key: cord-0857150-6476xzps authors: Duch, R. M.; Barnett, A.; Filipek, M.; Roope, L.; Violato, M.; Clarke, P. title: Cash versus Lotteries: COVID-19 Vaccine Incentives Experiment date: 2021-07-28 journal: nan DOI: 10.1101/2021.07.26.21250865 sha: 3ef264e76939c71e9e018525afd922af2c1c6c5a doc_id: 857150 cord_uid: 6476xzps Governments are considering financial incentives to increase vaccine uptake to end the COVID-19 pandemic. Incentives being offered include cash-equivalents such as vouchers or being entered into lotteries. Our experiment involved random assignment of 1,628 unvaccinated participants in the United States to one of three 45 second informational videos promoting vaccination with messages about: (a) health benefits of COVID-19 vaccines (control); (b) being entered into lotteries; or (c) receiving cash equivalent vouchers. After seeing the control health information video, 16% of individuals wanted information on where to get vaccinated. This compared with 14% of those assigned to the lottery video (odds ratio of 0.82 relative to control: 95% credible interval 0.57-1.17) and 22% of those assigned to the cash voucher video (odds ratio of 1.53 relative to control: 95% credible interval 1.11-2.11). These results support greater use of cash vouchers to promote COVID-19 vaccine uptake and do not support the use of lottery incentives. to the behavioural change toolkit" [14] . Evidence from recent surveys indicates that financial incentives may play a role in increasing rates of COVID-19 vaccination. Financial incentives, rather than information and appeals to the common good (or even personal advantage), convinced experimental subjects to subscribe to a COVID-19 contact tracing app [30] . A Kaiser Family Foundation study found that about one in four would be more likely to get vaccinated if they were entered into a lottery with a chance to win one million dollars [19] . Initial evidence from a German conjoint experiment suggests that a hypothetical financial incentive of 50 Euros, as part of a mass vaccination scenario, could increase vaccine uptake among the hesitant [20] . Similarly, a third of the unvaccinated respondents in a U.C.L.A. COVID-19 Health and Politics Project survey experiment indicated cash payments would make them more likely to get vaccinated [44] . Many U.S. state governments have adopted a variety of cash and lottery incentives in order to encourage citizens to get a COVID-19 vaccine. The National Governors Association has published a description of the various COVID-19 vaccines available in the different U.S. states [32] . While these schemes have garnered considerable enthusiasm amongst policy makers, a number of reservations have been raised regarding their morality, efficacy and possible negative consequences [45, 23, 21] . A recent study of vaccination rates before and after the implementation of lottery incentives in Ohio did not find significant increases in adult COVID-19 vaccination rates [46] . This has led some [13] to argue for abandoning lotteries in favour of other approaches including cash vouchers, which have been shown to be more effective than lottery incentives for other preventative behaviours such as chlamydia screening [33] . However, prominent scholars, such as Richard Thaler, have argued that lotteries are an effective way to increase vaccination rates [39, 34] . There is an evidence gap with respect to incentives. While we have abundant experimental clinical evidence on the effectiveness of COVID-19 vaccines, surprisingly, we have limited experimental evidence informing policies designed to encourage COVID-19 vaccine uptake. This study conducts a randomized experiment to inform the policy debate regarding the efficacy of cash and lotteries as incentives for individuals to get a COVID-19 vaccination. Vaccine incentive policies have varied greatly across U.S. states, which facilitated the design of both the informational treatment videos and the outcome measure [24] . We implemented a randomized experiment to determine which of three messaging strategies is most effective for encouraging the non-vaccinated to seek information on how to get vaccinated: a standard CDC-inspired message 3 that identifies the personal and public health benefits of COVID-19 vaccination; a message that highlights the chance of winning a lottery if vaccinated; and finally a message that indicates that first-time vaccinated would receive a cash voucher. Our approach was based on the assumption that a basic public health statement stressing the importance, efficacy and safety of COVID-19 vaccines would occur in any real-world information provision. We therefore tested the effects of additional pieces of information that highlighted the availability of cash and lottery incentives for being vaccinated. The online sample of just over 1,500 non-vaccinated subjects were randomly assigned (1:1:1) to the informational treatment and control videos. After viewing the assigned video, participants were given the opportunity to consult additional information on being vaccinated in their state. We treat this digital expression of interest as our outcome variable. We had two primary preregistered hypotheses: 1) Participants in Video Treatment 2 (lottery message) would be more likely to click through to the vaccination information web page than participants assigned to the Treatment 1 control group (standard CDC-inspired health message); 2) Participants in Video Treatment 3 (cash voucher message) would be more likely to click through to the vaccination information web page than participants assigned to the Treatment 1 control group (standard CDC-inspired Health message). Our secondary hypothesis was that there would be no difference in click through rates between Treatment 2 and Treatment 3. 1 Figure 1 describes the details of the sampling and treatment assignment. A total of 3,416 online participants were recruited to participate in the CANDOUR Incentive study (3, 2 In our pre-registration calculations we established a sample size objective of 500 respondents in each of the three treatment groups. Due to the adaptive nature of the random assignment algorithm we recruited a total of 8% more subjects than anticipated. survey is available in the Online Supplemental Materials). Participants were then randomly assigned to one of the three video messages with the pre-registered objective of approximately 500 participants per video. We implemented sequential covariate-adaptive randomization. For each incoming respondent we calculate the Mahalanobis distance for four key covariates to ensure the random assignment resulted in balance across the three treatment videos [29] : race, gender, age and education. As each new participant entered the online experiment we adjusted the treatment assignment probabilities to ensure balance. 3 . The three information treatments were delivered in a short 45 second video. Treatment 1 (Control): Standard CDC-inspired COVID-19 vaccine promotional and information video. Treatment 2: Lottery treatment -the first 30 seconds are identical to the control video -the last 15 seconds inform viewers that in some states you can be entered into a lottery and win over 1 million dollars if you get vaccinated. Treatment 3: Cash voucher treatment -the first 30 seconds are identical to the control video -the last 15 seconds inform viewers that in some states you can earn money or a money equivalent of up to $100 that can be spent online or in stores. 4 Our treatment design simulates real world decision making situations in which videos play an increasingly important role in delivering information related to public policy [18, 6, 15, 31] . We pre-specified a single primary outcome measure -the digital expression of further interest in vaccination information. After subjects viewed the assigned treatment video, they were provided with a choice between 1) ending the survey or 2) obtaining further information about how they can get vaccinated in their state. The outcome measure is their response to this simple choice. If they clicked on the link for further information, they were directed to this webpage, that we prepared, and which points to vaccination information sources in each state: http: //www.didyougetcovidvaccine.com. 5 This approach builds on online experimental designs that treat digital traces as outcome measures [8, 35, 40, 16] . Their principal advantage is that individuals are not asked to give opinions which can exaggerate experimenter demand effects [36, 31] . Table 1 summarizes the distribution of the covariates for each of the three treatment groups. In addition to age, gender, race and education, we asked state of residence. The percentage of each state's vote that was won by President Donald Trump in the 2020 presidential race is employed to generate a state Trump vote variable [26] . The state Trump variable scales the state Trump vote share by subtracting the mean percentage and dividing this by 10 percent (State % Trump vote (+10%)). We employ a scaled age variable in the analysis -subtracting 40 from the observed age and dividing the result by 10 (Age (+10 years)). Generally, relative to the U.S. population, the online convenience sample is younger, more white and disproportionately female. The distribution of these covariates within each of the three treatment groups suggests that the non-vaccinated subjects were randomly allocated across the three treatments. 4 Treatment 1 (control) video: https://youtu.be/V3DUCC8xnD0. Treatment 2 video: https://youtu.be/ QIm_vbpe_go. Treatment 3 video: https://youtu.be/gF045EaPj-o. 5 The screen shots of this text are available in the Online Supplemental Material. 7 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 28, 2021. ; https://doi.org/10.1101/2021.07.26.21250865 doi: medRxiv preprint Cash incentives, not lotteries, appeal to the non-vaccinated. Table 1 presents the click through rates for participants assigned to the three treatments. Of the 516 participants assigned to the control, i.e. the CDC-inspired health video, 81 (16%) subsequently expressed interest in more vaccine information. A slightly lower 14% of those assigned to the lottery video subsequently clicked on the link for further information. The treatment odds ratio for this lottery video is 0.82 (95% credible interval of 0.57 to 1.17, p-value = 0.27). Hence, there is no compelling evidence that lottery incentives perform any better than a standard health message. A much higher 22% of participants assigned to the cash voucher video subsequently clicked through to the further information page. With an odds ratio of 1.53 (95% credible interval of 1.11 to 2.11, p-value = 0.009), the odds of these participants expressing interest in information is about 1.5 times greater than the odds of the control subjects. We estimate a Bayesian logistic regression model with a dichotomous dependent variable indicating whether the respondent clicked through for more vaccine information. Table 3 in Methods presents the results for a basic specification with the Cash and Lottery treatment variables (Model 1); a Model 2 that includes the full set of covariates; and a Model 3 that adds random effects for States and participant pools. Treatment have an estimated odds ratio with credible confidence intervals that encompass 1providing little evidence that the lottery had any benefit. Note also that the credible intervals on the Cash Voucher and Lottery odds ratios just barely intersect -reasonable evidence that Cash rather than Lotteries appeals to the non-vaccinated. Estimated odds ratios for the covariate adjustments, from Model 3 in Table 3 in Methods, are also presented in Figure 2 . Two of the covariate adjustments are statistically significant. Whites have an odds ratio of 0.68 (95% credible interval of 0.47 to 0.97) indicating they were less likely than Blacks to click through for more information after the video treatment. Age has an odds ratio of 1.19 per 10 year increase (95% credible interval of 1.07 to 1.32) indicating older participants were more likely to click through for more vaccine information. 8 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; https://doi.org/10.1101/2021.07.26.21250865 doi: medRxiv preprint Heterogeneity We specified heterogeneous treatment effects for three covariates: gender, race and education. The specifications are presented in Equations 1 to 4 (Methods). The interaction terms in these models indicate whether exposure to one of the two treatments (Cash or Lottery) induces an increase in vaccination information interest, relative to the control in that same subgroup. Hence, in the case of education, we present the relative effect of changing treatment within the same education level: High Education and Cash vs High Education and Control; Medium Education and Cash vs Medium Education and Control; Low Education and Cash versus Low Education and Control. In Tables 4 to 6 (Methods), we present the estimated odds ratios with 95% credible intervals and Bayesian p-values. In Figure 3 , we present these estimated odds ratios along with their credible intervals for Education, Gender and Race. The set of three Education and Cash Voucher odds ratios are similar to the overall Cash odds ratio of 1.5. The Education and Lottery odds ratios are consistent with our overall findings -the Lottery treatment effect for the most part is indistinguishable from the control. There is some evidence in Figure 3 that the Lottery treatment effect for the High Education group is significantly lower than the control treatment for the highly educated. There is a gender interaction. For female participants, the Cash Voucher odds ratio with respect to the control is approximately 1.8 and quite precisely estimated, while for male participants the odds ratio is close one. The race interactions suggest interesting differences between Whites and Blacks. For Whites, the Cash Voucher odds ratio is approximately 1.5 and precisely estimated (they are the dominant racial group in the sample). The Lottery odds ratio for Whites is less than 1.0 and also precisely estimated. Relative to the control Health video, Whites are clearly more likely to click through for more information when they view the Cash Voucher information treatment and less is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; https://doi.org/10.1101/2021.07.26.21250865 doi: medRxiv preprint vote share responded to the Cash and Lottery treatments similarly to those in states with an average Trump 2020 vote share. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; for cash prizes and for products including cars and even guns [42] . Our U.S. experiment suggests that incentives to be vaccinated can be effective, but that the type of incentive matters -at least in the U.S., cash is effective while lotteries are not. Achieving greater compliance with the COVID-19 vaccination campaigns is an urgent public health challenge. Experimental research that builds on this simple design can inform policy efforts attempting to advance this goal. Vaccine incentives may, or may not, be an appropriate policy tool in many other countries facing this COVID-19 vaccination challenge. Our experimental design offers a relatively efficient approach to answering this question -random assignment of brief videos describing incentives on offer compared to standard health messaging. Moreover, this simple design can be extended to provide much richer insights into policies that promote vaccination uptake. An obvious design enhancement would be to vary the size of the incentive payments to determine a dose response relationship. This has been observed for incentives for other preventative health behaviours [5] ; but also in survey experiments concerning cash payments and COVID-19 vaccine intentions [44] . Our results also suggest that there maybe some heterogeneity in the treatment effects of both cash and lotteries in different sub-groups of the population. For example, there is a significant difference between cash and lotteries among whites, but not blacks. Conducting further experiments with a sample size large enough to detect differences in sub-groups where vaccine uptake is currently at lower level would allow us to assess the cost-effectiveness and equity impacts of different financial incentives. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; https://doi.org/10.1101/2021.07.26.21250865 doi: medRxiv preprint Our experimental design focuses very narrowly on cash versus lottery financial incentives for COVID-19 vaccinations. There is evidence that a range of other factors affect vaccine uptake, including the convenience of access to vaccination and the personal freedoms associated with proof of vaccination [20] . Some governments, in fact, are imposing increasingly negative incentives (such as the recent announcement of the French government to mandate proof of vaccination for visiting restaurants and cafes [10] ). A more informative, although challenging, experimental design would allow us to identify the causal impact of a full range of policy incentives, both financial and non-financial, on COVID-19 vaccination decisions. A potential limitation of our study -as with most vaccine uptake studies -is that we do not observe the participants' vaccine decisions. Our outcome measure is their decision to seek out additional vaccination information. An extension of our design that linked information treatments to actual vaccination decisions would be a more powerful design. The experimental results reported in this article provide some guidance for incentive policies. First, we provide evidence that COVID-19 vaccination messaging that highlights financial incentives will likely have a bigger motivational impact on the non-vaccinated than is the case for standard COVID-19 messaging that focuses solely on the health benefits. Second, the non-vaccinated are significantly more likely to be motivated by messaging that highlights cash voucher incentives than they are by lottery incentives. In fact, lottery messaging is no more effective than standard CDC health messaging. 14 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. Research Prime Panels has a large and diverse pool of online respondents [9] . Additional individuals were also recruited from online Facebook advertisements (297) [48] . The Lucid Fulcrum Exchange provided 364 participants -this is an online participant pool that matches the characteristics and attitudes of other online pools such as MTurk [11] . The survey experiment began 15 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. The three information treatments were delivered in a short video: . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; The Introduction is common across all three treatment arms. All subjects get common health information about COVID-19 vaccines. In some states, those who have at least one COVID-19 dose will automatically be entered into a lottery. Some of these lotteries will pay winners over $1 million in cash. If you get vaccinated, you could get rewarded. If you are interested in more information about getting a COVID-19 vaccine shot in your state please click on the link that will appear once this video ends. The US is working hard to distribute the COVID-19 vaccines, free for everyone with no strings attached. COVID-19 vaccines are safe and effective. After you've been fully vaccinated, you can resume activities that you did prior to the pandemic. In some states, those who have had at least on COVID-19 vaccine shot will be given cash or a voucher worth between 50 and 100 dollars that can be spent online or in a wide range of stores. If you get vaccinated in those states you will get rewarded. If you are interested in more information about getting a COVID-19 vaccine shot in your state please click on the link that will appear once this video ends. 17 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; https://doi.org/10.1101/2021.07.26.21250865 doi: medRxiv preprint As indicated in the main text, participants were randomly assigned to one of the three video treatments with the pre-registered objective of approximately 500 participants per treatment. We implemented sequential covariate-adaptive randomization. For each incoming respondent we calculate the Mahalanobis distance for four key covariates to ensure the random assignment resulted in balance across the three treatment videos [29] : race, gender, age and education. As each new participant entered the online experiment, we adjusted the treatment assignment probabilities to ensure balance. Assignment probabilities change based on the covariate profiles of units already in the three treatment groups. In practice, this procedure seeks to create similar covariate distributions in the treatment groups by biasing the current unit's treatment assignment in favor of treatment conditions with fewer units with similar profiles. This approach is especially adapted to online experiments such as ours in which we do not know the precise characteristics of respondents who will make up the final sample [43] . At any point in time during the survey period, we only have information on participants who have already taken the survey and the subject that has just arrived. We implement the sequential covariate adaptive randomization using the seqblock algorithm in the blocktools R package [29, 28, 27, 7] . The adaptive algorithm calculates the Mahalanobis distances and ranks them to bias the randomization toward treatment groups with high scores (the value of Mahalanobis distance). We use the Qualtrics Web Service feature to run this co-variate adaptive randomization remotely using an API. 6 Balance. Table 1 in the text summarized the covariate distributions within each of the three treatment groups. We supplement this by modelling treatment assignment as a function of our five covariates: gender, race, education, age, and State % Trump Vote High (>50%). Table 2 presents the results from a multinomial logit regression with the three treatment assignments (CDC Health, Lottery and Cash Voucher) as the values for the dependent variable. These results simply confirm our earlier observations regarding the covariate balance in treatment assignments we observed in Table 1 . As Table 2 indicates, all of the estimated credible intervals encompass zero. This suggests that there is no strong evidence of treatment assignment being correlated with any of our covariates. 6 Diag Davenport describes how to deploy the API and embed it into a Qualtric's survey https:// diagdavenport.com 18 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; We summarize the treatment effects using a Bayesian logistic regression model with clickthrough as the dependent variable. We used a multiple variable model by including the treatment and potential predictors of click-through (age, gender, race, education and Trump vote). We also included random effects for each state and the three sample pools. We used vague priors for all parameters. In the heterogeneity analyses we included interactions between the treatment effect and the four key variables of: gender, race, education and state % Trump vote. We present the estimates as odds ratios with 95% credible intervals and Bayesian p-values. The Bayesian p-values estimate the probability that the odds ratio is equal to one. We compared the model fit after adding additional covariates to the logistic regression model using the Deviance Information Criterion (DIC) [41] . A smaller DIC indicates a better fit and a difference between models of 10 or more is a strong indicator of a better model. The Bayesian model was fitted in JAGS version 4.3.0 [37] . We used two chains each with 5,000 samples thinned by 5 with a burn-in of 5,000. We visually checked the convergence and mixing of the two chains. The odds ratios with 95% credible intervals for the three models are presented in Table 3 . The Bayesian p-values for Model 3 are presented in the last column of Table 3. 19 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; Heterogeneity Figure 3 in the main text presents the interaction effects for Gender, Race and Table 4, Table 5 , Table 6 , and Table 7 . is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. All four models include the random effects of state and pool defined using Normal distributions with vague gamma priors for the inverse-variance: State j ∼ Normal(0, σ 2 s ), j = 1, . . . , 50, . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 28, 2021. ; . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 28, 2021. ; https://doi.org/10.1101/2021.07.26.21250865 doi: medRxiv preprint 23 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 28, 2021. ; https://doi.org/10.1101/2021.07.26.21250865 doi: medRxiv preprint 24 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 28, 2021. ; https://doi.org/10.1101/2021.07.26.21250865 doi: medRxiv preprint Challenges in creating herd immunity to SARS-CoV-2 infection by mass vaccination Messaging Interventions that Increase COVID-19 Vaccine Willingness in Latin America The Shot, the Message, and the Messenger: COVID-19 Vaccine Acceptance in Latin America The false promise of herd immunity for COVID-19 Effect of Different Financial Incentive Structures on Promoting Physical Activity Among Adults: A Randomized Clinical Trial Fact-checking" videos reduce belief in but not the sharing of "fake news" on Twitter Implementing Blocked Randomization in Online Survey Experiments Measuring public opinion via digital footprints Online panels in social science research: Expanding sampling methods beyond Mechanical Turk France mandates Covid health pass for restaurants and cafes Validating the demographic, political, psychological, and experimental results obtained from a new source of online survey respondents Do Voters Update when Audit Reports are Informative Out of luck -COVID-19 vaccination lotteries don't work The Effectiveness of Financial Incentives for Health Behaviour Change: Systematic Review and Meta-Analysis Does Counter-Attitudinal Information Cause Backlash? Results from Three Large Survey Experiments Less than you think: Prevalence and predictors of fake news dissemination on Facebook Promoting Completion of Health Risk Assessments with Lottery Incentives The Upside of Accents: Language, Inter-group Difference, and Attitudes toward Immigration Kaiser Family Foundation. 2021 What incentives can spur Covid-19 vaccination uptake Problems With Paying People to Be Vaccinated Against COVID-19 A global survey of potential acceptance of a COVID-19 vaccine Why Paying People to be Vaccinated Could Back Fire Adaptive diversification of COVID-19 policy Unmasking Partisanship: Polarization Undermines Public Response to Collective Risk Block, Assign, and Diagnose Potential Interference in Randomized Experiments Multivariate Continuous Blocking to Improve Political Science Experiments Blocking for Sequential Political Experiments Tracking and promoting the usage of a COVID-19 contact tracing app Improving Experimental Treatments in Political Science National Governors Association Vouchers versus Lotteries: What works best in promoting Chlamydia screening? A cluster randomised controlled trial Vaccine Lotteries and Beyond Shifting attention to accuracy can reduce misinformation online Beyond Attitudes: Incorporating Measures of Behavior in Survey Experiments JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling Authors Triggering altruism increases the willingness to get vaccinated against COVID-19 Wait, Vaccine Lotteries Actually Work Bit by Bit: Social Research in the Digital Age The deviance information criterion: 12 years on Donuts, drugs, booze, and guns: what governments are offering people to take covid-19 vaccines Randomisation doesn't guarantee similarity of groups; minimisation does $100 as Incentive to Get a Shot? Experiment Suggests it can Pay Off Incentives for Immunity -Strategies for Increasing Covid-19 Vaccine Uptake Lottery-Based Incentive in Ohio and COVID-19 Vaccination Rates Vaccination as a social contract: The case of COVID-19 and US political partisanship Quota sampling using Facebook advertisements