ArticlePDF Available

A 61-Million-Person Experiment in Social Influence and Political Mobilization



Human behaviour is thought to spread through face-to-face social networks, but it is difficult to identify social influence effects in observational studies, and it is unknown whether online social networks operate in the same way. Here we report results from a randomized controlled trial of political mobilization messages delivered to 61 million Facebook users during the 2010 US congressional elections. The results show that the messages directly influenced political self-expression, information seeking and real-world voting behaviour of millions of people. Furthermore, the messages not only influenced the users who received them but also the users' friends, and friends of friends. The effect of social transmission on real-world voting was greater than the direct effect of the messages themselves, and nearly all the transmission occurred between 'close friends' who were more likely to have a face-to-face relationship. These results suggest that strong ties are instrumental for spreading both online and real-world behaviour in human social networks.
LETTER doi:10.1038/nature11421
A 61-million-person experiment in social influence
and political mobilization
Robert M. Bond
, Christopher J. Fariss
, Jason J. Jones
, Cameron Marlow
, Jaime E. Settle
& James H. Fowler
Human behaviour is thought to spread through face-to-face social
networks, but it is difficult to identify social influence effects in
observational studies
, and it is unknown whether online social
networks operate in the same way
. Here we report results from
a randomized controlled trial of political mobilization messages
delivered to 61 million Facebook users during the 2010 US con-
gressional elections. The results show that the messages directly
influenced political self-expression, information seeking and real-
world voting behaviour of millions of people. Furthermore, the
messages not only influenced the users who received them but also
the users’ friends, and friends of friends. The effect of social trans-
mission on real-world voting was greater than the direct effect of
the messages themselves, and nearly all the transmission occurred
between ‘close friends’ who were more likely to have a face-to-face
relationship. These results suggest that strong ties are instrumental
for spreading both online and real-world behaviour in human
social networks.
Recent experimental studies
have attempted to measure the
causal effect of social influence online. At the same time, there is
increasing interest in the ability to use online social networks to study
and influence real-world behaviour
. However, online social networks
are also made up of many ‘weak-tie’ relationships
that may not
facilitate social influence
, and some studies suggest that online
communication may not be an effective medium for influence
open question is whether online networks, which harness social
information from face-to-face networks, can be used effectively to
increase the likelihood of behaviour change and social contagion.
One behaviour that has been proposed to spread through networks
is the act of voting in national elections. Voter turnout is significantly
correlated among friends, family members and co-workers in obser-
vational studies
. Voter mobilization efforts are effective at increas-
ing turnout
, particularly those conducted face-to-face and those that
appeal to social pressure
and social identity
. There is also evidence
from one face-to-face field experiment that voting is ‘contagious’, in
the sense that mobilization can spread from person to person within
two-person households
. Although anecdotal accounts suggest that
online mobilization has made a big difference in recent elections
meta-analysis of email experiments suggests that online appeals to vote
are ineffective
Voter mobilization experiments
have shown that most methods
of contacting potential voters have small effects (if any) on turnout
rates, ranging from 1% to 10%. However, the ability to reach large
populations online means that even small effects could yield behaviour
changes for millions of people. Furthermore, as many elections are
competitive, these changes could affect electoral outcomes. For
example, in the 2000 US presidential election, George Bush beat Al
Gore in Florida by 537 votes (less than 0.01% of votes cast in Florida).
Had Gore won Florida, he would have won the election.
To test the hypothesis that political behaviour can spread through
an online social network, we conducted a randomized controlled trial
with all users of at least 18 years of age in the United States who
accessed the Facebook website on 2 November 2010, the day of the
US congressional elections. Users were randomly assigned to a ‘social
message’ group, an ‘informational message’ group or a control group.
The social message group (n560,055,176) was shown a statement at
the top of their ‘News Feed’. This message encouraged the user to vote,
provided a link to find local polling places, showed a clickable button
reading ‘I Voted’, showed a counter indicating how many other
Facebook users had previously reported voting, and displayed up to
six small randomly selected ‘profile pictures’ of the user’s Facebook
friends who had already clicked the I Voted button (Fig. 1). The
informational message group (n5611,044) was shown the message,
poll information, counter and button, but they were not shown any
faces of friends. The control group (n5613,096) did not receive any
message at the top of their News Feed.
The design of the experiment allowed us to assess the impact that the
treatments had on three user actions; clicking the I Voted button,
clicking the polling-place link and voting in the election. Clicking
the I Voted button is similar to traditional measures of self-reported
voting, but here users reported their vote to their social community
rather than to a researcher. We therefore use this action to measure
political self-expression, as it is likely to be affected by the extent to
which a user desires to be seen as a voter by others. In contrast, social
desirability should not affect other user actions in the same way.
Clicking the polling-place link took users to a separate website that
helped them to find a polling location, and this action was not reported
to the user’s social community. We therefore use this action to measure
a user’s desire to seek information about the election. Finally, we used
a group-level process to study the validated voting behaviour of
6.3 million users matched to publicly available voter records (see
Supplementary Information).
We first analyse direct effects. We cannot compare the treatment
groups with the control group to assess the effect of the treatment on
self-expression and information seeking, because the control group did
not have the option to click an I Voted button or click on a polling-
place link. However, we can compare the proportion of users between
the two treatment groups to estimate the causal effect of seeing the
faces of friends who have identified themselves as voters (Fig. 1). Users
who received the social message were 2.08% (s.e.m., 0.05%; t-test,
P,0.01) more likely to click on the I Voted button than those who
received the informational message (20.04% in the social message
group versus 17.96% in the informational message group). Users
who received the social message were also 0.26% (s.e.m., 0.02%;
P,0.01) more likely to click the polling-place information link than
users who received the informational message (Fig. 1).
Although acts of political self-expression and information seeking
are important in their own right, they do not necessarily guarantee that
a particular user will actually vote. As such, we also measured the effect
that the experimental treatment had on validated voting, through
examination of public voting records. The results show that users
Political Science Department, University of California, San Diego, La Jolla, California 92093, USA.
Psychology Department, University of California, San Diego, La Jolla, California 92093, USA.
Science, Facebook, Inc., Menlo Park, California 94025, USA.
Medical Genetics Division, University of California, San Diego, La Jolla, California 92093, USA.
13 SEPTEMBER 2012 | VOL 489 | NATURE | 295
Macmillan Publishers Limited. All rights reserved
who received the social message were 0.39% (s.e.m., 0.17%; t-test,
P50.02) more likely to vote than users who received no message at
all. Similarly, the difference in voting between those who received the
social message and those who received the informational messagewas
0.39% (s.e.m., 0.17%; t-test, P50.02), suggesting that seeing faces of
friends significantly contributed to the overall effect of the message on
real-world voting. In fact, turnout among those who received the
informational message was identical to turnout among those in the
control group (treatment effect 0.00%, s.e.m., 0.28%; P50.98), which
raises doubts about the effectiveness of information-only appeals to
vote in this context.
These results show that online political mobilization can have a
direct effect on political self-expression, information seeking and
real-world voting behaviour, and that messages including cues from
an individual’s social network are more effective than information-
only appeals. But what about indirect effects that spread from person
to person in the social network? Users in our sample had on average
149 Facebook friends, with whom they share social information,
although many of these relationships constitute ‘weak ties’. Past
research indicates that close friends have a stronger behavioural effect
on each other than do acquaintancesor strangers
. We therefore
expected mobilization to spread more effectively online through
‘strong ties’.
To distinguish users who are likely to have close relationships, we
used the degree to which Facebook friends interacted with each other on
the site (see Supplementary Information for more detail). Higher levels
of interaction indicate that friends are more likely to be physically
proximate and suggest a higher level of commitment to the friendship,
more positive affect between the friends, and a desire for the friendship
to be socially recognized
. We counted the number of interactions
between each pair of friends and categorized them by decile, ranking
them from the lowest to highest percentage of interactions. A validation
study (see SupplementaryInformation)shows that friends in the highest
decile are those most likely to be close friends in real life (Fig. 2a).
We then used these categories to estimate the effect of the mobil-
ization message on a user’s friends. Random assignment means that
any relationship between the message a user receives and a friend’s
behaviour is not due to shared attributes, as these attributes are not
correlated with the treatment (see Supplementary Information). To
measure a per-friend treatment effect, we compared behaviour in the
friends connected to a user who received the social message to beha-
viour in the friends connected to a user in the control group. To
account for dependencies in the network, we simulate the null distri-
bution using a network permutation method (see the Supplementary
Information). Monte Carlo simulations suggest that this method
minimizes the risk of false positives and recovers true causal effects
without bias (see Supplementary Information).
Figure 2 shows that the observed per-friend treatment effects increase
as tie-strength increases. All of the observed treatment effectsfall outside
the null distribution for expressed vote (Fig. 2b), suggesting that they are
significantly different from chance outcomes. For validated vote
(Fig. 2c), the observed treatment effect is near zero for weak ties, but
it spikes upwards and falls outside the null distribution for the top two
deciles. This suggests that strong ties are important for the spread of
real-world voting behaviour. Finally, the treatment effect for polling
place search gradually increases (Fig. 2d), with several of the effects
falling outside the 95% confidence interval of the null distribution.
To simplify the analysis and reporting of results, we arbitrarily
define ‘close friends’ as people who were in the eightieth percentile
or higher (decile 9) of frequency of interaction among all friendships in
the sample (see the Supplementary Information). ‘Friends’ are all other
Facebook friends who had less interaction. A total of 60,491,898 (98%)
users in our sample had at least 1 close friend, with the average user
having about 10 close friends (compared with an average of 139 friends
who were not close).
The results suggest that users were about 0.011% (95% confidence
interval (CI) of null distribution 20.009% to 0.010%) more likely to
engage in an act of political self-expression by clicking on the I Voted
button than they would have been had their friend seen no message.
Similarly, for each close friend who received the social message, an
individual was on average 0.099% (null 95% CI –0.042% to 0.048%)
more likely to express voting.
We also found an effect in the validated vote sample. For each close
friend who received thesocial message, a user was 0.224% (null 95% CI
–0.181% to 0.174%) more likely to vote than they would have been had
their close friend received no message. Similarly, for information-
seeking behaviour we found that for each close friend who received
the social message, a user was 0.012% (null 95% CI –0.012% to 0.012%)
more likely to click the link to find their polling place than they would
have been had their close friends received no message. In both cases
there was no evidence that other friends had an effect (see
Supplementary Information). Thus, ordinary Facebook friends may
affect online expressive behaviour, but they do not seem to affect
private or real-world political behaviours. In contrast, close friends
seem to have influenced all three.
The magnitude of these contagion effects are small per friend, but it
is important to remember that they resultfrom a single message, and in
many cases it was not possible to change the target’s behaviour. For
example, users may have already voted by absentee ballot before
Election Day, or they may have logged in to Facebook too late to vote
or to influence other users’ voting behaviour. In other words, all effects
measured here are intent-to-treat effects rather than treatment-on-
treated effects, which would be greater if we had better information
about who was eligible to receive the treatment.
Informational message
Social message
friends have voted.
Today is Election Day What’s this?
People on Facebook Voted
Find your polling place on the U.S.
Politics Page and click the "I Voted"
button to tell your friends you voted.
VOTE l Voted
10 155376
Today is Election Day What’s this?
People on Facebook Voted
Find your polling place on the U.S.
Politics Page and click the "I Voted"
button to tell your friends you voted.
VOTE l Voted
10 155376
Direct effect of treatment
on own behaviour (%)
Search for
Jaime Settle, Jason Jones, and 18 other
Figure 1
The experiment and direct effects. a,b, Examples of the informational message and social message Facebook treatments (a) and their direct effect on
voting behaviour (b). Vertical lines indicate s.e.m. (they are too small to be seen for the first two bars).
296 | NATURE | VOL 489 | 13 SEPTEMBER 2012
Macmillan Publishers Limited. All rights reserved
Moreover, the scale of the number of users, their friendship
connections and the potential voters in a given election is very large.
We estimated the per-user effect (the per-friend effect multiplied by
the average number of friends per user) and the total effect (the
per-user effect multiplied by the total number of users) on the
behaviour of everyone in the sample (see Supplementary Informa-
tion). The results suggest that friends generated an additional
886,000 expressed votes (11.4%, null 95% CI 21.1% to 1.1%),
and close friends generated a further 559,000 votes (10.9%, null
95% CI –0.3% to 0.3%). In the Supplementary Information we also
show that close friends of close friends (2 degrees of separation)
generated an additional 1 million expressed votes (11.7%, null 95%
CI –0.8% to 0.9%). Thus, the treatment clearly had a significant impact
on political self-expression and how it spread through the network,
and even weak ties seem to be relevant to its spread.
However, the effect of the social message on real-world validated
vote behaviour and polling-place search was more focused. The results
suggest that close friends generated an additional 282,000 validated
votes (11.8%, null 95% CI –1.3% to 1.2%) and an additional 74,000
polling-place searches (10.1%, null 95% CI –0.1% to 0.1%), but there
is no evidence that ordinary friends had any effect on either of these
two behaviours. In other words, close friendships accounted for all of
the significant contagion of these behaviours, in spite of the fact that
they make up only 7% of all friendships on Facebook.
To put these results in context, it is important to note that turnout
has been steadily increasing in recent US midterm elections, from
36.3% of the voting age population in 2002 to 37.2% in 2006, and to
37.8% in 2010. Our results suggest that the Facebook social message
increased turnout directly by about 60,000 voters and indirectly
through social contagion by another 280,000 voters, for a total of
340,000 additional votes. That represents about 0.14% of the voting
age population of about 236 million in 2010. However, this estimate
does not include the effect of the treatment on Facebook users who
were registered to vote but who we could not match because of
nicknames, typographical errors, and so on. It would be complex to
estimate the number of users on Facebook who are in the voter record
but unmatchable, and it is not clear whether treatment effects would be
of the same magnitude for these individuals, so we restrict our estimate
to the matched group that we were able to sample and observe. This
means it is possible that more of the 0.60% growth in turnout between
2006 and 2010 might have been caused by a single message on
The results of this study have many implications. First and foremost,
online political mobilization works. It induces political self-expression,
but it also induces information gathering and real, validated voter
turnout. Although previous research suggested that online messages
do not work
, it is possible that conventional sample sizes may not
be large enough to detect the modest effect sizes shown here. We
also show that social mobilization in online networks is significantly
more effective than informational mobilization alone. Showing
familiar faces to users can dramatically improve the effectiveness of
a mobilization message.
Beyond the direct effects of online mobilization, we show the
importance of social influence for effecting behaviour change. Our
Probability of being the closest friend
Decile of user–friend interactions
Decile of user–friend interactions Decile of user–friend interactions
Decile of user–friend interactions
Increase in probability of vote (%)
Observed value
Simulated null
95% CI
Increase in probability of expressed vote (%)
Observed value
Simulated null
95% CI
Increase in probability of
polling-place search (%)
Observed value
Simulated null
95% CI
a b
c d
Figure 2
The effect of mobilization treatment that a friend received on a
user’s behaviour. ad, A validation study shows that at increasing levels of
interaction, Facebook friends are more likely to have a close real-world
relationship (a; see also the Supplementary Information). As the interaction
increases, so does the observed per-friend effect of friend’streatment on a user’s
expressed voting (b), validated voting (c) and polling-place search (d). Blue
diamonds indicatethe observed treatment effect. Horizontalgrey bars show the
null distribution derived from simulations of identical networks in which the
topology and incidence of the behaviour and treatment are the same but the
assignments of treatment are randomly reassigned.
13 SEPTEMBER 2012 | VOL 489 | NATURE | 297
Macmillan Publishers Limited. All rights reserved
validation study shows that close friends exerted about four times
more influence on the total numberof validated voters mobilized than
the message itself. These results are similar to those from a prior
network simulation study based on observational data that suggested
each act of voting on average generates an additional three votes as this
behaviour spreads through the network
. Thus, efforts to influence
behaviour should pay close attention not only to the effect a message
will have on those who receive it but also to the likelihood that the
message and the behaviour it spurs will spread from person to person
through the social network. And, in contrast to the results for close
friends, we find that Facebook friends have less effect. Online
mobilization works because it primarily spreads through strong-tie
networks that probably exist offline but have an online representa-
tion. In fact, it is plausible that unobserved face-to-face interactions
account for at least some of the social influencethat we observed in this
More broadly, the results suggest that online messages might
influence a variety of offline behaviours, and this has implications
for our understanding of the role of online social media in society.
Experiments are expensive and have limited external validity, but the
growing availability of cheap and large-scale online social network
means that these experiments can be easily conducted in the
field. If we want to truly understand—and improve—our society,
wellbeing and the world around us, it will be important to use these
methods to identify which real world behaviours are amenable to
online interventions.
Received 17 April; accepted 18 July 2012.
1. Berger, J. & Le Mens, G. How adoption speed affects the abandonment of cultural
tastes. Proc. Natl Acad. Sci. USA 106, 8146–8150 (2009).
2. Borgatti, S. P. et al. Network analysis in the social sciences. Science 323, 892–895
3. Ohtsuki, H., Hauert, C., Lieberman, E. & Nowak, M. A. A simple rule for the
evolution of cooperation on graphs and social networks. Nature 441, 502–505
4. Onnela, J. P. & Reed-Tsochas, F. Spontaneous emergence of social influence in
online systems. Proc. Natl Acad. Sci. USA 107, 18375–18380 (2010).
5. Apicella, C. L., Marlowe, F. W., Fowler, J. H. & Christakis, N. A. Social networks and
cooperation in hunter-gatherers. Nature 481, 497–501 (2012).
6. Aral, S. & Walker, D. Identifying influential and susceptible members of social
networks. Science 337, 337–341 (2012).
7. Centola,D. An experimental study of homophily in the adoption of healthbehavior.
Science 334, 1269–1272 (2011).
8. Gonzalez-Bailon, S., Borge-Holthofer, J., Rivero, A. & Moreno, Y. The dynamics of
protest recruitment through an online network. Sci. Rep. 1, 197 (2011).
9. Christakis, N. A.& Fowler, J. H. Social contagion theory: examining dynamic social
networks and human behavior. Preprint at
10. Aral, S., Muchnik, L. & Sundararajan, A. Distinguishing influence-based contagion
from homophily-driven diffusion in dynamic networks. Proc. Natl Acad. Sci. USA
106, 21544–21549 (2009).
11. Christakis, N. A. & Fowler, J. H. Thespread of obesity in a large social network over
32 years. N. Engl. J. Med. 357, 370–379 (2007).
12. Fowler, J. H. & Christakis, N. A. The dynamic spread of happiness in a large social
network. Br. Med. J. 337, a2338 (2008).
13. Christakis, N. A. & Fowler, J. H. The collectivedynamics of smoking in a large social
network. N. Engl. J. Med. 358, 2249–2258 (2008).
14. Salganik, M. J., Dodds, P. S. & Watts, D. J. Experimental study of inequality and
unpredictability in an artificial cultural market. Science 311, 854–856 (2006).
15. Centola, D. The spreadof behavior in an online social networkexperiment. Science
329, 1194–1197 (2010).
16. Fowler, J. H. & Christakis, N. A. Cooperative behavior cascades in human social
networks. Proc. Natl Acad. Sci. USA 107, 5334–5338 (2010).
17. Lazer, D. et al. Computational social science. Science 323, 721–723 (2009).
18. Ellison, N. B., Steinfield, C. & Lampe, C. The benefits of facebook ‘‘friends:’’ Social
capital and college students’ use ofonline social network sites. J. Comput. Mediat.
Commun. 12, 1143–1168 (2007).
19. Traud, A. L., Kelsic, E. D., Mucha, P. J. & Porter, M. A. Comparing community
structure to characteristics in online collegiate social networks. SIAM Rev. 53,
526–543 (2011).
20. Granovetter, M. The strength of weak ties. Am. J. Sociol. 78, 1360–1380 (1973).
21. Christakis, N. A. & Fowler, J. H. Connected: The Surprising Power of Our Social
Networks and How They Shape Our Lives (Little, Brown, and Company, 2009).
22. Nickerson, D. W. Does email boost turnout? Q. J. Polit. Sci. 2, 369–379 (2007).
23. Kenny, C. B. Political participation andeffects fromthe social environment. Am. J.
Pol. Sci. 36, 259–267 (1992).
24. Huckfeldt, R. & Sprague,J. Citizens, Politics, and Social Communication: Information
and Influence in an Election Campaign (Cambridge Univ. Press, 1995).
25. Gerber, A. S. & Green, D. P. Does canvassing increase voter turnout? A field
experiment. Proc. Natl Acad. Sci. USA 96, 10939–10942 (1999).
26. Gerber, A. S., Green, D. P. & Larimer, C. W. Social pressure and voter turnout:
evidence from a large-scale field experiment. Am. Polit. Sci. Rev. 102, 33–48
27. Bryan, C. J., Walton, G. M., Rogers, T. & Dwecka, C. S. Motivating voter turnout by
invoking the self. Proc. Natl Acad. Sci. USA 108, 12653–12656 (2011).
28. Nickerson, D. W. Is voting contagious? Evidence from two field experiments. Am.
Polit. Sci. Rev. 102, 49–57 (2008).
29. Vitak, J. et al. It’s complicated: Facebook users’ political participation in the 2008
election. Cyberpsychol. Behav. Soc. Netw. 14, 107–114 (2011).
30. Fowler, J. H. in The Social Logic of Politics: Personal Networks as Contexts for Political
Behavior (ed. Zuckerman, A.) 269–287 (Temple Univ. Press, 2005).
Supplementary Information is available in the online version of the paper.
Acknowledgements We are grateful to S. Aral, J. Berger, M. Cebrian, D. Centola,
N. Christakis, C. Dawes, L. Gee, D. Green, C. Kam, P. Loewen, P. Mucha, J. P. Onnela,
M. Porter, O. Smirnov and C. Volden for comments on early drafts. This work was
supported in part by the James S. McDonnell Foundation, and the University of Notre
Dame and the John Templeton Foundation as part of the Science of Generosity
Author Contributions All authors contributed to study design, data collection, analysis
and preparation of the manuscript. J.H.F. secured funding.
Author Information Reprints and permissions information is available at The authors declare no competing financial interests.
Readers are welcome to comment on theonline version of the paper. Correspondence
and requests for materials should be addressed to J.H.F. (
298 | NATURE | VOL 489 | 13 SEPTEMBER 2012
Macmillan Publishers Limited. All rights reserved
... Politics have already made intense use of social media for election campaigns. Such posts can influence real-life voting behavior [87]. Thus, political elections could be increasingly determined by social media or more concisely: P28. ...
... Bond et al. state that social media influence, through peer pressure, encourages other users to vote [87]. Their experiment points out that social media users, who viewed a photo and message of their friends voting, were more likely to vote themselves than users, who only received an informational message without a photo of their friends. ...
Full-text available
Over the past two decades, social media have become a crucial and omnipresent cultural and economic phenomenon, which has seen platforms come and go and advance technologically. In this study, we explore the further development of social media regarding interactive technologies, platform development, relationships to news media, the activities of institutional and organizational users, and effects of social media on the individual and the society over the next five to ten years by conducting an international, two-stage Delphi study. Our results show that enhanced interaction on platforms, including virtual and augmented reality, somatosensory sense, and touch- and movement-based navigation are expected. AIs will interact with other social media users. Inactive user profiles will outnumber active ones. Platform providers will diversify into the WWW, e-commerce, edu-tech, fintechs, the automobile industry, and HR. They will change to a freemium business model and put more effort into combating cybercrime. Social media will become the predominant news distributor, but fake news will still be problematic. Firms will spend greater amounts of their budgets on social media advertising, and schools, politicians, and the medical sector will increase their social media engagement. Social media use will increasingly lead to individuals’ psychic issues. Society will benefit from economic growth and new jobs, increased political interest, democratic progress, and education due to social media. However, censorship and the energy consumption of platform operators might rise.
... Fake news stories pose a vital danger considering social media's potential for social influence and political mobilization. Bond et al. (2012), for instance, conducted a large-scale experiment on Facebook that involved 61-million people. During the 2010 US congressional elections, they showed that political mobilization messages directly influence the voting behavior of millions of people. ...
... Such political messages influence not only the individual but also their friends and their friends of friends. Bond et al. (2012) showed that nearly all transmission occurred between close friends with a high likelihood of having a face-to-face relationship. They conclude that strong ties are instrumental in spreading behavior in human social networks -online and offline. ...
Full-text available
Users are increasingly interacting with machine learning (ML)-based curation systems. YouTube and Facebook, two of the most visited websites worldwide, utilize such systems to curate content for billions of users. Contemporary challenges such as fake news, filter bubbles, and biased predictions make the understanding of ML-based curation systems an important and timely concern. Despite their political, social, and cultural importance, practitioners' framing of machine learning and users' understanding of ML-based curation systems have not been investigated systematically. This is problematic since machine learning - as a novel programming paradigm in which a mapping between input and output is inferred from data - poses a variety of open research questions regarding users' understanding. The first part of this thesis provides the first in-depth investigation of ML-based curation systems as socio-technical systems. The second part of the thesis contributes recommendations on how ML-based curation systems can and should be explained and audited. The first part analyses practitioners' framing of ML by examining how the term machine learning, ML applications, and ML algorithms are framed in tutorials. The thesis also investigates the beliefs that users have about YouTube and introduces a user belief framework of ML-based curation systems. Furthermore, it demonstrates how limited users' capabilities for providing input data for ML-based curation systems are. The second part evaluates different explanations of ML-based systems. This evaluation uncovered an explanatory gap between what is available to explain ML-based curation systems and what users need to understand such systems. Informed by this explanatory gap, the second part of this thesis demonstrates that audits of ML systems can be an important alternative to explanations. This demonstration of audits also uncovers a popularity bias enacted by YouTube's ML-based curation system. Based on these findings, the thesis recommends performing audits to ensure that ML-based systems act in the public's interest.
... The largescale data that result from such experiments enable the detection of subtle but economically important effects across subpopulations. For example, previous work has employed large-scale field experiments to identify social influence (e.g., Aral and Walker 2011, Bakshy et al. 2012, Muchnik et al. 2013, Bond et al. 2012) and estimate the moderating effects of individual (Aral and Walker 2012), dyadic (Aral andWalker 2014, Taylor et al. 2015), and behavioral characteristics (Iyengar et al. 2015, Huang 2016. But, no experiment that we are aware of examines the heterogeneity in social influence across products. ...
Full-text available
We collaborated with a large online social network to conduct a randomized field experiment measuring social ad effectiveness across 71 products in 25 categories among more than 37 million users.
... Multidimensional user data on massive, concentrated user platforms provide material to statistical and machine learning techniques that have the potential to profile, infer, and predict people's actions in ways that have turned out to be incredibly profitable (Hendricks & Vestergaard, 2019). Improving the accuracy of user profiles also enables precisely targeted marketing, behavior engineering, attention harvesting, and swaying one's moods, among an endless number of other benign and not so benign uses (Kramer, Guillory, & Hancock, 2014;Bond et al., 2012;Hendricks & Vestergaard, 2019). ...
Full-text available
While much has been written about the personal, social, and democratic benefits of networked communities and partici-patory learning, critics have begun to draw attention to the ubiquitous data collection and computational processes behind mass user platforms. Personal and behavioral data have become valuable material for statistical and machine learning techniques that have the potential to profile, infer, and predict people's needs, values, and behavior. As a response, researchers are calling for data literacies and computational thinking to facilitate people's capacity and volition to make informed actions in their digital world. Yet, efforts and curricula towards a greater understanding of computational mechanisms of new media ecology are sorely missing from K12-education as well as from teacher education. This paper provides an overview of tensions that teachers and educators will face when they attempt to bridge participatory learning with a more robust understanding of machine learning and algorithmic production of social and cultural practices.
... Our framework encompasses a large number of examples from the literature where interference naturally occurs: information campaigns (Banerjee et al., 2013;Bond et al., 2012;Jones et al., 2017), cash transfer programs (Egger et al., 2019), health-programs Figure 1: Data on insurance adoption in rural China from Cai et al. (2015). The left panel reports the probability of insurance adoption as a function of the percentage of treated individuals, showing decreasing marginal effects. ...
Full-text available
This paper discusses experimental design for inference and estimation of individualized treatment allocation rules in the presence of unknown interference, with units being organized into large independent clusters. The contribution is two-fold. First, we design a short pilot study with few clusters for testing whether base-line interventions are welfare-maximizing, with its rejection motivating larger-scale experimentation. Second, we introduce an adaptive randomization procedure to estimate welfare-maximizing individual treatment allocation rules valid under unobserved interference. We propose non-parametric estimators of direct treatments and marginal spillover effects, which serve for hypothesis testing and policy-design. We discuss the asymptotic properties of the estimators and small sample regret guarantees of the estimated policy. Finally, we illustrate the method's advantage in simulations calibrated to an existing experiment on information diffusion.
... As such, they are unlikely to influence each other's treatment take-up decisions, even if they may impose employment externalities on one another. IOR is also reasonable in online settings where other subjects' take-up decisions are unobserved (Anderson et al., 2014;Bond et al., 2012;Eckles et al., 2016) or confidential (Yi et al., 2015). ...
Full-text available
This paper shows how to use a randomized saturation experimental design to identify and estimate causal effects in the presence of social interactions--one person's treatment may affect another's outcome--and one-sided non-compliance--subjects can only be offered treatment, not compelled to take it up. Two distinct causal effects are of interest in this setting: direct effects quantify how a person's own treatment changes her outcome, while indirect effects quantify how her peers' treatments change her outcome. We consider the case in which social interactions occur only within known groups, and take-up decisions do not depend on peers' offers. In this setting we point identify local average treatment effects, both direct and indirect, in a flexible random coefficients model that allows for both heterogenous treatment effects and endogeneous selection into treatment. We go on to propose a feasible estimator that is consistent and asymptotically normal as the number and size of groups increases.
... Social networks are increasingly becoming the primary channel for people to acquire information and form opinions. In an experiment involving 60m Facebook users prior to the 2010 US elections, Bond et al. [2012] showed they could generate 340,000 additional votes using a social message that informed a user about friends that had voted, compared to an informational message without social network information. Unlike traditional media or gatherings in the local church, club or pub, online social networks make it very easy for a user to "unfollow" someone who does not share their opinion. ...
In recent years online social networks have become increasingly prominent in political campaigns and, concurrently, several countries have experienced shock election outcomes. This paper proposes a model that links these two phenomena. In our set-up, the process of learning from others on a network is influenced by confirmation bias, i.e. the tendency to ignore contrary evidence and interpret it as consistent with one's own belief. When agents pay enough attention to themselves, confirmation bias leads to slower learning in any symmetric network, and it increases polarization in society. We identify a subset of agents that become more/less influential with confirmation bias. The socially optimal network structure depends critically on the information available to the social planner. When she cannot observe agents' beliefs, the optimal network is symmetric, vertex-transitive and has no self-loops. We explore the implications of these results for electoral outcomes and media markets. Confirmation bias increases the likelihood of shock elections, and it pushes fringe media to take a more extreme ideology.
... Investigating the processes of social influence on OSNs is a challenging task that can, nonetheless, bring new insight into our knowledge of how individuals form their opinions (Bond et al., 2012;Ravazzi et al., 2020). This is primarily because OSNs offer an unprecedented opportunity to observe the opinion dynamics of users unobtrusively, avoiding the interviewer effect (Barberá, 2014b). ...
Full-text available
In this paper, we present an empirical study of the opinion dynamics of a large-scale sample of online social network users. Opinions of users are estimated based on their subscriptions to information sources and we analyze how friendship connections affect the dynamics of these estimations. Our analysis has revealed traces of both positive (assimilative) and negative (repulsive) influence in how users decide whether to change their opinions, while magnitudes of changed opinions appear to obey a positive influence. We also found that individuals with networks that include many ideologically opposite users are less prone to radicalization than those with homogenous ones. Our results indicate that the opinion dynamics of online social network users cannot be described solely by any concrete theoretical opinion formation model; rather, they include patterns of multiple models simultaneously.
... The study of causality has demonstrated a multi-disciplinary impact at all four levels. Network science is another example of common issues and includes topics such as studying influence in networks [49][50][51][52] and network propagation methods [53,54] . Ultimately, a widespread practical deployment of any technology might produce "adverse side effects" misusing the knowhow. ...
Technological advances of virtually every kind pose risks to society including fairness and bias. We review a long-standing wisdom that a widespread practical deployment of any technology may produce adverse side effects misusing the knowhow. This includes AI but AI systems are not solely responsible for societal risks. We describe some of the common and AI specific risks in health industries and other sectors and propose both broad and specific solutions. Each technology requires very specialized and informed tracking, monitoring and creative solutions. We postulate that AI systems are uniquely poised to produce conceptual and methodological solutions to both fairness and bias in automated decision-making systems. We propose a simple intelligent system quotient that may correspond to their adverse societal impact and outline a multi-tier architecture for producing solutions of increasing complexity to these risks. We also propose that universities may consider forming interdisciplinary Study of Future Technology Centers to investigate and predict the fuller range of risks posed by technology and seek both common and AI specific solutions using computational, technical, conceptual and ethical analysis
Public sector recruitment is an urgent and prevailing challenge in both research and practice. Public employer branding is an important subject in the theoretical debate, but the mechanisms behind how certain signals of public employers affect individuals’ interest in a job are under‐researched. By bridging signaling theory, social identity theory, and personnel economics, this study analyzes the effects of signals in advertisements related to societal impact, job security, and performance orientation on different gender‐/age‐based target groups. This series of pre‐registered social media field experiments (n = 196,822 persons) with four public employers examines the degree to which these signals affect individuals’ interest in a job at a public employer. The results do not show an overall impact of the signals but target group‐specific effects—gender has a significant effect and age for certain public employers. Compared to the societal impact signal, the job security signal has a slightly stronger effect.
Full-text available
Members of the same household share similar voting behaviors on average, but how much of this correlation can be attributed to the behavior of the other person in the household? Disentangling and isolating the unique effects of peer behavior, selection processes, and congruent interests is a challenge for all studies of interpersonal influence. This study proposes and utilizes a carefully designed placebo-controlled experimental protocol to overcome this identification problem. During a face-to-face canvassing experiment targeting households with two registered voters, residents who answered the door were exposed to either a Get Out the Vote message (treatment) or a recycling pitch (placebo). The turnout of the person in the household not answering the door allows for contagion to be measured. Both experiments find that 60% of the propensity to vote is passed onto the other member of the household. This finding suggests a mechanism by which civic participation norms are adopted and couples grow more similar over time.
The prevalence of obesity has increased substantially over the past 30 years. We performed a quantitative analysis of the nature and extent of the person-to-person spread of obesity as a possible factor contributing to the obesity epidemic. We evaluated a densely interconnected social network of 12,067 people assessed repeatedly from 1971 to 2003 as part of the Framingham Heart Study. The body-mass index was available for all subjects. We used longitudinal statistical models to examine whether weight gain in one person was associated with weight gain in his or her friends, siblings, spouse, and neighbors. Discernible clusters of obese persons (body-mass index [the weight in kilograms divided by the square of the height in meters], > or =30) were present in the network at all time points, and the clusters extended to three degrees of separation. These clusters did not appear to be solely attributable to the selective formation of social ties among obese persons. A person's chances of becoming obese increased by 57% (95% confidence interval [CI], 6 to 123) if he or she had a friend who became obese in a given interval. Among pairs of adult siblings, if one sibling became obese, the chance that the other would become obese increased by 40% (95% CI, 21 to 60). If one spouse became obese, the likelihood that the other spouse would become obese increased by 37% (95% CI, 7 to 73). These effects were not seen among neighbors in the immediate geographic location. Persons of the same sex had relatively greater influence on each other than those of the opposite sex. The spread of smoking cessation did not account for the spread of obesity in the network. Network phenomena appear to be relevant to the biologic and behavioral trait of obesity, and obesity appears to spread through social ties. These findings have implications for clinical and public health interventions.
This paper examines the distinction made by Huckfeldt (1979, 1986) and Giles and Dantico (1982) between individually and socially based forms of participation as affected by social environment. Using survey responses from the 1984 South Bend study, the relationship between political discussion partners is explicitly modeled and estimated for several forms of individually and socially based participatory acts. The evidence indicates that certain types of both individually based and socially based participation are affected by those in the immediate social environment, suggesting that a modification of this distinction is in order.
Political campaigns are just now learning how to put the Internet to best use. Low transaction costs and huge economies of scale tempt campaigns to move traditional activities online, but the effectiveness of virtual campaigns is unknown. This paper conducts 13 field experiments on 232,716 subjects to test whether email campaigns are effective for voter registration and mobilization. Both registration and turnout were unaffected, suggesting that email, while inexpensive, is not cost-effective. Learning to use television as a campaign medium took politicians years, and candidates are now beginning to figure out how to use the Internet. An intuitive place to begin is by using the Internet to accomplish work previously done with older technology such as mail, phones, or face-to-face canvassing. The Internet's low transaction costs and massive economies of scale could alter the strategies parties employ in every facet of campaigning. The same economics that push businesses to move online are also present in the political realm. Unfortunately, studies of Internet usage during campaigns generally report on the content of websites (Farmer and Fender 2003, Norris 2003, Ward and Gibson 2003, Farmer and Fender 2005, Xenos and Foot 2005) and do not measure how voters respond to the online campaign activities.1 This paper evaluates the effectiveness of email as a voter mobilization tool by conduct- ing 13 field experiments. Direct mail has been shown to be an effective, albeit expensive, means of increasing voter turnout (Gerber et al. 2003). Examining all known randomized
The abstract for this document is available on CSA Illumina.To view the Abstract, click the Abstract button above the document title.
Democratic politics is a collective enterprise, not simply because individual votes are counted to determine winners, but more fundamentally because the individual exercise of citizenship is an interdependent undertaking. Citizens argue with one another, they inform one another, and they generally arrive at political decisions through processes of social interaction and deliberation. This book is dedicated to investigating the political implications of interdependent citizens within the context of the 1984 presidential election campaign as it was experienced in the metropolitan area of South Bend, Indiana. Hence, this is a community study in the fullest sense of the term. National politics is experienced locally through a series of filters unique to a particular setting. And this study is concerned with understanding that setting and its consequences for the exercise of democratic citizenship. Several different themes structure the undertaking: the dynamic implications of social communication among citizens, the importance of communication networks for citizen decision making, the exercise of citizen purpose in locating sources of information, the constraints on individual choice that arise as a function of contexts and environments, and the institutional and organizational effects that operate on the flow of information within particular settings. (PsycINFO Database Record (c) 2012 APA, all rights reserved)
The Social Logic of Politics: Personal Networks as Contexts for Political Behavior. Edited by Alan S. Zuckerman. Philadelphia: Temple University Press, 2005. 368p. $72.50 cloth, $25.95 paper. The central premise of this edited collection is set out with admirable clarity on the first page of the opening chapter: “It is both obvious and well-known that the immediate social circumstances of people's lives influence what they believe and do about politics. Even so, relatively few political scientists incorporate these principles into their analysis” (p. 3). Alan Zuckerman tackles this problem with a selection of chapters, written by authors with a range of intellectual pedigrees, that set out to show how what is “both obvious and well-known” can be incorporated into rigorous political science. The individual chapters are too numerous and diverse to review in detail here. What is perhaps more useful is to consider the extent to which, taken together, they map out a potentially fruitful line of future development for the discipline.