Private Partisan, Public Moderate: Preference Falsification on Twitter

Abstract

While there is a substantial body of research demonstrating the phenomenon of preference falsification, or inconsistencies between private and public preferences dependent on social context, there has been little examination of if and how these dynamics manifest in everyday online communication. This paper seeks to determine whether the public personas that individuals present on social media are ideologically consistent with their privately held beliefs, and under what social network conditions and for whom public-private discrepancies are most likely to arise. I experiment with a number of neural network architectures, particularly variations on the Long Short Term Memory network, to build a state-of-the-art model for detecting ideological slant in Tweets and embedded articles publicly posted on the social media platform Twitter. I achieve strong overall prediction accuracies of over 85 percent on both articles and Tweets and produce well-validated estimates of the public preferences of both political elites and media outlets. These predictors are used to construct estimates of the slant of the public preferences of Twitter users, which are compared with well-validated techniques for estimating private ideological preferences to gain a deeper understanding of how the composition of an individual’s social network influences the degree and frequency of their preference falsification. I find that the position of the ideological centers and the degree of ideological homogeneity of users’ follower networks have little effect on their falsification patterns; rather, there is a strong overall trend towards expressing moderate beliefs, irrespective of social context.