Social Media and Elections

The Effect of Social Media on Elections:
Evidence from the United States*
Thomas Fujiwara† Karsten Müller‡ Carlo Schwarz§
July 2023
Abstract
We study how social media affects election outcomes in the United States. We use variation in the
number of Twitter users across counties induced by early adopters at the 2007 South by Southwest
(SXSW) festival, a key event in Twitter’s rise to popularity. We show that this variation is unrelated
to observable county characteristics and electoral outcomes before the launch of Twitter. Our results
indicate that Twitter lowered the Republican vote share in the 2016 and 2020 presidential elections,
but had limited effects on Congressional elections and previous presidential elections. Evidence from
survey data, primary elections, and text analysis of millions of tweets suggests that Twitter’s relatively
liberal content may have persuaded voters with moderate views to vote against Donald Trump.
* We are grateful to Pablo Barbera, Levi Boxell, Matt Gentzkow, Ro’ee Levy, Alexey Makarin, Maria Petrova, Jesse
Shapiro, James Snyder, Ekaterina Zhuravskaya, as well as seminar participants at Princeton University, Imperial College,
Warwick University, Bocconi University, University of California San Diego, Insper, University of São Paulo, Inter-American
Development Bank, IPEG/Pompeu Fabra, University of Zurich, OPESS, Università di Bergamo, the 2021 ASSA Conference,
NBER Political Economy Spring 2021 Meeting, NBER IT and Digitization Summer Institute 2021, and the ENS Lyon for
their helpful comments. We thank Pedro Cabra-Acela and Wenqing Yu for excellent research assistance. Müller acknowledges
support from a Singapore MoE Start-up/PYP grant (number A-0003319-01-00).
†
Princeton University, Department of Economics and SPIA, and NBER, fujiwara@princeton.edu.
‡
National University of Singapore, Business School, kmueller@nus.edu.sg.
§
Università Bocconi, Department of Economics and IGIER, and PERICLES, carlo.schwarz@unibocconi.it.
1 Introduction
Does social media affect election outcomes? A popular narrative holds that Twitter played a
decisive role in both recent American presidential elections and the United Kingdom’s “Brexit”
referendum. Many see this as part of social media’s broader influence on political polarization and
the re-emergence of populist politicians in many countries. The U.S. Federal Election Commissioner,
for example, has argued that Facebook “has no idea how seriously it is hurting democracy” (NPR,
2020a).1
An alternative view suggests that social media platforms are biased against conservatives (e.g.,
NPR, 2020b; Wall Street Journal, 2020) and that its younger, relatively left-leaning user base is
unlikely to tilt elections towards right-wing politicians (e.g., Boxell et al., 2017, 2018). However,
there is limited evidence that can be used to evaluate these contrasting (causal) claims.
This paper focuses on the effects of Twitter, a platform used by almost a quarter of American
adults. We estimate how a county’s number of Twitter users affects election results by exploiting a
persistent network effect sparked by early Twitter adoption, building on Müller and Schwarz (2019).2
Although it was launched in March 2006, Twitter’s popularity increased rapidly after its advertising
campaign at the South by Southwest festival (SXSW) in March 2007. The SXSW festival was also
key for Twitter’s geographical diffusion: counties with more SXSW followers who joined during
the 2007 festival saw disproportionately higher growth of Twitter adoption compared to counties
with SXSW followers who already joined before the festival. Consistent with path dependence in
technology adoption, this difference in Twitter use across counties persists.
Our identification strategy leverages the 2007 SXSW festival as a shock to early Twitter
adoption that is uncorrelated with pre-existing election results. Conditional on geographic controls
and previous interest in the SXSW Twitter account, a county’s number of SXSW followers who
joined in March 2007 is essentially uncorrelated with a host of county characteristics. It is also
unrelated to election outcomes before Twitter’s launch (going back as far as 1924) and during the
period it had fewer users (between 2006 and 2012). However, the number of SXSW followers who
joined in March 2007 is correlated with Twitter usage in 2016, and has predictive power for the
2016 and 2020 presidential election results.
1
See, for example, The New Yorker (2016); New York Times (2017); Allcott and Gentzkow (2017); The Guardian
(2018); UK Parliament (2019).
2
Enikolopov et al. (2020) use a similar empirical strategy based on spatial variation in early adopters of the social
media network VK in Russia to study its effects on protests.
1
We estimate that a 10% increase in a county’s number of Twitter users lowered the vote share
of Republican presidential candidate Donald Trump by 0.2 percentage points (p.p.) in both the
2016 and 2020 presidential elections. The implied persuasion rates are 8.6% and 9.4%, respectively.
These estimates are smaller than the estimated pro-Republican effect of Fox News (DellaVigna
and Kaplan, 2007; Martin and Yurukoglu, 2017), the pro-Democrat effect of the Washington Post
(Gerber et al., 2009), or the effect of get-out-the-vote canvassing on turnout (Gerber and Green,
2000), but larger than the effect of an independent anti-Putin Russian TV channel on vote shares
(Enikolopov et al., 2011) or the effect of TV rollout on turnout (Gentzkow, 2006).
For presidential elections before 2016, we find effects that are small and statistically indis-
tinguishable from zero. The same holds true for House and Senate races, including the 2016 and
2020 elections. We thus detect a negative effect of Twitter adoption on Trump’s vote share but
do not do so for Republican candidates in congressional races in the same election. This pattern
bolsters confidence that our estimates capture an effect of Twitter, which contains more content on
presidential than congressional candidates.
To shed light on the mechanisms behind these results, we estimate Twitter’s effect on vote
choices reported in the Cooperative Congressional Election Survey (CCES), primary presidential
candidates’ approval in the Gallup Daily Tracker, and county-level results in the 2016 and 2020
presidential primaries. Further, we explore data on the partisanship of political content on Twitter.
These exercises yield three findings. First, the CCES results indicate that Twitter’s effect
is driven by independents and moderates switching their votes towards the Democratic candidate
(Hillary Clinton). This is consistent with Bayesian updating, since moderates presumably have
weaker priors and are thus more likely to be persuaded.
Second, we find that Twitter also lowered Trump’s vote share during the 2016 primaries, a
finding we confirm using individual-level Gallup candidate approval ratings. We find that Twitter
decreased Trump’s approval ratings and increased Clinton’s with only small effects on relatively
more moderate Republican candidates.3
Third, we document that political content on Twitter has a pro-Democratic slant. We classify
the slant of tweets based on two complementary approaches: one based on the network users follow,
and one using the text of tweets in a machine-learning approach in the spirit of Gentzkow and
Shapiro (2010). We apply these methods to the over 460 million tweets mentioning the presidential
3
We also estimate effects for the 2016 and 2020 Democratic primaries, detecting a (positive) effect for Bernie
Sanders in 2020.
2
candidates in the 2012, 2016, and 2020 elections. We find that the number and attention (proxied
by “likes”) of tweets mentioning Trump was substantially larger than that of those mentioning
Clinton and Joe Biden. Moreover, tweets about Trump in 2016 and 2020 70% more likely to have
Democratic rather than Republican slant.
Overall, our results indicate an effect that is specific to Trump and not other Republican
candidates. One potential interpretation is that Trump’s own behavior on Twitter created a “backlash”
effect of moderate voters against him. Trump adopted Twitter as his preferred social media platform
and differed from most high-profile candidates in its use, both in the amount of tweeting and
also in its tone and content (Bursztyn et al., 2020). Another interpretation is that Twitter has
relatively pro-Democratic content persuading moderate voters about a high-profile candidate. These
two interpretations are not only complementary but also likely to feed on each other: Trump’s
tweets generate many other tweets that paint him in negative light (accusing of misinformation or
inappropriate speech), which can in turn have persuasive effects.
Our work contributes to the literature on the impact of media on political outcomes. Expansions
of traditional media such as newspapers, radio, broadcast television, and cable news have been
associated with changes in voter turnout, polarization, and electoral outcomes.4 While a set of papers
studies the effect of overall internet access, the effects of social media per se received less attention.5
A nascent literature studies the political effects of social media on protest participation
(Howard et al., 2011; Enikolopov et al., 2020; Acemoglu et al., 2017; Fergusson and Molina, 2021),
xenophobia (Müller and Schwarz, 2020; Müller and Schwarz, 2019; Bursztyn et al., 2019) and
mental health (Braghieri et al., 2021).6 Additionally, a burgeoning field of experimental research
focuses on social media. Bond et al. (2012) and Jones et al. (2017) provide evidence that online
messages on social networks affect voter turnout. Allcott et al. (2020) and Mosquera et al. (2020) find
that individuals who deactivate Facebook react along many dimensions, including some measures
of political polarization. Levy (2021) studies the effect of randomly assigning Facebook users
subscriptions to conservative or liberal media outlets. Bail et al. (2018) estimate the effect of paying
Twitter users to follow a bot with messages of the opposing political ideology. Bessone et al. (2022)
4
See, for example, Gentzkow (2006); Huber and Arceneaux (2007); DellaVigna and Kaplan (2007); Gerber et al.
(2009, 2011); Gentzkow et al. (2011); Enikolopov et al. (2011); Campante and Hojman (2013); DellaVigna et al. (2014);
Larcinese and Miner (2017); Martin and Yurukoglu (2017); Spenkuch and Toniatti (2018); Chen and Yang (2019).
5
There is evidence that broadband internet (Falck et al., 2014; Gavazza et al., 2019; Campante et al., 2017; Lelkes
et al., 2017) and mobile internet (Manacorda and Tesei, 2016; Guriev et al., 2020) exert political effects.
6
For reviews, see DellaVigna and Gentzkow (2010), Napoli (2014), Strömberg (2015), Enikolopov and Petrova
(2015), and DellaVigna and Ferrara (2015) and in particular Zhuravskaya et al. (2020) for the case of social media.
3
studies how Facebook affected Brazilian politicians’ behavior. Most related to our paper is recent
unpublished work by Rotesi (2019), who finds social media negatively affected the Democratic
vote share in the 2008-2016 presidential elections using variation in Twitter adoption resulting from
transfers of NBA players with Twitter accounts.7
Existing research thus provides an incomplete picture. On one hand, social media has been
painted as a key force behind political change, and experimental studies indeed suggest that social
media affects individuals’ self-reported political beliefs. On the other hand, it remains unclear
whether social media can indeed persuade voters and affect election results on a larger scale. Our
paper sheds light on this question by focusing on how Twitter affects federal elections in the United
States.
2 Background: Social Media and Politics

Most Americans use social media platforms or messaging applications. Data from the Pew Research
Center suggest that the most popular services are YouTube (used by 73% of adults in the U.S.),
followed by Facebook (69%), and Instagram (37%) (Pew Research Center, 2019c). 22% of adults
in the U.S. use Twitter, a rate similar to that of Snapchat (24%) and Whatsapp (20%) users. On
average, adult users spend more than an hour a day using social networks (eMarketer, 2019).8
One popular perspective is that online networks, and social media in particular, may give rise to
so-called “filter bubbles” (Pariser, 2011) or “echo chambers” (Sunstein, 2017). The idea is that social
media—unlike traditional mass media outlets—may facilitate the provision and consumption of
one-sided information, either through the use of algorithms or by allowing individuals to self-select
into preferred content. While there is considerable empirical evidence supporting this idea (e.g.
Conover et al., 2011; Weber et al., 2013; Bessi et al., 2015; Del Vicario et al., 2016; Halberstam
and Knight, 2016; Schmidt et al., 2017; Levy, 2021), other studies have found that individuals are
exposed to a wide range of political opinions on social media (Barberá, 2014; Bakshy et al., 2015;
Nelson and Webster, 2017; Beam et al., 2018), perhaps even more so than via traditional media
7
Our paper differs from Rotesi (2019) not only in its research design, but in that we estimate effects for a larger set
of election years (including 2020) as well as Congressional elections. Rotesi (2019) only reports estimates pooling the
2008-2016 presidential elections, differently from our paper. This does not allow to see if his instrument is uncorrelated
with election results before Twitter’s launch and how its possible effects evolved over time.
8
Pew bases its usage measures on the share of respondents who state they have ever used one of the online platforms.
Twitter reported around 69 million monthly active users in 2019 (see Statista, 2019), which yields a slightly higher share
of around a third of the 210 million adults in the U.S.
4
outlets or personal interactions (Gentzkow and Shapiro, 2011). Some work also challenges the
notion that increased polarization due to online channels is quantitatively important (Flaxman et al.,
2016; Guess, 2018; Boxell et al., 2019).
Much of the recent public discussion about the role of social media platforms has been shaped
by controversies, including Cambridge Analytica’s involvement in political campaigns (e.g., The
Guardian, 2018); the Russian Internet Research Agency’s efforts to support Trump’s campaign
(e.g., New York Times, 2017); and the role of misinformation (“fake news”) (e.g., Allcott and
Gentzkow, 2017). Both Clinton and Trump have argued that these factors were instrumental in the
2016 election outcome, as has former president Barack Obama (The New Yorker, 2016). As Brad
Parscale, Trump’s digital media director in 2016, put it: “Facebook and Twitter were the reason we
won this thing. Twitter for Mr. Trump. And Facebook for fundraising” (Wired, 2016). In Appendix
Figure A.3, we document that discussions of social media have become increasingly frequent in
major American news outlets.
Such controversies intensified in the aftermath of the 2020 election. Twitter permanently
suspended Trump’s account in the aftermath of the January 6, 2021, invasion of the Capitol. In
October 2022, Elon Musk took ownership of the platform, reversed Trump’s suspension, and released
what he labeled the “Twitter Files.” The journalists that first received the files argued they showed a
pattern of Twitter taking politically motivated decisions to aid Democrats’ electoral chances, such as
moderation of news stories related to Biden’s son’s “laptop controversies” and providing relatively
less visibility to conservative accounts (e.g., NPR, 2022). The files also documented that Twitter
regularly communicated with the FBI about addressing the spread of misinformation and dealing
with foreign influences. Others have argued that the “Twitter Files” themselves were a political tool
and, if anything, showed the platform’s desire to remain politically neutral and limit the spread of
misinformation (e.g., New York Magazine, 2022).
Under Musk’s leadership, Twitter implemented a series of changes to the platform, most
notably affecting which Tweets receive more visibility and what content creates ground for a
suspension for the platform. Many argued this further politicized the platform and moved the content
most users see towards a conservative slant (e.g., USA Today, 2023; The Guardian, 2023). At the
time of writing, the landscape of American social media is rapidly changing, with new entrants such
as Bluesky and Threads providing platform with user interfaces that are similar to Twitter.
These events highlight both the prominence of social media’s effects on political speech and the
perceived need for regulation over social media platforms’ moderation policies. However, whether
5
social media actually affects electoral outcomes is largely unknown, and some have suggested that
concerns about its effects may be overblown. As one example, in the 2016 presidential election,
Trump received fewer votes from demographic groups with higher propensities to use social media
or the internet more broadly (The Hill, 2016; Boxell et al., 2017, 2018). Indeed, Trump’s broadest
support came from older white voters without college education in rural areas, who are among the
least likely people to use social media actively (Hargittai, 2015; Pew Research Center, 2015, 2018).
These patterns seem difficult to square with the idea that online channels were an important driver
of the 2016 presidential election result, although such observations also do not rule this out.
Further, most social media users—particularly on Twitter—appear to be disproportionately
left-leaning. This is not surprising given that most Twitter users are relatively younger, educated,
and urban. While there appears to be a cluster of right-wing networks, Pew Research Center
(2019d) estimates that, in 2018, 60% of Twitter users identified as Democrat and only 35% as
Republican. Among Democrats, those on Twitter are considerably more liberal and focus less on
finding common ground with Republicans (Pew Research Center, 2020). In 2019, 26% of American
Twitter users followed Obama, and 19% followed Trump (Pew Research Center, 2019a). Survey
evidence suggests that 80% of Twitter content is produced by people who strongly disapprove of
Trump (Pew Research Center, 2019b). “Liberals” are also more likely to get political news on
Twitter or Facebook and follow more media and political accounts compared to “conservatives”
(Pew Research Center, 2014; Eady et al., 2019). Twitter and Reddit, which are often said to be
pro-Trump factors, were considerably more popular among Clinton supporters before the 2016
election (Hargittai, 2015). Although social media allows users to partially select which content they
see, Twitter content disproportionately leans toward the Democratic party.
We provide additional evidence for the composition of political content on Twitter by analyzing
the Twitter reach of Democratic and Republican politicians. We collected data on the Twitter
accounts of all Senators and House Representatives from the 110th to the 115th Congresses (2007-
2019). In Figure 1, we plot the average number of tweets and followers that members of each party
have on Twitter, as well as the average number of retweets and “likes” their tweets receive. The
patterns here again clearly indicate that Democratic politicians are more active on Twitter and have
larger follower bases than their Republican counterparts. Tweets by Democrats also receive, on
average, three times the number of “likes.”9
9
In Appendix Figure A.2, we confirm that these patterns are not driven by a small group of Congress members by
showing that they also hold when we compare the median Twitter reach of Democrats and Republicans.
6
Figure 1: Twitter Reach by Party
(a) Average # of tweets (per account) (b) Average # of followers (per account)
5,000 600,000
4,000
400,000
3,000
2,000
200,000
1,000
0 0
Democrat Republican Democrat Republican
(c) Average # of likes (per tweet) (d) Average # of retweets (per tweet)
800
400
600
300
400
200
200 100
0 0
Notes: This figure plots data on the Twitter reach of Congress members. The sample includes all 901 senators and
House representatives who were in office between 2007 and 2019 for whom we could identify a Twitter account.
For each account, we plot the average number of tweets and followers, and the average number of “likes” and
retweets of their tweets. Appendix Figure A.2 replicates the figure using medians instead of averages. The data
were collected from Twitter in November 2019.
3 Data
The main analysis is based on a county-level dataset on election outcomes, political opinions, and
Twitter use. It covers 3,065 counties in 48 states (we exclude Alaska and Hawaii) and the District of
Columbia (except in congressional elections). County-level election results are from Dave Leip’s
Atlas of U.S. Presidential Elections and the MIT Election Lab. We complement our analysis with
individual-level survey data on approval ratings from the Gallup Daily Tracker and voting data from
the Cooperative Congressional Election Study (CCES). Our measure of Twitter usage is derived
from an archive of 475 million geo-located tweets compiled by Kinder-Kurlanda et al. (2017). We
combine this with newly collected data on Twitter’s early adopters at the 2007 SXSW festival; data
on the Twitter activity of U.S. Congress members; and a large corpus of tweets related to the 2012,
2016, and 2020 presidential elections. Additional county characteristics were obtained from the U.S.
Census, the U.S. Religious Census, the American Community Survey (ACS), and the Bureau of
Labor Statistics (BLS). We describe the individual data sources in more detail below. Appendix
Table A.1 provides additional details and summary statistics.
Election Outcomes. We use county-level data on presidential election outcomes between 1924
and 2020 from Dave Leip’s Atlas of U.S. Presidential Elections. From the same source, we also
7
obtained county-level voting data for the Republican and Democratic primaries in 2016 and 2020.
We complement this with county-level results on Senate and House elections from the MIT Election
Lab for the 1996-2020 period. In all cases, we focus on two-party vote shares.10
Individual-Level Voting Decisions. The Cooperative Congressional Election Study (CCES) is a

nationwide survey that collects information on voter behavior in two waves (before and after the
election). We focus on votes for Trump and Clinton in 2016 and 2020. The CCES contains a rich
set of individual characteristics, including political affiliation, family income (in 12 bins), gender,
race, education (in 6 bins), marital status, age, and interest in the news. Table A.2 provides summary
statistics (weighted by sample weights). The CCES also uses administrative data on turnout records
to verify its respondents have voted.
Presidential Candidate Approval. The Gallup Daily Tracker provides individual-level survey
data for a sample of 1,000 individuals per day since 2009.11 During the 2016 presidential campaign,
it fielded survey items regarding the approval of Republican and Democratic presidential candidates.
This allows us to investigate Trump’s pre-election approval relative to other candidates (e.g., Clinton
or Ted Cruz). The data also include a rich set of individual characteristics, including political
affiliation, county of residence, income (in 10 bins), gender, race, marital status, age, and education
(in 6 bins). Table A.3 in the Appendix provides summary statistics.12
Twitter Usage. We construct a measure of county-level Twitter usage based on a sample of 475
million geo-coded tweets collected by Kinder-Kurlanda et al. (2017).13 The tweets were collected
between 2014 and 2015 using the Twitter Streaming API by selecting a geographic bounding box
around the mainland US. The Streaming API continuously returns all geo-located tweets within the
bounding box as long as the sample does not exceed 1% of all tweets. Information on a tweet’s
geo-location either come from the GPS coordinates of a mobile phone or from a WiFi/IP address in
case a computer is used. Both types of information allow for a precise assignment to a US county.
At the start of the data collection by Kinder-Kurlanda et al. (2017) in 2014, Twitter had not yet
10
While senatorial and presidential elections are decided at the state level and House elections at the congressional
district level, counties are usually smaller geographic units and far more numerous. Additionally, unlike congressional
districts, county boundaries are fixed over our sample period, allowing us to observe changes across years.
11
The Gallup Daily Tracker for the 2020 election is not available at the time of writing.
12
For some estimations, we also collapse responses about approval of Trump to the county level using weighted
averages based on the number of survey respondents in each county.
13
These data are available in the Gesis Datorium at https://datorium.gesis.org/xmlui/handle/10.7802/1166.
8
introduced the feature that allowed users to tag specific places in their tweets. This avoids collecting
geo-located tweets based on arbitrary decisions by users (e.g., tagging their holiday location) and
prevents users from tagging wrong locations (either intentionally or by accident).
The individual tweets from this dataset are already assigned to counties. Additionally, we
collected the underlying user profiles for each tweet in the database. This allows us to construct a
user-based measure by assigning users to the county from which they tweet most frequently. The
resulting measure, which we use throughout the paper, is a proxy for the number of Twitter users
per county, based on 3.7 million individual users (around 7% of the Twitter population in 2015).
Figure 2a plots the number of Twitter users per capita across counties. Each user profile further
provides us with a short biography and the date that each user joined Twitter. We use the join dates
to construct a time-varying proxy of Twitter usage based on how many of the Twitter users had
opened an account at each point in time.
The great advantage of this dataset is that it allows us to provide individual-level evidence for
the adoption of Twitter in a county following the SXSW festival and further compare user profiles
in different counties. The drawback of using geo-located Twitter data is that only a sub-sample
of tweets is geo-located. To overcome concerns of measurement error in our Twitter measure, we
validate the data in two ways. First, Appendix Figure A.1a shows that our Twitter usage measure’s
evolution closely tracks the number of daily Twitter users from Statista (2019), which were directly
obtained from the platform. Secondly, our measure of county-level Twitter usage also strongly
correlates with the number of Twitter users in a county based on the GfK Media Survey (see
Figure A.1b). Lastly, note that measurement error is less of a concern in our setting because we
largely rely on 2SLS estimation throughout the paper. We return to the discussion of measurement
error in Section 5.
Twitter Data for the South by Southwest Festival. We collected data for our instrument for
Twitter usage, based on early adoption during the SXSW festival, through the Twitter API. More
specifically, we scraped the account data for 658,240 users who followed the Twitter account of
SXSW Conference & Festivals (@SXSW) at the time of collection (January 2019). We assign these
users to counties based on the location people report in their user profile.14
A user profile contains the month and year that they joined Twitter, which allows us to
determine the number of SXSW followers in each county that joined Twitter in a particular month.
14
Of the 44,625 SXSW followers who joined between 2006 and 2008, we are able to geo-code 25,830 (58%).
9
The two key variables in our analysis are: i) the number of SXSW followers that joined Twitter in
the month of March 2007 and ii) the number of SXSW followers that joined Twitter during 2006
(the year the platform was launched). We refer to (ii) as the number of SXSW followers who joined
before the March 2007 festival. We also scraped the follower lists of SXSW followers who joined in
March 2007, which allows us to investigate the connections of Twitter users to the SXSW festival.
Further, we additionally collected tweets mentioning the festival, based on the term “SXSW,” as
well as a proxy for overall Twitter activity based on the 100 common English words. We use these
measures to document the SXSW festival’s impact on local Twitter adoption.15
Figure 2: Twitter Usage and Identifying Variation

(a) Twitter Usage per Capita (b) Identifying Variation
1st Quintile
SXSW followers March 2007
2nd Quintile
3rd Quintile SXSW followers pre-period
4th Quintile SXSW followers March 2007 & pre-period
5th Quintile No SXSW followers
Notes: These maps plot the proxy for social media usage based on data from Twitter and the identifying
variation of our instrument. Panel (a) plots quintiles of the number of Twitter users per capita. Panel (b) plots
the three types of counties relevant for our identification strategy: 1) the 47 counties with SXSW followers
that joined Twitter both in March 2007 and the “pre-period” (light red); 2) the 108 counties with SXSW
followers that joined in March 2007, but none in the “pre-period” (dark red); and 3) the 20 counties with
SXSW followers that joined in the “pre-period,” but none in March 2007 (blue).
Data on Political Twitter Activity. We scraped the tweets, user profiles, and followers of the
901 Senators and House Representatives from the 110th to 115th (2007-2019) Congress who have
Twitter accounts. This includes 424 Democrats and 465 Republicans.16 In total, the data contain
4,300,579 tweets, which we use to analyze the Twitter reach of Democratic and Republican Congress
members.
We complement this dataset with election-related tweets to shed light on the overall partisan
slant of Twitter activity during the 2012, 2016, and 2020 elections. For each election, we obtained the
universe of tweets mentioning the last name of a Democrat and Republican presidential candidates.17
15
Data on SXSW 2007 attendants (e.g., their county of residence) is not available, despite our efforts to obtain it from
the SXSW organizers on multiple occasions.
16
The remaining 12 politicians are either Independents or switched their party affiliation.
17
For the 2012 election, we use data collected by Diaz et al. (2016), comprising 24 million tweets containing either
“Obama” or “Romney” for the period from July 1, 2012 through November 7, 2012. For 2016, we use the archive from
10
To determine the likely political affiliation of Twitter users, we create two measures of political
slant. The first measure is based on the political accounts a user is following. In particular, we check
whether a user follows more Democrat or Republican Congress members on Twitter. If they follow
more Republican than Democrats, all their tweets would be classified as Republican. In case a user
either does not follow any Congress members or an equal number of Congress members from either
party, their tweets are classified as neutral.18
The second measure of political slant is based on the similarity of the text of tweets to those
sent by Republican or Democratic Congress members. We train a L2 regularized logistic regression
model separately for each election based on 901 Congress members’ Twitter accounts to classify
whether a tweet contains language frequently used by either Republican or Democratic politicians.19
We then use this classifier to predict a partisan score between 0 and 1 for each of our election-related
tweets. These scores can be interpreted as the probability of a tweet with the same content being sent
by a Republican. As such, our approach is similar to how Gentzkow and Shapiro (2010) measure
newspaper slant. Both approaches lead to similar overall slant classifications for the election tweets
in our data.
Additional County Characteristics We collect county-level demographic control variables from

the U.S. Census and the ACS. In particular, we use information on population, population share
by age group and ethnicity, poverty rates, and education levels. We also obtained industry-level
employment shares and unemployment rates from the BLS. Additional controls on county media
usage patterns are from Simply Analytics. We also construct geographical controls such as the
distance from Austin, TX, where SXSW takes place every year; population density; and county size
(in square miles). For one set of results, we also use donation data from OpenSecrets.
Littman et al. (2016), which contains 280 million tweets, collected between July 13, 2016, and November 10, 2016. The
2020 election tweets are based on the archive from Chen et al. (2020), which covers the period from March 2020 to
November 2020. To make these datasets comparable, we restrict the 2016 election sample to tweets mentioning either
“Clinton” or “Trump” (112 million tweets). Similarly, we restrict the 2020 data set to the time period from July 1, 2020
through November 3, 2020 and tweets mentioning either “Biden” or “Trump” (339 million tweets).
18
The idea of using the Twitter network to determine a user’s ideology is inspired by Barberá (2015).
19
We clean the text of the tweets by removing common words (stopwords) and by reducing the words in each tweets
to their morphological roots (lemmatizing). The input is based on unigrams, bigrams, and trigrams from these tweets.
We choose the optimal normalization strength using 10-fold cross-validation. The resulting classifier achieves high
out-of-sample F1-scores, e.g. 0.904 for the tweets during the 2020 presidential election. We provide additional details
regarding the machine learning classifier in Online Appendix A.1., which also visualizes the most predictive terms
identified by the classifiers.
11
4 The 2007 South by Southwest Festival and Early Twitter Adop-
tion
The empirical strategy behind our main results exploits a shock to early-stage Twitter adoption
connected to the 2007 SXSW festival, as in Müller and Schwarz (2019). This section discusses
the key role of the festival in boosting the platform’s popularity and documents how it created a
persistent effect on its spatial diffusion.20
Founded in March 2006, Twitter was largely unknown before SXSW 2007. Twitter’s popularity
increased dramatically after the festival, where Twitter strategically placed screens in the conference
hallways and allowed users to sign-up by simply sending a text message to a predefined number.
As a result, speakers and bloggers in attendance broadcasted the platform to the outside world, and
Twitter went on to win the South by Southwest Interactive Web Award Prize.
The importance of SXSW 2007 has also been stressed by the platform’s founders. As co-
founder Evan Williams explained in a post on Quora (Quora, 2011):
“We didn’t actually launch Twitter at SXSW – SXSW just chose to blow it up. We
launched it nine months before – to a whimper. By the time SXSW 2007 rolled around,
we were starting to grow finally, and it seemed like all of our users (which were probably
in the thousands) were going to Austin that year ... I don’t know what was the most
important factor, but networks are all about critical mass, so doubling down on the
momentum seemed like a good idea. And something clicked.”21
SXSW’s immediate impact on Twitter’s popularity in early 2007 can be seen in Figure 3a, which
plots our proxy for the daily number of tweets as well as the number of tweets explicitly mentioning
SXSW. The figure shows that Twitter’s growth rate accelerated during the festival, visible as the
spike in SXSW-related tweets. The month-to-month growth rate of Twitter quadrupled with the start
of the SXSW festival.22 After SXSW 2007, Twitter experienced further rapid growth (Venture Beat,
2008). The platform went from an average of 5,000 tweets a day in 2007 to 300,000 in 2008, and
2.5 million in 2009 (Twitter, 2010). In 2019, users sent roughly 500 million tweets a day.
20
SXSW is an annual conglomeration of parallel film, interactive media, and music festivals and conferences organized
jointly that take place in March in Austin, TX.
21
Appendix Figure B.1 provides Williams’ full post describing the role of SXSW 2007.
22
Our proxy for Twitter usage is created by scraping tweets that contain any of the 100 most common English words.
Our data contain any tweet that contains at least one of these words. We should therefore obtain a large fraction of the
English-speaking tweets at that point in time.
12
Figure 3: South by Southwest (SXSW) 2007 and the Spread of Twitter
(a) Twitter Activity Around SXSW 2007 (b) SXSW and Local Twitter Adoption
400 40,000 0.15 SXSW festival starts
SXSW tweets (left axis) All tweets (right axis)
Change in Twitter activity (in SD)

0.10
300 30,000
0.05
200 20,000
0.00
100 10,000
-0.05
9
10
11
12
13
14
ek
ek
ek
ek
ek
ek
ek
ek
ek
ek
ek
ek
we
we
we
we
we
we
we
we
we
we
we
we
0 0
07
07
07
07
07
07
07
07
07
07
07
07
20
20
20
20
20
20
20
January February March April May
20
20
20
20
20
Notes: Panel (a) plots the total number of tweets and the number of tweets containing the term “SXSW” over
time, smoothed using a P P estimates of βτ from the panel event study
7-day moving average. Panel (b) plots the
regression tweetsct = τ βτ SXSWcM arch2007 ×1(t = τ )+ τ δτ SXSWcP re ×1(t = τ )+θc +γt +εct
where tweetsct is the log of (one plus) the number of tweets in county c on week t, SXSWcM arch2007 is
the logarithm of (one plus) the number of SXSW followers in county c that joined Twitter on March 2007
and SXSWcP re is a similarly defined variable for followers that joined Twitter before March 2007. We
standardize the variables to have a mean of zero and standard deviation of one. The whiskers represent 95%
confidence intervals based on standard errors clustered by state.
We exploit that the SXSW festival had persistent effects on Twitter’s spatial diffusion. This is
likely the result of network effects that are key to the social media experience, as a larger number of
users makes it more interesting for potential new users to join. Such a mechanism also applies at the
local level. For example, a boost in the number of neighbors, personal connections, local businesses
and/or people who play a prominent role in an area should also boost the value of joining the
platform for those living there. As Evan Williams’ quote above notes, “networks are all about critical
mass,” and initial differences in adoption can lead to persistent differences in network adoption.
Enikolopov et al. (2020) document a similar mechanism for spatial dispersion in the adoption of the
Russian VK social network: the home towns of the first university students invited to be users in
2006 exhibited a higher number of users in 2011.
We provide further support for this hypothesis by investigating whether the inflow of early-
stage adopters put these counties on a differential growth path of Twitter usage. Figure 3b plots the
estimates of βτ from the following panel event study regression at the county (c) and week (t) level:
X X
tweetsct = βτ SXSWcM arch2007 × 1(t = τ ) + δτ SXSWcP re × 1(t = τ ) + θc + γt + εct .
τ τ
13
where tweetsct is the log of (one plus) the number of tweets in county c on week t, SXSWcM arch2007
is the logarithm of (one plus) the number of SXSW followers in county c that joined Twitter on
March 2007 and SXSWcP re is a similarly defined variable for followers that joined Twitter before
March 2007. βτ thus illustrates, conditional on county and week fixed effects, the difference in the
number of tweets sent from counties with relatively larger numbers of SXSW followers that joined
on March 2007. The variables are standardized to have a mean of zero and standard deviation of
one. The whiskers represent 95% confidence intervals based on standard errors clustered at the state
level. The sample includes the period between the third and fourteenth week of 2007.
Figure 3b illustrates that home counties of SXSW followers who joined during the festival
in March 2007 saw a rapid, disproportionate increase in Twitter usage around the time of SXSW.
Importantly, however, this increase came only after the SXSW festival, and we find no evidence
for pre-existing trends. In Appendix Figure B.2 we show further that, in line with our expectations,
March 2007 is also a clear outlier when it comes to the number of people who started following the
SXSW festival. This is consistent with the idea that SXSW was a catalyst for the spread of Twitter
in the United States.
Appendix Figure B.3a presents additional evidence on the long-term adoption effect of the
2007 SXSW festival. It plots estimates from a similar regression as the one in Figure 3b but in a
county-quarter panel covering the period from Twitter’s launch in 2006 to 2016. The dependent
variable is substituted by the number of Twitter users per capita in a county based on our baseline
measure. The resulting S-shaped pattern in the figure is consistent with models of technology
adoption in the presence of network effects. More importantly, we find that the amount of early
adopters in a county still matters for the amount of Twitter usage today.23
5 Empirical Framework
Our identification strategy leverages the 2007 SXSW festival as a shock to early Twitter adoption.
We show that, conditional on a set of controls (described in further detail below), a county’s number
of SXSW followers that joined Twitter in March 2007 is uncorrelated with levels and trends in
election outcomes before Twitter’s launch and during its early years. It is also uncorrelated with a
23
Additionally, Figure B.3b shows just how dominant Twitter users connected to the SXSW festival were among early
adopters. In 2007, we estimate that around 60% of Twitter users either followed the SXSW festival or followed someone
who followed SXSW and joined in March 2007. As the number of Twitter users increased over time, the importance of
SXSW followers in the platform declined. But as Figure B.3a shows, the festival created persistent differences at the
county level. The next section outlines how we use the SXSW festival in our 2SLS estimates.
14
host of observable county characteristics. This feature of the data can be interpreted as idiosyncratic
factors (e.g., who attended the 2007 SXSW, who decided to join Twitter at the time), giving us a
“natural experiment” or “exogenous shock” in Twitter adoption that allows to estimate its effect on
election outcomes. This interpretation is, of course, not self-evident, and we provide several pieces
of evidence to support it.
An important concern is that counties whose populations are more interested in the SXSW
festival (and its Twitter account) may be systematically different from other counties. To address this
issue, our empirical strategy exploits variation in the exact timing of when Twitter users interested
in SXSW joined the platform across counties. In particular, our regressions control for the number
of SXSW followers who joined in the months before the festival. Intuitively, our empirical strategy
compares a “treatment” group of counties with SXSW followers that joined in March 2007 (during
the festival) against a “control” group of counties with followers that joined before. While both
groups of followers were interested in SXSW, we show that only the number of followers that joined
on March 2007 are predictive of later Twitter adoption, consistent with the evidence that users that
joined during the festival were key in the platform’s diffusion. In contrast, counties with more users
that joined before the festival do not have more additional Twitter users in subsequent years.24
The “treatment” and “control” counties are similar along several characteristics. Table A.4
compares the average characteristics of three types of counties relevant for our identification strategy:
1) the 47 counties with SXSW followers that joined Twitter both in March 2007 and the “pre-
period;” 2) the 108 counties with SXSW followers that joined in March 2007 (but none in the
“pre-period”’); and 3) the 20 counties with SXSW that joined in the “pre-period” (but none in March
2007). Differences in vote shares in the 1996 presidential election, demographics (e.g., race, age,
education), and media consumption (e.g., share that watches Fox News) are quantitatively small or
zero. This is particularly true for groups (2) and (3) — which are key to the identification — with
t-tests indicating that differences between the two groups are not statistically different from zero.25
The geographical variation in the three groups of counties is shown in Figure 2b. As the results in
Table A.4 suggest, the counties do not differ systematically in size and how distant they are from
major American cities.
24
An alternative approach is to compare the counties of users who signed up for Twitter during SXSW 2007 with
those of users who signed up during other festivals in the same year. We discuss the results from such an exercise in the
robustness section below.
25
Given the large number of county characteristics, we report Šidàk-corrected t-statistics, which are smaller than
those generated by applying the Bonferroni correction.
15
Moreover, observable individual characteristics of SXSW followers who joined Twitter in
March 2007 and the “pre-period” are also similar. We validate this using data on Twitter user profiles
we obtained from the platform. Table B.1 shows that followers who joined in March 2007 have
similar first names and profile descriptions compared to those that joined before: users in both
groups tend to have common names (e.g., “Michael” or “Chris”) and use words such as “founder” or
“tech” to describe themselves in their profiles. The correlations of the frequency of first names and
terms used in their bios between the two groups are 0.63 and 0.89, respectively. We also investigate
differences in the political leanings of the two groups using the network-based methods we outline
in Section 3. In particular, we test whether the users in March 2007 follow more Democrats or
Republicans than the users in the “pre-period”. We find that the political leanings of the two groups
are nearly identical. A t-test rejects differences in the average political slant with a p-value of 0.93.
Specification. Motivated by the evidence above, our main results are based on estimating the
following two equations:
T witter usersc = α + β · SXSWcM arch2007 + γ · SXSWcP re + Xc δ + ξc (1)
yc = α0 + β 0 · SXSWcM arch2007 + γ 0 · SXSWcP re + Xc δ 0 + ζc , (2)
where c indexes counties, SXSWcM arch2007 is the logarithm of (one plus) the number of SXSW
followers in county c that joined Twitter on March 2007, and SXSWcP re is a similarly defined
variable for followers who joined Twitter before March 2007. Xc is a vector of control variables that
hold constant geographical factors (e.g., population density, distance from Austin, TX), demographic
factors (e.g., the share of the population in different age and ethnic groups), socioeconomic factors
(e.g., the share of adults with a high school degree or employed in IT), factors related to the “China
shock” (e.g., exposure to Chinese import competition), and previous election results. Note that
the right-hand side of both equations is similar. T witter usersc is the logarithm of the number of
Twitter users in the county (during 2014-2015). yc are election outcomes (e.g., vote shares), which
we estimate in both levels and changes (e.g., yc can be the vote share in 2016 or the change in vote
shares between 2000 and 2016).
16
In a 2SLS framework, equations (1) and (2) are the first-stage and reduced form, while the
second stage is
yc = φ + θ · T witter
\ usersc + π · SXSWcP re + Xc ρ + εc , (3)
where T witter
\ usersc is predicted from the first stage regression in equation (1). We weigh
observations by turnout (total number of votes cast) in the 2000 presidential election.26 We cluster
standard errors at the state level.27
Identification. Formally, the identification condition for the effect of Twitter users (θ) is that
E(SXSWcM arch2007 · εc ) = 0 holds. Intuitively, this states that, conditional on the SXSWcP re and
other controls (Xc ), the number of SXSW followers who joined in March 2007 is uncorrelated with
other determinants of political outcomes yc , implying that it only affects political outcomes via
Twitter usage (the “exclusion restriction”).
We provide six pieces of evidence in support of this condition. First, as discussed above,
including the SXSWcP re control implies that the identifying variation comes from comparing
counties with similar observable characteristics.
Second, the coefficient of SXSWcP re is small and statistically insignificant in our first stage
regressions. This provides us with a “placebo” test based on checking if it is also unrelated to
political outcomes in the reduced form and 2SLS regressions. Intuitively, we have two variables
that are correlated with interest in the SXSW festival among early Twitter adopters, but only one
predicts Twitter users in later years, allowing us to disentangle interest in the festival from its effect
via more Twitter users.
Third, we provide additional placebo tests for other festivals in 2007 and show that none of
these other festivals is either correlated with higher Twitter adoption or changes in the Republican
vote share today. This allows us to rule out that our effects are simply driven by the selection of
users into attending festivals that are highly similar to SXSW.
Fourth, estimating equations (2) and (3) for different time periods shows that SXSWcM arch2007
does not correlate with both levels and trends in election outcomes before Twitter’s launch in 2006
26
We weigh observations to make our sample representative of national election results. For example, due to many
less populated counties tilting Republican, the unweighted average Republican vote share across counties in the 2016
elections is 64%. Weighing makes our sample mean match the national average of 46%. Moreover, we obtain similar
results without using weights or using election-year turnout as weights (Appendix Table C.2), suggesting that effect
heterogeneity along dimensions correlated with weights do not play an important role in our context (Solon et al., 2015).
27
We consider spatial standard errors using the methods described in Colella et al. (2019) for robustness.
17
and in its early years, when the platform had few users and was unlikely to affect election outcomes.
Intuitively, outcomes in “treatment” and “control” counties behaved similarly before Twitter could
plausibly affect elections.
Fifth, we detect a statistically significant effect of SXSWcM arch2007 on Trump’s vote share in
2016 and 2020 but not on House and Senate elections (neither in 2016 nor 2020 or other periods
between 2000 and 2018). This pattern is consistent with an effect of Twitter, since there is more
content on presidential candidates than on congressional elections in the platform.
Sixth, results based on survey data suggest the effects are concentrated among moderate or
independent voters, which is also the expected pattern from Twitter having a causal effect due to
voter persuasion.
Stated differently, a violation of the identification condition would require an omitted variable
that correlates with SXSWcM arch2007 , T witter usersc , and yc but is uncorrelated with: i) SXSWcP re
as well as followers of other festivals, ii) levels and trends in election results before Twitter’s launch
and rise to popularity, iii) the observable variables presented in Table A.4, and iv) election results in
congressional elections both during the Trump elections and before, while also v) being correlated
with vote choices of moderate voters. Our argument is that the existence of such an omitted variable
is implausible to an extent that allows us to interpret θ as the effect of Twitter users on election
outcomes.
Measurement error in county-level Twitter usage and SXSWcM arch2007 is also unlikely to
explain an effect in 2016 and 2020 presidential elections, but no effect in previous presidential
elections or congressional elections. Moreover, SXSWcP re and the measures for other festivals are
constructed similarly as SXSWcM arch2007 and should thus have similar measurement error. However,
SXSWcP re and the other festivals are uncorrelated with Twitter usage and election outcomes.
Lastly, another possible concern is that the SXSW adoption shock led to differences in the
composition of Twitter users when compared to other U.S. counties. In particular, one might be
concerned that the SXSW festival lead to a more liberal Twitter population in the treated counties.
While this would not influence the causal interpretation of our findings, it could make the local
average treatment effect harder to interpret. Three pieces of evidence suggest that this appears to
be an unlikely concern. First, as we show in Appendix Figure B.3b, Twitter’s user base became
less connected to the SXSW festival over time and, in this process, likely reached people from
more diverse backgrounds. Second, the findings of Müller and Schwarz (2019) indicate that the
SXSW adoption shock was associated with an increase in hate crime with Trump’s presidential
18
run. This suggests that the shock eventually reached even the right-wing fringes of the political
spectrum. Third, we can directly address this concern by comparing the profiles of Twitter users in
SXSW home counties with those in the rest of the country. The results are presented in Appendix
Table B.2. We find that the user profiles in SXSW counties are highly similar to the general Twitter
population. If the Twitter population in SXSW counties was significantly more liberal, their Twitter
and names and biographies should also be different, as Oliver et al. (2016) document that names
predict political ideology. We find similar results when we look at which politicians users in the
different counties follow. If anything, Twitter users in the “pre-period” counties appear to have a
slightly more liberal Twitter network.
To be transparent, we want to stress what our findings do not imply. First, they cannot speak
about social media platforms other than Twitter, such as Facebook. Our empirical strategy exploits a
“shock” specific to early Twitter adoption, and we do not have a credible research design to estimate
the effects of other platforms. While many other platforms share similarities with Twitter, such
as being popular among younger and more educated people in urban areas (Pew Research Center,
2019c), other platforms may have different effects on political outcomes. Second, our research
design cannot separate the effect of particular types of social media content on Twitter (e.g., foreign
governments or misinformation), but rather speaks to the overall effect of Twitter exposure. Third,
like other papers in media economics, we estimate a “partial equilibrium” effect. In our case, we
estimate the effect of adding Twitter users to a county while keeping other counties’ Twitter use
constant. We thus cannot address whether Twitter had a national-level effect on the election (e.g.,
Trump’s tweets driving traditional media content).
6 Results
6.1 Main results

First-stage. Table 1 reports results from estimating equation (1) with different sets of control
variables. The results indicate that counties with more SXSW followers who joined Twitter in March
2007 have higher numbers of Twitter users during 2014-2015. Since the variables are in logs, the
coefficients can be interpreted as elasticities. A 10% increase in SXSW followers in March 2007
is associated with 5.2% more Twitter users. The results do not seem to be sensitive to the set of
included covariates. For example, the distance from Austin, Texas (the location of SXSW) has no
19
significant explanatory power. Importantly, the coefficients on SXSW followers before the 2007
festival are statistically insignificant and small in size: Twitter usage in 2014-2015 is not higher in
areas with more SXSW followers who joined Twitter before March 2007.
Figure 4 presents the graphical representation of the estimates in column (5) of Table 1.
Specifically, we show a binned scatter plot of T witter usersc against SXSWcM arch2007 after both
variables are “residualized” by partialling out the control variables. The figure is constructed by
dividing the x-axis variable into 40 equal-sized bins and plotting the average values of both variables
in each bin.28
Table 1: South by Southwest 2007 and Twitter Usage
Dep. var.: Log(Twitter users)

(1) (2) (3) (4) (5)
Log(SXSW followers, March 2007) 0.726*** 0.683*** 0.563*** 0.524*** 0.523***
(0.087) (0.079) (0.055) (0.048) (0.048)
Log(SXSW followers, Pre) 0.104 0.110 0.059 0.059 0.058
(0.101) (0.076) (0.098) (0.082) (0.082)
Population deciles Yes Yes Yes Yes Yes
Census region FE Yes Yes Yes Yes Yes
Geographical controls Yes Yes Yes Yes
Demographic controls Yes Yes Yes
Socioeconomic controls Yes Yes Yes
China shock controls Yes Yes
1996 election control Yes
Observations 3,065 3,065 3,064 3,064 3,064
R2 0.92 0.93 0.95 0.95 0.95
Mean of DV 8.22 8.22 8.22 8.22 8.22
p-value: March 2007 = Pre 0.00 0.00 0.00 0.00 0.00
Notes: This table presents county-level regressions where the dependent variable is the number of Twit-
ter users (in natural logarithm). Log(SXSW followers, March 2007) is the number of Twitter users (in
logs, with 1 added inside) who joined in March 2007 and follow South by Southwest (SXSW). SXSW
followers, Pre is the number of SXSW followers who registered at some point in 2006, defined simi-
larly. The bottom row reports p-values from F-tests for the equality of these coefficients. Regressions
include the indicated control variables (see the Online Appendix for their descriptions). Observations
are weighted by turnout in the 2000 presidential election. Standard errors in parentheses are clustered
by state. *** p < 0.01, ** p < 0.05, * p < 0.1.
Reduced Form and 2SLS Estimates. Table 2 shows the reduced form estimates from equation
(2) and both OLS and 2SLS estimates of equation (3), focusing on the Republican vote share
in the 2016 and 2020 presidential elections. The specifications across columns match those in
Table 1. Panel B indicates that the number of SXSW followers who joined Twitter in March 2007 is
associated with a lower Republican vote share. Panel C presents the 2SLS effects of Twitter usage on
vote shares. The 2SLS estimate in column (5) indicates that a 10% increase in the number of Twitter
28
The fitted line is based on the unbinned data. Observations are weighted by turnout in the 2000 presidential election.
This procedure guarantees the slope of the fitted line matches the estimate in column (5) of Table 1.
20
Figure 4: First Stage – South by Southwest (SXSW) and Twitter Usage
0.50
Log(Twitter users), residualized

0.25
0.00
-0.25
-0.50
-1.00 -0.50 0.00 0.50 1.00
Log(1 + SXSW followers, March 2007), residualized
Notes: This figure presents a binned scatter plot of the relationship between Twitter users in 2014-2015 and
the number of SXSW followers who joined Twitter in March 2007. Variables are residualized by partialling
out SXSW followers who joined before March 2007, population deciles, Census region fixed effects, as well
as geographical, demographic, socioeconomic, China shock, and 1996 election control variables (see Online
Appendix for control variable definitions). The figure is constructed by dividing the x-axis variable into 40
equal-sized bins and plotting the average values of both variables in each bin. The fitted line is estimated
using the unbinned data.
users in a county lowers Trump’s vote share by 0.21 p.p. (e.g., a reduction from a 46.1% vote share
to 45.8%).29 We discuss potential differences between LATE and ATE as well as heterogeneous
treatment effects at the end of this section. The results for the 2020 presidential election shown in
Table 2 column (6)-(10) are nearly identical.
Figure 5 plots the reduced form and OLS estimates from Table 2 graphically, specifically the
models in columns (5) and (10). These figures are constructed similarly to Figure 4 but show the
Republican vote share on the y-axis. The estimated slopes are negative.
Magnitudes and Persuasion Rates. Given the Electoral College and the closeness of individual
state races, our effects are consistent with Twitter potentially affecting election results. For example,
our results indicate that increasing the number of Twitter users by 30% lowered Trump’s vote share
by 0.63 p.p. This is larger than the margins by which he won Michigan, Wisconsin, and Pennsylvania
in 2016. Had Clinton won those states, she would have been elected president.
While the effects reported on Table 2 appear modest at first pass, they have to be interpreted
taking into account that only a quarter of Americans are active Twitter users (Pew Research Center,
2019c). To interpret the magnitudes of our estimates, we calculate a persuasion rate following
29
The F-statistic of our estimated first-stage range from 70 to 120. This suggests that estimation and inference
concerns related to weak instruments are unlikely to apply in our case.
21
Figure 5: South by Southwest, Twitter, and the Republican Vote Share
2016 Election
(a) Reduced Form (b) OLS
0.02
0.02
Republican vote share, 2016 (residualized)

0.01
0.01
0.00
0.00
-0.01
-0.01
-0.02
-0.02
-1.00 -0.50 0.00 0.50 1.00 -1.50 -1.00 -0.50 0.00 0.50 1.00
Log(1 + SXSW followers, March 2007), residualized Log(Twitter users), residualized
2020 Election
(c) Reduced Form (d) OLS
0.04
0.02
0.01
0.02
0.00
0.00
-0.01
-0.02
-0.02
-1.00 -0.50 0.00 0.50 1.00 -1.50 -1.00 -0.50 0.00 0.50 1.00
Log(1 + SXSW followers, March 2007), residualized Log(Twitter users), residualized
Notes: Panel (a) presents a binned scatter plot of the relationship between the Republican vote share in the
2016 presidential election and the number of SXSW followers who joined Twitter in March 2007. Variables
are residualized with respect to SXSW followers who joined before March 2007, population deciles, Census
region fixed effects, as well as geographical, demographic, socioeconomic, China shock, and 1996 election control
variables. The figure is constructed by dividing the x-axis variable into 40 equal-sized bins and plotting the
average values of both variables in each bin. The fitted line is estimated using the unbinned data. Panels (b) and
(d) replicate the exercise using Twitter users in 2014-2015 as the x-axis variable. Panel (c) and (d) show results for
the 2020 election.
22
t
DellaVigna and Gentzkow (2010). The persuasion rate can be approximated as θ · e(1−y)
, where θ is
the 2SLS estimate, y is the average Republican vote share, e is the average exposure of American
adults to Twitter, and t is the share of adults that turn out to vote. Using the estimate for θ from
columns (5) and (10) of Table 2, the persuasion rate is 8.6% and 9.4% for 2016 and 2020, respectively.
It implies that, in 2016, one out of every twelve active Twitter users that voted for Clinton would not
have done so if they had not been exposed to the platform.30
Note also that we estimate county-level effects and persuasion rates, which may capture local
spillovers and social interaction effects. In other words, they implicitly assume that one additional
Twitter user can only affect one person’s vote. However, if this additional user also exposes other
non-users to content she sees on Twitter (e.g., via in-person conversations), then an “individual-level”
persuasion rate would be smaller. Guriev et al. (2020) show that persuasion rate formulas scale by a
factor 1/N where N is the number of voters in a county “exposed” given each additional Twitter
user. For example, if each new user has a 30% chance of exposing another voter in the county to the
persuasive content she sees on Twitter, our county-level persuasion rates must be divided by 1.3 to
be interpreted as individual-level persuasion rates.
The persuasion rates of Twitter are smaller than the persuasion rates of 3G internet on voting
for establishment parties, which Guriev et al. (2020) estimate to be 27%. It is also smaller than the
estimated pro-Republican persuasion rate of Fox News, which Martin and Yurukoglu (2017) estimate
to range between 27% and 58% (depending on the year).31 As a further comparison, Gentzkow et al.
(2011) estimate a persuasion rate of 12.9% for the effect of reading a local newspaper in the 1869-
1928 period on voter turnout. All these persuasion rates are not on the “individual level,” meaning
they are based on effects estimated at a regional or county level. However, Twitter’s persuasion
rates are also somewhat smaller than others from studies that estimate effects with individual-level
randomization: Gerber et al. (2009) report a 19.5% pro-Democrat persuasion rate for reading
the Washington Post and DellaVigna and Gentzkow (2010) survey randomized get-out-the-vote
canvassing experiments and calculate persuasion rates in the 10-15% range.
30
The persuasion rate re-scales effect sizes by how many individuals are exposed to the platform and how many are
not already persuaded. For marginal changes in exposure, the formula is f = dy t
de · 1−y (DellaVigna and Gentzkow,
2010). Since our estimate θ is the semi-elasticity dy t
de · e, we obtain f = θ · e(1−y) . In 2016, y = 0.46, t = 0.55, while
y = 0.47 and t = 0.62 in 2020. We assume e = 0.25 for both periods. This implicitly assumes that Twitter usage among
voters is the same as the overall population. If voters are over-represented among Twitter users, the persuasion rate
would be smaller. On one hand, Twitter users are younger (which is associated with lower turnout) but more educated
(which is associated with higher turnout) than the general population.
31
DellaVigna and Kaplan (2007) estimate a smaller persuasion rate of 11.6% for Fox News.
23
Robustness to Controls and Coefficient Stability. Although the estimated effects are always
negative and significant at the 1% level, comparing columns shows that the effect sizes vary
somewhat with the inclusion of controls, especially demographic and socioeconomic ones. This
sensitivity to controls is perhaps expected given their explanatory power over vote shares. The
results in other papers that explore effects on vote shares in similar frameworks, such as DellaVigna
and Kaplan (2007), Martin and Yurukoglu (2017), and Autor et al. (2020), show a similar sensitivity
to controls.
To further probe the sensitivity of our results, we provide three separate pieces of evidence.
First, we apply two approaches, from Altonji et al. (2005) and Oster (2019), designed to gauge
the potential importance of unobservable variables in driving the results. We compare our reduced
form (Panel B) specifications with all controls (columns 5 and 10) to those with the fewest controls
(columns 1 and 6). We obtain an “Oster-δ” of approximately 7 for both 2016 and 2020. This
suggests, in order for the true effect to be zero, unobservable variables would have to be seven
times as “important,” in terms of driving selection into treatment and explaining the outcome, as the
(numerous) controls added in columns (5) and (10). We also obtain an “Altonji et al-δ” of 1.7 for
2016 and 1.6 for 2020, which have a similar interpretation as the “Oster-δ;” note that these values
are smaller because they do not take into account how much of the variance in the treatment (in our
case, Twitter usage) is explained by the controls.32
Both approaches indicate that our results are robust when compared to usual cutoffs and to
other papers published in economics journals. Both Altonji et al. (2005) and Oster (2019) argue
that δs above one are an appropriate “robustness standard.” Moreover, Oster (2019) calculates the
“Altonji et al-δ” for 45 results published in the American Economic Review, Quarterly Journal of
Economics, and Journal of Political Economy. Only 20% of these results have a “Altonji et al-δ”
above the ones we find in our context. Moreover, Oster (2019) reports only 9.1% of these results
survive the cutoff of an “Oster-δ” above one.
Second, Appendix Table C.1 provides further evidence for the robustness of our findings
to the inclusion of different controls. It allows for more flexible interactions between observable
32
The R2 of regressions with the fewest controls (columns 1 and 6) is approximately 0.6, while it is 0.94 for the
regressions with the most controls (columns 5 and 10). Thus observable controls explain a large part of the variation
in vote shares but only generate a modest change in coefficient sizes. Intuitively, this is what generates large δs. The
difference between the calculated “Altonji et al-δ” and “Oster-δ” is that the former only considers the coefficients and R2
of the two specifications being compared, while the latter takes into account how much of the variance in treatment can
be explained by observed controls. Note that the “Oster-δ” is derived under more general conditions than the “Altonji et
al-δ.” In all calculations, we set Rmax (the hypothetical R2 that one would obtain if all relevant unobserved controls
were included) to one. This is the most conservative assumption possible.
24
Table 2: Twitter and the 2016/2020 Republican Vote Share
Dep. var.: Republican vote share in 2016 Dep. var.: Republican vote share in 2020
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Panel A: OLS
Log(Twitter users) -0.065*** -0.067*** -0.013*** -0.011*** -0.007** -0.058*** -0.064*** -0.012*** -0.011*** -0.008***
(0.009) (0.008) (0.004) (0.003) (0.003) (0.009) (0.009) (0.004) (0.003) (0.003)
Panel B: Reduced form
Log(SXSW followers, March 2007) -0.053*** -0.058*** -0.019*** -0.014*** -0.011*** -0.046*** -0.055*** -0.017*** -0.013*** -0.011**
(0.011) (0.012) (0.005) (0.004) (0.004) (0.009) (0.010) (0.006) (0.005) (0.005)
Log(SXSW followers, Pre) -0.021 -0.003 -0.000 -0.002 0.001 -0.022 -0.005 -0.002 -0.004 -0.001
(0.016) (0.013) (0.006) (0.006) (0.004) (0.016) (0.013) (0.007) (0.007) (0.005)
Panel C: 2SLS
Log(Twitter users) -0.072*** -0.085*** -0.034*** -0.027*** -0.021*** -0.064*** -0.080*** -0.031** -0.025*** -0.020**
(0.016) (0.018) (0.010) (0.008) (0.008) (0.015) (0.017) (0.011) (0.009) (0.009)
25
Log(SXSW followers, Pre) -0.014 0.007 0.002 -0.001 0.002 -0.015 0.004 -0.000 -0.002 0.000
(0.020) (0.016) (0.007) (0.006) (0.005) (0.020) (0.015) (0.008) (0.007) (0.006)
Population deciles Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Census region FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Geographical controls Yes Yes Yes Yes Yes Yes Yes Yes
Demographic controls Yes Yes Yes Yes Yes Yes
Socioeconomic controls Yes Yes Yes Yes Yes Yes
China shock controls Yes Yes Yes Yes
1996 election control Yes Yes
Observations 3,065 3,065 3,064 3,064 3,064 3,065 3,065 3,064 3,064 3,064
Mean of DV 0.46 0.46 0.46 0.46 0.46 0.47 0.47 0.47 0.47 0.47
Robust F-stat. 69.57 74.65 106.08 118.21 121.18 69.57 74.65 106.08 118.21 121.18
Notes: This table presents county-level regressions where the dependent variable is the Republican vote share in the 2016 or 2020 presidential election. Log(SXSW followers, March
2007) is the number of Twitter users (in logs, with 1 added inside) who joined in March 2007 and follow South by Southwest (SXSW). SXSW followers, Pre is the number of SXSW
followers who registered at some point in 2006, defined similarly. Twitter users are the number of users in 2014-2015. Regressions include the indicated control variables (see the Online
Appendix for their descriptions). The first-stage regressions for 2SLS results (Panel B) are presented in Table 1, with the F-stat for the excluded instrument in the bottom row. The
“Oster-δ” for gauging the possible role of unobservables in driving coefficient stability is 6.93 (comparing column 5 to 1) and 7.01 (comparing column 10 to 6); see text for further details.
Observations are weighted by turnout in the 2000 presidential election. Standard errors in parentheses are clustered by state. *** p < 0.01, ** p < 0.05, * p < 0.1.
characteristics by using partialling-out LASSO variable selection (Chernozhukov et al., 2015a,b).33
More specifically, we estimate regressions in which we allow for state fixed effects and all possible
interactions of the control variables (i.e., up to 654 controls). The LASSO then selects controls that
either predict the vote shares, SXSW participation, or Twitter usage. All estimates are statistically
significant and of similar magnitude as our baseline estimates.
Third, Figure D.3a plots the estimated θ of our 2SLS equation (3) while flexibly allowing the
included control variables to vary. The resulting “specification curves” suggest that our results (for
both the 2016 and 2020 elections) are robust to how our regressions are specified. The estimated
coefficients are always negative, almost always statistically significant at the 5% level, and in the
overwhelming number of specifications considerably more negative than our “baseline estimates,”
which is marked by the vertical line.
Placebo Test: Interest in SXSW Before March 2007. The coefficients on SXSWCP re in Table 2
are statistically insignificant and substantially smaller than those on SXSWcM arch2007 . As discussed
in Section 5, this provides support for our identification condition (exclusion restriction). Suppose
that our instrument merely captured that counties with an interest in SXSW’s Twitter account during
the platform’s early years also differ in (unobservable) characteristics that predict the 2016 election
outcome. If this was the case, the coefficients on SXSWCP re should be similar in size to those on
SXSWcM arch2007 . Intuitively, we have two variables that are correlated with interest in the SXSW
festival, but only one predicts Twitter users in later years, allowing us to disentangle interest in the
festival (and its correlates) from its effect via more Twitter users.
Placebo Test: Other Festivals in 2007. We also provide an additional placebo check by investi-
gating five other festivals that are similar in nature to SXSW (Burning Man, Coachella, Pitchfork,
Austin City Limits, and the Electric Daisy Carnival).34 One of these placebo festivals, Austin City
Limits, takes place in the same city as SXSW. For this exercise, we construct analogous measures
to our instrument for other festivals using the same exact procedure: the Twitter followers of the
respective festival that joined in the month the festival took place in 2007. In Table 3, we then show
that none of these festival variables, including that for Austin City Limits, has predictive power for
future Twitter usage. More importantly, the variables are also essentially uncorrelated with 2016 and
33
Note that, unlike most other regressions reported in the paper, these regressions are unweighted.
34
These festivals take place in the Black Rock Desert (NV), Indio (CA), Chicago (IL), Austin (TX), and Las Vegas
(NV), respectively.
26
2020 presidential vote shares. These results provide further support for our identifying assumption
and suggest that a spurious relationship between counties’ Twitter users interest in festivals during
2007 and future election outcomes cannot explain our findings.
Table 3: Placebo Tests for Other Festivals
Twitter Followers of Festival joining in Festival Month

SXSW Burning Man Coachella Pitchfork EDC ACL
(1) (2) (3) (4) (5) (6)
Panel A: First Stage (Dep. Var.: Twitter Usage)
Followers Festival Month 0.167*** -0.005 0.011 -0.006 0.010 -0.009
(0.015) (0.012) (0.012) (0.006) (0.008) (0.018)
Panel B: Reduced form (Dep. Var.: Rep. Vote Share 2016)
Followers Festival Month -0.004*** 0.001 -0.000 -0.000 0.000 -0.000
(0.001) (0.001) (0.001) (0.000) (0.001) (0.001)
Panel C: Reduced form (Dep. Var.: Rep. Vote Share 2020)
Followers Festival Month -0.003** 0.000 -0.001 -0.000 0.000 -0.000
(0.001) (0.001) (0.001) (0.000) (0.001) (0.001)
Census region FE Yes Yes Yes Yes Yes Yes
Population controls Yes Yes Yes Yes Yes Yes
Geographical controls Yes Yes Yes Yes Yes Yes
Demographic controls Yes Yes Yes Yes Yes Yes
Socioeconomic controls Yes Yes Yes Yes Yes Yes
China shock controls Yes Yes Yes Yes Yes Yes
1996 election control Yes Yes Yes Yes Yes Yes
Observations 3,064 3,064 3,064 3,064 3,064 3,064
Notes: This table presents the first stage and reduced from estimates using equation (3). The dependent
variable is either Twitter usage (Panel A) or the vote share of the Republican party in the 2016 and 2020
presidential elections (Panels B and C). The independent variables are number of followers of different
festivals that joined in the respective festival month (in logs with 1 added inside). To make the coefficients
comparable, we standardized these variables to have mean zero and standard deviation one; this means the
results for SXSW are not directly comparable in magnitude to those in the main text. Similarly to our
baseline specification, all regressions control for the number of festival followers that joined Twitter before
the festival. As the festivals take place after the SXSW festival, we additionally control for the number
of SXSW followers from March 2007. All regressions are weighted by turnout in the 2000 presidential
election. All regressions include the controls from columns 5 and 10 in Table 2. EDC is the Electric
Daisy Carnival and ACL is Austin City Limits. Standard errors in parentheses are clustered by state. ***
p < 0.01, ** p < 0.05, * p < 0.1.
Further Robustness and Additional Tests. The Online Appendix presents a number of additional
sensitivity checks. In Table C.2, we consider changes to the baseline regression specification. In
particular, we allow for unweighted regressions; weighting by the relevant election-year turnout
(2016 or 2020); alternative functional forms of the pre-SXSW user variable; restrict the sample to
the sub-sample of counties where we observe either SXSW followers who joined in March 2007 or
the pre-period; allow for flexible spatial correlation in the standard errors; and per capita measures
of Twitter usage. Further, Table C.4 replaces our baseline time-invariant measure of Twitter usage
(making use of all available user data) with a time-varying measure based on how many users were
27
using the platform in a given year. None of these adjustments make a substantial difference in the
magnitudes or statistical significance of the estimates.
Online Appendix Table D.1 reports results for additional outcome variables, all of which
support the idea that Twitter exerts a pro-Democrat effect. In column (1), we use a probit IV model
to investigate the likelihood of a county switching from Obama in 2008 to Trump in 2016. The
coefficient suggests that a one standard deviation increase in Twitter usage is associated with a -24%
lower probability of a county electing Obama in 2008 and Trump in 2016. Columns 2 and 3 look at
changes in campaign donations to Democrats and Republicans between 2000 and 2016. We find a
positive and statistically significant effect for donations to Democrats, and no effect for Republicans.
Lastly, columns 4 and 5 look at approval ratings for President Trump in 2017 based on data from the
Gallup Daily Poll. We find that exposure to social media is associated with a decrease in Trump’s
popularity, and more so among Republicans.
The 2007 SXSW Shock and Previous Election Outcomes Table 4 repeats the analysis from
column (5) of Table 2 using Republican vote share in the previous presidential elections of 2000,
2004, 2008, and 2012 as the dependent variable.35 All estimates for those years are substantially
smaller than the ones for 2016 and 2020 and statistically insignificant. For 2000, 2004, and 2008,
this can be interpreted as a placebo or pre-trends test: conditional on the covariates, our instrument
is uncorrelated with outcomes before Twitter’s launch (2000 and 2004) and when the platform
had few users (2008). In 2012, Twitter already had a substantial user base, so our estimates can
be interpreted as a genuine “zero effect”; we return to a comparison of the 2020, 2016, and 2012
results in Section 6.3. The estimates for 2000, 2004, 2008, and 2012 are also jointly statistically
insignificant (p-value=0.329). The average effect in 2016 and 2020 is statistically different from the
average effect in 2000, 2004, 2008, and 2012 (p-value=0.002).
In Appendix Figure C.1, we further show the reduced form estimates for presidential election
going back as far as 1924. We find that our instrument is also uncorrelated with any of the earlier
presidential election results. As discussed in Section 5, this result lends additional support for our
exclusion restriction. If our instrument merely captured uncontrolled differences across counties,
these should also correlate with vote shares in previous elections.
35
Note that the right-hand side of the regressions remains the same as in the previously reported regressions. That is,
the instrument is the same, and the “endogenous variable” in the 2SLS is Twitter users measured in 2014-2015. We
explore the role of time-varying Twitter user measures on Appendix Table C.4.
28
While these findings make it unlikely that our instrument is correlated with pro-Democratic
attitudes at the county level, a possible concern could be that we are picking up “anti-populist”
attitudes, which could have harmed Trump’s electoral results. To address this concern, we turn to
the historical case study of Ross Perot’s political campaign in 1992 and 1996. Perot, a billionaire
businessman, also ran as a “third-party candidate” on a populist platform. However, when we replace
the dependent variable with the third-party vote share in the 1992 and 1996 presidential election
(see Appendix Table C.5), we find no evidence that our instrument is associated with lower vote
shares for Ross Perot. This makes it unlikely we are capturing differences in “demand for populism”
across counties.36
Effects on Vote Share Changes. We also consider specifications of equations (2) and (3) using
vote share changes instead of levels as the dependent variable. All our estimates based on changes
take differences relative to the base year 2000 (akin to the approach in Autor et al. (2020)) and
use the full set of controls (as in columns 5 and 10 of Table 2). Figure 6a plots the reduced form
estimates for changes in the Republican vote share in presidential elections.
The results corroborate the previously presented evidence based on specifications in levels.
Our instrument is essentially uncorrelated not only with levels but also with changes (or trends)
in election outcomes during the 2000-2012 period. Given our arguments above, this also lends
support for our identification strategy. Again the estimates for 2000, 2004, 2008, and 2012 are
also jointly indistinguishable from 0 (p-value=0.398). The reduced form effects for 2016 and 2020
are similar to the one estimated using levels. For example, the estimated effect (θ) for 2016 using
changes is -0.017, similar to the one estimated in levels (-0.021).37 Unsurprisingly, the average
effect in 2016 and 2020 is statistically different from the average effect in 2000, 2004, 2008, and
2012 (p-value=0.001).
Effects on Turnout and Congressional Elections. Figure 6b, Figure 7a, and Figure 7b replicate
the exercise in Figure 6a using voter turnout and vote shares in House and Senate elections as the
outcomes. We do not find a statistically significant association between our instrument and election
turnout except for 2020. Before 2020, the estimated point effects are usually small. For example,
the upper bound on the 95% confidence interval of the 2SLS estimate for the effect of turnout in the
36
In Appendix Table C.6, we also investigate the vote shares in the 2020 democratic primaries. Here we find a positive
association between Twitter exposure and the vote share of Bernie Sanders, often described as a left-wing populist. This
further speaks against the hypothesis of “anti-populist” sentiment.
37
Appendix Table C.3 present the OLS and 2SLS estimates.
29
2016 election implies that a 10% increase in Twitter users raises turnout by 0.036 p.p. (Appendix
Table D.2).
In the 2020 election, which saw the highest turnout rate in more than a century (NPR, 2020c),
we find that Twitter is associated with a larger fraction of the voting-age population casting their
ballot. Why did Twitter have an effect on 2020 turnout but not in the previous election? One possible
explanation could be that calls to turn out, and vote were widespread on the platform, partially
because of an initiative by Twitter itself that was not present in 2016 (Twitter, 2020). Further,
the 2020 election was unique in its prevalence of mail and early voting because of the Covid-19
pandemic. Another possibility is that Twitter served as a platform to convey information on how to
vote by mail or before election day.
The coefficients for House and Senate races are more noisily estimated, particularly for the
smaller sample of Senate races (where only a third of seats is renewed every two years). Overall,
there is little evidence suggesting an effect of Twitter on congressional elections, including in 2016,
the 2018 midterm election, and 2020. Finding an effect on presidential vote shares but not in these
“down-ballot” races is perhaps expected since content on presidential candidates (and in particular
on Trump in 2016 and 2020) is more common on Twitter than content on congressional races.38
Discussion of Identification Condition. As discussed in more detail in Section 5, there are five
pieces of evidence supporting our identification condition: i) our empirical strategy compares
relatively similar counties (Section 5); ii) the placebo test based on the coefficient on SXSWcP re
and other festivals; iii) the instrument being uncorrelated with election outcomes in the 1924-2012
period; and iv) absence of a statistically significant effect of Twitter on House and Senate races;
while at the same time v) the instrument being correlated with vote choices of moderate voters in
particular.
Given this, a violation of the identification condition would require an omitted variable that
correlates with the instrument, Twitter usage, and Trump’s vote share in 2016 and 2020 but is
uncorrelated with: i) SXSWcP re and other festivals, ii) levels and trends in election results before
Twitter’s launch and rise to popularity, iii) the observable variables presented in Table A.4, and iv)
election results in congressional elections. At the same time, such omitted variable would also v) be
38
Appendix Table D.3 presents the reduced form estimates for House and Senate races. To accommodate the Senate’s
six-year terms, we take changes relative to 2000, 1998, and 1996, instead of always using 2000 (as we do for other
outcomes).
30
more strongly correlated with the vote choices of moderates and independents than “partisans.” Our
argument is that the existence of such omitted variable is unlikely.
Table 4: Twitter and the Republican Vote Share, 2000-2020
Dep. var.: Republican vote share in...

2000 2004 2008 2012 2016 2020
(1) (2) (3) (4) (5) (6)
Panel A: Reduced form
Log(SXSW followers, March 2007) -0.003 -0.005 -0.006 -0.003 -0.011*** -0.011**
(0.002) (0.003) (0.004) (0.004) (0.004) (0.005)
Log(SXSW followers, Pre) 0.001 0.001 -0.000 -0.002 0.001 -0.001
(0.004) (0.003) (0.006) (0.005) (0.004) (0.005)
Panel B: 2SLS
Log(Twitter users) -0.005 -0.009 -0.011 -0.007 -0.021*** -0.020**
(0.004) (0.006) (0.009) (0.008) (0.008) (0.009)
Log(SXSW followers, Pre) 0.001 0.001 0.000 -0.001 0.002 0.000
(0.004) (0.004) (0.006) (0.005) (0.005) (0.006)
Observations 3,064 3,064 3,064 3,064 3,064 3,064
Mean of DV 0.48 0.51 0.46 0.47 0.46 0.47
Robust F-stat. 121.18 121.18 121.18 121.18 121.18 121.18
Notes: This table presents county-level regressions where the dependent variable is the Republican vote
share in presidential elections. Log(SXSW followers, March 2007) is the number of Twitter users (in logs,
with 1 added inside) who joined in March 2007 and follow South by Southwest (SXSW). SXSW followers,
Pre is the number of SXSW followers who registered at some point in 2006, defined similarly. Twitter users
are the number of users in 2014-2015. All regressions control for population deciles, Census region fixed
effects, and the full set of controls (as in columns 5 and 10 of Table 2). The first-stage regression for 2SLS
results (Panel B) are presented in column (5) of Table 1, with the F-stat for the excluded instrument in the
bottom row. On Panel (a), the coefficients on Log(SXSW followers, March 2007) for 2000, 2004, 2008, and
2012 are jointly statistically insignificant (p-value=0.329). Further, the average effect in 2016 and 2020 is
statistically distinct from the average effect in 2000, 2004, 2008, and 2012 (p-value=0.002). Observations
are weighted by turnout in the 2000 presidential election. Standard errors in parentheses are clustered by
state. *** p < 0.01, ** p < 0.05, * p < 0.1.
Towards an Average Treatment Effect As with any instrument, our 2SLS results identify a local
average treatment effect (LATE). In our setting, the “compliers” are counties with higher Twitter
usage as a result of the inflow of SXSW attendees. While the negative treatment effect of Twitter
usage for these counties is in itself an interesting finding, the ATE for the US overall—and therefore
the overall impact of Twitter on elections—may differ. However, two pieces of evidence suggest
that our estimates, despite this concern, allow us to infer information about the ATE.
A first indication comes from comparing the OLS and 2SLS results in Table 2. Both estimates
are always negative and relatively similar in magnitude. Second, we build on the approach suggested
by Andrews and Oster (2019) and more formally investigate the external validity bias of our estimates.
For experimental settings, Andrews and Oster (2019) suggest using the observable heterogeneity in
estimated treatment effects within the experimental sample to learn about the ATE in the overall
population. Similarly, we can use the heterogeneity of treatment effects within counties that provide
31
Figure 6: Twitter and Presidential Elections – Reduced Form
(a) Changes in Republican Vote Share (b) Change in Voter Turnout
0.005 0.015
Point estimate on Log(SXSW followers, March 2007)

0.010
0.000
0.005
-0.005
0.000
-0.010
-0.005
-0.015
-0.010
2000-04 2000-08 2000-12 2000-16 2000-20 2000-04 2000-08 2000-12 2000-16 2000-20
Notes: These figures plot reduced form estimates β̂ 0 from county-level regressions as in equation (2). They
measure the effect of Log(1 + SXSW followers, March 2007), while controlling for Log(1 + SXSW followers, Pre),
on changes in the Republican vote share in presidential elections relative to the year 2000 in Panel (a), and changes
in the ratio of voter turnout to voting-age population relative to 2000 in Panel (b). All regressions control for
population deciles and Census region fixed effects, and the full set of controls (as in columns 5 and 10 of Table 2).
On Panel (a), the estimates for 2004, 2008, and 2012 are jointly statistically insignificant (p-value=0.398). Further,
the average effect in 2016 and 2020 is statistically distinct from the average effect for 2004, 2008, and 2012
(p-value=0.001). Regressions are weighted by turnout in the 2000 presidential election. Whiskers represent 95%
confidence intervals based on standard errors clustered by state.
the variation that identify our results to approximate the ATE for the United States as a whole. Using
all included control variables from our main specification for the prediction of heterogeneity of the
treatment effect, the Andrews and Oster (2019) approach suggests that the ATE should, if anything,
be larger than the LATE we estimate in our baseline results. This seems plausible as the more urban
counties for which we have variation in our instrument tend to be Democratic strongholds, and thus
likely have fewer independents, for which we find the largest persuasion effects (in survey data). We
provide additional details on our approach in Appendix E39
6.2 Effects Are Concentrated Among Moderate Voters

If social media indeed matters for election outcomes, we would expect there to be heterogeneous
effects across groups of voters. In particular, Bayesian updating suggests that voters who do not hold
strong priors about a particular party should be more likely to be persuaded. We test this prediction
39
As our setting differs from the one discussed in Andrews and Oster (2019), some adjustments to our baseline
estimation were required. We estimate the treatment effect exclusively in the subset of counties for which either
SXSWcM arch 2007 or SXSWcP re are not equal to zero. Then, we define a treatment indicator variable equal to 1 for
counties with SXSW followers who joined in March 2007.
32
Figure 7: Twitter and Congressional Election Results – Reduced Form
(a) House Elections
0.06
0.04
0.02
0.00
-0.02
2000-02 2000-04 2000-06 2000-08 2000-10 2000-12 2000-14 2000-16 2000-18 2000-20
(b) Senate Elections

0.04
0.02
0.00
-0.02
-0.04
1996-02 1998-04 2000-06 1996-08 1998-10 2000-12 1996-14 1998-16 2000-18 1996-20
Notes: These figures plot reduced form estimates β̂ 0 from county-level regressions as in equation (2). They
measure the reduced form effect of Log(1 + SXSW followers, March 2007), while controlling for Log(1 + SXSW
followers, Pre), on the Republican vote share in House and Senate elections since 2000. For House elections in
Panel (a), the dependent variable is the change in the Republican vote share since 2000. For Senate elections in
Panel (b), the dependent variable is the change in the Republican vote share from six, twelve, or eighteen years ago
(to accommodate senators’ 6-year terms). All regressions control for population deciles and Census region fixed
effects and the full set of controls (as in columns 5 and 10 of Table 2). Regressions are weighted by turnout in the
2000 presidential election. The whiskers represent 95% confidence intervals based on standard errors clustered by
state.
33
using individuals’ voting decisions from the 2016 and 2020 CCES. In particular, we estimate the
following instrumental variable Probit regression:
yic = φ + θ · T witter
\ usersc + π · SXSWcP re + Xic ρ + εic , (4)
where yic is an indicator variable equal to 1 if an individual i living in county c voted for Trump
in the 2016 or 2020 election and 0 for Clinton.40 The definition of the county-level variables
T witter usagec and SXSWcP re remains unchanged. Xic is now a vector of individual-level control
variables including age, gender, race, family income, and education. We again instrument for
county-level Twitter usage based on the SXSW followers who joined in March 2007.
Table 5 presents results from estimating equation (4). In Panel A and B, Column (1) suggests
county-level Twitter usage has a statistically significant negative effect on the likelihood to vote for
Trump. The marginal effect implies that a 10% increase in the number of Twitter users in a county
would lower Trump’s vote share by 0.49 p.p. in the 2016 and 0.46 p.p. in the 2020 presidential
election.41
Columns (2)-(6) report results estimated separately by voters’ reported party affiliation. The
effect is strongest for voters who identify as independents, and thus likely to not hold strong priors.
The results suggest weaker or zero effects for those with stronger political views, whether Republican
or Democrat.
Indeed, we find larger marginal effects for Log(T witter users) among younger voters. For
independents, the estimate for young voters is 20% larger than for older voters (−0.071 compared to
−0.059). Among moderate Republicans and Democrats, the estimated coefficients are close to zero
for voters aged 50+ but sizeable for younger voters (although they are not statistically significant
at conventional levels). Because young voters are less likely to vote for Trump, this implies larger
elasticities of vote outcomes with respect to Twitter for those below 50 relative to the baseline
probabilities.
A potential concern with this exercise is that party affiliation may itself be affected by Twitter
usage. We thus present further support for Twitter having persuasion effects on moderates using
40
Note that we use data on all CCES respondents, not only those in the 2016 wave (i.e., respondents from the 2018
and 2020 waves were asked how they voted in 2016). The results are similar if we only use the 2016 wave. In unreported
regressions, we do not find an effect on votes for Jill Stein.
41
This effect size is within the range of the county-level estimates presented in Table 2. Appendix Table C.7 shows
this baseline result is robust to using only individuals with validated turnout and/or who stated that they originally
intended to vote for Trump in the pre-election wave of the CCES.
34
county-level data that is not subject to such concerns. In particular, we estimate our county-level
specification (equation 3) for the 2016 elections, but split counties based on how consistently they
voted for either the Republican or the Democratic party. Specifically, we define “swing counties” as
counties that were not consistently won by one party in the presidential elections from 2000 to 2012.
Because we find no effect of Twitter on vote shares before 2016 (see Table 4), splitting the sample
by swing counties is not subject to potential concerns that this variable might itself be affected by
social media.
Appendix Table D.4 shows the results. For the 2016 presidential elections, we find that Twitter
usage only negatively impacts the Republican vote share in “swing” counties. We find no evidence
for Republican or Democratic strongholds, where people likely have the strongest priors. The
patterns are similar in 2020. But here, we also find a small effect on counties that usually vote
Republican, which could suggest that the effect on moderate Republicans we find in the CCES also
apply to the county level.
Table 5: Twitter and Individuals’ Vote Decisions in 2016/2020
Dep. var.: Voted for Trump in ...

Full Strong Mod. Mod. Strong
Sample Dem. Dem. Indep. Rep. Rep.
(1) (2) (3) (4) (5) (6)
Panel A: 2016 Election
Log(Twitter users) -0.135*** 0.018 -0.079 -0.209*** -0.045 -0.022

(0.045) (0.054) (0.067) (0.058) (0.043) (0.037)
Observations 146,579 44,241 30,745 13,625 27,865 28,559

Mean of DV 0.492 0.026 0.111 0.611 0.919 0.980
Marginal effect [-0.049] [0.001] [-0.014] [-0.073] [-0.006] [-0.001]
Panel B: 2020 Election
Log(Twitter users) -0.130*** -0.069 -0.002 -0.186*** 0.000 -0.080

(0.042) (0.075) (0.120) (0.063) (0.041) (0.079)
Observations 43,617 13,108 9,670 4,430 7,388 8,395

Mean of DV 0.475 0.012 0.074 0.516 0.908 0.978
Marginal effect [-0.046] [-0.002] [0.000] [-0.069] [0.000] [-0.004]
Notes: This table presents results estimated using IV probit models, as in equation (4). The dependent
variable is a dummy for individuals in the CCES who voted for Trump in 2016 or 2020. Log(Twitter
users) is instrumented using the (log) number of SXSW followers that joined Twitter in March 2007.
All regressions control for the (log) number of SXSW followers that joined Twitter at some point in
2006, family income, gender, education levels, marital status, news interest, and age, as well as county-
level population deciles and Census region fixed effects. Regressions are weighted by survey weights.
Standard errors in parentheses are clustered by state. *** p < 0.01, ** p < 0.05, * p < 0.10.
35
6.3 Potential Mechanisms
The findings above suggest that Twitter had an effect in 2016 and 2020, but not during previous
presidential elections. We address three potential explanations for this pattern: lack of familiarity
with social media (a learning channel), changes in social media’s “slant” (a content channel), and
Trump’s role as an outsider candidate (a political shock channel).
The first factor could be the reach of and familiarity with social media content. In 2008, social
media was a relatively new type of technology. Only a quarter of American adults used any social
media platform and only 10% of internet users posted political commentary on social media (Pew
Research Center, 2009, 2011). Figure A.1a shows that Twitter, which was founded in 2006, only
had around one million users during the 2008 elections, compared to 40 million in 2012, 67 million
in 2016, and 69 million in 2020 (Statista, 2019, 2020). Twitter’s limited reach and novelty might
have initially restricted its impact on voters.
The second possible explanation is that social media’s content changed between 2008 and
2016. It is conceivable that, similar to changes in the slant of cable news (e.g., Martin and Yurukoglu,
2017), the content to which Twitter users are exposed has become more left-leaning over time.
A third reason is that Trump’s political rise constituted a considerable shock to the U.S.
political system. In this view, Twitter may not have partisan effects per se. Instead, the platform
may have served as a conduit for spreading sentiments or information about Trump, either because
of his own prominent behavior on the platform or because of other users’ content about him.
Two pieces of evidence presented above are consistent with the political shock channel. First,
if Twitter had little effect before 2016 because it was not widely used, its effect on vote shares
should systematically increase over time. We do not find evidence supporting this idea in effects for
presidential, House, and Senate elections (Figures 6 and 7). Instead, we find a discontinuous negative
effect in the 2016 presidential election that persists in 2020. Second, we do not find significant
effects for the 2016, 2018, and 2020 House and Senate elections. This implies that Twitter usage
lowered Trump’s vote share without significantly affecting other Republican candidates on the same
election day.
Results from the 2016 Republican Primaries. We provide additional evidence for a Trump-
specific effect of Twitter exposure by investigating the 2016 county-level Republican primaries
results. The primaries allow us to focus on the favorability of different candidates among Republican
voters. The results from this exercise are presented in Table 6. We find that Twitter usage is
36
associated with a lower vote share for Trump. We also find a positive effect on the vote share of
John Kasich, the most moderate of the major Republican candidates.42
Table 6: Twitter and Vote Shares in the 2016 Republican Primaries
Dep. var.: Vote share in Republican Primary of...

Trump Cruz Rubio Bush Kasich
(1) (2) (3) (4) (5)
Log(SXSW followers, March 2007) -0.030** 0.004 0.016 -0.002 0.017**
(0.012) (0.009) (0.010) (0.001) (0.008)
Panel B: 2SLS
Log(Twitter users) -0.044** 0.005 0.024 -0.003 0.025**
(0.017) (0.013) (0.015) (0.002) (0.011)
Observations 2,831 2,831 2,831 2,831 2,831
Mean of DV 0.48 0.23 0.09 0.01 0.15
Robust F-stat. 69.54 69.54 69.54 69.54 69.54
Notes: This table presents county-level regressions where the dependent variable is the vote
share of the indicated candidate in the Republican party primaries in 2016. Log(SXSW follow-
ers, March 2007) is the number of Twitter users (in logs, with 1 added inside) who joined in
March 2007 and follow South by Southwest (SXSW). SXSW followers, Pre is the number of
SXSW followers who registered at some point in 2006, defined similarly. Twitter users are the
number of users in 2014-2015. All regressions control for population deciles, Census region
fixed effects, and the full set of controls (as in columns 5 and 10 of Table 2). The first-stage
regressions for 2SLS results (Panel B) are analogous to the one presented in Table 1, except
for the different sample of counties for which primary results are available. The F-stat for
the excluded instrument is provided in the bottom row. Observations are weighted by turnout
in the 2000 presidential election. Standard errors in parentheses are clustered by state. ***
p < 0.01, ** p < 0.05, * p < 0.1.
Results from Gallup Approval Ratings. A similar pattern emerges when we use data from the
Gallup Daily Tracker, which contains approval ratings for three other Republican presidential
candidates who ran alongside Trump during the primaries (Ted Cruz, Marco Rubio, and Kasich).
Table 7 shows the results of running individual-level IV probit regressions as in equation (4), where
the dependent variable is an indicator variable equal to one if the respondent approved of a specific
candidate. As in Table 5, we differentiate between respondents’ political affiliation.43
Table 7 confirms our main county-level result from general elections and primaries: Twitter
usage is associated lower approval of Trump, especially among independents (who are presumably
more likely to be persuaded by social media content). We also find lower approval of Cruz,
42
In Appendix Table C.6, we also investigate the voting behavior in the 2016 and 2020 Democratic primaries. Twitter
usage appears to be associated with a higher support for Bernie Sanders in 2020.
43
We pool people who identify as leaning Republicans or Democrats with independents, because—in contrast to the
CCES data—only a few individuals in the survey are classified as “leaners.”
37
who is substantially more right-wing than other presidential primary candidates in recent years
(FiveThirtyEight, 2015). We find no link between Twitter use and approval of the more moderate
Republican candidates, Rubio and Kasich. For the Democrat candidates, we find an effect for
Clinton’s but not Sanders’ approval.44
Taken together, these results are consistent with Twitter turning voters against voting for
Trump in particular and not against the Republican party more generally. Our results may also
explain the absence of an effect in the 2008 and 2012 elections: Obama’s opponents John McCain
and Mitt Romney were widely considered to be moderate Republicans (e.g., more similar to Kasich
than Trump or Cruz).45
Slant of Election-Related Tweets. We provide further support for the hypothesis that Trump’s
2016 campaign and presidency triggered opposition on Twitter by analyzing the content of more
than 460 million tweets mentioning the last name of presidential candidates during the 2012, 2016,
and 2020 presidential campaigns.
First, we classify the tweets’ slant as Republican, Democrat, or neutral using two approaches
described in Section 3. In the first case, we classify the political affiliation of Twitter users by
counting the number of Democrat and Republican Congress members they follow. If a user follows
more Democrats than Republicans, they are classified as Democrat, and vice-versa. Tweets sent
by a user classified as Democrat are classified as Democrat, and so forth.46 In the second case, we
classify individual tweets (not users) following an approach in the spirit of Gentzkow and Shapiro
(2010) and train a L2 regularized logistic regression classifier to predict whether a tweet is more
likely to have a Democrat or Republican slant, depending on its content’s similarity to tweets sent
44
Note that Appendix Table C.6 reports a significant positive effect for Sanders county-level primary vote share in
2020, but not in 2016. As mentioned earlier, Gallup Daily Tracker data is not available for 2020 at the time of writing.
Note also that Gallup collects data on candidate approval (as opposed to vote choice) for the entire population (as
opposed to Democratic primaries votes), which may explain some differences between these results and the ones based
on county-level primaries results.
45
As previously discussed, the number of Twitter users in 2008 was relatively small, but by 2012, it was relatively
close to its 2016 level (Figure A.1a).
46
Users who follow an equal number of Democrats and Republican or no Congress members are classified as neutral.
This approach is similar in spirit to Barberá (2015), who uses the network of Twitter users to create a measure of
ideology. Because we are only interested in a binary measure of partisan slant and not the ideological distance of users,
we do not estimate the full Bayesian ideal point mode. The advantage of our simplified approach is that it is faster to
compute and the resulting measure is easier to interpret.
38
Table 7: Twitter and Candidate Approval during the 2016 Primaries
Dep. var.: Approved of candidate during primaries

Trump Cruz Rubio Kasich Sanders Clinton
(1) (2) (3) (4) (5) (6)
Panel A: Republicans
Log(Twitter users) -0.108*** -0.086** -0.051 0.018 0.031 0.148***

(0.030) (0.035) (0.060) (0.050) (0.039) (0.041)
Observations 19,974 11,959 8,344 8,995 16,099 20,983
Mean of DV 0.647 0.698 0.779 0.665 0.238 0.092
Marginal effect [-0.038] [-0.029] [-0.014] [0.006] [0.009] [0.022]
Panel B: Independents and Leaners
Log(Twitter users) -0.065** -0.006 -0.015 0.050 0.059 0.154***

(0.028) (0.035) (0.043) (0.042) (0.043) (0.036)
Observations 22,852 12,135 8,080 8,280 17,356 23,813
Mean of DV 0.329 0.392 0.516 0.581 0.595 0.380
Marginal effect [-0.021] [-0.002] [-0.006] [0.019] [0.021] [0.054]
Panel C: Democrats
Log(Twitter users) -0.052 -0.116** -0.036 0.076 0.004 0.081**

(0.051) (0.054) (0.056) (0.051) (0.050) (0.038)
Observations 20,866 11,098 7,460 7,547 16,059 21,454
Mean of DV 0.107 0.195 0.271 0.502 0.808 0.807
Marginal effect [-0.009] [-0.030] [-0.012] [0.029] [0.001] [0.021]
This table presents results estimated using IV probit models, as in equation (4). The dependent
variable is a dummy for individuals who approved the respective presidential candidate during
the presidential primaries in 2015 and 2016. We restrict the sample to the period before Trump
became the presumptive nominee in June 2016. Log(Twitter users) is instrumented using the
number of SXSW followers that joined Twitter in March 2007. All regressions control for the
(log) number of SXSW followers that joined Twitter at some point in 2006, income, gender,
education, and marital status, as well as county-level population deciles and Census region
fixed effects. Regressions are weighted by survey weights. Standard errors in parentheses are
clustered by state. *** p < 0.01, ** p < 0.05, * p < 0.10.
by Congress members. If a tweet’s content has higher similarity with those of Democratic Congress
members, it is classified as Democratic, and as Republican otherwise.47
Figure 8 plots the amount of Twitter attention directed at the Republican and Democratic
presidential candidates in the 2012, 2016, and 2020 elections, as well as the tweets’ estimated
slant based on users’ following of Congress members. To account for the attention and popularity
of tweets, we base the graphs on the number of “likes” the tweets mentioning the last name of
candidates received. In Appendix Figure D.1, we confirm that the results are similar using the
number of tweets and when we base the slant classification on the text of the tweets (Figure D.2).48
47
See Section 3 and Appendix A.1. in the Appendix for more details. In unreported robustness checks, we confirm
that these findings are robust using different slant cutoffs for classifying Republican and Democratic tweets, such as
only classifying tweets for which the class probabilities are above 75% or 90%.
48
Twitter users can choose to “like” each tweet they see in their “timeline.” A user can only “like” a particular tweet
once. “Likes” thus provide an useful metric since they capture the popularity and attention that the content received. For
39
Panel A shows the number of “likes” for tweets mentioning the Republican presidential
candidate (Romney and Trump), while Panel B provides similar evidence for the Democrats (Obama,
Clinton, and Biden). There are three noteworthy facts presented in the figure. First, there was a
sizable growth in the overall volume of Twitter content mentioning presidential candidates. Second,
the number of “likes” for tweets mentioning Trump is larger than those mentioning his opponents
(the difference is fourfold in 2016 and almost threefold in 2020). Note that a “like” for a tweet
mentioning a candidate can occur for tweets that are positive or negative about the candidate, so the
overall size of the bars are not informative about slant or sentiments of Twitter content.
Third, Figure 8 also breaks down the share of tweets mentioning the candidates by slant.
The content sent by users classified as Democrats are more sizable than that from those classified
as Republicans. In particular, the amount of attention (proxied via “likes”) on Twitter content
mentioning Trump posted by Democrats is almost twice as large as the amount posted from
Republicans (for both 2016 and 2020). On the other hand, content on Biden was more likely to
have a Democrat slant and content on Clinton was almost equally likely to have a Democrat or
Republican slant.
This pattern of results suggests that, in 2016 and 2020, Twitter became a vehicle for spreading
opinions, particularly from Democratic-slanted users, on Trump. This may, in turn, have persuaded
voters with weaker priors—independents and perhaps more moderate Republicans—to vote against
Trump in the presidential election. This is likely due to a combination of i) Twitter (and other social
media) users being more likely to be young, well-educated, live in dense urban areas, and support
the Democratic party (see Section 2), and ii) Trump’s prominent behavior on the platform himself,
which could also have had its own direct effect as moderate voters’ “backlash” against the content
of his tweets.
It is difficult to separate whether our results are better represented as “Presidential-race-
specific” or “Trump-specific.” The 2012 election differed from 2016 and 2020 both in terms of
having less Twitter content referring to it (as Figure D.2 indicates) and also for having a more
moderate Republican candidate (Romney) that behaved differently on social media and may not
have attracted as much Democrat-slanted content.
example, if an account sent millions of Republican-slanted tweets about Clinton, but such account had few followers
and thus few users who can “like” the message, it would not meaningfully affect measures based on “likes,” but could
potentially do so for measures based on number of tweets.
40
Figure 8: Twitter’s Partisan Slant Across Presidential Elections
(a) Tweets about Republican Presidential Can- (b) Tweets about Democratic Presidential
didates Candidates
24%
600
26% 600
50%
Number of Likes (in millions)

500 500
400 400
300 300 18%

37%
200 200 45%
15%
26%
59% 14%
100 27% 100 29%
44%
11% 23%
42%
0
62% 0
48%
Election 2012 Election 2016 Election 2020 Election 2012 Election 2016 Election 2020
Democratic Slant Republican Slant Neutral Slant Democratic Slant Republican Slant Neutral Slant
Notes: These figures present the number of “likes” received by tweets that contain the last name of the candidates
in the 2012, 2016 and 2020 presidential elections, depending on whether the tweet was classified as having a
Republican, Democratic, or neutral slant. We classify the slant of a tweet based on the Twitter network of the
user who sent the tweet. If the user follows more Democratic than Republican Congress members, they will be
classified as a Democrat, and vice versa. Users who follow an equal number of Democrats and Republican or no
Congress members are classified as neutral.
7 Conclusion
Election officials around the globe are concerned about social media’s increasing influence on voting
decisions (e.g. NPR, 2020a). At the time of writing, there is a heated debate about whether platform
providers should “moderate” election-related content in the U.S. (e.g. Politico, 2020). Exploiting
variation based on a shock to Twitter’s initial rise to popularity, our paper provides some of the first
empirical evidence that social media can affect election outcomes.
We find that Twitter lowered the Republican party’s vote share in the 2016 and 2020 presiden-
tial elections. While this finding runs counter to a popular narrative that places social media at the
heart of Trump’s election win, it is consistent with a growing body of evidence showing that social
media users were less, not more likely to vote for Trump in 2016 or hold polarized views (Boxell
et al., 2017, 2018).
We also provide support for the idea that the demographics of Twitter users may account for
the platform’s partisan effects. People who use Twitter are 25 percentage points more likely to
identify as Democrats rather than Republicans, and Democratic politicians are more popular on
Twitter than Republican ones. Our work suggests that this environment not only reflects selection
41
of like-minded individuals, but also affects voting decisions, particularly for people with more
moderate views.
42
References
Acemoglu, D., T. A. Hassan, and A. Tahoun (2017, 08). The Power of the Street: Evidence from
Egypt’s Arab Spring. The Review of Financial Studies 31(1), 1–42.
Allcott, H., L. Braghieri, S. Eichmeyer, and M. Gentzkow (2020, March). The Welfare Effects of
Social Media. American Economic Review 110(3), 629–76.
Allcott, H. and M. Gentzkow (2017, May). Social Media and Fake News in the 2016 Election.
Journal of Economic Perspectives 31(2), 211–36.
Altonji, J. G., T. E. Elder, and C. R. Taber (2005). Selection on observed and unobserved variables:
Assessing the effectiveness of catholic schools. Journal of political economy 113(1), 151–184.
Andrews, I. and E. Oster (2019). A Simple Approximation for Evaluating External Validity Bias.
Economics Letters 178(C), 58–62.
Autor, D., D. Dorn, G. Hanson, K. Majlesi, et al. (2020). Importing Political Polarization? The
Electoral Consequences of Rising Trade Exposure. American Economic Review 110(10), 3139–83.
Bail, C. A., L. P. Argyle, T. W. Brown, J. P. Bumpus, H. Chen, M. B. F. Hunzaker, J. Lee, M. Mann,
F. Merhout, and A. Volfovsky (2018). Exposure to Opposing Views on Social Media Can Increase
Political Polarization. Proceedings of the National Academy of Sciences 115(37), 9216–9221.
Bakshy, E., S. Messing, and L. A. Adamic (2015). Exposure to Ideologically Diverse News and
Opinion on Facebook. Science 348(6239), 1130–1132.
Barberá, P. (2014). How Social Media Reduces Mass Political Polarization. Evidence from Germany,
Spain, and the US.
Barberá, P. (2015). Birds of the Same Feather Tweet Together: Bayesian Ideal Point Estimation
Using Twitter Data. Political analysis 23(1), 76–91.
Beam, M. A., M. J. Hutchens, and J. D. Hmielowski (2018). Facebook News and (De)polarization:
Reinforcing Spirals in the 2016 US Election. Information, Communication & Society 21(7),
940–958.
Bessi, A., F. Zollo, M. D. Vicario, A. Scala, G. Caldarelli, and W. Quattrociocchi (2015, 08). Trend
of Narratives in the Age of Misinformation. PLOS ONE 10(8), 1–16.
Bessone, P., F. R. Campante, C. Ferraz, and P. Souza (2022). Social media and the behavior of
politicians: Evidence from facebook in brazil. Technical report, National Bureau of Economic
Research.
Bond, R. M., C. J. Fariss, J. J. Jones, A. D. Kramer, C. Marlow, J. E. Settle, and J. H. Fowler (2012).
A 61-Million-Person Experiment in Social Influence and Political Mobilization. Nature 489(7415),
295.
Boxell, L., M. Gentzkow, and J. M. Shapiro (2017). Greater Internet Use Is Not Associated with
Faster Growth in Political Polarization Among US Demographic Groups. Proceedings of the
National Academy of Sciences of the United States of America, 201706588.
Boxell, L., M. Gentzkow, and J. M. Shapiro (2018, 07). A Note on Internet Use and the 2016 U.S.
Presidential Election Outcome. PLOS ONE 13(7), 1–7.
Boxell, L., M. Gentzkow, and J. M. Shapiro (2019). Cross-Country Trends in Affective Polarization.
Working Paper.
Braghieri, L., R. Levy, and A. Makarin (2021). Social Media and Mental Health. Available at SSRN:
https://ssrn.com/abstract=3919760 or http://dx.doi.org/10.2139/ssrn.3919760.
43
Bursztyn, L., G. Egorov, R. Enikolopov, and M. Petrova (2019, December). Social Media and
Xenophobia: Evidence from Russia. Working Paper 26567, National Bureau of Economic
Research.
Bursztyn, L., G. Egorov, and S. Fiorin (2020). From Extreme to Mainstream: How Social Norms
Unravel. American Economic Review 110(11), 3522–3548.
Campante, F., R. Durante, and F. Sobbrio (2017, 12). Politics 2.0: The Multifaceted Effect of Broad-
band Internet on Political Participation. Journal of the European Economic Association 16(4),
1094–1136.
Campante, F. and D. Hojman (2013). Media and Polarization: Evidence from the Introduction of
Broadcast TV in the US. Journal of Public Economics 100, 79–92.
Chen, E., A. Deb, and E. Ferrara (2020). # election2020: The first public twitter dataset on the 2020
us presidential election. arXiv preprint arXiv:2010.00600.
Chen, Y. and D. Y. Yang (2019, June). The Impact of Media Censorship: 1984 or Brave New World?
American Economic Review 109(6), 2294–2332.
Chernozhukov, V., C. Hansen, and M. Spindler (2015a). Post-selection and post-regularization
inference in linear models with many controls and instruments. American Economic Review 105(5),
486–90.
Chernozhukov, V., C. Hansen, and M. Spindler (2015b). Valid post-selection and post-regularization
inference: An elementary, general approach. Annu. Rev. Econ. 7(1), 649–688.
Colella, F., R. Lalive, S. O. Sakalli, and M. Thoenig (2019). Inference with Arbitrary Clustering.
IZA Discussion Papers 12584, Institute of Labor Economics (IZA).
Conover, M. D., J. Ratkiewicz, M. Francisco, B. Gonçalves, F. Menczer, and A. Flammini (2011).
Political Polarization on Twitter. In Fifth International AAAI Conference on Weblogs and Social
Media.
Del Vicario, M., A. Bessi, F. Zollo, F. Petroni, A. Scala, G. Caldarelli, H. E. Stanley, and W. Quattro-
ciocchi (2016). The Spreading of Misinformation Online. Proceedings of the National Academy
of Sciences 113(3), 554–559.
DellaVigna, S., R. Enikolopov, V. Mironova, M. Petrova, and E. Zhuravskaya (2014, July). Cross-
Border Media and Nationalism: Evidence from Serbian Radio in Croatia. American Economic
Journal: Applied Economics 6(3), 103–32.
DellaVigna, S. and E. L. Ferrara (2015). Chapter 19 - Economic and Social Impacts of the Media. In
S. P. Anderson, J. Waldfogel, and D. Strömberg (Eds.), Handbook of Media Economics, Volume 1
of Handbook of Media Economics, pp. 723 – 768. North-Holland.
DellaVigna, S. and M. Gentzkow (2010). Persuasion: Empirical Evidence. Annual Review of
Economics 2(1), 643–669.
DellaVigna, S. and E. Kaplan (2007). The Fox News Effect: Media Bias and Voting. The Quarterly
Journal of Economics 122(3), 1187–1234.
Diaz, F., M. Gamon, J. M. Hofman, E. Kıcıman, and D. Rothschild (2016). Online and Social Media
Data as an Imperfect Continuous Panel Survey. PloS one 11(1), e0145406.
Eady, G., J. Nagler, A. Guess, J. Zilinsky, and J. A. Tucker (2019). How Many People Live in
Political Bubbles on Social Media? Evidence From Linked Survey and Twitter Data. SAGE
Open 9(1), 2158244019832705.
eMarketer (2019). US Time Spent with Social Media 2019, By Debra Aho Williamson.
Enikolopov, R., A. Makarin, and M. Petrova (2020). Social Media and Protest Participation:
Evidence from Russia. Econometrica 88(4), 1479–1514.
44
Enikolopov, R. and M. Petrova (2015). Chapter 17 - Media Capture: Empirical Evidence. In S. P.
Anderson, J. Waldfogel, and D. Strömberg (Eds.), Handbook of Media Economics, Volume 1 of
Handbook of Media Economics, pp. 687 – 700. North-Holland.
Enikolopov, R., M. Petrova, and E. Zhuravskaya (2011). Media and Political Persuasion: Evidence
from Russia. The American Economic Review 101(7), 3253–3285.
Falck, O., R. Gold, and S. Heblich (2014, July). E-lections: Voting Behavior and the Internet.
Fergusson, L. and C. Molina (2021, April). Facebook Causes Protests. Documentos CEDE 018002,
Universidad de los Andes - CEDE.
FiveThirtyEight (2015). Let’s Be Serious About Ted Cruz From The Start: He’s Too Extreme And
Too Disliked To Win.
Flaxman, S., S. Goel, and J. M. Rao (2016, 03). Filter Bubbles, Echo Chambers, and Online News
Consumption. Public Opinion Quarterly 80(S1), 298–320.
Gavazza, A., M. Nardotto, and T. Valletti (2019). Internet and Politics: Evidence From UK Local
Elections and Local Government Policies. The Review of Economic Studies 86(5), 2092–2135.
Gentzkow, M. (2006). Television and Voter Turnout. The Quarterly Journal of Economics 121(3),
931–972.
Gentzkow, M. and J. M. Shapiro (2010). What Drives Media Slant? Evidence from US Daily
Newspapers. Econometrica 78(1), 35–71.
Gentzkow, M. and J. M. Shapiro (2011). Ideological Segregation Online and Offline. The Quarterly
Gentzkow, M., J. M. Shapiro, and M. Sinkinson (2011, December). The Effect of Newspaper Entry
and Exit on Electoral Politics. American Economic Review 101(7), 2980–3018.
Gerber, A. S., J. G. Gimpel, D. P. Green, and D. R. Shaw (2011). How Large and Long-lasting Are
the Persuasive Effects of Televised Campaign Ads? Results from a Randomized Field Experiment.
The American Political Science Review 105(1), 135–150.
Gerber, A. S. and D. P. Green (2000). The Effects of Canvassing, Telephone Calls, and Direct Mail
on Voter Turnout: A Field Experiment. The American Political Science Review 94(3), 653–663.
Gerber, A. S., D. Karlan, and D. Bergan (2009). Does the Media Matter? A Field Experiment
Measuring the Effect of Newspapers on Voting Behavior and Political Opinions. American
Economic Journal: Applied Economics 1(2), 35–52.
Guess, A. (2018). (Almost) Everything in Moderation: New Evidence in Americans’ Online Media
Diets.
Guriev, S., N. Melnikov, and E. Zhuravskaya (2020, 12). 3G Internet and Confidence in Government.
The Quarterly Journal of Economics.
Halberstam, Y. and B. Knight (2016). Homophily, Group Size, and the Diffusion of Political
Information in Social Networks: Evidence from Twitter. Journal of Public Economics 143, 73 –
88.
Hargittai, E. (2015). Is Bigger Always Better? Potential Biases of Big Data Derived from Social
Network Sites. The Annals of the American Academy of Political and Social Science 659, 63–76.
Howard, P. N., A. Duffy, D. Freelon, M. Hussain, W. Mari, and M. Maziad (2011). Opening Closed
Regimes: What Was the Role of Social Media During the Arab Spring? Working Paper.
Huber, G. A. and K. Arceneaux (2007). Identifying the Persuasive Effects of Presidential Advertising.
American Journal of Political Science 51(4), 957–977.
45
Jones, J. J., R. M. Bond, E. Bakshy, D. Eckles, and J. H. Fowler (2017, 04). Social Influence
and Political Mobilization: Further Evidence From a Randomized Experiment in the 2012 U.S.
Presidential Election. PLOS ONE 12(4), 1–9.
Kinder-Kurlanda, K., K. Weller, W. Zenk-Möltgen, J. Pfeffer, and F. Morstatter (2017). Archiving
Information from Geotagged Tweets to Promote Reproducibility and Comparability in Social
Media Research. Big Data & Society 4(2), 2053951717736336.
Larcinese, V. and L. Miner (2017, June). The Political Impact of the Internet on US Presidential
Elections. STICERD - Economic Organisation and Public Policy Discussion Papers Series 63,
Suntory and Toyota International Centres for Economics and Related Disciplines, LSE.
Lelkes, Y., G. Sood, and S. Iyengar (2017). The Hostile Audience: The Effect of Access to
Broadband Internet on Partisan Affect. American Journal of Political Science 61(1), 5–20.
Levy, R. (2021, March). Social Media, News Consumption, and Polarization: Evidence from a Field
Experiment. American Economic Review 111(3), 831–70.
Littman, J., L. Wrubel, and D. Kerchner (2016). 2016 United States Presidential Election Tweet IDs.
Manacorda, M. and A. Tesei (2016, May). Liberation Technology: Mobile Phones and Political
Mobilization in Africa. CEPR Discussion Papers 11278, C.E.P.R. Discussion Papers.
Martin, G. J. and A. Yurukoglu (2017, September). Bias in Cable News: Persuasion and Polarization.
Mosquera, R., M. Odunowo, T. McNamara, X. Guo, and R. Petrie (2020). The Economic Effects of
Facebook. Experimental Economics 23(2), 575–602.
Müller, K. and C. Schwarz (2019). From Hashtag to Hate Crime: Twitter and Anti-Minority
Sentiment. Working Paper.
Müller, K. and C. Schwarz (2020). Fanning the Flames of Hate: Social Media and Hate Crime.
Journal of the European Economic Association.
Napoli, P. M. (2014). Measuring Media Impact. The Norman Lear Center. http://www.learcenter.
org/pdf/measuringmedia.pdf .
Nelson, J. L. and J. G. Webster (2017). The Myth of Partisan Selective Exposure: A Portrait of the
Online Political News Audience. Social Media + Society 3(3), 2056305117729314.
New York Magazine (2022). The ‘Twitter Files’ Is What It Claims to Expose .
New York Times (2017). The Fake Americans Russia Created to Influence the Election, By Scott
Shane.
NPR (2020a). Facebook Clamps Down On Posts, Ads That Could Undermine U.S. Presidential
Election.
NPR (2020b). Facebook Keeps Data Secret, Letting Conservative Bias Claims Persist.
NPR (2020c). President-Elect Joe Biden Hits 80 Million Votes In Year Of Record Turnout.
NPR (2022). Elon Musk Is Using the Twitter Files to Discredit Foes and Push Conspiracy Theories.
Oliver, J. E., T. Wood, and A. Bass (2016). Liberellas versus Konservatives: Social Status, Ideology,
and Birth Names in the United States. Political Behavior 38(1), 55–81.
Oster, E. (2019). Unobservable Selection and Coefficient Stability: Theory and Evidence. Journal
of Business & Economic Statistics 37(2), 187–204.
Pariser, E. (2011). The Filter Bubble: What The Internet Is Hiding From You. Penguin Books
Limited.
Pedregosa, F., G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Pretten-
hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and
46
E. Duchesnay (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning
Research 12, 2825–2830.
Pew Research Center (2009). The Internet’s Role in Campaign 2008. Technical report.
Pew Research Center (2011). Social Networking Sites and Our Lives. Technical report.
Pew Research Center (2014). Political Polarization & Media Habits. Technical report.
Pew Research Center (2015). Americans’ Internet Access: 2000-2015. Technical report.
Pew Research Center (2018). For Most Trump Voters, ‘Very Warm’ Feelings for Him Endured.
Technical report.
Pew Research Center (2019a). About One-In-Five Adult Twitter Users in the U.S. Follow Trump.
Technical report.
Pew Research Center (2019b). National Politics on Twitter: Small Share of U.S. Adults Produce
Majority of Tweets. Technical report.
Pew Research Center (2019c). Share of U.S. Adults Using Social Media, Including Facebook, Is
Mostly Unchanged Since 2018. Technical report.
Pew Research Center (2019d). Sizing Up Twitter Users. Technical report.
Pew Research Center (2020). Democrats On Twitter More Liberal, Less Focused On Compromise
Than Those Not On The Platform. Technical report.
Politico (2020). Biden Campaign Lashes Out at New York Post.
Quora (2011). What is the process involved in launching a start-up at SXSW?
Rotesi, T. (2019). The Impact of Twitter on Political Participation. Technical report.
Schmidt, A. L., F. Zollo, M. Del Vicario, A. Bessi, A. Scala, G. Caldarelli, H. E. Stanley, and
W. Quattrociocchi (2017). Anatomy of News Consumption on Facebook. Proceedings of the
National Academy of Sciences 114(12), 3035–3039.
Solon, G., S. J. Haider, and J. M. Wooldridge (2015). What Are We Weighting For? Journal of
Human resources 50(2), 301–316.
Spenkuch, J. L. and D. Toniatti (2018, 05). Political Advertising and Election Results. The Quarterly
Statista (2019). Twitter: Number of Monthly Active U.S. Users 2010-2019.
Statista (2020). Leading countries based on number of Twitter users as of October 2020.
Strömberg, D. (2015). Media and Politics. Annual Review of Economics 7(1), 173–205.
Sunstein, C. R. (2017). # Republic: Divided Democracy in the Age of Social Media. Princeton
University Press.
The Guardian (2018). Revealed: 50 Million Facebook Profiles Harvested for Cambridge Analytica
in Major Data Breach, by Carole Cadwalladr and Emma Graham-Harrison.
The Guardian (2023). Twitter’s Rightwing Takeover Is Complete. Why Are Liberals Still on It?
The Hill (2016). Stop Blaming Facebook for Trump’s Election Win, By Keith N. Hampton and
Eszter Hargittai.
The New Yorker (2016). Obama Reckons with a Trump Presidency, By David Remnick.
Twitter (2010). Measuring Tweets, by Kevin Weil.
Twitter (2020). Empowering US Voters on National Voter Registration Day, by Bridget Coyne.
UK Parliament (2019). Disinformation and ‘fake news’: Final Report, By Digital, Culture, Media
and Sport Committee.
USA Today (2023). Is Elon Musk’s Twitter More Conservative? Marjorie Taylor Greene, Other
GOP Stars Win Big.
Venture Beat (2008). Hitwise: Twitter Traffic Is, in Fact, Going up but Still Not Big, by Eric Eldon.
47
Wall Street Journal (2020). Twitter’s Partisan Censors.
Weber, I., V. R. K. Garimella, and A. Batayneh (2013). Secular vs. Islamist Polarization in Egypt on
Twitter. In Proceedings of the 2013 IEEE/ACM International Conference on Advances in Social
Networks Analysis and Mining, ASONAM ’13, New York, NY, USA, pp. 290–297. Association
for Computing Machinery.
Wired (2016). Here’s How Facebook Actually Won Trump the Presidency, By Issie Lapowsky.
Zhuravskaya, E., M. Petrova, and R. Enikolopov (2020). Political Effects of the Internet and Social
Media. Technical report.
48
Online Appendix
This appendix presents further details on data construction, the SXSW festival, additional robustness
exercises, further results, and the LATE extrapolation:
• Appendix A provides additional details on the data.
• Appendix B discusses the SXSW instrument.
• Appendix C shows additional robustness checks.
• Appendix D provides further results.
• Appendix E discusses extrapolation for the average treatment effect.
A Additional Details on Data
Figure A.1: Validation of Twitter usage measure

(a) Twitter Usage over Time (b) Twitter Usage by County (Gesis vs GfK)
80 12
Statista (right axis)
5
Log(# of households using Twitter)
Number of Twitter users (in millions)
10
60
4
3 8
40
2
6
20
1 Gesis (left axis)
4
0 0 2 4 6 8 10
2007q1 2009q1 2011q1 2013q1 2015q1 2017q1 2019q1 Log(Twitter users)
Notes: This graph shows two validation exercises for the Twitter usage measure in the Gesis data (Kinder-
Kurlanda et al., 2017). Panel (a) plots the number of Twitter users in the Gesis data and the number of active
monthly users reported by Statista based on Twitter’s own reporting. Panel (b) plots the percentiles of the
number of Twitter users in the Gesis data at the county-level against the number of users based on the GfK
Media Survey.
1
Figure A.2: Twitter Reach by Party (Median)
(a) Median # of tweets (per account) (b) Median # of followers (per account)
4,000
25,000
3,000 20,000
15,000
2,000
10,000
1,000
5,000
0 0
(c) Median # of likes (per tweet) (d) Median # of retweets (per tweet)
30 150
20 100
10 50
0 0
Notes: This figure plots data on the Twitter reach of Congress members. The sample includes all 901 senators and
House representatives who were in office between 2007 and 2019 for whom we could identify a Twitter account.
For each account, we plot the median number of tweets and followers, and the median number of “likes” and
retweets of their tweets. The data were collected from Twitter in November 2019.
Figure A.3: News Reports About Social Media
4000
Twitter
Number of news reports on social media
Facebook
3000 2016 Election
2000 2012 Election
1000
0
1
1
m
m
04
06
08
10
12
14
16
18
20
20
20
20
20
20
20
20
Notes: This graph plots the number of times the terms “Twitter” and “Facebook” are mentioned
in USA Today, The Washington Post, The New York Post, and The New York Times based on
data from Nexis.
2
Table A.1: Summary Statistics (County-Level)
Mean Std. Dev. Min. Median Max. N

Vote outcomes and Twitter data
Republican two-party vote share (2016) 0.46 0.17 0.08 0.45 0.95 3,065
Change in Republican two-party vote share, 2000-2016 -0.02 0.10 -0.33 -0.03 0.45 3,065
Change in Republican two-party vote share, 2000-2020 -0.01 0.10 -0.34 -0.02 0.48 3,065
Log(Twitter users) 8.22 1.99 0.00 8.45 12.35 3,065
Log(SXSW followers, March 2007) 0.69 1.13 0.00 0.00 4.98 3,065
Log(SXSW followers, Pre) 0.33 0.73 0.00 0.00 3.61 3,065
Geographical controls
Population density 1925.15 6342.94 0.10 508.30 69468.40 3,065
Log(County area) 6.72 0.92 3.26 6.64 9.91 3,065
Distance from Austin, TX (in miles) 1731.21 653.61 5.04 1750.86 3098.88 3,065
Distance from Chicago (in miles) 1246.18 813.51 7.16 1103.75 3040.38 3,065
Distance from NYC (in miles) 1600.47 1255.14 6.48 1285.98 4191.67 3,065
Distance from San Francisco (in miles) 2841.16 1231.98 41.11 3157.16 4565.01 3,065
Distance from Washington, DC (in miles) 1448.51 1175.55 7.88 1047.13 3983.08 3,065
Demographic controls
% aged 20-24 0.07 0.02 0.01 0.06 0.27 3,065
% aged 25-29 0.07 0.01 0.03 0.07 0.15 3,065
% aged 30-34 0.07 0.01 0.03 0.06 0.12 3,065
% aged 35-39 0.06 0.01 0.03 0.06 0.10 3,065
% aged 40-44 0.06 0.01 0.02 0.06 0.10 3,065
% aged 45-49 0.06 0.01 0.02 0.06 0.09 3,065
% aged 50+ 0.36 0.06 0.11 0.35 0.75 3,065
Population growth, 2000-2016 0.14 0.19 -0.43 0.10 1.32 3,065
% white 0.65 0.21 0.03 0.68 0.98 3,065
% black 0.12 0.12 0.00 0.08 0.85 3,065
% native American 0.01 0.03 0.00 0.00 0.90 3,065
% Asian 0.05 0.06 0.00 0.03 0.37 3,065
% Hispanic 0.15 0.15 0.01 0.09 0.96 3,065
% unemployed 5.31 1.42 1.80 5.10 24.10 3,065
Socioeconomic controls
% below poverty level 15.11 5.34 1.40 15.10 53.30 3,065
% employed in IT 0.02 0.02 0.00 0.02 0.21 3,065
% employed in construction/real estate 0.07 0.03 0.00 0.07 1.00 3,065
% employed in manufacturing 0.11 0.08 0.00 0.08 0.72 3,065
% adults with high school degree 28.13 7.41 8.30 27.50 54.80 3,065
% adults with college degree 20.98 3.66 8.70 20.90 35.60 3,065
% watching Fox News 0.26 0.01 0.23 0.26 0.30 3,064
% watching prime time TV 0.43 0.01 0.40 0.43 0.47 3,064
China shock controls
Exposure to Chinese import competition 2.63 2.02 -0.63 2.10 43.08 3,065
Share of routine occupations 31.87 2.36 22.23 32.14 36.66 3,065
Average offshorability index -0.02 0.50 -1.64 0.09 1.24 3,065
1996 election control
Notes: This table presents descriptive statistics for the main estimation sample, weighted by the turnout in the 2000 presiden-
tial elections.
3
Table A.2: Summary Statistics (2016 CCES Individual-Level)

Vote outcome
Voted for Trump 0.49 0.50 0.00 0.00 1.00 146,579
Twitter data
Individual control variables
Log(Age) 3.89 0.37 2.89 3.99 4.61 146,579
Family income dummies 7.12 3.65 1.00 7.00 13.00 146,579
Female dummy 1.52 0.50 1.00 2.00 2.00 146,579
Education dummies 3.54 1.54 1.00 3.00 6.00 146,579
Marital status dummies 2.36 1.71 1.00 1.00 5.00 146,579
Interest in news dummies 1.61 0.98 1.00 1.00 7.00 146,579
Notes: This table presents descriptive statistics for the CCES estimation sample, weighted by survey
weights.
Table A.3: Summary Statistics (Gallup Individual-Level)

Candidate approval outcomes
Approve of Trump? 0.34 0.48 0.00 0.00 1.00 64,764
Approve of Kasich? 0.60 0.49 0.00 1.00 1.00 8,735
Approve of Rubio? 0.50 0.50 0.00 1.00 1.00 6,201
Approve of Cruz? 0.41 0.49 0.00 0.00 1.00 11,504
Approve of Sanders? 0.57 0.50 0.00 1.00 1.00 27,137
Approve of Clinton? 0.43 0.50 0.00 0.00 1.00 36,367
Twitter data
Individual control variables
Income dummies 6.99 2.38 1.00 7.00 10.00 64,764
Female dummy 1.50 0.50 1.00 2.00 2.00 64,764
Education dummies 3.58 1.60 1.00 4.00 6.00 64,764
Marital status dummies 1.98 0.94 1.00 2.00 5.00 64,764
Age deciles 4.45 2.68 1.00 4.00 10.00 64,764
Notes: This table presents descriptive statistics for the Gallup estimation sample, weighted by survey
weights.
4
Table A.4: Instrument Balancedness
March 2007 March 2007 Pre Difference

and Pre only only in means Šidàk
(1) (2) (3) (2) - (3) p-value p-value
Population density 5192.27 1021.39 1998.35 -976.96 0.07* 0.91
Log(County area) 6.30 6.63 6.54 0.09 0.73 1.00
Distance from Austin, TX (in miles) 1775.99 1749.38 1626.64 122.74 0.48 1.00
Distance from Chicago (in miles) 1439.45 1329.47 1214.42 115.05 0.53 1.00
Distance from NYC (in miles) 1685.31 1594.99 1510.05 84.94 0.78 1.00
Distance from San Francisco (in miles) 2751.83 2900.11 2833.01 67.10 0.83 1.00
Distance from Washington, DC (in miles) 1558.55 1450.23 1397.05 53.18 0.85 1.00
% aged 20-24 0.07 0.08 0.08 0.00 0.92 1.00
% aged 25-29 0.09 0.07 0.07 -0.00 0.51 1.00
% aged 30-34 0.08 0.07 0.07 -0.00 0.58 1.00
% aged 35-39 0.07 0.06 0.06 -0.00 0.82 1.00
% aged 40-44 0.06 0.06 0.06 0.00 0.82 1.00
% aged 45-49 0.07 0.06 0.06 0.00 0.89 1.00
% aged 50+ 0.32 0.35 0.35 -0.00 0.97 1.00
Population growth, 2000-2016 0.18 0.18 0.15 0.03 0.56 1.00
% white 0.50 0.65 0.67 -0.02 0.62 1.00
% black 0.18 0.12 0.08 0.04 0.20 1.00
% native American 0.01 0.01 0.02 -0.02 0.02** 0.45
% Asian 0.10 0.05 0.05 -0.01 0.55 1.00
% Hispanic 0.20 0.16 0.15 0.01 0.80 1.00
% unemployed 4.86 5.05 4.51 0.54 0.07* 0.91
% below poverty level 15.71 15.82 13.69 2.14 0.17 1.00
% employed in IT 0.04 0.02 0.02 -0.00 0.98 1.00
% employed in construction/real estate 0.06 0.07 0.07 0.01 0.39 1.00
% employed in manufacturing 0.07 0.09 0.07 0.02 0.16 0.99
% adults with high school degree 21.76 25.99 25.77 0.22 0.88 1.00
% adults with college degree 18.89 21.16 21.20 -0.04 0.97 1.00
% watching Fox News 0.25 0.26 0.26 -0.00 0.91 1.00
% watching prime time TV 0.42 0.43 0.43 0.00 0.91 1.00
Exposure to Chinese import competition 2.55 2.46 2.79 -0.32 0.54 1.00
Share of routine occupations 32.47 31.38 31.25 0.13 0.82 1.00
Average offshorability index 0.37 -0.07 -0.05 -0.02 0.84 1.00
Republican two-party vote share (1996) 0.36 0.42 0.42 -0.00 0.90 1.00
Notes: This table presents the averages for the main control variables separately for the three types of counties relevant
for our identification strategy: 1) the 47 counties with SXSW followers that joined Twitter both in March 2007 and
the “pre-period”; 2) the 108 counties with SXSW followers that joined in March 2007 (but none in the “pre-period”);
and 3) the 20 counties with SXSW that joined in the “pre-period” (but none in March 2007). The demographic and
socioeconomic controls are measured in 2016. We report p-values from a two-sided t-test for the equality of means
between the counties with the key identifying variation, as well as Šidàk-corrected values to account for multiple
hypothesis testing. *** p < 0.01, ** p < 0.05, * p < 0.1.
5
A.1. Additional Details on the Logistic Regression Classifier
We train a separate machine learning classifier for each of the three election years in our data
using the Python sci-kit package (Pedregosa et al., 2011). These classifiers help us to determine
whether tweets are more likely to be sent by Democratic-leaning or Republican-leaning users. The
classification process starts with the preparation of the underlying Twitter data. The inputs are the
text of each of the 4,300,579 tweets from U.S. Congress members. To focus on election-related
tweets, we restrict the sample to tweets that were sent either in the election year or in the year
leading up to the election and mention at least one of the presidential candidates.
The target variable y for the classifier is an indicator variable equal to one for tweets sent
by Republicans and zero otherwise. The feature matrix X for the classifier are created by count-
vectorizing the texts of the tweets. In other words, we transform the text of the tweets into n × v
matrix, where n is the number of tweets and v is the number of unique 1,2-grams that occur in the
tweets. In preparation for this step, we removed common words (stopwords), links, and special
characters from the tweets. Additionally, we reduced the words in the tweets to their morphological
roots using a lemmatizer, which improves the performance of the classifier. As an example, the
lemmatizer changes words like “walking” and “walked” to “walk”. Lastly, we reweight the n-grams
v of tweet i using term frequency–inverse document frequency (tf-idf):
1+T
tf idf (fi,v ) = (1 + ln(fi,v ) · (ln( ) + 1) (A.1)
1 + dv
where dv is the number of tweets n-gram v appears in. This reweighting reduces the importance of
words that appear frequently in many tweets, which help little to discriminate between tweets. The
tf-idf vectorizer also normalizes the feature matrix by its L2-norm. The vector y and the matrix X
then serve as the input for a L2 regularized logistic regression classifier. The classifier minimizes
the following cost function49 :
n
X 1
min C log(exp(−yi (XiT w + c) + 1) + wT w (A.2)
w,c
i=1
2
where w are the weights (coefficients) of the logistic regression, c is a constant (intercept), and
C is the inverse of the regularization strength. Larger values of C imply weaker regularization.
For C −→ ∞, the classifier converges towards a normal logistic regression. As is standard in
most machine learning applications, we choose the optimal regularization strength C using 10-fold
cross-validation. This involves randomly splitting the training data into ten equal slices. Nine of
49
Note that this formulation of the cost function assumes that yi takes values −1; 1. We use this formulation in line
with the sci-kit documentation.
6
Figure A.4: Most Predictive Terms Of “Republican” Tweets by Election
(a) 2012 Election (b) 2016 Election
(c) 2020 Election
Notes: This word cloud plots the n-grams most predictive of tweets sounding like those of Republican Congress
members, as identified by the logistic regression classifier for each election cycle. The size of the word represents
the magnitude of the coefficients.
the ten slices are then used to train the classifier, while the out-of-sample performance is evaluated
against the remaining slice using F1-scores.
The final classifiers achieves an out-of-sample F1-score of 0.916 in 2012, of 0.843 in 2016,
and of 0.904 in 2020. The classifiers, therefore, accurately predict the party affiliation of Congress
members. We then take these classifiers and apply it to the universe of tweets sent during the 2012,
2016, and 2020 presidential elections. For each tweet in the election data, the classifiers provide
us with a predicted class (either Democrat or Republican) and a probability for this class label. To
avoid that our results are driven by tweets for which the classifier is “uncertain,” we code tweets
with a predicted class probability below 60% as neutral. This adjustment has no bearing on our
findings. In spirit, this approach is similar to the work of Gentzkow and Shapiro (2010). While they
identify expressions that are more frequently used by Democrats and Republicans by hand, we use a
machine learning classifier to identify n-grams in the tweets of Congress members that help us to
differentiate between the two parties.
We visualize the most predictive n-grams identified by the classifiers for each election cycle in
Figure A.4. Overall, the classifiers identify words, hashtags, and Twitter handles that are intuitively
7
associated with a Republican slant for each election year. Among the most predictive term are
the hashtags “tcot” (Top Conservatives on Twitter) and “maga” (Make America great again) and
particularly in 2020 many references to Donald Trump’s Twitter account (“realdonaldtrump”).
B Additional Details on the SXSW Festival
Figure B.1: Screenshot Quote from Twitter Founder
Notes: This screenshot shows the full post of Twitter co-founder Evan Williams posted on Quora on January
4, 2011 describing the role of the SXSW festival in the platform’s rise to popularity (Quora, 2011).
8
Figure B.2: Number of SXSW Follower Joining each Month
Number of followers joining Twitter
800
600
400
200
0
06
06
06
em 6
cto 6
ov r 06
ec r 06
Fe y 07
07
07
0
y0
t0
r0
r0
ch
ril
ay
ne
ry
ch
s
be
be
be
be
l
ar
Ju
gu
ua
Ap
M
ar
Ju
ar
nu
em
em
Au
br
M
M
Ja
pt
O
Se
D
N
Notes: The figures shows the number of SXSW Follower joining in each month.
Figure B.3: Additional Evidence for the Impact of the SXSW Festival
(a) Long-term Effects of the 2007 SXSW on
Twitter Adoption (b) Connections to the SXSW festival
0.6
Diff-in-diff estimate of SXSW on Twitter usage
0.006 SXSW festival starts

Share of Twitter users with SXSW connection
0.4
0.004
0.002 0.2
0.000 0.0
2006q2 2009q1 2012q1 2015q1 2007q1 2009q1 2011q1 2013q1 2015q1
Notes: The figures provide evidence on the long-term impact of the SXSW festival Pon Twitter usage across the United
M arch2007
States. Panel (a) plots the β τ from the panel event study regression usersct = τ βτ SXSWc × 1(t =
P re
P
τ ) + τ δτ SXSWc × 1(t = τ ) + θc + γt + εct where usersct is the number of Twitter users per capita in county c
on quarter t, SXSWcM arch2007 is the logarithm of (one plus) the number of SXSW followers in county c who joined
Twitter in March 2007 and SXSWcP re is a similarly defined variable for followers who joined Twitter before March
2007. We standardize the variables to have a mean of zero and standard deviation of one. The whiskers represent 95%
confidence intervals based on standard errors clustered by state, where 2006q4 serves as excluded period. While the
confidence intervals for 2006q2 and 2006q3 cannot be seen, they include zero. Panel (b) plots the share of Twitter who
either follow SXSW or follow an user that follows SXSW.
9
Table B.1: Balancedness of SXSW Counties’ User Characteristics
First names (Corr. = 0.63) Terms used in bio (Corr. = 0.89)

Pre-period March 2007 Pre-period March 2007
michael michael http http
paul john founder com
mike chris com digital
chris jeff tech founder
eric matt product medium
justin brian co director
ryan david digital tech
kevin alex director music
jeff jason design social
david kevin social marketing
Notes: This table presents the ranking of the most common first names
and terms used in a Twitter user’s “bio” among users who follow “South
by Southwest” on Twitter, depending on whether they joined during
March 2007 or in the pre-period.
Table B.2: Are Twitter Users in Counties With SXSW Followers Different?
User first names Terms used in user bio

(Corr. = 0.97) (Corr. = 0.94)
Other counties SXSW counties Other counties SXSW counties
michael michael love co
chris david life love
john chris co life
david john http http
sarah alex http co http co
mike mike god music
emily matt ig lover
ryan sarah music ig
matt ryan university de
alex andrew like like
Notes: This table compares the individual characteristics of Twitter users
from counties with “South by Southwest” followers who joined in March
2007 (“SXSW counties”) to Twitter users from all other U.S. counties (“Other
counties”). We plot the ranking of the most common first names and terms
used in a Twitter user’s “bio”.
10
C Additional Robustness Checks
Table C.1: LASSO Variable Selection for 2016/2020 Election
Dep. var.: Republican vote share in...

Census Region FE State FE
2
Controls Controls Controls Controls2
(1) (2) (3) (4)
Panel A: 2SLS 2016 Election
Log(Twitter users) -0.030** -0.039*** -0.031*** -0.046***
(0.012) (0.013) (0.012) (0.013)
Observations 3,064 3,064 3,064 3,064
Nr. Controls 48 609 93 654
Nr. selected controls 30 70 50 87
Panel B: 2SLS 2020 Election
Log(Twitter users) -0.026* -0.035*** -0.029** -0.039***
(0.014) (0.013) (0.013) (0.013)
Observations 3,064 3,064 3,064 3,064
Nr. Controls 48 609 93 654
Nr. selected controls 31 73 51 91
Notes: This table presents county-level regressions where the dependent vari-
able is the Republican vote share in the 2016 or 2020 presidential election.
Log(Twitter users) is instrumented using the number of users who started
following SXSW in March 2007 (in logs with 1 added inside). Columns 1
includes census region fixed effects and allows all potential control variables
discussed in the text (48 controls) to be selected by the LASSO procedure.
Columns 2 includes census region fixed effects and allows for interactions
of all control variables with each other (609 potential controls). Column 3
includes state fixed effects and allows all control variables to be selected
(93 potential controls). Column 4 includes state fixed effects and allows all
interactions of control variables with each other to be selected (654 poten-
tial controls). In all regressions the included controls are selected using the
partialing-out LASSO procedure from Chernozhukov et al. (2015a,b). Stan-
dard errors in parentheses are clustered at the state level. *** p < 0.01, **
p < 0.05, * p < 0.1.
11
Table C.2: Twitter and the Republican Vote Share – Robustness
No Election Pre-period Pre-period No zero Spatial Per Capita

regression year control control SXSW user standard Twitter
weights weights polynomial deciles counties errors Usage
(1) (2) (3) (4) (5) (6) (7)
Panel A: Republican vote share in 2016
Log(Twitter users) -0.037*** -0.021** -0.019*** -0.020*** -0.030** -0.037***
(0.011) (0.008) (0.007) (0.007) (0.012) (0.011)
Log(Twitter users p.c.) -0.091**
(0.044)
Observations 3,064 3,064 3,064 3,064 165 3,064 3,064
Mean of DV 0.64 0.46 0.46 0.46 0.34 0.64 0.46
Robust F-stat. 72.94 94.50 125.40 114.91 23.46 54.13 24.50
Panel B: Republican vote share in 2020
Log(Twitter users) -0.036** -0.020** -0.018** -0.019** -0.033** -0.036***
(0.014) (0.010) (0.008) (0.009) (0.016) (0.012)
Log(Twitter users p.c.) -0.086*
(0.047)
Observations 3,064 3,064 3,064 3,064 165 3,064 3,064
Mean of DV 0.65 0.47 0.47 0.47 0.35 0.65 0.47
Robust F-stat. 72.94 88.47 125.40 114.91 23.46 54.13 24.50
Notes: This table presents 2SLS results estimated using equation (3). The dependent variable is the vote share of the
Republican party in the 2016 and 2020 presidential elections in panel A and B, respectively. Log(Twitter users) is
instrumented using the number of users who started following SXSW in March 2007 (in logs with 1 added inside). All
regressions except columns 1, 2, and 6 are weighted by turnout in the 2000 presidential election. Column 1 omits
regression weights. Column 2 weights by the turnout in the election year (2016 or 2020) instead of 2000. Columns
3 and 4 control for a fifth-order polynomial and deciles of SXSW followers who joined Twitter before the SXSW
2007 event, respectively. Column 5 drops all counties that had no SXSW followers joining Twitter in March 2007 or
in the period before. Column 6 uses spatial standard errors based on the method proposed in Colella et al. (2019),
implemented in Stata as acreg, using a 200 miles cutoff. Column 7 uses the number of Twitter users per capita (in logs
with 1 added inside). All regressions include the controls from columns 5 and 10 in Table 2. In columns 1 to 5, 7, and
8, standard errors in parentheses are clustered by state. *** p < 0.01, ** p < 0.05, * p < 0.1.
12
Figure C.1: Twitter and the Republican Vote Share, 1924-1996 (Reduced Form)
1924
1928
1932
1936
1940
1944
1948
1952
1956
1960
1964
1968
1972
1976
1980
1984
1988
1992
1996
-.04 -.02 0 .02 .04
Notes: This figure plots reduced form estimates β̂ 0 from county-level regressions as in equation (2). These
estimates reflect the correlation of Log(1 + SXSW followers, March 2007) with the Republican vote share
in presidential elections while controlling for Log(1 + SXSW followers, Pre). All regressions control for
population deciles and Census region fixed effects, and the full set of controls except 1996 Election controls
(same as columns 4 and 9 of Table 2). Regressions are weighted by turnout in the 2000 presidential election.
Whiskers represent 95% confidence intervals based on standard errors clustered by state.
13
Table C.3: Twitter and Changes in Republican Vote Share, 2004-2020
Dep. var.: ∆Republican vote share between...

2000-04 2000-08 2000-12 2000-16 2000-20
(1) (2) (3) (4) (5)
Log(SXSW followers, March 2007) -0.002 -0.003 -0.001 -0.009** -0.008**
(0.002) (0.003) (0.003) (0.004) (0.004)
Log(SXSW followers, Pre) -0.000 -0.001 -0.002 0.000 -0.002
(0.003) (0.005) (0.003) (0.004) (0.005)
Panel B: 2SLS
Log(Twitter users) -0.004 -0.006 -0.002 -0.017** -0.015**
(0.004) (0.006) (0.006) (0.007) (0.007)
Log(SXSW followers, Pre) 0.000 -0.001 -0.002 0.001 -0.001
(0.003) (0.005) (0.003) (0.005) (0.005)
Observations 3,064 3,064 3,064 3,064 3,064
Mean of DV 0.03 -0.02 -0.01 -0.02 -0.01
Robust F-stat. 121.18 121.18 121.18 121.18 121.18
Notes: This table presents county-level regressions where the dependent variable is the change in
the vote share of the Republican party between 2000 and the indicated year. Log(SXSW followers,
March 2007) is the number of Twitter users (in logs, with 1 added inside) who joined in March
2007 and follow South by Southwest (SXSW). SXSW followers, Pre is the number of SXSW fol-
lowers who registered at some point in 2006, defined similarly. Twitter users are the number of
users in 2014-2015. All regressions control for population deciles, Census region fixed effects, and
the full set of controls (as in columns 5 and 10 of Table 2). The first-stage regressions for 2SLS
results (Panel B) are presented in Table 1, with the F-stat for the excluded instrument in the bottom
row. On Panel (a), the coefficient on Log(SXSW followers, March 2007) for 2004, 2008, and 2012
are jointly statistically insignificant (p-value=0.398). Further, the average effect in 2016 and 2020
is statistically distinct from the average effect in 2004, 2008, and 2012 (p-value=0.001). Observa-
tions are weighted by turnout in the 2000 presidential election. Standard errors in parentheses are
14
Table C.4: Time-Varying Twitter Usage and Changes in Vote Shares
Dep. var.: ∆Republican vote share between...

2000-04 2000-08 2000-12 2000-16 2000-20
(1) (2) (3) (4) (5)
Panel A: First stage
Log(SXSW followers, March 2007) 0.644*** 0.644*** 0.546*** 0.523*** 0.523***
(0.055) (0.055) (0.047) (0.048) (0.048)
Panel B: Reduced form
Log(SXSW followers, March 2007) -0.002 -0.003 -0.001 -0.009** -0.008**
(0.002) (0.003) (0.003) (0.004) (0.004)
Log(SXSW followers, Pre) -0.000 -0.001 -0.002 0.000 -0.002
(0.003) (0.005) (0.003) (0.004) (0.005)
Panel C: 2SLS
Log(Twitter users) -0.003 -0.005 -0.002 -0.017** -0.015**
(0.003) (0.005) (0.005) (0.007) (0.007)
Log(SXSW followers, Pre) 0.000 -0.001 -0.002 0.001 -0.001
(0.003) (0.005) (0.003) (0.005) (0.005)
Observations 3064 3064 3064 3064 3064
Mean of DV -0.024 -0.008 -0.018 -0.009
Robust F-stat. 135.88 135.88 134.88 121.18 121.18
Twitter usage measured in 2008 2008 2012 2016 2016
Notes: This table presents county-level regressions where the dependent variable is the change in the
vote share of the Republican party between 2000 and the indicated year (except for Panel B, where the
dependent variable is Twitter users. Log(SXSW followers, March 2007) is the number of Twitter users
(in logs, with 1 added inside) who joined in March 2007 and follow South by Southwest (SXSW).
SXSW followers, Pre is the number of SXSW followers who registered at some point in 2006, defined
similarly. Differently from other tables, Twitter users varies over time, as opposed to being fixed to
2014-2015. All regressions control for population deciles, Census region fixed effects, and the full
set of controls (as in columns 5 and 10 of Table 2). Observations are weighted by turnout in the 2000
presidential election. Standard errors in parentheses are clustered by state. *** p < 0.01, ** p < 0.05,
* p < 0.1.
15
Table C.5: Twitter and the Ross Perot Vote
Dep. var.: Vote share Ross Perot in...

1992 1996
(1) (2) (3) (4)
Log(SXSW followers, March 2007) 0.003 0.003 0.000 -0.000
(0.003) (0.003) (0.002) (0.002)
Panel B: 2SLS
Log(Twitter users) 0.006 0.007 0.000 -0.001
(0.006) (0.006) (0.003) (0.003)
Population deciles Yes Yes Yes Yes
Census region FE Yes Yes Yes Yes
Geographical controls Yes Yes Yes Yes
Demographic controls Yes Yes Yes Yes
Socioeconomic controls Yes Yes Yes Yes
China shock controls Yes Yes Yes Yes
1996 election control Yes Yes
Observations 3,064 3,064 3,064 3,064
Mean of DV 0.20 0.20 0.10 0.10
Robust F-stat. 118.21 121.18 118.21 121.18
Notes: This table presents county-level regressions where the dependent variable
is the third party vote share in the 1992 or 1996 presidential election. Log(SXSW
followers, March 2007) is the number of Twitter users (in logs, with 1 added in-
side) who joined in March 2007 and follow South by Southwest (SXSW). SXSW
followers, Pre is the number of SXSW followers who registered at some point in
2006, defined similarly. Twitter users are the number of users in 2014-2015. The
first-stage regressions for 2SLS results (Panel B) are presented in Table 1, with the
F-stat for the excluded instrument in the bottom row. Observations are weighted
by turnout in the 2000 presidential election. Standard errors in parentheses are
16
Table C.6: Twitter and Vote Shares in Democratic Primaries
Dep. var.: Vote share in Democratic Primary of...

Clinton Sanders Warren Biden Sanders Buttigieg Bloomberg Klobuchar
2016 2016 2020 2020 2020 2020 2020 2020
(1) (2) (3) (4) (5) (6) (7) (8)
Log(SXSW followers, March 2007) 0.005 -0.004 0.002 -0.009 0.017*** -0.003 0.001 -0.003
(0.010) (0.010) (0.006) (0.014) (0.006) (0.002) (0.004) (0.002)
Panel B: 2SLS
Log(Twitter users) 0.008 -0.006 0.003 -0.014 0.025** -0.004 0.001 -0.004
(0.014) (0.014) (0.009) (0.021) (0.010) (0.003) (0.006) (0.002)
Observations 2,656 2,656 2,769 2,769 2,769 2,769 2,769 2,769
Mean of DV 0.55 0.43 0.06 0.56 0.24 0.02 0.06 0.01
Robust F-stat. 67.94 67.94 73.68 73.68 73.68 73.68 73.68 73.68
Notes: This table presents county-level regressions where the dependent variable is the vote share of the indicated candidate in the Demo-
cratic party primaries in 2016 or 2020. Log(SXSW followers, March 2007) is the number of Twitter users (in logs, with 1 added inside) who
joined in March 2007 and follow South by Southwest (SXSW). SXSW followers, Pre is the number of SXSW followers who registered at
some point in 2006, defined similarly. Twitter users are the number of users in 2014-2015. All regressions control for population deciles,
Census region fixed effects, and the full set of controls (as in columns 5 and 10 of Table 2). The first-stage regressions for 2SLS results
(Panel B) are analogous to the one presented in Table 1, except for the different sample of counties for which primary results are available.
The F-stat for the excluded instrument is provided in the bottom row. Observations are weighted by turnout in the 2000 presidential election.
Standard errors in parentheses are clustered by state. *** p < 0.01, ** p < 0.05, * p < 0.1.
Table C.7: Twitter and Vote Decisions in the 2016 CCES – Robustness
Dep. var.: Voted for Trump in 2016

(1) (2) (3) (4) (5)
Intended Intended
Verified Vote Trump other
Baseline vote intention vote vote
Log(Twitter users) -0.135*** -0.154*** -0.133** -0.231*** -0.064*

(0.045) (0.054) (0.052) (0.082) (0.034)
Observations 146,579 56,375 46,418 14,723 24,354
Mean of DV 0.492 0.497 0.455 0.991 0.137
Marginal effect [-0.049] [-0.055] [-0.048] [-0.005] [-0.013]
Notes: This table presents results estimated using IV probit models, as in equation (4).
The dependent variable is a dummy for individuals in the CCES who voted for Trump
in 2016. Log(Twitter users) is instrumented using the number of SXSW followers that
joined Twitter in March 2007. All regressions control for the (log) number of SXSW
followers that joined Twitter at some point in 2006, family income, gender, education
levels, marital status, news interest, and age, as well as county-level population deciles
and Census region fixed effects. Regressions are weighted by survey weights. Standard
errors in parentheses are clustered by state. *** p < 0.01, ** p < 0.05, * p < 0.10.
17
D Further Results
Table D.1: Additional Outcomes

Switching prob. ∆Campaign don., 2000-16 Trump approval, 2017
Obama to Trump Democrats Republicans Democrats Republicans
(1) (2) (3) (4) (5)
Log(Twitter users) -0.138*** 0.866*** 0.168 -0.011** -0.037***
(0.048) (0.185) (0.235) (0.005) (0.009)
Observations 3,065 2,250 2,446 2,727 2,920
Mean of DV 0.105 1.943 1.096 0.066 0.850
Robust F-stat. 74.65 62.57 64.34 60.20 66.44
Notes: This table presents results from county-level regressions of equation (3). Column 1 shows results
from an IV probit regression where the dependent variable is a dummy equal to 1 for the 217 counties for
which both Obama and Trump gained the majority of votes in 2008 and 2016, respectively. In columns 2 and
3, the dependent variable is the difference in the natural logarithm of campaign donations to the Democratic
and Republican party, respectively, between 2000 and 2016. In columns 4 and 5, the dependent variable
is the share of respondents in the Gallup Daily Poll approving of Trump in 2017. Log(Twitter users) is
instrumented using the number of users who started following SXSW in March 2007. All regressions control
for population deciles and Census region fixed effects and geographical controls. Regressions are weighted
by turnout in the 2000 presidential election. Standard errors in parentheses are clustered by state. *** p <
0.01, ** p < 0.05, * p < 0.1.
Table D.2: Twitter and Changes in Voter Turnout, 2004-2020

∆Votes cast/voting age pop.
2000-04 2000-08 2000-12 2000-16 2000-20
(1) (2) (3) (4) (5)
Log(SXSW followers, March 2007) -0.000 -0.000 -0.001 0.001 0.007**
(0.003) (0.003) (0.004) (0.004) (0.003)
Log(SXSW followers, Pre) -0.000 0.000 -0.002 -0.001 -0.005
(0.006) (0.004) (0.005) (0.005) (0.005)
Panel B: 2SLS
Log(Twitter users) -0.000 -0.000 -0.001 0.002 0.014**
(0.006) (0.006) (0.008) (0.008) (0.006)
Log(SXSW followers, Pre) -0.000 0.000 -0.002 -0.001 -0.006
(0.006) (0.005) (0.005) (0.005) (0.005)
Observations 3,063 3,063 3,063 3,063 3,063
Mean of DV 0.088 0.079 0.053 0.057 0.126
Robust F-stat. 121.23 121.23 121.23 121.23 121.23
Notes: This table presents county-level regressions where the dependent variable is the change
in the voter turnout (as a share of voting age population) between 2000 and the indicated year.
Log(SXSW followers, March 2007) is the number of Twitter users (in logs, with 1 added inside)
who joined in March 2007 and follow South by Southwest (SXSW). SXSW followers, Pre is the
number of SXSW followers who registered at some point in 2006, defined similarly. Twitter users
are the number of users in 2014-2015. All regressions control for population deciles, Census
region fixed effects, and the full set of controls (as in columns 5 and 10 of Table 2). The first-stage
regressions for 2SLS results (Panel B) are presented in Table 1, with the F-stat for the excluded
instrument in the bottom row. Observations are weighted by turnout in the 2000 presidential
election. Standard errors in parentheses are clustered by state. *** p < 0.01, ** p < 0.05, *
p < 0.1.
18
Table D.3: Twitter and Congressional Elections – Reduced Form Estimates
Panel A: House elections

∆Republican vote share in House election between...
2000-02 2000-04 2000-06 2000-08 2000-10 2000-12 2000-14 2000-16 2000-18 2000-20
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Log(SXSW followers, March 2007) 0.010 0.002 0.017* 0.018* 0.015 0.017 0.024* 0.009 0.019* 0.017
(0.011) (0.010) (0.010) (0.009) (0.011) (0.011) (0.012) (0.012) (0.011) (0.011)
Observations 3,046 3,007 3,035 2,993 3,027 3,047 3,042 3,044 3,047 3,046
Mean of DV 0.02 0.01 -0.04 -0.06 0.03 -0.01 0.02 0.00 -0.04 -0.01
Panel B: Senate elections
19
∆Republican vote share in Senate election between...
1996-02 1998-04 2000-06 1996-08 1998-10 2000-12 1996-14 1998-16 2000-18 1996-20
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Log(SXSW followers, March 2007) 0.009 0.007 0.003 0.005 -0.008 0.010 -0.013 -0.009 -0.004 -0.008
(0.011) (0.014) (0.010) (0.010) (0.012) (0.014) (0.014) (0.016) (0.010) (0.009)
Observations 2,247 2,049 1,832 2,247 2,049 1,832 2,247 2,049 1,832 2,247
Mean of DV 0.01 -0.02 -0.06 -0.05 0.02 -0.06 0.02 -0.07 -0.10 -0.00
Notes: This table presents reduced form results, as in equation (2). For House elections in Panel A, the dependent variable is the change in the Republican
vote share since 2000. For Senate elections in Panel B, the dependent variable is the change in the Republican vote share from six, twelve, or eighteen
years ago (to accommodate senators’ 6-year terms). The main independent variable is the number of users who started following SXSW in March 2007 (in
logs with 1 added inside). All regressions control for the (log) number of SXSW followers that joined Twitter at some point in 2006, population deciles and
Census region fixed effects and the full set of controls (as in columns 5 and 10 of Table 2). Regressions are weighted by turnout in the 2000 presidential
election. Standard errors in parentheses are clustered by state. *** p < 0.01, ** p < 0.05, * p < 0.1.
Table D.4: Twitter and the Republican Vote Share in Swing and Safe Counties
Swing Republican Democratic Safe

counties counties counties counties
(1) (2) (3) (4)
Panel A: ∆Republican vote share 2000-2016
Log(Twitter users) -0.073*** -0.006 -0.008 -0.005
(0.024) (0.008) (0.008) (0.006)
Log(SXSW followers, Pre) 0.013 -0.001 0.000 -0.002
(0.015) (0.014) (0.008) (0.005)
Observations 716 1,990 358 2,348
Mean of DV -0.033 0.021 -0.040 -0.012
Robust F-stat. 14.70 11.97 105.57 99.87
Panel B: ∆Republican vote share 2000-2020
Log(Twitter users) -0.066*** -0.017** -0.013 -0.007
(0.019) (0.007) (0.009) (0.006)
Log(SXSW followers, Pre) 0.006 -0.005 -0.001 -0.003
(0.012) (0.013) (0.009) (0.006)
Observations 716 1,990 358 2,348
Mean of DV -0.027 0.026 -0.026 -0.002
Robust F-stat. 14.70 11.97 105.57 99.87
Notes: This table presents results estimated using 2SLS, as in equation (3). The depen-
dent variable is the change in the vote share of the Republican party between the 2000
and 2016/2020 presidential elections in Panels A and B, respectively. Swing counties
are those that were not consistently won by either Republicans or Democrats between
2000 and 2012; Republican and Democratic counties are those who voted consis-
tently. Safe counties are the counties from columns (2) and (3) combined. Log(Twitter
users) is instrumented using the number of users who started following SXSW in
March 2007. SXSW followers, Pre is the number of SXSW followers who registered
at some point in 2006. All regressions control for population deciles and Census
region fixed effects and the full set of controls (as in columns 5 and 10 of Table 2).
Regressions are weighted by turnout in the 2000 presidential election. Standard errors
in parentheses are clustered by state. *** p < 0.01, ** p < 0.05, * p < 0.1.
20
Figure D.1: Twitter’s Partisan Slant (Tweet Measure)
(a) Tweets about Republican Presidential Can- (b) Tweets about Democratic Presidential
didates Candidates
34%
28%
250 37% 250
Number of Tweets (in millions)

200 200
150 150
24% 31%
100 27% 100 27%
49% 42%
22%
50 35% 50 38% 39%
14% 36% 39%
51% 36%
0 0
Notes: These figures present the number of tweets (as opposed to the number of “likes” of such tweets in Figure 8)
that contain the last name of the candidates in the 2012, 2016 and 2020 presidential elections, depending on
whether the tweet was classified as having a Republican (instead of Democratic) slant. We classify the slant of
a tweet based on the Twitter network of the user who sent the tweet. If the user follows more Democratic than
Republican Congress members, they will be classified as a Democrat, and vice versa. Users who follow an equal
number of Democrats and Republican or no Congress members are classified as neutral.
21
Figure D.2: Twitter’s Partisan Slant (Text-Based Classifier)
(a) Likes for Tweets about Republican Presi- (b) Likes for Tweets about Democratic Presi-
dential Candidates dential Candidates
11%
600
38% 600
51%

500 500
400 400
300 300 10%

47%
200 200 43%
17%
39%
13%
100 9% 44% 100 12%
65%
12% 36%
22%
0
79% 0
52%
(c) Tweets about Republican Presidential Can- (d) Tweets about Democratic Presidential
didates Candidates
12%
53%
250 35% 250
200 200
150 150
15% 7%
100 30% 100 47%
55% 46%
11%
50 9% 50 14% 51%
12% 40% 38%
78% 46%
0 0
Notes: These figures present the number of “likes” received by tweets, or the number of tweets, that contain the
last name of the candidates in the 2012, 2016 and 2020 presidential elections, depending on whether the tweet was
classified as having a Republican (instead of Democratic) slant. We classify the slant of a tweet based on similarity
in the language to that of a congressional Republican or Democrat, using a L2 regularized logistic regression
classifier using the tweets sent by Congress members. Optimal normalization strength is chosen using 10-fold
cross-validation. Tweets with a predicted class probability below 60% are coded as neutral. See Appendix A.1.
for details.
22
Figure D.3: Specification Curve
(a) 2016 Presidential Election Results
0
-.02
Coefficient Twitter Usage (IV)
-.04
-.06
-.08
-.1
Census_Region
Socioeconomic
Demographic
China_Shock
Geographic
Population
Election
Ethnic
TV
-.12
Population
Census_Region
Geographic
pad
Demographic
Ethnic
Socioeconomic
TV
China_Shock
Census_Region
Socioeconomic
Demographic
China_Shock
Geographic
Population
Election
Ethnic
TV
(b) 2020 Presidential Election Results

0
-.02
Coefficient Twitter Usage (IV)
-.04
-.06
-.08
-.1
Census_Region
Socioeconomic
Demographic
China_Shock
Geographic
Population
Election
Ethnic
TV
Population
Census_Region
Geographic
pad
Demographic
Ethnic
Socioeconomic
TV
China_Shock
Census_Region
Socioeconomic
Demographic
China_Shock
Geographic
Population
Election
Ethnic
TV
Notes: These figures plot the 2SLS estimates and 95% confidence intervals from a regression of the Repub-
lican vote share in 2016 on Log(T witter users), instrumented with SXSW M arch 2007 . All regressions
include population deciles, census region fixed effects, and SXSW P re . The combination of the other
included control variables is shown at the bottom; filled circles mean a set of controls was included. The
baseline specification with all controls is marked by the vertical line.
23
E Additional Details on the Extrapolation for the Average Treat-
ment Effect
Andrews and Oster (2019) show how selection into participating in an experiment can be used to
make extrapolations regarding the external validity of an experiment. When a set of covariates X is
observed for both the “experimental sample” and “population,” Andrews and Oster (2019) provide a
procedure that uses effect heterogeneity based on X estimated within the experimental sample to
extrapolate to the average treatment effect for the “population.”
We build on their procedure and argue that we can similarly use heterogeneity in the treatment
effect within the counties that “identify” our results to extrapolate the treatment effect to all other
counties in the US. Column (5) of Table C.2 show that we obtain similar estimates to our baseline
when we only compare counties with SXSW followers that joined Twitter in March 2007 to counties
with followers that joined in the pre-period, while excluding those counties in neither group. We
can use this subsample of counties as the “experimental sample,” and extrapolate effects to the
“population” of all other counties.
Since Andrews and Oster (2019) approach is designed for a binary treatment, we adjust our
regression framework by defining a treatment indicator variable equal to 1 for counties with SXSW
followers who joined in March 2007 and 0 for the counties with followers that joined in the pre-
period. We estimate the treatment effect for the subsample of counties that do not have zero SXSW
followers in both periods using the regression specification yc = α+β · 1[SXSWcM arch2007 > 0]+c .
The resulting treatment effect estimate is −0.075, which is similar Table 2 Panel B column (1). We
then perform a linear prediction of this treatment effect based all observable variables in Table A.4
within this subsample. The resulting predicted treatment effect is −0.085. Last, we extrapolate the
treatment effect for the rest of US counties. Based on the variation in observable characteristics we
would predict an ATE of −0.184 for the US overall.
Note that this extrapolation is based on adjusting our reduced-form estimates to use a binary
indicator variable for treatment thus the coefficients are not directly comparable to our baseline
estimates. The approach further assumes quasi-random treatment assignment with in the counties
with SXSW variation. Taken this into account, the extrapolation should therefore be viewed as
suggestive, but confirming the notion that the effect for all US counties would be larger.
24

Social Media and Elections

Uploaded by

Copyright:

Available Formats

Social Media and Elections

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Social Media and Elections

Uploaded by

Copyright:

Available Formats

The Effect of Social Media on Elections:

Evidence from the United States*

Thomas Fujiwara† Karsten Müller‡ Carlo Schwarz§

2 Background: Social Media and Politics

Individual-Level Voting Decisions. The Cooperative Congressional Election Study (CCES) is a

Figure 2: Twitter Usage and Identifying Variation

Additional County Characteristics We collect county-level demographic control variables from

Change in Twitter activity (in SD)

T witter usersc = α + β · SXSWcM arch2007 + γ · SXSWcP re + Xc δ + ξc (1)

yc = α0 + β 0 · SXSWcM arch2007 + γ 0 · SXSWcP re + Xc δ 0 + ζc , (2)

6.1 Main results

Dep. var.: Log(Twitter users)

Log(Twitter users), residualized

-1.00 -0.50 0.00 0.50 1.00

Log(1 + SXSW followers, March 2007), residualized

Republican vote share, 2016 (residualized)

Log(1 + SXSW followers, March 2007), residualized Log(Twitter users), residualized

Republican vote share, 2020 (residualized)

Log(1 + SXSW followers, March 2007), residualized Log(Twitter users), residualized

Table 3: Placebo Tests for Other Festivals

Twitter Followers of Festival joining in Festival Month

Table 4: Twitter and the Republican Vote Share, 2000-2020

Dep. var.: Republican vote share in...

Point estimate on Log(SXSW followers, March 2007)

6.2 Effects Are Concentrated Among Moderate Voters

(b) Senate Elections

Table 5: Twitter and Individuals’ Vote Decisions in 2016/2020

Dep. var.: Voted for Trump in ...

Log(Twitter users) -0.135*** 0.018 -0.079 -0.209*** -0.045 -0.022

Observations 146,579 44,241 30,745 13,625 27,865 28,559

Log(Twitter users) -0.130*** -0.069 -0.002 -0.186*** 0.000 -0.080

Observations 43,617 13,108 9,670 4,430 7,388 8,395

Table 6: Twitter and Vote Shares in the 2016 Republican Primaries

Dep. var.: Vote share in Republican Primary of...

Dep. var.: Approved of candidate during primaries

Log(Twitter users) -0.108*** -0.086** -0.051 0.018 0.031 0.148***

Log(Twitter users) -0.065** -0.006 -0.015 0.050 0.059 0.154***

Log(Twitter users) -0.052 -0.116** -0.036 0.076 0.004 0.081**

Number of Likes (in millions)

300 300 18%

• Appendix A provides additional details on the data.

• Appendix B discusses the SXSW instrument.

• Appendix C shows additional robustness checks.

• Appendix D provides further results.

• Appendix E discusses extrapolation for the average treatment effect.

A Additional Details on Data

Figure A.1: Validation of Twitter usage measure

2007q1 2009q1 2011q1 2013q1 2015q1 2017q1 2019q1 Log(Twitter users)

Figure A.3: News Reports About Social Media

3000 2016 Election

2000 2012 Election

Mean Std. Dev. Min. Median Max. N

Mean Std. Dev. Min. Median Max. N

Table A.3: Summary Statistics (Gallup Individual-Level)

Mean Std. Dev. Min. Median Max. N

March 2007 March 2007 Pre Difference

(c) 2020 Election

B Additional Details on the SXSW Festival

Figure B.1: Screenshot Quote from Twitter Founder

Log(Twitter users) -0.135* 0.018 -0.079 -0.209* -0.045 -0.022

Log(Twitter users) -0.130* -0.069 -0.002 -0.186* 0.000 -0.080

Log(Twitter users) -0.108* -0.086 -0.051 0.018 0.031 0.148***

Log(Twitter users) -0.065 -0.006 -0.015 0.050 0.059 0.154*

Log(Twitter users) -0.052 -0.116 -0.036 0.076 0.004 0.081

Log(Twitter users) -0.135* -0.154* -0.133 -0.231* -0.064*