Silence of the Tweets: incorporating social media activity drop-offs into crisis detection

Although extreme events are inevitable, the associated cost to infrastructure and human life is not. We can mitigate these costs through improving the information available to emergency responders during and after crisis events via social media. Recent research has identified a correlation between spikes of Twitter activity and the infrastructural damage incurred during natural disasters. This research, however, overlooks emergencies occurring in areas in which people have lost power, lack the ability to connect to the internet, or, due to differences in social media perceptions, are uncompelled to Tweet during a disaster. To assess the prevalence of Twitter activity decreases and the relative importance of those decreases in detecting areas in crisis, we study crisis-driven Twitter activity deviations from "normal" in nine cities affected by the 2017 Atlantic hurricane season. In analyzing more than 1.1 million Tweets across the season, we find that there is a stronger, more significant correlation between infrastructure damage and a metric that prioritizes both increases and decreases in Twitter activity than one that prioritizes only Twitter activity increases. These findings indicate that social media drop-offs could be representative of significant distress, and accounting for the apparent survivorship bias in social media will be critical to the equitable use of social media in crisis applications.