key: cord-0968833-li3slzw6 authors: McGough, S. F.; Gan, R. W.; Tibshirani, R.; Meyer, A.-M. title: Modeling COVID19 mortality in the US: Community context and mobility matter date: 2020-06-20 journal: nan DOI: 10.1101/2020.06.18.20134122 sha: e302fa274466ba9d83e8ff157efc9dd2df9adcdf doc_id: 968833 cord_uid: li3slzw6 The United States has become an epicenter for the coronavirus disease 2019 (COVID-19) pandemic. However, communities have been unequally affected and evidence is growing that social determinants of health may be exacerbating the pandemic. Furthermore, the impact and timing of social distancing at the community level have yet to be fully explored. We investigated the relative associations between COVID-19 mortality and social distancing, sociodemographic makeup, economic vulnerabilities, and comorbidities in 24 counties surrounding 7 major metropolitan areas in the US using a flexible and robust time series modeling approach. We found that counties with poorer health and less wealth were associated with higher daily mortality rates compared to counties with fewer economic vulnerabilities and fewer pre-existing health conditions. Declines in mobility were associated with up to 15% lower mortality rates relative to pre-social distancing levels of mobility, but effects were lagged between 25-30 days. While we cannot estimate causal impact, this study provides insight into the association of social distancing on community mortality while accounting for key community factors. For full transparency and reproducibility, we provide all data and code used in this study. One-sentence summary 24 County-level disparities in COVID19 mortality highlight inequalities in socioeconomic and 25 community factors and delayed effects of social distancing. 26 27 Abstract 28 The United States has become an epicenter for the coronavirus disease 2019 (COVID-19) 29 pandemic. However, communities have been unequally affected and evidence is growing that 30 social determinants of health may be exacerbating the pandemic. Furthermore, the impact and 31 timing of social distancing at the community level have yet to be fully explored. We investigated 32 the relative associations between COVID-19 mortality and social distancing, sociodemographic 33 makeup, economic vulnerabilities, and comorbidities in 24 counties surrounding 7 major 34 metropolitan areas in the US using a flexible and robust time series modeling approach. We 35 found that counties with poorer health and less wealth were associated with higher daily 36 mortality rates compared to counties with fewer economic vulnerabilities and fewer pre-existing 37 health conditions. Declines in mobility were associated with up to 15% lower mortality rates 38 relative to pre-social distancing levels of mobility, but effects were lagged between 25-30 days. 39 While we cannot estimate causal impact, this study provides insight into the association of 40 social distancing on community mortality while accounting for key community factors. For full 41 transparency and reproducibility, we provide all data and code used in this study. 42 The United States (US) has become an epicenter for the coronavirus disease 2019 (COVID-19) 45 pandemic, with over 2 million confirmed cases and 115,000 confirmed deaths as of June 15, 46 2020. However, impact of the disease across the US has been unequal, with documented 47 geographic heterogeneities in cases and deaths and excess burden in certain subgroups of the 48 population. Clinical reports have noted the increased incidence and severity of COVID-19 49 among individuals with diabetes, obesity, and chronic lung conditions (1, 2) . Other influential 50 community factors include social and economic vulnerabilities (3, 4), referred to in the 51 epidemiologic literature as "social determinants of health" (5 For example, social distancing policies have been adopted by state and local governments at 60 different points during the outbreak, and have also been shown to impact COVID-19 growth 61 rates (8). However, there is evidence to suggest that some communities initiated social 62 distancing before formal implementation of policies (9), and may have relaxed prior to the official 63 releases (10). Aggregated mobility data can capture these changes over time and offer a richer 64 resolution for the continuum of community-level behavioral modifications (11). With economic 65 and political pressures to release restrictions, it is critically important to assess how community-66 level sociodemographics and mobility play a role in COVID-19 mortality in order to develop 67 effective response strategies. 68 69 Our study combines both community context and social distancing effects and evaluate COVID-70 19 mortality rates over time across US counties. By simultaneously incorporating population-71 level comorbidities, sociodemographic makeup, and social distancing over time, our approach 72 identifies vulnerable communities at higher risk of mortality. Our study also quantifies the 73 independent associations between declining mobility and subsequent reductions in mortality 74 rates as well as mortality from community-level factors. 75 76 We investigated 24 counties surrounding 7 major metropolitan areas in the US that exhibit 77 heterogeneities in sociodemographic characteristics and comorbidities ( Figure 1A ) as well as 78 differences in timing of the adoption of social distancing over the course of the outbreak ( Figure 79 1B). The metropolitan areas -Chicago, Detroit, Los Angeles, New Orleans, New York City, San 80 Francisco, and Seattle -are similar to those examined in the CDC MMWR on COVID-19 and 81 mobility (12). They were selected for their substantial size in COVID-19 cases and deaths, 82 heterogeneity in social distancing adoption, and length of outbreak available to study. Counties 83 were studied up to and including May 13, 2020, and the epidemiologic profile of each county is 84 detailed in Table S1 . We observed strong correlations between county-level characteristics 85 including comorbidities and demographics ( Figure S1 , Table S2 ). To address these correlations, 86 . CC-BY-NC 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 June 20, 2020. . we applied principal components (PC) analysis to collapse potential COVID-19 county-level risk 87 factors into four main representative constructs that explain 80% of the variation in 88 sociodemographics across counties: "health and wealth," "housing inequality" "old age and 89 clinical care," and average daily "PM 2.5 " (Figures S2, S3; described in Supplementary Materials, 90 Materials and Methods). Together with data on daily human mobility, the principal components 91 generated in this data pre-processing step were then used as predictors in a model for COVID-92 19 mortality. 93 94 We used a generalized additive model (GAM), a flexible modeling approach (13), which 95 uniquely and simultaneously addresses three key challenges not fully addressed in current 96 literature: First, it allows for the quantification of a nonlinear relationship in COVID-19 mortality 97 rates over time, while accounting for county-specific temporal evolution of COVID-19 deaths in 98 these 24 counties. Second, it allows for quantification of the effect of the four PC county-level 99 sociodemographic constructs. Third, it allows for quantification of the lagged relationship of 100 social distancing over time using a penalized distributed lag framework, which can account for 101 the time delay that occurs between infection and mortality (14). Here, we discuss the two major 102 elements of the GAM model: 1) county-level sociodemographics and 2) social distancing as 103 measured by changes in county-level mobility. 104 105 Economic vulnerabilities and poor baseline health indicators ("health and wealth") had the 106 strongest associations with COVID-19 mortality rates across counties, contributing nearly 10 107 times as much explanatory power to the model compared to the next PC predictor, "housing 108 inequality" (F-statistic= 31.7 for "health and wealth" compared to 3.3 for "housing inequality"). 109 Higher prevalence of comorbidities was collectively associated with higher COVID-19 daily 110 mortality rates, including diabetes, obesity, and adult smoking ( Figure 2 ; Figure S3 ). In contrast, 111 higher median household income, higher levels of college education, and lower unemployment 112 rates were collectively associated with lower COVID-19 daily mortality rates ( Figure 2 ; Figure 113 S3). These effect estimates also incorporate any changes in social mobility over time. 114 115 Mortality rates associated with "health and wealth" were between 5 to 35 times higher in New 116 Orleans-surrounding counties (St. Bernard, Jefferson, St. Tammany, and Plaquemines 117 Parishes) compared to California Bay Area counties (San Francisco, San Mateo, Santa Clara, 118 and Alameda Counties) after accounting for mobility. Though, St. Bernard, a smaller parish near 119 New Orleans, experienced highly fluctuating numbers of deaths, resulting in wider uncertainty at 120 the tails of mortality rate estimates ( Figure 2 ). 121 122 It is important to note race/ethnicity was not implicitly included as an independent variable in 123 county-level sociodemographic constructs. Rather, we included measures such as racial 124 segregation, proportion non-native English speakers, income, and housing quality measures 125 (Table S2) , which represent the underlying community influences that contribute to vulnerable 126 racial and ethnic minority groups. Communities that scored high on variables contributing to 127 racial and ethnic subgroup vulnerabilities were at exceptionally high risk of COVID-19 mortality 128 in this analysis, such as Detroit and New Orleans-area counties (Figure 2 ; Figure S3 ). Our 129 results are complementary to current case and clinical reports that document a disproportionate 130 . CC-BY-NC 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 June 20, 2020. . amount of severe cases and deaths in disadvantaged populations, primarily African American 131 and Hispanic communities (4, 15-17) and those populations with chronic underlying health 132 conditions (18) including diabetes, hypertension, and obesity. 133 134 There has been considerable debate on the second component of our model, namely, the 135 impact and effectiveness of social distancing on the COVID-19 outbreak. We used Google's 136 human mobility data (19) as a proxy for social distancing, as it objectively reflects the localized 137 changes in movement early on in the outbreak before state and local social distancing policies 138 were put into place. Because social distancing policies were issued in response to exponential 139 growth in the epidemic, we observed a paradoxical relationship between the declining trajectory 140 of mobility against a rising trajectory of COVID-19 deaths ( Figure 1B) , also observed in other 141 populations (20). In this analysis, we controlled for the time evolution of the outbreak and 142 modeled mobility as multiple distributed lags of 0 to 30 days to account for this real-time trend. 143 This framework allows for a sufficient number of lag days without over-parameterizing the 144 model. 145 146 Our model shows that reduced mobility in all study counties was associated with steady, linear 147 declines in mortality rates over a 30-day period, after controlling for county-level principal 148 components. Moreover, we found that it may take approximately 25 to 30 days to observe 149 reductions in mortality rates following declines in mobility from baseline (reference point: 0% 150 change in mobility, Figure 3 ). In the beginning of the observed epidemic, declines in mobility 151 occur concurrently with a rise in mortality rates ( Figure 1B) , which is why studying the delayed 152 (lagged) effect of mobility is critical to capture associations over time. The largest reductions in 153 mortality rates were observed at the highest levels of social distancing (i.e. the highest 154 reductions in mobility). For example, mortality rate reductions of 13% (95% CI: 7%, 17%) and 155 15% (95% CI: 9%, 24%) after 30 days were associated with decreases in mobility of 50% and 156 80%, respectively, compared to more modest mortality rate reductions after 30 days associated 157 with decreases in mobility of 25% and 10% ( Figure 3) . Conversely, increasing mobility by 10% 158 was associated with a small, but steady increase in mortality rate over the 30-day lag period. 159 160 It is possible that these relationships will change or strengthen over time as the outbreak 161 persists, providing longer time periods in which to observe changes in mobility and lagged 162 effects. However, these findings suggest that the relationships between mobility and mortality 163 rates are complex, and it can take at least four weeks to observe notable reductions in deaths. 164 These findings are consistent with our hypothesis that as communities in the US return to pre-165 social distancing levels of mobility, increases in mortality rates will likely be observed around 166 four weeks later. Most importantly, these findings suggest that mortality rates in the initial days 167 following social policy implementations may appear paradoxically higher compared to rates pre-168 social distancing, and take time to decline. 169 170 Sensitivity analyses were conducted to assess the robustness of findings around the principal 171 components and lagged mobility. Results were similar when we: 1) conducted our statistical 172 analysis using two different lagged modeling approaches for mobility, 2) excluded New York 173 . CC-BY-NC 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 June 20, 2020. . City because of its unusually high disease burden, and 3) truncated individual counties to the 174 same period of observation time (Figures S4-S8 ). 175 176 Our results demonstrate it is critical to study COVID-19 in a broader context. By combining 177 community-level variables and mobility data, we are able to understand their relative impact at a 178 population level. As more data on COVID-19 emerge, evidence continues to mount on the 179 sociodemographic disparities of the COVID-19 outbreak in the US and globally (4, 15, 21). As 180 observed by Chung and colleagues (16), the scientific community is witnessing a 181 "socioeconomic gradient" in the COVID-19 outbreak, in which disadvantaged communities are 182 at higher vulnerability to COVID-19 infection and death. These social determinants of health are 183 well described in the broader health literature (22) This analysis also adds to the growing body of evidence on the relationship between mobility 192 and COVID-19 (20, 24, 25), but takes a step further to contextualize the evolution of the 193 outbreak at the county-level in the US. We show that social distancing was associated with 194 reductions in COVID-19 mortality rates of up to 15% after 25-30 days, relative to baseline 195 ( Figure 3) . lagged effects of mobility). Additional lagged terms at the same temporal and spatial resolution 216 may be added to the existing model to explore other potential risk factors including 217 . CC-BY-NC 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 June 20, 2020. . meteorological variables and air pollution. And, as more data related to the pandemic become 218 available, it will be possible to account for additional mitigating factors unavailable at this time, 219 including other public health actions (e.g. wearing masks, contact tracing) and indicators of 220 health system preparedness (e.g. hospital surge capacity, ventilators). Finally, the current 221 analysis focuses on counties surrounding major metropolitan areas, and may not capture risk 222 factors relevant to rural communities in the US. To better understand the generalizability of 223 these results nationally and even globally, an extension of this work to additional locations is an 224 important next step. 225 226 To the best of our knowledge, this analysis is the first to differentiate communities at-risk for 227 COVID-19 mortality and, in the context of these factors, assess the relationship between social 228 distancing and reductions in mortality over time. We identified potential community-level risk 229 factors of COVID-19 mortality, including pre-existing health conditions (smoking, diabetes, repository with links to data and all scripts to replicate and extend this analysis at: 237 https://github.com/phcanalytics/covid19_epi_model. 238 . CC-BY-NC 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 June 20, 2020. . CC-BY-NC 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 June 20, 2020. . CC-BY-NC 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 June 20, 2020. CC-BY-NC 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 June 20, 2020. . thank the broader "TPP2" teams within Diagnostic Information Solutions and Personalized 348 Health Care for their collaboration and support on data, engineering, and meaningful scientific 349 discussions on how to best study the COVID-19 pandemic and develop critical insight. We also 350 thank N. Pal for code review and G. Simpson for his input on our use of the "mgcv" R package. Table S1 -S2 364 Fig S1 -S10 365 . CC-BY-NC 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 June 20, 2020. . https://doi.org/10.1101/2020.06.18.20134122 doi: medRxiv preprint Severe Outcomes Among Patients with Coronavirus Disease 2019 (COVID-268 19) -United States Low-Income and Communities of Color at Higher Risk of Serious Illness 276 if Infected with Coronavirus COVID-19 and Underinvestment in the Health of the US Population Social conditions as fundamental causes of health 282 inequalities: theory, evidence, and policy implications Fundamental causes" of 285 social inequalities in mortality: a test of the theory Louisiana Coronavirus COVID-19 | Department of Health | State of Louisiana Hospitalization Rates and Characteristics of Patients Hospitalized with 320 Laboratory-Confirmed Coronavirus Disease 2019 -COVID-NET, 14 States The effect of human mobility 327 and control measures on the COVID-19 epidemic in China As Coronavirus Deepens Inequality, Inequality Worsens Its Spread Social determinants of health inequalities. The Lancet The Incubation Period of Coronavirus Disease Publicly Reported Confirmed Cases: Estimation and Application We thank the Roche COVID-19 task force for prioritizing this research, and the leadership at 346Roche and Genentech for supporting this scientific effort for the benefit of society. We also 347