	Subsidies to firms for workers and owners tfe(ENDW,“selected sectors”,REG)	n.a.
	Corporate taxes tfe(“Capital”,“selected sectors”,REG)d	n.a.

^aVariables labeled with a preceding “v” represent the sum of changes in the quantity and price relative to the numeraire. For instance, vg(i,r) is the percent change in the value of government purchases relative to the model numeraire: vg(i,r) = qg(i,r) + pg(i,r) - ptradewld. Note adding ptradewld to this equation does nothing, it is added for those who might be concerned about homogeneity when shocking so called ‘values’.

^bGTAP sectors cns, ele, ofi, otp and cmn.

^cGTAP sectors cmn and ele.

^dWhile tfe is not the most appropriate tax to use for corporate taxes, there is no explicit corporate tax rate in the GTAP model and the other alternative, to (production tax), was being used for mandatory closures, making tfe the simpler option.

Production is then shocked by the share of the period where no production takes place (Table B-1), after taking into account the ability to telework within an industry. For instance, if we take information from Table B-1 we see that the beverages and tobacco industry closed down for the equivalent of 2.9 percent of the year in the first six months. The gradual reopening of businesses meant that businesses were closed for a further 1.3 percent, or a total of a 4.2 percent (annualized) reduction in production in 2020_1. As sectors reopen or are closed for shorter periods in the second and later semi-annual periods, these phantom taxes are removed gradually according to the rate of reopening. In the second period, the beverage sector was closed for less than 0.01 percent (annualized) in total for the period, hence relative to the previous period, production of beverages rose, which we implement through a fall in the ‘phantom tax’. In the first period of 2021, the sector was fully opened and the endogenously determined rent/phantom tax is removed completely. In two cases (recreational services and accommodation and food services) mandatory business closures extend into the until the first half of 2021, the rent/phantom tax only being fully removed in the second half of 2021.

This must be done in several iterations to take account of the indirect effects of closing some sectors on other sectors. For instance, if restaurants are forced to close, demand for fruit and vegetables or beverages and tobacco used in producing restaurant meals also declines. In some cases (beverages and tobacco for instance), these indirect effects from the mandatory closures (of restaurants, for instance) are larger than the share of that sector subject to the mandatory closure, and, hence, we allow these indirect effects to dominate and sectoral production to decline by more than the share of the sector subject to the mandatory closure. This can be seen by looking at the sign of the endogenously determined rent/phantom tax (to). If the endogenously determined rent/phantom tax falls (i.e., to is positive), then this was an indication that the model predicted a larger decline in production than the period during which businesses in that sector was closed. These sectors (and countries/regions) would then be removed from the list of shocked sectors and the simulation re-run and rechecked. It can take several iterations to obtain the list of sectors that need to be shocked to ensure all the sectors decline by at least the time of the mandatory closures. In the case or re-openings, we also need to ensure that the sector does not expand beyond the share of the sector forced to close during that period. We take account of any potential expansion in production over this share; however, it is captured under pent-up demand, rather than under mandatory closures. Below, in Table D-4, is a list of sectors by region that were shocked in period 1.28

Table D4.

Shocks to capture the Mandatory Closure of Businesses in the first period (2020_1)

Variable shocked	Shocks (based on Table B-1)
qo(“rpp","USA")	−16.96
qo(“nmm","USA")	−17.55
qo(“nmm","China")	−1.86
qo(“ele","USA")	−19.52
qo(“omf","USA")	−18.26
qo(“ros","USA")	−25.10
qo(“osg","China")	−0.67
qo(“osg","ROW")	−1.70
qo(“edu","China")	−12.13

In addition to the business closures, we assume that any income from foregone consumption due to the mandatory closures was saved; savings rates therefore rise via a shock to the share if private consumption (dppriv).

Decline in the workforce due to deaths and illness

The decline in the workforce due to deaths and illness is implemented as a decline in the supply of labor and population, as well as a decline in demand for essential goods and services (Table D-3). Since the workforce and population are already exogenous, no closure changes were required for this simulation. Changes in the labor force reflect the number of deaths, as well as illnesses (including workday losses due to caring for sick family members), of the working age population as a share of the labor force (164,600,000) and the number of working days (250 per working person) in a year; while the change in population reflects all fatalities, outlined in Appendix B, as a portion of the total population. The shocks are provided in Table D-5 below.

Table D5.

Shocks to US labor force and population due to illness and death due to COVID in Scenario 1 (percent)

	Labor force (qedusup, not in GTAP)	Population (pop)
2020_1	−0.12	−0.04
2020_2	−0.21	−0.08
2021_1	−0.17	−0.06
2021_2	−0.02	−0.02
2022_1	−0.05	−0.08

We assume all labor types, regardless of education level, are impacted by the same proportion. Thus, we assume that the pandemic does not discriminate, even though evidence may suggest that lower-skilled workers in customer-facing occupations may have been affected more than other workers. Given the impact of these deaths and illnesses on the workforce and the overall economy is very small relative to the other causal factors, this assumption is unlikely to impact our results significantly.

In addition, private demand for essential goods was also reduced to reflect the decline in demand due to the decline in the population.

Avoidance

Avoidance behavior impacted both the supply of labor and the demand for certain goods and services. These were implemented directly as changes in labor supply and changes in private consumers’ preferences for goods and services, and, in the case of health care, government preferences (Table D-3). In the case of the changes to private and government demand, closures swaps are required. In this case we chose to endogenize the relevant distribution parameters (preferences) to reduce demand exogenously rather than use the phantom tax approach. We also assume that total government spending is fixexfd and shocked by the total increase in spending, and that any income from foregone consumption due to avoidance was saved. As avoidance dissipated, preferences and savings return to normal.

The shocks are taken from Table 3 and are then applied to the following variables:

• -Effective supply of labor by education (qedusup, not in the GTAP model) is shocked by −3.6443 (or (4.58 + 2.71)/2)29 in the first period.
• -Private demand (qp) for education (−1.59/2), health care (−29.16/2), trade (−5.85/2), recreational services (−21.17/2), accommodation and food services (-(26.5 + 9.67)/2), air (−57.48/2) and other (−4.21/2) transport are applied to the relevant sector. Income not spent on private consumption because of avoidance is allocated to savings rather than to other goods and services, hence we also shock the share of private expenditure in income (dppriv) to reflect the decline in private consumption, which increases the savings rate.
• -Government demand (qg) for health care (−29.16/2). Again, government revenue not spent on health care services is assumed to increase savings (or reduce the government deficit), hence we also shock government spending (yg) by the decline in government spending, causing savings to rise.

Avoidance is then assumed to reverse in the second half of 2021 which is simulated as the removal of the effective labor supply and the distribution parameters.

Increased demand for health care

Increased demand for health care due to COVID was also implemented as an increase in demand (preferences) for health care by private and government consumers (Table D-3). Both variables are then shocked by the same percentage changes (1.45 percent in 2020_1, 2.67 percent in 2020_2, 2.07 percent in 2021_1, 0.45 in 2021_2 and 1.45 in 2022_1).30 It was assumed that any health care expenses were paid for out of savings, resulting in a reduction in the savings rate. We therefore also shock the share of private expenditure (through dppriv) and total government spending by the increase in total spending to ensure that savings fall. Again, these preferences are reversed as the demand for health care returns to normal.

Pent-up demand

Pent-up demand was simulated as an increase in demand for goods and services that final consumers had been unable to purchase due to mandatory closures or avoidance. Pent-up demand was found by comparing actual changes in demand by final consumers to model estimates. Where demand was found to be stronger than the model predicted, an increase in demand was simulated. Table B-6 provides estimates of demand for selected goods relative to the fall in demand at the height of COVID (April 2020). Where the model predicts a lower demand for these commodities than the actual, we assume that the difference is due to pent-up demand. Based on the model results we obtain the following shocks for the U.S. (Table D-6). As in previous simulations we need to shock demand by private consumers for specific commodities and we again use the relevant distribution parameters to exogenise demand (Table D-3).

Any increased private consumption was assumed to be funded from savings using shocks to the share of private consumption (dppriv). Again, the rise in preferences and savings are reversed as pent-up demand subsides. Since we were only able to obtain estimates of pent-up demand for 2020 and the first half of 2021, further pent-up demand in late 2021 and 2022 could lead to further increases in real GDP.

Table D6.

Shocks to capture pent-up demand (based on Table B-6 and model results) (percent)

Variable shocked	2020_1	2020_2	2021_1
qp(“trd","USA")	6.39	–	1.60
qp(“afs","USA")	41.60	19.83	–
qp(“atp","USA")	3.47	–	24.31
qp(“rsa","USA")	14.42	–	31.60
qp(“ros","USA")	6.97	–	3.75
qp(“hht","USA")	14.65	–	0.75

Fiscal policy

The various rounds of fiscal stimulus were decomposed according to both round and mechanism for the U.S. Unlike the other causal effects, where we also created estimates for China and the rest of the world, we did not implement any fiscal policy changes in China and the rest of world. The impacts of the fiscal stimulus were separated into three parts: 1) Rounds 1–4, implemented in early 2020, 2) Round 5, implemented in late 2020 and early 2021, and 3) Round 6, implemented in early 2021. When we prepare the input data for the CGE modeling, detailed provisions in each of these rounds were categorized by the type of fiscal stimulus implemented. Not all rounds used all these categories. The closures and shocks required to implement the fiscal stimulus depended on the type of assistance and are summarized in Table D-3 and include:

• -Unemployment and other direct payments and assistance to individuals are implemented as shocks to private household consumption via the distribution parameter (dppriv). The shocks are given in Table D-7. Since these payments are temporary, the shocks are assumed to reverse in subsequent periods.
• -
  Government spending includes:

  • o
    increases in direct government spending on a product (vg);
  • o
    increases in public service sector purchases of the commodity (vf); or
  • o
    increases in investment (cgds) spending on electronics and communications.

These expenditures are swapped with the relevant distribution parameters shown in Table D-3. The choice between using direct government purchases (vg) or funneling money through the public sector (vf) depended on which of these bought these goods or services in the base data. For instance, government buys mostly public services, education and health. The public service then purchases electronics, construction, transportation, communication and financial services.

The direct shocks to government spending over time are provided in Table D-8 below. Again, the increases are one-offs, leading to their removal in subsequent periods.

• -Assistance to businesses and corporate tax relief are implemented as subsidies to businesses' labor and capital (tfe) or, while a business was subject to mandatory closures, direct payments to households (dppriv). Amounts received were converted into changes in the power of the tax and implemented as shocks and reversed as subsidies were later removed.

The fiscal stimulus was paid for by increased borrowing (i.e., using savings). As total government spending increases, savings fall automatically.

The shocks are obtained by taking the increase in cash payments or government spending as a share of the total annual private consumption or government spending of the commodity, respectively. We assumed that a portion of the unemployment benefits and cash payments were saved and hence had no impact on the results. The portion saved depended on the type of payment and was based on the data collected on savings rates discussed in Section 5. Using these data as a guide, we assumed that the share of the cash payment saved (36%) was higher than the share saved from unemployment benefits (7.8%), reflecting the fact that unemployed households were found to have lower savings rates during the pandemic than households in general.31 Further information on the fiscal policy simulations and results is available in Walmsley et al., 2022a.

Table D7.

Shocks to private expenditure to capture unemployment benefits, direct cash payments and relevant business subsidies

	Rounds 1-4	Round 5	Round 6
2020_1	1.37	0.00	0.00
2020_2	1.37	0.00	0.00
2021_1	−2.21	0.98	0.15
2021_2	0.00	0.00	3.69
2022_1	−0.53	−0.96	−3.49
2022_2	−0.01	0.00	0.36
2023_1	−0.01	0.00	0.36
2023_2	−0.01	0.00	0.02

Table D8.

Shocks to government spending

	Public sector servicesa	Education	Healthb
2020_1	5.85	0.26	3.48
2020_2	5.52	0.26	3.39
2021_1	−6.51	3.09	1.53
2021_2	21.52	2.24	3.87
2022_1	−15.39	2.23	−5.23
2022_2	0.00	0.00	0.00
2023_1	−3.18	−1.70	−3.57
2023_2	0.00	0.00	0.00

^aPublic sector then used this money to increase spending on electronics, construction, transportation, communication and the financial services.

^bHealth allocated this to pharmaceuticals.

Appendix E. Sectoral Impacts – Appendix to Section 7

Table E1.

Percent Changes in Sectoral Production in First Semi-annual Period of 2020 Due to Pandemic, Decomposed by Causal Effect (cumulative percent differences from baseline)

	Mandatory closures	Avoidance	Labor	Health	Rounds 1-4	Round 5	Round 6	Pent-up demand	Total
Wheat	−3	−2	0	0	9	0	0	0	5
Cereal grains nec	−10	−5	0	0	9	0	0	0	−6
Vegetables, fruit, nuts	−7	−2	0	0	9	0	0	−3	0
Oil seeds	−5	−2	0	0	9	0	0	0	2
Sugar cane, sugar beet	−12	−7	0	0	10	0	0	1	−9
Plant-based fibers	−16	−6	0	0	12	0	0	−1	−11
Crops nec	−19	−8	0	1	33	0	0	−1	0
Bovine cattle, sheep and goats, horses	−8	−5	0	0	11	0	0	−1	−2
Animal products nec	−10	−5	0	0	10	0	0	0	−6
Raw milk	−10	−8	0	0	10	0	0	0	−8
Wool, silk-worm cocoons	−26	−8	0	1	45	0	0	0	0
Forestry	−30	−10	0	1	11	0	0	0	−30
Fishing	−16	−11	0	1	21	0	0	3	−9
Coal	−19	−9	0	1	10	0	0	2	−19
Oil	−31	−15	0	1	15	0	0	2	−32
Gas	−17	−6	0	1	29	0	0	0	0
Minerals nec	−25	−5	0	0	13	0	0	1	−20
Bovine meat products	−7	−4	0	0	11	0	0	−1	−2
Meat products nec	−8	−4	0	0	11	0	0	0	−2
Vegetable oils and fats	−12	−5	0	0	14	0	0	1	−4
Dairy products	−11	−9	0	1	11	0	0	1	−10
Processed rice	−12	−2	0	0	10	0	0	−2	−5
Sugar	−12	−7	0	1	11	0	0	1	−10
Food products nec	−11	−6	0	0	8	0	0	0	−9
Beverages and tobacco products	−42	−25	0	3	10	0	0	1	−51
Textiles	−36	−18	0	2	17	0	0	−1	−38
Wearing apparel	−43	−26	0	3	21	0	0	−4	−47
Leather products	−40	−21	0	2	35	0	0	−2	−34
Wood products	−28	−6	0	1	9	0	0	0	−26
Paper products, publishing	−30	−17	0	1	15	0	0	3	−32
Petroleum, coal products	−31	−16	0	1	11	0	0	2	−35
Chemical products	−31	−13	0	1	16	0	0	0	−30
Basic pharmaceutical products	−15	−35	0	2	22	0	0	7	−31
Rubber and plastic products	−35	−12	0	1	17	0	0	1	−33
Mineral products nec	−37	−7	0	1	18	0	0	1	−30
Ferrous metals	−29	−6	0	1	17	0	0	0	−21
Metals nec	−29	−5	0	1	23	0	0	1	−17
Metal products	−28	−7	0	1	15	0	0	1	−23
Computer, electronic and optical products	−39	−6	0	1	32	0	0	1	−25
Electrical equipment	−29	−5	0	1	18	0	0	0	−19
Machinery and equipment nec	−26	−4	0	0	15	0	0	0	−17
Motor vehicles and parts	−32	−10	0	1	13	0	0	−1	−31
Transport equipment nec	−25	−3	0	0	34	0	0	0	−2
Manufactures nec	−38	−15	0	1	20	0	0	2	−35
Electricity	−19	−10	0	1	10	0	0	2	−19
Gas manufacture, distribution	−16	−9	0	1	10	0	0	1	−15
Water	−11	−7	0	0	8	0	0	1	−10
Construction	−22	−2	0	0	6	0	0	0	−19
Trade	−38	−20	0	2	13	0	0	10	−43
Accommodation, Food and service activities	−46	−46	0	5	21	0	0	47	−63
Transport nec	−31	−15	0	1	13	0	0	2	−33
Water transport	−31	−13	0	1	11	0	0	2	−32
Air transport	−36	−20	0	2	21	0	0	11	−37
Warehousing and support activities	−32	−15	0	1	14	0	0	2	−33
Communication	−34	−19	0	2	12	0	0	3	−39
Financial services nec	−37	−20	0	2	11	0	0	2	−43
Insurance	−42	−26	0	3	11	0	0	−2	−51
Real estate activities	−36	−26	0	3	12	0	0	49	−46
Business services nec	−28	−17	0	1	16	0	0	4	−30
Recreational and other services	−51	−32	0	4	15	0	0	22	−60
Public Administration and defense	−14	1	0	−1	13	0	0	0	−2
Education	−32	−8	0	1	7	0	0	−2	−33
Human health and social work activities	0	−53	0	4	14	0	0	22	−45

Supplementary data

The following is the Supplementary data to this article:

Data availability

Some data inputs can be provided, but not all data and models.