Social Media and Political Polarization

By Matthew Gentzkow

Posted on August 5, 2016


As the 2016 election moves closer, America seems ever more divided. Those pulling for single-payer national health insurance, free college for all, and higher top tax rates (all guided by the benevolent hand of President Bernie Sanders) might as well be on a different planet from those dreaming of Obamacare’s repeal, a Mexican border wall, flat income taxes, and the inauguration of President Donald J. Trump.


What brought us to this impasse? Rising inequality, bombings in Paris, racial divisions, Megyn Kelly... the list goes on. But what stands out for many people as an especially obvious culprit is the explosion of digital media, with its echo chambers and filter bubbles allowing partisans of each side to live in worlds where their prejudices, preconceptions, and worst fears are amplified and repeated back to them a hundred times a day.


While the digital-media-as-villain story can seem so transparently true that actually looking at the data to confirm it would be superfluous, academic research thus far actually provides surprisingly little support. As often happens, popular perception gets the broad outlines right but the magnitudes and proportions all wrong.


Yes, the Internet makes available a much wider range of viewpoints than we ever had before, including loud voices on both the left and right which are sometimes shockingly extreme. But extreme sites account for only a small share of consumption. Most people get most of their digital news from large, mainstream sources, and many of those that top the list on the left (CNN, USA Today, Yahoo, etc.) rank similarly high for consumers on the right. Moreover, as hard as it may be for those steeped in the tech world to believe, digital media as a whole still account for a relatively small share of total news consumption - 8 percent, as of 2013, according to a McKinsey report, compared to 41 percent for television and 35 percent for print newspapers.


In a recent blog post, I discuss evidence on trends in polarization and the role of digital media in more detail.


While I find this research largely convincing (full disclosure: I wrote some of it myself), it suffers from one glaring omission: the role of social media. Many of the key data come from almost ten years ago: an eternity given the current pace of change, and before the social revolution had fully taken hold. Only recently have new studies emerged that begin to fill in the gaps and shed light on whether the rise of social media has significantly reversed the earlier conclusions.


Here are a few of these recent data points.


The Facebook News Feed


Possibly the best known study of the way people consume news and opinion on social media is analysis of Facebook data by Eytan Bakshy, Solomon Messing, and Lada Adamic, published in Science in 2015. The authors tackle head-on the hypothesis advanced by Eli Pariser and others that algorithmic filtering may place people in ideological “filter bubbles.”


The authors begin with data on 10.1 million US Facebook users who declare their political orientation to be “conservative,” “moderate,” or “liberal.”


They first show that, as expected, peoples’ friends tend to share their political views, though to a smaller extent than some might have supposed.


For both conservatives and liberals, roughly two thirds of friends (among those who declare some orientation) have the same political views, with the remainder evenly split among moderates and those with opposite views.


Next, the authors look at the overall distribution of news content shared within these networks between July, 2014 and January, 2015. They use a machine learning algorithm to zero in on “hard news” content, then assign each hard news article an ideological score based on the average orientation of users that shared it. The set of all such articles shared shows clear clusters of conservative and liberal articles, with those from sites such as Fox News consistently categorized as conservative, and those from sites such as Huffington Post consistently categorized as liberal. Overall, 45 percent of content is conservative, 40 percent is liberal, and the remainder is classified as “neutral.”


Finally, the authors measure the extent to which various stages of social media diffusion – what articles get shared within a friendship network, which of these the Facebook algorithm chooses for the news feed, and which of these the user ultimately chooses to click on – filter out ideologically cross-cutting content. Each of these stages indeed induces some filtering, but the magnitude – particularly the magnitude of filtering by the news feed algorithm – is quite small, at least relative to some visions of what could be happening (such as the one painted in Eli Pariser’s widely viewed TED talk).


Begin with conservatives. With no filtering at all – that is, if they read randomly selected articles from the universe of those shared by all Facebook users – 40 percent of what they see would be ideologically cross-cutting. Choosing randomly from what is shared would reduce this to about 35 percent. Restricting to what is actually shown in the news feed (the key step according to the Pariser argument) has a tiny effect, reducing the number by maybe a percentage point. And restricting to what the user actually clicks brings the number down to 29 percent. Not a nirvana of ideological open mindedness, but not exactly a dystopian filter bubble either.


The picture is similar for liberals, with the difference that liberals are actually less likely than conservatives to share cross-cutting content in the first place. Random selection would mean liberals see 45 percent cross-cutting stories. The friendship network cuts this down sharply to about 24 percent. The news feed again reduces it by maybe a percentage point. And selectivity in clicks brings it down to 20 percent.


Your Friends Are Not What You Think They Are


An earlier study by Sharad Goel, Winter Mason, and Duncan J. Watts in the 2010 Journal of Personality and Social Psychology offers one explanation for why we might imagine the ideological bubbles we and others live in to be more restricted than they really are.


The authors surveyed 2,504 Facebook users in early 2008, asking their views on a wide range of political and social issues, as well as their overall political orientation. Respondents were also asked to predict their friends’ responses to the same questions. The sample was designed so that, in many cases, when A was asked to predict the responses of her friend B, B was a respondent to the survey as well.


Not surprisingly, the results showed that friends tend to think alike: the probability of agreeing with a friend on a random question was 75 percent, significantly higher than the 63 percent we would see if friends were matched at random.


The striking finding, however, is that people think their friends agree with them even more than they really do: the rate at which respondents predict their friends agree with them is 80 percent.


This result echoes a large body of literature in psychology suggesting that people may project their own views onto others, and perceive higher levels of consensus than actually exists. That we imagine our friends all agree with us is not necessarily great news for the health of our democracy, but it does mean that the particular danger of like-minded echo chambers could appear bigger than it is.


Political Junkies on Twitter


Of course there is more to social media than just Facebook. In a recent working paper, Yosh Halberstam and Brian Knight study the patterns of information diffusion and follower links on Twitter.


Although data on Twitter posts and links are easily available, direct measures of Twitter users’ political views are not. A key innovation of this study is to develop a proxy: the party of the political candidates a user follows. Of course most users do not follow any political candidates, and so restricting attention to these users necessarily isolates those who are most politically partisan and engaged. Though it is far from a representative sample, this is a population of particular interest for understanding larger trends in political polarization.


The authors focus on a sample of 2.2 million users who follow at least one US House candidate in the 2012 election. They define users to be liberal if they follow only or mainly Democratic candidates, and conservative if the follow only or mainly Republican candidates. They measure all links among the users (defined as one following another), as well as a large sample of retweets of both posts by the House candidates and posts mentioning the candidates.


As we would expect, the authors find that users are significantly more likely to follow and engage with those of the same political orientation. The overall degree of ideological segregation in the follower network is similar to what my co-authors and I found in an earlier study for offline social networks, and significantly higher than what we found for news and opinion consumption online.


Partly as a result, users are much more likely to be exposed to like-minded retweets. 90 percent of candidate retweets liberals are exposed to originate with Democratic candidates, and 90 percent of retweets conservatives are exposed to originate with Republican candidates. Of course this may be less surprising since we have defined ideologies based on the kinds of candidates the users follow. Exposure to tweets mentioning candidate names is a bit more balanced, though still ideologically selected.


Putting Social Media in Context


A final recent data point comes from work conducted at Microsoft Research by Seth Flaxman, Sharad Goel, and Justin Rao. They study the browsing behavior of a sample of 1.3 million Internet Explorer (IE) users in March-May 2013, focusing on consumption of news and opinion articles. Their main question is how consumption via social media differs from consumption through other channels such as direct browsing, search, or news aggregators, and how social media affects overall patterns of ideological segregation.


For their main analysis, they focus on a small subset of users (roughly 4 percent) who read news and opinion regularly, and they limit attention to the 100 most visited sites.


They reach three striking conclusions. First, consistent with fears that social media are exacerbating polarization, they find that opinion content accessed via social media is indeed substantially more segregated ideologically – that is, more likely to be either consistently conservative or consistently liberal – than opinion content accessed via other channels. Editorials, op-eds, and other opinion pieces are popular fodder for Facebook feeds, and as Bakshy et al.’s (2015) study would suggest, it looks like people are much more likely to see and click on content that shares a consistent ideological profile (presumably one that matches the user’s own).


Second, while the opinion content people see through social media is on the whole less diverse, it actually includes more content from opposite extremes of the political spectrum than what they see through other channels. People may mostly see stories shared by like-minded friends, but they also occasionally bump into very different points of view. This is less likely to happen when people are navigating to news sites directly.


Finally, the net effect on peoples’ news and opinion diets is ultimately quite small. Opinion content accounts for only a sliver (6 percent) of total consumption and the ideological segregation we see in socially accessed opinion content does not hold for regular news. Moreover, the share of news and opinion that people reach through social media is actually very small: only about 6 percent of news and 10 percent of opinion. We may eventually reach a point where social sharing is a dominant mode of accessing news and information, but we are not there yet.


One way to understand this is to remember that sharing substantive news articles is just not the main thing people share on social media. Cat videos and embarrassing celebrity photos are of course far more popular. Only 1 out of every 300 outbound clicks from Facebook is to what Flaxman and co-authors identify as a hard news article.


This study is an important and exciting contribution, providing the first detailed look at the importance of social media in an overall news consumption landscape. It requires one important caveat however: The data come from a very unusual segment of the population: heavy news users who happen to have installed the IE toolbar. The later criterion may be especially important: by 2013, use of IE had declined significantly from its peak, and presumably only a subset of those who did use it installed the toolbar. To the extent that these users tended to be older or less inclined to adopt new technologies, the findings might understate the importance of social media for the US population as a whole.




Will social media deepen our divisions and lead us ever deeper into ideological echo chambers? The truth is, we don’t yet know. The studies above only scratch the surface of what is currently happening, and they of course cannot predict how the situation will change in the future.


To the extent we can tell, the data suggest that the degree of ideological segregation in digital media remains substantially lower than much of the popular discussion would suggest. There is no question that Facebook feeds and Twitter networks expose users to less ideologically cross-cutting content than they would see if they randomly sampled from what is available. This is true for all the reasons we would expect: people connect with those more likely to share their views, these users mainly share content they agree with, algorithmic selection like Facebook’s news feed may enhance the selection (though the data suggest only slightly), and what users actually choose to read is likely to tilt even further toward their own views.


But random selection is not the right benchmark. Before social media, people got news from direct navigation to websites or search, from content shared through email, from traditional media, and by actually talking to their friends and acquaintances. Selective exposure is strong in all of these channels. The evidence we have suggests it may be stronger in social media than in some alternatives (like directly access news) and weaker than in others (face-to-face relationships). Either way, the magnitude of the differences may be smaller than sometimes supposed.


More importantly, content mediated through social media probably remains a small part of most users’ news diets. Traditional media are still the most important by far, and within the digital realm direct navigation still swamps social. One day we may get all of our news through Facebook (along with, its shareholders presumably hope, our advertising, online purchases, movies, video games, and college courses). If we do, the repercussions for our democracy could be profound.


Remember, though, that American politics offers plenty of pressing problems that exist right now in 2016. Thankfully, it would seem a dramatic increase in polarization driven by social media can be removed from the list.



The preceding is republished on TAP with permission by its author, Professor Matthew Gentzkow, and by the Toulouse Network for Information Technology (TNIT). “Social Media and Political Polarization” was originally published in TNIT’s August 2016 newsletter.