A tale of two pandemics: The enduring partisan differences in actions, attitudes, and beliefs during the coronavirus pandemic

Early in the new coronavirus disease (COVID-19) pandemic, scholars and journalists noted partisan differences in behaviors, attitudes, and beliefs. Based on location data from a large sample of smartphones, as well as 13,334 responses to a proprietary survey spanning 10 months from April 1, 2020 to February 15, 2021, we document that the partisan gap has persisted over time and that the lack of convergence occurs even among individuals who were at heightened risk of death. Our results point to the existence and persistence of the interaction of partisanship and information acquisition and highlight the need for mandates and targeted informational campaigns towards those with high health risks.


Introduction
Early in the pandemic, information about the effectiveness of different risk-mitigating strategies was scarce, ambiguous and confusing. For example, the Centers for Disease Control (CDC) initially did not recommend masks, many people were wiping down groceries. At the same time, partisan rhetoric was beginning to form about the seriousness of the pandemic.
In this era, several studies documented partisan gaps in self-reported precautionary behaviors and observed social mobility (e.g., Allcott, Boxell, Conway, Gentzkow, Thaler andYang 2020, Andersen 2020, Barrios and Hochberg 2020, Fan, Orhun and Turjeman 2020, Painter and Qiu 2020, Gadarian, Goodman and Pepinsky 2021. These scholars called for uniform public health messaging, in the hopes of eliminating the partisan gap through accurate and increased information dissemination. Over time, more information about the virus and effective strategies to curb its spread became available. Given the high stakes the pandemic presented, standard economics models would predict partisan differences in beliefs and behaviors to diminish over time. However, given the importance of initial beliefs (see Benjamin 2019 for a review) and the role of partisan identity in belief revisions (e.g., Taber andLodge 2006, Van Bavel andPereira 2018), whether partisan gaps in beliefs and behaviors narrowed in the long run is an open question.
In this paper, we document partisan gaps in behaviors, attitudes, and beliefs regarding the COVID-19 pandemic across a time span from the beginning of the pandemic to February 2021. Our analyses rely primarily on an original and nationally representative survey conducted between April 1, 2020 and February 15, 2021 covering 13,334 respondents. We also analyze geo-location data from a large sample of smartphones provided by SafeGraph to examine differences in mobility due to work and non-work trips.
We present three main findings. First, we show economically significant and persistent partisan gaps that emerge early in the pandemic and do not decrease over time. Using smartphone geolocation data, we document a remarkably stable gap in mobility across counties with more Democrats versus more Republicans from the beginning of the pandemic to February 2021. This gap is mainly driven by differences in discretionary (non-work) mobility.
Using our survey data, we examine self-reported behaviors related to discretionary mobility, as well as other precautionary behaviors, attitudes towards economic activities, and beliefs regarding individual and systemic risks. We find that the difference between Democrats and Republicans in their likelihood of avoiding large gatherings or limiting seeing friends, as well as in feeling comfortable with engaging in economic activities (e.g., eat at a restaurant, go shopping) largely remains the same throughout the 10-months of our study, even as general engagement with these activities fluctuate. Democrats are more worried about the health and the economic well-being of themselves and those around them, and increasingly so over time. Democrats are also persistently more pessimistic than Republications in their predictions regarding their own health risks and the future death toll of the pandemic in the U.S. over time. Consistent with these findings, we also find an association between individual risk perceptions and behaviors, except during periods of mandates in which partisan gaps in behaviors were reduced while partisan gaps in beliefs persisted.
Second, we document that partisan gaps are large and persistent even among individuals who are objectively at a heightened risk of severe health outcomes (65 years or older, or have pre-existing health conditions that put them in the high-risk category). For this highrisk population, the returns to acquiring more accurate information are expected to be higher. However, we find that although for some measures the partisan gap among the high-risk population is somewhat smaller than that among the low-risk population during some periods, the gaps among the high-risk population across all measures (on protective behaviors, comfort with economic activities, and beliefs) are still substantial and do not decrease over time.
Third, the partisan gaps are largely nonexistent among individuals who choose to follow a variety of news outlets spanning the political spectrum and are also greatly reduced among individuals who report paying more attention to information about the pandemic from a variety of formal and informal sources. This finding points to an interaction between partisanship and demand for information about the pandemic. The previously documented persistent partisan gaps are mainly driven by individuals who are exposed to news outlets within a narrower political spectrum and those who choose not to pay much attention to information about the pandemic.
These results imply that partisan gaps in beliefs and behaviors did not narrow in the long run, even for individuals with greater costs of being misinformed or guided by political ideology. Furthermore, partisan gaps are mainly driven by individuals who are exposed to news outlets within a narrower political spectrum and those who choose not to pay much attention to information about the pandemic. Combined with our finding that partisan gaps in actions were absent only under mandates, during which partisan gaps in worries and beliefs still remained, these results suggest that policy intervention may be warranted above and beyond providing more acute information.
This paper contributes to three strands of literature. First, this paper adds to a large literature on partisan gaps or partisan bias in economics. Examples include Leigh (2008)  Second, this paper contributes to a nascent literature providing field evidence of persistent belief biases. Although experimental studies have shown that initial beliefs and partisan identity influence individuals' learning process, it is unclear whether the systematic biases discussed in this literature persist in the long run and exist in high-stakes contexts -a question that by its nature needs to be examined in the field. Several field studies have shown evidence of biased beliefs in contexts where individuals presumably observe signals that should challenge their beliefs. For example, Park and Santos-Pinto (2010), Hoffman and Burks (2020), and Huffman, Raymond and Shvets (2021) document persistent overconfidence among chess players chess players, truckers, and managers. We contribute to this literature by examining the extent to which partisan belief differences evolve over a long period of time in an information-rich environment where mistakes can have life or death consequences.
Finally, this paper is related to the recent literature studying individual behaviors and beliefs during the COVID-19 pandemic. Examples include Allcott et al. (2020), Andersen (2020), Painter and Qiu (2020), Bundorf et al. (2021), Gadarian, Goodman and Pepinsky (2021), and Heffetz and Ishai (2021). We contribute to the literature by documenting partisan gaps in a large set of measures on actions, attitudes, worries, and beliefs. We also show that despite more acute information over time, these gaps are stable and significant even among the most vulnerable individuals.
The rest of the paper proceeds as follows. We document partisan gaps over time in the observed mobility in Section 2. In Section 3, we show persistent partisan gaps in selfreported behaviors, attitudes, and beliefs and discuss the correlation between beliefs and behaviors/attitudes. We also show that the partisan gaps are stable even for the high-risk population, but do not exist for individuals who follow news outlets across the political spectrum or pay a lot of attention to pandemic-related information. Section 4 concludes.
2 Observed Mobility Differences: SafeGraph Data Results In this section, we present partisan differences in mobility and how such differences persist over time as a starting point for our investigation. The data come from SafeGraph, which is a data company that aggregates anonymized location data from numerous smartphone applications to provide insights about physical places (www.safegraph.com). Using geo-location data obtained from the activity of more than 35 million smartphones, SafeGraph provides an aggregated dataset (Social Distancing Metrics) to researchers. For each census block on a given day, the Social Distancing Metrics data report the total number of devices, the number of devices exhibiting full-time work behavior (spending more than 6 hours between 8 a.m. and 6 p.m. in one location outside the home), devices exhibiting part-time work behavior (spending more than 3 but less than 6 hours between 8 a.m. and 6 p.m. in one location outside the home), and devices that stay home all day. 1 We merge election results from the 2016 presidential election, demographic information, daily local COVID-19 cases and deaths, and daily minimum and maximum temperature, precipitation, and wind speed data after aggregating these data to the county level. 2 We examine how the shares of devices that left home for various reasons (for any reason, for full-time work, for part-time work, and for non-work reasons) vary with county 1 SafeGraph determines the home of a device by its common nighttime location over a 6-week period. 2 We exclude Alaska and territories of the U.S. Moreover, following SafeGraph notifications, we drop two dates (2/25/2020 and 1/27/2021) known to have missingness. To enhance privacy, SafeGraph excludes a census block group if it has fewer than five devices observed in a month. We obtain county election results from the 2016 presidential election (from MIT Election Data and Science Lab (2018)), county COVID-19 case and death numbers from The New York Times (2020), demographic information from Killeen et al. (2020), and weather data on temperature, precipitation, and wind speed from gridMET maintained by Abatzoglou (2013). demographics between February 2, 2020 and February 7, 2021. Our main specification is: where s ct is the share of devices in county c on day t exhibiting a particular behavior and P c is the percent voting for Clinton in the 2016 election (termed "Democrats" in the discussion of results). The coefficients of interest, β w(t) , are allowed to vary by week w(t). The control variables Z ct are also allowed to vary weekly in their influence. They include demographic and socioeconomic controls, daily weather controls, and variables to capture differences in COVID-19 risks across counties. 3 We include county fixed-effects (ζ c ) to control for crosssectional differences and state-date fixed effects (η s(c)t ) to control for state-wide changes in social-distancing policies. Because each observation is a share in a county, we weight observations by the total number of devices in each county. We also cluster errors at the county level. Figure 1 presents the estimated coefficient β w(t) . Each column of the figure corresponds to a share of devices exhibiting a certain behavior: leaving home for any reason, for non-work From the first column of Figure 1, we can see that while partisan differences in mobility were nonexistent during February 2020, these differences emerge quickly in early March and 3 Demographic control variables are population share of women, percent living under the poverty line, population shares of whites, blacks, and Asians, percent of adults with a bachelor's degree, percent of adults without a high-school degree, unemployment rate, and the employed share of the county population. Weather controls include daily minimum and maximum temperature, precipitation, and wind speed. Pandemic risk proxies include the logarithm of one plus the number of COVID-19 cases, logarithm of one plus the number of COVID-19 deaths, the logarithm of population density, population share over the age of 65, and a binary variable indicating whether the county is in a metropolitan area.
4 These shares are computed using the total number of devices in the sample as the denominator and the number of devices exhibiting certain behaviors as the numerators. SafeGraph reports a sampling bias in favor of detecting devices that are moving (SafeGraph 2020). Supplemental Appendix SC.2 shows the robustness of our conclusions to several different adjustment approaches. 5 https://www.washingtonpost.com/outlook/2020/10/01/debate-early-travel-bans-china/ remained stable until the end of our sample. The estimated β w(t) hovers around -9 throughout the study period. To put this estimate into context, we note that the difference between the 95th and 5th percentiles of Democratic vote share across counties is 0.5. Therefore, an estimated value of -9 for the coefficient β w(t) implies that going from a county with the 5th to the 95th percentile in Democratic vote share is associated with a decrease of 4.5 percentage points in the share of devices staying at home. Such a partisan gap is sizable compared to the overall change in mobility. During the first week of the sample, 74.6% of devices left home, while in July and August of 2020, 67.8% of devices left home on average. Therefore, the partisan gap of 4.5 percentage points is equivalent to 66% of the mobility reduction between the first week of the sample to the summer of 2020.
Looking into various reasons why people leave home, we can see that the partisan gap in discretionary (non-work) mobility (the second column of Figure 1) dominates partisan gaps in work-related mobility (the third and fourth columns of Figure 1) and the former drives the persistent partisan gap in the overall mobility. In particular, the partisan gap in social mobility for non-work purposes starts in early March, stabilizes in April, and remains at that level throughout the summer of 2020. Afterwards, when the school year starts and when many workers returned to their workplaces as economic activities resume across the nation, the partisan gaps in non-work mobility becomes smaller but never disappears. Turning to work-related mobility, we find that partisan gaps are much smaller, but also display patterns consistent with Democrats reducing mobility more when they can. Specifically, we find that people in counties with a higher share of Democrats are more likely to leave home for work when the economy is closed (before the end of the summer) and during holidays (Thanksgiving, Christmas), but less likely to do so as a larger part of the population returns to their workplaces (after the summer). While the former pattern is consistent with the possibility of a higher representation of Democrats in front-line professions that require working outside the home (e.g. healthcare, retail, and grocery stores), 6 the contrast between the former and the latter patterns means that barring constraints imposed by workplaces 6 In Supplemental Appendix SC.1, we confirm a positive correlation between the share of most front-line occupations and Democratic vote share in a county, and report results from a specification that includes occupation share controls. Results are presented in Figure SC.1 As expected, the partisan gaps in mobility associated with work trips during the early phase of the pandemic are less pronounced with these controls.  Notes: This figure plots the estimated coefficient β w(t) in equation (1) and the corresponding 95% confidence interval where the dependent variable is indicated at the top of each panel. The week of February 1, 2020 is taken as the baseline t = 0. The y-axis indicates the beginning date of the week for which the coefficients are reported. Observations are weighted by the number of candidate devices in the county, and standard errors are clustered at the county level. and profession, Democrats seem more likely to reduce mobility for work purposes.
In sum, our results suggest that partisan gaps in overall mobility are economically meaningful and persistent over time. These gaps are mainly driven by non-work-related mobility, which tends to be more discretionary. In the next section, we examine self-reported behaviors related to discretionary mobility, as well as other outcomes of interest. The survey asked respondents about their adoption of behaviors that help to curb the spread of the virus, comfort with engaging in certain activities after stay-at-home mandates were lifted, worries regarding health and economic well-being, and beliefs regarding the impact of the virus. Below, we detail these items. Supplemental Appendix SD presents the entire survey.
Protective Behaviors The survey elicited adherence to recommended health precautions by asking "Which of the following changes have you personally made to protect yourself from the coronavirus infection?" for a set of behaviors including: (1) do not meet friends or extended family; (2) avoid large gatherings and public transportation; (3) wash hands more often; and (4) wear a mask when out and about. The first two are social distancing actions that may contribute to the mobility differences documented in Section 2, and the latter two are health precautions individuals were encouraged to take. All four actions are difficult to glean from observational data.
Comfort with Economic Activities From June 2020 on, as most of the stay-at-home orders were lifted across the nation, we asked respondents whether they "feel comfortable in 7 In particular, we used https://luc.id/theorem/ to target a nationally representative sample. In each period, we fielded waves of the survey on certain dates. Those dates and period sample sizes are as follows: In 2020, Early April (April 1,8,15;N=3,483), Late April (April 22, 29; N=2,002), June (June 1, 15; N=2,515), August (August 30; N=2,534) and in 2021, February (February 15, N=2,800). An independent academic study, Coppock and McClellan (2019), examined the validity of Lucid in terms of respondent characteristics and treatment effect estimates and concluded that "subjects recruited from the Lucid platform constitute a sample that is suitable for evaluating many social scientific theories, and can serve as a drop-in replacement for many scholars currently conducting research on Mechanical Turk or other similar platforms." 8 The University of Michigan Institutional Review Board (IRB) reviewed the surveys and determined that they are exempt from ongoing review (HUM00148129, HUM00180582). The IRB has also approved the merge between survey responses and county-level data (HUM00180640). engaging in" the following activities: (1) eat in a restaurant with indoor seating; (2) eat in a restaurant with outdoor seating; (3) be part of a gathering with more than 10 people; (4) go to a coffee shop; (5) go to a bar; (6) go to a gym; (7) go grocery shopping; and (8) go shopping for non-food items.
Worries The survey also asked how worried respondents felt about their own health, and the health of their partner, kids, extended family, members of their community, and the whole U.S. (on a scale of 1 to 5, 1 being not worried at all, 5 being extremely worried). It also asked how worried they feel about the economic well-being of the same groups of people (using the same scale). In the regressions, we use standardized (z-score) worry measures as dependent variables for ease of comparability.
Beliefs The survey elicited two types of beliefs: (1) Individuals' predictions on their own health risks, i.e., the chance of becoming infected in the next three months and chance of having no/mild or serious symptoms should they get infected. (2) Their predictions of the number of U.S. deaths by a certain target date assuming the state policies remain the same. 9 Again, we use standardized expectations as dependent variables.

Persistent Partisan Gaps
To study partisan differences across time in these measures of protective behaviors, attitudes towards a variety of economic activity, worries, and beliefs, we use the following regression equation: where Y iτ is an outcome variable of interest as explained above for person i at time τ . P i is a vector of indicator variables representing person i's political affiliation (Democrat, Republican, or Independent). We allow the impact of partisanship, captured by α τ , to vary over time. The control variables X iτ include individual-specific covariates as summarized in Table SA.2 in Supplemental Appendix SA. They include demographic controls (categor-9 The target date was July 1, 2020 in the April waves, September 1, 2020 in the June waves, December 1, 2020 in the August wave, and May 1, 2021 in the February wave. ical variables for the respondent's gender, race, age, educational level, annual household income, presence of children) and an indicator for whether the respondent has any of the chronic conditions listed by the CDC as a high-risk factor. Control variables X iτ also include geography-specific covariates that proxy for local pandemic severity over time (natural logarithm of one plus the number of cumulative COVID-19 cases in the respondent's county at the time of the survey, natural logarithm of one plus the number of cumulative deaths in the respondent's county at the time of the survey, natural logarithm of the population density of the respondent's zip code). We include state-time fixed effects, µ s(i)τ , to capture systematic differences across states at the time. Democrats and Republicans in the probability of taking protective actions. 10 From this panel, we can see that although in the early days of the pandemic partisan gaps in social distancing and washing hands are negligible, they quickly grow to a substantial level by the end of April 2020 and remain high even by February 2021. The absence of partisan gaps in social distancing early in the pandemic correspond with stay-at-home mandates many states had in place. As these restrictions are lifted, partisan gaps emerge because although both Democrats and Republicans became more likely to socialize with family and friends, the increase is more pronounced for Republicans, as can be seen from the raw response patterns documented by Figure A in the Appendix. Regarding mask-wearing, the partisan gaps are already present early in the pandemic, grow larger during the summer of 2020, and cease to exist by February 2021. By 2021, most businesses and local governments had mandated masks indoors.
Overall, these results show that partisan gaps in protective behaviors are relatively small in the beginning, increase to substantial levels quickly and remain persistent throughout the pandemic, even as the public health officials repeatedly confirm that these protective actions are highly effective. Notable exceptions are times in which certain behaviors are restricted under mandates. The fact that the partisan differences are eroded during times of mandates is consistent with previous findings that government-imposed social distancing measures reduced the spread of COVID-19 (see, for example, Courtemanche et al. 2020). Figure 2 presents partisan differences in comfort with a variety of economic activities. Democrats are less likely to feel comfortable with engaging in economic activities overall. Except in the case of grocery shopping, the partisan gap is consistently between 10% and 20% across these activities, and stably so over time. 11 These partisan gaps are substantial in magnitude. For example, over time, the overall share of respondents reporting feeling comfortable with dining indoors increased from 29% in June 2020 to above 42% in February 2021 (see Figure A). A partisan gap of 20% is, therefore, equivalent to about half of the overall change in comfort with indoor dining over time.

Panel B of
As Panel C of Figure 2 shows, from the beginning of the pandemic to the last period of our survey, Democrats report being more worried than Republicans about health and economic well-being, ranging from their own health or economic well-being to their partner's, local community's, and the whole nation's. The partisan gap in health worries varies between around 20-40% of a standard deviation, and increases over time. The partisan gap in economic well-being worries is generally positive, though smaller in magnitude compared to the gap in health worries.
Finally, Panel D of Figure 2 presents partisan gaps in predictions regarding own or systemic health risks. From the figure, we can see that while the magnitude of the gaps varies over time, the gaps are almost always present: Democrats are consistently more pessimistic than Republicans. For example, despite being more likely to take precautionary actions, Democrats' prediction of their chance of becoming infected in the next three months from the time of the survey is, on average, 0.1 to 0.2 standard deviations higher than that of Republicans. Their predictions about their chance of getting serious symptoms conditional on getting infected are also 0.1 to 0.2 standard deviations higher. In terms of systemic risks, Democrats on average predict more deaths due to COVID-19 than Republicans by around 0.1 standard deviations.
Overall, the above results show that throughout the 10-month period, Democrats were consistently more worried about the health and economic impact of the pandemic, more  Notes: This figure plots the estimated Democrat -Republican partisan gaps obtained from the estimates of α τ in equation (2) and the corresponding 95% confidence intervals. The x-axis indicates the period τ . In Panel A, a positive estimate means that, ceteris paribus, Democratic respondents are more likely than Republican respondents to have taken an action indicated in the legend. The actions studied are "Wash Hands"-wash hands more often; "Wear Mask"-wear a mask when out and about; "Not See Friends"-do not meet any friends or extended family; "Avoid Gatherings"-avoid public transportation and large gatherings. In Panel B, a positive estimate means that, ceteris paribus, Democratic respondents are more likely than Republican respondents to feel comfortable with an activity indicated in the legend. Activities studied are "Restaurant, In"-eat in a restaurant with indoor seating; "Restaurant, Out"-eat in a restaurant with outdoor seating; "10+ ppl"-be part of a gathering with more than 10 people; "Cafe"-go to a coffee shop; "Bar"-go to a bar; "Gym"-go to a gym; "Shopping"-go shopping for non-grocery items; "Grocery"-go grocery shopping. In Panel C, a positive estimate means that, ceteris paribus, Democratic respondents worry more about the health well-being (in the left graph) or the economic well-being (in the right graph) of the group of people indicated in the legend. The groups of people are "Self"-respondent herself; "Partner"-respondent's partner; "Children"-respondent's kids; "Ext. Family"-respondent's extended family; "Comm."-members of the respondent's community; "U.S."-all people in the U.S. In Panel D, a positive estimate means that, ceteris paribus, Democratic respondents predict a larger number on the outcomes indicated in the legend.
Outcomes over which expectations are elicited are "U.S. Deaths"-total number of deaths in the U.S. by a target date; "Chance of Infection"-chances that the respondent will get infected with the coronavirus in the next three months; "Chance of Serious Illness"-chances that the respondent will have serious symptoms should she get infected. All measures in Panels C and D are z-scores.
cautious about engaging in economic activities, and congruently more likely to take social distancing and preventive actions.
These findings are consistent with significant and stable associations between individuals' individual risk beliefs and their actions and attitudes. We find a positive association between beliefs regarding the severity of disease and whether or not a person engages in a given protective behavior and a negative correlation between disease severity beliefs and feeling comfortable engaging in economic activity (Figure SA.2 in Supplemental Appendix SA).
For example, we find that one standard deviation increase in the perceived chance of serious illness is associated with a 2.9% -8.5% increase in self-reported refrain from seeing friends across the 10-month span, and a 5%-7.6% decrease in whether the individual is comfortable with dining indoors. Moreover, the associations between risk perceptions and protective behaviors and comfort with economic activities are stable over time. An exception is the correlation between beliefs and wearing masks. In the early periods of the pandemic, the correlation is substantial, but as mask mandates are instituted, this behavior-belief correlation disappears.
Earlier work has also documented correlations between beliefs and attitudes. In particular, Bundorf et al. (2021)  we focus on severity beliefs instead of infection risks beliefs because the latter may be a result of individual actions (e.g., one thinks they are less likely to be infected because one avoids socialization). We also present evidence of a sustained association over time. Therefore, our results should be interpreted as extending prior evidence both temporally, and using a different risk belief that is both individually relevant and less susceptible to reverse causation.
While we refrain from a causal interpretation, we note that a sustained association between behaviors and beliefs over a 10-month span is suggestive of the possibility that the lack of convergence in risk assessments is at least partially responsible for the lack of convergence in behaviors across the partisan line.
The association between individuals' expectations of severe health outcomes and their behaviors naturally leads to a question of whether partisan gaps are reduced among individuals with conditions that put them at a high risk for complications. We explore this question in the next section.

Partisan Gaps Among At-Risk Population
Having established stable partisan gaps over time, we now examine whether the gap and its persistence vary by health risks. Since the potential downsides of being misinformed or guided by political ideology are greater for high-risk individuals, one would expect high-risk individuals to have more incentives to respond to risks rather than ideology. Therefore, partisan gaps are expected to be smaller among the high-risk individuals. In addition, because information about the asymmetric impact of the virus on high-risk individuals became more wide-spread over time, one might also expect the partisan gap to shrink over time for the high-risk group.
We consider an individual to be a higher risk (HiRisk i = 1) if they are either 65 years or older, or have at least one high-risk health condition. The CDC noted age as a risk factor.
The CDC also published a list of health conditions that are associated with more severe health outcomes. 13 According to this definition, 57% of our respondents are considered high risk. We run the following regression to evaluate the heterogeneity in partisan gaps: where Y iτ (the outcome variable of interest), P i (the political affiliation dummy variables), and µ s(i)τ (the state-period fixed effects) are the same as those in equation (2). The control variables X iτ are also the same as those in equation (2), except that we now separate out HiRisk i from the vector X iτ and interact HiRisk i with the individual's political affiliation 13 The list of diseases were obtained from the CDC website on March 30, 2020: Moderate to severe asthma; COPD or other chronic lung diseases; Serious heart conditions; Diabetes; Conditions that can cause a person to be immuno-compromised including cancer treatment, smoking, bone marrow or organ transplantation, immune deficiencies, poorly controlled HIV or AIDS, and prolonged use of corticosteroids and other immune weakening medications; Severe obesity (BMI of 40 or higher); Chronic kidney disease and currently undergoing dialysis; Liver disease. P i . Therefore, while the coefficient α τ gives the partisan gaps for low-risk individuals, the sum α τ + γ τ measures that for high-risk individuals. Figure 3 plots the estimated Democrat -Republican partisan gap for low-risk individuals with hallow markers, and the gap for high-risk individuals with solid markers. The estimates reveal two patterns: first, across all variables of interests (precautionary actions, attitude towards economic activities, worries, and beliefs), the partisan gap among high-risk individuals is mostly statistically the same as (and sometimes smaller than) the partisan gap among low-risk individuals. Second, partisan gaps among high-risk individuals are substantial and, more importantly, stable over time. For example, high-risk Democrats are more likely to avoid seeing friends than high-risk Republicans by 10 percentage points (Panel A) and less likely to feel comfortable with a gathering with more than ten people by more than 10 percentage points (Panel B) throughout the sample. They are also more worried (Panel C) and hold more pessimistic beliefs (Panel D) than their Republican counterparts. These results are alarming as they indicate that as information about the disease becomes more accurate and widespread, partisan gaps remained strong even for the most vulnerable individuals.

Role of Information
Could the persistent partisan gaps in behaviors and beliefs be related to partisan gaps in information demand? Do the partisan gaps differ across individuals with different attention levels to pandemic-related information and different scopes of news exposure? In this section, we address these questions. In doing so, we shed some light on the interaction between partisanship and information demand.
Our survey asked individuals to indicate the top news outlets they regularly consume.
The survey also asked respondents about the amount of attention they pay to various information sources spanning from scientists/researchers to friends and family members regarding the pandemic specifically. The respondents rated their attention level on a 5-point scale ranging from 1 " not at all" to 5 "very much so." Summary statistics can be found in Notes: This figure plots the estimated Democrat -Republican partisan gaps for low-risk and high-risk respondents. In each panel, a hallow marker gives the estimated α τ in equation (3), which is the partisan gap among low-risk respondents, and a solid marker gives the estimated α τ + γ τ , which is the partisan gap among high-risk respondents. The segments give the corresponding 95% confidence intervals. The x-axis indicates the period τ . The legends are the same as those in Figure 2.
outlets and who people pay more attention to when it comes to the pandemic. 14 There are also some similarities. A considerable fraction of people consume news across the partisan line: 46% of Democrats indicate Fox News as a source they regularly watch, and 42% of Republicans consume a more liberal news outlet in addition to Fox News. 15 Moreover, both Democrats and Republicans vary greatly in the amount of overall attention they pay to others in terms of the pandemic, calculated as the average attention score across all items.
First, we examine partisan gap differences among individuals who consume news across the partisan line versus those who consume a politically narrow scope of news. We use 14 Republicans are more likely to consume Fox News, and to pay greater attention to their pastor, people they follow on Facebook or Twitter, and the President. Democrats are more likely to consume any news outlet except Fox News, and to pay greater attention to the CDC, scientists, and their Governor. There are no partisan differences in the level of attention paid to family or friends. 15 We consider CNN, NBC/MSNBC, NPR, Huff Post, The New York Times and Washington Post as liberal news outlets, but the results are robust to narrower classifications.
an equation analogous to equation (3) where we replace HiRisk with whether the individual consumes news across the partisan line, and report the estimation results in Figure 4.
Throughout the period we study, partisan gaps in precautions taken, comfort with economic activities, health worries, and beliefs among individuals who choose to consume news across the partisan line (indicated by solid markers) are small in magnitude and oftentimes insignificant. In contrast, while partisan gaps in actions and health worries are mostly non-existent in early April, they emerge quickly among individuals who do not consume news across the partisan line (indicated by hallow markers). Moreover, partisan gaps in comfort with economic activities and beliefs are also large and significant in this group. These results, combined with the initial partisan differences being very small in early April, lends support to the role of information. In particular, the willingness to expose oneself to a variety of (potentially conflicting) information seems to play a role in reducing partisan gaps.
Next, we examine whether the willingness to pay attention to information about the pandemic is associated with a lower level of partisan gap. The results in Figure 5 show that among individuals who pay a high level of attention to information from others regarding the pandemic (average attention score greater than 3.5 on a scale of 1 to 5), the partisan gaps in precautions taken, comfort with economic activities, health worries and beliefs are economically small and oftentimes insignificant. In contrast, individuals who report not paying much attention overall to others as a source of information about the pandemic show substantial partisan gaps in all metrics.
To summarize, in Section 3.2, we find significant and stable partisan gaps even among high-risk individuals. As argued before, such individuals are expected to have more incentives to acquire accurate information, update their beliefs, and adjust their behaviors. This result indicates these individuals do not necessarily acquire more information or do not acquire enough information to overcome the effect of partisanship. The heterogeneity result in this section suggests that for individuals who do pay attention to pandemic-related information and consume news across the political spectrum or are willing to do so, partisan gaps become narrower or disappear. These results, overall, point to the role of information in partisan gaps.  Notes: This figure plots the estimated Democrat -Republican partisan gaps for respondents who consume news across the political line and for those who do not. In each panel, a hallow marker gives the estimated α τ in equation (3), which is the partisan gap among consumers of narrow news, and a solid marker gives the estimated α τ + γ τ , which is the partisan gap among consumers of news across the political line. The segments give the corresponding 95% confidence intervals. The x-axis indicates the period τ . The legends are the same as those in Figure 2.  Notes: This figure plots the estimated Democrat -Republican partisan gaps for respondents whose average attention score on a 1-5 scale across other information sources (Friends, Family, Scientists, Pastor, Facebook or Twitter, CDC, Governor, President) is greater than 3.5, and those whose average attention score is 3.5 or lower. In each panel, a hallow marker gives the estimated α τ in equation (3), which is the partisan gap among those who do not pay a lot of attention to other information sources, and a solid marker gives the estimated α τ + γ τ , which is the partisan gap among those who pay high attention overall. The segments give the corresponding 95% confidence intervals. The x-axis indicates the period τ . The legends are the same as those in Figure 2.

Conclusion
We find economically significant and persistent partisan gaps in actions, attitudes, worries, and beliefs. The partisan gaps in the propensity to take protective actions (e.g., avoid seeing friends) and feel comfortable with a variety of economic activities (e.g., indoor dining) are persistent over time. Such differences are not only potentially harmful to the individuals themselves, but may also introduce substantial negative spillovers on public health. Moreover, we find that both the partisan gaps and the lack of convergence over time are present even for the high-risk population. However, such gaps are almost non-existent among those who pay attention to others regarding the pandemic and/or consume news from a politically broad range of outlets.
We document the persistence of systematic differences in a high-stakes context that was characterized by a paucity of information early on, which allowed for initial belief heterogeneity. In such a context, as highly salient and increasingly accurate information becomes available, models of rational information acquisition and updating would predict steep learning and fast convergence. It is, therefore, all the more telling that initial partisan gaps in beliefs and behaviors persist, even among high-risk individuals.
Taken together, our results point towards the need for mandates to align individual behaviors for the public good. Early in the pandemic, researchers (including us) documenting partisan differences in risk assessments and behaviors highlighted the need for consistent public messages about the COVID-19 pandemic to achieve an effective public health response.
Yet, even as more accurate information because easily accessible, partisan gaps remained stable over 10 months. Our results suggest that persistent partisan gaps are driven by individuals who are exposed to news outlets within a narrower political spectrum and individuals who pay less attention to information from others regarding the pandemic. These results suggest a role for information resistance in the persistence of partisan gaps. Therefore, we caution researchers and policy-makers that consistent public messaging and the availability of accurate information are not enough. Along with the well-known political divide in news consumption and the persistent slant in news supply (e.g., Gentzkow and Shapiro (2011) and Bakshy et al. (2015)), these results further underscore the difficulty of relying on free-market news and individual rationality dissemination to eradicate partisan gaps over time. Together with the fact that times in which mask-wearing mandates were common correspond to erosion of partisan gaps in mask-wearing, and times in which social mobility was restricted (early April) correspond to lower levels of partisan gaps in mobility and self-reported socialization suggests that mandates might play an important role in eliminating response differences.  Hands"-wash hands more often; "Wear Mask"-wear a mask when out and about; "Not See Friends"-do not meet any friends or extended family; "Avoid Gatherings"-avoid public transportation and large gatherings. In Panel B, the y-axis gives the fraction of respondents who feel comfortable with an activity indicated in the legend. The activities are "Restaurant, In"-eat in a restaurant with indoor seating; "Restaurant, Out"-eat in a restaurant with outdoor seating; "10+ ppl"-be part of a gathering with more than 10 people; "Cafe"go to a coffee shop; "Bar"-go to a bar; "Gym"-go to the gym; "Shopping"-go shopping for non-grocery items; "Grocery"-go grocery shopping. In Panel C, the y-axis is how much respondents worry (on a scale of 1 to 5, larger numbers corresponding to more worry) about the health or economic well-being of people indicated in the legend. The groups of people are "Self"-respondent herself; "Partner"-respondent's partner; "Children"-respondent's kids; "Ext. Family"-respondent's extended family; "Community"-members of the respondent's community; "U.S."-all people in the U.S. In Panel D, the y-axis is respondents' prediction on the outcomes indicated in the legend. Outcomes over which expectations are elicited are "U.S. Deaths"total number of deaths in the U.S. by a target date (in thousands); "Chance of Infection"-chances that the respondent will get infected with the coronavirus in the next three month (in %); "Chance of Serious Illness"-chances that the respondent will have serious symptoms should she get infected (in %). Lucid provided demographic variables of respondents.   Notes: This figure plots the estimated Independent -Republican partisan gaps obtained from the estimates of α τ in equation (2) and the corresponding 95% confidence intervals. The x-axis indicates the period τ . In Panel A, a positive estimate means that, ceteris paribus, independent respondents are more likely than Republican respondents to have taken an action indicated in the legend. The actions studied are "Wash Hands"-wash hands more often; "Wear Mask"-wear a mask when out and about; "Not See Friends"-do not meet any friends or extended family; "Avoid Gatherings"-avoid public transportation and large gatherings. In Panel B, a positive estimate means that, ceteris paribus, independent respondents are more likely than Republican respondents to feel comfortable with an activity indicated in the legend. Activities studied are "Restaurant, In"-eat in a restaurant with indoor seating; "Restaurant, Out"-eat in a restaurant with outdoor seating; "10+ ppl"-be part of a gathering with more than 10 people; "Cafe"-go to a coffee shop; "Bar"go to a bar; "Gym"-go to a gym; "Shopping"-go shopping for non-grocery items; "Grocery"-go grocery shopping. In Panel C, a positive estimate means that, ceteris paribus, independent respondents worry more about the health well-being (in the left graph) or the economic well-being (in the right graph) of the group of people indicated in the legend. The groups of people are "Self"-respondent herself; "Partner"-respondent's partner; "Children"-respondent's kids; "Ext. Family"-respondent's extended family; "Comm."-members of the respondent's community; "U.S."-all people in the U.S. In Panel D, a positive estimate means that, ceteris paribus, independent respondents predict a larger number on the outcomes indicated in the legend. Outcomes over which expectations are elicited are "U.S. Deaths"-total number of deaths in the U.S. by a target date; "Chance of Infection"-chances that the respondent will get infected with the coronavirus in the next three months; "Chance of Serious Illness"-chances that the respondent will have serious symptoms should she get infected. All measures in Panels C and D are z-scores.  Notes: This figure plots the correlation between an individual's belief about her chance of getting serious symptoms should she gets infected and her likelihood of taking a certain protective action (in the upper panel) and feeling comfortable with a certain economic activity (in the lower panel). The x-axis indicates the period. In the upper panel, the actions studied are "Wash Hands"-wash hands more often; "Wear Mask"-wear a mask when out and about; "Not See Friends"-do not meet any of friends or extended family; "Avoid Gatherings"-avoid public transportation and large gatherings. In the lower panel, activities studied are "Restaurant, In"-eat in a restaurant with indoor seating; "Restaurant, Out"-eat in a restaurant with outdoor seating; "10+ ppl"-be part of a gathering with more than 10 people; "Cafe"-go to a coffee shop; "Bar"-go to a bar; "Gym"-go to a gym; "Shopping"-go shopping for non-grocery items; "Grocery"go grocery shopping.  Notes: This figure plots the correlation between an individual's belief about her chance of getting infected in the next three months and her likelihood of taking a certain protective action (in the upper panel) and feeling comfortable with a certain economic activity (in the lower panel). The x-axis indicates the period. In the upper panel, the actions studied are "Wash Hands"-wash hands more often; "Wear Mask"-wear a mask when out and about; "Not See Friends"-do not meet any of friends or extended family; "Avoid Gatherings"avoid public transportation and large gatherings. In the lower panel, activities studied are "Restaurant, In"eat in a restaurant with indoor seating; "Restaurant, Out"-eat in a restaurant with outdoor seating; "10+ ppl"-be part of a gathering with more than 10 people; "Cafe"-go to a coffee shop; "Bar"-go to a bar; "Gym"-go to a gym; "Shopping"-go shopping for non-grocery items; "Grocery"-go grocery shopping.  Notes: This figure plots the correlation between an individual's prediction about the future death toll of the pandemic in the U.S. and her likelihood of taking a certain protective action (in the upper panel) and feeling comfortable with a certain economic activity (in the lower panel). The x-axis indicates the period. In the upper panel, the actions studied are "Wash Hands"-wash hands more often; "Wear Mask"-wear a mask when out and about; "Not See Friends"-do not meet any of friends or extended family; "Avoid Gatherings"avoid public transportation and large gatherings. In the lower panel, activities studied are "Restaurant, In"eat in a restaurant with indoor seating; "Restaurant, Out"-eat in a restaurant with outdoor seating; "10+ ppl"-be part of a gathering with more than 10 people; "Cafe"-go to a coffee shop; "Bar"-go to a bar; "Gym"-go to a gym; "Shopping"-go shopping for non-grocery items; "Grocery"-go grocery shopping. Notes: The upper panel reports the percentage (among all 13,334) of respondents who consume a particular news outlet. The bottom panel reports the average rating of the degree of attention individuals pay to different sources of information (ranging from 1 -Not at all to 5 -Very much so), if they have indicated that source as applicable. The number of individuals who indicate the source as applicable is reported in parentheses. The attention question was not included in the first wave, therefore 12,852 respondents received this question.

SC Robustness: SafeGraph Data Analyses
We conduct two sets of robustness analyses to show how our results in Section 2 vary with different adjustment methods and control variables.

SC.1 Adding Occupational Controls
We noted in the main manuscript that the representation of Democrats and Republicans may systematically differ across occupations. Since workers in some occupations, most notably front-line ones, may have less flexibility to work from home, such differences may contribute to the partisan gaps we observe in mobility for work purposes, especially early on in the pandemic. To fully account for these differences, we would need individual-level data, which is not possible given the anonymity of Safegraph data. Therefore, we obtain county-level occupational share data from the United States Census Bureau's American Community Survey 5-year Estimates (5-year ACS) for 2016-2019. Table SC.1 provides average shares of occupations across the 3,110 counties in our data, and the raw correlation between the shares of these occupations and the Democratic vote share in a county. Notes: The first column reports the average share of each occupation type across counties. The second column reports coefficients (and associated standard errors in parantheses) obtained from regressing the county occupation shares onto county democratic voter shares.

Part-time Work
Notes: This figure plots the estimated coefficient β w(t) in equation (1) and the corresponding 95% confidence interval where the dependent variable is indicated at the top of each panel, and additional occupation share controls are included in the regression. The week of February 1, 2020 is taken as the baseline t = 0. The y-axis markers indicate the beginning date of the week for which the coefficients are reported. Observations are weighted by the number of candidate devices in the county, and standard errors clustered at the county level.

SC.2 Different Adjustment Methods
SafeGraph reports the number of devices that "pinged" during a given day (active device count). In the main text, we measure the share of devices that remain at home as the unadjusted share s at-home = 100 × (completely-at-home devices/active devices). Safegraph notes that the number of completely-at-home devices could be under-reported due to a sampling bias. SafeGraph (2020) reports: "GPS data from smartphones is often subject to a sampling bias in favor of devices that are changing locations (i.e., moving). Collecting GPS data is battery-intensive, and software applications sometimes implement GPS data collection methods that depend on the movement of the device, rather than a fixed time interval.
This represents a sampling bias in favor of detecting devices that are moving." SafeGraph therefore also reports the number of all devices in its sample during a month, regardless of whether it saw any activity for them on a specific day within the month (candidate device count). However, it is not clear whether this number reflects the number of devices that could have been reporting on a given day, since Safegraph's sample of phones dynamically evolves over time.
We explore the robustness of our results to the following two alternative approaches of calculating shares of devices showing different types of activity: 1. Using the maximum number of devices that pinged in a week as the denominator and basis for adjustment. In this approach, we calculate the largest number of active devices each week for each county ("max active device"), assume it to be the latent true number of active devices for that week. As a result, the number of at-home devices are adjusted by the difference (max active devices -active devices) to account for potentially latent at-home devices that did not ping. We define the dependent variables as: s part-time = 100×( part-time work devices max active devices ), s full-time = 100×( full-time work devices max active devices ), s at-home = 100 × ( completely-at-home devices+(max active devices -active devices )} 2. Using candidate devices as the denominator and basis for adjustment. In this approach, we assume the number of candidate devices to be the latent true number of active devices for each day, and assume that any non-active device was at-home. As a result, the number of at-home devices are adjusted by the difference (candidatedevices − activedevices) to account for potentially latent at-home de-vices that did not ping. We define the dependent variables as: s part-time = 100 × ( part-time work devices max active devices ), s full-time = 100 × ( full-time work devices max active devices ), s leave-home = 100 − s at-home , where s at-home = 100×( completely-at-home devices+(candidate devices -active devices )} candidate devices ), and s not-work = 100 − s at-home − s part-time − s full-time . The lower part of Figure SC.2 reports the estimates.
Although the results change across specifications, the general message of a persistent partisan gap in social mobility remains the same.

Survey Consent
The survey starts with the following consent form: You are invited to participate in a research study about COVID-19. This is a 15-minute long survey that will ask about your perceptions, expectations and feelings about the disease, its effects on you and on our nation. If you agree to be part of the research study, you will be asked to provide your opinions on policies, risks, and will be asked to answer questions related to your current situation. Please pay attention to all questions. We will include several attention checks.
Benefits of the research to the public stem from your participation and honest answers.
Using this survey data, we hope to be able to provide guidelines for assessing and responding to differences across communities. Risks and discomforts: Thinking about COVID-19 and its impact may induce negative emotions, like anxiety or fear. These risks and discomforts are minimal for most people.
Participating in this study is completely voluntary. Even if you decide to participate now, you may change your mind and stop at any time. You may choose not to continue with the survey at any time and for any reason.
There is no deception or false information in this survey.
We will protect the confidentiality of your research records by not publishing any information that may identify you. Information collected in this project may be shared with other researchers, and may be connected to other aggregate datasets at the county level.
We will not share any information that could identify you. All results will be reported in aggregate.

SD.2 Survey Questions
In what follows, survey questions are in normal font while our notes are in italic.

Risk Tolerance
• Thinking about yourself, in general, how willing or unwilling are you to take risks?
Please use the scale below, ranging from 0 to 10, where 0 means "completely unwilling to take risks" and a 10 means you are "very willing to take risks." You can also use any number between 0 and 10 to indicate where you fall on the scale.

Comfort with Economic Activities
• (asked in June, August, and February) At this time, which of the following activities do you feel comfortable engaging in? [Yes/No options were available for the following list of activities: Eat at a restaurant (outside seating); Eat at a restaurant (inside seating); Go into a coffee shop; Go into a bar/pub; Use public restrooms; Go grocery shopping; Go shopping for non-food items (at the mall, hardware store, etc.); Go to the beach; Go to the gym or other sports facility; Be part of a gathering with more than 10 people (church, school, meetings, work, etc.).] • How worried are you feeling for the economic well-being of the following people? [Same options and groups as above.]

Beliefs About Infection Risk
As of April 20, 2020, CDC (Centers for Disease Control and Prevention) is reporting 776,093 confirmed coronavirus cases and 41,758 deaths in the U.S. Many cases go undetected. 16 Of course, infection rates depend on the community and the protection measures each person can take. Assuming that the social distancing policies and your personal efforts stay the same, what are the chances that you will get infected with the coronavirus in the next three months? [Choose one: 0% chance; 1-10% chance; 11-20% chance; 21-30% chance; 31-40% chance; 41-50% chance; 51-60% chance; 61-70% chance; 71-80% chance; 81-90% chance; 91-100% chance.]

Beliefs About the Effectiveness of State Restrictions and Own Precautions in
Reducing Infection Risk (asked only in April ) The next two questions embedded responses from previous questions. If the respondent indicated that their state has not introduced any social distancing measures, or that they have not taken any precautions, the relevant question was not displayed to the respondent.
• Earlier, the survey asked about the changes you personally made to protect yourself from the coronavirus infection. You indicated that you [all precautions that the respondent indicated as having taken]. Assuming that the state policies stay the same, what do you think your chances of becoming infected with coronavirus in the next three months would be if you did not make these changes? [Choose one: Same chance; 5% higher chance; 10% higher chance; 15% higher chance; 20% higher chance; 25% higher chance; 30% higher chance; 40% higher chance; 50% higher chance; My chance of being infected would increase by more than 50%.] • Imagine that you could still take the same measures you personally took to lower your chances of being infected with coronavirus, but your state had not introduced any social distancing measures. Assuming your personal efforts stay the same, what do you think your chances of becoming infected with coronavirus in the next three months would be if your state did not have social distancing measures? [Choose one: Same chance; 5% higher chance; 10% higher chance; 15% higher chance; 20% higher chance; 25% higher chance; 30% higher chance; 40% higher chance; 50% higher chance; My chance of being infected would increase by more than 50%.]

Beliefs About Health Outcomes Conditional on Being Infected
• According to the CDC (Centers for Disease Control and Prevention) report, about 7% of people diagnosed with the coronavirus are hospitalized, but do not need intensive care. About 1.5% of people are hospitalized and need intensive care. It is also suspected that a large percentage of people are symptom-free and/or have mild versions of the disease.
Most importantly, the chances are person-specific. The progression of the disease can be very different based on your age, health, pre-existing condition, living conditions, how much of the virus you are exposed to, etc. Although it's hard to know without data, you probably have a better understanding of your situation than anyone else. Therefore, we ask you to predict how the coronavirus is likely to affect you, should you get infected: Please make sure numbers add up to 100. Allocate points according to how big you think your chances are for each possibility. [Chances that I will be symptom-free are: (fill in, numerical); Chances that I will have a mild version of the disease are: (fill in, questions below were asked if the applicable precaution was not taken by the respondent. • You indicated that working from home was not one of the changes you made. Which of the following best describes why? [Choose one: I am retired/a student/currently not working; I could perhaps work from home, but it's better not to/I don't like to; I was not given the option to work from home (essential worker or employer needed me); I cannot work from home; Other: (fill in).] • You indicated that wearing gloves when you go shopping was not one of the changes you made. Which of the following best describes why? [Choose one: It's not necessary; I would like to, but cannot find gloves; I don't shop for anything, including groceries; Other: (fill in).] • You indicated that wearing masks was not one of the changes you made. Which of the following best describes why? [Choose one: Masks are not useful; I do not like wearing masks; I would like to, but cannot find or afford masks; I do not leave my house even to take a walk; Other: (fill in).] • You indicated that avoiding all public places and self-isolating was not one of the changes you made. Which of the following best describes why? [Choose one: It's too much. We need to keep functioning; I would like to, but I have to work outside the home; I would like to, but I have to leave my house regularly for doctor visits; I would like to, but I have to go get food and groceries -I cannot afford to have it delivered; I would like to, but I have to go get food and groceries -I don't want to have it delivered; I would like to, but I have to go get food and groceries -there are no deliveries available; Other: (fill in).] • You indicated that canceling travel plans was not one of the changes you made. Which of the following best describes why? [Choose one: I did not have any travel plans to begin with; I really wanted to go on the trip/did not feel like canceling; I could not get a refund and did not want to waste the money; I had to travel because of work obligations; I traveled because of family obligations; other: (fill in).]