Political polarization and cooperation during a pandemic

Abstract

In this paper, we examine the relationship between political polarization and individuals' willingness to contribute to the public good by engaging in preventative behaviors against COVID‐19. Using a sample of individuals from close‐election states, we first show that individuals engage in fewer preventative behaviors when the governor of their state is from the opposite party. We also show that this effect is concentrated among moderate individuals who live in polarized states, and that it is strongest when the state has been relatively forceful in combating COVID‐19. We estimate that the opposite‐party effect increased COVID‐19 cases by around 1%.

1. INTRODUCTION

Social scientists have long recognized that trust and social cohesion may be important for economic performance (e.g., Berg et al., 1995; Fukuyama, 1996; La Porta et al., 1997; Putnam, 1993). In a world where externalities are rife, transactions take time, and information is imperfect, market mechanisms alone will not guarantee efficiency. Trust in others' competence and decency, and concern for their well‐being, can help economic actors to coordinate their behavior and achieve better outcomes.

It is of some concern, then, that Americans feel a decreasing level of trust and increasing hostility along party lines. Gentzkow (2016) reports that, as of 2008, nearly half of Americans classified members of the other party as “selfish”, up from around 20% in 1960. The same report shows that 20%–30% of Americans would be upset if their son or daughter married a member of the other party, up from around 5% in 1960. The difference in individual's warmth toward their own party and their warmth toward the other party, each on a scale of 0–100, has increased from around 25 points in 1980 to 45 points today (Boxell et al., 2020).

This hostility may be particularly problematic when members of different parties have to work together to combat a crisis, as has recently occurred during the COVID‐19 pandemic. Social distancing poses a classic collective action problem: while the costs of preventative behavior accrue entirely to the individual, the benefits are diffused across a large number of people. It is in precisely this situation that social preferences might help to restore more efficient outcomes. If individuals are altruistic, they internalize some of the utility costs their actions have on other people. When altruism breaks down, society's ability to overcome the collective action problem may be compromised. While this argument has been made in the context of ethnic fractionalization (e.g., Poterba, 1997; Vigdor, 2004), until recently there has been little empirical work examining the consequences of political fragmentation for public goods provision. ¹ In this paper, we fill this gap by examining whether political polarization has undermined the response to the COVID‐19 pandemic in the United States.

A simple comparison of social distancing across more or less polarized individuals or states will not identify the impact of polarization, because both polarization and willingness to comply with social distancing may be related to other, unobserved variables (such as an individual's general agreeableness, or a state's demographic composition). We instead attempt to infer the impact of political polarization indirectly, by using quasi‐randomization in a feature of an individual's environment: the party of their state governor. Using both survey and cell‐phone location data in the set of states with close gubernatorial elections, we show that individuals comply significantly less with social distancing measures when the other party has narrowly won the election. This effect is bigger when the state government has mounted a stronger policy response against the pandemic, and seems to be driven by opponents of the governor reducing their compliance behavior (rather than supporters of the governor increasing their behavior). Reduced trust in the state government among opponents explains some, but not all, of our results.

Clearly, this response must be driven by political fragmentation to some degree. An opposite party governor could not have this type of effect if political parties were not salient to individuals. Given that political parties are a feature of most democracies, however, is this type of response inevitable? Or is the opposite‐party response exacerbated by the uniquely high level of political polarization in the United States today?

While we cannot answer these questions definitively without a source of random variation in polarization, we show that the patterns in the data are most consistent with the opposite‐party effect being exacerbated by polarized environments ‐ but only for individuals who are themselves more moderate. The most polarized individuals show no opposite‐party response, in either polarized or non‐polarized states. Moderate individuals, however, show a limited opposite‐party response in unpolarized states and a very strong opposite‐party response in polarized states. We believe that this is due to the fact that less polarized individuals have weaker prior beliefs over the correct course of action and are therefore more easily influenced by their environments. Overall, the effect of an opposite‐party governor is about twice as strong in polarized states.

While the relationship between the opposite‐party response and political polarization is correlational, and should therefore be interpreted with caution, it does suggest an answer to the question of whether an opposite party response is inevitable. The answer is no. This response does not seem to be an immutable feature of human nature, but one that arises under particular conditions ‐ conditions which are unusually prevalent in the United States at this moment in time.

2. RELATION TO THE LITERATURE

This paper relates most closely to a large body of evidence linking ethnic fractionalization to a lower quality of governance and reduced provision of public goods. This literature shows that ethnic diversity is correlated with worse policies and slower growth across countries (Alesina et al., 2003; Easterly & Levine, 1997; La Porta et al., 1999; Lassen, 2007) and across cities or regions within countries (Alesina et al., 1999; Alesina & Ferrara, 2000; Banerjee et al., 2005; Miguel & Gugerty, 2005; Vigdor, 2004). Algan et al. (2016) use random variation in assignment to public housing in France to show that more diverse housing areas have more vandalism and neglected upkeep. A first contribution of our analysis is to show that the relationship between social fragmentation and reduced willingness to contribute to public goods is more general, extending to political divisions as well as ethnic divisions.

A key limitation of the above‐cited field studies is that they are not able to identify the mechanism relating ethnic fractionalization to outcomes. It is unclear, for example, whether ethnic diversity matters because ethnic groups have different preferences over public goods (the “preference hypothesis”; e.g., Alesina et al., 1999), because different groups have a harder time coordinating to put pressure on the government (the “coordination hypothesis”; e.g., Hardin, 1995; Besley et al., 1993; Miguel & Gugerty, 2005), or because people simply are less willing to vote for public goods when some of the benefit accrues to individuals they don't like (the “empathy hypothesis”; e.g., Poterba, 1997; Vigdor, 2004). Habyarimana et al. (2007) use a laboratory experiment designed explicitly to distinguish between these three hypotheses, and find support for the idea that cooperation is less easily sustained between members of different ethnic groups. It is not known whether these results generalize to other settings, however.

We view our study as occupying a middle ground between the field research and the experimental work on social cohesion and public goods. As in the laboratory, we are able to directly observe individuals' decisions to contribute to a public good, as opposed to more complex outcomes such as economic growth or school funding. This, in combination with other features of our setting, allows us to be somewhat more confident about the mechanism behind our results. In particular, the number of groups in our analysis (two) is constant across states and is unaffected by which party wins a gubernatorial election. This suggests that our results are not driven by either the preference or coordination channels. Instead, our results strongly suggest that leadership by an out‐group member directly reduces people's willingness to contribute to public goods, which is consistent with the empathy hypothesis. In contrast to the laboratory experiments, however, we are also able to show that this animosity has important real‐world consequences.

Our paper also complements a growing literature on the partisan divide in social distancing behavior in the United States (e.g., Allcott et al., 2020; Barrios & Hochberg, 2020; Gadarian et al., 2020; Painter & Qiu, 2020). These papers show that Republicans engage in fewer preventative behaviors than Democrats, even controlling for other factors that might contribute to social distancing decisions. While this is true within our data as well, we also document a more subtle phenomenon: that both Democrats and Republicans reduce preventative behaviors when living in an opposite‐party state. While these papers document the effect of specific policy positions taken by the two major parties during the pandemic, we believe our results speak to a more general mechanism by which polarization affects willingness to contribute to public goods. In this respect, our paper most closely aligns with 3 contemporaneous papers. Painter and Qiu (2020) note that political alignment with officials may partially explain underlying differences in partisan compliance behavior. Because all of the measures of compliance they use are at the county‐level, however, the precise mechanism behind the behavior change is unclear. Druckman et al. (2020) and Allcott et al. (2020) both provide contemporaneous work using individual level survey data and similarly find that polarization is associated with reduced compliance with social distancing directives. We build on this work, showing that little of this response is due to diminished trust in institutions and government (except at the state level). Indeed, our results are concentrated among individuals who are less polarized than their peers, and argue that this may be driven by the environment of polarization at the state level.

3. DATA

3.1. MTurk sample

We ran a survey of Amazon's Mechanical Turk (MTurk) workers starting on March 24th of 2020. ² We recruited approximately 2000 workers from 24 states that had close election results (Republican share of the votes for the two major parties between 45% and 55%) in the last gubernatorial election. ³ We over‐sampled from the smaller states in order to ensure that we could calculate sample means with precision in each state. After dropping respondents who do not report voting consistently for either party, along with some respondents with missing or irregular data, our main analysis sample is made up of 1753 MTurk workers. ⁴

An excellent discussion of the use of MTurk in economics research is provided in Horton et al. (2011). They argue that experiments conducted over online settings (and specifically MTurk) can be as valid as those run in other experimental settings. In particular, they directly evaluate whether pro‐social behaviors observed in laboratory experiments can also be observed in MTurk samples. Using a one‐shot Prisoner's Dilemma game, Horton et al. (2011) show that there is no statistically significant difference between the likelihood of cooperation across the MTurk and physical lab samples. ⁵ MTurk surveys have also been widely used by economists to elicit policy preferences over tax schemes (Fisman et al., 2017) and redistribution (Kuziemko et al., 2015), and evaluating the performance of voting schemes (Casella & Sanchez, 2019).

As in all research using MTurk, it is important to note that our sample is not representative of the U.S. population. The sample is younger, more highly educated, more likely to be white and/or non‐Latinx, and is more heavily concentrated in the South than the American population at large. We would also expect that they are more internet‐literate than the rest of the population. This is a key reason why we also show that the opposing‐party result holds in cellphone based travel data from Unacast, which is more representative.

3.2. Unacast sample

Because our MTurk sample may not necessarily be representative of the wider U.S. population, we supplement our analysis with a more broad‐reaching Social Distancing Dataset, made available to us by Unacast (2020). This dataset leverages GPS locations from mobile phones to identify the changes in the average time spent in and around the home, and more importantly for our analysis, changes in the average distance traveled.

To proxy social distancing compliance, we use two measures provided by Unacast: the percent difference in daily total distance traveled, and the percent difference in daily visitation of non‐essential points of interest, ⁶ both averaged across all devices in the county. To estimate these distance metrics, Unacast first calculates the average visitation for each day of the week before the COVID‐19 outbreak, where the pre‐COVID‐19 period is defined as March 8th and prior. Our data span the time period from February 28th to May 4th, meaning that they tell us how the daily distance and visitation metrics changed, on average, between March 8th and May 4th, relative to Feb 28th‐March 8th. Each day of the week during the outbreak is then compared to a corresponding pre‐COVID‐19 days of the week to account for underlying travel patterns, and this difference captures the change to daily distance traveled.

A number of cell signal aggregation firms have made mobility data available to researchers in order to support COVID‐19 research. Gupta et al. (2020) investigate the comparability of these data, arguing this is crucial in understanding the generality of a result. They further provide an extensive review of papers using these various data sources to understand the effects of stay‐at‐home orders on human mobility. While the papers listed (and thus, the datasets used) are not directly comparable, the impact of stay‐at‐home orders on mobility when measured using Unacast data appear to be on the same order of magnitude as the 4 largest other sources of data (PlaceIQ, SafeGraph, Google Mobility, Apple Mobility). This is perhaps reassuring, given each of these data sets are in principle drawn from potentially different sets of users and mobile applications.

3.3. Election data

Data on gubernatorial election outcomes was gathered from the website Ballotopedia.com. Of the 24 states with close elections, 14 had narrowly elected Democratic governors: Connecticut, Kansas, Kentucky, Louisiana, Maine, Michigan, Montana, Nevada, North Carolina, Oregon, Virginia, Washington, West Virginia, and Wisconsin. ⁷ The remaining 10 states with narrowly elected Republican governors are: Florida, Georgia, Indiana, Iowa, Mississippi, Missouri, New Hampshire, Ohio, South Carolina, and South Dakota.

3.4. Political orientation and polarization

We ask MTurk survey respondents to indicate which party they usually vote for: Democrats, Republicans, or neither. We use this variable to classify respondents as Democrats or Republicans. We drop around 250 respondents who do not consistently vote for either party. ⁸

We also want to examine whether our results differ across respondents who are more or less polarized. We use a standard measure of affective polarization, which is the difference in the degree of “warmth” the respondents feel toward their own party and toward the other party (e.g., Boxell et al., 2020; Iyengar et al., 2019, 2012; Levy, 2021). Respondents were asked to rate their level of warmth for each party on a scale of 0–100, and we measure the difference in the ratings between the respondent's own party and the other party. Iyengar et al. (2012) were the first to argue that this measure was the most appropriate way to capture the extent to which political affiliation had become an important part of social identity (which is defined, in part, by positive affect toward one's in‐group members and negative affect toward out‐group members). Because social identity underlies the “empathy hypothesis” put forth in the earlier literature on social cohesion and public goods provision, this measure accurately captures the channel through which we expect polarization to affect cooperation.

We drop approximately 40 respondents who report higher warmth toward the other party. We create an indicator for “polarized individual” if this difference exceeds 40, which is the national median from the American National Election Survey in 2016. ⁹

Because the COVID‐19 epidemic may have affected political polarization, we also attempted to recontact a set of respondents who were included in a previous survey on political polarization in the fall of 2019. We were successful in re‐surveying approximately 150 of these respondents. The results are qualitatively similar, but numerically stronger, using pre‐COVID‐19 measures of polarization in this sample. The details of this analysis are included in Appendix Table A5.

3.5. State polarization

The willingness to contribute to a public good may depend not only on a respondent's polarization, but also on their perceptions of the polarization of their community. We therefore also examine whether the response to an opposite party governor differs by state polarization prior to the COVID‐19 outbreak. To capture state polarization, we use measures taken from the American National Election Study in 2016. The ANES contains a variable that is identical to our measures of warmth toward the two parties. We use this to construct the fraction of people in each state who are above the national median (40) in the difference between warmth to their own and the other party. We create an indicator for “polarized state” for states that are in the top half of this variable within our sample.

3.6. Outcomes

Our key outcomes are measures related to compliance with recommended measures to combat the spread of COVID‐19. We first asked whether the respondent had left their house in the previous 48 h, and whether this was for reasons classified as essential (work, groceries, pharmacy, medical care) or non‐essential (everything else). We use this to create our first outcome variable, “left home ‐ non‐essential”, which is an indicator for visiting any location other than work, the grocery store or pharmacy, the doctor's office, or the hospital.

Secondly, we ask individuals whether they are continuing to work outside the home, and if so, whether this was by choice. We use this question to construct an indicator for “working outside the home” for the population that had worked outside the home before the pandemic, as well as an indicator for “working outside the home ‐ by choice” for the population that indicated having a choice in the matter. Note that the population that indicates having a choice in the matter is relatively small, around 200 people, which is why we present results for both variables. In both the “work outside the home” and “work outside the home ‐ by choice” measures, we exclude individuals who did not work outside the home prior to the pandemic.

Finally, we ask individuals directly about the measures they had taken in the past week to limit the spread of COVID‐19. We presented a list of six measures and asked them to indicate any that applied. The six measures were: washing hands or using hand sanitizer more frequently, staying home more often, canceling planned travel, limiting contact with elderly or more high‐risk friends and family, wearing gloves or a mask while outside of the house, and “other”. We create an indicator variable for each answer, as well as a variable indicating the number of responses the individual reported.

Table 1 shows the means of our outcome variables for Democrats and Republicans living in states with a Democratic/Republican governor. Approximately 30% of the sample reported having left their home in the previous 48 h, which was similar across different groups. Around 40% of the sample who worked outside the home before the crisis was continuing to do so as at the time of the survey. For Democrats, this number was larger in Republican states, while the reverse was true for Republicans. The raw difference‐in‐difference estimate is large, at around 11 percentage points, although it is not statistically significant. The estimate is similar among people who report having a choice in the matter, although we should note that this is a relatively small sample of individuals (approximately 200).

TABLE 1.

Preventative health measures, by political party and governor's political party (close election sample)

	MTurk data
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Respondent:	Dem.	Dem.		Rep.	Rep.		Diff. in.
Governor:	Dem.	Rep.	Diff.	Dem.	Rep.	Diff.	Diff.
Left house ‐ non‐essential	30.0%	28.6%	−1.4%	33.3%	31.6%	−1.7%	0.3%
Work outside of home a	37.7%	40.2%	2.5%	46.3%	37.3%	−9.0%	11.5%
Work outside of home by choice a	28.0%	12.3%	−15.7**%	47.0%	20.9%	−26.1%***	10.4%
Wash hands	95.3%	92.7%	−2.6%**	86.1%	91.1%	4.9%**	−7.6%***
Stay home	94.0%	92.3%	−1.7%	83.6%	88.8%	5.2%**	−6.9%**
Cancel planned travel	42.7%	39.5%	−3.1%	35.0%	43.1%	8.2%**	−11.3%**
Limit contact w. high‐risk	76.3%	74.3%	−2.0%	69.4%	65.8%	−3.6%	1.5%
Wear PPE	21.3%	24.6%	3.2%	22.4%	27.2%	4.8%	−1.5%
Other	3.4%	3.1%	−0.4%	3.3%	1.9%	−1.4%	1.0%
Number of preventative behaviors	3.3	3.3	−0.1	3.0	3.2	0.2**	−0.3***
N	553	521	1074	366	313	679	1753

	Unacast data
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
County:	Dem.	Dem.		Rep.	Rep.		Diff. in.
Governor:	Dem.	Rep.	Diff.	Dem.	Rep.	Diff.	Diff.
Change: daily distance	−24.8%	−21.6%	3.2%	−20.2%	−19.2%	1.0%	2.2%
N	234	142	376	638	661	1299	1675
Change: non‐essential visits	−50.3%	−49.1%	1.3%	−40.5%	−40.0%	0.5%	0.8%
N	177	107	284	395	453	848	1132

Note: This table shows summary statistics for preventative behavior during the COVID‐19 pandemic, by the political party of the respondent (in the MTurk data) or of the county (in the Unacast data), and by the political party of the state's governor. An observation in the MTurk data is an individual respondent, all of whom come from “close election” states (where the Republican vote share in the previous gubernatorial election was between 45% and 55%). An observation in the Unacast data is a county, and the sample is the set of counties that are in close‐election states. A county is classified as “Democrat” if more than 50% of voters voted Democrat in the previous gubernatorial election. The statistical significance of differences in columns (3) and (6) is based on regressions with standard errors clustered at the state level, while the statistical significance of the differences in column (7) are based on regressions with standard errors clustered at the state by party level.

^a The sample for the work outside of the house measure is restricted to the respondents who indicated that they worked outside the home before the pandemic. The sample size for this variable is 932. The sample for the work outside by choice measure is restricted to respondents who worked outside the home prior to the pandemic, and who indicated that they had a choice in working from home or working outside of the home. The sample size for this variable is 201.

For the preventative measures, the results are much clearer. For three of the six preventative measures ‐ washing hands, staying home more frequently, and canceling travel ‐ Democrats show reduced compliance in Republican states, while Republicans show greater compliance in Republican states. The differences are statistically significant, and range from around 8% of the mean for washing hands or staying home to 25% of the mean for canceling travel. There are no significant differences in the other three variables. On net, the raw difference in the number of behaviors for an individual in an opposing‐party state is around −0.3 on a base of around 3. In our regression results, we explore whether these differences are robust to the inclusion of controls and fixed effects, and how they vary with individual and state polarization.

Table 1 also shows the mean of the Unacast travel variables, and how these differ across Democratic/Republican counties in Democratic/Republican states. Both Democratic and Republican counties show less of a reduction in daily distance traveled when the governor is Republican, although this difference is larger for Democrats. The difference‐in‐difference estimate is around 2.2 percentage points (compared to an average of around 25 percentage points), although it is not statistically significant. The non‐essential visitation variable shows similar patterns, although the effects are quantitatively much smaller.

4. EMPIRICAL STRATEGY

We begin by estimating the effect of an opposing‐party governor on the likelihood of engaging in social distancing measures and other preventative behavior. We estimate the following regression using the MTurk data:

$$Y_{\mathit{ips}}=\alpha +{\beta }_{1}Opposin{g}_{\mathit{ips}}+{\beta }_{2}Democra{t}_i+{\beta }_3{\mathbf{X}}_i+{\gamma }_s+{ϵ}_{\mathit{ips}}$$ (1)

where Y _ips is an outcome variable. In most specifications, this is a dummy variable equal to one if individual i, with political party/ideology p, in state s, engaged in each of the outcome measures. In other specifications, we add up the count of behaviors the individual reports. Opposing _ips takes the value of one if the respondent's political party p is different from the party of the state governor, and 0 otherwise. Democrat _i takes the value one if individual i aligns themselves with the Democratic party and 0 otherwise. X _i is a vector of demographic characteristics including age, age squared, an indicator for four education levels (high school or less, some college, bachelor's degree, post‐graduate degree), 4 racial categories (white, black, Asian, and other), indicator for Latinx and an indicator for female. ¹⁰ Because our key dependent variable varies with state and party, we cluster standard errors at the state by party level. ¹¹

For this regression to capture a causal effect of having an opposing party governor, it must be that there is no other reason for individuals who are in the political minority in their state to show reduced compliance with social distancing. This might not be the case if, for example, political minorities are unusually contrarian, or are more likely to have moved from out of state (and therefore have fewer social ties in the state). If either of these stories was true, we might see reduced compliance for political minorities, but this would not be the result of having an other‐party governor.

This concern is the reason we focus on the set of states with close elections. The strategy of comparing close elections was first made by Lee (2008), who used a regression discontinuity design in Congress elections to estimate the causal effect of incumbency on future election. Because we are examining outcomes at the state level (rather than at the congressional seat level), we have a much smaller number of observations in our sample than a typical close‐election comparison; for this reason, we do not use a formal regression discontinuity design in our main results. However, the Appendix Table A3 shows that results using an RD methodology remain statistically significant and are larger in magnitude than the OLS results we present here.

The underlying assumption in our regressions is that individuals who live in states where the Republican party won by a narrow margin are otherwise similar to members of the same party who live in states where the Democratic party won by a narrow margin. In support of this, we show in Table A2 that there are no significant differences in demographics across individuals within our sample who come from opposing‐party versus own‐party states. Eggers et al. (2015) provides additional evidence that this strategy is valid for estimating treatment effects in most contexts.

It is important to note that close elections, by our definition, are not the same as surprise wins. For example, South Carolina is in our sample with an 8 percentage point Republican win; prior to the election, the website fivethirtyeight.com estimated that the Republicans had a 97% change of winning. Similarly, Oregon has a Democratic governor that won with a 3.4 percentage point gap, but fivethirtyeight.com put the odds of a Democratic win at 82% (fivethirtyeight.com, 2018). The logic behind restricting the sample to close elections is to ensure that when we compare the behavior of political minorities to political majorities, we are comparing populations that are relatively similar in underlying characteristics. This is true even if we know the outcome of a race ahead of time. ¹²

Finally, a potential concern with the MTurk data is that, as noted previously, the MTurk sample is not representative of the general population. Additionally, the measures we examine are self‐reported and potentially subjective. Perhaps MTurk respondents who live in opposing party states are simply less motivated to report that they undertake preventative measures, even if they are actually behaving the same way as everyone else. To allay this concern, we also present results using the Unacast social distancing scoreboard at the county level. Our regression equations in this dataset are:

$$Y_{\mathit{cps}}=\alpha +{\beta }_{1}Opposin{g}_{\mathit{cps}}+{\beta }_{2}Democra{t}_c+{\beta }_3{\mathbf{X}}_c+{\gamma }_s+{ϵ}_{\mathit{cps}}$$ (2)

and

$$Y_{cs}=\alpha +{\beta }_{1}OpposingShar{e}_{cs}+{\beta }_{2}Democra{t}_c+{\beta }_3{\mathbf{X}}_c+{\gamma }_s+{ϵ}_{\mathit{cps}}$$ (3)

The first equation is exactly analogous to Equation (1). We regress a social distancing measure for county c, with a dominant political affiliation p, in state s, on an indicator for whether the state's governor is of the opposite political affiliation to the majority of the county. We assign counties to be Democrat or Republican based on which party dominated during the previous gubernatorial election. County‐level controls are taken from MIT's Election Data and Science Lab's 2018 dataset, which contains information on the population and age, educational, race, and gender distribution of each county in 2016. We weight these regressions by county population and cluster standard errors at the state by party level.

In the second version of this regression, we use the fact that, based on the individual relationship we document in our MTurk sample, counties with more political opponents should see bigger reductions in compliance. Rather than classifying counties as Democrat or Republican, we simply measure the fraction of each county that voted for the other party in the previous gubernatorial election. Because our key independent variable now varies within state by party cells, we no longer cluster the standard errors at this level. All other details of these regressions are the same as in Equation (2).

5. RESULTS

5.1. Results on preventative behaviors

The results in Panel one of Table 2 show the coefficients from a regression of preventative measures in our MTurk sample on an opposing party indicator, using regression Equation (1). The results are quite similar to those shown by the summary statistics in Table 1. Individuals in opposing‐party states are 5–9 percentage points more likely to work outside the home (by choice or otherwise), although this difference is only marginally statistically significant. They are significantly less likely to report washing their hands or staying home more often than usual and to have canceled planned travel. On net, they report about 0.1 fewer behaviors than their counterparts living in same‐party states. ¹³ This is approximately 10% of a standard deviation in the independent variable.

TABLE 2.

Regression results, MTurk and Unacast Samples

	Panel 1: Mturk data
	Dependent variable
	(1)	(2)	(3)	(4)	(5)
	Left home, non‐ess.	Work outside home	Work outside home, by choice	Wash hands	Stay home
Opposing	0.003	0.055*	0.085*	−0.036***	−0.036***
	(0.015)	(0.029)	(0.042)	(0.008)	(0.011)
N	1753	932	201	1753	1753
R2	0.051	0.089	0.247	0.051	0.061

	Dependent variable
	(1)	(2)	(3)	(4)	(5)
	Cancel travel	Limit contact	Wear PPE	Other	Number of behaviors
Opposing	−0.049***	0.008	−0.008	0.005	−0.115***
	(0.016)	(0.021)	(0.018)	(0.006)	(0.043)
N	1753	1753	1753	1753	1753
R2	0.029	0.066	0.051	0.066	0.050

	Panel 2: Unacast data
	Dependent variable
	(1)	(2)	(3)	(4)
	% Change, daily distance	% Change, daily distance	% Change, non‐ess. visits	% Change, non‐ess. visits
Opposing	0.836*		0.528
	(0.468)		(0.582)
Opposing share		3.662***		2.606*
		(1.076)		(1.555)
N	1675	1675	1132	1132
R2	0.678	0.679	0.639	0.640

Note: The first panel of this table shows the results of OLS regressions of the indicated preventative measures on an indicator that the opposing party holds the governorship in a respondent's state, using the MTurk data. Respondents are classified as Democrat or Republican based on self‐reported responses. These regressions also include a control for being a Democrat, state fixed effects (which absorb the effect of having a Republican governor, as well as the Republican vote share), and the following demographic controls: age, age squared, indicators for four education levels, indicators for 3 racial categories, an indicator for Latinx, and an indicator for female. Standard errors are clustered at the state by party level. All results are robust to Bonferroni and Sidak corrections, except for those on working from home. The second panel of this table shows the results of OLS regressions of the indicated measures on measures of county‐level opposition to the governor's party, using the Unacast county‐level data. In columns (1) and (3), the key independent variable is an indicator that the majority of a county's voters voted for the opposing party to the governor in the last gubernatorial election. In columns (2) and (4), the key independent variable is the share of the county's voters who voted for the opposing party, which ranges from 0 to 1. These regressions include a control for being a Democratic county (Democrat vote share >50%), state fixed effects (which absorb the effect of being in a Republican state, as well as the overall Republican vote share), county population, the share of the population that is female, white, black, Hispanic, has less than a high school education, has less than a college education, is under age 29, or is over age 65. Regressions are weighted by the county population, and standard errors are clustered at the state by party level in columns (1) and (3).

***p < 0.01, **p < 0.05, *p < 0.1.

Panel two of Table 2 shows the results from regressions of Unacast social distancing measures at the county level on indicators for an “opposing” county, and the vote share for the opposing party to the governor, respectively. The first two columns have the percentage change in daily distance traveled as the dependent variable. The coefficient on “opposing” shows that counties where the majority of people voted for the non‐winning party reduced their daily distance traveled by about 0.8 percentage points less than counties where the winning party dominated, with the difference significant at the 10% level. In the second column, which uses regression specification 3, the coefficient on “opposing share” is around 3.6 and is significant at the 1% level. This coefficient implies a difference of around 1 percentage point between the 25th and 75th percentiles of opposing share in our data (0.35 and 0.61), which is similar to the estimate in column (1). This is approximately 10% of a standard deviation in the independent variable.

Columns (3) and (4) use an alternative measure, which is the percentage change in visits to non‐essential locations. Note that this measure is available for a smaller set of counties than the daily distance measure. The results are qualitatively similar to the daily distance regressions, although statistically insignificant for the opposing indicator, and marginally significant for the opposing share. Using the opposing share version of the regression, moving from the 25th to 75th percentile on opposing share increases the daily visitation change by around 0.6 percentage points.

Figures 1 and 2 show the results on the Unacast data visually, by plotting the outcome measures against the county Democrat share, separately for states with Democrat/Republican governors. Figure 1 shows that in states with Democratic governors, there is a negative relationship between the daily distance measure and a county's share of Democratic voters, indicating that Democratic counties changed their behavior at the onset of the pandemic. As shown in the righthand panel of Figure 1, however, this relationship is essentially absent in states with Republican governors. For majority‐Republican counties, social distancing behavior is higher in Republican states (where we see larger reductions in distance traveled), while the opposite is true for Democratic counties. For the visitation change variable, shown in Figure 2, more Democratic countries see a larger change in both Republican and Democratic‐led states; the relationship is somewhat stronger in states with a Democratic governor (although this is hard to detect visually).

FIGURE 1.

Variation in distancing behavior by county Democrat share. This figure shows how the change in daily distance traveled varies with the Democrat share in a county, separately for states with a Democratic or Republican governors. Both the daily distance variable and the Democratic share are residuals from a regression of these variables on state fixed effects, county population, the share of the population that is female, white, black, Hispanic, has less than a high school education, has less than a college education, is under age 29, or is over age 65. The size of the circles in the figure represents the size of the county population

FIGURE 2.

Variation in visitation behavior by county Democrat share. This figure shows how the change in daily visitation varies with the Democrat share in a county, separately for states with a Democratic or Republican governors. Both the daily distance variable and the Democratic share are residuals from a regression of these variables on state fixed effects, county population, the share of the population that is female, white, black, Hispanic, has less than a high school education, has less than a college education, is under age 29, or is over age 65. The size of the circles in the figure represents the size of the county population

How economically significant are the effect sizes that we document here? To answer this question, we perform a back‐of‐the‐envelope calculation relating our coefficients from the daily distance traveled regressions to growth in the number of COVID‐19 cases. We first relate county‐level growth in cases to social distancing behavior 2 weeks earlier, and then use this to estimate how many fewer cases there would be if every “opposing” county instead behaved like a non‐opposing county. ¹⁴ The details of this calculation are in the Appendix. We estimate that between March 30‐April 27th, there would have been approximately 2000 fewer cases in close‐election states in this case. The total number of cases in these states over this time period was around 250,000, implying that eliminating the opposite‐party effect would have reduced cases by around 1%.

The MTurk and Unacast data tell a strikingly similar story: supporters of the party that lost the most recent gubernatorial election are less likely to comply with social distancing and other preventative measures, with effect sizes of around 10% of standard deviation. As we discuss below, the welfare and policy implications of this finding depend on why this effect occurs (i.e., are winning parties inducing better behavior among their supporters, or are opponents reducing their behavior?) and under what conditions (i.e., are there things we could do to ensure better compliance among political opponents of the governor?). We turn to these questions in the following subsections.

5.2. Mechanisms

Why do people in opposite‐party states comply less with social distancing measures? We begin by examining whether opposite‐party status influences potential intermediate variables, such as the perceived importance of social distancing, trust in government and health organizations, or exposure to information skeptical of social distancing. Table 3 provides this analysis. In the first two columns, we examine whether opposite‐party governance affects individuals' perception of the seriousness of the COVID‐19 crisis. The dependent variable in the first column is a respondent's 0–10 rating of the importance of social distancing, while the second is an indicator for whether the respondent indicated being unworried about either getting or transmitting the virus. These coefficients are insignificant and close to zero.

TABLE 3.

Regression results, potential mechanisms

	Imp. of SD	Unworried	Trust in med. orgs.	Trust in state govt.	Est. compliance	Skeptical info
Opposing	−0.019	−0.019	−0.050	−0.722***	0.424	−0.010
	(0.018)	(0.013)	(0.080)	(0.123)	(0.747)	(0.023)
N	1753	1753	1753	1753	1753	1753
R 2	0.080	0.054	0.096	0.103	0.043	0.051

Note: This table shows the results of OLS regressions of the indicated measures on an indicator that the opposing party holds the governorship in a respondent's state, using the MTurk data. Respondents are classified as Democrat or Republican based on self‐reported responses. All controls are the same as those used in the other tables. The dependent variable “Trust in medical organizations” is an average of questions asking about trust in the WHO, the CDC, and the AMA. Standard errors are clustered at the state by party level. The results on trust in state government remain after Bonferroni and Sidak corrections.

***p < 0.01, **p < 0.05, *p < 0.1.

In the third and fourth columns, we examine whether individuals report less trust in either medical organizations or the state government. ¹⁵ There is no significant impact on trust in medical organizations (although the coefficient is negative), but a large impact on trust in the state government. This variable does seem to explain some, but not all, of our main result: including it as a control in our main regressions reduces the coefficients on working outside of the house and the number of behaviors by around 25% in both cases (see Appendix Table A4 for the results). ¹⁶

In the fifth column, we examine a respondent's estimation of other people's compliance in their community. This variable ranges from 0 to 100. There is a small and insignificant increase in estimated compliance in opposing‐party states, suggesting that reduced compliance of others does not explain our results. Finally, we examine whether people in opposing‐party states receive different information about the COVID‐19 epidemic. In the sixth column, the dependent variable is an indicator for whether the respondent has seen any information skeptical about the importance of social distancing. There is no effect of opposite party‐status on this variable. ¹⁷

In sum, respondents living in states where the other party narrowly won the election show little difference in attitudes toward the COVID‐19 crisis or social distancing, do not appear to trust medical organizations less, do not believe their neighbors are “cheating”, and do not get exposed to more skeptical information about the social distancing. They do, however, report lower levels of trust in their state governments, which seems to explain around 25% of our total effect size. ¹⁸ Most of our effect, however, appears to be unexplained by these intermediate variables. Why, then, do opposing‐party supporters cooperate less?

We believe that the most likely explanation for our results is that individuals simply respond less to requests to stay home when those requests come from an out‐group member. In support of this interpretation, Table 4 shows that our result is concentrated in states that had stronger policy responses to the pandemic. In these regressions, we use the summary variable “number of behaviors” as our dependent variable. The first column of Table 4 shows that the response to an opposite party governor is stronger in states that declared emergencies earlier; while the interaction of “opposing” and date of emergency is not significant, the response at zero (the state with the earliest date of emergency declaration) is roughly twice as large as our main estimate. The next five columns interact “opposing” with indicators for school closures, bans on large gatherings, bar/restaurant closures, non‐essential business closures, and mandatory quarantine measures. The coefficients on the interaction terms are somewhat mixed, with negative and significant interactions on school closures (although note that this is identified by only 3 states) and quarantine measures, but mixed and insignificant results for the other variables. In order to summarize these results, we take the first principal component of the six policy measures and create an indicator for being above average on this measure; this indicates that a state had a relatively strong policy response. The interaction of “opposing” and “strong policy response” is −0.136 and significant at the 10% level, indicating that our result is bigger in the states with stronger policy responses.

TABLE 4.

The effect of opposing party status, by the strength of state policy responses

	Dependent variable: number of behaviors
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Opposing	−0.248**	0.064	−0.024	−0.212**	−0.144***	−0.058	−0.034
	(0.102)	(0.073)	(0.059)	(0.086)	(0.040)	(0.044)	(0.044)
Opposing x	0.015
date emerg. declared	(0.010)
Opposing x		−0.191*
schools closed		(0.094)
Opposing x			−0.102
bars/restaurants closed			(0.077)
Opposing x				0.140
gatherings banned				(0.093)
Opposing x					0.092
non‐ess. business closed					(0.078)
Opposing x						−0.131*
mandatory quarantine						(0.076)
Opposing x							−0.136*
strong response							(0.073)
N	1753	1753	1753	1753	1753	1753	1753
R 2	0.051	0.051	0.050	0.051	0.051	0.051	0.051

Note: This table shows the results from a regression of the number of preventative behaviors an individual reports on an indicator for having an opposing party governor, and on this indicator interacted with variables measuring the policy response of the state at the time of the survey. All states in our sample had declared emergencies as of the survey date; the variable “Date emerg. declared” is the date of the declaration relative to March 1st. The policy variables used in columns (2)‐(6) are indicators for the policy described. The variable used in column (7) is an indicator that the state is above average in the first principal component of the six policy variables, indicating a relatively strong policy response.

***p < 0.01, **p < 0.05, *p < 0.1.

One possible interpretation of the results in Table 4 is that the states that took strong action are primarily Democratic states, ¹⁹ and that this effect is therefore driven by a relatively bigger reaction to an opposing‐party governor among Republicans. We explore this hypothesis in Appendix Table A6, which allows us to calculate the opposing party effect separately for Democrats and Republicans and for these groups separately in strong versus weak policy response states. Column (3) of the table confirms that Republicans have a more negative response to a Democratic governor than Democrats do to a Republican governor: the opposing party effect is −0.18 for Republicans, but only −0.06 for Democrats. Breaking this down further, the opposing‐party response for Republicans in strong versus weak policy response states is −0.266 versus −0.073; for Democrats, these numbers are −0.101 and −0.014. In other words, the results in Table 4 are partially, but not entirely, driven by the stronger oppositional response of Republicans.

The fact that our results are concentrated in states where the governing party took strong action against the pandemic is consistent with the interpretation that individuals are less likely to comply with state directives when the leader of their state ‐ and the person who is typically the public face of efforts to combat COVID‐19 ‐ is from the other party. It is not immediately obvious from these results whether this occurs because the winning party is able to induce their supporters to comply, or whether opponents of the winning party reduce their efforts when the governor encourages preventative measures. Some evidence on this point comes from looking at mean behaviors among supporters and opponents of the winning party, in states with a stronger or weaker policy response. For supporters of the winning party, the mean number of behaviors is approximately 3.25, regardless of whether the state has a strong policy response. For supporters of the losing party, the number of behaviors is also around 3.25 when the state has a relatively weak policy response. When the state has a strong policy response, however, this number falls to 3.10. These results are more consistent with the idea that opponents of the governor reduce their behavior when an out‐group member encourages compliance. ²⁰

If our interpretation is correct, it would clearly be preferable that opponents of the governor did not behave in this way. Is this type of response to a political opponent inevitable? Or is this a feature of the uniquely polarized environment in the United States? In the next section, we attempt to answer this question.

5.3. Polarization and the opposite‐party effect

In Table 5, we examine how the response to an opposite party governor varies with an individual's level of affective polarization, as well as the state's level of polarization. As in Table 4, we focus here on the dependent variable “number of behaviors”. The first column of Table 5 shows that our main result changes little when we include a control for whether the individual is above the national mean in our measure of affective polarization (which we include in all other regressions in this table).

TABLE 5.

Regression results, by level of polarization

	Dependent variable: number of preventative behaviors
	(1)	(2)	(3)	(4)	(5)
Opposing	−0.116***	−0.250***	−0.079	−0.098	−0.085
	(0.043)	(0.076)	(0.051)	(0.078)	(0.088)
Opposing x		0.234***
polarized individual		(0.105)
Opposing x			−0.060	0.141	−0.327***
polarized state			(0.075)	(0.110)	(0.121)
Sample	All	All	All	Polarized	Non‐polarized
				individuals	individuals
N	1753	1753	1753	1014	739
R 2	0.050	0.053	0.050	0.052	0.141

Note: This table shows the results of regressions of the number of preventative behaviors an individual in the MTurk sample reports on an indicator for being of the opposing party to the state's governor, and this variable interacted with measures of individual and state polarization. The variable “polarized individual” is an indicator that the individual in our sample is above the national median (taken from the 2016 ANES) on a measure of affective polarization. The variable “polarized state” is an indicator that the state is above the median in the fraction of polarized individuals, as measured in the 2016 ANES. The variable “both” is the interaction between the two. The first column differs from our regression in Table 2 (panel 2, column (5)) only in that we have added the variable “polarized individual” as a control. This variable is added as a control in all regressions in this table. Otherwise, the controls and sample are identical to those used in Table 2, and we continue to cluster standard errors at the state by party level.

***p < 0.01, **p < 0.05, *p < 0.1.

In the second column of Table 5, we examine how the effect of an opposing party governor varies with an individual's relative hostility to the other party. Somewhat surprisingly, the effect of an opposing party governor is significantly weaker among more polarized respondents. Respondents in the bottom half of the polarization index show a response of −0.25 behaviors to an opposite party governor, while more polarized respondents show a response of around −0.02. We suspect that this result is driven by the fact that less polarized respondents have weaker prior beliefs about the appropriate course of action, either because they have less strong beliefs in general or because they tend to be less well‐informed (Palfrey & Poole, 1987). ²¹ Because we cannot separate these two variables ‐ hostility to the other party, and the fixedness of one's beliefs ‐ it is difficult to say whether individual polarization would increase the opposite‐party response ceteris paribus.

Even if individually polarized respondents show weaker responses, it is still possible that a polarized environment helps create the opposite‐party response that we see. Column (3) provides some evidence for this by examining how the response to an opposite party government varies in states that are more or less polarized. The point estimates suggest that the response is about twice as strong in more polarized states (although the difference is not statistically significant).

If our hypothesis about less polarized individuals being more responsive to their environment is correct, we should see a stronger response to state‐level polarization among this group. Columns (4) and (5) confirm this fact. In these columns, we repeat the state‐level polarization regressions from column (3), splitting the sample into more‐ and less‐polarized individuals. For polarized individuals, there is no effect of an opposing party governor, regardless of the level of polarization in the state. This is consistent with the idea that these individuals are relatively fixed in their behaviors, and do not respond strongly to the environment. For less polarized individuals, however, the opposing‐party effect is strongly concentrated in more polarized states. The effect of an opposing‐party governor for a moderate individual living in a non‐polarized state is around −0.08; for a moderate individual in a polarized state, it is around −0.4. This is approximately one‐third of a standard deviation of the independent variable.

Again, a comparison of means is informative here. For polarized individuals, and non‐polarized individuals living in non‐polarized states, the number of behaviors reported ranges from 3.17 to 3.28, depending on whether the individual opposes the state government and whether the state is polarized. For the group that shows an opposite‐party effect ‐ non‐polarized individuals living in polarized states ‐ the number of behaviors is slightly higher than this when their own party wins, at 3.34. When their party loses, however, the number of behaviors drops to 3.00. Again, these results are most consistent with the interpretation that these individuals reduce their behavior in response to an opposite‐party governor.

While these results are correlational, they are suggestive of the interpretation that polarized environments help create the opposite‐party response ‐ but only among individuals whose prior beliefs about the appropriate response are relatively weak. These individuals appear to reduce cooperative behavior when the governor is from the other party, and do so more in polarized environments. The effects of polarization therefore appear to “spill over” to individuals who remain relatively moderate.

6. CONCLUSION

In this paper, we have shown that individuals reduce compliance with social distancing measures when the other party holds the state governorship. This appears to be driven by individuals reducing their compliance behavior when the other party is in charge, rather than increasing behavior in response to their own party's efforts, which is consistent with the negative partisanship view of Klein (2020). The most likely explanation appears to be that individuals are simply less willing to comply with efforts to combat the pandemic when the effort is led by an out‐group member.

Clearly, this result implies that political identity plays some role in individuals' decisions to contribute to the public good. This is consistent with theories that relate social fragmentation to difficulty in resolving collective action problems. In support of this interpretation, we also show that the effects are stronger in more polarized states (although not among more polarized individuals, whose behavior appears to be relatively less susceptible to environmental influences). Most importantly, we show the opposing‐party effect is present only in certain conditions: among moderate individuals who live in polarized states. This implies that the tendency to reduce compliance in response to an opposing‐party leader is not an immutable feature of human nature, but one that can be reduced or eliminated in favorable conditions.

CONFLICT OF INTEREST

The author declares that there is no conflict of interest that could be perceived as prejudicing the impartiality of the research reported.

A. Pandemic Response

One possible concern in this empirical strategy could stem from the potential for delayed reaction in policy intervention by more polarized states. If states facing more political gridlock as a result of polarization are also slower to implement social distancing directives, this could bias us to finding lower compliance. To address this, we have collected data on the timing of three important COVID‐19 response policies. We focus specifically on the timing of declarations of a state of emergency (or public health emergency declarations), school closures, and limits to bar and restaurant operations.

Table A1 demonstrates that with limited exceptions, each of our sample states implemented each of these measures within the same week. One notable exception is Florida, which declared a state of emergency 2 weeks ahead of the last state in the sample to do so (Georgia). ²²

TABLE A1.

State COVID response policies and timing

State	Public State of emergency declared	School closures	Bar/Restaurant limits	Mandatory quarantine	Gatherings banned	Non‐ess. Bus. closed
Connecticut (D)	3/10/2020	3/17/2020	3/16/2020		3/12/2020	3/23/2020
Florida (R)	3/01/2020	3/17/2020	3/17/2020	3/23/2020
Georgia (R)	3/14/2020	3/16/2020	3/23/2020		3/23/2020
Indiana (R)	3/06/2020	3/19/2020	3/16/2020			3/23/2020
Iowa (R)	3/09/2020		3/17/2020
Kansas (D)	3/12/2020	3/17/2020			3/17/2020
Kentucky (D)	3/06/2020	3/16/2020	3/16/2020		3/17/2020	3/22/2020
Louisiana (D)	3/11/2020	3/13/2020	3/17/2020	3/22/2020	3/13/2020
Maine (D)	3/15/2020		3/18/2020		3/18/2020	3/25/2020
Michigan (D)	3/10/2020	3/16/2020	3/16/2020	3/24/2020	3/17/2020	3/24/2020
Mississippi (R)	3/14/2020	3/19/2020
Missouri (R)	3/13/2020	3/23/2020	3/23/2020		3/23/2020
Montana (D)	3/12/2020	3/15/2020	3/20/2020
Nevada (D)	3/13/2020	3/15/2020	3/17/2020		3/19/2020	3/20/2020
New Hampshire (R)	3/13/2020	3/15/2020	3/16/2020		3/16/2020
North Carolina (D)	3/10/2020	3/16/2020	3/17/2020		3/14/2020
Ohio (R)	3/09/2020	3/16/2020	3/15/2020	3/22/2020	3/17/2020
Oregon (D)	3/08/2020	3/16/2020	3/17/2020	3/23/2020	3/12/2020	3/23/2020
South Carolina (R)	3/13/2020
South Dakota (R)	3/13/2020	3/13/2020	3/23/2020
Virginia (D)	3/12/2020	3/24/2020	3/24/2020		3/24/2020
Washington (D)	2/29/2020	3/13/2020	3/15/2020	3/23/2020	3/25/2020	3/25/2020
West Virginia (D)	3/04/2020	3/13/2020	3/18/2020	3/24/2020		3/24/2020
Wisconsin (D)	3/12/2020	3/18/2020	3/20/2020	3/25/2020	3/17/2020	3/25/2020

Note: This table shows the close timing of COVID‐19 response across the 24 main states in our sample. All schools in Iowa went under voluntary closure on 3/15/2020 (eliminating the need for explicit mandate).

B. Robustness checks and supporting tables

TABLE A2.

Balancing tests

	Age	College	Female	Non‐white	Latinx
Opposing	0.781	−0.014	0.011	−0.012	0.011
	(0.508)	(0.014)	(0.018)	(0.022)	(0.010)
N	1753	1753	1753	1753	1753
R 2	0.031	0.017	0.021	0.061	0.044

Note: This table shows the results from regressions of demographic variables on an indicator for “opposing”, an indicator for “Democrat”, and state fixed effects, within the MTurk sample.

TABLE A3.

Regression discontinuity results, MTurk and Unacast Samples

	Panel 1: Mturk data
	Dependent variable
	Left home, non‐ess.	Work outside home	Work outside home, by choice	Wash hands	Stay home
Opposing	0.042	0.118**	0.090***	−0.049***	−0.002
	(0.062)	(0.055)	(0.033)	(0.018)	(0.026)
N	1753	932	201	1753	1753

	Dependent variable
	Cancel travel	Limit contact	Wear PPE	Other	Number of behaviors
Opposing	−0.076*	−0.144***	−0.036	0.014**	−0.333***
	(0.045)	(0.021)	(0.051)	(0.005)	(0.083)
N	1753	1753	1753	1753	1753

	Panel 2: Unacast data
	Dependent variable
	% Change, daily distance	% Change, non‐ess. visits
Opposing	−0.350	5.591*
	(2.310)	(3.385)
N	1675	1132

Note: This table shows regression discontinuity estimates of the effect of an opposing party on preventative behavior, using the MTurk and Unacast data. The running variable is opposing party share, which by construction is symmetric and there shows no bunching at the cutoff. We implement these regressions using the Stata package rdrobust (Calonico et al., 2014), which optimally selects the bandwidth. There are no controls in the regression. Standard errors are clustered at the state by party level.

***p < 0.01, **p < 0.05, *p < 0.1.

TABLE A4.

Regression results, MTurk sample: The effects of trust in state government

	Dependent variable
	Left home, non‐ess.	Work outside home	Work outside home, by choice	Wash hands	Stay home	Cancel travel	Limit contact	Wear PPE	Other	Number of behaviors
Opposing	0.000	0.0409	0.085*	−0.035***	−0.031***	−0.041***	0.011	0.004	0.003	−0.088**
	(0.0154)	(0.031)	(0.043)	(0.008)	(0.011)	(0.016)	(0.019)	(0.020)	(0.005)	(0.041)
N	1753	932	201	1753	1753	1753	1753	1753	1753	1753
R 2	0.050	0.099	0.248	0.052	0.063	0.031	0.066	0.058	0.028	0.055

Note: This table shows the results of OLS regressions of the indicated preventative measures on an indicator that the opposing party holds the governorship in a respondent's state, using the MTurk data. Respondents are classified as Democrat or Republican based on self‐reported responses. These regressions are identical to those presented in Table 2, and add a control for trust in state government. They also include a control for being a Democrat, state fixed effects (which absorb the effect of having a Republican governor, as well as the Republican vote share), and the following demographic controls: age, age squared, indicators for four education levels, indicators for 3 racial categories, an indicator for Latinx, and an indicator for female. Standard errors are clustered at the state by party level.

***p < 0.01, **p < 0.05, *p < 0.1.

TABLE A5.

Results with pre‐existing measures of polarization: re‐contacted sample

	Dependent variable: number of behaviors
Opposing	−0.321	−0.192
	(0.281)	(0.310)
Opposing x	0.441	0.272
polarized individual	(0.450)	(0.310)
N	146	146
R 2	0.232	0.234
Polarization measured in:	Nov 2019	Mar 2020

Note: This table shows the results from a regression of the number of preventative behavior an MTurk respondent engages in on an indicator for living in an opposing party state, and this indicator interacted with an indicator for being above the national median in a measure of affective polarization. The sample is a set of individuals who were initially surveyed for a separate project in November 2016, and who we were able to recontact for this project. These individuals do not necessarily live in close election states, but details of sample restriction and the regressions are otherwise identical to those in the main analysis tables.

***p < 0.01, **p < 0.05, *p < 0.1.

TABLE A6.

Differential response to an opposing‐party governor, Democrats versus Republicans and strong versus weak policy states

Dependent variable: number of behaviors
	(1)	(2)	(3)	(4)
Opposing	−0.115***	−0.120***	−0.179**	−0.073
	(0.059)	(0.057)	(0.073)	(0.063)
Opposing x			0.120*	0.059
Democrat			(0.069)	(0.108)
Opposing x				−0.194 *
strong response				(0.111)
Opposing x				0.107
Democrat				(0.142)
x strong response
N	1753	1753	1753	1753
R 2	0.050	0.043	0.043	0.045
State FE	Y	N	N	N

Note: This table explores whether the opposing‐party effect in states with a stronger policy response to COVID‐19 is driven by the differential response by Republicans versus Democrats. Column (1) replicates the result from Table 2; column (2) shows that this result is little altered when we replace state fixed effects with an indicator for being in a Democratic state, as required for this analysis. Column (3) shows how the opposing party effect differs for Democrats, while column (4) shows how it differs by both Democrat/Republican and strong policy response.

C. The effect of polarization on COVID‐19 cases

In order to understand how the social distancing effects we find translate to disease burden, we provide a quick back of the envelope calculation below. We first want to estimate the effect of mobility on COVID‐19 cases. To do this, we use case and Unacast mobility data for the weeks from 03/30/2020 to 04/27/2020 and estimate the following equation:

$$Y_{\mathit{cst}}=DistanceTravele{d}_{cs,t-2}+DistanceTravele{d}_{cs,t-2}\ast Populatio{n}_{cs}+NumCase{s}_{cs,t-1}+{\beta }_3{\mathbf{X}}_c+{\gamma }_s+{ϵ}_{\mathit{cps}}$$ (4)

where the outcome of interest, Y _cst is new COVID‐19 cases in county c, in state s, in week t. We regress this on a measure of the change in distance traveled (relative to a pre‐COVID‐19 benchmark) in week t − 2. ²³ Because new cases are likely dependent on the existing number of cases, we further control for the number of cases in a county in the previous week. The remainder of our controls mirror those used in our main Unacast specification, in Equation (2), however we cluster our standard errors at the county level.

Our results show that an additional 1% point reduction in distance traveled is associated with an additional 0.8 cases per 100,000 population, which is significant at the 5% level. We use this relationship, along with our estimated coefficient on “opposing” in the daily distance traveled regressions, to predict the number of avoidable cases for each opposing county, each week. Specifically, this number is CasesAvoidable _cst  = 0.836 ∗ (−0.052 + 0.841 ∗ Population _cs ). We then collapse this number across all opposing counties in our sample to arrive at our estimate of around 2000 avoidable cases.

D. MTurk survey details

D.1. Sampling and dates

We recruited workers from MTurk between March 24th and March 26th, 2020. The vast majority (around 99%) of workers answered between the 24th and 26th. However, because we allowed the survey to be posted for 1 week, there are a small number of respondents who answered later. Approximately 1200 of our respondents came from tasks that restricted respondents to those living in our close‐sample states; the number of respondents by state in this part of the sample approximately mirrors the population distribution across states. The remaining respondents were recruited from state‐specific tasks, in order to ensure a minimum number of respondents from each state. We were not always successful in recruiting the desired number of participants in each state, which is why we dropped Alaska and why we have a small number of respondents in states such as Maine or New Hampshire.

D.2. Sampling restrictions

We were able to recruit 2088 participants who had non‐missing data on all essential variables. We first drop respondents who report that they do not vote consistently for either party. This eliminates 287 respondents. We next eliminated state‐by‐party cells with fewer than five observations; this eliminated a total of 8 respondents. Finally, we dropped respondents who reported higher warmth toward the other party than the party they typically voted for. This eliminated 40 observations. This left us with our final sample of 1753 participants.

D.3. Survey text

ABOUT THIS RESEARCH: You are being asked to participate in a research study. Scientists do research to answer questions and learn new information. Some research might help change or improve the way we do things in the future. This consent information will tell you more about the study to help you decide whether you want to participate. Please read this information before agreeing to be in the study.

TAKING PART IN THIS STUDY IS VOLUNTARY: You may choose not to take part in the study or may choose to leave the study at any time. Deciding not to participate, or deciding to leave the study later, will not result in any penalty and will not affect your relationship with the University.

As an alternative to participating in the study, you may choose not to take part.

WHY IS THIS STUDY BEING DONE?: The purpose of this study is to understand more about individuals' responses to the COVID‐19 (Corona virus) epidemic.

You were selected as a possible participant because you chose to fulfill this task on MTurk.

HOW MANY PEOPLE WILL TAKE PART?: If you agree to participate, you will be one of 1000 participants taking part in this study.

WHAT WILL HAPPEN DURING THE STUDY?: If you agree to be in the study, you will be asked to do the following things:

• You will be asked to complete a survey which will take approximately 5 min of your time.

• The survey will ask for some basic demographic information and about your attitudes on a variety of issues. It will also ask about how your behavior and attitudes have changed as a result of the epidemic.

• You will be compensated $1.00 for completing the survey.

WHAT ARE THE RISKS OF TAKING PART IN THE STUDY?: While participating in the study, the potential risks include:

• Discomfort answering some of the questions in the survey. If you do feel uncomfortable at any time, you may discontinue the survey or skip any question.

• Loss of confidentiality in the data. To avoid this risk, responses to this survey are collected anonymously; no information that could identify you will be requested during the data collection phase. A limited number of research team members will have access to the data during data collection.

WHAT ARE THE POTENTIAL BENEFITS OF TAKING PART IN THE STUDY?: We don't expect you to receive any benefit from taking part in this study but we hope to learn things that will help scientists in the future.

HOW WILL MY INFORMATION BE PROTECTED?: Efforts will be made to keep your survey responses confidential. We cannot guarantee absolute confidentiality, but we have guarded against this risk by not asking for any information that could identify you. The data collected in this study will be stored in a secure location and will be accessible only to a limited number of researchers.

Organizations that may inspect and/or copy your research records for quality assurance and data analysis include groups such as the study investigator and his/her research associates, the University Institutional Review Board or its designees, and (as allowed by law) state or federal agencies, especially the Office for Human Research Protections, who may need to access the research records.

Your Mechanical Turk Worker ID will be used to distribute payment to you but will not be stored with the research data we collect from you. Please be aware that your MTurk Worker ID can potentially be linked to information about you on your Amazon public profile page, depending on the settings you have for your Amazon profile. We will not be accessing any personally identifying information about you that you may have put on your Amazon public profile page.

WILL MY INFORMATION BE USED FOR RESEARCH IN THE FUTURE?: Information collected in this study may be used for future research studies or shared with other researchers for future research. Since identifying information will not be stored with the data, we will not ask for your additional consent.

WILL I BE PAID FOR PARTICIPATION?: You will be paid $1.00 for the survey upon completion.

MTurk does not allow for prorated compensation. In the event of an incomplete HIT, you will not receive any compensation.

This study contains a number of checks to make sure that participants are finishing the tasks honestly and completely. As long as you read the instructions and complete the tasks, your HIT will be approved. If you fail these checks, your HIT will be rejected and you will not receive any compensation.

WORLD HEALTH ORGANIZATION (WHO) SHOULD I CALL WITH QUESTIONS OR PROBLEMS?:

For questions about your rights as a research participant, to discuss problems, complaints, or concerns about a research study, please contact Research Compliance.

PARTICIPANT’S CONSENT: In consideration of all of the above, I give my consent to participate in this research study. By proceeding, I confirm that I am 18 years old, and agree to take part in this study.

Questions:

1. In which state do you currently reside?

2. What is your year of birth?

3. What is the highest level of school you have completed or the highest degree you have received?

   • (a)Less than high school degree
   • (b)High school graduate (high school diploma or equivalent including GED)
   • (c)Some college but no degree
   • (d)Associate degree in college (2‐year)
   • (e)Bachelor's degree in college (4‐year)
   • (f)Master's Degree
   • (g)Doctoral Degree
   • (h)Professional degree (JD, ND)

4. Are you Spanish, Hispanic, or Latino or none of these?

   • (a)Yes
   • (b)None of these

5. Choose one or more races that you consider yourself to be:

   • (a)White
   • (b)Asian
   • (c)Black or African American
   • (d)Native Hawaiian or Pacific Islander
   • (e)American Indian or Alaska Native
   • (f)Other

6. What is your sex?

   • (a)Male
   • (b)Female

7. How would you rate your political views on the following scale?

   • (a)Extremely left‐wing
   • (b)Left‐wing
   • (c)Somewhat left‐wing
   • (d)Neither left‐ nor right‐wing
   • (e)Somewhat right‐wing
   • (f)Right‐wing
   • (g)Extremely right‐wing

8. Which party do you vote for most often?

   • (a)Democratic Party
   • (b)Republican Party
   • (c)Other
   • (d)Prefer not to say

9. On a scale of 0–100, please indicate how warmly you feel toward each of the following political parties. 0 indicates that you feel very coldly toward the party, while 100 indicates that you feel very warmly. If you do not know how you feel about a party, do not enter a number. The Democratic Party

10. On a scale of 0–100, please indicate how warmly you feel toward each of the following political parties. 0 indicates that you feel very coldly toward the party, while 100 indicates that you feel very warmly. If you do not know how you feel about a party, do not enter a number. The Republican Party

11. Which of the following statements do you agree with more?

    • (a)People can generally be trusted
    • (b)You can't be too careful when dealing with people

12. Please indicate how much you trust the following group on a scale of 1–10. A rating of “1” indicates a high level of distrust, while a rating of “10” indicates a high level of trust. The people who live in your neighborhood

    • (a)1
    • (b)2
    • (c)3
    • (d)4
    • (e)5
    • (f)6
    • (g)7
    • (h)8
    • (i)9
    • (j)10

13. Please indicate how much you trust the following group on a scale of 1–10. A rating of “1” indicates a high level of distrust, while a rating of “10” indicates a high level of trust. The federal government

    • (a)1
    • (b)2
    • (c)3
    • (d)4
    • (e)5
    • (f)6
    • (g)7
    • (h)8
    • (i)9
    • (j)10

14. Please indicate how much you trust the following group on a scale of 1–10. A rating of “1” indicates a high level of distrust, while a rating of “10” indicates a high level of trust. Your state government

    • (a)1
    • (b)2
    • (c)3
    • (d)4
    • (e)5
    • (f)6
    • (g)7
    • (h)8
    • (i)9
    • (j)10

15. Please indicate how much you trust the following group on a scale of 1–10. A rating of “1” indicates a high level of distrust, while a rating of “10” indicates a high level of trust. Your local government

    • (a)1
    • (b)2
    • (c)3
    • (d)4
    • (e)5
    • (f)6
    • (g)7
    • (h)8
    • (i)9
    • (j)10

16. Please indicate how much you trust the following group on a scale of 1–10. A rating of “1” indicates a high level of distrust, while a rating of “10” indicates a high level of trust. Large corporations

    • (a)1
    • (b)2
    • (c)3
    • (d)4
    • (e)5
    • (f)6
    • (g)7
    • (h)8
    • (i)9
    • (j)10

17. Please indicate how much you trust the following group on a scale of 1–10. A rating of “1” indicates a high level of distrust, while a rating of “10” indicates a high level of trust. The American Medical Association

    • (a)1
    • (b)2
    • (c)3
    • (d)4
    • (e)5
    • (f)6
    • (g)7
    • (h)8
    • (i)9
    • (j)10

18. Please indicate how much you trust the following group on a scale of 1–10. A rating of “1” indicates a high level of distrust, while a rating of “10” indicates a high level of trust. The Center for Disease Control (CDC)

    • (a)1
    • (b)2
    • (c)3
    • (d)4
    • (e)5
    • (f)6
    • (g)7
    • (h)8
    • (i)9
    • (j)10

19. Please indicate how much you trust the following group on a scale of 1–10. A rating of “1” indicates a high level of distrust, while a rating of “10” indicates a high level of trust. The WHO

    • (a)1
    • (b)2
    • (c)3
    • (d)4
    • (e)5
    • (f)6
    • (g)7
    • (h)8
    • (i)9
    • (j)10

20. During this pandemic, I am most worried about

    • (a)Getting COVID‐19
    • (b)Spreading COVID‐19
    • (c)Neither

21. I think social distancing is

    • (a)Extremely important: No one should be outside of their home except to go to work (if they must), the doctor, or the grocery store.
    • (b)Very important: I understand that people sometimes leave their house, but they should be taking extra precautions
    • (c)Somewhat overblown: As long as people wash their hands, I don't really see the problem
    • (d)Completely overblown: People need to just live their life as usual, and it will all die down

22. Have you left your home in the past 48 h?

    • (a)Yes
    • (b)No

23. If so, what places did you visit? Please check all that apply.

    • (a)The homes of family or friends
    • (b)Work
    • (c)Grocery store or pharmacy
    • (d)Doctor's office/hospital
    • (e)Other stores
    • (f)Restaurants, bars, or other entertainment venues
    • (g)Outdoors (walk, jog, playground, etc)
    • (h)Other
    • (i)I have not left my home in the past 48 h

24. Have any of your family members left their homes (for reasons other than work or visiting the doctor) in the past 48 h?

    • (a)Yes
    • (b)No
    • (c)I have no family members, or do not know

25. What percentage of people in your community do you think have left their homes in the past 48 h?

    • (a)0%–10%
    • (b)10%–20%
    • (c)20%–30%
    • (d)30%–40%
    • (e)40%–50%
    • (f)50%–60%
    • (g)60%–70%
    • (h)70%–80%
    • (i)80%–90%
    • (j)90%–100%

26. Please describe your current work situation.

    • (a)I did not work outside the home prior to the COVID‐19 pandemic, and this has continued.
    • (b)I worked outside the home prior to the COVID‐19 pandemic, and am continuing to do so at the request or requirement of my employer.
    • (c)I worked outside the home prior to the COVID‐19 pandemic, and am continuing to do so by choice.
    • (d)I worked outside the home prior to the COVID‐19 pandemic, and am now working remotely at the request or requirement of my employer.
    • (e)I worked outside the home prior to the COVID‐19 pandemic, and am now working remotely by choice.
    • (f)I worked outside the home prior to the COVID‐19 pandemic, and have lost my job.
    • (g)Other.

27. Please indicate whether you have taken any of the following step in the past week to limit the spread of COVID‐19. (Select all that apply).

    • (a)Washing hands or using hand sanitizer more frequently
    • (b)Wearing gloves or a mask while out of the house
    • (c)Staying home more often
    • (d)Limiting contact with elderly or high‐risk friends and family
    • (e)Cancel ling planned travel
    • (f)Other

28. Please indicate whether you have taken any of the following step since the pandemic first began to limit the spread of COVID‐19. (Select all that apply).

    • (a)Washing hands or using hand sanitizer more frequently
    • (b)Wearing gloves or a mask while out of the house
    • (c)Staying home more often
    • (d)Limiting contact with elderly or high‐risk friends and family
    • (e)Canceling planned travel
    • (f)Other

29. When did you begin taking these measures?

    • (a)This week
    • (b)Last week
    • (c)More than 2 weeks ago

30. My main source of information on COVID‐19 has been:

    • (a)Local government
    • (b)State government
    • (c)Federal government
    • (d)Health organization (CDC, WHO, etc)
    • (e)Newspaper or television news
    • (f)Social media: friends and family
    • (g)Social media: others

31. At any point since the pandemic began, have you received any information or social media posts that said social distancing was an unnecessary precaution?

    • (a)Yes
    • (b)No

32. At any point since the pandemic began, have you received any information that social distancing is a good way to prevent getting and spreading COVID‐19?

    • (a)Yes
    • (b)No

Cornelson, K. , & Miloucheva, B. (2022). Political polarization and cooperation during a pandemic. Health Economics, 31(9), 2025–2049. 10.1002/hec.4560

DATA AVAILABILITY STATEMENT

The survey data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. The mobility data that support the findings of this study are available from Unacast. Restrictions apply to the availability of these data, which were used under license for this study. Data are available through Unacast, and the authors are willing to assist with this process.