[Paper Review] Characterizing Twitter users behaviour during the Spanish Covid-19 first wave

Bernat Esquirol, Luce Prignano+2 more|arXiv (Cornell University)|Dec 11, 2020

Misinformation and Its Impacts13 references4 citations

TL;DR

This study analyzes Twitter activity during Spain's first Covid-19 wave, using segmented regression and Bayesian switchpoint detection to identify behavioral shifts. It reveals that journalists and general users abruptly increased tweeting after March 9, 2020, while politicians showed no activity surge but instead increased retweeting, indicating endogenous institutional dynamics rather than crisis-driven behavior.

ABSTRACT

People use Online Social Media to make sense of crisis events. A pandemic crisis like the Covid-19 outbreak is a complex event, involving numerous aspects of the social life on multiple temporal scales. Focusing on the Spanish Twittersphere, we characterized users activity behaviour across the different phases of the Covid-19 first wave. Firstly, we analyzed a sample of timelines of different classes of users from the Spanish Twittersphere in terms of their propensity to produce new information or to amplify information produced by others. Secondly, by performing stepwise segmented regression analysis and Bayesian switchpoint analysis, we looked for a possible behavioral footprint of the crisis in the statistics of users' activity. We observed that generic Spanish Twitter users and journalists experienced an abrupt increment of their tweeting activity between March 9 and March 14, in coincidence with control measures being announced by regional and State level authorities. However, they displayed a stable proportion of retweets before and after the switching point. On the contrary, politicians represented an exception, being the only class of users not experimenting this abrupt change and following a completely endogenous dynamics determined by institutional agenda. On the one hand, they did not increment their overall activity, displaying instead a slight decrease. On the other hand, in times of crisis, politicians tended to strengthen their propensity to amplify information rather than produce it.

Motivation & Objective

• To characterize how different classes of Twitter users in Spain changed their online behavior during the first wave of the Covid-19 pandemic.
• To identify critical transition points in user activity that may correspond to policy interventions or public awareness shifts.
• To assess whether the behavioral response to the pandemic crisis differs across user types such as journalists, politicians, and general users.
• To evaluate the reliability of standard activity pattern models during crisis events, especially under the influence of infodemics.
• To explore whether online social media can serve as an early warning system by detecting behavioral footprints of societal stress during crises.

Proposed method

• Collected and analyzed timelines of 1,000 Spanish Twitter users across three classes: journalists, politicians, and general users.
• Applied stepwise segmented regression to detect abrupt changes in daily tweet rates, identifying potential switchpoints in user activity.
• Used Bayesian switchpoint analysis to estimate the most probable time of behavioral shifts, assuming a Poisson process for tweet counts.
• Validated results using synthetic data with multiple switchpoints to assess model robustness and p-value interpretation.
• Compared Poisson-based switchpoint models with sigmoidal and multistate models to evaluate sensitivity and accuracy of change detection.
• Evaluated the proportion of retweets versus original tweets before and after detected switchpoints to assess shifts in content production vs. amplification.

Experimental results

Research questions

• RQ1 When did significant changes in Twitter activity occur for different user types during Spain’s first Covid-19 wave?
• RQ2 Do journalists, politicians, and general users exhibit distinct behavioral responses to the onset of pandemic control measures?
• RQ3 Is there a detectable shift in the balance between original content production and retweeting during the crisis, and does it vary by user group?
• RQ4 Can Bayesian switchpoint analysis reliably detect behavioral transitions in user activity during a crisis, even when multiple shifts are present?
• RQ5 How do the activity patterns of politicians differ from those of other users, and what does this imply about their role in crisis communication?

Key findings

• Journalists and general users exhibited a sharp increase in tweet activity between March 9 and March 14, 2020, coinciding with the announcement of national and regional lockdown measures.
• The proportion of retweets remained stable before and after the switchpoint for journalists and general users, indicating no significant shift in content consumption habits despite increased activity.
• Politicians did not experience a surge in overall activity; instead, they showed a slight decrease in tweet volume, suggesting a more restrained communication strategy.
• Politicians significantly increased their retweeting behavior during the crisis, indicating a strategic shift toward amplifying information rather than producing original content.
• The Bayesian switchpoint model detected a single dominant switchpoint in most cases, but p-values below 0.5 suggest that multiple switches may exist and could be missed by the model.
• The multistate model provided better fit for complex activity patterns but introduced more noise and computational cost compared to the simpler linear breakpoints model.