1. Existing evidence

A vast literature investigates the role of partisanship in covid-related behaviors and attitudes. Given the focus of our analysis, for space reasons we review pieces that compare effects across time, individual behaviors, or according to elite partisanship. Supplemental Table S1 summarizes each piece. A few generalizations are worth highlighting.

First, few studies examine variation in masking behavior over time and those that do suggest a relatively steady partisan gap through 2020 and early 2021 (Baxter-King et al., Reference Baxter-King, Brown, Enos and Vavreck2022; Gadarian et al., Reference Gadarian, Goodman and Pepinsky2022).Footnote ¹ Baxter-King et al. also show that Republicans’ mask-wearing is lower the more conservative is the geographic area and interpret this effect as reflecting prosocial behavior. Arguably the closest study to ours is Constantino et al. (Reference Constantino, Cooperman, Keohane and Weber2022), which examines with panel data how personal covid experience alters behavior and finds that adverse experience affects Democratic and Republican partisans similarly. Our analysis, by extending past the vaccine rollout and examining individual vaccination status, assesses how the partisan gap in mask-wearing changes once the public health conditions improve substantially.

Second, various scholarship suggests that the difference between Democrats and Republicans in isolating and social distancing behaviors increased between the first 2 months of the pandemic and the next 12. This finding comes from surveys through April 2021 (Gadarian et al., Reference Gadarian, Goodman and Pepinsky2022) and the summer-fall of 2020 (Clinton et al., Reference Clinton, Cohen, Lapinski and Trussler2021; Kaushal et al., Reference Kaushal, Lu, Shapiro and Jennifer2022) and from county-based social mobility data based on mobile phones (Allcott et al., Reference Allcott, Boxell, Conway, Gentzkow, Thaler and Yang2020; Gollwitzer et al., Reference Gollwitzer, Martel, Brady, Pärnamets, Freedman, Knowles and Van Bavel2020; Bisbee and Lee, Reference Bisbee and Lee2022).Footnote ² There is also evidence that personal and local health experience with COVID-19 affects social distancing, although the role of partisanship in shaping this relationship varies across the works. Barrios and Hochberg (Reference Barrios and Hochberg2021) find that the effect of local outbreaks on social mobility is lower for conservative counties, and Bisbee and Lee (Reference Bisbee and Lee2022) show that this relationship strengthens over time. By comparison, Constantino et al.’s (Reference Constantino, Cooperman, Keohane and Weber2022) panel analysis suggests that personal COVID-19 experience increases isolating more for Republicans, while Druckman et al. (Reference Druckman, Klar, Krupnikov, Levendusky and Ryan2021) find that affective polarization has a lower effect on covid-related behaviors in counties with higher outbreaks.Footnote ³ Combined, these studies do not provide a clear prediction on whether Democrats or Republicans were more affected by health risks, and therefore how the vaccine rollout period and individual vaccination status should alter the impact of partisanship. Our analysis shows that as the personal net benefits from social distancing decline, the partisan gap decreases precipitously, with more of an effect on Democrats.

Third, working remotely and its correlation with partisanship across time and other behaviors has not been a major topic of inquiry. Graham and Singh (Reference Graham and Singh2024) show that working from home, isolating and masking grew similarly by party in March and April 2020. However, as described above, the literature suggests that the partisan gap in other behaviors, particularly social distancing, developed in the ensuing months. We expect individuals to have less control over whether they work from home without incurring high costs and therefore include it as a point of comparison to the other responses.

Finally, studies of the effects of elite partisanship on covid-related behavior produce a range of findings. Gadarian et al. (Reference Gadarian, Goodman and Pepinsky2021) suggest partisan elite endorsements did not significantly alter covid-related attitudes about policies. However, Barrios and Hochberg (Reference Barrios and Hochberg2021) show that the Conservative Political Action Committee’s March 2020 announcement regarding self-quarantining due to COVID-19 exposure had a larger effect on social mobility in more conservative counties. On government approval, Pickup et al. (Reference Pickup, Stecula and van der Linden2020) find that in 2020 Democrats were less approving than Republicans of the federal government.Footnote ⁴ Our study leverages variation in gubernatorial policy actions to examine how out-party and in-party approval vary based on these actions relative to co-partisanship. Closer to this aim, Grossman et al. (Reference Grossman, Kim, Rexer and Thirumurphy2020) provide evidence that although both Republican and Democratic counties responded to governors’ stay-at-home recommendations, Democratic counties were substantially more likely to follow them, particularly when the state had a Republican governor. Our analysis of approval differs in important ways, including in expectations regarding Republicans. For instance, whereas Grossman et al. find an average increase in distancing by Republican counties after a stay-at-home recommendation, we would not expect Republicans to be more likely to approve of the state response when a stay-at-home order is recommended/enacted. Our findings show that this is indeed the case, along with other new findings that result from the different context examined, including how gubernatorial policy choices moderate partisan differences in approval.

2. Data and measurement

Our data are from Gallup’s COVID Tracking Survey, derived from the Gallup Panel, a probability-based database with approximately 100,000 members at any given time who are originally contacted via address-based sampling or random-digit dialing (using cell phones and landlines) before completing the survey on-line.Footnote ⁵ Gallup invited a random sample of panel members to participate in the COVID-19 survey in phases. After all eligible members had been invited one time, respondents were invited if it had been at least 4 weeks since their last invitation or they were new panel members. All participants received a $1 postpaid incentive for completion.

Gallup fielded the survey on March 13, 2020 and collected roughly 1,000 responses per day until April 26, when the sample declined to roughly 500 responses per day, and then starting August 17, 2020, Gallup stopped collecting data during the first two weeks of each month and reduced the total sample size per month from approximately 20,000 (from March 2020 to August 2020) to approximately 4,000 (from September 2020 onwards). Our data extend through September 16, 2021, thereby encompassing over a year and a half of the pandemic. Overall, the average survey completion rate is 94% and average response rate is 46%. Invitees had 1 week to complete the survey during the daily collection period and 2 weeks during the monthly period.

In total, we have 164,327 responses from 54,216 unique individuals. Gallup creates sample weights based on the respondents’ probability of selection into the panel, and additionally, Gallup statisticians adjust for nonresponse bias. The sample targets are based on Census Bureau-defined region, race, Hispanic ethnicity, five age groups, gender, and educational attainment. We use the sample weights in our analysis so that the underlying sample approximates national representation.

An advantage of the survey instrument is that over 90% of the observations are from individuals who responded multiple times. Over 40,000 individuals responded at least twice, over 30,000 individuals at least thrice, and almost 20,000 four or more times. Table S2 in the Supplemental Materials provides the complete distribution of the number of survey responses. This panel structure allows us to not only include individual fixed effects as a way of mitigating omitted variables bias, but also trace out the response of individuals over key periods of the pandemic, including when the vaccines became widely available.Footnote ⁶ As the supplemental materials show (Figure S3), starting in February 2021 the rate of fully vaccinated individuals rises such that by April 19, more than half of the country had received at least one shot of a COVID-19 vaccine.

Notably, the large-scale rollout of the vaccines had both direct and indirect effects on the health risks associated with COVID-19. First and most obviously, those who chose to become vaccinated significantly reduced their risk of severe illness or death from contracting the disease. Second and more indirectly, vaccines reduced the probability of transmission, even for those with break-through infections, and accordingly reduced the risk for some groups of contraction, including for the unvaccinated (e.g., Tan et al., Reference Tan, Kwan, Rodríguez-Barraquer, Singer, Park, Lewnard, Sears and Lo2023).

The key covid-related items include whether the respondent has recently worn a mask, mostly isolated themselves from non-household members, visited their place of work, and whether they are worried about contracting COVID-19. In particular, the items capture whether the respondent has “worn a mask on [their] face” in the past 7 days outside their home; has “completely” or “mostly” isolated themselves, having “no” or “very little contact with people outside [their] household” in the past 24 hours; visited their “place of work” in the past 24 hours; and whether the respondent is “very worried” that they “will get coronavirus (COVID-19).” The online supplemental materials provide the complete survey question wording (S4). For all variables, Supplemental Table S5 provides descriptive statistics and for these covid-related survey items, Table S6 provides summary statistics by party.Footnote ⁷

Our mix of behaviors captures variation in the extent to which actions carry different perceived net costs as well as in public symbolism. For most people, socially isolating has relatively high costs and affords less opportunity for public symbolism. We accordingly expect that as the personal health benefits from socially isolating decline following the vaccine rollout, the partisan gap in this behavior should decrease. By comparison, masking carries lower personal costs while enabling publicly expressive behavior. Given these qualities and that the vaccine rollout reduced the health-related costs of not masking, we expect the rollout to be associated with a larger gap between Democrats and Republicans. The third behavior, working remotely, arguably has the lowest element of choice among the three. For many individuals, refusing to go into the office carried the cost of termination. Therefore, we expect less across-time variation in the partisan gap in working remotely than in other, more choice-based behaviors. Finally, we include the item on worry to gauge how behaviors relate to individuals’ concern about contracting the disease.

To illustrate the raw descriptive statistics, Figure 1 depicts the over-time partisan variation between Democrats and Republicans. With each item, a partisan gap exists for a substantial portion of the pandemic, but the size of the gap varies by item across the period, particularly in relation to the rollout of vaccines, which is identified by the grey shading. We define the vaccine rollout as occurring between February 1, when more than 1% of the population had received the vaccine and April 19, when 50% of the adult population had received at least one shot and all adults were vaccine-eligible (AJMC, 2021).Footnote ⁸

Notes: Shaded area indicates vaccine roll-out. Starting in September 2020, Gallup moved from daily to monthly data collection.

Figure 1. Evolution of COVID-19 reponses, by political affiliation.

In Figure 1, the Democrat-Republican gap on mask-wearing ranges from 14 to 29 percentage points before March 2021 and then widens to 42–50 percentage points by summer. By comparison, the partisan gap for socially isolating declines from over 30 percentage points in February 2021 to less than half that by the summer, due to Democrats becoming substantially less likely to engage in this activity. Worry over COVID-19 follows a similar (if not more) dramatic pattern, with worry declining precipitously for Democrats throughout the spring. Finally, the partisan gap in working remotely closes somewhat after vaccines become widely available, but otherwise remains relatively consistent throughout the period.

Of course, Figure 1 does not hold the individual’s disposition constant or account for confounding factors, making these patterns quite tentative. Our subsequent analyses account for a variety of such confounders, including demographics, pandemic conditions, and state policies. The demographic factors are from the survey, including age, race, ethnicity, gender, employment, education, income, and whether the respondent is living with children. Details on the measurement and descriptive statistics are given in the Supplemental Materials (Table S5). In analyses with individual-level fixed effects, many of the demographic controls drop out since they are time invariant, although those that do change are included.

To capture pandemic conditions, we use daily 7-day totals on newly confirmed COVID-19 cases per capita in the county (per million inhabitants) from USAFacts, which pulls the original data from state health departments. As subsequently described, we also conduct analyses that control for individual vaccination status. State policies are from Oxford University’s COVID-19 Government Response Tracker, as adapted to the United States by Hallas et al. (Reference Hallas, Hatibie, Majumdar, Pyarali, Koch, Wood and Hale2020). Given the content of the covid-related survey responses, we focus on the policies of stay-at-home orders, restrictions on social gatherings, mask mandates, and workplace closings. For each, a binary variable indicates whether a restriction is mandated on that date. Supplemental S8 provides full coding details.

The analysis of approval of the state response is also from the Gallup COVID Tracking Survey. Specifically, the key survey item asks whether the respondent approves of the way the state government is “handling the response to the coronavirus in the U.S.” (Supplemental S2 provides the complete wording.) Unlike the earlier covid-related items, the approval item is only asked through April 2, 2020, meaning we cannot employ individual-level fixed effects in analysis of this response. Still, identification is facilitated by within-state variation in policies—the data capture periods before and after policies were adopted—and respondent partisanship as well as by cross-state variation in policies and gubernatorial party.

3. Individual COVID-19 behaviors and attitudes across the pandemic

We conduct two main types of analyses in this section. First, we analyze how covid-related responses change in response to the vaccine rollout period. This analysis encompasses that the vaccine rollout provides health benefits even to the unvaccinated. Second, we employ a DIDs strategy in which vaccination status is the treatment. For these latter models, we show results both for the entire population and only for those who are ultimately vaccinated given that the “never treated” population may differ from those who are ultimately treated.

Beginning with the analysis of the vaccine rollout, we first estimate regressions of the following form with individual-level fixed effects:

(1)\begin{equation}{y_{ict}} = \xi \left( {{P_{it}} \times Mont{h_t}} \right) + \gamma {P_{it}} + \zeta COVI{D_{ct}} + \phi {D_{it}} + \kappa {R_{st}} + {\lambda _t} + {\nu _i} + {\epsilon _{it}}\end{equation}

where $y$ denote individual $i$’s survey response and survey date $t$ for each covid-related response; $P$ denotes partisan affiliation (with Republicans the omitted category); $Month$ reflects the month in which the survey was conducted; c represents the original county of residence of individual $i$, $COVID$ denotes the logged number of new COVID-19 cases per capita in the county; $D$ denotes a vector of demographic controls; $R$ reflects pandemic-related policies in state s on date t; and $\lambda $ and $\nu $ denote fixed effects for survey date and individual, respectively. We estimate linear probability models given the fixed effects and to ease interpretation of results. We cluster standard errors at the county-level to allow for arbitrary degrees of local autocorrelation across individuals within county.

The key coefficients are those on the interactions between the party and month variables, as they capture how the gaps between partisans evolve over the pandemic, relative to the base month of the survey (i.e., the first month in which the survey was asked). As other research has described, early in the pandemic, elite cues were weaker (e.g., Clinton et al., Reference Clinton, Cohen, Lapinski and Trussler2021). The CDC does not recommend masking until April 3, following which Trump proceeds not to wear one in front of the press for months (Netburn, Reference Netburn2021). Given the evolving set of signals in March and April, we would expect more divergence between Democrats and Republicans after April 2020.

We limit the individual-level fixed effects analyses to respondents who answered in the first month in which the item was asked. Even with this restriction, we have over 500 observations per month for all survey items and months. Still, an advantage of analyzing the data without the fixed effects is that doing so incorporates individuals who completed the survey/item only once. We therefore also analyze Equation (1) replacing individual with county-level fixed effects.

Figure 2 presents the findings by plotting the parameter estimates for each interaction between the month and Democratic partisan affiliation. The dark blue dots and lines depict the estimated coefficient and the 95% confidence interval, respectively, for the analysis with the individual fixed effects and the light blue dots and lines without them. Notably, each captures the additional partisan gap between Democrats and Republicans over the baseline gap in the first month of the survey item. In other words, a positive value indicates Democrats are more likely than Republicans to respond affirmatively to the survey question than at the start of the pandemic. The figure’s notes give the baseline gap as estimated by the model without the individual fixed effects. For the model with them, an estimate for the base month cannot be calculated (because the baseline effect is captured by the panel effects), and we therefore focus on magnitudes involving the model without them.

Notes: The figure plots the estimated change in the partisan gap between Democratic respondents and Republican respondents relative to this gap in the base month of the survey item. The dots and lines reflect, respectively, the estimated coefficients and 95% confidence intervals. Estimates of the partisan gap in the base month are as follows for the model without individual fixed effects: for mostly isolating 0.061; mask-wearing, 0.169; worry, 0.105; and working remotely, 0.018. Standard errors clustered at the county level. Shaded area indicates vaccine roll-out. Starting in September 2020, Gallup moved from daily to monthly data collection.

Figure 2. Change in partisan gap in COVID-19 response.

A few broader findings are worth noting before delving into the specifics. First, the coefficients are remarkably stable between the specifications with and without the individual fixed effects. Second, the Democrat–Republican partisan gap varies across time and response. The dissimilarity in the post-vaccine period is particularly striking, with the partisan gap increasing for masking, decreasing for isolation and worry, and remaining steadier for remote work until September 2021, when it declines. Moreover, for masking it is only after the probability of contagion declines in association with the vaccine rollout that substantial increases from the base month emerge. All these results are consistent with our theoretical expectations that partisanship’s influence on behavior should be constrained by material consequences and exaggerated by expressive benefits.

Consider the top-left panel on isolating. The estimates indicate that the Democrat–Republican gap grows from the March baseline, when elite signals were still evolving and individuals had less information about the disease, continues at a heightened level through the next winter and then declines precipitously from this peak with the vaccine rollout. Table S9 in the supplemental materials gives the precise parameter estimates. As described there and in the figure’s notes, the baseline partisan gap is estimated to be 6 percentage points. Accordingly, where Figure 2 shows the increased gap in March 2021 to be over 20 percentage points, the total gap is 26 percentage points. By April 2021, the increase over the base month shrinks to 7 percentage points, and by September 2021, there is no longer any significant difference from March 2020. This pattern comports with the theoretical prediction that partisan differences will be lower when the net personal costs from an action are higher.

Moving on to the top-right panel, which captures mask-wearing, a quite different pattern emerges. The Democratic–Republican partisan gap grows in May–June 2020 over the base month, then returns to levels close to or below that of the base month before increasing substantially in the spring of 2021. Indeed, the partisan gap is over 25 points higher in August than initially.Footnote ⁹ The estimated baseline gap is 17 percentage points, suggesting Democrats are at least 42 percentage points more likely than Republicans to wear a mask in August 2021.Footnote ¹⁰ Again, this increased partisan gap comports with our theoretical expectations; with the vaccine rollout, masking offers less substantial benefits yet remains a low-cost and publicly symbolic activity, making partisan behavior more attractive.

One might ask whether the changes in partisan gaps are driven more by Republicans versus Democrats, and Supplemental Figures S14 and S15 show how each group’s behavior changes relative to that of Independents. For isolating, the downward trend in the partisan gap that occurs during the vaccine rollout is associated with more movement by Democrats than Republicans, although each partisan group becomes more similar to Independents. For masking, once the rollout occurs, both Democrats and Republicans move decisively away from Independents in opposite directions.

Continuing with Figure 2 for the results on worry, in the bottom-left panel, the estimates indicate a similar general pattern to isolating. In April 2020, the initial gap is 11 percentage points. It then increases in fall 2020, but once vaccines become widely available, declines such that by April–September 2021 it is lower than in April 2020.

Finally, in the bottom-right panel of Figure 2, the partisan gap for working remotely fluctuates less than the other items. The initial gap, without the panel effects, suggests Democrats were only 2 percentage points more likely to work remotely in May 2020 and as the pandemic evolves, the partisan gap is relatively stable until the vaccine rollout, when it declines from the initial month depending on the specification. A more pronounced change occurs in September 2021, when many companies and schools reopened in person (New York Times, 2021; Rubenstein, Reference Rubenstein2021). The gap then closes to a level only slightly greater than it was in the initial month. The general stability until the September back-to-office/school period is consistent with respondents not having full discretion over whether to work remotely and correspondingly, our expectation that the vaccine would be less associated with a change in the partisan gap in this behavior than in other, more choice-based actions.

Thus far we have focused on the vaccine rollout, given that this event affected all individuals. However, we have additional information on whether an individual was vaccinated at the time of the survey response. Gallup began collecting this information beginning January 25, 2021; Supplemental S4 describes the full question wording. We code vaccinated as an indicator based on whether an individual had received at least one dose, and for dates earlier than January 25 impute a 0.Footnote ¹¹

Multiple analyses are conducted with this variable as a predictor of the behaviors.Footnote ¹² In particular, we employ a DID strategy using both standard two-way fixed effects (TWFE) models and a stacked DID design. As Chiu et al. (Reference Chiu, Lan, Liu and Yiqing2025) note, most DID analyses in political science use TWFE. The stacked DID estimator was developed for contexts such as ours where observations are not treated simultaneously (e.g., Gormley and Matsa, Reference Gormley and Matsa2011). It involves creating separate datasets for each treated cohort by combining data from that cohort and data from the control cohort, and “stacking” them together into a single dataset. Chiu et al. (Reference Chiu, Lan, Liu and Yiqing2025) find that results from stacked DID (and other estimators developed for heterogenous treatment effects) are usually qualitatively similar to those from TWFE. In the main text, we show results from both TWFE and stacked DID analyzing all observations, including respondents who are never-treated given that Chiu et al. (Reference Chiu, Lan, Liu and Yiqing2025) suggest doing so for stacked DID. In the Supplemental Materials (S20), we show findings from analyzing only the ever-treated (i.e., ever-vaccinated), and as detailed below, the main substantive findings are consistent.Footnote ¹³ The Supplemental Materials (S21) also present evidence supporting the parallel trends assumptions.

Each model interacts vaccination status with partisanship, to allow for different effects of vaccination on behavior and attitudes by party. For comparability with earlier results, the omitted partisan category is Republicans. We anticipate vaccination status will have the same directional effects as the rollout period on behaviors given that it reduces the likelihood of contracting the illness. Therefore, we expect individuals to be less likely to isolate, worry, and work remotely. Because Democrats were more likely to engage in these behaviors initially, we expect the decrease to be largest for them. For masking, the lower cost and public symbolism of the action suggests a different calculus. For Republicans (and Independents), declines should occur but for Democrats the public symbolism may prevent any sort of significant decline.

Table 1 presents the results. The interaction with Democratic respondents represents the difference from Republican respondents, and the main effect of vaccination represents the estimated effect for Republicans, the omitted partisan category. Consistent with Chiu et al.’s (Reference Chiu, Lan, Liu and Yiqing2025) arguments, the results are remarkedly consistent between the TWFEs and stacked DID estimators. (Note that the high number of observations for the stacked DID results is due to the “stacking” of the data into cohort-panel units and is the norm for these analyses; see, e.g., Gormley and Matsa, Reference Gormley and Matsa2011.) In each case, as anticipated, the coefficients on the interaction for Democrats for isolating, worry, and working remotely are consistently negative, suggesting vaccination induced a larger decline for Democrats than Republicans. This differential estimate reduces the partisan gap between the groups, given the Democrats’ originally greater likelihood of isolating, worrying, and working remotely.Footnote ¹⁴

Table 1. Pre- versus postvaccination status, DID

Notes: Standard errors clustered by individual below coefficients. Sample includes never-treated individuals. Supplemental S20 shows results for only the ever-treated. ***p < 0.01, **p < 0.05, *p < 0.10, two-tailed.

However, for masking, not only were Democrats more likely than Republicans to engage in this behavior post-vaccine, but Democratic mask-wearing increases (e.g., for the stacked DID results, −0.067 + 0.168 = 0.101, p < 0.01). These findings comport with the earlier ones about the large increase in the partisan gap but also raise questions about why the estimate on Democrats is not merely null but significantly positive. One possibility is that as these individuals became less likely to isolate, they were motivated by the public symbolism of the act in addition to the increased health risks associated with no longer isolating.Footnote ¹⁵

The main effects on vaccination in Table 1, representing the estimates for Republicans, are consistently negative and significant at conventional levels for all but working remotely. However, as shown in Supplemental S20, when the analysis is limited to the ever-treated, the TWFE results suggest the main effect on vaccination, for Republicans, is only significant for masking and the stacked DID results suggest this main effect is significant only for masking and isolating. However, even when the sample is limited in this way, the differences between Democrats and Republicans continue to hold strongly.

Together, Table 1 and Figure 2 (along with the supplemental analyses) suggest that partisanship’s function as a perceptual screen is conditional, supporting the Bolsen et al. (Reference Bolsen, Druckman and Cook2014) view that partisan motivated reasoning is more likely to occur under some circumstances than others. For a public and lower-cost behavior such as masking, the gap between Democratic and Republican behavior grows in association with vaccination. This is the case both for the month-by-month analysis of the vaccine rollout and the DID analysis of vaccine status. By comparison, for higher cost and less publicly symbolic activities such as isolation and working remotely, the gap between Democrats and Republicans lessens in sync with the decreasing net personal benefits from engaging in the actions.

4. Partisanship, policy, and state government approval

Building on the above findings that suggest party cues do not pervasively overwhelm other information (e.g., Bullock, Reference Bullock, Suhay, Grofman and Trechsel2020), we exploit state-level and intertemporal variation in covid-related restrictions to study how partisan affiliation relates to an individual’s approval of their state government’s response. In all 50 states, emergency declarations enabled governors to quickly enact policies including stay-at-home orders and restrictions on gatherings. The governor was therefore the primary state policy actor. Furthermore, in many states, governors were the face of public communications, routinely giving press conferences and speeches (e.g., Doerr, Reference Doerr2021). We accordingly expect that if partisanship influences individual approval of a state government’s response, approval will be higher when an individual’s party matches that of the governor, regardless of the policy actions taken.

Additionally, we expect the relationship of approval to the policies enacted will vary by respondent partisanship, particularly given the earlier findings regarding the association of partisanship to health-related responses. Democratic respondents should be more likely than Republican ones to approve of the state response when the governor enacts covid-related restrictions and Republicans less likely to approve under these circumstances. Moreover, research on partisanship suggests these responses may depend on co-partisanship with the governor, with individuals more likely to favor policies endorsed by co-partisan elites (e.g., Lenz, Reference Lenz2012; Barber and Pope, Reference Barber and Pope2019).

To analyze these predictions, we focus on gatherings restrictions and stay-at-home orders (SAHOs), coding them identically as before. Recall that the Gallup data on approval of the state response extend only through April 2, 2020, and by that time no state had enacted a mask ordinance. However, many states had enacted gatherings restrictions and SAHOs, enabling a comparison pre and post the enactment of the policies. Figure S24 maps which states enacted each type of policy, along with the governor’s partisanship and whether the state voted for Trump or Clinton in 2016.Footnote ¹⁶ As the map shows, gubernatorial partisanship usually but not always matches state partisanship, and although most states enacted both types of restrictions, 13 enacted only gatherings restrictions, one only a stay-at-home order, and three neither.Footnote ¹⁷

For the main analyses, we estimate the following linear probability model for respondent i on date t, first for states with a Republican governor and then those with a Democratic one:

(2)\begin{equation}\begin{gathered} APPROV{E_{it}} = \gamma {P_{it}} + \xi \left( {{P_{it}} \times {R_{st}}} \right) + \kappa {R_{st}} + \zeta COVI{D_{ct}} + \phi {D_{it}} + {\eta _c} + {\lambda _t} + {\epsilon _{it}} \\ \end{gathered} \end{equation}

where the dependent variable is approval of the state response; $P$ again reflects the respondent’s party; and $\,R$ represents policy. The interaction between $R\,$and $P$ allows the relationship between approval and policy to vary by an individual’s partisan affiliation. The main effects of $P$ then capture the partisan gap in approval before any policy is enacted, and the main effects of R approval of the policy by the omitted partisan category. The previously described demographic controls are represented by $D$, and we also control for the local case count given evidence that Republican politicians suffered lower support as the local count rose (Warshaw et al., Reference Warshaw, Vavreck and Baxter-King2020). State-level fixed effects account for unobserved and time-invariant state factors not otherwise accounted for, and date fixed effects for national-level, intertemporal influences. Because of the shorter time series of the state approval survey item, we do not have individual-level panel data from before and after the policy actions for these analyses and therefore cannot include individual fixed effects.

We separate the analyses by party of governor given that it is constant during the timespan and to ease interpretation of the findings. Moreover, because of high collinearity between the interaction terms involving the two policies, we analyze them separately.

Table 2 reports the results. Columns 1 and 4 present the average baseline estimates of partisanship. In Column 1, for states with Republican governors, the negative coefficient on Democrats reflects that Democrats are on average less approving than Republicans, the omitted category, by approximately 25 percentage points. Similarly, for states with Democratic governors, the negative coefficient on Republicans in Column 4 suggests they are less approving than Democratic respondents by approximately 20 percentage points. In each case, the gap between Independents and co-partisans with the governor is substantially smaller than that for out-partisans, but a significant gap remains. All these results are in line with expectations.

Table 2. Individual approval of state response

Notes: Standard errors clustered by state below coefficients. Estimates on controls provided in supplemental materials (Table S26). ***p < 0.01, **p < 0.05, *p < 0.10, two-tailed.

The remaining columns consider how governors’ policy choices alter in- and out-party approval. Columns 2 and 3 present the estimates for Republican governors. Beginning with gatherings restrictions, the results suggest Democrats are substantially more approving after the policy enactment, consistent with a world in which approval of the government depends not only on partisanship but also the actions taken by officials. Specifically, gatherings restrictions are associated with higher approval of 19 percentage points by Democratic respondents. Republicans, on the other hand, neither reward nor punish a Republican governor for this policy action, as reflected by the main effect of the policy (given that Republicans are the omitted partisanship category). The partisan gap in approval is therefore significantly lower when the governor enacts the restriction. Supplemental figure S27 shows the marginal effects for these and other results in Table 2.

For stay-at-home requirements, none of the estimates for the policy enactment are significant for states with Republican governors. One possible reason is that in these states, almost all stay-at-home orders were enacted in conjunction with or after gatherings restrictions. We therefore attempted to analyze the first covid-related restriction (including workplace closings, school closings, and canceling of public events in addition to gatherings restrictions and SAHOs). However, the first policy is enacted so quickly that only 36 survey responses in states with Republican governors were taken before the first policy was enacted and only four responses in states with Democratic governors. What we can assess is whether the results on gatherings restrictions extend to cases in which they are not a first restriction, and Supplemental S28 shows they do.

Supplemental Table S28 also reports results regarding two additional considerations. First, it presents the results without the state fixed effects. In this case, the estimates on the main effects of the policies, representing how Republicans approve of the governor once the policies are enacted, are negative (p < 0.05 for SAHO and p < 0.1, two-tailed, for gatherings restrictions), consistent with a world where Republicans are more approving of governors who did not enact these restrictions. Second, we investigate whether the results are driven purely by teamsmanship, whereby Democrats are approving of gatherings restrictions because doing so reflects dissolution of the Republican party and rooting against Trump. Columns 4 and 5 of the supplemental table show the estimates for Democrats who did not consider avoiding gatherings versus those that did (see S28 for details). Interestingly, the results on gatherings restrictions for the former is negative and marginally significant (p < 0.1, two-tailed), further supporting the argument that out-party members are motivated in part by the governor’s policy actions in assessing government performance.

Moving onto states with Democratic governors, Table 2 suggests out-partisans (Republicans) are less approving of the state response after a SAHO is enacted (p < 0.1, two-tailed) and in-partisans more approving after a gatherings restriction is enacted. The results on SAHOs are consistent with those on gatherings restrictions for Republican governors in that out-partisans are more approving when the governor’s actions counter those of the national party. Also, there is not a significant coefficient on SAHOs for the subset of Republicans who report being very likely to isolate if public health officials recommended doing so (Supplemental S28 provides full details), and the results are robust to examining only states that did not enact SAHOs as a first policy.

Overall, Table 2 suggests that individual approval of the state’s COVID-19 response is conditioned by governors’ policy choices as well as individual partisanship. As expected, co-partisanship with the governor is significantly related to approval. Notably, however, there is evidence that out-partisan approval can increase when governors’ policies counter their national parties’ positions. These results are consistent with the perspective that party cues do not completely dominate other policy-related information (e.g., Bullock, Reference Bullock, Suhay, Grofman and Trechsel2020); although individual partisanship is a strong driver of government approval, partisan differences are tempered by the policies the government chooses.