Selective Exposure to Misinformation:Evidence from the consumption of fake news during the 2016 U.S.presidential campaignAndrew GuessDepartment of PoliticsPrinceton UniversityBrendan NyhanDepartment of GovernmentDartmouth CollegeJason ReiflerDepartment of PoliticsUniversity of ExeterDecember 20, 2017AbstractThough some warnings about online “echo chambers” have been hyperbolic, tendencies towardselective exposure to politically congenial content are likely to extend to misinformation and tobe exacerbated by social media platforms. We test this prediction using data on the factuallydubious articles known as “fake news.” Using unique data combining survey responses withindividual-level web tra�c histories, we estimate that approximately 1 in 4 Americans visited afake news website from October 7-November 14, 2016. Trump supporters visited the most fakenews websites, which were overwhelmingly pro-Trump. However, fake news consumption washeavily concentrated among a small group — almost 6 in 10 visits to fake news websites camefrom the 10% of people with the most conservative online information diets. We also find thatFacebook was a key vector of exposure to fake news and that fact-checks of fake news almostnever reached its consumers.
The combination of rising partisanship and pervasive social media usage in the United States havecreated fears of widespread “echo chambers” and “filter bubbles” (Sunstein, 2001; Pariser, 2011).To date, these warnings appear to be overstated. Behavioral data indicates that only a subset ofAmericans have heavily skewed media consumption patterns (Gentzkow and Shapiro, 2011; Barber´aet al., 2015; Flaxman, Goel, and Rao, 2016; Guess, 2016).However, the risk of information polarization remains. Research shows people tend to prefercongenial information, including political news, when given the choice (e.g., Stroud, 2008; Hartet al., 2009; Iyengar and Hahn, 2009; Iyengar et al., 2008), but these studies typically focus on howideological slant a↵ects the content people choose to consume; relatively little is known about howselective exposure extends to false or misleading factual claims. Research in political science andpsychology has documented that misperceptions are often systematically related to people’s politicalidentities and predispositions (Flynn, Nyhan, and Reifler, 2017). In this article, we thereforeevaluate whether people di↵erentially consume false information that reinforces their political viewsas theories of selective exposure would predict.We additionally consider the extent to which social media usage exacerbates tendencies towardselective exposure to misinformation. Though Messing and Westwood (2014) find that social en-dorsements can help overcome partisan cues when people are choosing news content, other researchindicates that tendencies toward selective exposure to attitude-consistent news and informationmay be exacerbated by the process of sharing and consuming content online (e.g., Bakshy, Mess-ing, and Adamic, 2015). In this way, social media consumption may also be a mechanism increasingdi↵erential exposure to factually dubious but attitude-consistent information.Finally, we analyze whether fact-checking — a new format that is increasingly used to counterpolitical misinformation — e↵ectively reached consumers of fake news during the 2016 election.Though fact-checks are relatively widely read and associated with greater political knowledge (e.g.,Gottfried et al., 2013), they are often disseminated online in a politically slanted manner that islikely to increase selective exposure and reduce consumption of counter-attitudinal fact-checks (Shinand Thorson, 2017). To date, however, no previous research has considered whether consumers offact-checks have been exposed to the claims that they evaluate. Does selective exposure underminethe e↵ectiveness of fact-checking?We evaluate these questions in the context of the rise of so-called “fake news,” a new form1
of political misinformation that features prominently in journalistic accounts of the 2016 U.S.presidential election (e.g., Solon, 2016). Data from Facebook indicates that these factually dubiousfor-profit articles were shared by millions of people (Silverman, 2016). Many people also reportbelieving the claims that fake news sites promoted in post-election surveys (Silverman and Singer-Vine, 2016; Allcott and Gentzkow, 2017).However, little is known scientifically about theco n s u m p t i o nof fake news, including who readit, the mechanisms by which it was disseminated, and the extent to which fact-checks reached fakenews consumers. These questions are critical to understanding how selective exposure can distortthefactualinformation that people consume — a key question for U.S. democracy.We therefore examine the prevalence and mechanisms of exposure to fake news websites in aunique dataset that combines pre-election survey responses and comprehensive web tra�c data froma national sample of Americans. Our design allows us to provide the first individual-level estimatesof visits to fake news websites, including who visited these websites, how much and which types offake news they consumed, and the probability that fact-checks reached fake news website readers.We can thus provide the first measures of the prevalence of selective exposure to misinformation inreal-world behavior.Specifically, we find that approximately one in four Americans visited a fake news website, butthat consumption was disproportionately observed among Trump supporters for whom its largelypro-Trump content was attitude-consistent. However, this pattern of selective exposure was heavilyconcentrated among a small subset of people — almost six in ten visits to fake news websites camefrom the 10% of Americans with the most conservative information diets. Finally, we specificallyidentify Facebook as the most important mechanism facilitating the spread of fake news and showthat fact-checking largely failed to e↵ectively reach consumers of fake news.Taken together, these results suggest a need to revisit the study of selective exposure usingmeasures of real-world media consumption and to consider the behavioral mechanisms by whichpeople are exposed to misinformation.2
Data and resultsData for the analyses below combine responses to an online public opinion survey from a nationalsample of 2,525 Americans with web tra�c data collected passively from their computers withtheir consent during the October 7–November 14, 2016 period. Our primary outcome variables arecomputed from web tra�c data and measure the type and/or quantity of websites publishing fakenews that respondents visited. We employ survey weights to approximate the adult population ofthe U.S. (Further details on the sample and the survey weights are provided in the Supplemen-tary Materials, where we show that the sample closely resembles the U.S. population in both itsdemographic characteristics and privacy attitudes.)The survey questions we administered allow us to examine the relationship between demographicand attitudinal variables (e.g., candidate preference) and visits to fake news websites. Addition-ally, we compute three key explanatory measures from respondents’ web tra�c data: the overallideological slant of a person’s online media consumption (or “information diet”), which we dividebelow into deciles from most liberal to most conservative using the method from Guess (2016); theirconsumption of “hard news” sites classified as focusing on national news, politics, or world a↵airs(Bakshy, Messing, and Adamic, 2015); and their Facebook usage, which we divide into terciles byhow often they visit the site.Of course, studying fake news consumption requires defining which websites are publishing fakenews. We define pro-Trump fake news websites as those that published two or more articles thatwere coded as fake news in Allcott and Gentzkow (2017), the first peer-reviewed study of fake newsin social science, and for which 80% or more of the fake news articles identified from the site werecoded as pro-Trump.1An identical approach is used to create our measure of pro-Clinton fake newssites. We exclude domains from these sets that were previously identified in Bakshy, Messing, andAdamic (2015) as focusing on hard news topics in order to concentrate on the new websites thatwere created around the election. Finally, we construct a measure of total fake news website visitsthat includes visits to both pro-Trump and pro-Clinton fake news websites as defined above.21In the Supplementary Materials, we present robustness tests using two alternate outcome measures. The resultsare highly consistent with those presented below.2Our measures of fake news consumption thus exclude more established but often factually dubious sites such asBreitbart. Due to restrictions in the Facebook API, we also cannot observe incidental exposure to fake news or otherkinds of dubious content such as “hyper-partisan” sites in the Facebook News Feed. In this sense, our estimatesrepresent a lower bound of fake news consumption.3
The fake news sites in question, which are listed in the Supplementary Materials, display littleregard for journalistic norms or practices; reporting suggests most were created to generate profits(Silverman and Alexander, 2017). Though they sometimes publish accurate information, theyalso frequently publish false claims, distort genuine news reports, and copy or repurpose contentfrom other outlets. It is important to note, however, that there is still considerable diversity inthe stories that these sites publish. Some content is deeply misleading or fabricated (e.g., the“Pizzagate” conspiracy theory), while other articles instead selectively amplify political events inan over-the-top style that flatters the prejudices of a candidate’s supporters.Total fake news consumptionWe estimate that 27.4% of Americans age 18 or older visited an article on a pro-Trump or pro-Clinton fake news website during our study period, which covered the final weeks of the 2016election campaign (95% CI: 24.4%–30.3%). While this proportion may appear small, 27% of thevoting age population in the United States is more than 65 million people. In total, articles on pro-Trump or pro-Clinton fake news websites represented an average of approximately 2.6% of all thearticles Americans read on sites focusing on hard news topics during this period. The pro-Trumpor pro-Clinton fake news that people read was heavily skewed toward Donald Trump — peoplesaw an average (mean) of 5.45 articles from fake news websites during the study period of October7–November 14, 2007. Nearly all of these were pro-Trump (average of 5.00 pro-Trump articles).Selective exposure to fake newsThere are stark di↵erences by candidate support in the frequency and slant of fake news websitevisits.3We focus specifically in this study on respondents who reported supporting Hillary Clintonor Donald Trump in our survey (76% of our sample) because of our focus on selective exposure bycandidate preference. People who supported Trump were far more likely to visit fake news websites— especially those that are pro-Trump — than Clinton supporters. Among Trump supporters,40% read at least one article from a pro-Trump fake news website (mean = 13.1, 95% CI: 7.8, 18.3)3Our analysis considers visits to fake news websites as defined above but we show in the Supplementary Materialsthat the results in Table 1 (below) are consistent if we instead only consider visits to specific article URLs thatAllcott and Gentzkow (2017) identify as being classified as false or misleading by fact-checkers. The results are alsoconsistent if we consider visits to websites identified by Silverman (2016) as publishing the most widely shared fakenews articles before the 2016 election (see the Supplementary Materials).4
compared with only 15% of Clinton supporters (mean = 0.51, 95% CI: 0.39, 0.64). Consumptionof articles from pro-Clinton fake news websites was much lower, though also somewhat divided bycandidate support. Clinton supporters were modestly more likely to have visited pro-Clinton fakenews websites (11.3%, mean articles: 0.85) versus Trump supporters (2.8%, mean articles: 0.05).The di↵erences by candidate preference that we observe in fake news website visits are even morepronounced when expressed in terms of the composition of the overall news diets of each group.Articles on fake news websites represented an average of 6.2% of the pages visited on sites thatfocused on news topics among Trump supporters versus 0.8% among Clinton supporters.The di↵erences we observe in visits to pro-Trump and pro-Clinton fake news websites by candi-date support are statistically significant in OLS models even after we include standard demographicand political covariates, including a standard scale measuring general political knowledge (Table1).4Trump supporters were disproportionately more likely to consume pro-Trump fake news andless likely to consume pro-Clinton fake news relative to Clinton supporters, supporting a selectiveexposure account. Older Americans (age 60 and older) were also much more likely to visit fakenews conditional on these covariates, including pro-Trump fake news.We also find evidence of selective exposurewithinfake news; pro-Trump voters di↵erentiallyvisited pro-Trump fake news websites compared with pro-Clinton websites. To help demonstratethis, we employ a randomization inference-style approach in which we randomly permute the coding(pro-Trump or pro-Clinton) of visits to fake news websites by Trump supporters in our panel. Totalconsumption of articles from pro-Trump fake news websites is as frequent as we observe or greaterin 4 of 1,000 simulations (p=0.004 one-sided; see the Supplementary Materials). We thus rejectthe null hypothesis that Trump supporters are no more likely to visit pro-Trump fake news contentthan pro-Clinton fake news content.Finally, we show that individuals who engage in high levels of selective exposure to online newsin general are also di↵erentially likely to visit fake news websites favoring their preferred candidate.In general, fake news consumption seems to be a complement to, rather than a substitute for,hard news — visits to fake news websites are highest among people who consume the most hardnews and do not measurably decrease among the most politically knowledgeable individuals. (See4We use OLS models due to their simplicity, ease of interpretation, and robustness to misspecification (Angristand Pischke, 2009), but we demonstrate in the Supplementary Materials that these conclusions are consistent if weinstead use probit and negative binomial regression models.5
Table 1: Who chooses to visit fake news websites (behavioral data)Pro-Trump fake Pro-Clinton fakenews consumption news consumptionBinary Count Binary CountTrump supporter 0.220** 13.121** -0.113** -1.100**(0.033) (3.576) (0.019) (0.181)Political knowledge 0.019* 1.013 0.003 -0.003(0.008) (0.609) (0.004) (0.039)Political interest 0.044* 1.744 0.027 0.378**(0.021) (1.028) (0.015) (0.117)College graduate -0.010 -2.655 0.015 -0.109(0.030) (1.771) (0.019) (0.157)Fe m a l e 0 . 0 4 7 4 . 5 6 5 0 . 0 2 1 0 . 2 0 0(0.028) (2.921) (0.020) (0.146)Nonwhite -0.057 5.519 -0.054* -0.633**(0.035) (4.876) (0.024) (0.180)Age 30–44 -0.038 -0.040 0.053* 0.369*(0.055) (1.306) (0.023) (0.167)Age 45–59 0.031 1.215 0.077** 0.801**(0.059) (1.472) (0.023) (0.225)Age 60+ 0.084 7.221* 0.107** 0.635**(0.056) (2.924) (0.024) (0.139)Constant -0.110 -16.568* -0.049 -0.692*(0.081) (7.558) (0.046) (0.302)R20.13 0.05 0.07 0.03N2167216721672167*p