November 2016


Moving forward: RQ2. How do we know what Facebook affects?

“Trump didn’t win because of X” has become a popular genre of punditry in the last two and a half weeks, along with any number of declarations backed up by little or no specific data.[1] In the wake of concern about “fake news” and partisan echo chambers online, fed by both Russian intelligence and American hoaxsters, Facebook (and social media more broadly) has become the focal point. Keith Hampton and Eszter Hargittai make this point, but like most such analyses, don’t have data specific to actual voters; instead, they note that the demographics of Trump support are negatively correlated with social media use, and that most people don’t click through from headlines in their Facebook feeds.

But this sort of supposition ignores a range of ways that we know information filters through even pre-Internet social networks, let alone the supercharged networking that is the core function of Facebook. The point here is not to say that Facebook did or didn’t do anything, but that stitching together population-level generalities is not going to provide anything like compelling evidence.

So how do we figure out what Facebook affected, if anything, and how it did it? It’s important to have some handle on what we mean here, because no what matter we do there are going to be lots of variables tangled up in a mess of colinearity. We also need to note that getting a look at actual Facebook content is difficult to impossible, but the online environment presents a lot of problems along these lines. Survey respondents might be able to recall how frequently they visited a major source; can they recall whether or not they ever read something from one of the minor partisan sources that use Facebook as their primary distribution platform?

If actual content is out, we’re going to need to contextualize Facebook use. One way to do this is at the model level, putting Facebook use for news into an mediation model with other media use, and online and offline political discussion. Some co-authors and I have a paper in development that takes one approach to this, essentially wrapping an online version of the communication mediation model in a Facebook-based container. We find no direct effects of Facebook news use on any outcomes outside of Facebook, but significant indirect effects running through links to other media and discussion behaviors. This sort of thinking also suggests examining the relationship of Facebook shares to prominence in other media, and especially major partisan media. Facebook may act as a conduit for stories that bubble up from 4chan, Reddit, or Twitter to make their way to Fox News and conservative talk radio, for example.

Understanding potential Facebook effects at the individual level requires understanding individuals within their network contexts, as both senders and receivers of information. This helps us get at the central complicating factor of measuring Facebook’s effects, which is that everyone’s Facebook experience is different. Unlike a measurement of how often one watches network news broadcasts, for example, just asking for Facebook use frequency tells us basically nothing. However, what if we also knew something about people’s networks? In a survey this would be imperfect self-reported data, but we could ask questions about political homogeneity of one’s network, along with things like tendency to engage with agreeing or disagreeing others. An interaction term between frequency of Facebook use for news and network homogeneity would give us a measure of Facebook as a filter bubble or echo chamber; putting that in a model with reflection, elaboration, and talk would start us toward a model of how a variety of influences affect individuals’ attitudes. I have another paper in progress that utilizes an interaction term like this, and one problem with it is that it’s basically an impossible measure to validate. But that’s a problem for another day!


[1] This is especially weird given the ultimate closeness of the election. Anything that could have cost Clinton 100,000 total votes across Pennsylvania, Michigan, and Wisconsin could be said to be the reason Trump won. The existence of multiple “but for” causes doesn’t make any single one invalid.

Filed: Super Special Questions || 13:32, November 26 || No Comments »


Moving forward: RQ1. Who were the mediated voters?

I love blogging, but it’s obviously not the right medium for me. I’m going to try to rectify that with this series of short posts using the chaos that currently ensnares us to develop some research questions for 2017 and beyond. Some of them are strongly journalism-focused, others about campaign organization, some about information systems. I’m going to dig at the necessary research designs a bit, but I’m not thinking too much here about doability; this is more about what we should be figuring out.

The first one came to mind reading this New York Times piece on voters and non-voters, and primarily black ones, in Milwaukee County:

“We went to the beach,” said Maanaan Sabir, 38, owner of the Juice Kitchen, a brightly painted shop a few blocks down West North Avenue, using a metaphor to describe the emotion after Mr. Obama’s election. “And then eight years happened.”

All four barbers had voted for Mr. Obama. But only two could muster the enthusiasm to vote this time. And even then, it was a sort of protest. One wrote in Mrs. Clinton’s Democratic opponent, Senator Bernie Sanders of Vermont. The other wrote in himself.

This sort of voter profile piece is a staple of post-election reporting, particularly when a candidate under or overperforms in an unexpected way, and is presented as a way to understand the broader scope of what happened in the election. At the same time, this piece managed to find just one reluctant Clinton voter in a city that cast 76% of its votes for her (“…as did many others here” is how the story puts it). This genre existed during the campaign as well, as noted frequently by Eric Boehlert of Media Matters for America:

In general, I understand the media’s desire to try to explain what’s driving the support for Trump, who’s obviously running a highly unusual campaign and marketing his run in openly bigoted language. For a lot of people that’s deeply troubling, so understanding the dynamic behind Trump represents an obvious story of interest.

What I’m baffled by is the media’s corresponding lack of curiosity about examining Clinton voters. After all, she has accumulated more votes than any other candidate this year and is leading a Democratic surge into key states. (Why hasn’t The New Yorker published an 8,000-word piece on why Virginia has turned into a deeply blue state over the last decade?)

And I’m not alone in noting the year’s long-running disparity. Journalism professor and Clinton supporter Jeff Jarvis recently admonished the media (emphasis in original): “I never hear from voters like me who are enthusiastic supporters. I never see reporters wading among eager backers at Clinton rallies to ask them how much they like her and why.”

So the question here is a simple one: Do the voters and areas presented in mainstream press profiles represent the actual electorate that votes in the election? If not, in what ways is the presentation biased? Some of these are fairly obvious — the view of Trump as a weird insurgent, at best, or a danger to the republic, at worst, make for a Man Bites Dog story regardless of what else is going on. However, “Former First Lady becomes first ever woman to win major party nomination” is also an unusual story. Projected swing states seem likely to have gotten more attention, but the Democratic movement of Arizona and Georgia is also compelling.

This sort of study would’ve been much easier to do 20 years ago. Identifying what qualifies as the national political press in 2016 is a study on its own, and then figuring out how to find all the relevant profile pieces from, for example, CNN.com is another extensive piece of work. The difficulty of systematic sampling and the breadth of how profiles are presented suggests a qualitative approach may be the most sensible, but any comparison with real election results will need more precision than that. Assuming we find a way through the sampling process, the work is a little easier. Coding for candidate support, enthusiasm, location, history, available demographic information, and anything else that helps form a picture of the voters being profiled can be aggregated up to a model of what the mediated electorate looks like. It’s very possible there aren’t enough data points available to do a true statistical analysis, but I think that capturing the picture in the coverage is really the goal of a study like this.

Filed: Super Special Questions || 7:39, November 23 || No Comments »