TL;DR
This paper investigates the intricate spatiotemporal relationships between social media activity, online/mobile donations, and traditional media coverage during natural disasters, using Hurricane Irene as a case study. The authors demonstrate that social media, particularly Twitter, can serve as a leading indicator for donation behavior, highlighting the importance of considering both space and time for a comprehensive understanding of disaster response dynamics.
What Problem Does This Paper Address?
Understanding how people respond to disasters, particularly in terms of charitable giving, is crucial for effective relief efforts. While previous research has explored the temporal or spatial aspects of social media use and donation behavior in isolation, there's a significant gap in understanding how these dimensions interact. This paper aims to bridge this gap by investigating the combined spatiotemporal patterns of social media activity, online and mobile donations, and traditional media coverage during a natural disaster, Hurricane Irene. The goal is to determine if social media can act as a predictive indicator for offline behaviors like donations and to inform how organizations can better target their messaging.
Methods
The authors employ a multi-faceted analytical approach:
- Temporal Analysis:
- Fixed Effects Model: Used to identify daily cycles and the impact of the hurricane on tweet frequency.
- Variable Autoregressive Model (AR(p)): Applied to model the relationship between mobile donations and tweets, incorporating message events as exogenous factors.
- Vector Autoregressive Model with Exogenous Variables (VARX): A more sophisticated approach to analyze the dynamic relationships between traditional media, social media, and donation behavior, accounting for potential endogeneity. This model uses differenced variables due to stability issues in one variable.
- Impulse Response Functions (IRFs): Used to visualize the dynamic effects of an impulse (e.g., an increase in tweets) on response variables (e.g., donations) over time.
- Spatial Analysis:
- Manual Regional Characterization: Dividing the US into "highly impacted," "somewhat impacted," and "not impacted" zones based on Hurricane Irene's track.
- Spatial Autocorrelation Statistics: Employing Moran's I and Geary's C to measure the spatial clustering of donation and tweet data at the ZIP code level.
- Spatiotemporal Analysis:
- Fixed Effects Model: A spatiotemporal model is developed to simultaneously estimate spatial and temporal influences of tweets on mobile donations, incorporating state-level and area-level effects.
- Spatiotemporal Cluster Analysis: A k-means clustering algorithm is used to identify spatial clusters of events over time, analyzing the volatility of different data series.
Datasets
The study utilizes the following datasets:
- Geotagged Tweets: Collected using TwEater, focusing on tweets containing keywords related to Hurricane Irene. Approximately 22,045 geotagged tweets were used, covering the Continental United States and reverse geocoded to state and ZIP code.
- Donation Data:
- Mobile Donations: 10,073 donations received via SMS.
- Online Donations: 28,214 donations received via the web.
- For mobile donations, ZIP codes and latitude/longitude were estimated from phone numbers. Online donation data included billing addresses for geocoding.
- Traditional Media Coverage: Hourly counts of news stories about Hurricane Irene from CNN, CBS News, and ABC News websites.
The data spans from August 26, 2011, to September 13, 2011, covering the period before and after Hurricane Irene's US landfall.
Key Insights and Ideas
- Social Media as a Leading Indicator: The research strongly suggests that social media activity, particularly Twitter, can precede and predict offline donation behavior. This is a significant finding for disaster relief organizations looking to optimize their outreach.
- The Importance of Spatiotemporal Dynamics: Analyzing temporal and spatial patterns in isolation is insufficient. The interplay between where events occur and when they happen is crucial for a comprehensive understanding of disaster response.
- Lagged Effects of Social Media: While tweets can signal increased donations, there's a substantial lag. This is attributed to the typical sequence of events: social media buzz during the storm, followed by donations after the immediate impact.
- Differential Impact of Media: Traditional media doesn't directly boost donations as much as social media but can indirectly increase social media activity, which in turn positively influences donations.
- Geographic Influence: Individuals in areas directly impacted by the hurricane exhibit higher Twitter activity and a greater propensity to donate, underscoring the localized nature of disaster response.
- Twitter's Volatility: Twitter data is found to be the most volatile among the analyzed series, reflecting its real-time, dynamic nature.
Key Results
- Temporal Models:
- A fixed effects model showed that message events (relief agency solicitations) were strong exogenous factors for mobile donation peaks.
- VARX models indicated that while additional tweets about the hurricane initially led to a decrease in donations, after approximately 37 hours, prior tweets resulted in a significant increase in donations.
- Traditional media stories had a limited long-term effect on donation behavior, showing an initial burst followed by a trough.
- Increasing donations or media stories had little permanent positive effect on future donation/media story counts.
- New tweets were predicted to result in more future tweets, capturing information diffusion on Twitter.
- Traditional media sources could lead to future tweets, sparking conversations that boost social media activity.
- Spatial Analysis:
- Areas directly impacted by Irene (Area A) showed higher Tweets/population, web donations/population, and mobile donations/population compared to less impacted areas.
- Geary's C analysis revealed that Twitter activity patterns were more similar to mobile donations than web donations, both exhibiting negative autocorrelation in affected areas.
- Spatiotemporal Analysis:
- A spatiotemporal model found significant differences between states, with directly impacted states making more donations.
- K-means clustering showed Twitter data to be the most volatile (z=0.244), followed by mobile donations (z=0.215) and web donations (z=0.192).
Limitations and Future Work
- Model Sensitivity: The k-means clustering algorithm's sensitivity to the initial choice of clusters and temporal window size is acknowledged.
- Regression Method Dependence: The chosen regression method and variable selection might influence results.
- Future Work:
- Explore other variable selection methods (e.g., Akaike, Bayesian Information Criterion).
- Investigate time series decomposition based on rates of change or predictability.
- Further explore the sensitivity of clustering parameters.
- Develop more actionable strategic results to better describe and predict donation behavior.
- Build a decision support tool for managers interested in using this data.
- Investigate alternative clustering algorithms (k-medoids, DBSCAN) and aggregation methods.
Summary updated