Introduction
Mitigating the COVID-19 pandemic requires individuals to comply with public health directives such as wearing a mask (Howard et al. Reference Howard, Huang, Li, Tufekci, Zdimal, van der Westhuizen, von Delft, Price, Fridman, Tang, Tang, Watson, Bax, Shaikh, Questier, Hernandez, Chu, Ramirez and Rimoin2021). Unfortunately, mask-wearing has become a partisan issue. Both former President Donald Trump and conservative media outlets have downplayed the need for and efficacy of various pandemic mitigation measures, including mask-wearing (Calvillo et al. Reference Calvillo, Ross, Garcia, Smelter and Rutchick2020; Yamey and Gonsalves Reference Yamey and Gonsalves2020). As a result, Republicans (Ingram et al. Reference Ingram, Chuquichamb, Jimenez-Leal and Olivera-LaRoa2020; Milosh et al. Reference Milosh, Painter, Van Dijcke and WrightN.d.) and conservatives (Utych Reference Utych2021) are considerably less likely to wear masks than Democrats and liberals and often resist messages encouraging them to do so (Gelfand et al. Reference Gelfand, Li, Stamkou, Pieper, Denison, Fernandez, Choi, Chatman, Jackson and Dimant2022).
In this environment, how can experts persuade the public to wear masks more often? One promising approach is to alert the public to descriptive norms among an important reference group (Bilancini et al. Reference Bilancini, Boncinelli, Capraro, Celadin and Di Paolo2020; Fisher and Karl Reference Fisher and Karl2022; Folmer et al. Reference Folmer, Kuiper, Olthuis, Kooistra, de Bruijn, Brownlee, Fine and van Rooij2020, Reference Folmer, Brownlee, Fine, Kuiper, Olthuis, Kooistra, de Bruijn and van Rooij2021; Kitamura and Yamada Reference Kitamura and Yamada2020; Kooistra and van Rooij Reference Kooistra and van Rooij2020). People who know that others in the reference group are engaging in pandemic mitigation behaviors might be more likely to adopt more pandemic mitigation practices themselves. This intervention could be especially effective if people underestimate the strength of a behavioral norm due to highly visible dissent and non-compliance (e.g., protests against mask-wearing and media coverage of the issue).
This article experimentally tests whether an intervention that alerts Americans to empirical data on the strength of behavioral norms of mask-wearing on downstream mask-wearing intentions, perceived mask effectiveness, and affective polarization.
Our results indicate that telling people most Americans report wearing masks increased self-reported intentions to wear masks and improved the perceived efficacy of masks relative to controls, though these effects were not statistically distinguishable from other treatment conditions. The marginal effects on mask-wearing were significant among Republicans, the partisan group with the lowest levels of mask-wearing intentions, and among participants who did not overestimate rates of self-reported mask-wearing. Finally, despite partisan controversy over the issue, learning about reported mask-wearing behavior among co-partisans or out-partisans did not differentially affect mask-wearing intentions.
Theoretical approach
Descriptive norms are observed patterns of actions that people take in a particular setting. These norms can have strong effects on behavior because people want to be accepted in social contexts and follow relevant behavioral expectations, which they can generally accomplish by acting like everyone else (Christensen et al. Reference Christensen, Rothgerber, Wood and Matz2004; Cialdini, Kallgren and Reno Reference Cialdini, Kallgren and Reno1991).
By the time our study was conducted, mask-wearing had become common. In a June 25–July 12, 2020 survey conducted about a month before the experiment reported here, 74% of respondents reported wearing a mask “all of the time” or “most of the time” when they left the house. Our first preregistered hypothesis predicted that making this norm salient by informing people how many Americans say they wear masks would increase mask-wearing intentions (H1A).
People may lack information about the strength of the mask-wearing norm due to, for example, the media emphasis on a minority of anti-mask activists, living in a community with below-average mask-wearing adherence, or people staying home and not observing public mask-wearing behaviors. We also predicted that the effect of the mask-wearing norm treatment would be particularly strong among people who previously underestimated reported mask-wearing among the public (H1B).
When descriptive norms differ between subgroups, as they do between Democrats and Republicans in pandemic mitigation, people may be especially responsive to in-group behavior (Goldring and Heiphetz Reference Goldring and Heiphetz2020). We thus expected that providing information about reported mask-wearing among co-partisans would increase mask-wearing intentions (H2A), particularly among those who previously underestimated mask-wearing among co-partisans (H2B).
To the extent Americans see opposing partisans’ motivations and behaviors as suspect (Munro, Weih and Tsai Reference Munro, Weih and Tsai2010; Waytz, Young and Ginges Reference Waytz, Young and Ginges2014), they may not wish to follow out-partisans’ general patterns of behavior or may even wish to resist out-partisan norms. Therefore, it remains an open question whether exposure to out-partisans’ reported mask-wearing tendencies influences individuals’ own intentions to wear masks (RQ1A). How partisans react to over- and under-estimating reported out-partisan mask-wearing is an open question (RQ1B) for the same reasons.
We also ask whether providing information about descriptive norms affects beliefs about the efficacy of masks (RQ2). Descriptive norms can influence behavior independently of underlying beliefs and attitudes about that behavior (Christensen et al. Reference Christensen, Rothgerber, Wood and Matz2004; Cialdini, Kallgren and Reno Reference Cialdini, Kallgren and Reno1991). Following descriptive norms may reflect an effort to preserve a sense of belonging with others regardless of agreement with the underlying reasons for a behavior (Göckeritz et al. Reference Göckeritz, Schultz, Rendón, Cialdini, Goldstein and Griskevicius2010). Alternatively, people who anticipate engaging in a behavior may adopt beliefs and attitudes to justify their actions and avoid cognitive dissonance (Festinger Reference Festinger1957). In this case, exposure to descriptive norms could prompt an update not only of behavioral intentions but also of corresponding beliefs.
Finally, though most Americans report engaging in pandemic mitigation behaviors such as wearing masks, Republicans do so less often (Druckman et al. Reference Druckman, Klar, Krupnikov, Levendusky and Ryan2020). Furthermore, Republican Party opinion leaders are more likely to criticize or disregard masks (Calvillo et al. Reference Calvillo, Ross, Garcia, Smelter and Rutchick2020; Yamey and Gonsalves Reference Yamey and Gonsalves2020). As a result, Republican mask-wearing intentions could potentially increase more than Democrats in response to the norms treatments. Alternatively, however, the norms interventions could backfire among Republicans because they conflict with messaging from co-partisan opinion leaders. We therefore explore whether the effect of learning descriptive norms varies by party (RQ3). Finally, we ask whether learning that the vast majority of Americans report wearing masks, including a majority of out-partisans, might highlight shared cross-party norms and thereby reduce affective polarization (RQ4; Gaertner et al. Reference Gaertner, Dovidio, Anastasio, Bachman and Rust1993; Levendusky Reference Levendusky2018).
Materials and methods
Experimental manipulation
We conducted this preregistered experiment in the third wave of a multi-wave panel study examining COVID-19 misperceptions in the USA (n = 2,982). Data were collected from July 28 to August 19, 2020 (see Appendix for additional details about the preregistration and the survey). At this point in the pandemic, the USA had experienced a summer surge in infections, with 7-day averages peaking at above 65,000 in mid-July and remaining above 45,000 during the data collection period (New York Times 2021). After initially discouraging widespread mask use at the pandemic’s outset, the US Centers for Disease Control (CDC) had changed its guidelines in April 2020, recommending mask-wearing in public locations. By mid-July, the CDC published a study suggesting masking curbed the transmission of COVID-19 and strengthened its recommendation to wear masks (Netburn Reference Netburn2021).
Participants in this study were randomly assigned either to a control condition that did not receive information about mask-wearing norms or to one of three treatment conditions. The American norms treatment condition informed respondents of the percentage of Americans (74%) who reported wearing masks “all of the time” or “most of the time.” The Democratic norms treatment reported the figure for self-identified Democrats and Democratic-leaning independents (89%) and the Republican norms treatment reported the figure for self-identified Republicans and Republican-leaning independents (56%).Footnote 1 Each statistic provided was calculated using data from a question administered during Wave 2 of the survey panel in the general population sample. We recoded the Democratic and Republican norms conditions into co-partisan and out-partisan norms treatments (true independents who do not lean toward a party were excluded from analysis).
It is important to note that the reference group (Americans, Democrats, or Republicans) and reported levels of mask-wearing (74%, 89%, or 56%) simultaneously varied by condition in this design. Reporting the same level of mask-wearing in each reference group would have simplified interpretation but would have been inconsistent with our commitment to provide accurate information to experimental participants and our goal of testing descriptive norm effects under real-world conditions. We explore this issue further in the discussion section below.
Outcome measures
After the randomization, respondents indicated their intentions to wear masks, perceptions of the effectiveness of masks, and attitudes toward voters from each political party:
-
Behavioral intention to wear masks: “In the future, how often will you do the following? Wear a mask in public” (five-point scale from “Not at all” to “All of the time”).
-
Perceptions of mask effectiveness: “Please indicate whether you believe the following statement is accurate or not. Masks are an effective way to prevent the spread of coronavirus” (four-point scale from “Not at all accurate” to “Very accurate”).
-
Affective polarization: Difference between in-party and out-party feeling thermometers for “[p]eople who support Democrats” and “[p]eople who support Republicans” (0–100 scales).
Analytic strategy
We tested our primary hypotheses and research questions using linear regression with HC2 robust standard errors. We used a lasso variable selection procedure to determine the set of prognostic covariates to include in models for each dependent variable (see Appendix for additional details). To limit the risk of false positives (Benjamin et al. Reference Benjamin, Berger, Johannesson, Nosek, Wagenmakers, Berk, Bollen, Brembs, Brown, Camerer, Cesarini, Chambers, Clyde, Cook, De Boeck, Dienes, Dreber, Easwaran, Efferson, Fehr, Fidler, Field, Forster, George, Gonzalez, Goodman, Green, Green, Greenwald, Hadfield, Hedges, Held, Ho, Hoijtink, Hruschka, Imai, Imbens, Ioannidis, Jeon, Jones, Kirchler, Laibso, List, Little, Lupia, Machery, Maxwell, McCarthy, Moore, Morgan, Munafó, Nakagawa, Nyhan, Parker, Pericchi, Perugini, Rouder, Rousseau, Savalei, Schönbrodt, Sellke, Sinclair, Tingley, Zandt, Vazire, Watts, Winship, Wolpert, Xie, Young, Zinman and Johnson2018), we conduct significance tests using the p < 0.005 and p < 0.01 thresholds in addition to p < 0.05 (see Appendix) and report the most stringent standard at which we can reject the null hypothesis in the text. We correspondingly report 95% and 99.5% confidence intervals for treatment effect estimates in each figure.
To assess the precision of any null results we observe for main effects, we report equivalence bounds using a two one-sided tests approach (Lakens, Scheel and Isager Reference Lakens, Scheel and Isager2018). When we observe null estimates of heterogeneous treatment effects, we instead report the coefficient value of the relevant interaction term that our model and sample size can detect with 80% power. These estimates were obtained using simulations in DeclareDesign assuming the sample size and standard deviation of the residuals observed in our each model (Blair et al. Reference Blair, Cooper, Coppock and Humphreys2019).
Preregistration and survey instruments
Our preregistered hypotheses, research questions, and analysis plan for this experiment are available at https://osf.io/wyb2e. The survey instrument is available at https://osf.io/248af/. We did not deviate from the preregistration but did conduct additional analyses, which are reported below and labeled as exploratory. All analyses not labeled exploratory are preregistered.
Results
Descriptive results
Table 1 reports descriptive results for respondents’ reported levels of mask-wearing, perceived levels of mask-wearing among others, intention to wear a mask in the future, perceptions of mask effectiveness, and levels of affective polarization from our July 28 to August 19, 2020 survey wave (the third wave in a multi-wave panel). In this wave, 79.5% of our total sample reported wearing masks most or all the time, including 93.1% of Democrats and 60.1% of Republicans. We define under- and over-estimation of mask-wearing intention as being more than 10 percentage points under or over the percentages of regular mask-wearing featured in our treatments.Footnote 2 Respondents are much more likely to underestimate reported mask-wearing than overestimate it (48.1% and 8.7%, respectively), although each partisan group underestimates mask-wearing among its opponents far more than among co-partisans. Both Democrats and Republican perceive masks to be effective, though perceived effectiveness is greater among Democrats (mean of 3.8 versus 3.0 for Republicans on a four-point scale). Finally, partisans in our sample are highly affectively polarized. On average, they rate their own party 56.5 points above the other party on 0–100 point feeling thermometers.
Descriptive Statistics
NOTES: Mask-wearing intention is measured on a five-point scale. Perceived mask effectiveness is measured on a four-point scale. Affective polarization is measured as the difference between the 100-point feeling thermometer ratings of supporters of the respondent’s preferred party and the other party. Standard deviations are provided in parentheses.
Experimental results
Our first hypothesis predicted that the American descriptive norms treatment would increase mask-wearing intentions (H1A). The estimated marginal effects of each treatment on mask-wearing intentions are presented in Figure 1 (see Table A1 for corresponding ordinary least squares (OLS) results).Footnote 3 Consistent with H1A, exposure to the true percentage of Americans who report wearing masks regularly increases mask-wearing intentions among partisans by 0.140 points on a five-point scale (d = 0.137, p < 0.005).Footnote 4 We also predicted that the co-partisan descriptive norms treatment would increase mask-wearing intentions (H2A). However, exposure to the true percentage of co-partisans who report wearing masks has no measurable effect on mask-wearing intentions (p > 0.05). We estimate equivalence bounds of (−0.135, 0.054) using two one-sided tests—in other words, our results allow us to rule out effects less than −0.135 or greater than 0.054 on a five-point scale. Finally, per RQ1A, exposure to out-partisan mask-wearing had no significant effect on mask-wearing intentions either (p > 0.05; equivalence bounds estimated using two one-sided tests: [−0.178, 0.013]).
Effect of norm treatments on mask-wearing intentions.
Covariate-adjusted average treatment effects of norm treatments (including 95% and 99.5% confidence intervals) on mask-wearing intentions among partisans. See Table A1 for corresponding OLS results and Table A2 for estimates of the American norms treatment effect among the full sample including true independents.
In an exploratory analysis, we tested whether these treatment effects differed significantly from one another. Our results indicate that the American descriptive norms treatment had a greater positive effect on mask-wearing intentions than the co-partisan norms treatment (p < 0.05), but we could not reject the null of no difference between the American and out-partisan norms treatments and the co-partisan and out-partisan norms treatments (p > 0.05 for both; equivalence bounds versus out-partisan norms estimated using two one-sided tests: American [−0.164, 0.017], co-partisan [−0.050, 0.134]).Footnote 5
Figure 2 next disaggregates the marginal effects of each treatment by whether respondents underestimated, overestimated, or accurately perceived mask-wearing norms among the reference group corresponding to their treatment condition. We hypothesized that the treatment effect would be especially strong among respondents who underestimated the percentage of Americans who report wearing masks (H1B). Marginal effects are depicted in the left panel of Figure 2 (see Table A3 for corresponding OLS results).Footnote 6 However, the estimated impact of the American norms treatment was not statistically discernible between those who underestimated mask norms and those who accurately estimated them (p > 0.05).Footnote 7 In this case, the marginal effect of the American norms treatment on mask-wearing intentions was positive and statistically significant among both those who underestimate reported mask-wearing among the public and those who perceive it accurately (p < 0.05 for those with accurate beliefs, p < 0.005 for underestimators).
Effect of norms treatments on mask-wearing intentions.
Covariate-adjusted average treatment effects of norm treatments (including 95% and 99.5% confidence intervals) on mask-wearing intentions among partisans. See Table A3 for corresponding OLS results and Table A4 for estimates of the American norms treatment effect among the full sample including true independents.
We hypothesized that the co-partisan descriptive norms treatment would be especially effective among those who underestimated reported mask-wearing among co-partisans (H2B). However, as shown in Table A3, we find no evidence to support this hypothesis. The marginal effect of the treatment among underestimators is not measurably different from zero and we cannot reject the null hypothesis of no difference in treatment effects compared to respondents who accurately perceived reported co-partisan mask-wearing behavior (p > 0.05).Footnote 8
We asked whether providing information about the prevalence of mask use among opposition partisans affects behavioral intentions to wear masks compared to the control group (RQ1A). Marginal effects are depicted in Figure 1 (see Table A1 for corresponding OLS results). We find that exposure to the true percentage of opposing partisans who report wearing masks has no significant effect on mask-wearing intentions (p > 0.05; equivalence bounds estimated using two one-sided tests: [−0.178, 0.013]). We further asked whether providing information about the prevalence of mask use among opposing partisans differed by prior beliefs about the prevalence of mask use among opposing partisans (RQ1B). Marginal effects are depicted in the third panel of Figure 2 and Table A3 in the Appendix. We find that the treatment did not affect mask-wearing intentions among respondents who had accurate perceptions of opposing partisan mask-wearing. Moreover, the out-partisan × underestimated interaction is not significant (p > 0.05).Footnote 9
We asked whether providing information about the prevalence of mask use among Americans, co-partisans, or opposing partisans affects belief in the efficacy of masks compared to respondents in the control group (RQ2). Treatment effect estimates by condition are depicted in Figure 3 (see Table A5 for corresponding OLS results). We find that the American descriptive norms treatment increased belief in the efficacy of masks among partisans by 0.085 points on a 4-point scale (p < 0.05). However, this finding is sensitive to the presence of covariates when we estimate results both among partisans and for the full sample (p > 0.05 without covariates, p < 0.05 with covariates; see Tables A5 and A6). Moreover, neither the co-partisan descriptive norms treatment nor the out-partisan descriptive norms treatment affect the perceived effectiveness of masks among partisans relative to the control (p > 0.05 for each; equivalence bounds estimated using two one-sided tests: co-partisan [−0.089, 0.072], out-partisan [−0.057, 0.103]).
Effect of norms treatments on perceptions of mask effectiveness.
Covariate-adjusted average treatment effects of norm treatments (including 95% and 99.5% confidence intervals) on mask-wearing intentions among partisans. See Table A5 for corresponding OLS results and Table A6 for estimates of the American norms treatment effect among the full sample including true independents.
We asked whether the effect of each norms treatment differs by party (RQ3). Estimated treatment effects by condition and party are depicted in Figure 4 (see Table A7 for OLS results). We begin examining treatment effects on intent to wear masks. We cannot reject the null of no difference in treatment effects by party—none of the interaction terms are statistically significant (p > 0.05 for each).Footnote 10 However, the marginal effect of the American norms treatment is only significant among Republicans. Among Republican Party identifiers and leaners, the treatment increases mask-wearing intentions (d = 0.184, p < 0.05), whereas it does not change mask-wearing intentions among Democrats, who potentially face a ceiling effect given higher baseline levels of mask-wearing intention (see left panel of Figure 4).Footnote 11
Mask-wearing intentions and norm treatment effects by party.
Left panel presents reported mask-wearing intentions by party. Right panel presents covariate-adjusted average treatment effects of norm treatments (including 95% and 99.5% confidence intervals) on mask-wearing intentions by party. See Table A7 for corresponding OLS results.
We also find no treatment × party interactions on the perceived effectiveness of masks (p < 0.05 for each; see Figure A1 and Table A7 in Online Appendix A).Footnote 12 In this case, though, the marginal effects are null for both Democrats and Republicans.
Finally, we also preregistered research questions exploring whether the descriptive norms treatments would change affective polarization and whether the descriptive norms treatments were moderated by exposure to fact-checks of false claims about COVID-19 in a separate experiment fielded in the same study. Results, which are presented in the Appendix in Tables A8 and A9, respectively, show that the norms treatments did not measurably change partisan affect and fact-check exposure significantly reduced the effect of the American norms treatments in only one of eight estimated models.
Discussion
Can treatments strengthening descriptive norms increase mask-wearing to help mitigate the spread of COVID-19? We find that alerting people to the fact that the vast majority of Americans report wearing masks regularly is broadly effective in increasing mask-wearing intentions. These marginal effects were significant among Republicans (but not Democrats) and those who previously underestimated or accurately perceived reported mask-wearing rates in the relevant reference group. Finally, learning that most Americans report wearing masks regularly also improves assessments of the effectiveness of masks. These results contribute both to our understanding of the effects of descriptive norms and to social science research investigating COVID-19 behaviors and attitudes.
Importantly, these results suggest that the importance of partisan reference groups in shaping COVID-19 behavior may have been overstated. Learning about rates of self-reported mask-wearing among Americans increased mask-wearing intentions in general and among Republicans. By contrast, learning about the mask-wearing habits of one’s co-partisans or opposing partisans has no measurable effect on mask-wearing intentions. However, as we note above, the effects of learning what Americans do are not always statistically discernible from the effects of learning what co-partisans or out-partisans do. Our findings therefore warrant further investigation.
These findings come with another important caveat. In our design, the strength of the co-partisan treatment was limited by moderate levels of reported mask-wearing among Republicans (56% versus 74% for all Americans and 89% for Democrats). We chose to use actual data from a previous survey in the treatments given our goal of testing messages that could potentially be deployed in the real world. However, this decision creates a confound between the group featured in the treatment and the percentage engaging in the descriptive norm behavior (mask-wearing). Future research should seek to isolate in-group effects by keeping the percentage of people who wear masks in treatments fixed across conditions.
Several other important limitations of this research should be acknowledged. First, the severity of COVID-19 spread, the behaviors that people engage in to protect themselves, and perceptions of other people’s behavior have varied widely over the course of the pandemic. Future research should seek to replicate these findings under differing conditions. Second, we are unable to measure actual mask-wearing behavior given our reliance on surveys; as in all survey research on health behavior, our results thus depend on the imperfect correspondence between these intentions and the actions take in the real world. The descriptive norm information provided to respondents in our treatments may be less salient than more visible or concrete messages about behaviors in reference groups.
Nonetheless, these findings provide important evidence that even a limited descriptive norms intervention can change mask-wearing intentions during a global pandemic. Despite the deep divides over the response to COVID-19, Americans are sensitive to how other people act and change their intended behaviors accordingly.
Supplementary material
To view supplementary material for this article, please visit https://doi.org/10.1017/XPS.2022.20
Data availability statement
The data, code, and any additional materials required to replicate all analyses in this article are available at the Journal of Experimental Political Science Dataverse within the Harvard Dataverse Network at https://doi.org/10.7910/DVN/PJ27BB (Carey et al. Reference Carey, Nyhan, Phillips and Reifler2022).
Acknowledgements
We thank the National Science Foundation (grant number 2028485) and the Economic and Social Research Council (grant number ES/V004883/1) for funding support. All conclusions and any errors are our own.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was funded by the National Science Foundation (grant number 2028485) and the Economic and Social Research Council (grant number ES/V004883/1).
Conflicts of interest
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Ethics statement
This research complies with all relevant ethical regulations and with APSA’s Principles and Guidance for Human Subjects Research. The study was approved by the human ethics review board at Dartmouth College (STUDY00032068). The University of Exeter recognized the approved protocol for the North American study. In particular, informed consent was obtained from all participants. At the conclusion of the study, respondents were referred to the Centers for Disease Control and Prevention for more information about COVID-19. The survey vendor, YouGov, compensated participants with reward points that can be redeemed for cash.
Open access
