Critical Slowing Down, Not Intensity, of Collective Anger Predicts Surges of Social Movement Activity

Technical Summary: Critical Slowing Down of Collective Anger Predicts Surges of Social Movement Activity

Problem Statement
Social movements are characterized by highly irregular temporal dynamics, often shifting from periods of relative inactivity to abrupt, difficult-to-anticipate surges in participation. While online platforms have become central arenas for coordination and expression (e.g., the #BlackLivesMatter movement), predicting the timing of these activity surges remains a challenge. Existing literature identifies anger as a primary driver of individual participation in collective action; however, it remains unclear whether the intensity of collective anger or its temporal dynamics (persistence) better predicts large-scale mobilization. This study investigates whether internal dynamics within a social movement, specifically the temporal evolution of collective anger, can serve as early warning signals for impending surges in activity, drawing on complex systems theory.

Methodology
The study utilized a large-scale dataset comprising approximately 60 million tweets containing the hashtags #BlackLivesMatter and #BLM collected between January 2015 and May 2020 (prior to the George Floyd killing to avoid outlier dominance). The data was aggregated into daily time series.

1. Surge Detection: The authors employed Seasonal-Trend decomposition using Loess (STL) to isolate the residual component of daily tweet volume, removing long-term trends and seasonal cycles. Surges were defined as sharp local deviations from the baseline, specifically where the z-score of the residual exceeded 2.5 above the local average of the preceding 14 days. This method identified the onset of surges while minimizing false positives from sustained elevated activity.
2. Emotional Measurement: Daily emotional scores for anger, anxiety, and sadness were calculated using the LIWC-22 dictionary applied to the raw tweet text.
3. Predictor Variables:
   • Intensity: The mean level of expressed emotion over rolling windows (5-day and 30-day).
   • Persistence (Critical Slowing Down Indicators):
     • Autocorrelation (Lag-1): Measuring the degree to which current emotional states depend on past states.
     • Variance (Standard Deviation): Measuring the variability of emotional intensity over time.
4. Statistical Modeling: The primary analysis modeled the "distance to the next surge" (number of days until the next identified surge onset) as a continuous dependent variable. Linear regression models tested whether mean intensity, autocorrelation, and variance predicted this distance, controlling for the time elapsed since the previous surge. Robustness checks included multivariate analyses with other emotions, non-linear modeling (log-linear, quadratic, spline), and a null model using shuffled time series to rule out preprocessing artifacts.

Key Results

• Intensity vs. Persistence: Mean levels of collective anger did not predict the proximity to activity surges. In contrast, increasing autocorrelation in collective anger reliably predicted a shorter time to the next surge.
• Variance Findings: Contrary to some classical conceptualizations of critical slowing down (which often predict increased variance), the study found that variance in anger decreased as surges approached. This pattern, combined with rising autocorrelation, aligns with recent literature suggesting that reduced emotional flexibility (reduced variance) and increased dependence on the past (increased autocorrelation) signal a loss of resilience.
• Specificity to Anger: While anxiety intensity showed a negative association with distance to surges (higher anxiety = closer surge) and sadness showed a positive association, only anger demonstrated a consistent pattern of critical slowing down (increased autocorrelation and decreased variance) predictive of surges.
• Non-linearity: The relationship between anger autocorrelation and proximity to surges was non-linear. Autocorrelation increased progressively faster as the system neared a surge, with the steepest rise occurring in the final days before mobilization.
• Robustness: The predictive signal of autocorrelation remained significant even after controlling for preprocessing artifacts via a shuffled null model, whereas the variance effect was more sensitive to noise.
• Offline Correlation: Daily #BlackLivesMatter tweet volume was strongly positively correlated with independently estimated offline protest attendance, validating the online activity as a proxy for broader mobilization intensity.

Significance and Claims
The paper claims to extend early warning signal research from non-human complex systems to large-scale human behavior. The central theoretical contribution is the demonstration that collective mobilization is driven not by the average magnitude of anger, but by the temporal dynamics of emotional persistence.

The authors argue that as a social movement approaches a critical transition (a surge in activity), the collective emotional system exhibits "critical slowing down," characterized by delayed recovery from perturbations. This manifests as increased autocorrelation (emotional states becoming more dependent on recent history) and reduced variance (a narrowing of emotional expression). These findings suggest that collective emotional dynamics can provide advance warning of transitions in social systems, offering a potential tool for activists and policymakers to identify periods of latent collective tension. The study emphasizes that the emotional dynamics predicting when a system transitions differ from those motivating individual participation, highlighting the need to distinguish between individual-level drivers and system-level transition signals.

Limitations
The authors note that the study is correlational and does not establish causality; anger dynamics may signal rather than drive structural changes. The reliance on dictionary-based sentiment analysis (LIWC) may miss culturally specific expressions, though the authors argue that aggregation across millions of observations mitigates measurement error. Additionally, the focus on a single movement and online data limits generalizability, though supplementary analyses confirmed a link between online volume and offline protest size.