Science Journals — AAAS Warshaw et al., Sci. Adv. 2020; 6 : eabd8564 30 October 2020 S C I E N C E A D V A N C E S | R E S E A R C H A R T I C L E 1 of 4 C O R O N A V I R U S Fatalities from COVID-19 are reducing Americans’ support for Republicans at every level of federal office Christopher Warshaw1*†, Lynn Vavreck2†, Ryan Baxter-King2† Between early March and 1 August 2020, COVID-19 took the lives of more than 150,000 Americans. Here, we ex- amine the political consequences of the COVID-19 epidemic using granular data on COVID-19 fatalities and the attitudes of the American public. We find that COVID-19 has led to substantial damage for President Trump and other Republican candidates. States and local areas with higher levels of COVID-19 fatalities are less likely to sup- port President Trump and Republican candidates for House and Senate. Our results show that President Trump and other Republican candidates would benefit electorally from a reduction in COVID-19 fatalities. This implies that a greater emphasis on social distancing, masks, and other mitigation strategies would benefit the president and his allies. INTRODUCTION COVID-19 has killed about 5 times as many Americans as were killed in the Korean War, over 3 times as many as in Vietnam, and 40 times as many Americans as were killed in the entire Iraq War. Americans broadly disapprove of the president’s handling of the pandemic (1), but as of yet, there has been no clear causal evidence about whether the rise in COVID-19 fatalities has led Americans to turn away from President Trump. A large academic literature has shown that the American public holds presidents accountable for their performance in office (2, 3). Among other things, the public penalizes a president and others in their party for casualties in war. Areas with more local casualties, for example, were among the first to turn against the Vietnam War be- tween 1965 and 1972 (4), and during the Iraq War, people who knew someone who died on 9/11 or in the Iraq War were consistently more likely to disapprove of George W. Bush (5). As a result, states with greater losses were more likely to vote against President Bush in the 2004 presidential election (6). Voters also punished Republican can- didates at other levels of office: Areas with higher casualties from the war in Iraq were more likely to support Democratic House and Senate candidates in the 2006 midterm elections (7, 8). Last, areas with higher casualties in the war in Afghanistan penalized Barack Obama’s Democratic successor in 2016, Secretary of State Hillary Clinton, by supporting Donald Trump in greater numbers (9). The U.S. president has likened his battle against COVID-19 to that of a “war-time president” (10). Voters may also see him that way. On the basis of previous studies of the political costs of war-time casualties, we hypothesize that the American public will be less like- ly to support President Trump and other Republican candidates for federal offices in areas with higher levels of COVID-19 fatalities. We examine whether Americans are penalizing the president and other Republicans for the fatalities due to COVID-19 using several granular data sources (see Materials and Methods for more details). We leverage both temporal and geographic variation in the magni- tude of the COVID-19 pandemic using local-level data on fatalities gathered by the New York Times. We use the Democracy Fund + UCLA Nationscape Project to measure the attitudes of the American pub- lic at a local level. This survey includes the responses of more than 300,000 people between the summers of 2019 and 2020 (11). RESULTS Figure 1 (below) examines the state-level association between cu- mulative COVID-19 fatalities as of 31 May 2020 and changes in Americans’ attitudes between the first 2 months of 2020 and June. It shows that states with more COVID-19 fatalities were less likely to support Republican candidates. For example, people in the states with the highest fatalities were about 6% less likely to approve of Presi- dent Trump’s performance in office than people in the states with the lowest level of fatalities (Fig. 1A). The states with the highest level of fatalities were about 3% less likely to support President Trump’s reelection in the presidential race against Democrat Joseph R. Biden (Fig. 1B). The hardest-hit states were nearly 13% less likely to sup- port Republican Senate candidates (Fig. 1C) and about 5% less likely to support Republican House candidates (Fig. 1D). These associations, however, could be confounded by other state- level factors and may be affected by sampling variability at the state level (particularly for smaller states). Thus, we move next to a more rigorous difference-in-differences regression design to assess the causal effect of COVID-19 fatalities on political preferences. This approach examines the effect of COVID-19 fatalities over the past 30 days in each respondent’s state or county on their attitudes about President Trump and other politicians. In addition to providing a more granular test, county-level results characterize the impact of the information environment surrounding the pandemic relative to the actual number of fatalities. We use fixed effects for geography and week of interview to account for area- and time-specific con- founders. We also control for a host of pre–COVID-19 individual- level attributes of the survey respondents, including 2016 vote choice, making our results net of factors such as race, education, gender, and partisan preference in 2016 (see Materials and Methods). We find consistent results at every level of geography and for every office (Fig. 2): The effect of fatalities is a drain on Republican vote share (see Materials and Methods for a variety of robustness checks and the Supplementary Materials for a table with the regression re- sults). Overall, areas with higher COVID-19 fatalities are signifi- cantly less likely to support President Trump and other Republican 1Department of Political Science, George Washington University, Washington, DC, USA. 2Department of Political Science, UCLA, Los Angeles, CA, USA. *Corresponding author. Email: warshaw@gwu.edu †These authors contributed equally to this work. Copyright © 2020 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works. Distributed under a Creative Commons Attribution NonCommercial License 4.0 (CC BY-NC). o n A p ril 5 , 2 0 2 1 h ttp ://a d va n ce s.scie n ce m a g .o rg / D o w n lo a d e d fro m http://advances.sciencemag.org/ Warshaw et al., Sci. Adv. 2020; 6 : eabd8564 30 October 2020 S C I E N C E A D V A N C E S | R E S E A R C H A R T I C L E 2 of 4 candidates. A doubling of COVID-19 fatalities (0.69 U on the natu- ral log scale) at the county level leads to a roughly 0.19% reduction in President Trump’s approval rating, and a doubling in fatalities at the state level leads to a 0.5% reduction in the president’s approval. In the presidential election, a doubling of COVID-19 fatalities at the county level makes Americans about 0.14% less likely to support President Trump against Joseph R. Biden and a doubling in fatali- ties at the state level leads to a 0.37% reduction in support for Trump. In Senate races, a doubling of COVID-19 fatalities at the county level makes Americans about 0.28% less likely to support Republi- can candidates and a doubling in fatalities at the state level leads to a 0.79% reduction in support for Republicans. Last, in House races, a doubling of COVID-19 fatalities at the county level makes Americans about 0.22% less likely to support Republican candidates and a doubling in fatalities at the state level leads to a 0.58% reduction in support for Republicans. DISCUSSION Our results show that the COVID-19 pandemic has already substan- tially damaged the political standing of President Trump. Just as the public penalizes the president for casualties during wars, the public AK AL AR AZCA CO CTDE FL GA HI IA ID ILINKS KY LA MAMD ME MI MN MO MS MT NC ND NE NH NJ NMNV NYOH OKOR PA RI SC SDTN TX UT VA VT WA WI WV WY ALAAL FLLL AAAA AAAAAA ILININNINSSKKS KYKCACACACAFFAAAAAAAA LALAL MAMMAMDMDILILILILMOM MSS MTMMT NN NDNDSSSS KKKK NJNJNNNNNNNHHOHOHHOOOO PPAAPPPPSCSCCCCC TNTN TXTXKKKK VVAAOOVVV −20% −15% −10% −5% 0% 5% 10% 15% 20% 0 1 2 3 4 5 Log(COVID deaths per 100,000) C ha ng e in T ru m p ap pr ov al fr om J an ua ry /F eb ru ar y to J un e Association between COVID and Trump approvalA AK AL AR AZ CA CO CT DE FL GA HI IA ID IL IN KS KY LA MA MD ME MI MN MO MSMT NC ND NE NH NJ NM NV NY OH OK OR PA RI SC SDTN TX UT VA VT WA WI WV WY ALAALCAACAAAAAA ILILKY CACA LAL MAMMA MDMD ILLILL MEE MO NCNC CC NJNJ NV ALAL MM NNNNN OKNCNC PAPP TXTX VVMVVMMMMAAMOMOVVVMM WAWW −10% −5% 0% 5% 10% 0 1 2 3 4 5 Log(COVID deaths per 100,000) C ha ng e in T ru m p vo te in te nt fr om J an ua ry /F eb ru ar y to J un e Association between COVID and Trump vs. Biden vote intent AK AL AR AZ CO DE GA IA ID ILKS KY LA MAME MI MN MS MT NC NE NH NJ NM OK OR RI SC SD TN TX VA WV WY AK LAZAAZALALA GG ILL Y MEE NENEKYKY N OROR RIRI SCKYKYNNKYKYKYKYDDSSDMEMEEEMM VAVV WVV −10% −5% 0% 5% 10% 15% 20% 0 1 2 3 4 5 Log(COVID deaths per 100,000) C ha ng e in R ep ub lic an v ot e in te nt fr om J an ua ry /F eb ru ar y to J un e Association between COVID and Senate vote intent AK AL AR AZCA CO CT DE FL GA HI IA ID ILINKS KY LA MA MD ME MI MNMO MS MT NC ND NE NH NJ NM NV NY OH OKOR PA RI SCSD TN TX UT VAVT WA WI WV WY AZAAZCAAAAAA DEDE ILIL KYY CACA LAL MAM MDMDILILILIL EEMEMEE MTM NCN CCN CCKYKYYY CCNCNCNN CCCCKYKYKYKYKYKYKYKY CC NJNJNNNNN ORO SCSCS TXTXMM VAV −15% −10% −5% 0% 5% 10% 15% 20% 0 1 2 3 4 5 Log(COVID deaths per 100,000) C ha ng e in R ep ub lic an v ot e in te nt fr om J an ua ry /F eb ru ar y to J un e Association between COVID and House vote intent C B D Fig. 1. Association between COVID-19 deaths and changes in political preferences. House vote Senate vote President vote Trump approval −1.25% −1% −0.75% −0.5% −0.25% 0% Effect of doubling in COVID deaths per 100,000 people County State Fig. 2. Effect of COVID-19 deaths on political preferences for various offices. This graph shows the results of regression models of the effect of a doubling in COVID-19 deaths per 100,000 people in the past 30 days in each state and county on Trump approval and whether respondents plan to vote for Republican candidates for president, Senate, and House. The dots show the point estimates, and the bars show 95% confidence intervals. o n A p ril 5 , 2 0 2 1 h ttp ://a d va n ce s.scie n ce m a g .o rg / D o w n lo a d e d fro m http://advances.sciencemag.org/ Warshaw et al., Sci. Adv. 2020; 6 : eabd8564 30 October 2020 S C I E N C E A D V A N C E S | R E S E A R C H A R T I C L E 3 of 4 is penalizing the president and other members of his party for local fatalities during the pandemic. The number of local fatalities due to COVID-19 appears to be at least as important as the local economy in Americans’ evaluations of their leaders (12, 13). COVID-19 could cost Trump and other Republicans several percentage points in the 2020 election. This could swing the presidential election and the U.S. Senate toward Democrats, with particularly high effects in swing states such as Michigan, Wisconsin, Pennsylvania, New Hampshire, Arizona, and Florida. All of these states had tight margins in the 2016 presidential election. Michigan’s margin was particularly narrow (0.2%), as was New Hampshire’s (0.4%), suggesting that COVID- related fatalities may be consequential not only at the individual level in 2020 but also in terms of Electoral College results. Similarly, there were very close U.S. Senate elections in 2018. In Florida, 0.2% of the vote separated the Republican winner from the Democrat. These narrow margins in 2016 and 2018, coupled with the reali- zation that fatalities from COVID-19 are not unlike casualties of war in voters’ minds, suggest that a winning strategy for President Trump and other Republican candidates on the ballot in 2020 should be to adopt mitigation strategies to limit the spread and consequences of COVID-19  in the American population. Increasing fatalities from the disease leads to losses for Republicans. MATERIALS AND METHODS This section describes the methods and data that we use in our paper. The first building block of our study is granular data on reported COVID-19 fatalities across geography and time. For this, we use data that the New York Times has collected on the basis of state web- sites and databases (see https://github.com/nytimes/covid-19-data). We then aggregate the county-level data on COVID-19 deaths at the state level. County-level population data are taken from the 2014 to 2018 American Community Survey (ACS). The next building block is data on attitudes of the American public about President Trump and vote intentions for the 2020 elections. For this, we use the Democracy Fund + UCLA Nationscape Project to measure the attitudes of the American public at a local level (11). This survey includes the responses of more than 300,000 people, about 6400 of whom were interviewed each week between the summers of 2019 and 2020 (through 29 July 2020). The survey is fielded on- line and is representative of the nation as a whole (14). The Nation- scape staff generate sampling weights for the weekly datasets. The technique is based on processes used by the American National Election Studies. In table S1, we show a detailed comparison of the weighted Nationscape sample with population targets. Overall, the weighted sample appears to be extremely representative of observable population targets. Owing to its large size, Nationscape can also be disaggregated to reflect opinions at the state and local levels. The survey asks about a variety of political attitudes and prefer- ences. We use four specific questions from the survey. First, we use data on whether respondents approve of President Trump’s job per- formance. We collapse this four-point question to a dichotomous variable. Second, we use data on whether people would vote for President Trump or Joseph R. Biden in a head-to-head matchup in the 2020 presidential election. Third, we use data on whether respondents plan to vote for the Republican or Democratic candidate in the 2020 House election in their district. Last, we use data on whether re- spondents plan to vote for the Republican or Democratic candidate in the 2020 Senate elections in their state (if they have one). For each, we are excluding individuals who answered “Not sure.” (Note that fig. S4 shows that the results are similar in models that include don’t know responses.) Our main paper reports the results of two sets of analyses. The next two sections describe the details for these analyses. Association between COVID-19 deaths and changes in political preferences at the state level First, we look at the state-level association between COVID-19 fa- talities and Americans’ attitudes about President Trump and their vote intentions in the 2020 election. For this analysis, the independent variable is the natural log of the number of COVID-19 fatalities per 100,000 people in each state before 1 June 2020. The outcome vari- able is the change in the public’s attitudes before the COVID-19 pandemic (defined as the first 2 months of 2020) and their attitudes after the arrival of COVID-19, between 1 June and 2 July 2020. We use the appropriate state-level sampling weights to calculate the public’s state-level attitudes in each time period. We then graph the relationship between COVID-19 fatalities and the changes in political attitudes in each state. By focusing on changes in political attitudes, our analysis implicitly accounts for time-invariant confounders (omitted variables) in each state and common shocks that affect all states. However, there is large sampling variability at the state level, particularly in smaller states, which we will address in further analyses. Causal effect of COVID-19 deaths on political preferences for various offices Next, we move to a more rigorous difference-in-differences regres- sion design. We use a linear probability model to examine the effect of COVID-19 fatalities over the past 30 days in each survey respon- dents’ state or county with their attitudes about President Trump and other politicians. For this analysis, the independent variable is the natural log of the number of COVID-19 fatalities per 100,000 people in the last 30 days (relative to the date each respondent was interviewed) in each geographic area. A 0.69-U increase on the nat- ural log scale can be interpreted as approximately a doubling of fa- talities (15). Here, we use fixed effects for geography and survey wave (week) to account for area- and time-specific confounders and identify the causal effects of COVID-19 on political attitudes (16). The geographic fixed effects account for the tendency of different areas to have varying levels of baseline support for President Trump and other Republican candidates. The temporal fixed effects ac- count for national-level changes in political attitudes due to the pandemic, the economy, and national events such as the Black Lives Matter movement. We also control for a host of individual-level pretreatment attributes of the survey respondents. These are not crucial for our identification strategy, but they reduce the variance in our results (17). Specifically, we control for respondents’ gender, race/ethnicity, education, Hispanic ancestry, and their vote choice in the 2016 presidential election. The SEs in our regression results are clustered at the state-day or county-day level depending on the model (18). We use national sampling weights in all our analysis. So, our results are representative of the American public at the national level. While our main analyses use a linear probability model, we find substantively similar results using logistic regression models. To validate our research design, we run a placebo check where we examine the effect of future COVID deaths on an index of ap- proval, presidential voting, senate voting, and house voting at the state level. Specifically, we look at future COVID deaths over the next 30 o n A p ril 5 , 2 0 2 1 h ttp ://a d va n ce s.scie n ce m a g .o rg / D o w n lo a d e d fro m https://github.com/nytimes/covid-19-data http://advances.sciencemag.org/ Warshaw et al., Sci. Adv. 2020; 6 : eabd8564 30 October 2020 S C I E N C E A D V A N C E S | R E S E A R C H A R T I C L E 4 of 4 and 90 days using survey data before the start of the COVID-19 pandemic, from between July 2019 and March 2020. Figure S1 shows that there is no effect of future COVID-19 deaths on political preferences. Note that the state of the art in panel research designs is con- stantly moving forward. In recent years, a number of scholars have conducted innovative work (19–21). However, to our knowledge, all of this work currently requires dichotomous treatment variables. So, overall, we believe that our design is the best available research design for our data and that our placebo checks validate a causal interpretation of our results. We have also run a number of robustness checks for our main research design and results. For simplicity, each of these robustness checks focuses on our state-level model using an index of our four outcome variables to capture aggregate political preferences. 1) First, we examine the results if we use several different num- bers of days as cutoffs rather than just 30 days. Specifically, we ex- amine cutoffs ranging from 10 to 90 days. In fig. S2, we find that the results are quite similar across models, although the point estimates decrease a bit for longer cutoffs. Overall, this suggests that our re- sults are not especially sensitive to the choice of cutoffs. They are also significant across all cutoffs. 2) Our next robustness check examines the results if we do not include any control variables in our analyses (fig. S3). We find that our results are slightly noisier without any control variables, but the results are still significant without controls. In our main analyses, we prefer to retain control variables because of the increase in effi- ciency that they provide. 3) In our main analysis, we dropped don’t knows. However, it is reasonable to think that don’t knows could be an important middle category, and voters could move into this category because of concern about COVID-19. To assess this possibility, we coded alternative variables for all our outcomes with don’t know as a middle category (0.5). Figure S4 shows the results at the state level. It indicates that the results are generally very similar with and without don’t knows, especially for the presidential race. The point estimates in Senate and House races are a bit smaller when we include don’t knows, but the results are significant both with and without don’t knows at all levels of geography. Likewise, our county level results are also simi- lar with and without don’t knows. 4) Last, we examine whether the results change if we drop each state one by one. Figure S5 shows that our results are not sensitive to dropping individual states. The point estimates are generally quite similar across models. The highest P value is in a model that drops Texas. Even in this model, however, we still find a P value of 0.02. Overall, these robustness checks indicate that our results are not sensitive to alternative regression specifications or driven by outliers. SUPPLEMENTARY MATERIALS Supplementary material for this article is available at http://advances.sciencemag.org/cgi/ content/full/6/44/eabd8564/DC1 REFERENCES AND NOTES 1. K. Karson, Approval of Trump’s coronavirus response underwater, as he returns to campaign trail: POLL. ABC News (2020). 2. S. Ashworth, Electoral accountability: Recent theoretical and empirical work. Annu. Rev. Polit. Sci. 15, 183–201 (2012). 3. A. Healy, N. Malhotra, Retrospective voting reconsidered. Annu. Rev. Polit. Sci. 16, 285–306 (2013). 4. S. S. Gartner, G. M. Segura, M. Wilkening, All politics are local. J. Confl. Resolut. 41, 669–694 (1997). 5. S. S. Gartner, Ties to the dead: Connections to iraq war and 9/11 casualties and disapproval of the president. Am. Sociol. Rev. 73, 690–695 (2008). 6. D. Karol, E. Miguel, The electoral cost of war: Iraq casualties and the 2004 U.S. presidential election. J. Polit. 69, 633–648 (2007). 7. C. R. Grose, B. I. Oppenheimer, The Iraq War, partisanship, and candidate attributes: Variation in partisan swing in the 2006 U.S. House elections. Legis. Stud. Q. 32, 531–557 (2007). 8. D. L. Kriner, F. X. Shen, Iraq casualties and the 2006 senate elections. Legis. Stud. Q. 32, 507–530 (2007). 9. D. L. Kriner, F. X. Shen, Battlefield casualties and ballot-box defeat: Did the Bush–Obama wars cost Clinton the White House? PS Political Sci. Politics 53, 248–252 (2020). 10. C. Oprysko, S. Luthi, Trump labels himself a wartime president combating coronavirus. Politico (2020). 11. C. Tausanovitch, L. Vavreck, Democracy Fund+ UCLA Nationscape Project, January 2–June 25, 2020. (2020). 12. A. Healy, G. S. Lenz, Presidential voting and the local economy: Evidence from two population-based data sets. J. Polit. 79, 1419–1432 (2017). 13. J. de Benedictis-Kessner, C. Warshaw, Accountability for the local economy at all levels of government in United States elections. Am. Polit. Sci. Rev. 114, 660–676 (2020). 14. C. Tausanovitch, L. Vavreck, T. Reny, A. R. Hayes, A. Rudkin, Democracy Fund+ UCLA Nationscape methodology and representativeness assessment (2019). 15. A. Gelman, J. Hill, Data Analysis Using Regression and Multilevel/Hierarchical Models (Cambridge Univ. Press, 2006). 16. J. D. Angrist, J.-S. Pischke, Mostly Harmless Econometrics: An Empiricist’s Companion (Princeton Univ. Press, 2008). 17. D. J. Hopkins, K. Parish, The medicaid expansion and attitudes toward the affordable care act: Testing for a policy feedback on mass opinion. Public Opin. Q. 83, 123–134 (2019). 18. A. Abadie, S. Athey, G.W. Imbens, J.Wooldridge, “When should you adjust standard errors for clustering?” (Technical Report, National Bureau of Economic Research, 2017). 19. A. Goodman-Bacon, “Difference-in-differences with variation in treatment timing” (Technical Report, National Bureau of Economic Research, 2018). 20. K. Imai, I. S. Kim, When should we use unit fixed effects regression models for causal inference with longitudinal data? Am. J. Polit. Sci. 63, 467–490 (2019). 21. Y. Xu, Generalized synthetic control method: Causal inference with interactive fixed effects models. Polit. Anal. 25, 57–76 (2017). Acknowledgments: We are grateful for feedback on this paper from D. Caughey and A. Cox. Funding: We also appreciate funding support for the Nationscape Survey from the Democracy Fund, the Klarman Family Foundation, and the UCLA Marvin Hoffenberg Chair in American Politics and Public Policy. Author contributions: L.V. secured funding and managed the fielding of the Nationscape survey. C.W. and R.B.-K. executed the data analysis. All authors participated in writing and editing the manuscript. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Replication data and code for this paper are available at the Harvard Dataverse at https://doi.org/10.7910/DVN/SN3XAM. Additional data related to this paper may be requested from the authors. Submitted 16 July 2020 Accepted 16 September 2020 Published 30 October 2020 10.1126/sciadv.abd8564 Citation: C. Warshaw, L. Vavreck, R. Baxter-King, Fatalities from COVID-19 are reducing Americans’ support for Republicans at every level of federal office. Sci. Adv. 6, eabd8564 (2020). o n A p ril 5 , 2 0 2 1 h ttp ://a d va n ce s.scie n ce m a g .o rg / D o w n lo a d e d fro m http://advances.sciencemag.org/cgi/content/full/6/44/eabd8564/DC1 http://advances.sciencemag.org/cgi/content/full/6/44/eabd8564/DC1 https://doi.org/10.7910/DVN/SN3XAM http://advances.sciencemag.org/ federal office Fatalities from COVID-19 are reducing Americans' support for Republicans at every level of Christopher Warshaw, Lynn Vavreck and Ryan Baxter-King DOI: 10.1126/sciadv.abd8564 (44), eabd8564.6Sci Adv ARTICLE TOOLS http://advances.sciencemag.org/content/6/44/eabd8564 MATERIALS SUPPLEMENTARY http://advances.sciencemag.org/content/suppl/2020/10/26/6.44.eabd8564.DC1 REFERENCES http://advances.sciencemag.org/content/6/44/eabd8564#BIBL This article cites 13 articles, 0 of which you can access for free PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions Terms of ServiceUse of this article is subject to the is a registered trademark of AAAS.Science AdvancesYork Avenue NW, Washington, DC 20005. The title (ISSN 2375-2548) is published by the American Association for the Advancement of Science, 1200 NewScience Advances License 4.0 (CC BY-NC). Science. No claim to original U.S. Government Works. Distributed under a Creative Commons Attribution NonCommercial Copyright © 2020 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of o n A p ril 5 , 2 0 2 1 h ttp ://a d va n ce s.scie n ce m a g .o rg / D o w n lo a d e d fro m http://advances.sciencemag.org/content/6/44/eabd8564 http://advances.sciencemag.org/content/suppl/2020/10/26/6.44.eabd8564.DC1 http://advances.sciencemag.org/content/6/44/eabd8564#BIBL http://www.sciencemag.org/help/reprints-and-permissions http://www.sciencemag.org/about/terms-service http://advances.sciencemag.org/