Personal hardship narrows the partisan gap in COVID-19 and climate change responses

Extensive research in psychology focuses on how people update their beliefs when confronted with new information. One influential theory holds that people update their beliefs in proportion to the prediction error—the difference between prior expectations and present experience—allowing them to correct inaccurate expectations and adapt to a changing world (8). In these accounts, experience is generally taken to be veridical or objective, and new experience that deviates substantially from expectations results in a strong updating signal.

A separate literature in psychology has identified different “modes” by which individuals process information (9, 10). The analytic and affective modes shape how people evaluate information accessed by direct experience. The affective decision mode uses human associative processing (11), which operates quickly and automatically in response to threats, converting adverse experience into affective responses (e.g., fear, dread, anxiety), and thus represents risk as a feeling rather than as a probability (12, 13). The calculation-based decision mode uses analytic processes that, in contrast, operate more slowly and require conscious effort and control. Personal adverse experience may thus lead to the updating of expectations but also drive protective actions through worry and fear responses.

A substantial body of evidence points to the importance of personal experience with extreme weather events and local temperature abnormalities for climate concern (14–17), behavioral intentions, and policy support (18). However, there is also evidence that learning from experience depends on an individual’s prior beliefs in ways different from the rational updating process described above. People do not impartially detect changes in their local environments. Instead, contextual and cultural factors and motivational drivers shape how people seek (or avoid), attend to, process, interpret, remember, and report new information (19–21). People tend to process information in a way that favors desired conclusions or beliefs, especially when these are strongly held, leading them to accept confirmatory evidence at face value while ignoring or challenging conflicting evidence (22). Such processes can result in motivated reasoning, leading individuals to reject new information when it challenges their beliefs. They also shape how people understand and evaluate new information (23). In the context of climate change, extreme weather is only “evidence” for climate change if people believe that the two are causally related (24, 25). Stronger belief in climate change may make people more likely to look for and thus see evidence supporting it (5, 6).

Using a longitudinal dataset to distinguish between evidence-based revision of beliefs and belief-motivated perceptions, Myers et al. (3) found evidence for belief-motivated perceptions in people with strong ideological commitments to climate change and evidence-based revision of beliefs among the uncommitted middle. These results are directionally consistent with Bayesian updating (26), such that beliefs should change with disconfirming evidence quickly when initial beliefs are weak and more slowly when beliefs are deeply held (14). Using signal detection theory, Dryden et al. (27) showed that attribution of hurricanes to climate change depends on preexisting climate beliefs, which influence the threshold used to classify hurricanes as “normal” vs. “evidence of climate change.” Respondents who were less certain about the existence of climate change required more evidence before attributing an unusual hurricane season to climate change. Prior beliefs thus acted as a lens through which experience was interpreted rather than blocking it altogether (27).

When belief-motivated perceptions are shaped by salient partisan identities, this interpretive filter is called “partisan-motivated reasoning” (28). Partisan-motivated reasoning shapes public opinion on a variety of social issues (28–30), including climate change (23, 31). Political ideology shapes interpretations of local temperature abnormalities and perceptions of the scientific consensus on climate change (15, 32). Research with farmers found that their recall of local temperature conditions was biased by their preexisting beliefs about climate change and political party (33). Partisan identities can overtake stated policy preferences, with individuals acting against their beliefs or preferences when receiving conflicting information from a trusted partisan source (34). For example, endorsement of climate action by copartisan messengers increases responsiveness to climate change, especially among Republicans (35, 36).

The partisan channel provides one explanation for polarized public views on both COVID-19 and climate change, despite mounting evidence, including through direct exposure, of their adverse consequences. Party affiliation predicts social distancing in response to COVID-19, measured through both survey responses and geolocation data (37–40). Throughout 2020, Democrats and Republicans continued to be divided on the utility and importance of masks and other protective measures recommended by epidemiologists, decisions to close or reopen local economies, and priorities embedded in the various COVID-19 stimulus plans. Likewise, perceptions and actions related to climate change vary significantly by partisan affiliation and political ideology (5, 41, 42). Democrats are more inclined than Republicans to believe that anthropogenic climate change is occurring and are more worried about it (43).

Based on this literature, we hypothesize that there will be substantial partisan polarization on our key outcome measures for both COVID-19 and climate change—worry, protective behaviors, and policy support—and that messaging by copartisan elites will shape policy support, especially among Republicans who are otherwise more likely to be opposed to these measures.

“Who you gonna believe, me or your own eyes?” is an old line attributed to Chico Marx of the Marx Brothers that serves as a metaphor for our theorizing in this paper. When individuals are confronted with novel, complex, and threatening situations, how do their experience and partisan worldviews interact to shape beliefs and preferences? When the partisan and experience channels contradict one another, which gives?

A study of Phoenix, AZ, residents found that political ideology predicted perceptions of regional temperature change, but that objective temperature variation predicted perceptions of local temperature changes in one’s neighborhood (44). These results suggest that prior beliefs affect individuals’ interpretation of ambiguous, distal, or abstract information, while personal experience may supersede prior beliefs when events are proximal, concrete, and immediate. The emergence of and subsequent polarization around COVID-19 offers a unique opportunity to study these interacting dynamics in the context of a consequential and rapidly spreading threat. In this context, experience with the virus was concrete and immediate, and it could either confirm or disconfirm the worldviews put forth by a respondent’s self-reported political group.

Experience is a powerful teacher and can change attitudes and actions. This is especially true when threats are immediate, concrete, and local, as was the case for COVID-19 in the first half of 2020. In view of strong polarization in initial beliefs, we hypothesize that personal experience with COVID-19 will play different roles among Democrats and Republicans. For Republicans, experience might provide a strong error correction signal, leading Republicans to update their beliefs that COVID-19 is a serious public health threat. By contradicting their initial beliefs, adverse experience provides novel information that leads to greater updating of prior beliefs. For Democrats, adverse experience with COVID-19 essentially confirms partisan messages and priors about the severity of the issue. Since Democrats are already worried and more likely to endorse personal and policy actions, we expect the impact of experience on these outcomes to be relatively weak or even nonexistent. We thus hypothesize that Republicans will respond more strongly to direct personal experience with the virus than Democrats.

Nevertheless, where experiential and partisan signals contradict each other, we could observe the opposite. Motivated reasoning could lead Republicans to overlook negative experience with COVID-19, perceive it to be less adverse than a Democrat experiencing the same impacts, remember or report it differently, or attribute the impacts to causes that are consonant with their views. Recent work has identified partisan differences in accuracy and directional motivations in survey responses, such that partisans may respond along party lines despite dissenting personal beliefs or attitudes (45). In an exploratory analysis, we test for evidence of partisan-motivated reasoning in our measure of experience by assessing whether Republicans report less adverse experience with COVID-19 as a function of county-level incidence rates than Democrats, controlling for relevant demographic factors. Even with similar interpretations of experience, Republicans may require more evidence to update strong and contradictory priors relative to Democrats. These arguments, for different reasons, suggest mechanisms for the alternative hypotheses that experience would have little impact on Republicans or might widen the partisan gap. Our study design is unable to test a null hypothesis that experience has no effect; however, we calculate two-tailed P values to allow for the possibility of the opposite relationship in the interaction between partisanship and experience.

Personal experience with COVID-19 might also have broader impacts on perceptions of key institutions and actors involved in the crisis. Trust in government and related institutions is low in the United States right now (ref. 46, p. 104) and divided along partisan lines (SI Appendix, Fig. S8). In exploratory analyses, we investigate these broader effects of experience with COVID-19 by focusing on trust in the Centers for Disease Control and Prevention (CDC), news media, and scientists—actors who played prominent roles in COVID-19 response and communication. We conduct parallel analyses for climate change.

Our sample is longitudinal with replenishment. The first wave was composed of new participants, and each subsequent wave had new and repeat participants. The second and third waves were first populated by respondents from the previous wave and then replenished strategically to meet our quotas. The overall attrition rate (loss of repeat respondents) was 51% between wave 1 and wave 2 and 26% between wave 2 and wave 3.^† Our sampling strategy means that, by wave 3, we had five distinct types of participants: those who completed all three waves (n $=$ 1,739), those who completed only waves 1 and 2 (n $=$ 1,078), those who completed only waves 2 and 3 (n $=$ 747), those who completed waves 1 and 3 (n $=$ 7), and those who only participated in a single wave (n $=$ 5,149 across the three waves).

The survey included six modules—extreme weather, climate change, COVID-19, trust and community, political attitudes, and demographics (see SI Appendix for question text)—repeated with minimal modification in each wave. Participants took 25 min, on average, to complete the survey. They were not informed of the module categories. The order of the climate change and COVID-19 blocks was randomized across participants.

Most analyses in the main text refer to data from June and July 2020 (wave 2), the period in which COVID-19 had spread across the United States. At this time, vaccines were not yet available, and the presidential election was a few months away. This is also the survey wave with the largest number of questions on a variety of attitudes, personally protective behaviors, and policy preferences. In SI Appendix, we show trends over time in our key explanatory and outcome variables across all three survey waves and conduct within-subjects analyses for variables present in multiple waves.

Our primary outcome variables are described in detail in SI Appendix, Fig. S1 for COVID-19 and SI Appendix, Fig. S2 for climate change. For example, worry about COVID-19 was answered on a four-point scale (“not at all” to “extremely”) in response to the question, “How worried are you about the coronavirus epidemic?” Worry about climate change was solicited on the same response scale but with a broader temporal focus in line with the longer-term nature of the hazard: “How worried are you about current and future global warming/climate change?”^‡

Our primary explanatory variables are indices measuring a respondent’s self-reported negative experience of adversity with COVID-19, assessed by experience with natural hazards. Using multiple survey questions, we constructed separate indices of negative personal experience with 1) COVID-19 and 2) natural hazards. In each wave, we asked participants about the economic, health, psychological, and social impacts of COVID-19 and natural hazards that they had experienced. We summed responses to obtain indices of the total negative impacts of COVID-19 and natural hazards across various dimensions of an individual’s life. We asked respondents about experience with natural hazards over a 10-y period since these are low-frequency, high-impact events. We use the term “natural hazards” instead of “extreme weather events” because the list of hazards included earthquakes. We conduct additional analyses removing respondents who experienced earthquakes and report these in the SI Appendix. We took the natural log of the aggregate impact measures to account for their skewed distributions. Our primary COVID-19 experience index combines measures of health and economic impacts, which could have led to very different attitudes toward social distancing and closing businesses. Disaggregating these two dimensions does not substantially change our results (SI Appendix, Tables S38–S41).

In additional analyses, we also use county-level COVID-19 incidence rates and natural hazard risk and exposure as proxies for self-reported experience. For COVID-19 analyses, we used cumulative cases per county from the Johns Hopkins University COVID-19 Data Repository. We deviate from the preregistration by using cumulative COVID-19 case rates as opposed to new cases, since our personal negative experience index measures cumulative experience. We calculate cumulative cases up to each respondent’s survey response date, convert these into values per 100,000 population using county-level population estimates from the 2019 US Census Bureau, and take the natural log to account for the skewed distribution. For climate change analyses, we construct a county-level natural hazard risk index by summing the risk level (zero to three) for each of 11 hazards from the US National Center for Disaster Preparedness. This index captures hazard risk but not necessarily occurrence over the past 10 y. We thus identify a proxy for county-level natural hazard exposure using data from the Federal Emergency Management Agency (FEMA). We sum all natural hazard disaster declarations for the 10-y time frame in wave 2: 1 January 2010 to 15 June 2020. Disaster declarations provide a proxy for the impact of hazards on local populations without overemphasizing regions with more valuable properties or industries, as cost measures might. However, disaster declarations are a step removed from actual exposure and subject to partisan bias due to politicized processes of disaster relief distribution (47, 48). See SI Appendix for more information about the construction of the individual experience indices and county-level measures for both issues.

We also included embedded experiments in our survey. To examine respondents’ willingness to pay for different policies, we randomly assigned subjects to different costs associated with policies to help mitigate COVID-19 or compensate victims. To examine the role of partisan elites’ messaging, we randomly assigned the partisan identity of the proponent of policies related to emissions reductions or COVID-19 mask mandates.^§ See SI Appendix, Fig. S1 for question wording.

All analyses, unless otherwise indicated, are based on weighted data to adjust for any deviations from our desired quotas. Deviations were generally small but differ across survey waves, with wave 3 showing the largest deviations. We compute survey weights using the American National Election Studies (ANES) method (49) as implemented by the anesrake package in the R programming language (50).^¶

We report the results of multiple linear regressions. All regressions include the following sociodemographic control variables: gender; age; ethnicity; race; income; education; region; a categorical variable indicating whether the respondent lives in an urban, suburban, or rural area; and a dummy variable for repeat respondents. We deviate from the preregistration by adding the dummy variable for repeat respondents to the previously specified list of controls. We include this to account for unobserved differences between respondents who choose to participate in repeated online surveys and those who do not. We include partisan identification as either a primary independent variable (H1) or as a moderator (H2 to H4), depending on the analysis.

Cross-sectional (within wave) analyses are captured by the following specification using weighted least-squares regression:where D_it is the dependent variable of interest for respondent i in wave t, I_it is the primary explanatory variable of interest, P_it is a moderator where applicable, ${\mathbf{X}}_{it}$ is a vector of controls, and Ω_i is a dummy variable for respondents who were repeaters. Most of our models involve the interaction of negative experience with partisan identity (H2 to H4), where Democrats are the omitted category. For example, Table 1 reports the following specifications for t = 2:

For H2 to H4, we standardize dependent variables and include the same control variables in all regressions so coefficients (β) are comparable in terms of SDs of the dependent variable. Statistical significance is evaluated at the p < 0.05 level. We use the Benjamini and Hochberg (52) and Bonferroni methods to account for multiple comparisons within each outcome category when we estimate the variance on β₃ for $\mathit{NegEx}{p}_{it}P\mathit{artisa}{n}_{it}$ (SI Appendix, Tables S44–S47).

We include additional robustness analyses to evaluate endogeneity in our self-reported experience variables. These include 1) analyses using Eq. 1 with a dependent variable of self-reported experience and an interaction between county-level COVID-19 incidence and partisan identity, 2) the analyses described above using county-level measures of exposure in place of self-reported measures of experience, and 3) longitudinal panel data analyses of repeat respondents. We assess change over time within respondents by looking at the difference-in-differences of variables of interest in an analysis of our repeater group. These analyses are ordinary least-squares regressions with the same demographic controls, but the dependent and independent variables of interest are now the differences between t₂ and t₁, and we include control variables measured at t₁ and the value of the dependent variable at t₁ to control for baseline effects.

In H1, we hypothesize that our respondents are polarized along partisan lines across three categories of dependent variables: 1) worry, 2) behavior (i.e., self-reported mask wearing and a host of other protective behaviors), and 2) policy preferences (i.e., willingness to pay to mitigate pandemic risks). Fig. 1 shows respondents’ mean responses on these variables as a function of political party for wave 2.

Fig. 1, Left shows that worry is lower for Republicans than Democrats, resulting in a substantial partisan gap that persists across all three waves and is consistent with other studies (SI Appendix, Fig. S3 and ref. 53).^# Fig. 1, Middle shows that partisan differences are also present in a key behavioral response to the pandemic, with Democrats more likely to report wearing masks. Partisan gaps persist across waves (SI Appendix, Fig. S4) and the other self-reported protective behaviors we assess (SI Appendix, Figs. S4 and S5), consistent with research on social distancing compliance (39, 54, 55). Fig. 1, Right shows partisan differences in willingness to pay to support policies to reduce the likelihood of pandemics (see SI Appendix, Figs. S6, S7, and S9 for partisan gaps across the other COVID-19 policies we assess). Support is highest for policies incurring no or low costs to respondents, but Democrats are more willing to support the policy at all price levels than Republicans or Independents.

Polarization on these outcome variables is likely influenced by messaging from party leaders. In an embedded survey experiment, we find that respondents are more likely to support emissions reduction policies that are proposed by leaders of their own political parties as opposed to identical statements from other messengers (Fig. 2). This is especially true for Republicans.^‖

In summary, we find evidence of partisan polarization on worry, behaviors, and policy preferences, and that the affiliation of political messengers can sway public support for contentious policies when issues are polarized.

In keeping with H2, we find that hardship with COVID-19 is associated with increased worry for all groups, but especially for Republicans. At high levels of reported negative experience, there is no longer a partisan gap in mean levels of worry about COVID-19 in the spring and summer of 2020 (waves 1 and 2; Fig. 3).** As shown in Table 1 for wave 2, and SI Appendix, Table S11 for all waves, using the Likert scale outcome, we find a significant negative coefficient (P < 0.01) on the base term for Republicans (i.e., lower worry than Democrats under no negative experience) and a significant positive coefficient (P < 0.01) on the interaction term (i.e., a steeper slope of negative experience on worry for Republicans than for Democrats). The same relationships hold for perceived current and future harm to a respondent’s community (SI Appendix, Table S12).

We find similar results when assessing the role of experience on attitudes toward four personally protective actions (H3): vaccination intent, mask wearing, importance of self-isolating, and the retrospective view that one was too slow to self-isolate (SI Appendix, Table S13). We consistently find a significant negative coefficient on the base term for Republicans and a significant positive coefficient on the interaction term.^†† These results suggest a strong association between experience and personally protective actions among Republicans. Republicans with high levels of adverse experience report behaviors that are more similar to those of Democrats than to those of Republicans with low levels of experience, resulting in a narrowing of the partisan gap.

We again find similar results when looking at support across distinct COVID-19 policy responses (H4): a retrospective evaluation about whether one’s state was too slow to mandate social distancing, support for caution in reopening the economy, and support for contributions by the United States to national and global efforts to fight the pandemic (SI Appendix, Table S14). Support for these policies is not related to personal experience for Democrats, but there is a positive and statistically significant coefficient on the interaction term (i.e., the slope on experience for Republicans vs. Democrats).^‡‡

Adverse experience is associated with more support for policies aimed at mitigating the risks or impacts of the virus, even when the policies incur personal costs. Fig. 4 shows that negative experience is strongly associated with the willingness to pay for a policy to mitigate COVID-19 for both Democrats and Republicans.^§§

With the caveat that the attribution of extreme weather events to climate change involves greater uncertainty for both experts and the general public (6), we briefly report the results of parallel analyses for climate change.

We find a substantial partisan gap on worry about climate change that stays constant across all three waves (SI Appendix, Fig. S15). In contrast to the COVID-19 results, we do not see a general decline in worry about climate change across waves. We also find that the county-level FEMA-based exposure measure is a positive and significant predictor of our negative experience index (SI Appendix, Table S53); this relationship is not statistically significant for county-level hazard risk (SI Appendix, Table S21).^## In keeping with our COVID-19 findings, partisan affiliation does not moderate the relationship between county-level exposure and self-reported experience with natural hazards. Like COVID-19, we expect that the effect of county-level natural hazard risk or exposure on responses also depends on the vulnerability of the respondent’s social networks, media discussion of hazards and variation in disaster response and preparedness policies that are only partially captured by our demographic controls.

Adverse experience with natural hazards is associated with a narrowing of the partisan gap on climate change worry across all three waves (SI Appendix, Fig. S16). The large substantive difference may be due to the low baseline worry about climate change among Republicans with no experience, which is lower than baseline worry about COVID-19. We also find significant positive relationships between reported experience and perceptions of current and future harm from climate change among Republicans (SI Appendix, Table S22).

When looking at comparable measures of individual behaviors, policy support, and trust in key actors, we find that the role of adverse experience parallels our COVID-19 findings in several ways. Among Republicans, adverse personal experience is positively and significantly associated with an increase in the perceived importance of taking individual action to mitigate climate change. We find significant positive effects of experience on support for national and international efforts to mitigate climate change (SI Appendix, Table S23) and trust in key actors (Environmental Protection Agency [EPA], news media, and scientists; SI Appendix, Table S24). Greater reported adverse experience among Republicans is associated with a narrowing of the partisan gap on these measures.