key: cord-0858192-8mb4zt48
authors: Dave, Dhaval; Friedson, Andrew I.; Matsuzawa, Kyutaro; Sabia, Joseph J.
title: When Do Shelter‐in‐Place Orders Fight COVID‐19 Best? Policy Heterogeneity Across States and Adoption Time
date: 2020-08-03
journal: Econ Inq
DOI: 10.1111/ecin.12944
sha: 5013b1037bb6f275df9b081f8bf58b69507180b7
doc_id: 858192
cord_uid: 8mb4zt48

Shelter in place orders (SIPOs) require residents to remain home for all but essential activities. Between March 19 and April 20, 2020, 40 states and the District of Columbia adopted SIPOs. This study explores the impact of SIPOs on health, with particular attention to heterogeneity in their impacts. First, using daily state‐level social distancing data from SafeGraph and a difference‐in‐differences approach, we document that adoption of a SIPO was associated with a 9 to 10 percent increase in the rate at which state residents remained in their homes full‐time. Then, using daily state‐level coronavirus case data collected by the Centers for Disease Control and Prevention, we find that approximately three weeks following the adoption of a SIPO, cumulative COVID‐19 cases fell by approximately 53.5 percent. Event‐study analyses confirm common COVID‐19 case trends in the week prior to SIPO adoption and show that SIPO‐induced case reductions grew larger over time. However, this average effect masks important heterogeneity across states — early adopters and high population density states appear to reap larger benefits from their SIPOs. Finally, we find that statewide SIPOs were associated with a reduction in coronavirus‐related deaths, but estimated mortality effects were imprecisely estimated.

5 modest short-run effects from statewide orders Sears et al. 2020: Abouk and Heydari 2020) .

A rapidly emerging literature has begun to study the short-run health effects of SIPOs. Friedson et al. (2020) focus specifically on California, which enacted the nation's first shelter-inplace order. Using a synthetic control approach, and a variety of matching strategies, find that California's SIPO was associated with approximately 125.5 to 219.7 fewer COVID-19 cases per 100,000 following the policy's first three weeks of enactment. To put their estimates in context of SIPO-related economic costs, they suggest that California's SIPO caused approximately 400 job losses per life saved.

While understanding the experience in a single state is important, the findings of work such as Friedson et al. (2020) may not generalize to jurisdictions with different population or outbreak characteristics. For example, California is an outlier, both as an early SIPO adopter and as a highly urbanized state with extraordinarily low COVID-19 case growth at the time of SIPO adoption. Given that an additional 40 states (including D.C.) adopted statewide SIPOs following California's enactment, understanding both the average effect of SIPOs and the heterogeneity of their impact based on the characteristics of the target location is of primary policy importance.

Moreover, the recent resurgence of new COVID-19 cases has compelled many states to suspend or push back their reopening plans. As states are again making decisions on whether to reimpose social distancing orders and re-issue mandates for shutting down non-essential businesses, it is critical for policy-makers to have information on the effectiveness of these policies and the conditions under which they are effective. We provide some of the first evidence on these questions in this study.

This article is protected by copyright. All rights reserved.

First, using daily state-level measures of social mobility from SafeGraph, Inc., we document that statewide SIPOs were associated with a 9 to 10 percent increase (relative to the pre-treatment period) in the share of the population that sheltered in place completely on any given day. This treatment-control differential increases during the first week following SIPO adoption and then remains constant or slightly declines. Next, turning to COVID-19, differencein-differences estimates show that the adoption of a SIPO had little effect on COVID-19 cases during the five (5) days following its enactment, corresponding to the median incubation period.

However, after the incubation period, and intensifying rapidly three weeks or more after the policy's adoption, SIPO adoption is associated with an approximate 53.5 percent decline in COVID-19 cases. Approximately weeks following SIPO adoption, this corresponds to approximately 3,073 fewer cumulative COVID-19 cases for the average Evidence from event study analyses is consistent with common pre-treatment trends. Our results persist when we (i) drop California from our panel, confirming that we are not simply replicating Friedson et al. (2020) , and (ii) when we control for state-specific growth in COVID-19 testing, which could affect the number of reported coronavirus cases. While statewide SIPOs were negatively related to coronavirus-related deaths, estimated mortality effects were imprecisely estimated.

Importantly, we find that the impact of the average state SIPO masks important statelevel heterogeneity. The earliest adopters of statewide SIPOs saw the largest declines in the rate of coronavirus cases, including declines in the rate of COVID-19-related mortality. In addition, more densely populated states also appear to reap relatively larger health benefits from their 7 SIPOs. 5 Consistent with these larger health impacts, we find that statewide SIPOs are far more effective at increasing social distancing among early adopting states and states with higher population densities. We conclude that there are important heterogeneous health impacts of statewide SIPOs across states and adoption time.

After being detected in Wuhan, China, in December 2019, the first confirmed case of COVID-19 in the United States was identified on January 20, 2020 in Washington State. 6 The disease spread exponentially over the next three months across the U.S., with confirmed cases at 778,328 as of April 20, 2020, accounting for 32 percent of the global caseload. Public health interventions to flatten this growth trajectory have mobilized around two complementary sets of policy responses. 7 Surveillance-based initiatives, such as expanding COVID-19 testing capacity and deploying antibody tests, seek to monitor the spread and intensity of the disease. 8 These efforts can be instrumental in identifying infected persons, and tracing and monitoring their contacts to limit further spread of the virus. In addition, mitigation and suppression policies aim at lowering the reproduction rate of the virus and slowing its spread by limiting interactions between individuals in the community and increasing social distancing (Ferguson et al. 2020) .

Components of such a response include shelter-in-place orders, closures of educational facilities, restrictions on mass gatherings, and closure of business and non-essential services. 5 This is consistent with findings from a case-study of county-level orders in Texas . 6 See Holshue et al. (2020) . 7 While numerous clinical trials for a COVID-19 vaccine and anti-viral treatments are underway, significant lag times with clinical testing and approval by the Food and Drug Administration (FDA) mean that an effective prophylactic and treatment are unlikely over the short-term. Hence, public health efforts centered on suppression and mitigation take on added relevance to prevent the surge in cases from overwhelming the healthcare system. 8 Also see: https://www.cdc.gov/coronavirus/2019-ncov/covid-data/covidview/purpose-methods.html. 8 Given that the authority for imposing sheltering-at-home orders and school or businesses closures rests with states and localities, the Federal response has focused on (i) providing funding to states to bolster preparedness and healthcare capacity, and (ii) surveillance-based policies aimed at expanding testing and tracking infection rates. Some suppression efforts have also included travel restrictions to limit infections from international exposure. For example, on January 31, 2020, the Trump administration enacted restrictions on all foreign nationals who had been in China over the past 14 days from entering the U.S. Then, following a surge in COVID-19 related deaths, the administration suspended travel from the Schengen Area to the U.S. starting on March 13, which was further extended to include the U.K. and Ireland three days later. 9 A global health advisory, advising U.S. citizens to avoid all international travel, was issued by the State Department on March 31. Also, the Centers for Disease Control and Prevention (CDC) has issued further guidelines for social distancing and personal protective measures (face covering, hand-washing, etc.) as part of a broader strategy for community mitigation while awaiting a vaccine or effective treatment.

A flurry of responses at the state and local levels also ensued. At the local level, one of the first actions taken by many jurisdictions was to declare a state of emergency, which typically frees up the state's office of emergency management to deploy resources to localities for immediate assistance. 10 The power for imposing the strongest mitigation and suppression policies lies with state and local authorities. Consequently, following the declaration of emergency, many states and jurisdictions started closing schools and shutting down non-essential businesses and services. 10 mortality from COVID-19, may take up to an additional 8 days , Wu et al. 2020 , Zhou et al. 2020 . 14 Other indirect behavioral pathways may also explain a link between SIPOs and coronavirus-related cases and deaths. For instance, SIPOs may affect confirmed cases by affecting selection into testing. Attempting to comply with the stay-at-home order or because of fear of getting exposed at medical facilities, infected persons who are unaware of their status may choose not to seek out medical care. Conditional on infection, SIPOs may also affect coronavirus-related mortality by reducing the demand for non-essential or elective medical procedures, thereby freeing up resources for care of COVID-19 patients.

This discussion underscores several key points that guide our empirical analyses. First, the incubation period for the virus and the lag from presentation of symptoms to ARDS imply important dynamics. SIPOs would not be expected to immediately dampen the growth curve given these dynamics, and strong effects may take some time to materialize (> 5 days for cases, and perhaps at least 14 days for deaths). Second, given that the effectiveness of SIPOs is driven by an increase in social distancing, this effectiveness may be moderated by factors such as urbanicity and population density that play in integral role in the spread of infections across communities. In other words, urbanicity and population density may serve as multipliers which can enhance the efficacy of a given level of social distancing. Third, given the exponential progression of infections, the effects of social distancing may magnify and accelerate over time if enacted early (Florida 2020; Friedson et al. 2020) . This suggests that health benefits of SIPOs can vary depending on whether they were enacted early or late during the outbreak cycle. Our study provides among the first national evidence on the effectiveness of statewide shelter-in-11 home orders in promoting social distancing, in decreasing infection rates and coronavirus-related deaths, and potential heterogeneity in the response based on timing of enactment and state characteristics.

We begin our analysis by examining whether SIPOs affect social mobility, drawing daily state-level data on social distancing for the period March 8, 2020 to April 17, 2020 from SafeGraph, Inc. For our analysis, we leverage this firm's anonymized population movement dataset representing 45 million smartphone devices that have opted into location tracking. These data have recently been used by the Centers for Disease Control and Prevention (CDC) to gather information on the degree to which social distancing has been practiced by individuals in the United States following the COVID-19 outbreak (Lasry et al. 2020) . From these data we collect a state-by-day measure of the percent of the state population who remain at home for the entire day. Specifically, this shelter-in-place index is the percentage point change in the number of cell phones staying at home relative to the baseline of February 6, 2020 through February 12, 2020. 15 A person's home is defined as a 153-meter by 153-meter area that receives the most frequent GPS pings during the overnight hours of 6pm to 7am over a six-week period.

While this measure of social distancing is imperfectfor instance, it does not capture whether an individual engages in social distancing while outside the home or if someone works at nightit is plausible to expect that having a higher percentage of the population who is -fully‖ sheltering in place is positively correlated with rates of social distancing. Given that our 12 focus is on changes within states over time, any measurement error introduced in these measures cross-spatially will not affect our results. As the data are based on smartphone users, they are not representative of the population that does not own smartphones. 16 Over our sample period, 35.7 percent of the population reported staying at home at all times (see Appendix Table 1 ). On average, 42.3 percent of individuals stayed at home on days when a state had a SIPO in place. This compares to 28.7 percent on state-days when a SIPO was not in effect.

Turning to our main analysis, we draw a panel of state-specific daily counts of COVID-19 cases from March 8, 2020 through April 20, 2020. These data are collected by the Centers for Disease Control and Prevention (CDC) and made public by the Kaiser Family Foundation. 17 By April 20, 2020 there were a total of 778,328 positive screenings for COVID-19 in the United States. In Appendix Table 2 , we show the day on which the first confirmed reported coronavirus case (and death) occurred in each state. The first known confirmed case was in Washington on January 20 followed by Illinois on January 24, though scientific knowledge on initial coronavirus arrival in the United States is evolving and earlier cases may yet be discovered. The state with the last initial case of reported coronavirus was West Virginia on March 17. Deaths followed a similar pattern, with a lag, as expected from the coronavirus's incubation period (Lauer et al. 16 Data from the 2018 American Community Survey indicate that 90 percent of households included at least one smartphone user, and 81 percent of Americans owned a smartphone based on a survey by the Pew Research Center in 2019 (see: https://www.pewresearch.org/internet/fact-sheet/mobile/). Smartphone ownership increases with education and income, decreases with age, and is roughly similar across race/ethnicity. Analyses by SafeGraph show that their sample correlates very highly with the true Census populations (0.97 across U.S. counties) and with the proportion of the population across educational attainment and income levels (correlation of 0.99). See: https://www.safegraph.com/blog/what-about-bias-in-the-safegraph-dataset. 17 See data available here: https://www.nytimes.com/interactive/2020/us/coronavirus-us-cases.html 13 2020) and time from first symptom to ARDS (Wang et al. 2020) . In Appendix Table 1 , we show that the mean rate of coronavirus cases per 100,000 population over our analysis period was 45.9. Figure 1 shows state-specific coronavirus case growth over our sample window. New York and New Jersey are clear outliers, on case growth trends that are higher than any other state. Over our sample period, the average increase in cases per 100,000 was about 8 times higher in New York and New Jersey as compared to the other 49 states (26.9 daily cases per 100,000 vs. 3.4 per 100,000). This is owed to the spread of COVID-19 in the high population density cities of New York City, Newark, and Jersey City (Rosenthal 2020; Warren 2020) . Table 2 describe coronavirus-related mortality from March 8 through April 20. The average COVID-19 death rate was 1.6 per 100,000 population. Figure 2 suggests a delay in the growth of deaths, as compared to cases. Total deaths did not start rising (even in New York and New Jersey) until late March or early April. This lag is consistent with the time period from infection to death.

We collect statewide shelter-in-place orders from Mervosh et al. (2020) as well as from our own search of state orders. 18 Table 1 lists the set of SIPOs enacted over our sample period and Appendix Figure 1A shows maps depicting the geographic and temporal adoption of SIPOs.

California was the first state to adopt a shelter in place order on March 19, 2020. Following California, the first cluster of states to adopt SIPOs was in the Midwest and parts of the Northeast, as well as Louisiana. Notably, many of these states were also in the midst of COVID-14 19 outbreaks during that time. Later adopters of SIPOs were largely concentrated in the mid-Atlantic and upper Midwest.

As of April 20, 2020, 40 states and the District of Columbia had adopted statewide SIPOs. Among those states who had not adopted a SIPO, six (6) had adopted some limited shutdown orders that fell short of full SIPOs, including mandates to close non-essential businesses (Arkansas, Iowa, Kentucky, Massachusetts 19 , North Dakota, and Wyoming), and more narrowly targeted SIPOs, which apply only to elderly individuals and those with underlying health conditions (Kentucky and Oklahoma). Only Nebraska, South Dakota, and Utah had not adopted a SIPO, a limited shutdown order, or a targeted SIPO. 20

For our main analyses we use a difference-in-differences design to estimate the association between state SIPOs and COVID-19 cases. Specifically, we estimate: ln (COVIDCASE st ) = β 0 + β 1 * SIPO_0to5 st + β 2 * SIPO_6to9 st + β 3 * SIPO_10to14 st + β 4 *SIPO_15to19 st + β 5 *SIPO_20plus st + β 6 *LIMITORDER st + β 7 *PARTORDER st + β 8 *TRAVEL st + β 9 *EMERG st + β 10 *TEMP st + β 11 *PRECIP st + α s + γ t + α s *t + ε st (1)

where ln (COVIDCASE st ) is the natural log of the count of COVID-19 cases in state s on day t, SIPO_0to5 st is an indicator set equal to 1 for the period 0 to 5 days following a SIPO's adoption, SIPO_6to9 st an analogous indicator for the period 6 to 9 days following adoption, SIPO_10to14 st 15 similarly indicating a period of 10 to 14 days following adoption, SIPO_15to19 st is also an analogous indicator for the period 15 to 19 days following adoption, and SIPO_20plus st is a final analogous indicator for 20 or more days following adoption. We are particularly interested in the periods (i) 6 to 14 days following adoption and (ii) two or more weeks following adoption, as these represent the periods following the median and 99 th percentile thresholds, respectively, in the incubation window for COVID-19 (Lauer et al. 2020 In addition, α s is a set of state fixed effects to control for fixed differences across states in COVID-19 infections due to, for example, baseline hospital capacity differences, population density, the presence of an important airport hub, or baseline testing capacity; γ t is a set of day fixed effects to control for national factors that commonly affect state COVID-19 infections such as national travel restrictions, announcement of Federal guidelines, expansion of COVID-19 testing capacity, general awareness and proliferation of concern regarding COVID-19, or 21 The period from 1-14 days is the 99 percent confidence interval for the incubation period for coronavirus (Lauer et al. 2020 Lee and Solon 2011) . However, we view this as a reasonable tradeoff to make in order to establish event studies which satisfy the common pre-trends assumption for difference-indifferences models, and we view our estimates as a lower bound on the treatment effect. To be transparent, we also report results with alternative approaches to controlling for unmeasured state-specific linear trends, such as controls for census region-or census division-specific day effects to account for common unmeasured spatial shocks and treatment state-specific pre-policy trends.

Identification of our key coefficients of interest, β 1 to β 5 , comes from within-state variation in SIPO adoption. Over the period under study, 40 states and the District of Columbia adopted SIPOs (see Table 1 ). It is important to note that our estimates of β 1 to β 5 capture the impact of the SIPO itself over and above any impacts from general increases in social distancing and avoidance behaviors common to treated and untreated states.

In a standard difference-in-differences research design that is capitalizing on the treatment turning on at different times, as with the differential timing of SIPO adoption across states and over time, Goodman-Bacon (2018) shows that the treatment effect is a weighted Accepted Article average of all possible two-group and two-period DD estimators. In other words, the main DD treatment effect is identified off many -mini‖ experiments comparing: 1) earlyand late-adopting states with never-adopting states as controls; 2) early-adopting states with late-adopting states as controls; and 3) late-adopting states with early-adopting states as controls. In the presence of dynamic treatment effects, using early adopters as a control for the treated later adopters may underestimate the treatment effect. This is because the trajectory of the early-adopting states, at the time when the late-adopting states enact their own SIPOs, is still being affected by the policy (that is, by the SIPOs in the early-adopting states).

In this context, Goodman-Bacon (2018) suggests that it may be better to compare treated early adopters with yet untreated later adopters or never adopters, and compare treated later adopters with never adopters. At the same time, however, specifically when it comes to shelterin-place orders, non-adopting states may be different than states that issue such an order. 23 This suggests that it may also be important draw on variation just among the ever-adopters, excluding the never adopters from contributing any identifying variation.

These considerations guide our main analyses and supplementary checks. In order to assess the sensitivity of our estimates to these different sources of timing-based identification, we draw on subsets of the treated cohorts to identify our effectsessentially validating that our effects are not driven by utilizing one particular cohort of states as a counterfactual. While the general issues outlined are valuable in thinking about which groups are identifying the effects, and which groups may be potentially problematic, event studies help to assess and alleviate some of these concerns of bias in a standard difference-in-differences analysis.

In light of this, we place an added focus on flexible event study formulations. We augment Equation (1) to estimate separate lead coefficients associated with windows spanning 1 to 2 days prior to SIPO adoption, 3 to 4 days pre-adoption, 5 to 6 days pre-adoption, and 7 or more days pre-adoption. Event studies, because they rely on different states turning on and off as treated vs. control for different leads and lags, are more sensitive in manifesting through the lead effects any issues with problematic controls. In the context of somewhat tricky timingbased identification and potentially strong dynamic effects, flat pre-trends in the event study analysis become an important signal of the presence (or lack thereof) of these issues. 24 We are careful in assessing the pre-policy trends and accounting for potentially unmeasured timevarying shocks that may be correlated with policy adoption.

In order to produce unbiased estimates of the effects of SIPOs on social distancing and COVID-19-related health, the common trends assumption must be satisfied. Threats to identification include (i) state-specific time-varying unobservables correlated with SIPOadoption and the outcomes under study, and (ii) policy endogeneity, whereby social distancing trends or trends in COVID-case or death growth induce the adoption of SIPOs.

We take a number of tacks to address each threat. With regard to state-specific time-varying unmeasured heterogeneity, one important concern is that changes in COVID-19 testing may conflate the effects of SIPOs on COVID-19 cases. To address this issue, we measure data on testing from the COVID Tracking Project, compiled by The Atlantic and Related Sciences from state public health authorities. 25 The variable TESTS st measures the cumulative number of COVID-19 tests conducted in state s on day t. We then explore (i) whether SIPOs are associated with changes in log testing rates, and (ii) how the estimated coefficient β1 changes when we control for state-specific changes in testing.

To take another example, it may be that other unobserved COVID-related policies or voluntary behaviors are both related to SIPO adoption and impact social distancing or COVIDrelated health. 26 To address this possibility we (i) partial out other state COVID policies that may be contemporaneously adopted, and (ii) control for state-specific linear time trends, which capture any state-specific unobservables that trend linearly that may include voluntary tastes for social distancing or heterogeneous infection growth across states.

Endogenous adoption of SIPOs is an important concern. For instance, some jurisdictions may adopt SIPOs in response to a noticeably accelerating COVID-19 outbreak. While controlling for state-specific time trends is one important way to ameliorate the possibility of policy endogeneity leading to biased estimates, the event study analysis further allows us to examine and address whether pre-treatment COVID-19 case trends were common across jurisdictions.

To explore the association between statewide SIPOs and COVID-19-related deaths, we turn to a negative binomial model. As can be gleaned from Appendix 

where COVIDDEATH st is the count of COVID-19 related deaths in state s on day t. We include the same controls as model (1) and use state-level population as an exposure measure. In addition, we also utilize a Tobit regression model and Poisson regression model and find results that are qualitatively similar. 27

All regressions described above are weighted using the state population and standard errors are corrected for clustering at the state-level (Bertrand et al., 2004) .

We begin by examining the effect of statewide SIPOs on social distancing. Table 2 presents estimates of the relationship between state SIPOs and the percent of individuals who stayed at home throughout the day. In our most parsimonious specification, which includes state fixed effects, and day fixed effects we find that the enactment of a SIPO is associated with a 2.1 percentage-point increase in stay-at-home rate. 28 Column 2 adds controls for state linear trends (to account for potential differential voluntary social distancing in the pre-treatment period). The estimate is largely unchanged, indicating that the enactment of a SIPO is associated with a 2.2 27 These findings are available upon request. 28 As the baseline for the shelter-in-place index is the same for all states (week ending February 12, 2020), and the pre-SIPO trends are nearly identical, the treatment effect can be interpreted as the increase in the percent of households who are staying at home relative to its counterfactual in the post-SIPO period.

percentage-point increase in staying at home. This marginal effect represents a 9.4 percent increase relative to the mean pre-treatment stay-at-home rate among future SIPO-adopting states.

Event study analyses shown in Figure 3 suggest an interesting pattern of results. While there is some indication of a weak positive differential trend in the pre-treatment period, consistent with somewhat faster growth in voluntary social distancing in SIPO states relative to non-SIPO states, there was a sharp and steep relative increase in stay-at-home rates in treatment versus control states in the week following the policy's adoption. We interpret these findings as evidence that the SIPO had an important short-run impact on social distancing.

In Panel (a), which excludes the state linear trends, the increase in stay-at-home behavior continues to magnify over the post-treatment period. In Panel (b), however, which accounts for state trends, following the first week after SIPO enactment, the stay-at-home differential between treatment and control states experienced a slight decline before leveling off. This result can be interpreted in several ways. First, the finding could suggest that residents become complacent over time and reverted back to usual habits. Such an effect could have been exacerbated by -cabin fever,‖ a belief that a week was sufficient time for a SIPO to have worked, or diminishing marginal utility (or perhaps disutility) of family time such that facing the expected risk of coronavirus is rationally preferred to staying at home. However, trends in the percent of individuals sheltering-at-home are positive throughout the sample period for the SIPO adopting states, and do not indicate an absolute decline in such social distancing behaviors. Second, a SIPO might have led to short-run panic, including an overestimation of the risk of serious COVID-19 illness. Additional time to overcome the negative emotional shock of being ordered to shelter in place, along with gathering of more health information, may have led to a more accurate assessment of risk of contracting serious illness from venturing outside of one's home.

Third, those who were sheltering in place full-time may have learned appropriate precautions to take to increase safety while venturing away from their residences.

Of course, the explanation could also reflect factors other than behavioral responses by residents of SIPO-adopting states. For instance, a lagged increase or catch-up in voluntary social distancing by those in control states, perhaps in response to widespread SIPO adoption in other states or general proliferation of awareness and concern regarding COVID-19, may have led to greater convergence in rates of staying at home. 29 Consistent with this explanation, Sears et al.

(2020) find that mobility, travel behavior, and social interactions had declined considerably by the time states were implementing stay-at-home orders, though these orders also modified travel behaviors further and led to additional declines in mobility and interactions. It is important to note, however, that even if SIPOs merely accelerated sheltering in place in treated versus untreated states, and both sets of states achieved the same level of social distancing eventually, there may be meaningful benefits to SIPO-adoption in the longer term by slowing spread of the illness earlier in the outbreak cycle.

Next, we test the sensitivity of our estimates to controls for other predictors of social distancing. Specifically, we include controls for other COVID-19 orders (column 3), travel restrictions, disaster declaration (column 4), and weather controls (column 5). It is validating that the results remain largely unchanged. We find consistent evidence of a 2.1 to 2.3 percentagepoint increase in stay-at-home rates, representing daily increases in social distancing of about 9.3 29 Trends in Appendix Figure 1C are consistent with this interpretation, showing a general increase in social distancing across all states that accelerated initially and then slowed down. Rate of growth is monotonic with the timing of SIPO adoption, with relatively faster growth in social distancing experienced by the early-adopting states. Appendix Figure 1D shows the trends in stay-at-home behavior across event time, and suggest diminishing returns in the increase in social distancing among the early-adopting states, as the non-adopting states are somewhat catching upthus partly narrowing the differential.

to 10 percent relative to the mean stay-at-home rate. These findings are largely consistent with those of Friedson et al. (2020) and Sears et al. (2020) . 30 The dynamics of the SIPO effects on stay-at-home behaviors, presented in Panel II, confirm the pattern evident in the event study analysis (Figure 3 ). We find strong significant effects in sheltering-in-place full time within 14 days post-adoption, with effect magnitudes representing an 8-13 percent increase. The effect sizes slightly diminish after this point, reflecting an increase in voluntary social distancing in the control states, though the differential for the treated states remains positive throughout the post-treatment period and statistically significant up to 19 days following the adoption of a SIPO. (5) and (6) are statistically equivalent.

Finally, in column (7), we drop New York and New Jersey from the analysis sample.

These states are outliers with respect to both COVID-19 case levels and annual growth rates, 30 Friedson et al. (2020) find that California's shelter-in-place order resulted in a 9-12 percent increase in the percent of individuals staying at home. Sears et al. (2020) , using data on change in average distance traveled, change in visits to non-essential businesses, and changes in the rate of human interactions, based on cellular location data provided by Unacast, similarly find that the state shelter-in-place mandates reduced travel behaviors by approximately 6-11 percent. 31 Complete sheltering-in-place full time, our main measure, is an extreme margin. And, we view this social distancing measure as a litmus check on whether or not the mechanism is active, but not as the full story on social distancing behavior as there is potential for a great number of margins of response. There may be other forms or social distancing and avoidance behaviors, induced by the SIPO, that are being followed and which can drive containment in the infection rate. We also assessed effects on two other measures: 1) median hours spent at home; and 2) percent full-time work outside the home. These effects are consistent with the estimates for stay-at-home behavior, and also suggest a significant increase in time spent at home and a decrease in work outside the home.

Accepted Article 24 owed to outbreaks in the high-population density cities of New York City and Jersey City. From

March 8 through April 20, the average daily increase in COVID-19 cases per 100,000 population was around 8 times higher in New York and New Jersey as compared to the remaining 49 (26.9

daily cases per 100,000 population vs. 3.4 daily cases per 100,000 population). In addition, these states were also early adopters of SIPOs, perhaps in response to the gathering storm of outbreak.

When we drop these states from our sample, the estimated effect of SIPO adoption on stay-athome rates are unchanged.

We begin our coronavirus case analysis with a sample including 48 states and the District of Columbia, excluding the two states on a very different case growth trajectory, New York and New Jersey. However, we will return to these states shortly.

Event study analyses of COVID-19 cases are visually presented in Panels (a) and (b) of Figure 4 , and these prefigure our main findings. In Panel (a) there is some indication of a distant differential pre-treatment trend, with the coefficient on the lead for 7 or more days prior to SIPO adoption being statistically significant. Panel (b) accounts for the state linear trend controls. In doing so, we now find very little evidence of differential pre-SIPO COVID-19 case trends in treatment and control states during all periods prior to adoption. Each of the leads is statistically indistinguishable from zero and each point estimate is near zero. Moreover, in both panels, there is a clear break in the trend between the SIPO and non-SIPO states following the mandate.

Estimated case reductions accelerate over time, becoming largest after 20 days following enactment of a SIPO. These findings are consistent with a causal interpretation and with exponential growth in short-run health benefits during the period of the shelter-in-place order.

In Table 3 , we present difference-in-differences estimates of the effect of SIPO adoption on COVID-19 cases, based on equation (1). 32 In our most parsimonious specification, we find little evidence that COVID-19 cases were affected during the 5 days following a SIPO's enactment. This is not too surprising given that transmission may not be as common during an asymptomatic incubation period. 33 However, the estimated coefficients on the SIPO policy become much larger after 6 to 14 days. After 14 days of adoption, enactment of a SIPO is associated with a 20.5 percent decline in COVID-19 cases, and the effect strengthens to 39.2 percent 20 or more days post enactment (column 1), though these effects are imprecisely estimated. 34

In column (2) we utilize controls for state-specific linear time trends. We believe this specification is preferable for three reasons. First, in a model with log cases as the outcome, state linear trends control for the state-specific exponential growth path of the outbreak (at least prior to reaching the peak number of cases), making the estimated effect of the SIPO deviations from that growth path. Second, with regard to the common trends assumption, the specification including state-specific linear time trends is more defensible when examining event studies, as the pre-treatment trends are generally flatter. 35 Third, the trend controls appear to soak up residual variance and improve precision of the estimated coefficients.

In column (2), we find that between 6-and 9-days following enactment of a SIPO, there was a 27.3 percent decline in COVID-19 cases, an effect that is statistically distinguishable from zero at the 10 percent level. Moreover, 10 to 14 days after SIPO adoption, we find that 26 coronavirus cases fell by 35.6 percent. The effect progressively gets stronger with the length of the post-treatment window, and after 20 days subsequent to SIPO adoption, we find that coronavirus cases fell by 52.2 percent, suggesting that the health benefits of SIPOs may grow larger in the periods following enactment.

The remaining columns of Table 3 show the robustness of the COVID-19 case results to observable state-level observable controls. 36 We find no evidence that other COVID-19-related shutdown or shelter policies (column 3), travel restrictions or major disaster emergency declarations (column 4), or weather (column 5) affected the estimated impact of SIPOs on COVID-19 cases. 37

In Table 4 , we explore the sensitivity of our main findings to the inclusion and exclusion of states from the analysis sample. Above, we argued for the exclusion of New York and New Jersey from our sample. In column (1) of Table 4 , we include these states in the analysis sample.

The estimated effect of SIPO adoption on coronavirus cases is similar to that reported in column (5) of Given that California's SIPO is significantly associated with a decline in coronavirus cases , we next excluded this state from the analysis to ensure that our findings were not driven by the earliest-adopting state for which there is already strong evidence 36 In alternate specifications, available upon request, we estimate negative binomial and tobit regressions of the effect of SIPOs on COVID-19 cases. The findings are qualitatively unchanged from those we report. 37 In Appendix Table 4 , we explore the lagged effects of non-essential business closure policies, which was a common policy adopted just prior to full SIPOs or instead of SIPOs. We find no evidence that non-essential business closure policies affect the number of COVID-19 cases. 38 Appendix Figure 3 shows an event-study analysis when we include New York and New Jersey in the main analysis sample. Note that as we approach the period over one week prior to adoption, the coefficient estimates decrease, suggesting a slight pre-treatment upward trend in cases when we include New York and New Jersey to our sample.

for SIPO-induced COVID-19 case reductions. Our results show that the average SIPO effect we detect is not driven by California, and continue to exhibit a similar pattern of case declines that grow larger in the weeks following enactment.

In columns (3) through (7), we exclude a number of states with high COVID-19 case levels (relative to the national mean), as well as states with relatively high COVID-19 case rates.

The states we drop include Washington (column 3), Massachusetts (column 4), Louisiana (column 5), District of Columbia (column 6), and Connecticut (column 7). The results show that the average SIPO effect we detect is not also driven by these states. We estimate that, after 20 days of enactment, SIPO adoption is associated with a 50.1 to 55.0 percent reduction in COVID-19 cases when we exclude these states from our sample.

In the final three columns of Table 3 , we drop several states with lower rates of coronavirus and low rates of coronavirus case growth: Oregon (column 8), Texas (column 9), and Minnesota (column 10). Again, we find no evidence that our main finding is changed.

One concern with the estimates presented thus far is that they may be biased if SIPO adoption were correlated with COVID-19 testing capabilities. This may be the case due to the evolution of testing over the period of analysis. As of March 13, only 15,000 tests had been conducted in the U.S. To address the low testing rate in the U.S., the Food and Drug

Administration approved a new COVID-19 test from the pharmaceutical company Roche (Arnold 2020). In the following days, Delaware, New York, Massachusetts and Texas, began implementing drive-up testing sites, which made testing more accessible (Yancey-Bragg 2020).

Despite these improvements in accessibility, many testing delays persisted due to laboratory capacity constraints (Brown and Court 2020).

This article is protected by copyright. All rights reserved.

SIPOs could affect COVID-19 testing in several ways. First, SIPOs may induce some who have flu-like symptoms to stay at home rather than seek medical attention, either due to perceived civic duty or a perception of greater adverse selection in patients who present at The first two columns of Table 5 show estimates of the effect of SIPOs on the natural log of COVID-19 tests. We find that SIPOs are negatively related to testing, but these effects are mainly not statistically distinguishable from zero at conventional levels. In light of these findings, it is perhaps not surprising that in columns (3) and (4) of Table 5 , we find no evidence that the estimated effect of enactment of a SIPO on COVID-19 cases is affected by the addition of a control for COVID-19 testing capacity. Given that testing is a potentially endogenous control (though we find little evidence that it is affected by a SIPO), we nevertheless take this descriptive evidence as suggestive of the hypothesis that our estimates are not biased due to state-level heterogeneity in growth of testing capacity.

The difference-in-differences estimates presented to this point have identified the effect of a SIPO on the state-specific change in cumulative cases of coronavirus. In Table 6 , we explore the effect of SIPO adoption on the -derivative‖ of cumulative COVID-19 cases, that is,

This article is protected by copyright. All rights reserved.

daily COVID-19 cases. And, in fact, we find evidence that the enactment of a SIPO also affected the rate of change in cumulative COVID-19 cases. The results suggest that state adoption of a SIPO was associated with a 51.1 to 54.1 percent decline in daily coronavirus cases after 20 or more days.

We undertake additional analyses to address specific issues. First, there may be some concern that the composition of states that identify each of the lagged effects is changing over time. We have at least 14 days of post-treatment data for all SIPO states. Furthermore, 29 of the 41 treated states contribute to identification of the coefficient of the longest lag window, that is effects over 20 or more days following adoption. Appendix Figure 2 plots the event study

analyses on a restricted sample of states that have 20+ days of post-treatment data following the adoption of a SIPO, showing largely similar results and dynamics.

Second, our main conclusions are not materially affected by alternate approaches to controlling for unmeasured time-varying state factors. Appendix Table 5 presents estimates that control for region by day and census division by day fixed effects. Comparing these estimates to our main analyses, we draw the following empirical conclusions. Our interpretation of the pattern of results across all models is similar: the estimated effect of SIPOs on COVID-19 cases grows over the post-treatment period, becoming largest in the window three weeks or more following enactment. 39 39 In Appendix Tables 6 and 7, we present analyses that utilize alternate controls for unobserved state trends. Appendix Table 6 utilizes a two-part correction, which estimates state-specific trends (separately for each state and alternately for each treatment cohort) using only the pre-policy observations and then partials out these extrapolated pre-policy trends from the post-policy periods (Goodman-Bacon 2018). These estimates are the largest (in absolute magnitude), indicating a 61-62 percent decline in COVID-19 cases, and a 85-87 percent decline in mortality three or more weeks following adoption. While this finding is consistent with the hypothesis that estimated policy effects from models including the state-specific linear time trend may understate the policy's impact, with an exponentially growing outbreak, extrapolating the pre-treatment trends forward using the pre-policy periods may also risk biasing the dynamics in the opposite direction if the intervention is effective and dampens the trajectory to sub-exponential growth. Moreover, the extrapolated functional form yields pre-trends that violate parallel trends in event study analyses. For instance, the large 34 to 37 percent increase in COVID-related mortality within 0-5 following SIPO Accepted Article 30 Third, in order to assess the sensitivity of our estimates to different sources of timingbased identification, we draw on subsets of the treated cohorts to identify our effects, in turn excluding late adopters, early adopters, and non-adopters from the sample). These results confirm that: 1) the effects of SIPOs in reducing COVID-19 cases are substantially larger among early-adopting states relative to later-adopting states; 2) the effects for the early adopters are progressively larger with the length of the post-treatment window; 3) these effects are robust to using non-adopters as controls or late-adopters as controls; and 4) the pattern of results is largely similar to those discussed above, utilizing all states in the analysis.

In our main analyses, we weight the models by state population to produce an average treatment effect at the population level. We also produce unweighted estimates in Appendix Table 8 to assess sensitivity to weighting, which may arise, for instance, if there are heterogeneous effects across larger vs. smaller treatment states (Solon, Haider and Wooldridge, 2015; Angrist and Pischke 2008) . 40 Our results and patterns remain largely similar in the unweighted models.

Next, we turn to an analysis of whether statewide SIPOs affected mortality. There are several channels through which mortality may be affected by SIPOs. If SIPOs reduce coronavirus cases, mortality will decline in the longer-run because fewer people will become adoption is not credible and reflects a continuation of existing positive pre-trends that are not effectively netted out with the extrapolated trend controls. Appendix Table 7 presents estimates that control for treatment state-specific linear pre-policy trends. While these models indicate similar patterns as with our main analyses, they also fail to generate credible pre-trends. The significant decrease, in this case, in mortality of 17-23 percent within the first few days of the enactment of the SIPO is a carry-over of existing pre-policy trends that are again not effectively being eliminated with these trend controls. In summary, while the patterns are consistent with our main analyses, these alternate approaches to controlling for unmeasured time-varying state factors fail to produce parallel pre-policy trends between the treated and control states and should be interpreted with much caution. 40 Appendix Figure 4 presents the event study analyses for the unweighted specifications.

This article is protected by copyright. All rights reserved.

infected with the coronavirus. Of course, these effects are likely to come with a much longer lag than cases given that the time from first symptoms until acute respiratory distress syndrome (ARDS) is, on average, about 8 days ). In addition, SIPOs may affect the likelihood that infected patients choose to stay at home rather than seek out testing and other medical care, having the unintended consequence of increasing serious illness and death.

Finally, SIPOs may also impact the availability of resources for medical care, as public resources are used to enforce SIPOs instead.

Negative binomial regressions of the effect of SIPO enactment on COVID-19-related mortality are shown in Table 7 . Models without state trends of any sort (column 1) and with state-specific linear time trends (column 2) suggest that after 20 days, SIPO adoption is associated with 36.1 to 50.2 percent reduction in mortality. Controlling for state-specific linear time trends, the estimated mortality effects are smaller, but continue to show long-run COVID-19 death declines. But because none of these estimates are statistically distinguishable from zero at conventional levels, we can only cautiously interpret these findings as evidence of mortality declines. We note here that due to the longer lag with which we expect mortality effects to materialize, the effects for the longer time windows (15-19 days post SIPO, > 19 days post SIPO) are identified off a few early-adopting states. Hence, more long-run data is necessary for a definitive conclusion.

In Figure 4 , Panels (c) and (d), we show the event study analysis on COVID-19 related mortality. We find no evidence of differential pre-treatment trends in mortality, with a longer delayed but imprecisely estimated potential decline in mortality.

Accepted Article This article is protected by copyright. All rights reserved.

The results presented above provide consistent evidence that the adoption of a statewide shelter-in-place order significantly reduced infection rates, with the strongest effects realized two or more weeks after enactment. This lag is consistent with the incubation period of the virus (2-14 days) over which transmission may be possible but less efficient due to lack of symptoms.

Given the exponential growth trajectory of infection, there may additional dynamics in terms of the effectiveness of the SIPO depending on when the policy is enactedwhether early or late over the cycle of disease progression. That is, the beneficial effects of social distancing on COVID-19 caseload may have an accelerating effect over time if enacted early. Friedson et al. (2020) find strong evidence that California's first-in-the-nation SIPO had a strong public health benefit, and continued to do so even after social distancing measures between California and its control narrowed. This is suggestive of persistent and magnified effects of enacting a SIPO early in the outbreak cycle. However, given the study's focus on California, Friedson et al. (2020) were not able to explicitly test for heterogeneity in the response across early vs. late adopters. 41 Table 8A presents effects separately across early adopting states (adopted on March 25 or earlier) and late adopting states (adopted after March 25). These results confirm that the effects are primarily driven by states which enacted the shelter-in-place orders relatively early, thus capitalizing on the magnified benefits of social distancing as the growth trajectory was rising but still relatively low compared with later adopters. 42 In addition, we now uncover some evidence that SIPOs are effective at reducing coronavirus-related mortality when they were adopted early in the U.S. COVID-19 outbreak. While some of this result could be 41 The states within each percentile group used to subset the data in the analyses to follow can be found in Appendix Table 9 . 42 Appendix Table 10 explores heterogeneity in the efficacy of SIPOs based on whether or not the state in question had a nonessential business closure or local SIPOs covering over 50 percent of the state population in place prior to the statewide SIPO. Statewide SIPOs are less effective when these other policies are enacted first. For further discussion of this phenomenon see Dave et al. (2020) .

This article is protected by copyright. All rights reserved.

have data between 2 and 3 weeks after SIPO adoption and early adopters see substantially larger health benefits than later adopters over this period. 43

In Table 8B , we explore whether COVID-19-related health benefits are larger among states with higher population density. To do this, we explore the interquartile range of population density rankings of U.S. states. 44 In the main, our findings suggest larger COVID-19related health benefits among those outside the lower 25 th percentile of state population density rankings. Those states in the middle 50 th percentile and upper 25 th percentile of population density tend to see larger reductions in COVID-19-related cases. This result is consistent with the hypothesis that stay-at-home orders are likely to generate greater health benefits when crowd-related contagion is avoided. The relationship between population density and the marginal effect of SIPOs on mortality appears more non-monotonic.

Together, the findings in Tables 8A and 8B generally suggest compounding and stronger effects of social distancing among the early adopters and higher population density states and provide some explanation as to why the SIPOs continue to have a beneficial effect on infection rates 6 to 20+ days post-enactment. 45 In Table 8C , we test this hypothesis by presenting estimates of the effect of SIPOs on stay-at-home rates by (i) whether the enacting state was an early or later adopter, and (ii) the interquartile range of population density. 43 Appendix Table 11 conducts a similar analysis to Table 8A , instead splitting states into groups based on their average daily growth rate of COVID-19 cases prior to their SIPO as a proxy for how early or late their adoption was relative to their state-specific outbreak. Results show that states which adopted relatively earlier in their outbreak cycle realized stronger effects. 44 We obtain our population density measures here: https://www.statista.com/statistics/183588/population-densityin-the-federal-states-of-the-us/ 45 Event-study analyses on early-adopting and more densely populated states, available upon request, produces a pattern consistent with coronavirus case differentials that trend near zero in the pre-treatment period and fall precipitously following SIPO enactment. Furthermore, there is sufficient variation in the distribution of population density across the early adopting states. Models that further separate out the effects of early adoption from population density indicate that states which adopted earlier and are more densely populated experienced the strongest decrease in COVID-19 cases. We lose precision in these models however.

First, consistent with our health results our findings in Panel I provide strong evidence that SIPOs were more effective at increasing social distancing in early adopting SIPO states as compared to later adopting states. The estimated social distancing effect was twice as large for early as compared to later-adopting states (2.6 percentage-points vs 1.3 percentage-points). One explanation for our finding may be that residents of later-adopting SIPO states already adopted some social distancing behaviors prior to their SIPO adoption or because later-adopting SIPO state residents were less compliant with social distancing mandates, perhaps because of political or ideological preferences (Painter and Qiu 2020).

Next, if the primary mechanism through which SIPOs impact case rates is through social distancing, then we would expect these benefits to be further magnified in states with a high population density. Population density is an independent predictor of community spread of infection; SIPOs therefore would have more capacity to reduce social distancing in these states, and a given decrease in social distancing would also translate into reduced modes of person-toperson contact. In Panel II Table 8C , we assess heterogeneity in the effects on stay-at-home behaviors across states with lower versus higher rates of population density, as measured by population per state area in square miles. Consistent with our findings on COVID-19, we find that SIPOs are more effective in promoting social distancing in more densely populated states.

The patterns uncovered in Table 8C , with respect to heterogeneity in social distancing, parallel the stronger health effects that we find among states that adopted the orders earlier and among more densely populated states.

Since March 19, 2020, 40 states and the District of Columbia have adopted shelter-inplace orders to attempt to hasten the spread of COVID-19, smooth illnesses and treatment over time, and prevent short-run medical capacity constraints from causing otherwise avoidable deaths. Critics, however, argue that SIPOs impose large, and perhaps long-lasting, job loss, strain social insurance programs beyond their capacity, and slow the progression toward community-level herd immunity. While the lack of long-run data does not yet permit us to answer all of the claims made by proponents and opponents of SIPOs, this paper proposes an important first step: to estimate the short-term impact of statewide SIPOs on COVID-19 cases and COVID-19-related mortality. That is, did these policies meet their immediate public health objective?

First, using GPS-based anonymized cell phone records from SafeGraph, Inc. to measure state-level daily social mobility, we find evidence that the enactment of a state SIPO resulted in a short-run increase in the percent of state residents remaining at home full time. While longer-run differentials in stay-at home rates for treatment and control states were smaller or slightly negative, the findings do suggest an important effect of SIPOs on social distancing. We view the social distancing results as an important first-stage outcome, but by no means the complete story with regards to policy and social distancing behavior. Our results capture a single, but important measure of social distancing, and studying other metrics of mobility (as done by Sears et al. 2020 ) will likely be a fruitful way to push this literature forward.

Second, using coronavirus case and mortality data from March 8, 2020 to April 20, 2020, and a difference-in-differences approach, we find that approximately three weeks after the enactment of a statewide SIPO, the average number of cumulative cases fell by approximately 53.5 percent. Roughly 3 weeks following SIPO adoption, this corresponds to approximately Accepted Article 3,073 fewer cumulative COVID-19 cases for the average SIPO-adopting state. Event study analysis suggests that this result was not driven by differential pre-treatment trends, nor was it driven by the California experience . We find that states that were highly urbanized and were early SIPO adopters saw the largest declines in cases, consistent with the hypothesis that early adopters with high population densities had the largest margins for larger short-run public health benefits. 46 With the U.S. seeing a resurgence in new COVID-19 cases, these results on the effectiveness of state orders targeting social distancing have continued relevance as states are starting to suspend or roll-back their reopening plans, and making decisions on re-imposing orders for non-essential business closures and sheltering-at-home.

An important limitation of our work is the preliminary examination of COVID-19-related deaths. While it is true that detecting more consistently negative death effects of SIPOs may take more time to uncover, effects on COVID-related deaths could be quite different from cases.

For instance, it could be that SIPOs may have important case-reducing effects mostly among populations at low-risk for mortality from this disease. In that case, the avoided infections may not generate substantial reduction in mortality because the high-risk infections, i.e. in nursing homes, still occurred. Future work might explore characteristics of marginal individual who avoided a COVID-19 infection due to a SIPO.

The larger efficacy of a statewide SIPO depends on a number of factors beyond the specific scope of our paper. But if we assume (i) SIPOs are temporary because the economic costs are substantial, (ii) development of a vaccine or effective treatment for COVID-19 is unlikely in the short-run (Fauci 2020) , and (iii) universal (or at least widespread) testing of 46 In this context, that our average estimate is smaller than the estimated effect of CA statewide SIPO ) is to be expected. As indicated by our heterogeneity analysis, early-adopting and more densely populated statessuch as CAstand more to gain from enacting a SIPO early during the outbreak cycle. Courtemanche et al. (2020a Courtemanche et al. ( , 2020b find results that are larger in magnitude by examining all social distancing policies in tandem (including the SIPO) in Kentucky and nationally.

Accepted Article 37 asymptomatic individuals is also very unlikely in the short-run (Center for Disease Control and Prevention 2020d), some of the short-run COVID-19 cases and deaths may simply be postponed to the near future when the SIPO is lifted. In that case, deaths and serious illnesses averted by avoiding short-run shortages of ventilators, hospital beds, and medical professionals may be a SIPO's most likely path to generating long-run public health benefits. A business closure order is an indicator for whether the state had issued a non-essential business closure order (that fell short of a SIPO). Partial SIPOs include a targeted SIPO for older individuals or those with underlying health conditions and an indicator for whether coverage of local (i.e. city or county) SIPO orders covered at least 50 percent of the state population. A travel restriction is an indicator for whether the state had issued restrictions on travel to or from the state. A major disaster declaration is an indicator for whether the state had received a major disaster emergency declaration from the Federal government. Finally, weather controls include the average temperature (in degrees Celsius) in the state and an indicator for whether measurable precipitation fell in the state. All models include state fixed effects and day fixed effects. Standard errors, clustered at the state-level, are reported in parenthesis.

This article is protected by copyright. All rights reserved. Notes: Estimates are obtained using weighted least squares regression. A business closure order is an indicator for whether the state had issued a non-essential business closure order (that fell short of a SIPO). Partial SIPOs include a targeted SIPO for older individuals or those with underlying health conditions and an indicator for whether coverage of local (i.e. city or county) SIPO orders covered at least 50 percent of the state population. A travel restriction is an indicator for whether the state had issued restrictions on travel to or from the state. A major disaster declaration is an indicator for whether the state had received a major disaster emergency declaration from the Federal government. Finally, weather controls include the average temperature (in degrees Celsius) in the state and an indicator for whether measurable precipitation fell in the state. All models include state fixed effects and day fixed effects. Standard errors, clustered at the statelevel, are reported in parenthesis. Notes: Estimates are obtained using weighted least squares regression. A business closure order is an indicator for whether the state had issued a non-essential business closure order (that fell short of a SIPO). Partial SIPOs include a targeted SIPO for older individuals or those with underlying health conditions and an indicator for whether coverage of local (i.e. city or county) SIPO orders covered at least 50 percent of the state population. A travel restriction is an indicator for whether the state had issued restrictions on travel to or from the state. A major disaster declaration is an indicator for whether the state had received a major disaster emergency declaration from the Federal government. Finally, weather controls include the average temperature (in degrees Celsius) in the state and an indicator for whether measurable precipitation fell in the state. All models include state fixed effects and day fixed effects. Standard errors, clustered at the state-level, are reported in parentheses. Notes: Estimates are obtained using weighted negative binomial regression. A business closure order is an indicator for whether the state had issued a non-essential business closure order (that fell short of a SIPO). Partial SIPOs include a targeted SIPO for older individuals or those with underlying health conditions and an indicator for whether coverage of local (i.e. city or county) SIPO orders covered at least 50 percent of the state population. A travel restriction is an indicator for whether the state had issued restrictions on travel to or from the state. A major disaster declaration is an indicator for whether the state had received a major disaster emergency declaration from the Federal government. Finally, weather controls include the average temperature (in degrees Celsius) in the state and an indicator for whether measurable precipitation fell in the state. All models include state fixed effects and day fixed effects. Standard errors, clustered at the state-level, are reported in parentheses. N 2100 2156 * Significant at the 10% level, ** Significant at the 5% level, *** Significant at the 1% level.

Notes: Estimates in column (1) are obtained from weighted least squares regression. Estimates in column (2) are obtained from a weighted negative binomial regression. All models include the following controls: an indicator for whether the state had issued a non-essential business closure order (that fell short of a SIPO) or a targeted SIPO for older individuals or those with underlying health conditions, an indicator for whether coverage of local (i.e. city or county) SIPO orders covered at least 50 percent of the state population, an indicator for whether the state had issued restrictions on travel to or from the state, an indicator for whether the state had received a major disaster emergency declaration from the Federal government, the average temperature (in degrees Celsius) in the state, an indicator for whether measurable precipitation fell in the state, state fixed effects, day fixed effects, and a state-specific linear time trend. Standard errors, clustered at the state-level, are reported in parentheses. States that enacted SIPO between March 19 and 25 are coded as early adopting states. States that enacted SIPO on March 26 or later are coded as late adopting states. Notes: Estimates in column (1) are obtained from weighted least squares regression. Estimates in column (2) are obtained from a weighted negative binomial regression. All models include the following controls: an indicator for whether the state had issued a non-essential business closure order (that fell short of a SIPO) or a targeted SIPO for older individuals or those with underlying health conditions, an indicator for whether coverage of local (i.e. city or county) SIPO orders covered at least 50 percent of the state population, an indicator for whether the state had issued restrictions on travel to or from the state, an indicator for whether the state had received a major disaster emergency declaration from the Federal government, the average temperature (in degrees Celsius) in the state, an indicator for whether measurable precipitation fell in the state, state fixed effects, day fixed effects, and a state-specific linear time trend. Standard errors, clustered at the state-level, are reported in parentheses. N 2091 * Significant at the 10% level, ** Significant at the 5% level, *** Significant at the 1% level.

Notes: Estimates are obtained using weighted least squares regression. The models include the following controls: an indicator for whether the state had issued a non-essential business closure order (that fell short of a SIPO) or a targeted SIPO for older individuals or those with underlying health conditions, an indicator for whether coverage of local (i.e. city or county) SIPO orders covered at least 50 percent of the state population, an indicator for whether the state had issued restrictions on travel to or from the state, an indicator for whether the state had received a major disaster emergency declaration from the Federal government, the average temperature (in degrees Celsius) in the state, an indicator for whether measurable precipitation fell in the state, state fixed effects, day fixed effects, and a state-specific linear time trend. Standard errors, clustered at the state-level, are reported in parentheses. States that enacted SIPO between March 19 and 25 are coded as early adopting states. States that enacted SIPO on March 26 or later are coded as late adopting states.

Accepted Article

The Immediate Effect of COVID-19 Policies on Social Distancing Behavior in the United States.‖ Available at SSRN

Bay Area Orders ‗Shelter in Place

Mostly harmless econometrics: An empiricist's companion

America Closed; Thousands of Stores, Resorts, Theaters Shut Down.‖ NPR

Australian Government Department of Health

At the Top of the Covid-19 Curve, How Do Hospitals Decide Who Gets Treatment?‖ The New York Times

To Enforce Coronavirus Distancing, Police Say Arrests Are Last Resort‖. ABC News

How much should we trust differences-in-differences estimates

Raffaela Di Napoli. 2020. -Features, Evaluation and Treatment Coronavirus (COVID-19).‖ NCBI

Stay-At-Home Order Now in Effect: What You Need to Know.‖ Dayton Now

Center for Disease Control and Prevention. 2020a. -Cases in U.S.‖ Available at

COVID-19): How It Spreads

sick/how-covidspreads.html?CDC_AA_refVal=https%3A%2F%2Fwww.cdc.gov%2Fcoronavirus%2F2019

Centers for Disease Control and Prevention. 2020c. -Social Distancing, Quarantine, and Isolation‖ Available at

Infection Control Guidance for Public Health Personnel Evaluating Persons Under Investigation (PUIs) and Asymptomatic Close Contacts of Confirmed Cases at Their Home or Non-Home Residential Settings

2020 -Governor Charlie Baker Orders All Non-Essential Businesses To Cease In Person Operation, Directs the Department of Public Health to Issue Stay at Home Advisory For Two Weeks‖

Did Social-Distancing Measures in Kentucky Help to Flatten the COVID-19 Curve?

Strong Social Distancing Measures In The United States Reduced The

Growth Rate.‖ Health Affairs

Weekly Jobless Claims Double to 6.6 Million.‖ CNBC

Were Urban Cowboys Enough to Control COVID-19? Local Shelter-in-Place Orders and Coronavirus Case Growth

Congress livid over lags in coronavirus testing.‖ Politico

Long Before COVID-19, Dr. Anthony Fauci 'Changed Medicine In America Forever

Rapid Expert Consultation on the Possibility of Bioaerosol Spread of SARS-CoV-2 for the COVID-19 Pandemic‖ The National Academies of Science, Engineering, Medicine

Bay Area Coronavirus: How Life Has Changed After 20 Days of Shelter-In-Place Order.‖ San Francisco Chronicle

Did California's Shelter-in-Place Order Work? Early Coronavirus-Related Public Health Effects.‖ NBER Working Paper No

As Coronavirus Spreads, Los Angeles County Scores an A in Social Distancing

social-distancing-according-to-a-gps-tracking-project Food and Drug Administration. 2020. -Coronavirus (COVID-19) Update: FDA Continues to Accelerate Development of Novel Therapies for COVID-19

Difference-in-Differences with Variation in Treatment Timing.‖ NBER Working Paper No. 25108

2020 -Directive to the State of Utah: The Governor's Coronavirus Directive for Utah -Stay Safe, Stay Home‖‖

New York City is Social Distancing‖

Coronavirus in NY: City Has Fewer Than 400 Free Intensive Care Beds.‖ New York Post

First case of 2019 novel coronavirus in the United States

House Oversight and Reform Committee. 2020. -Hearing on Coronavirus Response

Doctor in New York Details Dire Supply Shortages From the Front Lines of the Coronavirus Fight.‖ CNN

Hospitals Prepare for a COVID-19 Wave.‖ National Public Radio

Timing of Community Mitigation and Changes in Reported COVID-19 and Community Mobility -Four

The Incubation Period of Coronavirus Disease 2019 (COVID-19) from Publicly Reported Confirmed Cases: Estimation and Application

The fragility of estimated effects of unilateral divorce laws on divorce rates

Transmission of COVID-19 In The Terminal Stage of Incubation Period: A Familial Cluster

Infected but Feeling Fine: The Unwitting Coronavirus Spreaders.‖ The New York Times

What's a ‗Shelter In Place' Order, and Who's Affected?‖ Wired

Order of the Governor of the State of Maryland Number

See Which States and Cities Have Told Residents to Stay at Home.‖ The New York Times

Emergency Executive Order 20-20.‖ Available at

Police Hoping Public Adheres to Holcomb's Shelter-In-Place Order, But Will Be Vigilant About Enforcement.‖ Chicago Tribune

Political Beliefs affect Compliance

PPE: Types of Personal Protective Equipment Used to Combat COVID-19.‖ USA Today

The Pros and Cons of Sick Pay Schemes: Testing for Contagious Presenteeism and Noncontagious Absenteeism Behavior

2020. -Positive Health Externalities of Mandating Paid Sick Leave

Guidance on Social Distancing for Everyone in the U.K.‖ Available at

Public Health Agency of Canada. 2020. -Community-Based Measures to Mitigate the Spread of Coronavirus Disease (COVID-19) in Canada

As Coronavirus Numbers Rise, C.D.C. Testing Comes Under Fire.‖ The New York Times

Critical Supply Shortages -The Need for Ventilators and Personal Protective Equipment during the Covid-19 Pandemic

2020. -California Governor Issues Statewide Stay-At-Home Order‖ AP News

Density Is New York City's Big ‗Enemy' in the Coronavirus Fight.‖ The New York Times

Coronavirus: Half of Humanity Now on Lockdown as 90 Countries Call for Confinement.‖ Euronews

Miguel and Villas-Boas, Vasco and Villas-Boas, Sofia, Are We #StayingHome to Flatten the Curve?

-An Interactive Map Shows How Overloaded Hospitals in the US Could Get as Coronavirus Cases Grow, Region by Region.‖ Business Insider

Hospitals Increasingly Worried About Surge in COVID-19 Cases.‖ Time

The Lost Month: How a Failure to Test Blinded the U.S. to Covid-19.‖ The New York Times

What are we weighting for?

Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects‖ Working Paper

American Hospital Capacity and Projected Need for COVID-19 Patient Care.‖ Health Affairs

News Release: Unemployment Insurance Weekly Claims

Estimates of The Severity of Coronavirus Disease 2019: A Model-Based Analysis.‖ The Lancet: Infectious Diseases

Here's Why New Jersey and New York Are The Epicenter of The Coronavirus Pandemic.‖ NJ.com

CDC Director On Models For The Months To Come: ‗This Virus Is Going To Be With Us'.‖ National Public Radio

The President's Coronavirus Guidelines for America

Oklahoma Governor Kevin Stitt Issues ‗Safer-At-Home

Residents Snitch on Businesses, Neighbors Amid Shutdowns.‖ The Associated Press

The President's Coronavirus Guidelines for America

World Health Organization. 2020a. -Rolling Updates on Coronavirus Disease (COVID-19).‖ Available at

World Health Organization. 2020b. -Q&A: Similarities and Differences -COVID-19 and Influenza.‖ Available at

World Health Organization. 2020c. -Shortage of Personal Protective Equipment Endangering Health Workers Worldwide

Chunxue Bai, Junhua Zheng, and Yuanlin Song. 2020. -Risk Factors Associated With Acute Respiratory Distress Syndrome and Death in Patients With Coronavirus Disease

Characteristics of and Important Lessons From the Coronavirus Disease 2019 (COVID-19) Outbreak in China: Summary of a Report of

Clinical Course and Risk Factors For Mortality of Adult Inpatients With COVID-19 In Wuhan, China: A Retrospective Cohort Study

This article is protected by copyright. All rights reserved.

This article is protected by copyright. All rights reserved.

This article is protected by copyright. All rights reserved.

This article is protected by copyright. All rights reserved.