SCIENCE ADVANCES | RESEARCH ARTICLE

SOCIAL SCIENCES Copyright © 2019

The Authors, some
Less than you think: Prevalence and predictors of fake rights reserved;
exclusive licensee
news dissemination on Facebook American Association
for the Advancement
of Science. No claim to
Andrew Guess1*, Jonathan Nagler2, Joshua Tucker2 original U.S. Government
Works. Distributed
So-called “fake news” has renewed concerns about the prevalence and effects of misinformation in political cam- under a Creative
paigns. Given the potential for widespread dissemination of this material, we examine the individual-level char- Commons Attribution
acteristics associated with sharing false articles during the 2016 U.S. presidential campaign. To do so, we uniquely NonCommercial
link an original survey with respondents’ sharing activity as recorded in Facebook profile data. First and foremost, License 4.0 (CC BY-NC).
we find that sharing this content was a relatively rare activity. Conservatives were more likely to share articles
from fake news domains, which in 2016 were largely pro-Trump in orientation, than liberals or moderates. We also
find a strong age effect, which persists after controlling for partisanship and ideology: On average, users over 65 shared
nearly seven times as many articles from fake news domains as the youngest age group.

INTRODUCTION sults hold when alternate lists are used, such as that used by peer-­
One of the most discussed phenomena in the aftermath of the 2016 U.S. reviewed studies (2).

Downloaded from https://www.science.org on February 28, 2025

presidential election was the spread and possible influence of “fake Overall, sharing articles from fake news domains was a rare ac-
news”—false or misleading content intentionally dressed up to look tivity. We find some evidence that the most conservative users were
like news articles, often for the purpose of generating ad revenue. more likely to share this content—the vast majority of which was pro-
Scholars and commentators have raised concerns about the impli- Trump in orientation—than were other Facebook users, although
cations of fake news for the quality of democratic discourse, as well this is sensitive to coding and based on a small number of respon-
as the prevalence of misinformation more generally (1). Some have dents. Our most robust finding is that the oldest Americans, especially
gone so far as to assert that such content had a persuasive impact those over 65, were more likely to share fake news to their Facebook
that could have affected the election outcome, although the best ev- friends. This is true even when holding other characteristics—including
idence suggests that these claims are farfetched (2). While evidence education, ideology, and partisanship—constant. No other demographic
is growing on the prevalence (3), believability (2), and resistance to characteristic seems to have a consistent effect on sharing fake news,
corrections (4, 5) of fake news during the 2016 campaign, less is known making our age finding that much more notable.
about the mechanisms behind its spread (6). Some of the earliest jour-
nalistic accounts of fake news highlighted its popularity on social media,
especially Facebook (7). Visits to Facebook appear to be much more RESULTS
common than other platforms before visits to fake news articles in It is important to be clear about how rare this behavior is on social
web consumption data, suggesting a powerful role for the social net- platforms: The vast majority of Facebook users in our data did not
work (3), but what is the role of social transmission—in particular, share any articles from fake news domains in 2016 at all (Fig. 1), and
social sharing—in the spread of this pernicious form of false political as the left panel shows, this is not because people generally do not
content? Here, we provide important new evidence complementing the share links: While 3.4% of respondents for whom we have Facebook
small but growing body of literature on the fake news phenomenon. profile data shared 10 or fewer links of any kind, 310 (26.1%) respon-
dents shared 10 to 100 links during the period of data collection and
Data and method 729 (61.3%) respondents shared 100 to 1000 links. Sharing of stories
Our approach allows us to provide a comprehensive observational from fake news domains is a much rarer event than sharing links over-
portrait of the individual-level characteristics related to posting arti- all. The right panel of Fig. 1 reveals a large spike at 0, with a long tail
cles from fake news–spreading domains to friends on social media. that goes as far as 50 shares for a single Facebook user, and we see in
We link a representative online survey (N = 3500) to behavioral data Table 1 that over 90% of our respondents shared no stories from fake
on respondents’ Facebook sharing history during the campaign, avoid- news domains. According to our main measure of fake news con-
ing known biases in self-reports of online activity (8, 9). Posts con- tent, 8.5% of respondents for whom we have linked Facebook data
taining links to external websites are cross-referenced against lists shared at least one such article to their friends. Again referencing Fig. 1,
of fake news publishers built by journalists and academics. Here, we among those who shared fake news to their friends, more were Re-
mainly use measures constructed by reference to the list by Silverman publicans, both in absolute (38 Republican versus 17 Democratic
(7), but in the Supplementary Materials, we show that the main re- respondents) and in relative (18.1% of Republicans versus 3.5% of
Democrats in our sample) terms.
We further explore the factors that explain the variation in fake
news sharing behavior. As shown in Fig. 2A, Republicans in our sam-
1
Department of Politics and Woodrow Wilson School, Princeton University, Fisher ple shared more stories from fake news domains than Democrats;
Hall, Princeton, NJ 08544, USA. 2Wilf Family Department of Politics and Social Me- moreover, self-described independents on average shared roughly as
dia and Political Participation (SMaPP) Lab, New York University, 19 West 4th Street,
New York, NY 10012, USA. many as Republicans (0.506 and 0.480, respectively). A similar pat-
*Corresponding author. Email: [email protected] tern is evident for ideology (Fig. 2C): Conservatives, especially those

Guess et al., Sci. Adv. 2019; 5 : eaau4586 9 January 2019 1 of 8

SCIENCE ADVANCES | RESEARCH ARTICLE

Fig. 1. Distribution of total and fake news shares. (Left) Histogram of the total number of links to articles on the web shared by respondents in the sample who iden-
tified as Democrats, Republicans, or independents. (Right) Stacked histogram of the number of fake news articles shared by respondents who identified as Democrats,

Downloaded from https://www.science.org on February 28, 2025

Republicans, or independents using the measure derived from (7).

Table 1. Distribution of fake news shares.

0 1 2 3 4 5–10 11–50
1090 (91.5%) 63 (5.3%) 12 (1.0%) 8 (<1.0%) 5 (<1.0%) 9 (<1.0%) 4 (<1.0%)

identifying as “very conservative,” shared the most articles from fake association is also robust to controlling for party, as the various
news domains. On average, a conservative respondent shared 0.75 alternative specifications provided in the Supplementary Materials
such stories [95% confidence interval (CI), 0.537 to 0.969], and a very illustrate. In column 2, the coefficient on “Age: over 65” implies that
conservative respondent shared 1.0 (95% CI, 0.775 to 1.225). This is being in the oldest age group was associated with sharing nearly seven
consistent with the pro-Trump slant of most fake news articles pro- times as many articles from fake news domains on Facebook as those
duced during the 2016 campaign, and of the tendency of respon- in the youngest age group, or about 2.3 times as many as those in the
dents to share articles they agree with, and thus might not represent next-oldest age group, holding the effect of ideology, education, and the
a greater tendency of conservatives to share fake news than liberals total number of web links shared constant (e1.9 ≈ 6.69, e1.9−1.079 ≈ 2.27).
conditional on being exposed to it (3). This association is also found in the specifications using the alternate
Figure 2D shows that, if anything, those who share the most content peer-reviewed measure (2) as a dependent variable in columns 3 and 4,
in general were less likely to share articles from fake news–spreading with those over 65 sharing between three and four times as many
domains to their friends. Thus, it is not the case that what explains fake news links as those in the youngest age group.
fake news sharing is simply that some respondents “will share any- Aside from the overall rarity of the practice, our most robust and
thing.” These data are consistent with the hypothesis that people who consistent finding is that older Americans were more likely to share
share many links are more familiar with what they are seeing and articles from fake news domains. This relationship holds even when
are able to distinguish fake news from real news. (We note that we we condition on other factors, such as education, party affiliation,
have no measure as to whether or not respondents know that what ideological self-placement, and overall posting activity. It is robust
they are sharing is fake news.) Turning to a key demographic char- to a wide range of strategies for measuring fake news (see Materials
acteristic of respondents, a notable finding in Fig. 2B is the clear as- and Methods). Further, none of the other demographics variables in
sociation between age group and the average number of articles from our model—sex, race, education, and income—have anywhere close
fake news domains shared on Facebook. Those over 65 shared an average to a robust predictive effect on sharing fake news. We subject our
of 0.75 fake news articles (95% CI, 0.515 to 0.977), more than twice findings to a battery of robustness tests in the Supplementary Materials.
as many as those in the second-oldest age group (0.26 articles; 95% CI, Among them, we show that model specification, other predictors such
0.206 to 0.314). Of course, age is correlated with other characteristics, as political knowledge, and distributional assumptions about the de-
including political predispositions. Thus, we turn to a multivariate anal- pendent variable do not appear to be driving our results (tables S1 to
ysis to examine the marginal impact of individual characteristics. S8 and S13). Last, we show in table S14 that, when we try to explain
Table 2 shows that the age effect remains statistically significant patterns of hard news sharing behavior using the same approach, the
when controlling for ideology and other demographic attributes. The predictors are more varied and do not include age.

Guess et al., Sci. Adv. 2019; 5 : eaau4586 9 January 2019 2 of 8

SCIENCE ADVANCES | RESEARCH ARTICLE

Downloaded from https://www.science.org on February 28, 2025

Fig. 2. Average number of fake news shares (and 95% CIs) using the list of domains derived from (7). (A) Party identification, (B) age group, (C) ideological
self-placement, and (D) overall number of Facebook wall posts. Proportions adjusted to account for sample-matching weights derived from the third wave of the SMaPP
YouGov panel survey.

DISCUSSION pected. More puzzling is the independent role of age: Holding con-
Using unique behavioral data on Facebook activity linked to individual- stant ideology, party identification, or both, respondents in each age
level survey data, we find, first, that sharing fake news was quite rare category were more likely to share fake news than respondents in the
during the 2016 U.S. election campaign. This is important context next-youngest group, and the gap in the rate of fake news sharing
given the prominence of fake news in post-election narratives about between those in our oldest category (over 65) and youngest category
the role of social media disinformation campaigns. Aside from the is large and notable.
relatively low prevalence, we document that both ideology and age These findings pose a challenge and an opportunity for social sci-
were associated with that sharing activity. Given the overwhelming entists. Political scientists tend to favor explanations based on sta-
pro-Trump orientation in both the supply and consumption of fake ble, deeply held partisan or ideological predispositions (10, 11). The
news during that period, including via social pathways on Facebook predictive power of demographic traits evaporates when subjected
(3), the finding that more conservative respondents were more likely to multiple regression analyses that control for other characteristics cor-
to share articles from fake news–spreading domains is perhaps ex- related with those demographics. Yet, when an empirical relationship

Guess et al., Sci. Adv. 2019; 5 : eaau4586 9 January 2019 3 of 8

SCIENCE ADVANCES | RESEARCH ARTICLE

Table 2. Determinants of fake news sharing on Facebook. Quasi-Poisson models with YouGov’s sample-matching weights applied. Dependent variables are
counts of fake news articles shared using measures derived from (7) (columns 1 and 2) and (2) (columns 3 and 4). The reference category for ideology is “Not
sure.” “Number of links shared” refers to the number of Facebook posts by each respondent that includes a link to an external URL. A&G, Allcott and Gentzkow.
Number of stories shared Number of stories shared (A&G)
(1) (2) (3) (4)
Very liberal 0.487 0.387 1.634* 1.485*
(1.238) (1.209) (0.876) (0.800)
Liberal −1.127 −1.141 0.873 0.812
(1.439) (1.404) (0.886) (0.809)
Moderate 0.333 0.392 0.748 0.824
(1.186) (1.157) (0.875) (0.799)
Conservative 2.187* 2.248** 1.736** 1.800**
(1.155) (1.128) (0.868) (0.794)
Very conservative 2.366** 2.297** 2.231** 2.087***
(1.158) (1.132) (0.869) (0.795)
Age: 30–44 0.772 0.742 0.253 0.172

Downloaded from https://www.science.org on February 28, 2025

(0.811) (0.791) (0.390) (0.356)
Age: 45–65 1.136 1.079 0.602* 0.488
(0.765) (0.746) (0.359) (0.328)
Age: over 65 2.052*** 1.900** 1.389*** 1.152***
(0.766) (0.750) (0.362) (0.333)
Female −0.114 0.008 −0.329** −0.199
(0.217) (0.219) (0.155) (0.146)
Black −0.880 −0.806 −0.609 −0.536
(0.754) (0.736) (0.400) (0.366)
Education −0.085 −0.091 −0.021 −0.021
(0.081) (0.081) (0.055) (0.052)
Income −0.007 −0.007 0.003 0.003
(0.008) (0.008) (0.004) (0.003)
Number of links shared 0.001*** 0.001***
(0.0002) (0.0001)
Constant −3.416** −3.635*** −1.201 −1.502*
(1.379) (1.348) (0.931) (0.851)
N 1041 1040 1041 1040

*P < 0.1   **P < 0.05   ***P < 0.01.

such as the one documented here emerges, we are challenged to view and beyond, lacks the level of digital media literacy necessary to reli-
demographic traits not as controls to be ignored but as central ex- ably determine the trustworthiness of news encountered online (13, 14).
planatory factors above and beyond the constructs standard in the There is a well-established research literature on media literacy and
literature (12). This is especially the case with age, as the largest gener- its importance for navigating new media technologies (15). Build-
ation in America enters retirement at a time of sweeping demographic ing on existing work (16, 17), researchers should further develop
and technological change. Below, we suggest possible avenues for further competency-based measures of digital media literacy that encompass
research incorporating insights from multiple disciplines. the kinds of skills needed to identify and avoid dubious content de-
Given the general lack of attention paid to the oldest generations signed to maximize engagement. Research on age and digital media
in the study of political behavior thus far, more research is needed literacy often focuses on youth skills acquisition and the divide be-
to better understand and contextualize the interaction of age and tween “digital natives” and “digital immigrants” (18), but our results
online political content. Two potential explanations warrant further suggest renewed focus on the oldest age cohorts.
investigation. First, following research in sociology and media stud- Within this cohort, lower levels of digital literacy could be com-
ies, it is possible that an entire cohort of Americans, now in their 60s pounded by the tendency to use social endorsements as credibility

Guess et al., Sci. Adv. 2019; 5 : eaau4586 9 January 2019 4 of 8

SCIENCE ADVANCES | RESEARCH ARTICLE

cues (19). If true, this would imply a growing impact as more Amer- sibility. Similarly, it may be the case that the composition of older
icans from older age groups join online social communities. A second Facebook users’ News Feeds differs in systematically important ways
possibility, drawn from cognitive and social psychology, suggests a from that of younger users; while we lack the data in the present work
general effect of aging on memory. Under this account, memory de- to test this proposition, future research with access to these data could
teriorates with age in a way that particularly undermines resistance prove illuminating. These concerns aside, the evidence we have pre-
to “illusions of truth” and other effects related to belief persistence sented is strongly suggestive of an emerging relationship between not
and the availability heuristic, especially in relation to source cues (20–22). only ideological affiliation but also age and the still-rare tendency to
The severity of these effects would theoretically increase with the spread misinformation to one’s friends on social media platforms.
complexity of the information environment and the prevalence of If the association with age holds in future studies, there are a whole
misinformation. host of questions about the causes, mechanisms, and ramifications
We cannot definitively rule out the possibility that there is an omit- of the relationship that researchers should explore. First, how much
ted variable biasing our estimates, although we have included con- of the effect can be attributed to lack of digital or media literacy as
trols for many individual-level characteristics theoretically related to opposed to explanations rooted in cognitive abilities and memory?
acceptance of misinformation and willingness to share content online. Answering this question will require developing measures of digital
Even if our models are correctly specified, we use observational data literacy that can be behaviorally validated. Second, what is the role
that cannot provide causal evidence on the determinants of fake news of the (currently) unobserved News Feed and social network envi-
sharing. This study takes advantage of a novel and powerful new dataset ronment on people’s tendency to see, believe, and spread dubious
combining survey responses and digital trace data that overcomes content? How are consumption and spreading, if at all, related? How
well-known biases in sample selection and self-reports of online be- does social trust over networks mediate the relationship between age
havior (8, 9). However, we are still limited in our ability to collect and the sharing of misinformation?

Downloaded from https://www.science.org on February 28, 2025

these data unobtrusively. Despite our high response rate, half of our Last, if media literacy is a key explanatory factor, then what are
respondents with Facebook accounts opted not to share their profile the interventions that could effectively increase people’s ability to
data with us. Any inferences are therefore limited insofar as the like- discern information quality in a complex, high-choice media envi-
lihood of sharing data is correlated with other characteristics of interest. ronment replete with contradictory social and political cues? Both
In addition, while our approach allows for enhanced measure- theory and existing curricula could serve as the basis of rigorously
ment of online sharing behavior, we lack data on the composition of controlled evaluations, both online and in the classroom, which could
respondents’ Facebook News Feeds. It is possible, for instance, that then help to inform educational efforts targeted at people in differ-
very conservative Facebook users were exposed to more fake news ent age groups and with varying levels of technological skill. These
articles in their networks and that the patterns we observe are not efforts leave open the possibility that simple interventions, perhaps
due to differential willingness to believe or share false content across even built into online social environments, could reduce the spread
the political spectrum. While other evidence suggests the limits of of misinformation by those most vulnerable to deceptive content.
the “echo chambers” narrative (23), we cannot rule out this pos- Developing these innovations would be further aided by increased

Table 3. Comparison of samples. FB, Facebook.

Full sample Sample 2* Sample 3† P‡ Sample 4§
% Democrat 31 32 40 0.17 40
Mean ideology 2.98 2.89 2.76 0.01 2.75
(five-point)
% Vote intention 36 37 47 0.07 47
(Clinton)
% Voted in 2016 general 0.59 0.59 0.63 0.01 0.63
% Knowledge (0–2) 2.05 2.04 2.13 0.03 2.13
Mean age 51 49 49 0.16 49
% High school or less 23 20 22 0.17 22
Self-reported
% Post to FB several 26 28 0.28 28
times/day
% Look at FB often 65 67 0.42 68
N 3500 2711 1331 1191

*Column 2 summarizes characteristics of respondents who said in the survey that they have a Facebook account (i.e., they selected “Facebook” from the list of
response options to the question “Do you have accounts on any of the following social media sites?”).   †Column 3 subsets to respondents (regardless of
their answer in the previous question) who consented to share Facebook profile information with the researchers.   ‡P values are computed from t tests of
the difference in means between the sample of respondents who reported having a Facebook account and those who consented to provide access to their
profile data.   §The final column subsets to those who shared any Facebook data at all that we were able to link back to the survey.

Guess et al., Sci. Adv. 2019; 5 : eaau4586 9 January 2019 5 of 8

SCIENCE ADVANCES | RESEARCH ARTICLE

cooperation between academic researchers and the platforms them- genuine—posts that our respondents did not feel compelled to re-
selves (24). move at a later time. There may be an additional concern that some
types of people were more likely to delete fake news articles that they
posted, leading us to biased inferences. However, to the extent that
MATERIALS AND METHODS these characteristics are negatively correlated with the characteristics
Survey data that predict posting in the first place, such deletion activity (which
We designed and conducted a panel survey (fielded by online polling is likely very rare) should reduce noise in the data that would other-
firm YouGov) during the 2016 U.S. presidential election to under- wise be generated by split-second sharing decisions that are imme-
stand how social media use affects the ways that people learn about diately retracted.
politics during a campaign. In addition to including a rich battery of
individual-level covariates describing media use and social media use, Defining fake news
we were able to match many respondents to data on their actual The term fake news can be used to refer to a variety of different
Facebook behavior (see below). The survey had three waves. Wave 1 phenomena. Here, we largely adopted the use suggested in (25) of
was fielded 9 April to 1 May 2016 (3500 respondents), wave 2 was knowingly false or misleading content created largely for the purpose
fielded 9 September to 9 October 2016 (2635 respondents), and wave 3 of generating ad revenue. Given the difficulty of establishing a com-
was fielded 25 October to 7 November 2016 (2628 respondents). monly accepted ground-truth standard for what constitutes fake news,
our approach was to build on the work of both journalists and academ-
Facebook profile data ics who worked to document the prevalence of this content over the
We were able to obtain private Facebook profile data from a substan- course of the 2016 election campaign. In particular, we used a list of
tial subset of our survey respondents. Starting 16 November 2016, fake news domains assembled by Craig Silverman of BuzzFeed News,

Downloaded from https://www.science.org on February 28, 2025

YouGov e-mailed all respondents with a request to temporarily share the primary journalist covering the phenomenon as it developed (7).
information from their Facebook profiles with us. That request read, As a robustness check, we constructed alternate measures using a list
in part: “We are interested in the news people have read and shared on curated by Allcott and Gentzkow (2), who combined multiple sources
Facebook this year. To save time and to get the most accurate infor- across the political spectrum (including some used by Silverman) to
mation, with your permission we can learn this directly from Facebook. generate a list of fake news stories specifically debunked by fact-
Facebook has agreed to help this way. Of course, we would keep this checking organizations.
information confidential, just like everything else you tell us.” The Silverman list is based on the most-shared web domains during
Those who consented to do so were asked to provide access to a the election campaign as determined by the analytics service BuzzSumo.
Facebook web application and to specifically check which types of Silverman and his team followed up their initial results with in-depth
information they were willing to share with us: fields from their public reporting to confirm whether a domain appeared to have the hall-
profile, including religious and political views; their own timeline mark features of a fake news site: lacking a contact page, featuring a
posts, including external links; and “likes” of pages. We did not have high proportion of syndicated content, being relatively new, etc. We
access to the content of people’s News Feeds or information about took this list and removed all domains classified as “hard news” via
their friends. Respondents read a privacy statement that informed the supervised learning technique used by Bakshy et al. (23) to focus
them that they could deactivate the application at any time and that specifically on fake news domains rather than the more contested
we would not share any personally identifying information. The category of “hyperpartisan” sites (such as Breitbart). (The authors
app provided access for up to 2 months after respondents who chose used section identifiers in article URLs shared on Facebook that are
to share data agreed to do so. This data collection was approved by associated with hard news—“world,” “usnews,” etc.—to train a ma-
the New York University Institutional Review Board (IRB-12-9058 chine learning classifier on text features. They ultimately produced
and IRB-FY2017–150). a list of 495 domains with both mainstream and partisan websites
Of 3500 initial respondents in wave 1, 1331 (38%) agreed to share that produce and engage with current affairs.) The resulting list con-
Facebook profile data with us. The proportion rises (49.1%) when tains 21 mostly pro-Trump domains, including well-known purveyors
we consider that only 2711 of our respondents said that they use such as abcnews.com.co, the Denver Guardian, and Ending the Fed.
Facebook at all. We were successfully able to link profile data from In analyses using this list, we counted any article from one of these
1191 of survey respondents, leaving us with approximately 44% of domains as a fake news share. (See below for details on these coding
Facebook users in our sample. (See the “Sample details” section for a procedures and a list of domains in what we refer to as our main
comparison of sample characteristics on various demographic and BuzzFeed-based list.)
behavioral dimensions. Linked respondents were somewhat more The Allcott and Gentzkow list begins with 948 fact checks of false
knowledgeable and engaged in politics than those who did not share stories from the campaign. We retrieved the domains of the publish-
data.) For the purposes of this analysis, we parsed the raw Facebook ers originating the claims and again removed all hard news domains
profile data and identified the domains of any links posted by re- as described above. Then, we coded any article from this set of do-
spondents to their own timelines. mains as a fake news article. For robustness, in table S9, we used
We did not have access to posts that respondents deleted before only exact URL matches to any of the 948 entries in the Allcott and
consenting to temporarily share their data with us. It is theoretically Gentzkow list as a more restrictive definition of fake news, but one
possible that some respondents posted fake news articles to their pro- that does not require assuming that every article from a “fake news
files and then deleted them before we had the opportunity to collect domain” should be coded as fake news. Since the list contains the
the data. To the extent that this activity reflects second-guessing or researchers’ manual coding of the slant of each article, we also pre-
awareness of how fake news posting is perceived by social connec- sented models analyzing pro-Trump and pro-Clinton fake news sharing
tions, we interpret the sharing data that we were able to gather as activity only.

Guess et al., Sci. Adv. 2019; 5 : eaau4586 9 January 2019 6 of 8

SCIENCE ADVANCES | RESEARCH ARTICLE

Additional lists (5) truepundit.com

In addition to these primary measures, we report (below) analyses (6) redstatewatcher.com
using three supplementary collections of fake news articles produced (7) worldpoliticus.com
after the election. Two lists were also produced by Silverman and his (8) subjectpolitics.com
team at BuzzFeed (26), and the third is a crowdsourced effort headed (9) conservativestate.com
by Melissa Zimdars of Merrimack College. Our key results are es- (10) conservativedailypost.com
sentially invariant to whatever measure of fake news we use. (11) libertywritersnews.com
(12) worldnewsdailyreport.com
Modeling strategy and covariates (13) endingthefed.com
We aggregated all shares to the individual respondent level so that (14) donaldtrumpnews.co.
our dependent variables are counts (i.e., number of fake news sto- (15) yesimright.com
ries shared). To account for this feature of the data, as well as the (16) burrardstreetjournal.com
highly skewed distribution of the counts, we primarily used Poisson (17) bizstandardnews.com
or quasi-Poisson regressions to model the determinants of Facebook (18) everynewshere.com
sharing behavior. We conducted dispersion tests on the count data (19) departed.co.
and used quasi-Poisson models if the null hypothesis of no dispersion (20) americanmilitarynews.com
is rejected. Below, we included negative binomial and Ordinary Least (21) tmzhiphop.com
Squares (OLS) regressions to show that our results are generally not
sensitive to model choice. All models applied weights from YouGov Sample details
to adjust for selection into the sample. We specifically used sample-­ Table 3 reports raw proportions of characteristics and self-reported

Downloaded from https://www.science.org on February 28, 2025

matching weights produced for the third wave of the survey, which behaviors across various sample definitions. Knowledge ranged from
was closest to the Facebook encouragement sent to respondents (27). 0 to 4 and was constructed from a grid of questions about the
(Results also do not appear to be sensitive to the use of weights.) majority party in the House and Senate, in addition to questions
We included a mix of relevant sociodemographic and political about whether the uninsured rate and earnings had increased over
variables as predictors. These include age (reference category, 18 to the course of 2016. Voter turnout was verified by our survey provider,
29), race, gender, family income, and educational attainment. In all which matches individual respondents to the TargetSmart voter file.
models, we included either five-point ideological self-placement, three- A potential concern about sample-selection bias is that those who
point party identification, or both. Since these variables were correlated consented to share Facebook data were different from the rest of the
(r = 0.31), we addressed possible multicollinearity via transparency— sample along some dimension that is also related to our outcome of
we provided our main results all three ways. (In all models, the ref- interest (fake news sharing behavior). Table 3 suggests that, at least
erence case for party identification and ideology is “Not sure.” Specifi- on closely related observable characteristics, the subgroup for which
cations including additional racial/ethnic categories are statistically we have profile data is a valid cross section of the overall sample. In
and substantively unchanged; available from the authors.) Last, we particular, frequent self-reported Facebook sharing activity is roughly
included a measure of the total number of wall posts including a URL. indistinguishable between those who report having a Facebook ac-
This is intended to capture the overall level of respondents’ Facebook count and those who provided access (P = 0.28). The samples are
link-sharing activity regardless of political content or verifiability. also comparable on age, frequency of looking at Facebook, and vote
intention. Those who shared data were slightly more liberal on av-
Details on main BuzzFeed-based list erage (P = 0.01), but we controlled for this in our models and we
Our BuzzFeed-based list began with 30 domains identified by that expected differences between the samples to arise due to chance
site’s reporting as purveyors of intentionally false election-related alone. Last, it may not be surprising that those who provided access
stories generating the most Facebook engagement. To do this, the to profile data were also more likely to participate in elections, as
journalists, led by Craig Silverman, used keywords and existing lists measured by verified voter turnout in the 2016 general election. We
combined with the analytics service BuzzSumo. To ensure that our see this somewhat heightened political engagement in the Facebook
analysis stayed clear of websites that could be construed as partisan subsample as important to note, and we accounted for the effects
or hyperpartisan (rather than intentionally or systematically factu- of this difference when we controlled for overall posting activity
ally inaccurate), we additionally filtered out domains identified as hard in our models.
news by a supervised learning classifier developed by Bakshy et al.
(23). The nearly 500 hard news domains encompass a wide range of
news and opinion websites, both mainstream and niche. The classi- SUPPLEMENTARY MATERIALS
Supplementary material for this article is available at http://advances.sciencemag.org/cgi/­
fier was trained on the text features of roughly 7 million web pages content/full/5/1/eaau4586/DC1
shared on Facebook over a 6-month period by U.S. users, with train- Tables S1–S13. Determinants of fake news sharing on Facebook (alternate specification).
ing labels for hard and soft news generated using bootstrapped key- Table S14. Determinants of hard news sharing on Facebook.
word searches on the URLs. Once matches to this list of hard news Fig. S1. Average number of fake news articles shared by age group (with 95% confidence inter-
vals), using the URL-level measure derived from (2).
domains were removed (for example, Breitbart.com), we were left
with 21 domains, shown below.
(1) usanewsflash.com REFERENCES AND NOTES
1. D. M. J. Lazer, M. A. Baum, Y. Benkler, A. J. Berinsky, K. M. Greenhill, F. Menczer,
(2) abcnews.com.co M. J. Metzger, B. Nyhan, G. Pennycook, D. Rothschild, M. Schudson, S. A. Sloman,
(3) denverguardian.com C. R. Sunstein, E. A. Thorson, D. J. Watts, J. L. Zittrain, The science of fake news. Science
(4) rickwells.us 359, 1094–1096 (2018).

Guess et al., Sci. Adv. 2019; 5 : eaau4586 9 January 2019 7 of 8

SCIENCE ADVANCES | RESEARCH ARTICLE

2. H. Allcott, M. Gentzkow, Social media and fake news in the 2016 election. 22. B. Swire, U. K. H. Ecker, S. Lewandowsky, The role of familiarity in correcting inaccurate
J. Econ. Perspect. 31, 211–236 (2017). information. J. Exp. Psychol. Learn. Mem. Cogn. 43, 1948–1961 (2017).
3. A. Guess, B. Nyhan, J. Reifler, “Selective exposure to misinformation: Evidence from the 23. E. Bakshy, S. Messing, L. A. Adamic, Exposure to ideologically diverse news and opinion
consumption of fake news during the 2016 U.S. presidential campaign” (2016); on facebook. Science 348, 1130–1132 (2015).
http://www-personal.umich.edu/~bnyhan/fake-news-2016.pdf. 24. G. King, N. Persily, “A new model for industry-academic partnerships” (2018);
4. G. Pennycook, T. D. Cannon, D. G. Rand, Prior exposure increases perceived accuracy of https://gking.harvard.edu/files/gking/files/partnerships.pdf.
fake news. J. Exp. Psychol. Gen., in press. 25. J. A. Tucker, A. Guess, P. Barberá, C. Vaccari, A. Siegel, S. Sanovich, D. Stukal, B. Nyhan,
5. E. Porter, T. J. Wood, D. Kirby, Sex trafficking, Russian infiltration, birth certificates, and Social Media, Political Polarization, and Political Disinformation: A Review of the Scientific
pedophilia: A survey experiment correcting fake news. J. Exp. Polit. Sci. 5, 159–164 Literature (William and Flora Hewlett Foundation, 2018).
(2018). 26. C. Silverman, Here are 50 of the biggest fake news hits on facebook from 2016 (BuzzFeed,
6. S. Vosoughi, D. Roy, S. Aral, The spread of true and false news online. Science 359, 2016); www.buzzfeed.com/craigsilverman/top-fake-news-of-2016.
1146–1151 (2018). 27. D. Rivers, Sample Matching: Representative Sampling from Internet Panels (Polimetrix
7. C. Silverman, “This analysis shows how fake election news stories outperformed real White Paper Series, 2006).
news on facebook” (BuzzFeed, 2016); www.buzzfeed.com/craigsilverman/viral-fake-
election-news-outperformed-real-news-on-facebook.
Acknowledgments: This study has been approved by the New York University Institutional
8. A. M. Guess, Measure for measure: An experimental test of online political media
Review Board (IRB-12-9058 and IRB-FY2017-150). We thank S. Luks at YouGov for facilitating
exposure. Polit. Anal. 23, 59–75 (2015).
survey data collection and J. Ronen for development and technical assistance. Funding: This
9. A. Guess, K. Munger, J. Nagler, J. Tucker, How accurate are survey responses on social
research was supported by the INSPIRE program of the NSF (Award SES-1248077). Author
media and politics? Polit. Commun. 1–18 (2018).
contributions: A.G., J.N., and J.T. designed the research. J.N. and J.T. oversaw data collection.
10. A. Campbell, P. E. Converse, W. E. Miller, The American Voter (Wiley, 1960).
A.G. analyzed the data and produced the figures. A.G. drafted the paper, and all authors
11. D. Green, B. Palmquist, E. Schickler, Partisan Hearts and Minds (Yale Univ. Press, 2002).
contributed to revisions. Competing interests: J.T. received a small fee from Facebook to
12. C. A. Klofstad, The sociological bases of political preferences and behavior, in The Future
compensate him for time spent in organizing a 1-day conference for approximately 30
of Political Science: 100 Perspectives (Routledge, 2009).
academic researchers and a dozen Facebook product managers and data scientists that was
13. B. Schäffer, The digital literacy of seniors. Res. Comp. Int. Educ. 2, 29–42 (2007).
held at NYU in the summer of 2017 to discuss research related to civic engagement. He did not
14. B. B. Neves, F. Amaro, Too old for technology? How the elderly of lisbon use and perceive

Downloaded from https://www.science.org on February 28, 2025

provide any consulting services or any advice to Facebook as part of this arrangement. He also
ICT. J. Community Inform. 8, 1–12 (2012).
participated in other unpaid meetings of academics researchers and Facebook product
15. S. Livingstone, Media literacy and the challenge of new information and communication
managers and data scientists to discuss topics related to social media and politics. All of these
technologies. Commun. Rev. 7, 3–14 (2004).
events occurred after the data collection for the current paper. All other authors declare that
16. E. Hargittai, Survey measures of web-oriented digital literacy. Soc. Sci. Comput. Rev. 23,
they have no competing interests. Data and materials availability: Counts and survey
371–379 (2005).
variables at the respondent level needed to replicate all tables and figures in the paper are
17. E. Hargittai, An update on survey measures of web-oriented digital literacy. Soc. Sci.
available at doi.org/10.7910/DVN/IKDTPZ. All data needed to evaluate the conclusions in the
Comput. Rev. 27, 130–137 (2009).
paper are present in the paper and/or the Supplementary Materials. Additional data related to
18. S. Bennett, K. Maton, L. Kervin, The ‘digital natives’ debate: A critical review of the
this paper may be requested from the authors.
evidence. Br. J. Educ. Technol. 39, 775–786 (2008).
19. S. Messing, S. J. Westwood, Selective exposure in the age of social media: Endorsements
trump partisan source affiliation when selecting news online. Commun. Res. 41, Submitted 12 June 2018
1042–1063 (2012). Accepted 30 November 2018
20. E. L. Glisky, S. R. Rubin, P. S. R. Davidson, Source memory in older adults: An Published 9 January 2019
encoding or retrieval problem? J. Exp. Psychol. Learn. Mem. Cogn. 27, 1131–1146 10.1126/sciadv.aau4586
(2001).
21. I. Skurnik, C. Yoon, D. C. Park, N. Schwarz, How warnings about false claims become Citation: A. Guess, J. Nagler, J. Tucker, Less than you think: Prevalence and predictors of fake
recommendations. J. Consum. Res. 31, 713–724 (2005). news dissemination on Facebook. Sci. Adv. 5, eaau4586 (2019).

Guess et al., Sci. Adv. 2019; 5 : eaau4586 9 January 2019 8 of 8