Down the Rabbit Hole: Modeling Twitter Dynamics through Bayesian Inference

Social media usage, and its impact on people’s physical and mental health, is of interest to a diverse range of academic disciplines and everyday people. Despite this, we know very little about the ways in which, over months and years, someone’s social media use may escalate to consume significant amounts of their daily leisure time. Nor do we understand the ways in which a user’s posts may shift into toxic or unexpectedly abusive patterns. Understanding the long-term dynamics of use is complicated by the fact that day-to-day engagement has significantly non-normal statistics and may fluctuate by orders of magnitude—informally, users are sometimes driven to rare “binges” with lasting consequences for their future trajectory. To address this complex interplay of timescales, this work presents a Bayesian model for usage over time, flexible enough to capture a wide range of short and long-term temporal dependencies. Examining the “dose response” curves of a random sample of 500 users, we find that most users (≈ 90%) show evidence for a stable, equilibrium level of use. A smaller “high-risk” subset (≈ 10%) show evidence for instability: when short-term fluctuations drive their levels of use sufficiently high, they enter a new phase of sustained, run-away usage. Once we control for the levels of use, we find that “likes”, retweets, and other forms of feedback received from other users do not significantly impact future behavior. This casts doubt on the common heuristic that social media use is driven by an “addiction to likes": for example, there is no evidence that a user whose posts receive an unexpectedly low level of likes posts more to “make up the difference”. Finally, in looking at the dynamics of user toxicity, we find a tail-risk effect: prior toxic behavior rarely shifts the user’s median post, but rather increases the likelihood of (otherwise rare) extreme toxicity. Our flexible dynamical modeling approach reveals significant heterogeneity in the ways in which users adapt to social media systems, and opens the door for more qualitative investigations into the outsized effects social media may have on users across arbitrarily long timescales