TL;DR
This paper investigates whether COVID-19 led to an increase in cyberbullying discussions on Twitter. Using a novel Bayesian time-series model, they found a significant upward trend in cyberbullying-related tweets starting mid-March 2020, suggesting a correlation between the pandemic and increased online discourse about cyberbullying.
What Problem Does This Paper Address?
The COVID-19 pandemic, with its associated lockdowns and increased reliance on digital communication, raised concerns about a potential surge in cyberbullying. While anecdotal evidence and speculation exist, a rigorous quantitative analysis to confirm this hypothesis was lacking. Existing methods for analyzing time-series data, like simple frequentist approaches, often fail to capture the serial correlation inherent in tweet data and may miss subtle, non-abrupt changes. This paper aims to address this gap by providing a robust statistical framework to analyze trends in cyberbullying discussions on social media during the pandemic.
Methods
The authors employ a Bayesian time-varying autoregressive count (TVBARC) model to analyze the daily counts of tweets related to cyberbullying. This model is an extension of the linear Poisson autoregressive model and is specifically designed to handle:
- Time-varying parameters: Allowing trends to change smoothly over time, rather than assuming a constant rate.
- Autoregression: Accounting for the dependence of current tweet counts on past counts, a crucial aspect of time-series data.
- Poisson distribution: Suitable for modeling count data, where the mean and variance are related.
This Bayesian approach allows for the construction of posterior samples of the parameter functions over time, providing a more nuanced understanding of trend changes compared to traditional change-point analysis, which often looks for abrupt shifts. The model uses B-spline basis expansions for the time-varying functions and employs Markov Chain Monte Carlo (MCMC) methods for posterior computation.
Datasets
The study utilizes a dataset of 454,046 publicly available tweets collected from Twitter. These tweets were posted between January 1st, 2020, and June 7th, 2020, and contained keywords related to cyberbullying. The data was collected using the "Get Old Tweets API" and stored in MongoDB. The keywords were broadly categorized into three sub-classes: "cyber" (CY), "online/internet" (ON), and "twitter" (TW).
Key Insights and Ideas
- Subtle trends are important: Simple visual analysis or traditional change-point methods can miss important trends, especially when they are not abrupt. The Bayesian approach reveals a smoother, more continuous upward trend.
- COVID-19 correlates with increased cyberbullying discourse: The analysis strongly suggests an increase in discussions about cyberbullying on Twitter starting around mid-March 2020, coinciding with the onset of the pandemic and widespread social distancing measures.
- Dependence matters: Ignoring the serial correlation in tweet data can lead to unjustified conclusions. The autoregressive component of the model is crucial for accurately capturing the dynamics of tweet counts.
- Model flexibility: The TVBARC model's ability to handle time-varying parameters and count data makes it a powerful tool for analyzing social media trends during significant global events.
Key Results
The Bayesian trend analysis revealed a significant upward trend in cyberbullying-related tweets starting around mid-March 2020. This trend continued, with some fluctuations, throughout the observed period. While the AR(1) and AR(10) models showed comparable results, the AR(1) model indicated that the first lag accounted for most of the correlation. The mean trend functions for the total count and the three sub-classes (CY, ON, TW) all showed a similar pattern of increase around early March. The authors note that while this shows an increase in discussion of cyberbullying, it doesn't directly prove an increase in cyberbullying incidents themselves, but rather a correlation between the crisis and the online discourse surrounding the issue.
Limitations and Future Work
The authors acknowledge that their current work focuses on the discussion of cyberbullying and not the direct incidence of cyberbullying. They plan to conduct in-depth qualitative analysis of the tweet content to further understand the nuances of these discussions. They also note that the dataset covers only the initial phase of the pandemic, and future work could extend the analysis to later phases, which may have involved different trends or impacts. The study's findings are based on Twitter data, and extending the analysis to other social media platforms could provide a more comprehensive picture.
Summary updated