key: cord-0285535-h9v7rkln authors: Avery, A. J.; Wang, J.; Ma, X.; Pan, Q.; McGrady, E. E.; Yuan, Z.; Liang, Y.; Nugent, R.; Lakdawala, S. S. title: Variations in Non-Pharmaceutical Interventions by State Correlate with COVID-19 Disease Outcomes date: 2021-07-31 journal: nan DOI: 10.1101/2021.07.28.21261286 sha: b0eb4854f165aa390916da08d10aad57915c25b3 doc_id: 285535 cord_uid: h9v7rkln The COVID-19 pandemic highlighted the lack of understanding around effective public health interventions to curtail the spread of an emerging respiratory virus. Here, we examined the public health approaches implemented by each state to limit the spread and burden of COVID-19. Our analysis revealed that stronger statewide interventions positively correlated with fewer COVID-19 deaths, but some neighboring states with distinct intervention strategies had similar SARS-CoV-2 case trajectories. Additionally, more than two weeks is needed to observe an impact on SARS-CoV-2 cases after an intervention is implemented. These data provide a critical framework to inform future interventions during emerging pandemics. estimated basis-spline models underlying the state time series revealed geographical groupings, indicating that similar SARS-CoV-2 case trajectories may be based on state location in addition to 48 NPIs. Finally, we observed a statistically significant negative correlation between COVID-19 deaths and state NPI score, solidifying the importance of implementation of public health 50 interventions. 52 Heterogeneity in State NPI restrictions 54 To capture the heterogeneity in protective measures implemented by each state, including the District of Columbia (DC), we collected state mandated COVID-19 NPI orders from each 56 governor website for one year (March 11, 2020-March 31, 2021). NPI orders for the following five categories were included in this study: stay-at-home orders, non-essential business 58 restrictions, indoor gathering limitations, restaurant/bar restrictions, and mask/face covering mandates. Daily cumulative NPI scores were calculated for each state based on the stringency of 60 each intervention in these five categories (see methods for score rubric). Fig. 1 highlights the heterogeneity in the NPIs implemented by each state, at three distinct times during the pandemic. 62 A time-lapse movie of the NPI map with daily COVID-19 cases is provided in movie S1. Interestingly, implementation strategies varied by state with some states directly placing orders on 64 all residents, while others delegated the responsibility for ordering NPIs to county or city governments ( fig. S1A ). In some cases, states initially placed statewide orders on all residents, but 66 as the pandemic progressed, implementation of interventions was moved to a county or city level. Fig. S1 provides a snapshot of states that implemented statewide, county-level, or city-level NPIs, 68 note that the various levels of NPI implementation were not exclusive. In some states, statewide NPIs were implemented for all residents, but individual counties would extend or modify the NPIs 70 ( fig. S1 ). To explore the accessibility of NPI guidelines per state, we additionally measured the time it took for a new user to find NPI information for each state ( fig. S1B ). We observed that 72 information for the majority of states could be found in under five minutes ( fig. S1B ). All states, except for South Dakota where there weren't any mandatory statewide NPIs, 74 implemented a different combination of interventions statewide during the pandemic (www.phightcovid.org ). Initially, 43 states implemented stay-at-home orders, of which 93% were 76 lifted by June 15, 2020. Restaurant/bar closures were implemented in 49 states (including DC) between March 15 and April 3, 2020. These states, except Missouri, eased restaurant/bar 78 restrictions for the first time between April 24 and June 15, 2020, with the largest proportion doing so between May 9 and May 19, 2020. Missouri did not ease this NPI, rather the closure was 80 completely lifted on June 16, 2020. Between June 1, 2020, and January 2, 2021, 41 states and DC reissued restaurant/bar restrictions for some duration of time before again easing them. A similar 82 scenario of alternating between easing and re-issuing mandates occurred with indoor gathering limitations and non-essential business restrictions. Temporal restrictions by state are available in 84 an interactive format on www.phightcovid.org/graphs. 86 Daily SARS-CoV-2 cases were normalized to state population represented per 500,0000 people (black circles), and the seven-day rolling average is indicated by the grey line. State NPI score per day is indicated by the colored bar along the x-axis. The NPIs are labeled with the category abbreviation, and colored as magenta if restricting, or teal if easing. The duration of the stay-at-home order is indicated by the grey band; black vertical lines represent implementation of statewide mask mandate, labor-day, and each vaccine FDA approval dates. To compare the implementation of various NPIs per state over the last year to the trajectory 90 of COVID-19 cases, we created time series graphs, (including a seven-day rolling average) for each state as displayed in Fig. 2 spline. Estimated splines and weight coefficients are closely related in states with similar underlying case curves. We compared the estimated state splines by using the K-means algorithm 122 to cluster similar sets of weight coefficients, identifying groups of states with similar SARS-CoV-2 case trajectories. Returning to our three example states, we see that the coefficients for two states 124 in the same cluster (Texas, Georgia) are more similar than those in a different cluster (Maine) (fig. S3D&E compared to F). 126 Standard best-fit criteria for K-means and repeated simulation identified a set of seven clusters of states, each with a distinct case curve shape shared across each group. The most 128 common clustering is depicted in (Fig. 3A) . Interestingly, the clusters demonstrate that states grouped geographically even though no geographic or NPI information was included when 130 identifying the clusters, indicating that neighboring states tended to have similar case curve trajectories even if their NPI interventions differ (Fig. 3A) . 132 An assessment of state median NPI score from March 2020 to March 2021 is depicted in 134 Fig. 3B ; surprisingly, a similar clustering pattern was not observed between these two maps. For example, Minnesota and Wisconsin have distinct state median NPI scores but fall within the same 136 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 31, 2021. K-means clusters ( Fig. 3A and B), suggesting that implementation of various NPIs may not solely dictate the trajectory of SARS-CoV-2 cases. Aside from geographical location, environmental 138 conditions such as humidity and temperature are thought to contribute to the seasonality of influenza viruses (6). Recently, humidity and temperature have been implicated in SARS-CoV-2 140 spread (7). However, the identified clusters may not correlate directly to environmental conditions, rather it could be driven by interstate travel patterns, similarities in state demographics, or when 142 specific interventions were implemented(8, 9) . Mobility and social behavior have also been linked to transmission of COVID-19 and may also influence the spread between states(10-12). These 144 potential confounders all require further investigation. 146 Impact of NPI on case change occurs at greater than 2 weeks 148 Figure 4 : More than two-weeks is needed before a NPI has an impact on case change. Case change between the date a restrictive NPI intervention was issued two (A), four (B) and five (C) weeks later for Michigan. Interventions are represented by the points, with easing shown in red and restricting in blue. The circled areas show where the points will be located if there is a considerably greater (red), or less (blue) number of cases at 2, 4 and 5 weeks later. The line of identity, where x=y is the diagonal dashed line in the center of the graph. (D) Overlapping time series graphs showing the number of days since the first time a restaurant/bar easing was issued after May 2020, represented by day 0 on the x-axis. NPIs that were eased are represented by circles. NPIs that were restricted are represented by triangles. The Fourth of July is marked with an X for reference. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 31, 2021. ; Effective implementation of NPI requires an understanding of the impact of NPIs on the trajectory of SARS-CoV-2 transmission. During the COVID-19 pandemic, discussion around the 150 impact of interventions focused on case changes less than two weeks after an intervention. This included both easing or a restricting of an NPI (13, 14) . However, that may not be a realistic time 152 lag between NPI and case change. To examine this relationship, we graphed the number of cases at the time of an intervention versus the number of cases some point in the future, either 2-, 4-, and 154 5-weeks after an intervention (Fig. 4) . Michigan was chosen as a representative of states that implemented many NPI over the course of the last year. By examining the NPIs within Michigan, 156 we found that cases 2-weeks after either an easing or restricting NPI were similar to cases when this NPI was implemented, as the spots are still along the line of identity (x=y). The farther away 158 from the line of identity the greater the case change; we anticipate that as cases raise after lifting of restrictions the cases will increase and move toward the red circle. In contrast as cases reduce 160 after the implementation of an NPI, cases will decrease and move toward the blue circle. Fig. 4B and 4C examine case change after NPIs at 4-and 5-weeks later. Cases 4 weeks after a NPI is eased 162 have increased such that the red easing restrictions are in the red circle (Fig. 4B ). In contrast, cases 4 weeks after a restrictive NPI is implemented have lower cases, but they continue to decrease 5 164 weeks after the restrictive intervention (Fig. 4C ). The observations made in Michigan pose interesting implications regarding the duration of time an NPI should be left in place until it 166 becomes effective or when NPIs should be implemented. This analysis did not distinguish between the type of NPI and impact on cases, which could impact the effect on cases and requires further 168 exploration. To assess the impact of a specific type of NPI, we overlayed the SARS-CoV-2 case 170 trajectories from multiple states on the day a specific intervention was implemented. Using states from multiple different clusters as shown earlier (Fig. 3A) , we graphed the lag in days from the 172 first-time restaurant/bar restrictions were eased after May 2020. Two weeks after the first restaurant and bar easing in May all states were relatively similar in that all experienced little case 174 change. However, between two and four weeks, and even more so between four and five weeks, cases increased. This analysis further supports the observation that a lag of more than 2-weeks is 176 needed before a change in cases after an NPI is eased or implemented (Fig. 4D ). 178 COVID-19 disease outcome correlates with NPI score . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 31, 2021. ; To determine whether the implementation of various NPIs and their stringency impacted COVID-19 outcomes in each state we compared the median state NPI score for each state to 182 SARS-CoV-2 cases and COVID-19 deaths. A linear regression of cumulative SARS-CoV-2 cases normalized to state population and state median NPI score revealed a statistically significant 184 negative relationship (Fig. 5A, to remove the potential confounder of variable case reporting (Fig. 5B) . Linear regression indicated that median state scores were negatively correlated with cumulative COVID-19 deaths (p=0.02), 190 indicating that states with more interventions had reduced mortality rates in addition to fewer cases. 192 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261286 doi: medRxiv preprint While the trendlines are clearly significant between NPI score and SARS-CoV-2 cases or COVID-19 deaths, there are a number of states that reside outside of the 95% confidence interval 194 ( Fig. 5A and B) . In particular, we were surprised by the states that had high NPI scores, but a similar case or cumulative death value as states with much lower NPI scores, suggesting that there 196 may be an optimal NPI restriction threshold. To define the 'goldilocks' or 'just right', of NPI combinations, we compared the states with cumulative cases below the national cumulative case 198 or death average for common interventions, including mask orders, some level of restaurant/bar restrictions, and gathering limitations. Using our rubric, states with all three intervention 200 combinations at some stringency level for a prolonged period of time during the pandemic would have a median state NPI score of ~2.0-3.0, and a greater likelihood of lower mortality and fewer 202 cases than states without these interventions. For example, Maryland, shown in Fig. 2A , had a median score of 2.55 and had the 12 th lowest mortality rate out of all states. 204 Overall, 45% of states had a median NPI score above 2.5 and of these only 8 out of 23 had a mortality rate above the national average. In contrast, in states that had a median score below 206 2.5, 18 out of 25 had a mortality rate higher than the US average mortality rate (Fig. 5C ). It is interesting to note that in states with the highest population density, 14 of 16, had a median score 208 above 2.5, indicating stronger restrictions. In contrast, only 5 out of 24 states with the smallest population densities had median NPI scores above 2.5. To compare how population density may 210 contribute to COVID-19 mortality in the context of NPI orders, a regression between average cumulative deaths per state by population density was performed for states with median NPI. 212 Surprisingly, we did not find a statistically significant relationship between population density and cumulative mortality rate from June 2020 -March 31, 2021, for state median NPI scores either 214 above or below a median score of 2.5 ( fig. S4 ). 216 Discussion 218 During the course of the COVID-19 pandemic a number of studies have attempted to examine the effectiveness of NPIs on SARS-CoV-2 cases around the world and in the US. These studies 220 focused on NPIs implemented early in the pandemic (prior to June 2020) and found that NPIs reduced SARS-CoV-2 cases and mobility with both real and modeled data (1, 2, 10, 15, 18-21). 222 Comparison of 190 countries from January 23, 2020 to April 13, 2020 found that countries with NPIs (specifically: face mask, quarantines, social distancing, and travel restrictions) may have had 224 a lower Rt value, suggesting a decrease in secondary SARS-CoV-2 transmission events in these spaces (1). However, this broad view does not integrate social behavioral variations or climate 226 impacts that could also influence SARS-CoV-2 transmission. In other studies focused on U.S. counties or states, it was found that NPIs reduced mobility of the population and increased the 228 doubling time of SARS-CoV-2 infections (18, 19) . These studies again examined NPIs implemented early in the pandemic and could not account for issues with SARS-CoV-2 testing or 230 the multiple rounds of issuing and easing interventions throughout the course of the past year. In contrast, our study includes data from all 51 U.S. states (including DC) and encompasses the NPIs 232 implemented from March 2020 to April of 2021, to incorporate multiple waves of SARS-CoV-2 infections in states. 234 Taken together our results indicate that state implemented NPIs can lead to less COVID-19 cases 236 and mortality. While it was previously thought that more NPIs led to a better outcome of disease burden (15), we found that even a moderate amount of restrictions can have a substantial impact 238 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 31, 2021. ; on lowering COVID-19 transmission. In addition, we observed a clear geographical impact on SARS-CoV-2 trajectories, and it is likely that consideration of geographical neighbors should be 240 considered when designing future pandemic NPI plans. Finally, additional analysis into the impact of interventions in each state is needed to account for the numerous confounders that limit this 242 type of analysis. A more refined analysis examining the impact of interventions, when they were implemented, the state of the disease at time of implementation, the strength of interventions, and 244 implementation of similar interventions in neighboring states, is needed. 246 Methods 248 Data Collection COVID-19 case and death data were obtained from the COVID-19 Data Repository by the 250 Center for Systems Science and Engineering (CSSE) at Johns Hopkins University. For JHU CSSE COVID-19 Data see https://github.com/CSSEGISandData/COVID-19. Using this data, we 252 calculated the new daily cases/500,000 people for every state by normalizing a state's new confirmed cases to the state population then multiplying by 500,000. 254 NPI intervention data was collected from a variety of sources and evolved over time. Data was collected manually from state government and governor websites. Data obtained on these 256 websites was found in press releases, FAQ pages, reading through executive orders, or in the most accessible format, through an informative interactive portal detailing current restrictions and any 258 changes from past restrictions. Data obtained through various news sources or governor's twitter page, was validated through a search for an official state government or governor announcement 260 or order. Implementation of NPIs varied between statewide or county specific order, which will 262 impact the ability for residents to obtain this information and their awareness of the NPI in their region. To address the accessibility of the NPI orders in place for each state, we measured the time 264 it took to 1) find state issued COVID-19 restrictions and 2) determine if state government orders apply statewide or county-specific. Additionally, we identified if there were county government 266 issued NPIs and the specific stringency of the orders. The search process was consistent for all 50 states and took place Feb 17, 2021. Starting at google.com the keywords "[state of interest] covid 268 restrictions" was queried. The most relevant (state government page or governor page) first result was chosen. After entering the website, a process of looking for the above-mentioned questions 270 was carried out. The time was recorded upon finding each of the questions of interest. Furthermore, if the site was exceptionally easy to navigate and the current restrictions in place were found in 272 less than a minute without downloading and opening additional files, the state was noted as being outstanding in user accessibility (n=16, CA, CO, CT, DE, DC, ID, IL, MI, MN, NJ, NM, OH, OR, 274 PA, RI, and WA). Although states without any NPI orders or no detail on the NPI orders state were also easy to navigate. In situations such as this the state was not marked as being outstanding in 276 user accessibility due to their lack of information displayed at all (AR, FL, and SD). Details of the time, sources, and notes per state for this analysis is available as table S3. 278 In order to compare the stringency of interventions implemented in each state, we developed a detailed scoring system. The system scores on the following five categories: Stay at 282 Home Orders, Non-Essential Business Restrictions, Indoor Gathering Bans, Restaurant/Bar Restrictions and Mask/Face Covering Mandates. The highest score possible for each of the five 284 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 31, 2021. ; categories is 1.00, with the collective highest score possible being 5.00. The scoring criteria are presented in table S2 and briefly described here: 286 • Stay at home orders had a score impact of 1.0 when issued and -1.0 when removed. • Non-essential business sectors that we impacted the score on are as follows: retail (+/-288 0.25), hair salons/barbershops (+/-0.10), personal care services (+/-0.10), gyms/fitness centers (+/-0.2), indoor entertainment (+/-0.2). Each of these business sectors could have 290 between 0 and their max score depending on the level of the restrictions in place. For example, if hair salons were open but only at 50% occupancy, the score impact would be 292 0.05). • Indoor gathering bans primarily applied to indoor private social gatherings, however, as 294 states eased restrictions, indoor gathering sizes described were more closely associated with larger private events. See table S2, category 3 for the scores associated with gathering 296 sizes. Scoring of states that adopted a state government issued county risk-level metric (CA, CO, HI, IL, IN, NM, ND, OR, UT, WA) was more complicated than just scoring for the state as by itself. In 306 these states the NPIs scored on were whatever they were for the majority risk level out of all counties. For example, if Oregon had 16 counties at low, 6 at moderate and 2 at high risk levels, 308 then the entire state would be considered as being at a low risk level. The low risk level NPIs would be recorded and the state as a whole scored scoring accordingly. 310 Time series graphs were generated for each state to visualize an individual state's precise interventions (either restrictive or easing), NPI score (calculated based on the system described 314 above), new daily SARS-CoV-2 cases, and seven day rolling average line. Daily SARS-CoV-2 cases were normalized to state population (based on the U.S. census bureau 2017 population 316 estimates) and presented per 500,000 people. Interactive time series graphs are displayed on www.phightcovid.org, where the user can hover over each intervention and see the description of 318 the type of intervention. The code to generate each of the time series is available on the phightcovid Github repository (see code availability section). 320 Lag graphs were created for gaining a better initial understanding of the time needed between when an NPI is implemented and when cases begin to change as a result. Using the 324 programming language R we plotted today's new daily cases per 500,000 people on the x-axis and the new daily cases per 500,000 people at a lag of 2-, 4, and 5-weeks later, on the y-axis. We then 326 use larger circles to highlight where it was expected to see the points if there was a considerable case change. R code for the lag graphs can be found in our Github repository 328 CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 31, 2021. ; 414 416 418 Fig S1: Implementation strategies and accessibility of NPI in each state (A) State governments issuing NPI on all residents either with entire state orders or county-based orders. Indicated with a star are state governments that adopted a county risk-level metric and regularly changed specific county NPIs depending on the current COVID-19 risk level determined by the state (green). State governments delegating NPI responsibility to counties (red). County governments either delegating NPI responsibilities, or city and health departments deciding to implement NPI on their own (brown and blue respectively). Given the complexity of differentiating between health department and county level implementation, we were unable to capture all the states that implemented NPI through local health departments (B) To quantify accessibility of state COVID-19 NPI information, we timed how long we had to spend searching for NPI information on each state COVID-19 website. Data is from Feb 2021. Some states may have changed the NPI implementation strategy and website for ease. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261286 doi: medRxiv preprint Daily COVID deaths were normalized to state population represented per 1,00,0000 people (black circles) and the 7-day rolling average is indicated by the grey line. State NPI score per day is indicated by the colored bar along the x-axis. The NPIs are labeled with the category abbreviation, and colored as magenta if restricting, or teal if easing. Additionally, the duration of the stay-at-home order is highlighted by the grey band and black lines indicate implementation of statewide mask mandate, labor-day, and each vaccine FDA approval dates. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261286 doi: medRxiv preprint COVID-19: Effectiveness of Non-Pharmaceutical Interventions in the 376 United States before Phased Removal of Social Distancing Protections Varies by Region. medRxiv, 2020 The Immediate Effect of COVID-19 Policies on Social-Distancing Behavior in the United States The unintended consequences of inconsistent pandemic control policies. medRxiv Impacts of social distancing policies on mobility and COVID-19 case growth in the US Effect of non-pharmaceutical interventions to contain COVID-19 in China Challenges in control of COVID-19: short doubling time and long delay to effect of interventions Optimal, near-optimal, and robust epidemic control Which COVID policies are most effective? A Bayesian analysis of COVID-19 by jurisdiction An interactive web-based dashboard to track COVID-19 in real time State-level variation of initial COVID-19 dynamics in the United States Impacts of introducing and lifting nonpharmaceutical interventions on COVID-19 daily growth rate and compliance in the 398 United States Inferring the effectiveness of government interventions against 400 COVID-19 Inferring change points in the spread of COVID-19 reveals the 402 effectiveness of interventions