Superstorm Sandy: Looking at the Twitter Response

Event Description: 
Hurricane Sandy was a tropical cyclone that severely affected portions of the Caribbean, mid-Atlantic, and Northeastern United States. In diameter, it was the largest Atlantic hurricane on record, with winds spanning 1,100 miles. Sandy made its final landfall 5 miles (8.0 km) southwest of Atlantic City, New Jersey at about 8 p.m. EST on October 29, the center of the storm missing New York City by almost 100 miles. Sandy is estimated in early calculations to have caused at least $20 billion in damage. Affecting at least 24 states, from Florida to Maine, Hurricane Sandy brought tropical storm force winds stretching far inland and mountain snows in West Virginia. The hurricane brought a destructive storm surge to New York City, flooding numerous streets, tunnels, and subway lines in Lower Manhattan, Staten Island, Coney Island, the Rockaways, and other areas of the city and cutting off electricity to parts of the city and its suburbs. Severe damage occurred in New Jersey, especially in the communities along the Jersey Shore. (

Online Response by the Public: 
Data collected includes tweets from the public timeline containing the word “sandy” Tweets were posted between 20:45 UTC, October 28th and 23:59 UTC, November 10th, 2012. Over this period, more than 518,000 messages from over 363,000 unique users were collected. We show the estimated hourly posting rate for tweets containing “sandy” for the first 5 days of the event in the figure below. Night and day panels are distinguished by the background color (gray bars represent the nighttime hours of 8:00 pm EST to 8:00 am EST). Hourly seasonality is clearly visible as the rate of posting drops drastically during nighttime hours. Peak activity occurs around 8:00 p.m. on October 29th, 2012 — coinciding with the time the hurricane made landfall in New Jersey. In comparison with our previous work on the Waldo Canyon wildfire event, posting rates of tweets in this event were much higher, reaching over 550,000 per hour.


In addition to the rate of posting of event-related content we consider the most popular hashtags that were used in tweets containing the word “sandy.” The #sandy hashtag itself was the most popular hashtag in the collected tweets.

We create a wordcloud of the other hashtags used, removing #sandy, for visualization purposes. As seen in the figure to the right, many of the popular hastags make reference to “hurricane” or specific locations where the event was taking place.  We also find references to specific political figures and public officials. Some hashtags seem to request following relationships. If we consider the changes in hashtags used over time we find that one particular hashtag emerges post-event with high usage frequency — #prayforusa. Hashtag use is an interesting area for further study, however, due to limitation in data collection it is also subject to missing data. These preliminary findings, however, do demonstrate that event-specific hashtags are present and widely used.

 Online Response by National, State, Regional, and Local Targeted Accounts :
In addition to collected public tweets, we enumerated a set accounts from organizations and government officials involved in the response. 103 accounts in total were identified in the enumeration procedure. These accounts range from emergency management accounts such as “fema” and “redcrossny” to public officials such as “govchristie” and “scottmstringer.” We also include public utilities and other government office accounts. We begin by considering the rate of posting for these accounts of interest. Below, we show the estimated posting rate per hour per organization. As seen the number of tweets posted per hour per account doubles during the immediate post-event period (after November 4). We also find strong seasonality effects. Posting rates drop significantly during nighttime hours.

Next we consider the effect of the storm event on the number of followers of these accounts. In previous work on the Waldo Canyon wildfire, results indicated that local organizations/accounts tended to gain the highest number of followers relative to their initial follower count. In the figure below we show the proportion increase in followers for all 103 accounts of interest during the observation period. There are seven accounts in particular that more than double their follower count during this time. These accounts are labeled in the figure. Many of these accounts are related to infrastructure and public utilities.

  • redcrosssnorthnj – American Red Cross North Jersey Region
  • jcp_l – Jersey Central Power and Light
  • conedison – conEdison
  • psegdelivers - Public Service Enterprise Group Incorporated
  • mtainsider - Metropolitan Transportation Authority of the State of New York
  • nysbridge – New York State Bridge Authority
  • femaregion2 - FEMA Region: New York, New Jersey, Puerto Rico, and US Virgin Islands

As observed in previous work, very few accounts in our set of 103 actively add new ties (that is, following back new accounts) during the event period. Only four accounts show a large proportion increase in the number of accounts they follow. Interestingly, all four accounts that have a proportion increase of greater than 2 also have large follower increase. “redcrossnorthnj,” ” psegdelivers,” and “jcp_l” are three of these accounts that add outgoing following relationships. The account with the most follow back activity is “conedison” which shows an 11-fold increase in their outgoing ties.  On October 30th they were following just 10 accounts but by November 16th this number had risen to 113.

We also compare the proportion increase in followers across different organizational scales (by scale we mean local, state, or federal level accounts). During the Waldo Canyon wildfire local accounts experienced the highest increase of followers during the event. As seen on the right, the results differ for Hurricane Sandy. In this case, regional accounts gain this largest increase in followers, relative to their pre-event follower count. We might speculate that this different is due to the different in the scale of the actual event. Where the Waldo Canyon wildfire  affected a relatively small community within Colorado Springs, CO, Hurricane Sandy’s impact was felt across the entire Northeastern seaboard. Sandy affected many states and cities in the region. This is the most immediately obvious difference between the two events and is reflected in the results.

Our ongoing research has shown that posting behavior of government accounts differs from general public Twitter users. Here we ask about the tendency of our targeted accounts to follow Twitter developed conventions such as including hastags in tweets, directing content at other users, and retweeting others’ messages. Each of these content elements comprise what we call “relational microstructure.” In the figure below we show the proportion of tweets that contain each microstructure element. The set of tweets used in this analysis are posted by our set of targeted accounts between October 10th, 2012 and November 10th, 2012. We show a series of boxplots for these proportions based upon organizational scale, ranging from local to national.  We see similarities across the targeted accounts in terms of retweets, inclusion of a url, and lack of modified tweets.  But we notice that the national level accounts use hashtags in their posts more consistently than the others and the regional targeted accounts had more posts directed to individual users, signifying one to one communication activities.

Finally, the last aspect of the content posted by our set of targeted accounts we consider is the probability that messages will be “retweeted.” Retweeting is a term used to describe the behavior of re-posting a tweet. Such serial transmission of content is one means for diffusion of information through the following relationships on Twitter, since every message posted by an account is automatically delivered to his/her followers. In the figure below, we consider the probability that any given tweet posted by a specific account will be retweeted at least once. We consider the change in this probability pre-event and during the event (where pre-event is defined as any time before landfall of the hurricane and during is defined as after landfall). We observe large variance with some accounts showing striking increases in the probability that a message they post will be retweeted.  Importantly, some accounts are negatively affected and their retweet probability during the event decreases.

Along with tweet content, we also collect data on friend and follower relationships among targeted users. Here we consider the growth of the social network induced by our set of accounts of interest. Snapshots of the social ties are sampled daily. In the figure below, we illustrate the posting behavior and underlying follower network among our set of official accounts. Each node in the network represents a single organization/entity.  Nodes grow in size over time as they gain followers over time. Each directed tie (blue lines) represents a following relationship on Twitter. We find the induced social network becomes more dense over the course of the event. We also observe the emergence of clusters within the network. Nodes are highlighted in red when they post a message to Twitter. As tweets are automatically delivered to each of that account’s followers, we highlight these information exchange pathways in blue at the time of posting.



Lessons Learned:

  •  Twitter functions as a tool for risk communication, with heightened communication during the most critical points of a crisis event.
  • There is evidence that public attention on Twitter will increase significantly during crisis events, as shown by the increase in followers during the hurricane.
  •  While few organizations establish new friend relationships with public followers during a disaster, ties between organizations are imperative for maintaining communications in crisis.
  •  In comparison with previous HEROIC research on the Waldo Canyon fire event, there are differences event scale, magnitude, and effect over time, resulting in differing patterns of predicted retweets.
  •  Consistent with previous research, we find that few organizations have directed communications with followers, but instead use Twitter as a sort of redundant, broadcast mechanism.
  • There is evidence that patterns for content posting are emerging, as organizations include hashtags and links as part of their routine communication activities in disaster.

This material is based on research supported by the National Science Foundation under awards CMMI-1031853 and CMMI-1031779 and by the Office of Naval Research under award N00014-08-1-1015.

The above material is based on analysis completed by HEROIC team members. Please cite as follows.

Spiro, E., Sutton, J., Johnson, B., Fitzhugh, S. and Butts, C. (2012). “Superstorm Sandy: Looking at the Twitter Response.” Online Research Highlight.

Comments are closed.