Citizenship,Social Media,and Big Data

Benefits and Lingering Problems

New methods of inquiry have long been associated with major shifts in the way science is practiced. Telescopes, for example, eventually lead to a reconstruction of how humans think about their place in the universe (Kuhn, 1957). As Kuhn (2012) notes, paradigm shifts do not take place over night or through the work of one person. Instead, new methods of inquiry initially open the door to anomalies. The extent to which the old theories and methods are incapable of explaining new observations determines how, and when, new paradigms arise. Big data sets alone are not enough to represent such a shift in the area of political communication. Sometimes, the term big data is reserved for an area of research dominated by private companies for the purpose of selling products to consumers (McAfee, Brynjolfsson, Davenport, Patil, & Barton, 2012). One myth in this regard is that a scientific method, driven by causal inference and theoretical questions, is no longer essential. Instead, all that is necessary is to data mine relationships (Anderson, 2008; Jungherr, 2015; Mayer-Schönberger & Cukier, 2013). On the contrary, in order to better understand the fluid, high-speed world of social media and politics, we still rely on basic principles of social science: validity of constructs, theory that stresses context, generalizability, and ethical concerns.

Thus, it is important to remember what political communication research with big data looks like. Drawing on Shah, Cappella, and Neuman (2015), articles in this issue display the main features of what is referred to as “computational social science,” defined by:

(1) the use of large, complex datasets …; (2) the frequent involvement of “naturally occurring” social and digital media sources and other electronic databases; (3) the use of computational or algorithmic solutions to generate patterns and inferences from these data; and (4) the applicability to social theory in a variety of domains from the study of mass opinion to public health, from examinations of political events to social movements (p. 7).

Other questions remain for how to visualize and update results. For example, one study might rely on a visualization (or collection) tool that, due to outside reasons, may no longer be available or no longer valid. In other words, as the technology changes so might decisions about how to collect, visualize, and interpret data, which ultimately affects the means of replicating any given study. And by extension, it may also affect the means of moving social scientific disciplines forward, as it is harder to “stand on the shoulders” of previous research. Thus, the actual practice of big data research in social science is a moving target. Today’s representation is different than tomorrows. In short, the findings are also dynamic and in flux. I argue that the discipline needs to be more flexible here. Perhaps one solution is to be more willing to share tools and methods in a repository as a part of the online version of the journal. This way, it may be easier to update findings as technology changes. Unfortunately, this may be impossible, as some scholars have stressed that cost and access to big data tools prevent replication and correction (boyd & Crawford, 2012).

Large sample sizes also raise questions about the generalizability of a particular sample. If the sample size is inclusive of all users on a particular platform (Backstrom, Boldi, Rosa, Ugander, & Vigna, 2012), inferential statistics are not needed, since the entire sample is present. However, large sample sizes of this nature also render hypothesis testing pointless, since any relationship will appear statistically significant (Kramer, Guillory, & Hancock, 2014). On the other hand, even if we get a large sample, more care needs to be taken to compare how any population of social media users (for example Twitter users) might represent the general population. Vargo and Hopp (in this issue) address this concern by purposeful sampling of Twitter users based on data from official population estimates. This type of solution forces researchers to compare data points across multiple data sets.

Often, data sets are so large that manual coding is impossible. Machine learning algorithms are used to do the coding. In these cases, it is important that researchers understand, and at least make an effort to explain, how a particular algorithm performs on a particular data set. As Hindman (2015) argues, all algorithms may not perform equally. Understanding the features of your data goes a long way to determining the limits and benefits of the particular tool being used. Machine learning can also be applied to smaller data sets (say a subset of the original sample) to test for model fit and offer a check against results gleamed from the whole set.

As with machine learning, certain algorithms may also replace normal methods of data analysis (e.g., Tibshirani, 1996). Instead of a priori hypothesis testing, these tools allow researchers to data mine sources, by finding the most powerful predictors in any given model. In the past, social scientists posed questions, collected data, and tested a set of hypothesis. Big data tools offer researchers a chance to rethink these practices, particularly with regard to over emphasis on p value testing (Hindman, 2015; Ioannidis, Munafo, Fusar-Poli, Nosek, & David, 2014). However, big data do not relieve the social scientist of their responsibility to ensure the validity of their methods and place findings in an interpretable framework. Many of the articles in this volume address these concerns by combining traditional social science methods with big data. Others make sense of data through the lenses of well-established theories of political reasoning and behavior.

Finally, a lingering concern comes in the form of ethical issues. Most guidelines of ethical research are based on the Belmont Report, a guide to ethical treatment of human subjects in medical practice, released in 1978 (Cassell, 2000). The report outlined basic practices of benevolence, respect for persons, and justice. Issues of fair treatment of research subjects in big data research are not easily resolved based on these premises. In general, the field is in desperate need of revised ethical guidelines, particularly in regard to privacy, informational harm, and fair use of publically accessible information (boyd & Crawford, 2012; Fairfield & Shtein, 2014). For example, how do we go about attaining informed consent from millions of Twitter users? What if the data were not initially intended for research purposes? Further, the more an individual leaves behavioral traces in the digital world, the more susceptible they are to informational harm—where any breach of data could result in leaks of potentially sensitive personal information (Fairfield & Shtein, 2014). Researchers should at least consider potential ethical concerns, and when possible, report how they handle such concerns.

Articles in This Volume

Selections in this volume approach the analysis of political digital trace data from a variety of perspectives. The following brief summaries highlight how the authors each tackle some of the limitations of big data research.

Alejandro Ecker (this issue) examines the potential of Twitter data to estimate policy positions at the individual and party level in the Polish Sejm—Poland’s lower house of parliament. By comparing results with outside data sources, Ecker is able cross-validate results against “real-world” political outcomes. At the aggregate level of party politics, policy positions seem to be well represented on Twitter. However, at the level of the individual legislator, political language used on Twitter does not reliably predict how a particular legislator will vote. Results indicate that cross-validation helps avoid some of generalizability problems inherent in using Twitter data. Such efforts help establish the long path to establishing context-specific criteria for when Twitter feed might represent actual policy preferences and when they do not.

Based on the creation of an original Facebook page app, Chris Wells and Kjersten Thorson (this issue) introduce a creative means of capturing the actual flow of information an individual is exposed to on their Facebook news feed. Social media sites allow users to contour their news feed to their preferences. While many have argued such activity would lead to news avoidance, selective exposure, or even a decrease in inadvertent news exposure (Prior, 2007; Sunstein, 2009), Wells and Thorson are able to empirically asses these curated information flows. The authors ask respondents to agree to install an app on users Facebook page. When they agree, the app collects data about what a respondent follows, how many friends they have, and what content was in their newsfeed. As the app collects these data, respondents are surveyed about their political behaviors and levels of public affairs knowledge. After the survey is complete, the app is removed from the Facebook account. Big data sets often ignore individual behaviors, while surveys often rely on self-reported measures. Both approaches are limited in their ability to capture the news preferences of individuals on Facebook. By combining methods, the authors are able to verify survey results and rely on a more nuanced analysis of an individual’s information environment.

Qinfeng Zhu (this issue) uses Twitter data to track global conversations around the Hong Kong Occupy Central Movement. Using systems theory, Zhu challenges the assumptions that social network spaces are dominated by citizen to citizen, decentralized network activity. Instead, Zhu employs network analysis to argue that conversations about the occupy movement were dominated by elite institutions. In order to better contextualize these findings, the author also used multivariate analysis to account for a region’s income and level of political grievances. This additional layer of validation changes the interpretation of the results. Although large institutions dominate the global conversation, a region’s level of political grievance was more powerful predictor of network centrality than income.

Kwon and Cho (this issue) question long-held assumptions about civility in public discourse. Some scholars have noted how acidic online conversation spaces can be (Coe, Kenski, & Rains, 2014). However, the authors argue that some swearing is acceptable in political discussion because it raises emotional engagement and draws readers’ attention to the conversation. By measuring readers’ response to swearing, they are able to challenge the notion that civility is necessary for engagement. Swearing, they argue, might also re-enforce like-minded groups. Although the authors acknowledge that 83,000 comments might not be considered big data, their study is one of the most inclusive for a particular media market (South Korea). In addition, they build on current theories of the role of emotion in political engagement, while using large data sets as an empirical tool—not an end in itself.

Vargo and Hopp also address civility in political discourse. They compare census data with tweets from all 435 congressional districts in the United States. By combining these two different data sources, the authors are able to compare individual demographic and political traits with the type of political discourse in over 400,000 tweets. In addition, the authors took vote outcomes in each district as a measure of polarization. As opposed to conventional thinking on political polarization, those with lower level of polarization engaged in less civil conversations. Triangulating three data sets creates a picture of conversation on twitter that goes far beyond simple pattern recognition.

Brandtzæg (this issue) draws on over 21 million Facebook users to compare gender differences in political participation. In a rare glimpse of global social media habits, Brandtzæg applies developing theories on the role of expression online in spurring civic and political engagement. The author finds that gender gaps and differences in political preferences persist in social media, despite claims that the Internet can act as a gender “equalizer.”

Finally, Maireder and colleagues (this issue) introduce two new tools for mapping network connections on twitter, the audience diversity score and the communication connector bridging score. These measures assess the diversity of a particular actor’s followers and highlight the most influential actors in the network, who are potentially able to connect different spheres of communication. In the author’s analysis of the conversations on twitter around the Transatlantic Trade and Investment Partnership, they show how and when individuals might be able to spur protest without the help of mainstream media actors.

This volume speaks to the utility of digital trace data for political communication research. Social media and the Internet offer an alternative space to discuss political issues, consume news, and interact with elites—all in pursuit of a more personalized, interactive form of citizenship. As more and more of our lives moves online, communication scholars have a greater variety of tools and methods available for tracking human behavior than ever before. The era of computational social science is here. However, as Kuhn (2012) notes, the new paradigm will not be set at one point in time. The methods we use today, and more importantly, the theoretical rationale for employing those methods, take time to develop. We are still a long way from consensus in this area. Huge challenges remain, particularly with deriving valid interpretations and treating subjects with care. The articles in this volume represent some of this most forward think in these areas.