Will I get COVID-19? Partisanship, Social Media Frames, and Perceptions of Health Risk in Brazil *

Abstract

In these polarized times, not even perceptions of personal risk are immune to partisan considerations. We report results of a COVID-19 social media framing experiment with positive and negative partisan messages from high-level politicians. Descriptive results show that pro-government and opposition partisans report very different expectations of health and job risks. Job and health policy have become wedge issues that elicit partisan responses.

Full text

Will I get COVID-19?
Partisanship, Social Media Frames, and Perceptions
of Health Risk in Brazil ∗
Ernesto Calvo †Tiago Ventura ‡
November 21, 2020
Forthcoming at Latin American Politics and Society
Abstract
In these polarized and challenging times, not even perceptions of personal risk are im-
mune to partisanship. This paper introduces results from a new survey with an embedded
social media experiment conducted during the first months of the COVID-19 pandemic in
Brazil. Descriptive results show that pro-government and opposition partisans report very
different expectations of health and job risks. Job and health policy have become wedge
issues that elicit partisan responses. We exploit random variation in the survey recruitment
to show the effects of the President’s first speech on national TV on perceived risk and the
moderating effect of partisanship. We present a framing experiment that models key cogni-
tive mechanisms driving partisan differences in perceptions of health risks and job security
during the COVID-19 crisis.
∗This research is part of the Inter-American Development Bank project: “Transparency, trust, and Social Media”, 1300600-01-PEC. PI:
Ernesto Calvo, 2019-2020. We thank Elizabeth Zechmeister, Noam Lupu, and Maita Schade from LAPOP, who coordinated the probabilistic
selection of respondents from a Netquest panel of Brazilian voters. We also thank Julia Rubio, who contributed to the survey design and
Sandra Ley and Francisco Cantu, who are collaborators of the three surveys fielded in this project. We received invaluable feedback from
the members of the interdisciplinary Lab for Computational Social Science (iLCSS-UMD) as well as important suggestions from Fernando
Guarnieri, Antoine Banks, Natalia Aruguete, Maria Victoria Murillo, Isabella Alcaniz, Mario Pecheny, Mariana Carvalho, Amaro Grassi,
Lorena Barberia, Guilherme Russo, Jonathan Phillips and Marisa Van B¨
ulow. Replication materials for the article are available at: https:
//github.com/TiagoVentura/Calvo_Ventura_LAPS_2021
†University of Maryland, Government and Politics, UMD. Address: 3140 Tydings Hall, College Park, MD 20742, USA. Email:
ecalvo@umd.edu. Webpage: http://gvptsites.umd.edu/calvo/
‡University of Maryland, Government and Politics, UMD. Address: 4118 Chiconteague, College Park, MD 20742, USA. Email: ventu-
rat@umd.edu. Webpage: http://tiagoventura.rbind.io/
1

1 Introduction
Will partisan messages alter the voters’ perceived risk of contracting COVID-19? Will voters
internalize elite messages and align perceived risks with the policy preferences of their parties?
Since the seminal studies on framing and risks by behavioral economists Daniel Kahneman and
Amos Tversky(1982), researchers have documented framing effects in the subjective assessments
of risk and in policy preferences.1Frames can induce myopic responses when the messages
emphasize potential gains or losses, which are weighted differently by voters (Thaler et al.,1997;
Iyengar,1990). Frames may also alter perceptions of risk by increasing the salience and memory
accessibility of features of an event (Kahneman,2011). Accordingly, as polarization increases,
scholars have documented distinctive partisan responses that align with changes in perceptions
of risks (Iyengar and Westwood,2015;Green et al.,2004) and in perceived trust in political facts
and scientific evidence (Nisbet et al.,2015;Bullock et al.,2013;Kraft et al.,2015).
Political and public health responses to the COVID-19 pandemic provide significant anec-
dotal evidence of the effects of partisanship on risk perceptions, risky behavior, and policy
responses. Populists leaders such as Jair Bolsonaro in Brazil, Donald Trump in the United
States, and Manuel Lopez Obrador in M´exico, publicly challenged scientific recommendations
and the adoption of strict sanitary measures. Following these cues, government supporters in
all three countries publicly challenged and actively mobilized against social distancing rules, the
use of masks, and other measures that would limit the propagation of the virus. Nevertheless,
the extent to which partisan messages are associated with changes in subjective perceptions of
risk is less clear.
Understanding the effect of competing partisan frames on perceived (or subjective) risk is
1For an illuminating reading of this problem from behavioral economics see Kahneman (2011).

critical to manage the COVID-19 pandemic successfully. Since the early days of the pandemic,
social distancing became the most important public health response. Compliance with social
distancing measures, however, requires voters to accept the individual and collective risks that
may affect them personally. Accordingly, a successful health response needs to evaluate how
political beliefs affect perceived risks and interact with policy implementation (Kushner Gadarian
et al.,2020;Allcott et al.,2020;Barrios and Hochber,2020;Mariani et al.,2020;Ajzenman et al.,
2020).
Our research brings new and timely survey data2to understand subjective perceptions of
risks during the COVID-19 pandemic. We analyze how perceptions of risks vary among party
supporters, how sensitive voters are to information shocks, and how they react to social media
frames. To this end, we first introduce descriptive evidence of partisan3differences in perceived
risk. We show that supporters of the Bolsonaro administration in Brazil report lower subjec-
tive levels of job and health risks along with greater support for the government’s response
to the COVID-19 pandemic. The results are robust to several control variables and to model
specification.
Second, we take advantage of random variation in the recruitment of survey respondents and
model the effect of the first national address on COVID-19 by President Jair Bolsonaro. Using
a difference-in-difference design with respondents interviewed in the two days before/after the
2Our survey field started on March 23 and went on until May 8. The first official death due to COVID-19
in Brazil occurred only a week before, on March 17. Our timely survey collects a snapshot of citizens’ reactions
during the first months of the pandemic.
3Partisanship in Brazil is not a term that can be used without a proper discussion. In our survey, we collect
three different measures of party affinity: (a) vote intention “if the election is next week”, (b) self-reported
partisanship, and (c) self-reported anti-partisanship (Samuels and Zucco,2018). All three of the questions were
thoroughly tested. In the paper, we report results using (a) and (c). As a young democracy with a large menu
of parties, researchers have shown partisanship to be a weak predictor of voters’ attitudes and preferences. More
recent work by David Samuels and Cesar Zucco (2018), argues that partisanship is better captured by the pro and
anti-partisan feelings against the Workers Party. Most parties, other than the PT, failed to built strong labels
and score very low on party identification responses in survey data. In this article, we consider “vote intention”
as the best alternative to describe party affinity. However, models with the alternative variables are reported in
the SIF and/or can be requested from the authors.
2

speech, we find robust evidence of partisan updates of risk perceptions. Our results show that
among opposition voters, perceptions of job and health risk increased after Bolsonaro’s speech
compared to independents, while no changes are perceived among government’s partisans.
We conclude with an experimental design, with an IRB-approved and preregistered instru-
ment,4to detect the effect of social media frames on perceived health and job risks. Our exper-
iment exposes respondents to high-level politicians’ positive and negative social media messages
about COVID-19, and measures sharing and emotional responses, as well as their effects on
perceived risk.
Overall, we find critical partisan differences in risk perceptions. We also find a significant
uptake on risk perceptions after Jair Bolsonaro’s public speech. However, evidence of framing
effects from social media messages in our experiment are modest.5Our results align with similar
findings in the United States, raising questions about the level of sensitivity of the experimental
treatments (Kushner Gadarian et al.,2020).
To understand the modest effect of our social media frames on subjective risk, we conduct a
statistical autopsy of our experiment, unpacking the behavioral responses to the experimental
frames. The analysis reveals positive effects for the mediation mechanism (“anger”) on perceived
health risk and in lower support for the government. However, “angry” responses to social
media messages are not consistently higher for publications by out-group politicians – in-group
polarizing messages also elicited similar reactions. Therefore, while the mediation mechanism
elicited the expected response, the different frames did not.
As in the difference-in-difference analysis of Bolsonaro’s speech, “anger” was more readily
reported by partisans and increased perceived risks among independents. While our experimental
4Our pre-registration and pre-analysis plan, available here: https://osf.io/c67m3
5Only one of our four pre-registered hypotheses is confirmed by the data
3

design expected frames to increase partisan anger and anger to increase risk, f rames →anger →
risk, our findings only validate the effect of anger on perceived risk, frames 6→ anger →risk.
By troubleshooting our experiment, we are able to pinpoint the mediating factors that increase
perceived risks.
The results of our study have important public policy implications. Current studies in the
US and Brazil have shown that districts with high voter support for Trump and Bolsonaro
respectively, have steeper epidemiological curves for the COVID-19 spread (Ajzenman et al.,
2020;Mariani et al.,2020;Allcott et al.,2020). Our research shows that this is consistent with
government messages that made COVID-19 a wedge partisan issue. At a time when perceived
health risks is critical to manage the COVID-19 pandemic successfully, findings of this article
should be of interest to health policy and political communication experts.
The organization of this article is the following: First, we introduce the Brazilian case, how the
government has reacted to the COVID-19 pandemic, and partisan dynamics in the country. In
section three, we present descriptive evidence of partisan differences in government performance
assessments, perceptions of job security, and perceptions of health risks. The next section
presents evidence from the difference-in-difference models describing Bolsonaro’s speech’s effect
during our survey collection process. In section five, we introduce the hypothesis and survey
instruments, testing for the effect of negative and positive social media frames on perceptions of
risk. In section six, we describe our experimental findings. The last section discusses the paper’s
overall contribution for our understanding of how partisanship affects risk perceptions during
the COVID-19 pandemic.
4

2 Brazilian Populism, out and about
In the first weeks of January, 2020, news about the rapid spread of COVID-19 in the Hubei
province of China circulated around the world. As Chinese authorities quarantined millions
of citizens, governments worldwide struggled to assess the potential domestic damage of the
virus and identify the proper health emergency protocols to halt its spread. Timid responses in
February of 2020, both in Europe and the United States, included travel and trade restrictions
both to and from the affected areas. On March 11, 2020, the World Health Organization declared
the rapidly-spreading COVID-19 virus a pandemic, likely to affect every country on the globe.
While some governments promptly adopted social distancing protocols to mitigate the conse-
quences of the pandemic, leaders in a few countries resisted calls for swift action. The President
of the United States, Donald Trump; the President of Mexico, Lopez Obrador; and the president
of Brazil, Jair Bolsonaro, asked their citizens to dismiss the threat. Among these three leaders,
Bolsonaro’s response serves as a textbook example of a defiant, unflinching, and vocal challenge
to the scientific recommendations to address the crisis. As community spread of COVID-19
was confirmed in major cities of Brazil, Bolsonaro asked citizens to maintain their regular work
schedule and prop up the economy. On the offensive, he criticized the media for their ”hysteri-
cal” reporting on the virus and accused the political opposition of using COVID-19 for political
gain. As he actively impaired Brazil’s own federal agencies, Bolsonaro urged mayors and state
governors to roll back stay-at-home orders and, repeatedly, defied calls for social distancing. He
promoted meetings and local gatherings, walked the streets to defy stay-at-home orders, and
used his social media account and the bully pulpit of his office to dismiss the health consequences
of the virus.
Bolsonaro’s supporters were equally vocal, sharing his social media posts, echoing his business-
5

as-usual demeanor, defying stay-in-place orders, and minimizing the health risks of the crisis. In
contrast, the opposition, the media, and most health professionals criticized the President for po-
larizing messages that failed to respond to the health crisis’s challenges. Anti-Bolsonaro activists
pushed back against the President’s message, circulating their own distinct health messages.
As a young democracy with an extensive and fragmented menu of parties, researchers con-
sidered that partisanship in Brazil is a weak predictor of voters’ attitudes and preferences.
The Brazilian party system was frequently described as weakly institutionalized (Mainwaring,
1991,1999;Mainwaring and Scully,1995), with candidate-centered incentives driving politicians’
electoral behavior (Samuels,2003;Ames,2001). Recent studies have begun to challenge some
of these preconceptions, confirming that partisan and anti-partisan sentiments affect candidate
evaluation and policy preferences (Samuels and Zucco,2018;Power and Rodrigues-Silveira,2018;
Baker et al.,2016). Our findings bring further support to these views, with partisan preferences
having measurable effects on perceptions of job and health risk during the COVID-19 pandemic.
3 Partisanship and perceived risk to COVID-19
As in the United States, partisan assessments of health and job risks are markedly different.
Figure 1vividly portrays differences in perceived risks by supporters of President Bolsonaro
and supporters of the opposition’s candidate Fernando Haddad.6For our outcome variables, we
consider three main questions. These questions capture perceptions about personal risk during
the COVID-19 pandemic, and the respondents’ assessments about the government’s performance
during the crises 7
6We consider respondents as Bolsonaro supporters, Haddad supporters, or independents depending on their
reported voting preference for our question: If the runoff presidential election “were to take place next week” who
would you vote for? We provided respondents with three possible choices: Jair Bolsonaro, Fernando Haddad, and
Other/Blank/Null. We considered independents the respondents who selected the latter option. Results are robust
to other specifications either using respondents’ voting preference in the first round, or positive partisanship.
7The wording of all three questions is presented below:
6

A total of 29% and 23% of respondents who support Haddad consider it very likely that
they will lose their jobs or become infected with COVID-19. By contrast, Bolsonaro supporters
reported a much lower probability, 22% and 12%, respectively. The differences are even more
salient when reporting their evaluation of the government’s response to the crisis, resulting in
20 percentage points of difference between supporters of the government and of the opposition
that consider the government response very appropriate. Measures of positive and negative
partisanship towards the Workers Party, (Samuels and Zucco,2018), yield broader differences
on risk assessments, with 33% of pro-PT supporters losing their job and 25% reporting being
very likely to become infected by COVID-19, compared to 22% and 14% for anti-PT respondents.
•Question 1: How likely is it that your health would be affected by COVID-19? (very likely, somewhat
likely, somewhat unlikely, very unlikely)
•Question 2: Given the current health and economic crisis produced by the Coronavirus COVID-19, how
likely is it that you could lose your job? (very likely, somewhat likely, somewhat unlikely, very unlikely)
•Question 3: Has the government response been appropriate when faced with the Coronavirus COVID-19?
(Very appropriate, somewhat appropriate, somewhat unappropriated, very inappropriate).
7

Figure 1 Survey assessments of the quality of the Government response, perceptions of personal health risk, and perceptions of personal job
security, March 23 through May 4, 2020.
8

Figure 2 Regression Estimates for Partisan Effects on Risk Perceptions and Government Assessment
during the Covid-19
a) Partisan Effects
b) Negative Partisanship
9

We also present results from linear models regressing the three outcome variables on partisan
preferences and a set of socio-demographic variables such as age, income, education, occupation
in the labor market, and gender. Our regression estimates using both the voter choice for the last
presidential election and positive and negative partisanship towards the Workers Party render
similar results. These results hold when the models are estimates controlling by age, gender,
income, occupation, and education of the respondents. Figure 2presents the results.
Descriptive evidence is overwhelming, with significant inter-party differences in perceptions
of risk and government response assessment. In Table 4 of the appendix file, we report the effect
of the controls. Controls for the models show that employed and highly educated respondents
report lower perceived job risks and higher health risks than unemployed and less educated
respondents. Also, as age increases, perceptions of job and health risk increase. In particu-
lar, older voters see a considerably larger increase in their perceived likelihood of losing their
job. By contrast, there are no statistically significant differences in assessments of government
performance and age. Full results are presented in Section B, Table 4 of the SIF file.
4 Beyond Description: Modeling the effect of Bolsonaro’s Speech
Descriptive results show dramatic partisan differences in reported health and job risks, as well
as in subjective assessments of the government responses. In this section, we take advantage
of a public speech by Bolsonaro during data collection and causally identify changes in the
respondents’ risk perceptions due Bolsonaro’s discourses denying the COVID-19 pandemic.8
Bolsonaro, in both social media posts and his public appearances, urged local authorities
8At the time we pre-registered our experimental design, we could not anticipate that our ongoing survey
recruitment would allow us to perform the difference-in-difference analysis measuring the effects of the Bolsonaro’s
speech on risk perceptions. Therefore, our empirical analysis and theoretical expectations for this section were
not pre-registered.
10

to prioritize growth, challenged (and fired) his Minister of Health, and minimized the potential
health risks of the pandemic. On March 24, President Bolsonaro gave one of his more widely
publicized, and dismissive, messages on the COVID-19 crisis and on his administration’s re-
sponse. In a nationally televised address to the country, which was also his first presidential
speech dedicated solely to the COVID-19 pandemic, Bolsonaro displayed this confrontational
tone. Contrary to most pundits’ beliefs that he would moderate his attacks and hedge his polit-
ical bets, the President accused governors of overreacting, challenged social distancing policies,
criticized schools closures, described himself as an athlete who would “not even notice” if he got
infected, and labelled the virus, in the worst case, as just a little flu.
We make use of the granularity of our survey data over time to model the effect of Bolsonaro’s
dismissive behavior about the COVID-19 pandemic during its first days in Brazil. Modeling this
event, at the beginning of the pandemic, allows us to measure risk perceptions when the number
of cases was still modest. Our survey field started on March 23, allowing us to collect a small part
of our sample two days before the Presidential Pronouncement. As before, we focus the analysis
on the differential effects among partisans and non-partisans of the President. To identify the
effects, we use a differences-in-differences approach on a narrow window of days before and after
the event, described by the following estimation:
yit =αi+β1·Haddad +β2·Independents +β3·P ost −March −24+
τ·Haddad ∗P ost −March −24 + β4·I ndependents ∗P ost −M arch −24 + it
(1)
where yit is the survey responses on risk perceptions and assessments of government responses,
and the partisan variables come as answers to who the respondent would be likely to vote if
elections were to be held in the following week. To make our sample before and after more
11

comparable, we limit the analyzes for the time window between 23 and 26 of March 9. Our
parameter of interest is τ, which measures the differences in the outcomes comparing Bolsonaro
voters (depicted by the intercept in equation 10) and Haddad supporters.
The effect of Bolsonaro’s speech on perceptions of risk
Table 1presents our results. The first three (restricted) models use no control variables, while
the remaining three control for the respondents’ age, gender, occupation, education, and income.
Among Haddad’s supporters, perceptions of job and health risk increased after Bolsonaro’s
speech compared to government supporters. The estimates for Health Risk are statistically
significant at p<.05, while the effects for job risk are statistically significant at p<.1. More
interestingly, results show that Haddad voters did no change their overall assessment of the
government’s performance. By contrast, we observe a small decline of-0.441 in evaluations of
the government performance among pro-government voters, significant at p<0.1. The models
that include all controls provide substantively similar, although slightly stronger, statistical
results.
The findings provide support for the effect of contextual partisan events on perceptions
of risk. Related research has found robust evidence that Bolsonaro’s denial about COVID-
19 increases the spread of the disease and reduced levels of compliance to social distance in
pro-government localities (Ajzenman et al.,2020;Mariani et al.,2020). Our results provide
a behavioral explanation for these shocking findings; as the President sends dismissive signals
about the pandemic risks, although risk perceptions overall increase, his supporters do not report
the same concerns as the rest of the population. Significantly, partisans of the opposition increase
9Such decision reduces the chance our estimate is capturing some omitted factor varying over time. In such a
small time interval, it is unlikely something else has affected perceptions of risk about the COVID other than the
Presidential speech
12

their risk perceptions, while government supporters keep their business as usual, decreasing social
distancing policies’ effectiveness and facilitating the spread of the disease.
13

Table 1 The Effects of Bolsonaro’s Presidential Pronouncement of March 24 on Risk Assessments
Dependent variable:
Job Risk Health Risk Government Assessment Job Risk Health Risk Government Assessment
(1) (2) (3) (4) (5) (6)
Intercept 2.062∗∗∗ 2.538∗∗∗ 3.091∗∗∗ 1.811∗∗∗ 2.316∗∗∗ 2.427∗∗∗
(0.132) (0.102) (0.115) (0.393) (0.303) (0.348)
Post-March 23 −0.362 −0.338 −0.441∗−0.393 −0.378∗−0.331
(0.272) (0.210) (0.238) (0.278) (0.215) (0.246)
Haddad Voters 0.524∗∗ 0.242 −1.310∗∗∗ 0.589∗∗ 0.121 −1.166∗∗∗
(0.212) (0.164) (0.186) (0.234) (0.181) (0.207)
Independent Voters 0.127 −0.048 −0.600∗∗∗ 0.129 −0.165 −0.535∗∗∗
(0.197) (0.152) (0.172) (0.214) (0.166) (0.190)
Post-March 23 x Haddad Voters 0.867∗0.740∗∗ 0.297 0.799∗0.967∗∗∗ 0.246
(0.452) (0.350) (0.397) (0.468) (0.361) (0.414)
Post-March 23 x Independent Voters 0.273 0.248 0.050 0.253 0.304 −0.119
(0.390) (0.301) (0.342) (0.398) (0.307) (0.353)
Controls No No No Yes Yes Yes
Observations 210 210 211 195 195 195
Adjusted R20.062 0.042 0.220 0.089 0.058 0.226
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
14

Up to this point, our paper has shown robust descriptive evidence for partisanship moderating
risk perceptions in Brazil. We identify strong partisan differences on risk perceptions in Brazil,
and a direct effect of the Bolsonaro’s speech denying the severity of the COVID-19 on in-group
risk updates. Next, we use an online experiment to discuss how partisanship interacts with
framing in the context of social media positive and negative messages about the pandemic.
5 Framing and Risk Perceptions during the COVID-19 Pandemic
Following Entman (1993), we define framing as the act of selecting “some aspects of a per-
ceived reality and mak[ing] them more salient in communicating text, in such a way as to promote
a particular problem definition, causal interpretation, moral evaluation and/or treatment rec-
ommendation for the item described” (Entman, 1993: 5). In social media networks, partisan
messages frame events by altering the frequencies of words, handles, and images (frame ele-
ments) that focus the attention of users on particular partisan traits(Aruguete and Calvo,2018;
Lin et al.,2014). Posts are made accessible to users when peers publish content that makes
salient moral evaluations of blame attribution by increasing the frequency of loaded terms (e.i.
the “Chinese virus”), as well as cognitive assessments of likely threats (i.e. just a cold [“Uma
gripezinha ou resfriadinho”]) (Banks et al.,2020). Framing is critically dependent on the will-
ingness of individuals to share content they observe in their social media feeds (i.e. cascading
activation in networks (Aruguete and Calvo,2018)). Once activated, peers observe social media
messages that “promote a particular problem definition”.
Since Kahneman and Tversky’s (1982) landmark studies on framing and risk, we have come
to understand that presenting questions to voters in terms of losses yields responses that are
substantively different from the responses produced by the same questions presented in terms
15

of gains. Similarly, competing frames that focus the attention of distinct issues, such as job
losses or health risks, alter the weights that voters attach to the negative economic or health
consequences of COVID-19.
Consider first how voters may perceive a politician’s message, such as, “we need to work
together to address this crisis.” In this case, the speaker’s willingness to cooperate with political
rivals provides novel information to voters about the seriousness of the crisis as well as the
importance of investing in reducing health and economic costs, thereby converting enemies into
allies. Now, compare the previous message with one that attributes responsibility to out-group
politicians, such as, “the government response has been careless.” The second message contains
less information, as attacks are interpreted by constituents as a politics-as-usual jab among
contenders. Negative messages, therefore, activate partisan identities and trigger a politically
congruent affective response (Iyengar et al.,2012;Iyengar and Westwood,2015;Mason,2016).
In polarized political environments, ‘cross-the-aisle’ frames and congruent messages from in-
group politicians provide new information to voters about the severity of COVID-19. On the
other hand, negative framing by out-group politicians activates partisan identities and reduces
the informative value of political or scientific facts being reported (Nisbet et al.,2015).
As in Banks et al. (2020), who models the effect of anger on preferences, our experiment
presents respondents with a particular type of frame, procedural or generic, which alters the
perceived legitimacy of the actors’ response to a crisis (Entman,1993). We then inquire on
the extent to which negative and positive frames alter the voter’s evaluations of government
performance and, more importantly, their relative perceptions of job security and health risk.
As in Iyengar and Westwood (2015) and Nisbet et al. (2015), our interest lies in understanding
how partisanship shapes voters’ beliefs about likely outcomes.
16

Hypotheses
We develop a social media framing experiment with positive and negative partisan messages
from high-level politicians to understand the effects of partisan preference and framing on risk
perceptions during the COVID-19 pandemic. In this section, we present the pre-registered
hypothesis and our instruments.
The first set of pre-registered hypotheses tests for the effect of social media content on per-
ceptions of risk and government performance. We consider the effects of negative and positive
messages and the extent to which the effect interacts with partisan cognitive congruence or
dissonance between the authors of the tweet and the respondents’ preferences.
Positive messages bring to voters the willingness of political elites to cooperate with rivals to
fight the COVID-19 pandemic. In an era of high polarization, these messages provide voters with
novel information, reinforcing the importance of unity and cooperation to address the crises. The
negative frames blame political opponents for sowing conflict and weakening the needed response
to the crises. By contrast, positive messages minimize party identity responses and signal that
politicians do not behave as in a ”politics-as-usual” way. Consistent evidence show people weight
negative messages more heavily when compared to positive information (Arceneaux and Nick-
erson,2010), and, when thinking about risk, negative messages frame risks as dynamic losses
for respondents affecting the attention to the topic (Kahneman and Tversky,1982). The first
hypothesis of the experiment predicts negative messages on average to increase perceptions of
personal risk and induce partisan responses in reported support for the government’s response
to the pandemic.
•Hypothesis 1: We predict that negative messages, compared to positive ones, will increase
perceptions of risk and decrease support for the government’s response to the COVID-19
17

pandemic.
A broad literature in political behavior shows that partisanship is central to attitude forma-
tion, in areas as distinctive as candidates evaluation, economic perceptions, support for democ-
racy and authoritarianism, and policy preferences (Green et al.,2004;Arceneaux,2008;Slothuus
and De Vreese,2010;Evans and Andersen,2006;Zaller,1992). Based on this literature, we expect
the framing effect from negative and positive messages to be conditional on partisan identities.
In our second hypothesis, we argue that a “politics-as-usual” polarizing message from elites elic-
its a partisan identity response from voters. We expect that cognitive dissonance between the
respondents’ preferences and the author of the tweets will ensure that health risks and job losses
will be interpreted as wedge issues that separate the parties. We expect cognitive dissonance
to mitigate responses to the social media message when framing in a ”crossing-the-isle” style
politics. Consequently, respondents who observe a “cross-the-aisle” message from a politician
from a different color (T1 and T3) will decrease risk perceptions and increase support for the
government, moderating partisan responses.
•Hypothesis 2: Cognitive dissonance and calls for greater collaboration between politicians
will decrease party identity responses, decrease perceptions of risk, and increase support
for the government.
We expect the opposite effects when cognitive dissonance interacts with negative social media
content. As shown in (Banks et al.,2020), exposure to negative dissonant social media messages
increases contrast effects (Merrill et al.,2003) and heightens perceived polarization, increasing
party identity responses and reducing support for the government. After being exposed to
negative messages by an out-group politician, Banks et al. (2020) show that voters perceived
ideological distance increases (contrast), driving responses to align further with their in-group
18

beliefs. Similar dynamics have been found on previous studies with a focus on political behavior
during a health crisis (Adida et al.,2018). Following this intuition, we expect that to the
extent respondents observe a dissonant partisan signal with a negative frame, partisan identity
responses will be exacerbated. Opposition voters will report heightened risks and lower marks for
government’s response. The opposite effects are expected from Bolsonaro supporters, lowering
their risk exposure and increasing support for the government:
•Hypothesis 3: Cognitive dissonance and negative frames will heighten partisan identity
responses. When exposed to cognitive dissonant negative frames:
–H3a: Respondents aligned with the opposition will report higher health and job risks
and lower performance scores for the government.
–H3b: Respondents aligned with the government will report lower health and job risk,
and greater performance scores for the government.
Experimental Design
Our experiment implements a four-arm treatment assignment in which each respondent is
randomly exposed to one of four different tweets, with a variation on the content and the au-
thor of the message. Each respondent was exposed to only one tweet, and after the treatment
assignment, responded to our outcome variables 10. In order to prime respondents in our ex-
periment, we edited tweets. Although we reduce the experiment’s external validity by not using
real tweets for our treatment conditions, we carefully chose the wording of the tweets based on
actual public statements and social media activity to maximize the validity of the treatment
10The experiment was included in a national online survey in Brazil with 2.400 respondents. The survey is
fielded by Netquest-Vanderbilt, with probabilistic samples drawn by the LAPOP team in Vanderbilt from users
registered with Netquest. The experiment received the approval of the University of Maryland Institutional Board
Review 1552091-3
19

conditions. Internal validity is achieved by randomization, and Section A of the appendix shows
a balanced sample of respondents across a range of socio-demographic and attitudinal variables
between the treatment arms.
We vary only two features of each tweet, the author and the content. For the author, we
use two prominent political figures: Eduardo Bolsonaro, congressmen and son of President Jair
Bolsonaro, and Fernando Haddad, the front-runner candidate of the Workers’ Party in the 2018
national election. We choose high-level politicians to ensure congruence or dissonance between
the message and the respondents’ preferences.
For the content, we vary between a positive and negative framing of COVID-19. In the
positive, we use precisely the same wording for each author, in which the tweets mainly highlight
the existence of a crisis and the importance of President Bolsonaro’s leadership of institutional
efforts to fight the pandemic. For the negative tweets, we created one for each sender, mimicking
their political preferences, thus maximizing external validity for the experiment. With regard to
Eduardo Bolsonaro, the tweets reinforce the argument that the crisis is not serious and that the
opposition and the media are responsible for the ”hysteria” around the spread of the virus. For
Fernando Haddad, the tweet criticizes the government and Bolsonaro’s statements, minimizing
the consequences of the crises. Appendix C presents the wording of each treatment, and the
tweets as the respondents read in Portuguese.11
6 Results: Framing Risk Perceptions
We now turn to our experimental results. We manipulate our four treatment arms to iden-
tify the effects previously described, expecting negative messages from out-group politicians to
11All the respondents were debriefed that the tweets were not factual by the end of the survey.
20

increase perceptions of risk among opposition voters and reduce them among government sup-
porters. The proposed mechanism rests on angry reactions to negative out-group politicians
altering the interpretation of the COVID-19 questions to better follow the policy cues of their
preferred parties.
For presentation purposes, we concentrate on describing the relevant comparisons of all treat-
ments as reported in Figure 3, and report the p-values for the statistically significant and theoret-
ical relevant comparison. Earlier in the article, figure 2showed significant inter-party differences
in evaluations of the governments and in perceptions of job and health risks. For visualization
purposes, estimates in Figure 3manipulate those average results by demeaning our dependent
variables, and showing inter-party deviations when respondents are treated with any of the
different frames.12
Consider the first row of Figure 3, which reports differences in the variables of interest for
each treatment for all respondents. In the top plot of the left, we see that a negative tweet
by Eduardo Bolsonaro reduces reported perceptions of government responses while a negative
tweet by Fernando Haddad does the opposite. In fact, respondents move on average counter
to the political leaning of the author of the tweet, with perceptions of government performance
increasing when Haddad posts a message and decreasing with Bolsonaro (p < 0.05). Results
also show that, on average, negative tweets by Bolsonaro increase perceptions of personal job
risk (“losing your job”) while negative tweets by Haddad reduce job risk (p= 0.12). Health risks,
however, do not seem affected by the different treatments.
12All the respondents in our survey are exposed to at least one tweet. Therefore, we do not have a classic
control group with no information. To model this particular design, we estimate a simple linear regression of
the demeaned outcomes on all the four frames. To capture the effects for all the frames, the model is estimated
without a constant. We report the point-estimates of the model and compare each point-estimate against each
other, using a t-test to assess their statistical difference.
21

Figure 3 Framing Estimates by likely Vote
22

In the second row, we present estimates for the subsample of Bolsonaro voters. Like the full
sample, negative messages by Eduardo Bolsonaro decrease overall perceptions of government
response to the crisis and increase perceptions of job risk. This is an unexpected result, as
respondents treated with negative tweets by Eduardo Bolsonaro are not activating a partisan
response by the in-group. In the third row, we present the estimates of Haddad (Workers’ Party)
voters. Messages by Eduardo Bolsonaro increase perceptions of job risks. As it was the case of
Bolsonaro, we find no significant results on health risks. Social media frames, therefore, have
measurable effect in perceptions of job insecurity among voters of the opposition, as argued
in hypothesis 3b. We find a large gap on job risk perceptions comparing negative message by
Bolsonaro with positive cross-the-isle message by Haddad (p < 0.05).
Finally, the fourth row presents the estimates for independent voters, who preferred to mark
blank in the run-off election rather than voting for either Bolsonaro or Haddad. We had no pre-
registered expectation for this group, however, we believe the discussion and results are worthy
of being reported. Among independents, we see that messages by Haddad increase evaluations
of the government while messages from Bolsonaro decrease them (p < 0.05). Different from
partisans, the most interesting finding is that positive messages modestly increase perceptions
of job and health risks. We interpret this as independents identifying partisan messages as
posturing, thereby reducing the message’s information value while considering positive messages
as informative.
Figure 4re-estimates our models for the sub-samples of self-identified partisans of the Work-
ers’ Party (PT), negative partisans (anti-PT), and others. Results align well with those in
Figure 3. Results indicate that self-identified anti-PT respondents are particularly sensitive to
the treatments, with a significant decline in support for the government and an increase in job
23

Figure 4 Framing Estimates by Negative Mass Partisanship
24

risk assessment when treated to negative messages by Bolsonaro (p < 0.05). In other words, in
the broader partisan group of anti-petistas, a political factor that was crucial for Bolsonaro’s
election in 2018, his polarizing message is indeed increasing perceptions of risk, and hurting his
support.
Overall, our survey experiment finds no robust evidence for our pre-registered hypothesis. Al-
though we find consistent and robust partisan differences on non-experimental survey responses
to risk and support for the government during the pandemic, exposure to distinct framing on
social media seems to alter little how citizens’ update their beliefs. Only one of our hypotheses
is confirmed: voters of the opposition fell more at risk when treated with a negative message
by a high-level politician aligned the Bolsonaro’s government. Giving that we conduct multiple
tests, and do not confirm most of our pre-registered hypotheses, we report our experimental
results as indicating null effects for framing. As discussed in the introduction, this finding goes
in the direction of previous investigations about COVID-19 pandemic in the American context
(Kushner Gadarian et al.,2020), and suggests an environment where respondents are dealing
with a saturated social media environment, which would explain why framing and endorsement
effects have no effects on risk perceptions.
In addition, two risk differences are robust in our experiment, and are purely exploratory since
we did not pre-registered these expectations. First, among independents, positive messages are
read as posturing, increasing their job risk perception and support for the government. In a
polarized environment, crossing-the-isle seems to signal to independents that the crisis is rather
serious. Second, negative messages minimizing the risks of COVID-19 sent by core members of
the government seem to hurt Bolsonaro’s popularity, and increase job risk perceptions, among
his voters, and partisans anti-petistas.
25

7 Why “null findings”? An Autopsy of our Experiment
The experimental results in the previous section are, at first sight, disappointing. The de-
scriptive evidence in the opening pages showed significant party differences in perceptions of
health and job risk. Then, the difference-in-difference analysis of Bolsonaro’s speech gave new
support for the proposed argument, showing that voters are sensitive to partisan messages, with
heightened perceptions of health risk after the aggressive standing of Bolsonaro. The sample size
of the experiment is large and comfortably exceeds power requirements, even for the subsamples
for party voters and independents.13 So, why are the results of the framing experiment modest
and why do we find support for only one of our three hypotheses? Luckily for us, we included
in the survey a number of validation checks that allow us to explore the mechanisms behind the
modest results of the survey experiment.
What could have failed...or not?
Null findings are always important if they disprove theories but are less interesting if they
reflect poor design choices. Therefore, it is important to know what could have failed. There
are three different reasons that may explain weak findings in our experiment: First, (1) respon-
dents could have failed to interpret and/or react to the partisan message of the four different
frames. In that case, weak findings would be explained by the failure of the frame/signal to
which respondents had to react. We may test for this potential problem because we included a
validation check in the experiment, asking respondents if they would“like”, “retweet”, “reply”, or
ignore the tweet. Therefore, we can observe whether partisans’ behavior aligns with the content
13Finding the minimum sample size that would prevent Type I errors is difficult prior to collecting survey
results. We expected a survey of 2,400 respondents to exceed power requirements. This assumption is justified in
the analyses in 7.3 of this section, showing that power requirements were sufficient to test for the effect of anger
on risk.
26

of the frames.
Second, findings may be weak because frames did not elicit the expected emotional response
to the negative or positive tweets posted by the in-group or out-group politician. Following Banks
et al. (2020), we expect the mediation mechanism (“anger”) to activate partisan identities and
increase perceptions of risk among opposition and independent voters. This would be consistent
with results from the difference-in-difference analysis of Bolsonaro’s speech. However, if the
“angry” response to the social media frames is not consistently higher for negative messages by
the out-group politician, there would be modest differences in perceived risk among respondents
exposed to the different frames. While our experimental design expects frames to increase
partisan anger and, in turn, expects anger to increase perceived risk, f rames →anger →
risk, failing to elicit the correct behavioral response would dissociated the frames from risk
perceptions, f rames 6→ anger →risk.
Finally, it is possible that the treatment frames are properly interpreted by the respondents
and that they elicit the expected response, “anger”, without this emotional response changing
risk perceptions, f rames →anger 6→ risk. In this case, the expected hypothesis would be
thoroughly rejected.
In this autopsy, proving the first problem would highlight a design failure (e.g. poor frames);
support for the second problem would amount to an expectation failure, failing to elicit the
correct reaction to the frames (e.g. “anger”). Finally, support for the third problem would
disprove the theory, with the emotional trigger failing to affect perceptions of risk. We proceed
now to troubleshoot our experiment and isolate the source of the reported weak findings.
27

Autopsy of Problem 1: Were the frames properly designed frames
To evaluate if there is a failure to communicate the partisan content of the frames, we
can take advantage of one of the survey questions that asked respondents whether they would
“like”, “retweet”, “reply”, or “ignore” the tweet they had just seen. Descriptive information in
Figure 5shows that, as expected, decisions to “like” or “retweet” follow clear partisan lines, with
voters supporting the government considerably more likely to retweet the negative and positive
messages of Bolsonaro. Similarly, voters of the PT (Workers Party) were considerably more
likely to share messages by Haddad.
More interestingly, results show a clear preference by voters to “like” and “retweet” positive
partisan messages. While government supporters shared 43% of the negative Bolsonaro post,
sharing increased to 63% for the positive post. Numbers also increased among Haddad voters
from 11% to 22%, and among independents from 11% to 34%. Figure 5also shows that sup-
porters of Bolsonaro and independents were considerably more likely to share positive messages
by Haddad.
Third, sharing behavior also reflects a much higher propensity by independents to share
messages from Haddad compared to those of Bolsonaro. Finally, while negative and cognitively
dissonant messages trigger “reply” behavior by out-group voters, this is only true of Haddad
voters in response to negative Bolsonaro messages. By contrast, there is no equivalent change
in “reply” rates when government supporters read a negative Haddad message.
Overall, sharing behavior shows that the treatments were properly interpreted by respondents
and triggered the expected sharing response. Results rule out that the source of the weak findings
was a failure to communicate the partisan content of the tweets. Respondents understood and
reacted as expected to each of the four treatments.However, there is a clear inclination for
28

positive messages among Bolsonaro voters. This will be relevant when troubleshooting the
second problem.
29

Figure 5 Favs, Retweets, Replys in response to each of the four treatments.
30

Autopsy of Problem 2: did the experiment elicit “anger” for negative out-group
messages?
Results in Figure 5already hint that something is not quite as expected and that the affective
reaction to the treatments may be more nuanced than anticipated. Positive tweets by both
Bolsonaro and Haddad collected more shares than negative ones. Further, reply rates for negative
messages by out-group respondents were not particularly high. Both issues suggest that positive
frames by out-group politicians and negative frames from in-group politicians have a larger
presence in the data than what we expected.
We can do considerable more to see how voters react to the content of the tweets and test
for “anger” as a mediator, because, after we asked respondents if they would share a tweet, we
asked them how did the tweet “make them feel.” The “angry” response to this question collected
≈8% in the positive frames and ≈19% in the negative frames. While sharing behavior is higher
for positive messages, “angry” responses were indeed higher for the negative tweets. Table 2
presents descriptive evidence using logistic models for the effects of the four frames eliciting
“anger” among our respondents. In the overall sample, negative messages from pro-government
and anti-government officials induced similar levels of anger and positive messages have no
statistically significant effect. However, while negative posts elicit more angry reactions, the
effects filtering by the partisan groups are not consistent: among Haddad voters, both in-group
and out-group message elicits similar amounts of angry response, while among Bolsonaro voters
both out-group negative and positive messages induced “anger”.
Results in the Table 2provide compelling evidence of a disconnect between the content of the
frames and the expected emotional response. The evidence confirms that the negative frames
increased “angry” responses. However, the difference between the negative and positive tweets
31

is modest and “anger” was frequent after reading in-group and out-group messages. Therefore,
the proposed frames failed to elicit the expected response, f rames 6→ anger. While “anger” was
shown to have a crucial effect on subjective risk, the effect of the frames was modest.
Autopsy of Problem 3: Is there a disconnect between “anger” and risk?
Table 2estimates the determinants of our dependent variables perceived health risk, job risk,
and government performance. As covariates we include the four treatments, a variable taking
the value of 1 if the respondent indicated that they felt anger after reading the tweet, and a
latency variable that measured the time, log(miliseconds), that respondents took to answer
the “how did you feel” question. As in the pre-approved plan, we expect automatic and fast
responses to be associated with heightened perceptions of risk. What Kahneman (2011) defines
as a “System 1” response.
Table 2shows evidence to rule out problem number 3 in the health equation. Results show, as
expected, that “anger” is associated with an increase in the users’ perceptions of health risk. The
effect of “anger” increases perceptions of subjective health risk among Haddad and independent
voters.
Results in Model 1 of Table 2show that an angry response yield a statistically significant 0.241
increase in perceived risk for the full sample. Model 2 in Table 2shows that the effect of anger
on perceptions of health risk holds when controlling for vote intention, a statistically significant
0.187. Results also show that a lower time to express how they feel is associated with higher
perceptions of risk. The effect of time is consistent with a decision that “operates automatically
and quickly [...] originating from impressions and feelings”(Kahneman 2011, pp.21).
While results of this autopsy provide compelling evidence of a positive effect of “anger” on
perceptions of health risk, validating the mediating mechanism, this is not the case for subjective
32

job risk. The effect of “anger” on job risk is not statistically significant. We interpret the lack
of significance as supporting evidence for Problem 3 being an issue in the job equation. That
is, frames that elicit anger are not having the hypothesized effect on perceived job risk. Finally,
we do find an effect of the Anger mediator decreasing support for the government’s response to
the pandemic, as shown in model 5 in table 2. However, the effect shrinks and loses significance
when controlling for party vote, which is a consequence of exteme levels of polarization in the
support for the government during the pandemic. .
To summarize, the autopsy on key validation checks of our experiment allows us to discard
problems in the interpretation of the frames (Problem 1) and shows a disconnect between the
frames and our expected emotional responses by respondents (Problem 2). In the case of health
and support for the government, our model supports the interpretation that frames 6→ anger →
risk, instead of the expected f rames →anger →risk.
33

Table 2 Regression Models: Effects of Anger on Risk and Support for the Government
Health Risks Job Risks Support for the Government
(1) (2) (3) (4) (5) (6)
Anger 0.241∗∗∗ 0.187∗∗∗ 0.013 −0.032 −0.246∗∗∗ −0.046
(0.053) (0.053) (0.072) (0.072) (0.062) (0.055)
Latency −0.109∗∗∗ −0.101∗∗∗ −0.039 −0.033 0.099∗∗ 0.077∗∗
(0.036) (0.035) (0.048) (0.047) (0.041) (0.036)
Negative Bolsonaro −0.079 −0.071 0.039 0.046 −0.045 −0.073
(0.052) (0.051) (0.070) (0.069) (0.060) (0.053)
Negative Haddad −0.045 −0.035 −0.069 −0.060 0.108∗0.074
(0.051) (0.051) (0.069) (0.069) (0.060) (0.052)
Positive Haddad −0.024 −0.010 −0.026 −0.013 0.064 0.027
(0.051) (0.050) (0.068) (0.068) (0.059) (0.052)
Haddad Voters 0.026 −0.0005 −0.331∗∗∗
(0.045) (0.061) (0.047)
Bolsonaro Voters −0.316∗∗∗ −0.295∗∗∗ 0.863∗∗∗
(0.042) (0.057) (0.043)
Constant 3.060∗∗∗ 3.142∗∗∗ 2.595∗∗∗ 2.681∗∗∗ 1.868∗∗∗ 1.722∗∗∗
(0.123) (0.123) (0.165) (0.167) (0.143) (0.127)
Observations 2,354 2,354 2,352 2,352 2,352 2,352
Adjusted R20.011 0.042 −0.001 0.013 0.010 0.245
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
34

8 Concluding Remarks
In a time when social distancing is the primary health response to the COVID-19 pandemic,
understanding subjective assessments of health and job risks is essential. In countries such as
Brazil, Mexico, and the United States, health and job policies have become deeply contested
issues that separate partisans and trigger identity responses. In this article, we (1) provide
descriptive evidence of large differences in perceptions of risk by pro-government and opposition
voters; (2) test for the effect of public discourses by Bolsonaro on perceptions of individual
risks, and (3) test for the effect of negative and positive social media frames on perceptions of
individual risk.
Our results verify the existence of partisan differences in perceptions of risks; a heightened ef-
fect of government speeches on opposition voters perceptions of personal risk; and a bounded par-
tisan identity response to negative social media messages, particularly against pro-government
messages denying responsibility for the crisis. Evidence of framing effects from social media mes-
sages in our experiment are modest, and mostly null considering our initial hypothesis for the
effect of negative content and of crossing-the-isle positive social media message on risk percep-
tions. However, we find evidence of backlash against negative messages by in-group politicians
for the government supporters in Brazil. Rather than triggering partisan responses, negative
messages from in-group politicians triggered opposite responses. Bolsonaro voters exposed to
negative messages by Bolsonaro increased their perceptions of job and health risks, and decrease
their support for the government. Similarly, Haddad voters exposed to negative messages by
Haddad reduced their perceptions of job and health risks. Our experiment provides evidence for
citizens’ behavioral reactions to different narratives during the first months of the COVID-19
pandemic in Brazil, and suggests polarization was not being received as an effective strategy by
35

the core supporters of the government.
While the COVID-19 crisis lingers, political acts such as rallies, party meetings, and fundrais-
ing move to the virtual world. In a context of restricted physical mobility, social media and
technologically mediated information exchanges become increasingly important. Beyond the
pre-registered findings, our research provides novel evidence on the partisan online behavior
of negative and positive social media messages. Measures of the social media response to our
treatments provide clear evidence that positive messages were more extensively shared by all
voters, in-group and out-group, and that negative messages activated a smaller number of in-
tense voters. Negative social media messages, therefore, both induce identity responses by strong
partisans but also reduce participation by less committed voters. This is an important effect
that is worth exploring in future research, as it provides evidence of content in social media
data being considerably more partisan than that expected from in-group voters. Therefore, at
least in Brazil’s case during the first months of the pandemic, activating partisan identities to
energize the base also reduces overall support for the government among its own constituency.
36

References
Adida, C. L., Dionne, K. Y., and Platas, M. R. (2018). Ebola, elections, and immigration: how
politicizing an epidemic can shape public attitudes. Politics, Groups, and Identities, pages
1–27.
Ajzenman, N., Cavalcanti, T., and Da Mata, D. (2020). More than Words: Leaders’ Speech and
Risky Behavior During a Pandemic. Working Paper.
Allcott, H., Boxell, L., Conway, J., Gentzkow, M., Thaler, M., and Yang, D. Y. (2020). Po-
larization and Public Health: Partisan Differences in Social Distancing during COVID-19.
Available at SSRN 3570274.
Ames, B. (2001). The Deadlock of Democracy in Brazil. University of Michigan Press, Ann
Arbor.
Arceneaux, K. (2008). Can partisan cues diminish democratic accountability? Political Behavior,
30(2):139–160.
Arceneaux, K. and Nickerson, D. W. (2010). Comparing negative and positive campaign mes-
sages: Evidence from two field experiments. American Politics Research, 38(1):54–83.
Aruguete, N. and Calvo, E. (2018). Time to #Protest: Selective Exposure, Cascading Activation,
and Framing in Social Media. Journal of Communication, 68(3):480–502.
Baker, A., Sokhey, A. E., Ames, B., and Renno, L. R. (2016). The dynamics of partisan
identification when party brands change: The case of the workers party in brazil. Journal of
Politics, 78(1).
37

Banks, A., Karol, D., Calvo, E., and Telhami, S. (2020). #polarized feeds: Two experiments on
polarization, framing, and social media. International Journal of Press/Politics.
Barrios, J. and Hochber, Y. (2020). Risk Perception Through the Lens of Politics in the Time
of the COVID-19 Pandemic. Available at SSRN 3568766.
Bullock, J. G., Gerber, A. S., Hill, S. J., and Huber, G. A. (2013). Partisan bias in factual beliefs
about politics. Technical report, National Bureau of Economic Research.
Coppock, A. (2019). ri2: Randomization inference for randomized experiments. R package
version 0.1, 2.
Entman, R. M. (1993). Framing: Toward clarification of a fractured paradigm. Journal of
communication, 43(4):51–58.
Evans, G. and Andersen, R. (2006). The political conditioning of economic perceptions. The
Journal of Politics, 68(1):194–207.
Gerber, A. S. and Green, D. P. (2012). Field experiments: Design, analysis, and interpretation.
WW Norton.
Green, D. P., Palmquist, B., and Schickler, E. (2004). Partisan hearts and minds: Political
parties and the social identities of voters. Yale University Press.
Iyengar, S. (1990). Framing responsibility for political issues: The case of poverty. Political
behavior, 12(1):19–40.
Iyengar, S., Sood, G., and Lelkes, Y. (2012). Affect, not ideologya social identity perspective on
polarization. Public opinion quarterly, 76(3):405–431.
38

Iyengar, S. and Westwood, S. J. (2015). Fear and loathing across party lines: New evidence on
group polarization. American Journal of Political Science, 59(3):690–707.
Kahneman, D. (2011). Thinking, fast and slow. Macmillan.
Kahneman, D. and Tversky, A. (1982). The psychology of preferences. Scientific American,
246(1):160–173.
Kraft, P. W., Lodge, M., and Taber, C. S. (2015). Why people “don’t trust the evidence”
motivated reasoning and scientific beliefs. The ANNALS of the American Academy of Political
and Social Science, 658(1):121–133.
Kushner Gadarian, S., Goodman, S. W., and Pepinsky, T. B. (2020). Partisanship, Health Be-
havior, and Policy Attitudes in the Early Stages of the COVID-19 Pandemic. Health Behavior,
and Policy Attitudes in the Early Stages of the COVID-19 Pandemic (March 27, 2020).
Lin, Y.-R., Keegan, B., Margolin, D., and Lazer, D. (2014). Rising tides or rising stars?:
Dynamics of shared attention on Twitter during media events. PloS one, 9(5).
Mainwaring, S. (1991). Politicians, Parties, and Electoral Systems: Brazil in Comparative
Perspective. Comparative Politics, 24(1):21–43.
Mainwaring, S. (1999). Rethinking Party Systems in the Third Wave of Democratisation: The
Case of Brazil. Stanford University Press, Stanford, California.
Mainwaring, S. and Scully, T. (1995). Building democratic institutions: party systems in Latin
America. Stanford University Press, Stanford, Calif.
Mariani, L., Gagete-Miranda, J., and Retti, P. (2020). Words can hurt: How political commu-
nication can change the pace of an epidemic. CEPR PRESS.
39

Mason, L. (2016). A cross-cutting calm: How social sorting drives affective polarization. Public
Opinion Quarterly, 80(S1):351–377.
Merrill, S., Grofman, B., and Adams, J. (2003). Assimilation and contrast effects in voter
projections of party locations: Evidence from norway, france, and the usa. European Journal
of Political Research, 40(2):199–221.
Nisbet, E. C., Cooper, K. E., and Garrett, R. K. (2015). The partisan brain: How dissonant
science messages lead conservatives and liberals to (dis) trust science. The ANNALS of the
American Academy of Political and Social Science, 658(1):36–66.
Power, T. J. and Rodrigues-Silveira, R. (2018). The political right and party politics 1. Routledge
Handbook of Brazilian Politics, pages 251–268.
Samuels, D. (2003). Ambition, federalism, and legislative politics in Brazil. Cambridge University
Press, Cambridge ; New York.
Samuels, D. J. and Zucco, C. (2018). Partisans, antipartisans, and nonpartisans: Voting behavior
in Brazil. Cambridge University Press.
Slothuus, R. and De Vreese, C. H. (2010). Political parties, motivated reasoning, and issue
framing effects. The Journal of Politics, 72(3):630–645.
Thaler, R. H., Tversky, A., Kahneman, D., and Schwartz, A. (1997). The effect of myopia
and loss aversion on risk taking: An experimental test. The quarterly journal of economics,
112(2):647–661.
Zaller, J. (1992). The nature and origins of mass opinion. Cambridge university press.
40

Will I get COVID-19?
Negative Partisanship, Social Media Frames, and Perceptions of
Health Risk in Brazil
Supporting Information Files (SIF)
41

Section A: Survey Information
Our paper presents observational, quasi-experimental, and a framing experiment using novel
data from an national on-line survey fielded by Netquest-Vanderbilt. The survey uses proba-
bilistic samples drawn by the LAPOP team in Vanderbilt implemented with the panel of users
registered with Netquest. The entire survey and the embedded framing experiment received the
approval of the University of Maryland Institutional Board Review 1552091-3.
The survey was carried out from March 23 to May 08, 2020 from a national poll of 2,360
respondents. Completion for the survey took on average 28 minutes. We provide two different
incentives for respondents to engage in the survey. Beyond the survey experiment described here,
our survey asked a series of questions about trust, policy preferences, social media consumption
and standard demographic information. Several of these pre-treatment variables were used and
described in different sections for this paper.
Below, we present information about some of the survey variables used throughout the paper.
In a later section in this appendix, we describe in detail the treatment conditions, and outcomes
variables.
Table 3 Survey Questions - Demographic Information
Variable Wording Responses
Age What is your age Binned (From 18-25 up to more than 66)
Gender What is your gender? Male/Female
Education What education level have you achieved? Graduate Studies)
Employment During last week, did you work or study
at least one hour, in some paid activity?
Yes/no
Income magine a staircase with 10 steps. In the
first step, people with lower income are
located, and in step 10, people with
higher income are located. Where would
you be located?
0-10
Income
Assistance
During last month, did you or a member
of your household received
Nominal with government programs
42

Table 4 Survey Questions -Political Attributes and Behavioral Responses to the Treatment
Variable Wording Responses
Likely to Vote
(First Round)
Which candidate would you support if the
presidential election “were to take place
next week” ?
All presidential candidates from 2018
Likely to
Vote (Runoff)
Which candidate would you support if the
runoff presidential election “were to take
place next week” ?
Jair Bolsonaro, Fernando Haddad, Null
Positive
Partisanship
Which party do you like the most? List of Political Parties in Brazil
Negative
Partisanship
Which party do you dislike more ? List of Political Parties in Brazil
Ideological
Placement
Imagine a scale that goes from “very
conservative” to “very progressive”, were
would you place yourself?
0-10
Emotions to
the
Treatment
Thinking about the tweet we just showed
you, do you feel
Angry, Happy, Disgusted, Optimistic,
Stressed, Sad, Fearful, Indifferent
Reactions to
the
Treatment
Thinking about the tweet we showed you.
Would you?
Fav, Retweet, Reply, Ignore
To guarantee that our randomization procedure worked properly, we present below demo-
graphic information for our respondents across the four treatment conditions of our framing
experiment. As the reader can assess, there are no significant differences across the treatment
groups in our sample. Since most of these variables are nominal, the values do not have a direct
interpretation.
43

Table 5 Demographics Across the Treatment Arms
Variable Quantity Negative Bolsonaro Negative Haddad Positive Bolsonaro Positive Haddad
Age
Mean 3.01 3.12 3.11 3.08
Standard Error 3.36 3.30 3.20 3.39
Education
Mean 2.15 2.11 2.19 2.18
Standard Error 1.50 1.55 1.54 1.56
Gender
Mean 4.36 4.57 4.50 4.50
Standard Error 0.63 0.63 0.63 0.62
Ideological Placement
Mean 5.24 5.55 5.22 5.36
Standard Error 1.28 1.27 1.22 1.26
Occupation
Mean 6.47 6.62 6.41 6.32
Standard Error 0.96 0.96 0.97 0.94
Income Assistance
Mean 1.50 1.47 1.47 1.49
Standard Error 2.18 2.00 2.12 2.02
Relative Income
Mean 1.75 1.73 1.78 1.71
Standard Error 0.50 0.50 0.50 0.50
Total Cases
Total Number of Cases 571.00 588.00 590.00 613.00
44

Section B: Negative Partisanship and Risk Perceptions
In this section, we provide further descriptive evidence for deeper partisan divisions on risk
perceptions and government assessment. We first replicate figure 1 in the paper but using a
measure for negative and positive partisanship towards the Workers Party (PT). As argued by
Samuels and Zucco (2018), mass partisanship in Brazil is strongly connected to voters’ assessment
about the PT. Therefore, we test for this explanation to increase the robustness of our findings.
Figure 6presents the results. We manipulate positive and negative partisanship, as suggested
in Samuels and Zucco (2018), and use the excluded cases as others in our sample. 32% of Pro-
PT supporters report fell very likely chance of losing their job and 24% of becoming infected
by COVID-19, compared respectively to 22% and 13% for anti-PT respondents. In terms of
assessing government responses, half of our sample of PT supporters considered them very
unappropriate, while only 29% among anti-petistas have the same assessment.
We also provide in table 6the numerical results from the models summarized on figure 2.
To make the presentation more intuitive, we use Bolsonaro voters, and Anti-Petistas, as the
reference group for the models. In the main paper, we do not explore the results for the control
variables, yet their interpretation provides some interesting correlational insights about factors
associated with risk perceptions in Brazil. Older, wealthier men report across all the models
lower risk perceptions. On the other side, more education decreases risks on the job market,
but increases fear of being infected by COVID-19. A similar effect is detected when comparing
employed versus unemployed respondents, with the former predicting higher health risk, and
lower perception regarding the labor market.
45

Figure 6 Survey assessments conditional on Negative Partisanship of the quality of the Government response, perceptions of personal health risk,
and perceptions of personal job security, March 23 through May 4, 2020.
46

Table 6 Regression models of perception of risk and government assessments with full controls
Dependent variable:
Job Risk Health Risk Government Assessment Job Risk Health Risk Government Assessment
(1) (2) (3) (4) (5) (6)
Intercept 3.309∗∗∗ 2.514∗∗∗ 3.001∗∗∗ 3.349∗∗∗ 2.655∗∗∗ 2.793∗∗∗
(0.112) (0.084) (0.087) (0.115) (0.087) (0.097)
Voters Haddad 0.202∗∗∗ 0.337∗∗∗ −1.203∗∗∗
(0.063) (0.047) (0.049)
Voters Independents 0.238∗∗∗ 0.296∗∗∗ −0.868∗∗∗
(0.058) (0.044) (0.046)
Petistas 0.225∗∗∗ 0.192∗∗∗ −0.922∗∗∗
(0.079) (0.059) (0.067)
Others (Non-Partisans) 0.118∗∗ 0.121∗∗∗ −0.591∗∗∗
(0.055) (0.042) (0.047)
Income −0.055∗∗∗ −0.035∗∗∗ −0.012 −0.056∗∗∗ −0.036∗∗∗ −0.007
(0.012) (0.009) (0.010) (0.012) (0.009) (0.010)
Gender:Male −0.037 −0.072∗−0.025 −0.052 −0.097∗∗ 0.048
(0.050) (0.038) (0.039) (0.050) (0.038) (0.042)
Employed −0.155∗∗∗ 0.122∗∗∗ 0.055 −0.146∗∗∗ 0.132∗∗∗ 0.025
(0.052) (0.039) (0.041) (0.053) (0.040) (0.045)
Education −0.055∗∗∗ 0.055∗∗∗ −0.039∗∗ −0.050∗∗ 0.058∗∗∗ −0.060∗∗∗
(0.020) (0.015) (0.016) (0.020) (0.015) (0.017)
Age −0.140∗∗∗ −0.038∗∗∗ 0.021 −0.141∗∗∗ −0.047∗∗∗ 0.034∗∗
(0.016) (0.012) (0.013) (0.017) (0.013) (0.014)
Observations 2,159 2,163 2,158 2,142 2,146 2,142
Adjusted R20.074 0.057 0.247 0.070 0.035 0.115
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
47

Section C: Tweets for the Treatment Conditions
Table 7presents the treatment conditions in English. Figure 7presents the images, as the
respondents read in Portuguese, of the tweets used in each of the treatment conditions.
Table 7 Treatment Conditions
Positive Tweet Negative Tweet
Eduardo
Bolsonaro
The world is currently living an
unprecedented crisis. Countries rally
together to fight against Coronavirus.
It is the responsibility of President
@jairbolsonaro to coordinate our
answers. He needs to act together
with Congress, Business leaders, and
civil society. This is what we expect
in such critical times.
The world is currently living an
unprecedented crisis. Countries rally
together to fight against Coronavirus.
However, we have seen other diseases
before, some way more dangerous
than Coronavirus, that did not lead
to all this hysteria. Only that it was
the PT’s government at that time.
No panic. Switch off from the
pandemic of misinformation from the
media
Fernando
Haddad
The world is currently living an
unprecedented crisis. Countries rally
together to fight against Coronavirus.
It is the responsibility of President
@jairbolsonaro to coordinate our
answers. He needs to act together
with Congress, Business leaders, and
civil society. This is what we expect
in such critical times.
The world is currently living an
unprecedented crisis. Countries rally
together to fight against Coronavirus.
President @jairbolsonaro is delayed
in answering. He is only concerned
about attacking his opponents and
take part in protests that put in risk
the Brazilian people.
48

Figure 7 Tweets for the Treatment Conditions
a) Eduardo Bolsonaro x Positive Tweet (T1) b) Eduardo Bolsonaro x Negative Tweet (T2)
a) Fernando Haddad x Positive Tweet (T3) b) Fernando Haddad x Negative Tweet (T4)
Section E: Robustness Checks for the Effects of Bolsonaro’s Speech
This section provides some robustness checks for the effects of the Bolsonaro’s national pro-
nouncement on March 24 discussed in the paper. Our results’ main inferential threat comes
from the chance that our measures might capture random fluctuations over time of respondents’
risk perceptions. Therefore, to increase the robustness of our findings we examine the extant to
which our point estimates differ from changes in our dependent variable over time. We perform
49

a set of placebo checks to analyze this possibility.
We estimate the same model, as in section four of the paper, but using as a placebo for the
treatment effect each other day after March 24. In other words, we simulate as if Bolsonaro
speech had happened in all the remaining 45 days we have in our sample. As in the main paper,
we estimate the models using data from two days before, and two days for each placebo test.
Figure 8presents the results. We color red the treatment results presented in the main
paper and two other presidential pronouncements made by Bolsonaro to discuss the COVID-
19 pandemic on TV. Our results suggest strong support for our argument that the effects of
Bolsonaro’s speech on March 24 is hardly a random variation from respondents updating their
risk assessment over time. For the Job perceptions, only the other two point-estimates, out of 45
placebos, are positive and statistically different from zero, as it is the true treatment effect. As
a matter of fact, both estimates happen exactly in the following days of another pronouncement
of Bolsonaro. For the Health models, only three out of 45 placebos are positive and statistically
different from zero. Overall, the placebo checks give strong support for the robustness of our
findings.
Finally, we use randomization inference to asses covariates balance between our survey re-
spondents before and after March 24 (Gerber and Green,2012;Coppock,2019). Figure 9
plots a histogram of the observed F-statistic, and the null distribution of F-statistics calculated
through randomization inference, and using a linear probability model regressing the treatment
assignment (answering the survey between March 24-25, after Bolsonaro’s speech) on a set of
demographic and political information collected during the survey (age, gender, occupation, ed-
ucation, income, ideology, positive and negative partisanship, and voting choices). As in the
main models, we limit the analysis to respondents who answered two days before (control), and
50

Figure 8 Placebo Checks for the Effects of Bolsonaro Speech on March 24.
51

two days after (treatment) Bolsonaro’s speech.
Randomization inference provides a strategy to calculate p-values for hypothesis test using
randomization techniques. The null hypothesis for our robustness check is that a set of socio-
demographics and political covariates do not explain if the respondents answered to the survey
before or after March 24. The results are presented on figure 9. The distribution of F-statistics
indicates that the null hypothesis (covariates have no effect on treatment assignment) cannot be
rejected. Approximately 75% (P-value=0.75) of the simulated F-statistics were larger than the
observed F-statistic in the true model. The vertical red line on both graphs denotes the observed
F-statistic, while shaded regions denote simulated estimates more extreme than the one observed.
I used 5.000 simulations under the the null hypothesis, implied by random assignment, that no
covariates is correlated with answering the survey before or after March 24.
Figure 9 Randomization Inference for Covariate Balance Before and After Bolsonaro’s Speech.
Section F: Effect of Frames on “Anger”
52

Table 8 Regression Models: Effects of the four-frame treatments on “angry” response.
All Sample Bolsonaro Voters Haddad Voters
(1) (2) (3)
Constant 1.090∗∗∗ 1.020∗∗∗ 1.169∗∗∗
(0.014) (0.019) (0.032)
Negative Bolsonaro 0.091∗∗∗ 0.044 0.154∗∗∗
(0.020) (0.027) (0.045)
Negative Haddad 0.106∗∗∗ 0.161∗∗∗ 0.117∗∗∗
(0.020) (0.027) (0.045)
Positive Haddad −0.007 0.060∗∗ −0.049
(0.020) (0.026) (0.044)
Observations 2,362 855 658
Adjusted R20.021 0.039 0.036
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
References
Adida, C. L., Dionne, K. Y., and Platas, M. R. (2018). Ebola, elections, and immigration: how
politicizing an epidemic can shape public attitudes. Politics, Groups, and Identities, pages
1–27.
Ajzenman, N., Cavalcanti, T., and Da Mata, D. (2020). More than Words: Leaders’ Speech and
Risky Behavior During a Pandemic. Working Paper.
Allcott, H., Boxell, L., Conway, J., Gentzkow, M., Thaler, M., and Yang, D. Y. (2020). Po-
larization and Public Health: Partisan Differences in Social Distancing during COVID-19.
Available at SSRN 3570274.
Ames, B. (2001). The Deadlock of Democracy in Brazil. University of Michigan Press, Ann
53

Arbor.
Arceneaux, K. (2008). Can partisan cues diminish democratic accountability? Political Behavior,
30(2):139–160.
Arceneaux, K. and Nickerson, D. W. (2010). Comparing negative and positive campaign mes-
sages: Evidence from two field experiments. American Politics Research, 38(1):54–83.
Aruguete, N. and Calvo, E. (2018). Time to #Protest: Selective Exposure, Cascading Activation,
and Framing in Social Media. Journal of Communication, 68(3):480–502.
Baker, A., Sokhey, A. E., Ames, B., and Renno, L. R. (2016). The dynamics of partisan
identification when party brands change: The case of the workers party in brazil. Journal of
Politics, 78(1).
Banks, A., Karol, D., Calvo, E., and Telhami, S. (2020). #polarized feeds: Two experiments on
polarization, framing, and social media. International Journal of Press/Politics.
Barrios, J. and Hochber, Y. (2020). Risk Perception Through the Lens of Politics in the Time
of the COVID-19 Pandemic. Available at SSRN 3568766.
Bullock, J. G., Gerber, A. S., Hill, S. J., and Huber, G. A. (2013). Partisan bias in factual beliefs
about politics. Technical report, National Bureau of Economic Research.
Coppock, A. (2019). ri2: Randomization inference for randomized experiments. R package
version 0.1, 2.
Entman, R. M. (1993). Framing: Toward clarification of a fractured paradigm. Journal of
communication, 43(4):51–58.
54

Evans, G. and Andersen, R. (2006). The political conditioning of economic perceptions. The
Journal of Politics, 68(1):194–207.
Gerber, A. S. and Green, D. P. (2012). Field experiments: Design, analysis, and interpretation.
WW Norton.
Green, D. P., Palmquist, B., and Schickler, E. (2004). Partisan hearts and minds: Political
parties and the social identities of voters. Yale University Press.
Iyengar, S. (1990). Framing responsibility for political issues: The case of poverty. Political
behavior, 12(1):19–40.
Iyengar, S., Sood, G., and Lelkes, Y. (2012). Affect, not ideologya social identity perspective on
polarization. Public opinion quarterly, 76(3):405–431.
Iyengar, S. and Westwood, S. J. (2015). Fear and loathing across party lines: New evidence on
group polarization. American Journal of Political Science, 59(3):690–707.
Kahneman, D. (2011). Thinking, fast and slow. Macmillan.
Kahneman, D. and Tversky, A. (1982). The psychology of preferences. Scientific American,
246(1):160–173.
Kraft, P. W., Lodge, M., and Taber, C. S. (2015). Why people “don’t trust the evidence”
motivated reasoning and scientific beliefs. The ANNALS of the American Academy of Political
and Social Science, 658(1):121–133.
Kushner Gadarian, S., Goodman, S. W., and Pepinsky, T. B. (2020). Partisanship, Health Be-
havior, and Policy Attitudes in the Early Stages of the COVID-19 Pandemic. Health Behavior,
and Policy Attitudes in the Early Stages of the COVID-19 Pandemic (March 27, 2020).
55

Lin, Y.-R., Keegan, B., Margolin, D., and Lazer, D. (2014). Rising tides or rising stars?:
Dynamics of shared attention on Twitter during media events. PloS one, 9(5).
Mainwaring, S. (1991). Politicians, Parties, and Electoral Systems: Brazil in Comparative
Perspective. Comparative Politics, 24(1):21–43.
Mainwaring, S. (1999). Rethinking Party Systems in the Third Wave of Democratisation: The
Case of Brazil. Stanford University Press, Stanford, California.
Mainwaring, S. and Scully, T. (1995). Building democratic institutions: party systems in Latin
America. Stanford University Press, Stanford, Calif.
Mariani, L., Gagete-Miranda, J., and Retti, P. (2020). Words can hurt: How political commu-
nication can change the pace of an epidemic. CEPR PRESS.
Mason, L. (2016). A cross-cutting calm: How social sorting drives affective polarization. Public
Opinion Quarterly, 80(S1):351–377.
Merrill, S., Grofman, B., and Adams, J. (2003). Assimilation and contrast effects in voter
projections of party locations: Evidence from norway, france, and the usa. European Journal
of Political Research, 40(2):199–221.
Nisbet, E. C., Cooper, K. E., and Garrett, R. K. (2015). The partisan brain: How dissonant
science messages lead conservatives and liberals to (dis) trust science. The ANNALS of the
American Academy of Political and Social Science, 658(1):36–66.
Power, T. J. and Rodrigues-Silveira, R. (2018). The political right and party politics 1. Routledge
Handbook of Brazilian Politics, pages 251–268.
56

Samuels, D. (2003). Ambition, federalism, and legislative politics in Brazil. Cambridge University
Press, Cambridge ; New York.
Samuels, D. J. and Zucco, C. (2018). Partisans, antipartisans, and nonpartisans: Voting behavior
in Brazil. Cambridge University Press.
Slothuus, R. and De Vreese, C. H. (2010). Political parties, motivated reasoning, and issue
framing effects. The Journal of Politics, 72(3):630–645.
Thaler, R. H., Tversky, A., Kahneman, D., and Schwartz, A. (1997). The effect of myopia
and loss aversion on risk taking: An experimental test. The quarterly journal of economics,
112(2):647–661.
Zaller, J. (1992). The nature and origins of mass opinion. Cambridge university press.
57