Positive and Negative Effects of Republicanism on US citizens’ Mobility During the First Period of the COVID Crisis

While Republican states have been criticized for their limited efforts to contain the spread of COVID-19, it is important to consider that political orientation can modify human behaviour via complex effects that are still poorly understood. During the first period of the pandemic, we found that the association of republicanism with US citizens' mobility varied depending on the nature of the exposure being considered. First, republicanism was associated with increased mobility when the stringency of anti-COVID measures increased. Second, republicanism was associated with decreased mobility when COVID-related deaths increased. Third, republicanism was associated with increased mobility over time, i.e. as time went by, citizens living in Republican states were more mobile than those in Democratic states. These findings raise caution on any over-interpretation of the impact of polarization in US politics on COVID-related behaviour. They prompt consideration of persuasive tools that emphasize risk perception to promote social distancing in Republican states, rather than relying heavily on stringent anti-COVID interventions. Our study provides an analysis of the association between republicanism and US citizens' mobility during the first period of the COVID-19 pandemic. Using mobility to transit stations as our dependent variable, we found that republicanism was associated with increased mobility when the stringency of anti-COVID measures increased, but decreased mobility when the number of COVID deaths increased. Over time, we observed a positive effect of republicanism on mobility: as time went by, citizens living in Republican states were more mobile than those in Democratic states. We retrieved relatively similar results when running sensitivity analysis on outcomes, modulo subtle differences that could be due to our use of a rather general measure for the strength of anti-COVID interventions.

parties, accurate or not, about the need to distance and the severity of the disease 20,21 , and more recently about vaccination risks and effectiveness.
Little is known about these moderating effects of political orientation on citizens' behaviour in response to the COVID crisis. Their investigation would however be of paramount importance to better understand the natural tendencies and fine-grained mechanisms underlying the influence of political affiliation on people's behaviour, and further, how to counteract them when they result in significant health and social burden. In the current study, we sought to determine how republicanism affected US citizens' behaviour during the first period of the pandemic, via its effect on citizens' compliance with anti-COVID measures, on their perception of COVID-related risks, as well as over time. We chose mobility to transit stations as our outcome variable since that measure would theoretically be closely linked to a higher spread of the virus 22 . As proposed by others, we used the local cumulative number of COVID deaths and cases as proxies for informational inputs about general perception of COVID risks during the pandemic 19 . Finally, we chose to circumscribe our analysis to the beginning of the pandemic (from Mid-February to end of May 2020) as we did not want the US presidential elections campaign to amplify the bias inherent to political polarization on citizens' behavioural response.

Data
Outcome variable (Mobility to transit stations) We used the Google COVID-19 Community Mobility Reports database to extract state-wise daily percentage change in mobility to transit stations 23 . This database has been used extensively to assess various interventions during the pandemic. It provides time series cross sectional daily data for six particular categories of mobility data: (1) transit stations; (2) retail and recreation places; (3) workplaces; (4) groceries and pharmacies; (5) residential places; (6) parks. Observations are based on cellular phone, laptop and tablet signals.
Specifically, the data shows how visits and time spent in the above places change compared to so-called "baseline days". For each day of the week, the baseline day is the median value from the 5-week period Jan 3 -Feb 6 2020 (i.e. before the pandemic hit the US, hence before restrictions on movement) for a specific location and a specific mobility category. For each day, an integer value gives the percent change in mobility (positive or negative) compared to that same day during that 5-week period (Jan 3 -Feb 6).
In the current study, we chose mobility to transit stations as our main outcome of interest and used other measures in our sensitivity analysis 24 . We chose to restrict our time series to the period from February the 15th 2020 (start date of the database) till May the 31th 2020. The latter is often taken as the end of the first wave of the pandemic. In total, our time series amounts to 107 days. Overall, our data comprised N = 51 states x 107 days = 5457 observations. There were no missing values.

Political orientation
Each state's political orientation was defined as each state trifecta as of 2020 (Democratic vs. divided vs. Republican) and retrieved from the ballotpedia website 25 . In the Supplementary Materials 1 , we provide a list of US states according to their political affiliation.

Stringency of anti-COVID measures
For each state and each day, a measure of the stringency of anti-COVID interventions was extracted from the Oxford COVID-19 Government Response Tracker (OxCGRT) 26 . In short, the OxCGRT provides a measure of stringency on a scale from 0 to 100, publicly available on GitHub 27 . This measure is based on a summary of seven indicators on policies regarding social isolation and confinement, including school, workplace, and public transport closures, public events cancellations, stay at home requirements, and restrictions on gatherings and internal movement. Data is collected from publicly available sources such as news articles and government press releases and briefings. Each indicator measures the stringency of each policy or intervention on an ordinal scale of severity or intensity (from no measures taken to simple recommendations and implementation of the policy). Note that OxCGRT measures for US states do not include federal policies that apply to the country as a whole (e.g., international travel bans).
We reasoned that the effect of anti-COVID measures would be expected on citizens' mobility on the day they were introduced, as in 28 , therefore we did not include any lags in the "Stringency" time series.

Cumulative number of COVID deaths
We used the daily cumulative number of COVID deaths in each state as a reflection of the amount of risks due to COVID perceived daily by citizens in each state 19 . Indeed, we reasoned that a great marginal increase in the number of COVID-related deaths would inform citizens that the current status of the pandemic may be dangerous. And in turn, a large (negative) effect of the number of COVID deaths on mobility would suggest that citizens take great precautions against the virus by not going out. For each state and each day, the cumulative number of confirmed COVID-related deaths was extracted from the Oxford COVID-19 Government Response Tracker, and lagged by one day 27 . Note that for the purpose of our research question, we indeed reasoned that an optimal way to model the impact of COVID deaths on citizens' mobility would be to use the raw number of deaths rather than rates. This is because our hypothesis relies on the impact of COVID deaths on mobility via people's perceptions of the danger to go out due to an increased number of COVID deaths. In that sense, we hypothesized that a raw number would be more meaningful to people than a (relatively small) rate.

Temperature
We included a climate variable in our statistical model as we reasoned that climate might confound the relationship between COVID deaths and mobility. We retrieved the mean temperature (in degrees Fahrenheit) of each state's capital city for each day of our time series from the wunderground website 29  investigated. The variance inflation factor was below 7.5 for each regressor, and below 5 for our three regressors of interest. When we further scaled the temperature variable (linear and quadratic terms), the variance inflation factor was below 5 for each regressor, without any change in the mean coefficient value for each parameter of interest ( , and ). β 2 β 4 β 7 Cluster robust standard errors were estimated to account for the correlation of data within each state. Note that we did not include a main effect of republicanism, nor a single linear temporal trend in our statistical model as these would be collinear with states and days fixed effects, respectively.

Event-studies
We performed additional analysis where variables Stringency , Deaths and Trend were binearized. Specifically, Stringency , Deaths and Trend were discretized into 10 categorical bins, such that each bin contained enough data to perform additional regression analyses (see Supplementary Materials 3 for a table summarizing the number of observations contained in each bin for each political group). Doing so, the chances of data sparsity were reduced, which improves the interpretability of our results. Another reason why we performed these additional analyses is that they provide better insights on the effect modification of Stringency , Deaths and Trend by republicanism, because binearization allows better visualization of subtle non-linear effects.
We specified the following linear equation: We then reported so-called "event-study plots" to illustrate the effect of each binearized variable, modified by republicanism, on the outcome.

Sensitivity analysis
We ran the following sensitivity analysis to check that subtle changes in measurement would not invalidate our conclusions.
Outcome variable.
Here, we substituted mobility to transit stations to that of other places where signals were recorded in the Google COVID-19 Community Mobility Reports database: retail and recreation places, workplaces, groceries and pharmacies, residential places, and parks.
We hypothesized that our estimates of interest would not substantially change as compared to those obtained with mobility to transit stations when outcome variables were mobility to retail and recreation places, workplaces, and groceries and pharmacies. For residential places, we hypothesized that estimates would be of an opposite sign than those obtained with transit stations. Finally, we hypothesized that estimates would be non-significant for mobility to parks, as we reasoned that these would be much less impacted by subtle exposure changes.

Measurement of republicanism
Here, we changed our strategy for measurement of republicanism. Using data from the Pew Research Center, we differentiated states in terms of their percentage of voters for the Republican candidate at the 2016 presidential elections (former President Trump), rather than a rough differentiation of their political group as we did in the main analysis (Democratic vs. divided vs. Republican states) 35 .
Measurement of risk perception.
We re-ran our main analysis using cumulative COVID cases instead of cumulative COVID deaths as a measure of perception of COVID risk.

Accounts of temporality
We first tested a model taking into account a specific correlation structure of the error term, as sequences of non-seasonal Autoregressive Moving Average (ARMA) of daily values for each state, with p=1 lag.
Second, to test whether our results were robust to a longer period of analysis, we reproduced our event-study plots (aming to illustrate the effect of each variable of interest, modified by republicanism) over a longer period of analysis (till the end of February 2021 rather than end of May 2020).
Analyses were performed R version 4.0 and packages daggity , fixest and nlme .

Results
Our multivariate time series analysis aimed to linearly fit, among US states, daily mobility to transit stations (obtained from the publicly available Google COVID-19 Community Mobility Reports database) to anti-COVID stringency interventions (the so-called "Stringency" term in our statistical model), number of COVID deaths ("Deaths" term), as well as their interaction with an index of republicanism ("Republicanism" term). Republicanism was based on each state's government trifecta, and was defined as a linear increase from Democratic trifectas to divided governments and further, to Republican trifectas. We were also interested in the effect of republicanism on mobility over time ("Trend" term). Note that our model also included an index of climate (daily temperature in each state -raw and squared terms), as well as states and days fixed effects. The latter allowed to remove the variance related to state-and time-wise factors that could have had an impact on mobility to transit stations and that could have confounded the effect of party affiliation (e.g. unemployment rate, ethnicity, population density, national holidays etc...). In terms of our sensitivity analysis on outcomes, we retrieved relatively similar coefficients for Stringency x Republicanism, Deaths x Republicanism and Trend x Republicanism when using mobility to retail and recreation places, workplaces, or groceries and pharmacies as our dependent variable. Compared to mobility to transit stations, specifying models with mobility to residential places as our outcome demonstrated effects of an opposite sign. Finally, mobility to parks did not demonstrate any significant effect of Stringency x Republicanism, nor Deaths x Republicanism.

Discussion
Our study provides an analysis of the association between republicanism and US citizens' mobility during the first period of the COVID-19 pandemic. Using mobility to transit stations as our dependent variable, we found that republicanism was associated with increased mobility when the stringency of anti-COVID measures increased, but decreased mobility when the number of COVID deaths increased. Over time, we observed a positive effect of republicanism on mobility: as time went by, citizens living in Republican states were more mobile than those in Democratic states. We retrieved relatively similar results when running sensitivity analysis on outcomes, modulo subtle differences that could be due to our use of a rather general measure for the strength of anti-COVID interventions.
Our study adds to the growing body of literature showing that political polarization impacts health-related behaviours, from enrollment in health care insurance to perceptions about the safety of vaccines 7, 8,12,17,[36][37][38][39][40][41] . Yet, our design also allowed us to improve our understanding of the effect of political orientation on citizens' response to the pandemic at a fine-grained level. We first showed that, in terms of mobility to transit stations, republicanism was associated with a lower compliance with anti-COVID interventions. This confirms previous findings that alignment with political ideas could bias the formation of opinions and behaviours, even when these ideas are not substantive 39 , and that party endorsement mitigates adherence to evidence-based policy directives 42 . More specifically, psychological underpinnings typically related to republicanism might be inherently unaligned with highly constraining health-related policies, even though they could benefit the whole community 43 .
For instance, while being usually referred to as people of honor and duty, that respect authority, religion and traditions 44,45 , Republican citizens demonstrate significant antisocial traits [46][47][48] and, when reflecting upon public health-related matters, show a low level of trust in their governments and healthcare providers 19,30,31,49,50 . Likewise, from the perspective of cultural norms, Republican citizens demonstrate less communion traits than their Democratic counterparts, and more individual, agency traits 51 . Overall, these psychological 18 characteristics might have biased citizens from Republican states towards more reluctance to comply with constraining, "high-cost" anti-COVID measures taken by their local governments.
Second, in line with previous theories and research, we reasoned that, irrespective of government policies, learning the number of local COVID deaths would impact self-protective behaviours, because of a heightened perception of risk to catch the virus, and/or of the disease severity 15 . Confirming these predictions, we found an association between a greater number of COVID deaths and a decreased likelihood of going out to transit stations. Yet, we also demonstrated variations in such risk assessments across political groups, with individuals from Republican states being overall more cautious vis-a-vis the virus. Given previous research, we hypothesize that this over-cautiousness associated with republicanism might relate to: (1) a different age structure, i.e. a higher rate of older aged people who would be more risk-averse towards the virus 52 ; (2) a specific cognitive profile, with less extraversion and risk-taking behaviours associated with republicanism 31,32 ; and/or (3) a heightened response to fearful stimuli in Republican citizens [53][54][55] . Further research should explore how each of these factors may contribute to the perception of COVID-related risks in citizens that are differently affiliated to political parties.
Our third finding pertains to the significant difference in the temporal variation of mobility to transit stations across political groups, with an increased temporal trend in mobility associated with republicanism 8,41 . One potential explanation for such a temporal pattern is that the economic burden that mobility restrictions increasingly impose on individuals would be differentially appraised, and palliated for, in Republican vs. Democratic states. As an example, Republican citizens and governors demonstrated more reluctance to get federal unemployment aid 56 , which, in turn, could have resulted in increased mobility over time related to job hunting. A second hypothesis is that protective, yet constraining behaviours that relate to low levels of mobility may be differently integrated to social norms 57 in Democratic vs. Republican states. Indeed, some have shown that Democratic citizens were in general more motivated to socially distance 58 , while others, more likely to be Republicans, would demonstrate a heightened degree of so-called "pandemic fatigue" [59][60][61] .
Thirdly, these temporal variations may have resulted from specific events, a.k.a. discontinuities in the time series. Typically, political speeches or interventions on social media, such as statements from President Trump that "America wasn't built to be shut down", or that "We cannot let the cure be worse than the problem itself", may have influenced citizens from Republican states more than those from Democratic states, which in turn would have resulted in the former being less inclined to socially distance than the latter 20,21,62 .
Our study has some limitations. First, we used a cross-sectional time series database and in this context, causal inferences are difficult to make compared to cohort datasets. because there is a high degree of uncertainty in the lags j and k, which in turn would typically result in some form of simultaneity bias.
Fifth, in our event-studies, while we were careful that each stratum of covariates contained a substantial number of data for each of the three political categories (to avoid data sparsity), one could still argue that our analysis would be biased by the fact that Democratic states presented with a higher strength of anti-COVID interventions and a greater number of COVID deaths than Republican states. However, we found a significant effect of republicanism at low stringency levels and at low levels of COVID deaths, which suggests that our reported estimates are likely not fallacious.
In short, our parameter estimates of interest should be interpreted as representing estimates of adjusted associations rather than causal estimates per se. With these precautions in mind, we found republicanism to be associated with increased mobility when the stringency of anti-COVID measures increased, as well as over time. In contrast, republicanism was associated with decreased mobility when COVID-related deaths increased. These results raise caution on any simplistic over-interpretation of the impact of polarization in US politics on citizens' behaviour, and more particularly towards over-blaming and stigmatizing Republican states. They also suggest that specific measures could be implemented in Republican states to ensure the lowest possible rates of infection. Relying too much on constraining, "high-costs" measures that are tough to be scrupulously respected, such as business, university and school closures, or stay-at-home orders, might not represent an optimal solution in Republican states. Rather, citizens living in Republican states might be more easily persuaded to socially distance by emphasizing COVID risks based on COVID deaths or cases. Tools such as those provided by behavioural economics, e.g. using scare tactics, might better counteract the negative effects of republicanism on adherence to anti-COVID interventions, than increasing their stringency 63 .