Exposure to opposing views on social media can

increase political polarization

Christopher A. Baila,1 , Lisa P. Argyleb , Taylor W. Browna , John P. Bumpusa , Haohan Chenc , M. B. Fallin Hunzakerd ,
Jaemin Leea , Marcus Manna , Friedolin Merhouta , and Alexander Volfovskye
a
Department of Sociology, Duke University, Durham, NC 27708; b Department of Political Science, Brigham Young University, Provo, UT 84602; c Department
of Political Science, Duke University, Durham, NC 27708; d Department of Sociology, New York University, New York, NY 10012; and e Department of
Statistical Science, Duke University, Durham, NC 27708

Edited by Peter S. Bearman, Columbia University, New York, NY, and approved August 9, 2018 (received for review March 20, 2018)

There is mounting concern that social media sites contribute to challenges for the study of social media echo chambers and
political polarization by creating “echo chambers” that insulate political polarization, since it is notoriously difficult to establish
people from opposing views about current events. We surveyed whether social media networks shape political opinions, or vice
a large sample of Democrats and Republicans who visit Twit- versa (27–29).
ter at least three times each week about a range of social Here, we report the results of a large field experiment designed
policy issues. One week later, we randomly assigned respon- to examine whether disrupting selective exposure to partisan
dents to a treatment condition in which they were offered information among Twitter users shapes their political attitudes.
financial incentives to follow a Twitter bot for 1 month that Our research is governed by three preregistered hypotheses. The
exposed them to messages from those with opposing political first hypothesis is that disrupting selective exposure to parti-
ideologies (e.g., elected officials, opinion leaders, media orga- san information will decrease political polarization because of

SOCIAL SCIENCES
nizations, and nonprofit groups). Respondents were resurveyed intergroup contact effects. A vast literature indicates contact
at the end of the month to measure the effect of this treat- between opposing groups can challenge stereotypes that develop
ment, and at regular intervals throughout the study period to in the absence of positive interactions between them (30). Stud-
monitor treatment compliance. We find that Republicans who ies also indicate intergroup contact increases the likelihood of
followed a liberal Twitter bot became substantially more con- deliberation and political compromise (31–33). However, all of
servative posttreatment. Democrats exhibited slight increases these previous studies examine interpersonal contact between
in liberal attitudes after following a conservative Twitter bot, members of rival groups. In contrast, our experiment creates
although these effects are not statistically significant. Notwith- virtual contact between members of the public and opinion lead-
standing important limitations of our study, these findings have ers from the opposing political party on a social media site.
significant implications for the interdisciplinary literature on polit- It is not yet known whether such virtual contact creates the
ical polarization and the emerging field of computational social
science.
Significance
political polarization | computational social science | social networks |
social media | sociology Social media sites are often blamed for exacerbating political
polarization by creating “echo chambers” that prevent people
from being exposed to information that contradicts their pre-

P olitical polarization in the United States has become a central

focus of social scientists in recent decades (1–7). Americans
are deeply divided on controversial issues such as inequality, gun
existing beliefs. We conducted a field experiment that offered
a large group of Democrats and Republicans financial com-
pensation to follow bots that retweeted messages by elected
control, and immigration—and divisions about such issues have officials and opinion leaders with opposing political views.
become increasingly aligned with partisan identities in recent Republican participants expressed substantially more conser-
years (8, 9). Partisan identification now predicts preferences vative views after following a liberal Twitter bot, whereas
about a range of social policy issues nearly three times as well Democrats’ attitudes became slightly more liberal after fol-
as any other demographic factor—such as education or age (10). lowing a conservative Twitter bot—although this effect was
These partisan divisions not only impede compromise in the not statistically significant. Despite several limitations, this
design and implementation of social policies but also have far- study has important implications for the emerging field of
reaching consequences for the effective function of democracy computational social science and ongoing efforts to reduce
more broadly (11–15). political polarization online.
America’s cavernous partisan divides are often attributed to
“echo chambers,” or patterns of information sharing that rein- Author contributions: C.A.B., L.P.A., T.W.B., J.P.B., H.C., M.B.F.H., J.L., M.M., F.M., and
force preexisting political beliefs by limiting exposure to oppos- A.V. designed research; C.A.B., L.P.A., T.W.B., H.C., M.B.F.H., J.L., M.M., and F.M. per-
formed research; C.A.B., T.W.B., H.C., J.L., and A.V. contributed new reagents/analytic
ing political views (16–20). Concern about selective exposure tools; C.A.B., L.P.A., T.W.B., H.C., M.B.F.H., J.L., M.M., F.M., and A.V. analyzed data; and
to information and political polarization has increased in the C.A.B., L.P.A., T.W.B., M.B.F.H., M.M., F.M., and A.V. wrote the paper.
age of social media (16, 21–23). The vast majority of Ameri- The authors declare no conflict of interest.
cans now visit a social media site at least once each day, and a This article is a PNAS Direct Submission.
rapidly growing number of them list social media as their primary
This open access article is distributed under Creative Commons Attribution-
source of news (24). Despite initial optimism that social media NonCommercial-NoDerivatives License 4.0 (CC BY-NC-ND).
might enable people to consume more heterogeneous sources Data deposition: All data, code, and the markdown file used to create this report
of information about current events, there is growing concern will be available at this link on the Dataverse: https://dataverse.harvard.edu/dataverse.
that such forums exacerbate political polarization because of xhtml?alias=chris bail.
social network homophily, or the well-documented tendency of 1
To whom correspondence should be addressed. Email: christopher.bail@duke.edu.y
people to form social network ties to those who are similar to This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
themselves (25, 26). The endogenous relationship between social 1073/pnas.1804840115/-/DCSupplemental.
network formation and political attitudes also creates formidable

www.pnas.org/cgi/doi/10.1073/pnas.1804840115 PNAS Latest Articles | 1 of 6

same type of positive mutual understanding—or whether the in the following paragraph and thereby the overall likelihood of
relative anonymity of social media forums emboldens people opinion change.
to act in an uncivil manner. Such incivility could be partic- We received 1,652 responses to our pretreatment survey (901
ularly rife in the absence of facial cues and other nonverbal Democrats and 751 Republicans). One week later, we randomly
gestures that might prevent the escalation of arguments in offline assigned respondents to a treatment condition, thus using an
settings. “ostensibly unrelated” survey design (43). At this time, respon-
Our second hypothesis builds upon a more recent wave of dents in the treatment condition were offered $11 to follow a
studies that suggest exposure to those with opposing politi- Twitter bot, or automated Twitter account, that they were told
cal views may create backfire effects that exacerbate political would retweet 24 messages each day for 1 mo. Respondents
polarization (34–37). This literature—which now spans sev- were not informed of the content of the messages the bots would
eral academic disciplines—indicates people who are exposed retweet. As Fig. 2 illustrates, we created a liberal Twitter bot and
to messages that conflict with their own attitudes are prone to a conservative Twitter bot for each of our experiments. These
counterargue them using motivated reasoning, which accentu- bots retweeted messages randomly sampled from a list of 4,176
ates perceived differences between groups and increases their political Twitter accounts (e.g., elected officials, opinion lead-
commitment to preexisting beliefs (34–37). Many studies in this ers, media organizations, and nonprofit groups). These accounts
literature observe backfire effects via survey experiments where were identified via a network-sampling technique that assumes
respondents are exposed to information that corrects factual those with similar political ideologies are more likely to follow
inaccuracies—such as the notion that Saddam Hussein possessed each other on Twitter than those with opposing political ideolo-
weapons of mass destruction prior to the 2003 US invasion of gies (44). For further details about the design of the study’s bots,
Iraq—although these findings have failed to replicate in two please refer to SI Appendix.
recent studies (38, 39). Yet our study is not designed to evaluate To monitor treatment compliance, respondents were offered
attempts to correct factual inaccuracies. Instead, we aim to assess additional financial incentives (up to $18) to complete weekly
the broader impact of prolonged exposure to counterattitudinal surveys that asked them to answer questions about the content
messages on social media. of the tweets produced by the Twitter bots and identify a picture
Our third preregistered hypothesis is that backfire effects of an animal that was tweeted twice a day by the bot but deleted
will be more likely to occur among conservatives than liberals. immediately before the weekly survey. At the conclusion of the
This hypothesis builds upon recent studies that indicate con- study period, respondents were asked to complete a final survey
servatives hold values that prioritize certainty and tradition, with the same questions from the initial (pretreatment) survey.
whereas liberals value change and diversity (40, 41). We also Of those invited to follow a Twitter bot, 64.9% of Democrats and
build upon recent studies in cultural sociology that examine 57.2% of Republicans accepted our invitation. Approximately
the deeper cultural schemas and narratives that create and sus- 62% of Democrats and Republicans who followed the bots were
tain such value differences (34, 26). Finally, we also build upon able to answer all substantive questions about the content of mes-
studies that observe asymmetric polarization in roll call voting sages retweeted each week, and 50.2% were able to identify the
wherein Republicans have become substantially more conserva- animal picture retweeted each day.
tive whereas Democrats exhibit little or no increase in liberal
voting positions (42). Although a number of studies have found Results
evidence of this trend, we are not aware of any that examine Fig. 3 reports the effect of being assigned to the treatment condi-
such dynamics among the broader public—and on social media tion, or the Intent-to-Treat (ITT) effects, as well as the Complier
in particular. Average Causal Effects (CACE) which account for the differen-
tial rates of compliance among respondents we observed. These
Research Design estimates were produced via multivariate models that predict
Fig. 1 provides an overview of our research design. We hired respondents’ posttreatment scores on the liberal/conservative
a professional survey firm to recruit self-identified Republi- scale described above, controlling for pretreatment scores on this
cans and Democrats who visit Twitter at least three times each scale as well as 12 other covariates described in SI Appendix.
week to complete a 10-min survey in mid-October 2017 and We control for respondents’ pretreatment liberal/conservative
1.5 mo later. These surveys measure the key outcome vari- scale score to mitigate the influence of period effects. Negative
able: change in political ideology during the study period via a scores indicate respondents became more liberal in response to
10-item survey instrument that asked respondents to agree or treatment, and positive scores indicate they became more con-
disagree with a range of statements about policy issues on a servative. Circles describe unstandardized point estimates, and
seven-point scale (α = .91) (10). Our survey also collected infor- the horizontal lines in Fig. 3 describe 90% and 95% confidence
mation about other political attitudes, use of social media and intervals. We measured compliance with treatment in three ways.
conventional media sources, and a range of demographic indi- “Minimally Compliant Respondents” describes those who fol-
cators that we describe in SI Appendix. Finally, all respondents lowed our bot throughout the entire study period. “Partially
were asked to report their Twitter ID, which we used to mine Compliant Respondents” are those who were able to answer
additional information about their online behavior, including at least one—but not all—questions about the content of one
the partisan background of the accounts they follow on Twit- of the bots’ tweets administered each week during the survey
ter. Our research was approved by the Institutional Review period. “Fully Compliant Respondents” are those who success-
Boards at Duke University and New York University. All respon- fully answered all of these questions. These last two categories
dents provided informed consent before participating in our are mutually exclusive.
research. Although treated Democrats exhibited slightly more liberal
We ran separate field experiments for Democratic and Repub- attitudes posttreatment that increase in size with level of compli-
lican respondents, and, within each group, we used a block ran- ance, none of these effects were statistically significant. Treated
domization design that further stratified respondents according Republicans, by contrast, exhibited substantially more conserva-
to two variables that have been linked to political polarization: tive views posttreatment. These effects also increase with level
(i) level of attachment to political party and (ii) level of interest of compliance, but they are highly significant. Our most cautious
in current events. We also randomized assignment according to estimate is that treated Republicans increased 0.12 points on a
respondents’ frequency of Twitter use, which we reasoned would seven-point scale, although our model that estimates the effect of
influence the amount of exposure to the intervention we describe treatment upon fully compliant respondents indicates this effect

2 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1804840115 Bail et al.

 Initial Survey Randomization Weekly Surveys Post-Survey

Respondents were offered $11 One week later, respondents Respondents in treatment Respondents were
to provide their Twitter ID and were assigned to treatment conditions informed they are offered $12 to repeat
complete a 10-minute survey and control conditions within eligible to receive up to $6 each the pre-treatment
strata created using pre- week during the study period
about their political attitudes, survey one month
treatment covariates that for correctly answering
social media use, and describe attachment to party, questions about the content of after initial survey.
media consumption habits frequency of Twitter use, messages retweeted by Twitter
(demographics provided by and overall interest in .Bots.
survey firm). current events.

Treatment Treatment

SOCIAL SCIENCES
Offered $11 to
follow Twitter
bot that retweets
24 messages from
Republicans liberal accounts each day for
1 month

Control Control

Treatment Treatment
Offered $11 to
follow Twitter
bot that retweets
24 messages from
conservative accounts each day for
Democrats 1 month

Control Control
Fig. 1. Overview of research design.

is substantially larger (0.60 points). These estimates correspond Discussion and Conclusion
to an increase in conservatism between 0.11 and 0.59 standard Before discussing the implications of these findings, we first note
deviations. important limitations of our study. Readers should not interpret

Bail et al. PNAS Latest Articles | 3 of 6

 Elected Officials Presidential Candidates
Collect Twitter handles

Lisa Murkowski (R-AK) @lisamurkowski Ben Carson @RealBenCarson and presidential

Don Young (R-AK) @repdonyoung Hillary Clinton @HillaryClinton candidates.
Jon Tester (D-MT) @SenatorTester Carly Fiorina @CarlyFiorina
Steve Daines (R-MT) @stevedaines Lawrence Lessig @Lessig
Mike Enzi (R-WY) @SenatorEnzi Martin O’Malley @martinomalley
John Barrasso (R-WY) @SenJohnBarrasso Donald Trump @realDonaldTrump
...etc ...etc ...etc ...etc

Extract the names of

2
Hillary Clinton Lisa Murkowski Steve Daines Donald Trump all Twitter accounts
that these 563 elected
Tim Kaine
Mike Pence candidates follow
Mike Pence
Ivanka Trump (n=636,738).
Sarah Sanders
Mike Pence
Tim Kaine Mike Pence
Sarah Sanders Sarah Sanders
Ivanka Trump

(Small Network Component Pictured)

3 Create directed
network of all elected
Tim Kaine
candidates, and
everyone they follow;
Lisa Murkowski
Hillary Clinton dropping non-elected

less than 15 as well as

Mike Pence Twitter accounts from
Planned U.S. government agencies,
CNN Heritage
Parenthood
Foundation and accounts that
originate outside the U.S.
(n=4,176).
(Small Network Component Pictured)

Sarah Sanders
Ivanka Trump FOX Steve Daines
4 Create adjacency
matrix that describes
following patterns of
the 4,176 “opinion
v ....
leaders” and conduct
Tucker Carlson
0 0 1
Correspondence
Donald Trump Analysis. Adjust
scores of accounts with
A= 1 0 1
large no. of followers
(see Supp. Materials).
0 1 0
....

Liberal/Conservative Scale
5
component to create
liberal/conservative
0 .14 .28 .42 .57 .71 .85 ideology score for
4,176 opinion leaders.
Liberal Conservative
Bot #1 Bot #2 Create bots that
6
tweet a random
sample of tweets
from the 1-3 (liberal)
and 5-7 (conservative)
quantiles of the
distribution .

Fig. 2. Design of study’s Twitter bots.

our findings as evidence that exposure to opposing political views shape political polarization in other parts of the world. In addi-
will increase polarization in all settings. Although ours is among tion, we did not study people who identify as independents, or
the largest field experiments conducted on social media to date, those who use Twitter but do so infrequently. Such individuals
the findings above should not be generalized to the entire US might exhibit quite different reactions to an intervention such
population, because a majority of Americans do not use Twit- as ours. Future studies are needed to further evaluate the exter-
ter (24). It is also unclear how exposure to opposing views might nal validity of our findings, because we offered our respondents

4 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1804840115 Bail et al.

 More Liberal More Conservative More Liberal More Conservative

Fig. 3. Effect of following Twitter bots that retweet messages by elected officials, organizations, and opinion leaders with opposing political ideologies for
1 mo, on a seven-point liberal/conservative scale where larger values indicate more conservative opinions about social policy issues, for experiments with
Democrats (n = 697) and Republicans (n = 542). Models predict posttreatment liberal/conservative scale score and control for pretreatment score on this

SOCIAL SCIENCES
scale as well as 12 other covariates described in SI Appendix. Circles describe unstandardized point estimates, and bars describe 90% and 95% confidence
intervals. “Respondents Assigned to Treatment” describes the ITT effect for Democrats (ITT = −0.02, t = −0.76, p = 0.45, n = 416) and Republicans (ITT = 0.12,
t = 2.68, p = 0.008, n = 316). “Minimally-Compliant Respondents” describes the CACE for respondents who followed one of the study’s bots for Democrats
(CACE = −0.04, t = −0.75, p = 0.45, n of compliant respondents = 271) and Republicans (CACE = 0.19, t = 2.73, p < 0.007, n of compliant respondents =
181). “Partially-Compliant Respondents” describes the CACE for respondents who correctly answered at least one question, but not all questions, about the
content of a bot’s tweets during weekly surveys throughout the study period for Democrats (CACE = −0.05, t = −0.75, p = 0.45, n of compliant respondents =
211) and Republicans (CACE = 0.31, t = 2.73, p <.007, n of compliant respondents = 121). “Fully-Compliant Respondents” describes the CACE for respondents
who answered all questions about the content of the bot’s tweets correctly for Democrats (CACE = −0.14, t = −0.75, p = 0.46, n of compliant respondents =
66) and Republicans (CACE = 0.60, t = 2.53, p < 0.01, n of compliant respondents = 53). Although treated Democrats exhibited slightly more liberal attitudes
posttreatment that increase in size with level of compliance, none of these effects were statistically significant. In contrast, treated Republicans exhibited
substantially more conservative views posttreatment that increase in size with level of compliance, and these effects are highly significant.

financial incentives to read messages from people or organiza- psychology, communications, and information science. Although
tions with opposing views. It is possible that Twitter users may we found no evidence that exposing Twitter users to opposing
simply ignore such counterattitudinal messages in the absence views reduces political polarization, our study revealed signif-
of such incentives. Perhaps the most important limitation of our icant partisan differences in backfire effects. This finding is
study is that we were unable to identify the precise mechanism important, since our study examines such effects in an exper-
that created the backfire effect among Republican respondents imental setting that involves repeated contact between rival
reported above. Future studies are thus urgently needed not only groups across an extended time period on social media. Our
to determine whether our findings replicate in different popula- field experiment also disrupts selective exposure to informa-
tions or within varied social settings but to further identify the tion about politics in a real-world setting through a combina-
precise causal pathways that create backfire effects more broadly. tion of survey research, bot technology, and digital trace data
Future studies are also needed because we cannot rule out collection. This methodological innovation enabled us to col-
all alternative explanations of our findings. In SI Appendix, we lect information about the nexus of social media and politics
present additional analyses that give us confidence that our results with high granularity while developing techniques for measuring
are not driven by Hawthorne effects, partisan “learning” pro- treatment compliance, mitigating causal interference, and veri-
cesses, variation in the ideological extremity of messages by party, fying survey responses with behavioral data—as we discuss in SI
or demographic differences in social media use by age. At the Appendix. Together, we believe these contributions represent an
same time, we are unable to rule out other alternative explana- important advance for the nascent field of computational social
tions discussed in SI Appendix. For example, it is possible that our science (46).
findings resulted from increased exposure to information about Although our findings should not be generalized beyond party-
politics, and not exposure to opposing messages per se. Similarly, identified Americans who use Twitter frequently, we note that
increases in conservatism among Republicans may have resulted recent studies indicate this population has an outsized influence
from increased exposure to women or racial and ethnic minori- on the trajectory of public discussion—particularly as the media
ties whose messages were retweeted by our liberal bot. Finally, itself has come to rely upon Twitter as a source of news and a
our intervention only exposed respondents to high-profile elites window into public opinion (47). Although limited in scope, our
with opposing political ideologies. Although our liberal and con- findings may be of interest to those who are working to reduce
servative bots randomly selected messages from across the liberal political polarization in applied settings. More specifically, our
and conservative spectrum, previous studies indicate such elites study indicates that attempts to introduce people to a broad
are significantly more polarized than the general electorate (45). range of opposing political views on a social media site such as
It is thus possible that the backfire effect we identified could be Twitter might be not only be ineffective but counterproductive—
exacerbated by an antielite bias, and future studies are needed to particularly if such interventions are initiated by liberals. Since
examine the effect of online intergroup contact with nonelites. previous studies have produced substantial evidence that inter-
Despite these limitations, our findings have important impli- group contact produces compromise and mutual understanding
cations for current debates in sociology, political science, social in other contexts, however, future attempts to reduce political

Bail et al. PNAS Latest Articles | 5 of 6

polarization on social media will most likely require learning replication materials, additional robustness checks, and an extended discus-
which types of messages, tactics, or issue positions are most likely sion of alternative explanations of our findings. Our research was approved
to create backfire effects and whether others—perhaps deliv- by the Institutional Review Boards at Duke University and New York
ered by nonelites or in offline settings—might be more effective University.
vehicles to bridge America’s partisan divides.
ACKNOWLEDGMENTS. We thank Paul DiMaggio, Sunshine Hillygus, Gary
King, Fan Li, Arthur Lupia, Brendan Nyhan, and Samantha Luks for helpful
Materials and Methods conversations about this study prior to our research. Our work was sup-
See SI Appendix for a detailed description of all materials and methods ported by the Carnegie Foundation, the Russell Sage Foundation, and the
used within this study as well as links to our preregistration statement, National Science Foundation.

1. DiMaggio P, Evans J, Bryson B (1996) Have American’s social attitudes become more 25. McPherson M, Smith-lovin L, Cook JM (2001) Birds of a feather: Homophily in social
polarized? Am J Sociol 102:690–755. networks. Annu Rev Sociol 27:415–444.
2. Iyengar S, Westwood SJ (2015) Fear and loathing across party lines: New evidence on 26. Edelmann A, Vaisey S (2014) Cultural resources and cultural distinction in networks.
group polarization. Am J Polit Sci 59:690–707. Poetics 46:22–37.
3. Baldassarri D, Gelman A (2008) Partisans without constraint: Political polarization and 27. Lazer D, Rubineau B, Chetkovich C, Katz N, Neblo M (2010) The coevolution of
trends in American public opinion. Am J Sociol 114:408–446. networks and political attitudes. Polit Commun 27:248–274.
4. Sides J, Hopkins DJ (2015) Political Polarization in American Politics (Bloomsbury, 28. Centola D (2011) An experimental study of homophily in the adoption of health
London). behavior. Science 334:1269–1272.
5. Baldassarri D, Bearman P (2007) Dynamics of political polarization. Am Sociol Rev 29. Vaisey S, Lizardo O (2010) Can cultural worldviews influence network composition?.
72:784–811. Social Forces 88:1595–1618.
6. Fiorina MP, Abrams SJ (2008) Political polarization in the American public. Annu Rev 30. Pettigrew TF, Tropp LR (2006) A meta-analytic test of intergroup contact theory. J
Polit Sci 11:563–588. Pers Soc Psychol 90:751–783.
7. DellaPosta D, Shi Y, Macy M (2015) Why do liberals drink lattes? Am J Sociol 120: 31. Huckfeldt R, Johnson PE, Sprague J (2004) Political Disagreement: The Survival
1473–1511. of Diverse Opinions Within Communication Networks (Cambridge Univ Press,
8. Levendusky M (2009) The Partisan Sort: How Liberals Became Democrats and Cambridge, UK).
Conservatives Became Republicans (Univ Chicago Press, Chicago). 32. Mutz DC (2002) Cross-cutting social networks: Testing democratic theory in practice.
9. Mason L (2018) Uncivil Agreement: How Politics Became Our Identity (Univ Chicago Am Polit Sci Rev 96:111–126.
Press, Chicago). 33. Grönlund K, Herne K, Setälä M (2015) Does enclave deliberation polarize opinions?
10. Dimock M, Carroll D (2014) Political polarization in the American public: How increas- Polit Behav 37:995–1020.
ing ideological uniformity and partisan antipathy affect politics, compromise, and 34. Bail C (2015) Terrified: How Anti-Muslim Fringe Organizations Became Mainstream
everyday life (Pew Res Cent, Washington, DC). (Princeton Univ Press, Princeton).
11. Achen CH, Bartels LM (2016) Democracy for Realists: Why Elections Do Not Produce 35. Lord CG, Ross L, Lepper MR (1979) Biased assimilation and attitude polarization:
Responsive Government (Princeton Univ Press, Princeton). The effects of prior theories on subsequently considered evidence. J Pers Soc Psychol
12. Erikson RS, Wright GC, McIver JP (1993) Public Opinion and Policy in the American 37:2098–2109.
States (Cambridge Univ Press, Cambridge, UK). 36. Nyhan B, Reifler J (2010) When corrections fail: The persistence of political
13. Fishkin JS (2011) When the People Speak: Deliberative Democracy and Public misperceptions. Polit Behav 32:303–330.
Consultation (Oxford Univ Press, Oxford). 37. Taber CS, Lodge M (2006) Motivated skepticism in the evaluation of political beliefs.
14. Bennett WL, Iyengar S (2008) A new era of minimal effects? The changing Am J Polit Sci 50:755–769.
foundations of political communication. J Commun 58:707–731. 38. Wood T, Porter E (2016) The elusive backfire effect: Mass attitudes’ steadfast factual
15. Sunstein C (2002) Republic.com (Princeton Univ Press, Princeton). adherence. Polit Behav, 1–29.
16. Bakshy E, Messing S, Adamic LA (2015) Political science. Exposure to ideologically 39. Wood T, Porter E (2018) The elusive backfire effect: Mass attitudes’ steadfast factual
diverse news and opinion on Facebook. Science 348:1130–1132. adherence. Polit Behav, 10.1007/s11109-018-9443-y.
17. Sunstein C (2001) Echo Chambers, Bush v. Gore, Impeachment, and Beyond (Princeton 40. Graham J, Haidt J, Nosek BA (2009) Liberals and conservatives rely on different sets
Univ Press, Princeton). of moral foundations. J Pers Social Psychol 96:1029–1046.
18. King G, Schneer B, White A (2017) How the news media activate public expression 41. Jost JT, et al. (2007) Are needs to manage uncertainty and threat associated with
and influence national agendas. Science 358:776–780. political conservatism or ideological extremity? Pers Soc Psychol Bull 33:989–1007.
19. Berry JM, Sobieraj S (2013) The Outrage Industry: Political Opinion Media and the 42. Grossmann M, Hopkins DA (2016) Asymmetric Politics: Ideological Republicans and
New Incivility (Oxford Univ Press, Oxford). Group Interest Democrats (Oxford Univ Press, Oxford).
20. Prior M (2013) Media and political polarization. Annu Rev Polit Sci 16: 43. Broockman D, Kalla J (2016) Durably reducing transphobia: A field experiment on
101–127. door-to-door canvassing. Science 352:220–224.
21. Pariser E (2011) The Filter Bubble: How the New Personalized Web Is Changing What 44. Barberá P (2014) Birds of the same feather tweet together: Bayesian ideal point
We Read and How We Think (Penguin, New York). estimation using twitter data. Polit Anal 23:76–91.
22. Conover M, Ratkiewicz J, Francisco M (2011) Political polarization on twitter. ICWSM 45. Abramowitz AI, Saunders KL (2008) Is polarization a myth?. J Polit 70:542–555.
133:89–96. 46. Lazer D, et al (2009) Life in the network: The coming age of computational social
23. Boxell L, Gentzkow M, Shapiro JM (2017) Greater Internet use is not associated with science. Science 323:721–723.
faster growth in political polarization among us demographic groups. Proc Natl Acad 47. Faris R, et al. (2017) Partisanship, Propaganda, and disinformation: Online media and
Sci USA 114:10612–10617. the 2016 U.S. presidential election, (Berkman Klein Cent Internet Soc Harvard Univ,
24. Perrin A (2015) Social media usage: 2005-2015 (Pew Res Cent, Washington, DC). Cambridge, MA).

6 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1804840115 Bail et al.