Can We Use Social Media Data to Predict Our Political Behaviors? A Mixed-methods Study

Previous research has utilized both network and content features of social media data to predict real-life public opinion and political behaviors (e.g. Adamic & Glance, 2005; Bollen et al., 2011; Boutet et al., 2012; Burgess & Bruns, 2012; Conover et al., 2011; Garcia et al., 2012). Still, there is no clear consensus about the utility of such approaches, as the predictive power of social media-derived estimates varies widely (Gayo-Avello, 2013). In this study, based on both survey and Twitter data from 251 respondents recruited on Amazon Mechanical Turk, we explore the feasibility of using social media data to predict online and offline political participation. The predictors include psychological features extracted from tweets, measures of online network influence, as well as demographic factors. Based on the 82,275 tweets of the 251 respondents collected during Nov. 2015 to Jan. 2016, we use LIWC lexicon (Pennebaker et al., 2015) to conduct text analysis and extracted three categories of psychological features (i.e., drives, affects, and cognitive process). In detail, five types of drives (i.e., affiliation, achievement, power, reward, and risk), five types of affects (i.e., positive emotion, negative emotion, anxiety, anger, and sadness), six types of cognitive process (insight, causation, discrepancy, tentative, certainty, differentiation) were derived from the tweets. Online network influence was measured by four types of centrality metrics (i.e., [in-/out-] Degree centrality, Betweenness centrality, Closeness centrality, Eigenvector centrality) generated by social network analysis (Freeman, 1978), PageRank (Chen et al., 2007), Follower/Followee ratio (Gayo Avello & Brenes Martínez, 2010), and Retweet/Tweet ratio (Hawthorne et al., 2013). Online political participation and offline political participation are measured by the scales modeled after a study by Gil de Zúñiga et al. (2012).

We found that five psychological features (i.e., positive emotion, insight, affiliation, and reward) together significantly predicted online political participation (model 1), R2 = .102, F (5,245) = 5.56, p < .001, while only one psychological factor (i.e., affiliation) significantly predicted offline political participation (model 2), R2 = .024, F (1,249) = 5.994, p < 0.05. Thus, we conclude that user generated contents in social media could provide certain psychological cues to predict one’s online and offline political behavior; however, online network influence, despite of the various measures utilized to capture its dimensions, failed to help in predicting political behaviors, both online and offline.