The cost of business restrictions during the COVID-19 pandemic

Introduction

COVID’s most notable impact has been illness and the loss of life, but the pandemic has also imposed a substantial societal cost. The economy has grabbed headlines, as cumulative losses during just the first two quarters of 2020 amounted to between 1 and 7% of annual GDP across the globe, and approximately 3% of annual GDP in the United States.¹ Other societal costs include reduced care for health conditions other than COVID,^2,3 although that impact may have been largely limited to the initial several months of the pandemic.⁴ Finally, evidence suggests that lost in-person schooling will permanently diminish economic opportunities for today’s students; moreover, the impact is greatest for students from the most disadvantaged backgrounds.⁵

The pandemic’s economic costs suggest that COVID-19 medical interventions could confer societal benefits exceeding the value of their health benefits.⁶ Based on stock market movements in apparent response to positive news about vaccines in development, for example, one analysis estimated a pandemic-ending therapy to be worth between 5 and 15% of total wealth.⁷ In the United States alone, that result implies a value ranging from $6 to $17 trillion.⁸

Assessing the societal value of COVID therapies that have a less dramatic impact on the population as a whole, such as antiviral treatments⁹ and monoclonal antibodies,¹⁰ presents a distinct challenge. These therapies do not directly address population disease transmission and therefore have less potential to return life to “normal”. Instead, they aim to prevent progression to severe disease or speed recovery.

We hypothesize that these clinical benefits, if the data hold up, nonetheless confer a societal benefit by diminishing the need for government-imposed restrictions on economic activity designed to increase social distancing. By “flattening the curve”, economic restrictions and other distancing measures aim in part to prevent the health care system from becoming overwhelmed.^11,12 The link between hospital utilization in particular and government-imposed restrictions seems evident in headlines from late 2020. For example, one headline noted, “If NY hospitals get overwhelmed, state could shut down again, Cuomo says”.¹³ Another stated, “California’s Governor Warns of ‘Drastic Action’ as Hospitals Near Crisis”.¹⁴ A December 3 executive order outlined economic restrictions in California when a region’s free intensive care unit capacity dropped below 15% – including a ban on hotel stays for out-of-state guests, suspension of indoor dining at restaurants, and a 20% capacity limit on retail establishments.¹⁵ When Massachusetts governor, Charlie Baker, loosened business restrictions toward the end of January 2021, he identified a drop in COVID hospitalizations as a key factor informing his decision.¹⁶

Because high hospital utilization levels trigger restrictions, therapies that reduce hospitalizations or hospital length of stay can also reduce the need for restrictions. This paper’s aim is to estimate the cost of these restrictions and hence the potential value conferred by COVID-19 therapies that help avert hospital admissions by preventing progression to severe disease or that speed recovery.

Methods

To assess the cost of government-imposed economic restrictions, we combined high frequency economic data, information on the pandemic’s progression (case incidence and mortality), and information about when state governments imposed and lifted restrictions on business activities. Rather than including data from spring, 2020, when the pandemic’s initial onslaught increased unemployment claims and reduced consumer spending to an extent unrepresentative of subsequent periods during the pandemic,^17,18 we instead focused on the fall and winter of 2020-2021. All of the analysis was conducted in R v4.0.3, utilizing the R nlme package for regression analysis (R package: nlme, RRID:SCR_015655).

Descriptive statistics: We report the number of observations, mean, and standard deviation for all outcome quantities and all explanatory variables. We stratify the data by the level of regulatory measures in place during that week. As described below (see Data), we classify each U.S. state during each week as either “mostly closed” due to government-imposed activity restrictions, or as not “mostly closed.”

Base case analysis: We used regression to predict the influence of restrictions on consumer spending and weekly initial unemployment claims. We controlled for daily COVID case count, COVID mortality, and the change in these two rates over the preceding two weeks. More formally, we developed three sets of models, one set for each of three outcomes:¹ percent change since January, 2020 in total seasonally adjusted consumer spending,² percent change in seasonally adjusted consumer spending on restaurants and accommodations since January, 2020, and³ weekly unemployment claims (not seasonally adjusted) as a percent of the 2019 US workforce.

Our base case regression specification models each outcome as,

${\mathit{\text{Outcome}}}_i^s=\alpha +{\beta }_1{\mathit{\text{Restricted}}}_i^s+{\beta }_2{\mathit{\text{NewCases}}}_i^s+{\beta }_3{\mathit{\text{DeltaNewCases}}}_i^s$+${\beta }_4{\mathit{\text{NewDeaths}}}_i^s+{\beta }_5{\mathit{\text{DeltaNewDeaths}}}_i^s$,

where the subscript $i$ denotes the week and the superscript $s$ denotes one of the 50 U.S. states (and Washington, D.C. for the initial unemployment claims). Explanatory variables included ${\mathit{\text{Restricted}}}_i^s$ (a binary variable indicating if state $s$ was “mostly closed” during week $i$; see data sources below); ${\mathit{\text{NewCases}}}_i^s$ and ${\mathit{\text{NewDeaths}}}_i^s$, which refer to state-specific, population-normalized daily case incidence and mortality; and $\mathit{\text{Delta}}{\mathit{\text{NewCases}}}_i^s$ and $\mathit{\text{Delta}}{\mathit{\text{NewDeaths}}}_i^s$, which refer to state-specific, population-normalized changes in these values over the last two weeks. We included the COVID case and mortality rates, and their rate of change to address the possibility that reduced economic activity may in part be due to voluntary disengagement from the economy in response to reports of higher rates of COVID illness and death.

All regressions are linear mixed models that account for nesting within state and correlation across observations that are close in time. To account for temporal correlation, the base case model implemented a first-order, autoregressive (lag 1-week) covariance structure. This covariance structure assumes correlation is greatest for observations adjacent in time, with decreasing correlation between observations more distant in time.

Sensitivity analysis: We conducted a series of sensitivity analyses to explore how alternative assumptions influence our conclusions.

First, we conducted an analysis limited to the 10 states designated “mostly closed” at some point between September 15 through December 27, 2020 (California, Kentucky, Illinois, Michigan, Minnesota, New Mexico, Oregon, Pennsylvania, Rhode Island, and Washington).

Second, we explored alternative model specifications that omitted some of the explanatory variables in the base case model, including¹ the daily COVID mortality rate,² the daily COVID case rate, or³ both the case rate and the mortality rate. We conducted this sensitivity analysis because inclusion of the case rate and mortality rate in the regression model may lead to “over-control” of phenomena correlated with government restrictions, and that could artificially reduce the estimated impact of government restrictions on economic activity. This set of sensitivity analyses explores the potential magnitude of this type of potential error.

Third, we considered alternative criteria for designating a week as subject to government-imposed restrictions. The first alternative designates a state as subject to restrictions only after restrictions have been in place for at least 14 days. This alternative explores the potential for a lag (of two weeks) between a state’s imposition of restrictions and the impact of those restrictions on economic activity. For example, if a state enacted restrictions from the 1^st to the 21^st of October, this sensitivity analysis designates the week of October 1^st and the week of October 8^th as not $\mathit{\text{Restricted}}$, but designates the week of October 15^th as $\mathit{\text{Restricted}}$. The second alternative recognizes only those restrictions that remain in place for at least three weeks. For the preceding example, that would mean that all three weeks (starting October 1^st, October 8^th, and October 15^th) would be designated as $\mathit{\text{Restricted}}$. On the other hand, if the state had imposed restrictions from October 1^st to October 20^th (just under 3 weeks), this sensitivity analysis would designate all three weeks as not $\mathit{\text{Restricted}}$. That is, we treat states imposing only restrictions shorter than three weeks in duration as unrestricted at all times. The third alternative set of criteria recognizes only those restrictions that include a complete ban on indoor dining at restaurants.

Finally, we considered alternative model correlation structures, including 2^nd, 3^rd, and 4^th order autoregressive models (i.e., AR², AR³, and AR⁴), 1^st and 2^nd order moving average (i.e., MA¹ and MA²), and a series of autoregressive-moving average models (ARMA) with different combinations of first- and second-order AR and MA components (i.e., AR1MA1, AR1MA2, AR2MA1, and AR2MA2).

We report the Akaike Information Criterion (AIC) statistic as an indicator of goodness of fit for each model (lower or more negative values indicate better fit).

Data: We report our time series data as weekly values. For spending data reported by the original source as daily seven-day moving averages, we retained one value from each seven-day period to represent that week.

Outcomes – Outcomes data come from the Opportunity Insights COVID-19 Economic Tracker project at Harvard University.^19,20 The Opportunity Insights Economic Tracker data repository sourced its spending data from Affinity Solutions, and its initial unemployment claims data from the Department of Labor. Selection of these outcomes reflects the expectation that government restrictions affect consumer spending in general and restaurant dining and accommodations in particular. These impacts, in turn, contribute to higher initial unemployment claims.

For state $s$ and week $i$, outcomes include percent change in seasonally adjusted spending on restaurants and accommodations compared to January 2020 (${\mathit{\text{SpndRst}}}_i^s$), and percent change in seasonally adjusted total consumer spending compared to January 2020 (${\mathit{\text{SpndTtl}}}_i^s$). To eliminate weekend and weekday periodicity, the Opportunity Insights Economic Tracker group computed seven-day moving average values for the spending data, as described above. To eliminate effects of state size, they also normalized all spending values by population size. The spending data span the period September 15 to December 27, 2020. Our third outcome, for state $s$ and week $i$, was initial unemployment claims (reported on Thursday of each week) per 100 people in the 2019 labor force (${\mathit{UE}}_i^s$). This data series spanned the period September 19 to December 19, 2020.

High frequency data are well suited to assessing restrictions imposed at moderately different times across states and remaining in effect for durations often measured in weeks during the fall and winter of 2020-2021. Moreover, because states imposed and lifted these restrictions only shortly before we conducted our analysis, frequently updated data best cover salient periods. Conventional economic activity datasets, like government estimates of gross domestic product (GDP) growth, undergo updating too infrequently to assess the impact of short-term restrictions and the recent imposition of restrictive government measures.

Explanatory variables – As with outcomes, we index explanatory variables by state $s$ and week $i$. Our primary explanatory variable of interest, designated ${\mathit{\text{Restricted}}}_i^s$, is binary. We classified state businesses as restricted or not restricted based on narrative descriptions of state-level closures maintained by the New York Times, which reported for each day whether businesses in each state were “mostly closed.”²¹ Based on review of these narratives, we identified when each state changed from not “mostly closed” to “mostly closed”, or the reverse. Our base case analysis designated business in a state as “restricted” during week $i$ if, according to the New York Times database, businesses were “mostly closed” on Monday (for the consumer spending outcomes) or Thursday (for the initial unemployment claim outcome) of that week.

We included other explanatory variables in an effort to account for how news about the pandemic might influence voluntary, spontaneous reductions in economic activity. These variables included seven-day average, population-normalized¹ daily new COVID cases (${\mathit{\text{NewCases}}}_i^s$),² daily COVID mortality (${\mathit{\text{NewDeaths}}}_i^s$),³ the increase in daily new COVID cases over the last two weeks (${\mathit{\text{DeltaNewCases}}}_i^s$), and⁴ the increase in daily COVID mortality over the last two weeks (${\mathit{\text{DeltaNewDeaths}}}_i^s$). Daily COVID cases and deaths were provided from the New York Times COVID-19 Repository to the Opportunity Insights Economic Tracker data repository.²²

Results

Descriptive statistics: Table 1 summarizes the data. Outcomes (consumer spending, unemployment claims) are more favorable during weeks without restrictions (first three numeric columns) than during weeks with restrictions (last three numeric columns). Case counts and mortality are lower during weeks without restrictions than during weeks with restrictions. The rate at which case counts grow (increase per two-week period) is higher during weeks without restrictions. On the other hand, the change in mortality is similar across these two sets of weeks.

Base case analysis: Table 2 reports regression analysis results for our base case models. The estimated impacts of government-imposed restrictions (row labeled “Restricted”) are statistically significant and in the expected direction for all three outcomes – i.e., a reduction in consumer spending, a reduction in restaurant and accommodations spending, and an increase in initial unemployment claims.

Sensitivity analysis: Our first sensitivity analysis (Table 3) explored the impact of restricting attention to the 10 states that all experienced periods during which the government imposed restrictions and other periods during which the government did not impose restrictions. For these 10 states, government restrictions had the same impact on total consumer spending (coefficient of -0.02 for Restricted, i.e., a 2% reduction) as they had for all 50 states (Table 2). The impacts on restaurant and accommodations spending (coefficient of -0.04 in Table 3 vs. coefficient of -0.05 in Table 2) and on initial unemployment claims (coefficient of 0.20 in Table 3 vs. coefficient of 0.21 in Table 2) were slightly smaller than the corresponding impacts in the base case.

Tables 4, 5, and 6 show results for the second set of sensitivity analyses – i.e., for alternative models that do not control for COVID case rate, COVID mortality, or both. The impacts of government restrictions on total consumer spending (Table 4) remain similar to the corresponding base case impacts (coefficient of around -0.02). Impacts on restaurant and accommodations spending (Table 5) likewise remain similar to the corresponding base case impacts (coefficient of around -0.05). For initial weekly unemployment claims (Table 6), not controlling for case rate and death rate modestly increased the estimated impact of restrictions on initial unemployment claims to 0.25 percent of the 2019 workforce from 0.21 percent in the base case.

Tables 7, 8, and 9 present regression model results for the third set of sensitivity analyses, which explore alternative definitions for the $\mathit{\text{Restricted}}$ variable. Models designating weeks as $\mathit{\text{Restricted}}$ only after two weeks following enactment of state-imposed distancing measures (left panels in Tables 7, 8, and 9) estimated that government restrictions had smaller estimated impacts than the corresponding impacts in our base case model. In Table 7, the left panel coefficient for $\mathit{\text{Restricted}}$ is -0.01, compared to -0.02 for the corresponding base case coefficient in Table 2. In Table 8, the left panel coefficient for $\mathit{\text{Restricted}}$ is -0.02, compared to -0.05 for the corresponding base case coefficient in Table 2. In Table 9, the left panel coefficient for $\mathit{\text{Restricted}}$ is 0.04, compared to 0.21 for the corresponding base case coefficient in Table 2. As detailed in the last rows in Tables 7, 8, and 9, these models had higher AIC values than the base case model, indicating that they do not fit the data as well as the base case model.

Finally, we report results for other model correlation structures (Tables 10, 11, and 12). Not all alternative correlation structures we evaluated produced models that converged. For those that did, the estimated impacts of government-imposed restrictions were similar to or modestly larger than the corresponding base case estimates. The AIC goodness of fit statistics indicate, however, that models that produced estimates for the impact of government-imposed restrictions that differed the most from the corresponding base case estimates did not fit the data as well as the base case model (less negative AIC value).

Discussion

We found that government-imposed restrictions during the fall and early winter of 2020-2021 were associated with reduced economic activity. Descriptive statistics revealed lower total consumer spending, lower consumer spending on restaurants and accommodations, and a greater number of initial unemployment claims during weeks with restrictions than during weeks without restrictions (Table 1 – compare outcome means for weeks without and with restrictions). Regression analysis revealed an association with restrictions even after controlling for the number of new COVID cases, COVID mortality, and the change in those rates over the preceding two weeks (Table 2). That influence appears to be robust to modeling assumptions, as indicated by the modest impact of our sensitivity analyses on the estimated impact of government-imposed restrictions. Where sensitivity analyses yielded the most notable departures from the base case, the AIC value suggested the alternative model did not fit the model as well.

Several factors complicated assessment of the impact of government restrictions on economic outcomes. Importantly, the restrictions imposed by state governments differed across states and over time. Our sensitivity analyses using alternative definitions for government restrictions suggest, however, that this issue does not invalidate our findings. A second complication is the fact that government-imposed restrictions tend to coincide with the events that cause voluntary reductions in economic activity. In short, the same news headlines that cause state governments to impose restrictions also cause many consumers to voluntarily stay home and reduce their spending. We attempted to isolate the impact of government-imposed restrictions by including in our models daily COVID cases and mortality, and the changes in those statistics over the preceding two weeks.

We estimated the dollar cost of restrictions as follows. Our base case analysis results imply that restrictions reduce consumer spending by 2%, reduce restaurant and accommodations spending by 5%, and increase weekly unemployment claims by 0.21% of the 2019 work force. Nationally, total annualized consumer spending amounted to $13.3 trillion in the fourth quarter of 2019;²³ restaurant and accommodations spending totals approximately $1.2 trillion annually, including $863 billion spent in restaurants²⁴ and $300 billion spent on accommodations.²⁵ The U.S. labor force in 2019 totaled 163 million individuals.²⁶ The model results hence imply that restrictions, if imposed by all states for a full year, would reduce total consumer spending by $2\%\times \mathit{\text{\$}}13.3\mathit{\text{trillion}}\approx \mathit{\text{\$}}270\mathit{\text{billion}}$, of which $5\%\times \mathit{\text{\$}}1.2\mathit{\text{trillion}}=\mathit{\text{\$}}60\mathit{\text{billion}}$ could be attributed to a reduction in spending on restaurants and accommodations. The model results also imply that restrictions imposed by all states would increase initial weekly unemployment claims by $0.21\%\times 163\mathit{\text{million}}\approx \mathrm{340,000}$ initial claims each week.

The computations just described yield cost estimates for a hypothetical, nationwide, year-long imposition of business restrictions. We scale these estimates down to characterize the impact of the restrictions actually imposed. During the fall and winter of 2020-2021, when the U.S. experienced the pandemic’s third wave, 10 states (CA, IL, KY, MI, MN, NM, OR, PA, RI, and WA) imposed restrictions for between 21 and 86 days, with an average (and GDP-weighted average) duration of 49 days. Those states represent approximately one-third of the country’s GDP.²⁷

Taking a third of the projected nationwide, annual costs ($90 billion for total spending and $20 billion for restaurant and accommodations spending), and then scaling these values down by a further 86.5 percent to impute the costs associated with the imposition of restrictions for 7 weeks (13.5 percent the year) yields costs of $12 billion in total consumer spending, of which $2.7 billion would be attributable to decreased spending on restauraunts and accommodations. Assuming employment is approximately proportional to GDP, these states would together experience an additional 114,000 initial weekly unemployment claims each week during the imposition of these restrictions. (We note that the unemployment data do not reflect seasonal adjustment, but it is not evident that this limitation introduces any particular bias; nonetheless, it does introduce some uncertainty.)

While a total cost of $12 billion can seem modest in the context of a global pandemic with estimated costs likely amounting to trillions of dollars in lost economic activity, the losses estimated here are indeed substantial. Keep in mind that they occur over a limited time period (during the 1 or 2 months at the peak of a surge in COVID cases) and that the population can experience multiple case surges during a pandemic; as of this writing, the United States has experienced three such surges and is in the midst of a fourth due to the Delta variant. Assuming that late 2020 is representative of tendencies for the government to restrict economic activity in response to high hospital utilization, therapies reducing hospital utilization have the potential to confer substantial societal value.

Whether prices charged for therapies should reflect this component of value is another question, as decision makers must also consider a range of issues, including affordability.²⁸ Importantly, this analysis shows that estimating the cost of government-imposed restrictions is feasible, hence making it possible to extend health technology assessments beyond the typical and often exclusive focus on the valuation of health benefits.

Conclusions

Government restrictions reduced consumer spending by 12 billion dollars and increased unemployment claims by 114,000 each week during the Fall/early Winer surge of COVID-19. Sensitivity analysis showed these findings to be robust to alternative modeling assumptions. The criteria for implementing business restrictions in various U.S. states, such as California and New York, suggest that increases in hospital utilization during COVID pandemic surges trigger such government restrictions. As COVID-19 therapies that can lower the risk of hospitalization and severe disease are developed and become available, government business restrictions may become a less necessary measure for lowering the hospital burden. Therefore, our analysis suggests that investment in the development and use of such therapies may not only provide a substantial health benefit, but also avert some of the economic impact of government restrictions, providing a substantial societal benefit.

While a therapy’s health benefits will no doubt remain the core element of health technology assessment, for therapies addressing threats that affect others in addition to those who become ill, recognizing societal benefits will help allocate resources to promote innovations that address the most important risks society faces.

Data availability statement