1. Introduction
The risks posed by many natural hazards are dynamic in that the threat and information available about it evolve. When a hurricane threatens a coastline, for example, its position and intensity changes, and forecast and preparedness information is refined as the storm approaches. People’s assessments of and responses to natural hazard risks are also dynamic, as individuals process information and interact with each other to communicate about, interpret, and respond to the changing threat. These dynamic individual and social processes are fundamental aspects of how people perceive and respond to natural hazards (Morss et al. 2017). Thus, it is essential to understand them in order to develop effective risk communication and emergency response policies that help protect people from harm.
Although previous research notes that people may iteratively assess and respond to a threat (Lindell and Perry 2012; Mileti and Sorensen 1990), there is little empirically based understanding of how this actually occurs for an evolving risk, such as that posed by an approaching hurricane. This lack of knowledge derives partly from the difficulty in gathering data from people about what they know, perceive, feel, and do at multiple times as a threat is occurring. These challenges are exacerbated by the inherent uncertainties associated with whether, where, and when hazardous weather events will occur, which make it difficult to know in advance who will be affected and thus to design and implement real-time data collection with those populations throughout a threat.
The prevalence of social media use offers new opportunities for studying the dynamic risk information ecosystem that emerges when hazards threaten (Morss et al. 2017). Although in the weather community social media is often discussed in terms of its potential for authorities to distribute risk information, it can offer much more. Social media offers intrinsically participatory platforms where users actively access, discuss, create, and share information about many topics, including information about their situations, attitudes, and behaviors related to risks (Palen et al. 2010; Neeley 2014). Moreover, when people post on social media more than once during the course of a threat, their posts offer a chronological record of how they are assessing and responding to a real-world, evolving risk from their context and perspective.
Here, we utilize social media data as a lens for examining individuals’ risk information behaviors, risk perceptions, and responses as a hazardous weather threat unfolded over several days. We investigate these processes by qualitatively analyzing Twitter narratives created by people at risk from Hurricane Sandy1 during the time period leading up to and during the storm’s landfall. The Twitterers analyzed were located in the Far Rockaway neighborhood of New York City, which was in a mandatory evacuation zone during Sandy. Complementing recent work that utilizes Twitter data for macrolevel analyses of what hazard information is shared, how much, and by whom across broader populations (see section 2), we perform an in-depth analysis that aims to build a rich understanding about how individuals experience evolving risks.
Our study addresses three research questions:
- How do people interact with different types of information related to the hurricane threat?
- How do they perceive and respond to the risks posed by the hurricane?
- How do these processes evolve and interact as the threat unfolds?
The goal of this analysis is to gain a deeper understanding of the complex and evolving ways that people assess and respond to their risk from the dynamic threat of an approaching hurricane by analyzing social media narratives from a sample of individuals who were at high risk leading up to Sandy. In doing so, we aim to augment knowledge gained from other studies that utilize complementary methods, samples, and theories to investigate behavioral responses to risks. We further aim to provide a foundation for additional research investigating similar questions using social media data, including broader Twitter datasets.
This study adds to the weather risk, natural hazards, and social media literatures in several ways. First, it reveals what is salient to people who are at risk from an approaching hurricane as they process the ubiquitous pieces of risk-related information available to them, evaluate the risk, and decide how to respond. Because the Twitter narratives we investigate provide a new type of data for understanding how people assess and respond to risks, the analysis reveals aspects of these processes that have not previously been well described in the theoretical and empirical literatures. It also develops new knowledge about the dynamics of these processes, for example, what factors are important at different times and how these intersect and change as a threat evolves. Finally, the analysis illustrates the potential value not only of social media data, but also of social media narratives, for building understanding about how people interact with information and perceive and respond to evolving threats.
2. Background and study scope
A number of previous studies have investigated how people assess and respond to hurricane risks [see, e.g., reviews by Baker (1991), Dash and Gladwin (2007), Lindell (2012), Lazo et al. (2015), and Huang et al. (2016)]. This body of work provides important knowledge about how people’s perceptions of hurricane risks and their protective decisions are influenced by a variety of factors, ranging from sociodemographic characteristics to situational factors to risk messages. Much of this research utilizes data gathered through surveys, survey-based experiments, or interviews to understand how people perceive risks and make decisions at a specific point in time or integrated across a hurricane threat. Fewer studies have focused on understanding how people’s decision processes evolve over the lifetime of a hurricane using, for example, retrospective interviews (Gladwin et al. 2001; Taylor et al. 2009; Morss and Hayden 2010), simulations (Christensen and Ruch 1980; Meyer et al. 2013; Wu et al. 2015a,b) or multiple phone surveys conducted during the same hurricane threat (Meyer et al. 2014). Because hurricane threats and responses are dynamic, as Meyer et al. (2014, p. 1402) note, “additional attempts to conduct real-time measurements of responses to natural hazards” are needed—particularly multiple measurements from the same sample of people. Social media data provide one means for potentially filling this gap, whereby an individual’s postings and shares over time constitute “multiple measurements” for analysis.
Social media is a useful resource for studying hazards and disasters because it is a participatory platform that users actively and creatively leverage during disruptive, uncertain situations (Palen et al. 2010; Neeley 2014; Houston et al. 2015). As such, social media helps make visible the individual and social processes that have long been thought to contribute to how people assess and respond to risks. The information that people share on social media is quasi real time, in that it can reveal what they observe, think, or feel at that moment, or it can be summative and reflective. In the context of hazards, this information may be about the threat, situations, attitudes, perceptions, and behaviors pertaining to one’s own or others’ risk.
Much of the research noted above collects data about how people assess and respond to hurricane risks using questions structured by the researcher. Social media provides a different type of data, with its content determined by what a person chooses to convey in quasi real time from their perspective. The collection of information from a social media user over a period of time constitutes a narrative, that is, a “written text giving an account of an event/action or series of events/actions, chronologically connected” (Creswell 2007, p. 70). These narratives can be analyzed to investigate timing, changes, and causal connections in what people share. In short, social media leave “digital traces” of individuals’ perspectives when faced with real-world, changing risks, providing researchers a window into people’s evolving risk assessments and decision-making (Palen et al. 2010; Morss et al. 2017).
Twitter is one social media platform that is particularly conducive to research because the data are publicly available (Twitter 2016). Tweets are limited to 140 characters (as of November 2017, tweets may now contain up to 280 characters), but they nevertheless can provide a rich source of information, including the tweet text itself and embedded emoticons or emojis along with Internet links to websites, photos, Facebook posts, and so forth. Thus, researchers are leveraging Twitter to investigate different hazard and crisis events. Twitter research of weather hazards includes studies of winter weather, floods, tornadoes, and hurricanes, including Hurricane Sandy. One area of emphasis within this body of research is the temporal and geospatial patterns in hazard-related Twitter data and their covariation with other factors, such as National Weather Service watches and warnings or economic damages (Lachlan et al. 2014a; Ripberger et al. 2014; Shelton et al. 2014; Kryvasheyeu et al. 2016). Scholars also have studied Twitter activity by public safety organizations and news media, including their tweet content and citizen engagement (Cates et al. 2013; Hughes and Palen 2014; Lachlan et al. 2014b; St. Denis et al. 2014; Sutton et al. 2015; Rice and Spence 2016). Other areas of research include classifying “useful information”2 or expressions of emotion in citizen’s tweets (Brynielsson et al. 2013; Lachlan et al. 2014a,b; Spence et al. 2015) and investigating local versus nonlocal information sources and behaviors (Shelton et al. 2014; Kogan et al. 2015).
These Twitter-focused research studies provide valuable knowledge about the Twitter information ecosystem during weather hazards. In most studies, however, the units of observation and analysis are the individual tweet (or a subset of the individual tweet), and thus the tweet datasets are derived accordingly (e.g., by using only certain keywords or hashtags, by only gathering a set number of most recent tweets at a point in time). Consequently, other tweet content that may be relevant goes uncollected and unanalyzed. Moreover, individual and social evolutions in and connections among risk information, perceptions, responses, and other factors—which are key topics of interest in our research questions—are less able to be investigated.
Here, we add to the extant research by studying a sample of people who were at risk from a dynamic weather threat and examining how they evaluated and managed this evolving risk. As discussed in detail in section 3b, we identified users who were at high risk of Sandy and collected all of the tweets in their narratives (and associated images and other content) over a multiday period as the storm approached and made landfall. This approach allows us to acquire and analyze tweets in which users discuss the risk of Sandy but that do not necessarily use any researcher-defined keyword. This approach also allows us to analyze the context of, associative and causal connections among, and evolutions in a user’s tweets.
We analyze these narratives qualitatively to address the three research questions presented in the introduction. Our analysis draws empirically on knowledge from previous research on hurricane risk communication and evacuation decision-making, including the studies noted above. We also draw on concepts and theories from the risk communication and hazards literatures, including Lindell and Perry’s (2012) Protective Action Decision Model (PADM), which is a multistage, iterative framework that models people’s responses to environmental hazards. More specifically, the analysis draws on representations of information sources (forecast and warning messages, social cues, environmental cues), threat perceptions (cognitive and affective), and responses (communication, protective, and emotion focused) from the PADM and other literature (e.g., Mileti and Sorensen 1990; Peacock et al. 2005; Trumbo et al. 2016; Demuth et al. 2016). However, the analysis ultimately is based on the content of the Twitter narratives. Thus, in addition to elucidating how people discuss these concepts in ways that are consistent with what is characterized in the existing literature, our analysis reveals new aspects of how people experience hazard threats.
3. Methods
a. Hurricane Sandy: Summary of the event
Hurricane Sandy formed as a tropical depression in the Caribbean Sea on 22 October 2012. Over the next two days, the five-day hurricane track and cone-of-uncertainty forecasts from the U.S. National Hurricane Center projected Sandy moving north and making landfall in Jamaica, Cuba, and the Bahamas, and then curving northeast out to the Atlantic Ocean (Fig. 1a). Subsequent forecasts showed the hurricane recurving to the north, with indications that the northeastern U.S. coast could be affected (Fig. 1b). By late morning on 25 October, Sandy was projected to make landfall near New Jersey (Fig. 1c), and the forecast track and landfall location remained consistent over the subsequent days. On 28 October just before 1530 UTC, New York City’s Mayor Bloomberg announced a mandatory evacuation order for zone A of New York City, which included Far Rockaway (see section 3b). Sandy made landfall on 29 October at approximately 2330 UTC along the New Jersey coast, approximately 140 km south-southwest of Far Rockaway. Far Rockaway experienced sustained winds of approximately 50 kt (25.7 m s−1) and a storm surge of approximately 1.5–1.8 m (5–6 ft.) above ground level (NOAA 2013a,b).
The National Hurricane Center’s 5-day forecast track and cone of uncertainty for Sandy, issued (a) Wednesday, 1200 UTC 24 Oct 2012; (b) Thursday, 0900 UTC 25 Oct 2012; and (c) Thursday, 1500 UTC 25 Oct 2012.
Citation: Weather, Climate, and Society 10, 3; 10.1175/WCAS-D-17-0126.1
b. Hurricane Sandy Twitter data collection, sampling, and analysis
As Palen and Anderson (2016, p. 225) articulate, “a tempting myth is that large volumes of social media data alone will reveal patterns of behaviors.” Yet, in order to develop robust, meaningful findings, data from social media (like other forms of research data) must be extracted and sampled in ways that match the research questions. To develop such a dataset for this study, we used a multistep process (Palen and Anderson 2016).
First, beginning on 24 October 2012, we used Twitter’s public streaming application program interface (API) to collect, in real time, all tweets3 that included any of the following Sandy-related keywords: frankenstorm, hurricane, hurricanesandy, perfectstorm, sandy, sandycam, stormporn, and superstorm (Kogan et al. 2015; Morss et al. 2017). This collection yielded 15.9 million tweets worldwide through 16 November 2012.
Initial exploration of this Twitter dataset suggested that there were indications of information use, risk perceptions, and decision-making related to Sandy, but that mentions of these constructs were rare and difficult to extract from the global keyword dataset. After exploring several ways of sampling the Twitter data to support analysis of people’s risk assessments, we focused our analysis on tweet streams provided by Twitterers who resided in geographic areas that were at high risk of strong winds and storm surge from Sandy and thus were asked to evacuate. In other words, we focused on people who were sufficiently exposed to Sandy that they might be deciding whether to take protective action—and thus utilizing Twitter in assessing their risk—prior to landfall. After exploring multiple ways to identify such a population in the dataset (Morss et al. 2017), we utilized a geographically based sampling strategy by selecting Twitterers who (leading up to Sandy) resided in an area that was at high risk from the hurricane and that experienced significant impacts. This approach is similar to that often used in posthurricane interview and survey studies, in which people are sampled from selected geographic areas that were at risk and/or significantly affected by the storm.
We selected Far Rockaway, New York, to investigate our research questions for several reasons. Far Rockaway is a neighborhood on the Rockaway Peninsula of New York City, and thus it is at significant risk from storm surge flooding from coastal storms. Consequently, Far Rockaway is fully within New York City’s evacuation zone A, and thus the entire neighborhood was under a mandatory evacuation order for Sandy (Fig. 2), and it experienced significant impacts from Sandy (Shelton et al. 2014; Superstorm Research Lab 2013). Far Rockaway also is a sufficiently distinct geographic “place” that Twitterers made reference to it, which we determined empirically by exploring its mention within the Sandy keyword-based dataset (described in the next paragraph) relative to other New York City neighborhoods that were under a mandatory evacuation order. The neighborhood distinction allowed us to subsample from the Sandy-keyword dataset using localized, place-based terms (Palen and Anderson 2016) without obtaining an overwhelming number of nonlocal Twitterers.
(a) Study location of Far Rockaway, located in the Queens neighborhood of New York City, New York, and (b) map showing the issuance and location of the mandatory evacuation of zone A for Hurricane Sandy along with the location of the other nonevacuated zones (New York Times 2012).
Citation: Weather, Climate, and Society 10, 3; 10.1175/WCAS-D-17-0126.1
Identifying the set of Far Rockaway Twitterers was informed and refined by our exploration of the Twitter data; we searched the Sandy-keyword-based dataset for mentions of “farrockaway,” “far rockaway,” “far rock,” and “farrock”4 in either the users’ tweet content or their metadata during the period from 24 October through 7 November. Next, using the full-archive, historical search from the Gnip API, we pulled every tweet authored by each of these Twitterers during the same time period. These “contextual” tweet streams provide a Twitterer’s narrative, which is important for interpreting tweets in context rather than in isolation and for finding tweets that are relevant to Sandy but that do not explicitly mention the Sandy keyword search terms used (Palen and Anderson 2016; Morss et al. 2017). This place-based, contextual dataset yielded 307 Twitterers with approximately 144 000 tweets [see Anderson et al. (2016) for more details].
The first two authors then read through the contextual tweet streams for the 307 Twitterers to identify those who were located in the Far Rockaway area leading up Sandy’s landfall (rather than tweeting about it from afar). Users also were required to be tweeting primarily in English (for readability by the analysts), to have some original tweet content [versus all retweets (RTs)], and to have at least one original-content tweet pertaining to Sandy before or during landfall to allow for analysis of their risk assessments during this period. This process of pulling contextual tweet streams to identify a relevant Twitter user sample is what Palen and Anderson (2016, p. 225) refer to as “mak[ing] ‘Big’ data bigger, then smaller.” The resulting dataset for this article consists of 53 Far Rockaway–area Twitterers who generated 8660 tweets from 24 October through 7 November. The number of tweets per user during this period ranges from 6 to 1040 and the overall distribution is right skewed (median = 78.0, mean = 163.4, std dev = 228.0).
Per our research questions, we qualitatively analyzed these Twitter narratives, focusing on the Sandy-related information that the Twitterers attended to and shared, their perceptions of the risk to themselves and to others, their responses to the threat, and evolutions in these processes as Sandy approached and made landfall. Each tweet was analyzed in the context of the user’s full tweet narrative and in the context of the Sandy threat. When analyzing the data, we examined the tweet text itself as well as emoticons/emojis and linked content (images, Facebook posts, etc.), when publicly available. Our data analytic approach is iteratively deductive, drawing on theories of behavioral responses to risks and knowledge from previous empirical studies (section 2), and inductive, guided by what the 53 Far Rockaway users chose to tweet about, when, and how.
Per Twitter’s terms of service, unless individuals choose to protect their tweets, all tweets are visible to anyone with or without a Twitter account. Twitter’s privacy policy further explains that the company shares data with universities (Twitter 2016). The Twitter data analyzed here thus are publicly available. However, this does not absolve researchers from responsibly treating Twitterers (Boyd and Crawford 2012; Zimmer and Proferes 2014; Bica and Anderson 2016). We have therefore taken several steps in our data presentation in accordance with our ethical considerations to respect and minimize risk of harm to the Twitterers. For all tweets presented, tweet authors were anonymized, web links were removed, profanity in tweets was redacted, and user names of other Twitterers who were explicitly named (i.e., @mentions) were anonymized except for Twitterers who clearly maintain a public profile (e.g., media professionals). Also, to make it more difficult to search for the Twitterers presented in section 4c, the tweet text in the narratives was modified in minor ways that do not alter the meaning. The tweet content was not otherwise modified (e.g., punctuation, capitalization, and misspellings were not corrected). Last, we have taken care to avoid presenting tweet content with identifying or sensitive information, and we focus on presenting tweets that illustrate points that are central to the research purpose.
4. Results
This section examines how people interacted with information and perceived and responded to risks leading up to and during Sandy’s landfall, as revealed by our analysis of the Far Rockaway Twitter narratives. The findings illustrate what was most salient to these high-risk individuals as Sandy threatened, as indicated by what they chose to tweet about.
The main informational, perceptual, and response-related themes in the data are depicted in Fig. 3. The information themes are discussed in section 4a, and the perception and response themes are discussed in section 4b. Exemplar tweets illustrating key points for these two sections are presented in Tables 1–9 and are referenced with alphanumeric identifiers. Although we use individual tweets as examples, these were interpreted in the context of the individuals’ narratives.
Model of key types of hurricane risk information, perceptions, and responses and their interactions that emerged from analysis of Twitter data from Hurricane Sandy.
Citation: Weather, Climate, and Society 10, 3; 10.1175/WCAS-D-17-0126.1
Example tweets about weather forecast information.
Example tweets about evacuation information.
Example tweets about social cues.
Example tweets about natural and built environmental cues.
Example tweets about perceived exposure.
Example tweets about perceived severity.
Example tweets about affective and emotional perceptions and responses.
Example tweets about preparatory and protective actions.
Example tweets about coping responses.
The primary emphasis in sections 4a and 4b is on characterizing the main concepts in Fig. 3, although aspects of their evolution and interactions are discussed. In section 4c, we discuss how these concepts evolve and interact in greater depth by examining four example Twitter narratives, presented in Tables 10–13 and again referenced with alphanumeric identifiers.
Tweet narrative from Twitterer33. Tweets that are emphasized in the text are shown in boldface. (n/a = not applicable)
Tweet narrative from Twitterer53. Tweets that are emphasized in the text are shown in boldface. (n/a = not applicable)
Tweet narrative from Twitterer48. Tweets that are emphasized in the text are shown in boldface. (n/a = not applicable)
Tweet narrative from Twitterer6. Tweets that are emphasized in the text are shown in boldface.
a. Risk information
Analysis of the Twitter narratives reveals that people attended to four major types of risk information as the hurricane threatened: weather forecast information, evacuation orders, social cues, and environmental cues (Fig. 3).
Many Twitterers made reference to information regarding the weather forecast for Sandy (Table 1). Most of these mentions were implicit, in that no specific forecast products or sources were named, yet the tweets indicate that the Twitterers had obtained some type of forecast information. Examples include references to the hurricane’s forecasted timing of landfall and impacts (A1–A3), track (e.g., “a hurricane coming towards my crib again” in A4), and severity of impacts (A5). These examples illustrate how Twitter narratives can reveal the influence of forecasts even when people do not specifically mention them in the language typically used by forecasters.
Fewer Twitterers made explicit references to forecast information from formal sources or products. Those who did shared information about Sandy’s forecasted timing of landfall (A6) and physical impacts, including storm surge (A7, A8) and wind gusts along the coast (A9, A10). Most of the explicit forecast mentions were RTs, such as those shown in A7 and A10 from private weather providers and in A8 from a public official. Some, however, tweeted in their own words about explicit forecast information that they obtained, for instance, from the National Weather Service (A6) or from television meteorologists (A9, Twitterer34’s reference to being “in the 60–80 [mph] part” of the forecast map shown in Fig. 4).
Photo shared by Twitterer34 of a television meteorologist showing the forecast wind gusts from Sandy (see tweet A9 in Table 1).
Citation: Weather, Climate, and Society 10, 3; 10.1175/WCAS-D-17-0126.1
The mandatory evacuation order was a salient piece of information for many of the Twitterers (Table 2). Mayor Bloomberg held a press conference to notify people about New York City’s evacuation order, which included all of Far Rockaway, shortly before 1530 UTC on 28 October. Four people in our sample tweeted about the evacuation order within one minute of the announcement (B1–B4), and over one-fifth tweeted about it within three hours; many others followed suit in the subsequent hours. Most of these Twitterers also conveyed that they were considering what the evacuation order meant for them personally, for instance, through indications of being displeased about needing to leave (B3–B7) as well as barriers they face in doing so (B3, B6). Although some, including those who tweeted about the mandatory evacuation order, ultimately decided not to evacuate for various reasons (see section 4b), expressions of defiance of the evacuation order (B8) were rare.
Social cues, that is, observations of others’ behavior and other information from the social environment, have long been recognized to play an important role in people’s risk assessments (Mileti and Sorensen 1990; Krimsky and Golding 1992; Renn 2008; Lindell and Perry 2012). Indeed, they are a central tenet of the social amplification of risk framework, which theorizes that risk assessments are amplified or attenuated through, among other things, social “stations” (e.g., opinion leaders, personal networks, organizations), which can affect the salience of a risk through the volume of information and interpretations about it (Kasperson et al. 1988, 2005). However, it is not well understood what social cues people find most significant when hazards threaten and how these cues influence risk assessments. This is true in general, and especially in the social media context, which expands and adds complexity to this dynamic social space.
Three key types of social cues emerged as prevalent in our data analysis (Table 3). The first type of social cue, and that perhaps most commonly described in the literature, is cues from peers (family, friends, neighbors, and others). These cues comprise protective and preparatory actions others are taking, such as boarding up one’s home (C1), purchasing supplies (C2), or more general behaviors (e.g., “evrybdy is goin krzy” in C3). These cues also include information about others’ perceptions, such as the lack of concern conveyed to Twitterer34 by his neighbor who “rode out” Hurricane Irene the previous year (C4). Although most tweets about peer cues are mentions of what others are thinking and doing (descriptive norms), some are messages from the tweet authors that aim to cue others about what they should be thinking and doing (injunctive norms). For instance, on the morning of landfall, Twitterer38 directs people to “Get out of #farrockaway before it’s too late!” along with a photo (not shown) of flooded roads (C5).
A second type of social cue is cues from businesses, such as the closing of coffee shops, stores, and restaurants for Sandy (C6–C8). These types of cues often were important amplifiers of the risk, as indicated by comments like “You know Rockaway’s in trouble when Pickles and Pies is closed” (C7), especially if the Twitterers experienced that business staying open a year prior during Hurricane Irene (C8).
Third, protective steps taken by local jurisdictions served as cues from government for many. Examples include closing public areas such as the boardwalk (C9), suspending public transportation (C10, C11), closing bridges (C12), and shutting off utilities (C13, C14). As the latter tweets indicate, such governmental actions spurred feelings of anger and distrust for some (see also section 4b; Anderson et al. 2016; Lazrus et al. 2017).
As with social cues, environmental cues are recognized to play important roles in people’s assessments of environmental risks (Taylor et al. 2009; Lindell and Perry 2012; Lazrus et al. 2016; Demuth 2018). However, less is known about which environmental cues are important, when, and how. This perhaps is because environmental cues do not manifest regularly or clearly for all types of risks. Weather hazards, however, intrinsically present such cues, and more than half of the Far Rockaway users tweet about them. Thus, the social media data analyzed here offer insight into the salience of such cues to people as they assess a threat (Table 4).
Many people tweeted about natural environment cues related to Sandy, most commonly the strong or strengthening winds (D1–D8). Often they mentioned wind along with related cues, such as the rough ocean (D3), the smell of ocean spray (D4), falling trees (D5), and animal behavior (D6). Others mentioned different natural cues, like rain (D7) and darkness (D8). Built environment cues also were important indicators to some, as water poured from street drains (D9), roads were flooded (D10; Fig. 5), and especially as power flickered or went out (D11, D12). Most environmental cues mentioned in these data were observed by the person tweeting, although some were from information relayed by others. Mentions of environmental cues often were associated with negative emotions, which are further discussed in section 4b.
Photo shared by Twitterer30 of flooding as an environment cue (see tweet D10 in Table 4).
Citation: Weather, Climate, and Society 10, 3; 10.1175/WCAS-D-17-0126.1
As might be expected given the different types of risk information available at different times leading up to Sandy (section 3a), the information that people tended to tweet about evolved with the threat. Mentions of weather forecast information began on 25 October, were common on 26–27 October, and continued until landfall as people tweeted about where Sandy was and when it would make landfall with increasing specificity. Mentions of evacuation orders primarily clustered on 28 October, the day the mandatory order was issued for Far Rockaway. Mentions of social cues were most common around the time of the evacuation order, as the risk became more certain, and as more people considered and engaged in preparatory and protective actions; social cue mentions also extended into 29 October, the day of landfall. Environmental cues were mentioned by a few Twitterers on 28 October, but most tweets about environmental cues were on the day of landfall, and they increased in frequency as Sandy approached.
b. Risk perceptions and responses
Although risk perception and response typically are parsed theoretically, we found that these processes often are indistinguishable in the Twitter narratives, especially as the hazard approaches. Thus, we combine discussion of them into this section. The data analysis reveals that, as people gathered information about and assessed the evolving threat, several aspects of their cognitive and affective risk perceptions emerged as important, along with their emotional, preparatory and protective, and coping responses (Fig. 3).
One aspect of people’s cognitive risk perceptions that emerged was their perceived exposure to hurricanes (Table 5), that is, their beliefs about the natural and built characteristics of the environment that influence how they could be affected by a hurricane (Zhang et al. 2004; Wilhelmi and Hayden 2010; Lazo et al. 2015; Morss et al. 2016). Many Twitterers referenced their perceived geographical exposure. Some indicated that they thought their location’s elevation (E1) or proximity to the ocean (E2) put them at greater risk. Others thought that their location was at lower risk, for example, because it “never floods” (E3) or has not in previous storms (E4). Some Twitterers discussed their perceived vertical exposure, based on what floor in a building they reside on (e.g., perceived lower risk above the first floor; E5, E6). And, some discussed their perceived structural exposure based on the type of building they reside in. In the Far Rockaway Twitter data, this was often expressed through the notion that being in a “building” (E7–E9), meaning public housing provided by the New York City Housing Authority, was safer than a single-family home or other smaller structure (Lazrus et al. 2017).
Our analysis of the Twitter narratives also revealed a second type of cognitive judgment: people’s perceived severity of Sandy and its impacts (Table 6). These perceptions were expressed through many of the forecast information tweets discussed in the previous section, such as A1 (“we’re looking to get pretty wet on Monday”) and A7 (“Sandy expected to bring life-threatening storm surge flooding”). Another example is F1 (which includes Instagram text associated with the tweet) downplaying the intensity of Sandy and suggesting that the hurricane might pose more harm if it were stronger (“a buck 10”, i.e., with 110 mile-per-hour winds). Others referred to perceptions of Sandy’s severity in terms of potential impacts to their home (F2) and to Far Rockaway (F3), including concern that the area might “look like new orleans after Katrina” (F4). Some described the potential for specific negative impacts, such as their home being “underwater” (F2), but most people’s expressions of the possible severity were vague, holistic, and idiomatic, such as through comments about Sandy getting “real” (F5, F6), “crazy” (F7), or being no joke (F8). These tweets and those in the previous paragraph reveal the range of ways that people think about and express their views about how likely they are to experience a risk and how bad it could be.
In addition to cognitive risk perceptions, our analysis revealed several important aspects of people’s affective responses to the hurricane threat, as well as specific emotion states (Table 7). The majority of expressions of affect and emotion were negative, including worry, fear, anger, and other unspecified negative emotions. Worry often emerged a day or more prior to landfall, which the context of the Tweet narratives indicate is due to the prospect of the negative effects from Sandy (G1, G2). As Sandy approached and made landfall, many people indicated fear, either implicitly (G3, G4) or explicitly (G5–G7). Fear was often spurred by environmental cues, including strong winds, power loss, and flooding (G4–G7). For instance, Twitterer34 sends, in rapid succession, a series of tweets expressing surprise and negative emotions along with a photo as floodwaters reach his home (G4, Fig. 6). Another commonly expressed emotion was anger, prompted, for example, by perceived poor decision-making by officials (G8; see also section 4a), jokes from others (G9), and having to evacuate (G10).
Photo shared by Twitterer34 of flooding outside his front gate as Sandy made landfall (see tweet G4 in Table 7).
Citation: Weather, Climate, and Society 10, 3; 10.1175/WCAS-D-17-0126.1
The Twitter narratives also included expressions of nonnegative affect and emotion, such as excitement about the hurricane and humor (G11–G14). As discussed in Parkhill et al. (2011), humor is one way that people cope when faced with risks. Thus, while humorous comments about Sandy may seem to be dismissive of the risk (Knox et al. 2016), they may also be an indication of risk perceptions and help reveal important aspects of “the ways in which people experience and live with risk” (Parkhill et al. 2011, p. 352). Other ways of coping are discussed below.
As discussed in section 2, dozens of studies have examined people’s behavioral responses to hurricane risks. Many of the actions to prepare and protect life and property that are well known from the literature were seen in the Far Rockaway Twitter data (Table 8). These include references to actions taken, such as boarding up one’s home (H1), gathering supplies (H2), preparing to leave (H3, H4), and evacuating (H5–H7), as well as actions not taken, such as decisions not to evacuate (H8–H10). Although some Twitterers explicitly say that they are or are not evacuating, other mentions of evacuation decisions are less apparent and require the user’s narrative to accurately determine the meaning. For example, “Rolling deep with minivans right now. #escapefromrockaway” (H5), is the way that Twitterer49 conveys he is evacuating, but his other tweets, including those after Sandy makes landfall, are needed to ascertain this. The value of having a user’s narrative is further shown in section 4c.
In addition to these protective actions commonly discussed in the literature, additional types of preparatory and protective behaviors emerged, especially among nonevacuees. For instance, to prepare to be without utilities, Twitterer34 trimmed his beard (i.e., “trimmed my ‘Alan Moore,’” H11), and Twitterer6 showered while hot water was still available (H12). As Sandy approached and conditions worsened, these same two people took additional protective actions by moving the family dog into a bathroom away from windows (H13) and by pumping flood waters out of the basement (H14). After losing power due to the storm, others tweeted about preserving computer and mobile phone communication capabilities by powering off devices (H15).
Although coping in the aftermath of a disaster is well chronicled, it is less frequently studied in the predisaster phase. Our analysis revealed several ways in which people coped with Sandy’s evolving threat before and during landfall (Table 9). In the days leading up to Sandy’s landfall, some people engaged in leisure activities, such as having a party (I1), watching sports (I2, I3), and enjoying the outdoors (I4). As the storm approached and made landfall, some mentioned spending the time stuck indoors by cooking, eating, drinking, or binge watching television or movies (I5). Others coped through prayer (I6, I7) or found ways to have fun or distract themselves (I8). As discussed above, people also used humor to cope with Sandy’s threat, from a few days before up to landfall (G11–G14).
c. Individuals’ evolving information use, risk assessments, and responses
The analysis presented in sections 4a and 4b illustrates the different, nuanced ways that people interacted with risk information, perceived their risk, and responded as Sandy approached and arrived. In this section, we explore in greater depth the complex, interwoven nature of those processes and their evolution by presenting and discussing segments of four of the Twitter narratives. These four narratives were selected to illustrate the dynamic ways in which different people attended to different risk information, interpreted and responded to the risk, and made decisions, including how these processes interacted with the unique context of people’s lives. Each sequence of tweets discussed is only a subset of that person’s full tweet narrative; all tweets are shown for the period of time selected unless otherwise indicated.
Twitterer33 is an example of someone who indicates early awareness about the risk of Sandy; as the threat evolves, he attends closely to weather information, evacuation orders, and social and environmental cues, and then evacuates (Table 10). His Sandy-relevant tweets begin four days before landfall, with a retweet of information from Mayor Bloomberg about potential evacuation (J1). The next day, he indicates that he has been actively seeking and obtaining weather information (J2). Over the next few days, he tweets almost exclusively about Sandy, with increasing frequency, sharing photos that document his experience. The morning of 28 October, he anticipates an impending evacuation order based on the evacuation of nearby areas (J9), and then he tweets about the mayor’s evacuation order for his area during the press conference (J11). Twenty minutes later, in conversation with another Twitterer, he tweets about packing and his plans to evacuate (J13). After several tweets about social cues (J14–18), he then tweets that he is evacuating, less than three hours after Bloomberg announced the evacuation order (J21, J22, J24), using a car that he purchased the week prior (J24). His tweets also reveal coping behaviors through humor (J12) and purchasing beer (J26). After evacuating, he continues to seek information about Sandy by watching news coverage (J28, Fig. 7).
Photo shared by Twitterer33 of the television news coverage he is watching of his neighborhood after evacuating (see tweet J28 in Table 10).
Citation: Weather, Climate, and Society 10, 3; 10.1175/WCAS-D-17-0126.1
Twitterer53 is a young adult who is less focused on risk information than Twitterer33 and has different concerns and constraints, but who also evacuates (Table 11). The segment of her narrative shown begins shortly after the mayor announces the evacuation order for her area. Although she does not mention forecast information and makes only a vague reference to the evacuation order, her Twitter narratives reveal that she recognizes her risk from Sandy. For example, she tweets four times in 13 minutes about her perceived exposure, including the exposure of different locations relative to the ocean and different types of structures (K1–K4). The risk perception content in these tweets is interspersed with consideration about where she should evacuate to, at times expressing humor and at times frustration. She then tweets that she is evacuating to her brother’s dormitory (K5) with her mother (K7). Over the next few hours, she tweets about packing to evacuate (K8, K17) and social cues in the form of class cancellations (K12), interwoven with tweets about past hurricane experience, evacuation decision-making, hurricane risks, and associated uncertainty (K10, K19). Although she expresses concern about potentially evacuating unnecessarily (K10), she twice says that she would rather be safe than sorry (K10, K20). After a short delay waiting for her mother to leave work (K21), she tweets that they evacuate Far Rockaway that evening, noting the environmental cues that indicate worsening conditions (K23–K26).
Twitterer48 is aware of and worried about the risk of Sandy several days before landfall, although she does not explicitly reference forecasts or other official risk information. Initially she decides not to evacuate, but she then changes her mind and moves somewhere safer hours before landfall (Table 12). She first tweets about Sandy on 26 October, revealing that she is aware of Sandy’s threat and expressing (through an emoji) negative affect (L1). A few minutes later, another tweet reveals her awareness that Sandy is forecast to make landfall on “Monday,” 29 October and her associated risk perceptions in the form of Far Rockaway’s exposure (L2). She then tweets almost exclusively about topics other than Sandy, until 28 October. That morning begins with her tweeting in conversation with someone about having difficulty sleeping, which she says is probably due to worry about Sandy (L3, L4). Several hours later, she retweets information about a governmental social cue (suspension of local public transportation, L6), and then tweets about her decision not to evacuate (L8). After several tweets unrelated to Sandy, on 29 October (the morning of landfall), she tweets about environmental and social cues: the large ocean waves (L10), a photo of herself on the beach (L11, photo not shown), and a photo of a boarded-up gas station surrounded by flooded streets (L12, Fig. 8). Within a few minutes, her Twitter narrative reveals that she has moved to a relative’s home in a “projects building,” which she perceives as structurally safer (L13). Although we do not know for certain, this sequence of tweets suggest that her decision to move was triggered to some extent by these environmental and social cues and associated affect (e.g., symbols of a person praying in L11). After returning home briefly (L15), she gets settled at her relative’s home (L18) and tweets about food, Sandy’s winds, and concern about Far Rockaway as the storm arrives (L19–L21).
Photo shared by Twitterer48 of flooding and a boarded-up gas station (see tweet L12 in Table 12).
Citation: Weather, Climate, and Society 10, 3; 10.1175/WCAS-D-17-0126.1
Twitterer6 draws on his past experience with damage from a hurricane, presumably from Hurricane Irene from the previous year, and he decides not to evacuate for Sandy but again suffers losses (Table 13). His first Sandy-relevant tweet comes about two hours after the mayor’s announcement of the evacuation order. In conversation with several other Twitterers, he tweets about deciding whether to leave or stay (M1, M2). He does not explicitly reference forecast information or the evacuation order; however, he tweets about uncertainty in official information based on his experience “last time [when] it wasn’t as bad as they said,” and says he thinks he could have prevented damage had he not evacuated then (M2). Although Twitterer6 does not explicitly say that he decides not to evacuate, his subsequent tweet narrative reveals that he indeed stayed. As Sandy approaches and makes landfall, he tweets about protective actions, including pumping water out of his basement (M6), and natural and built environmental cues, including winds, flooding, and losing power (M8–M10). Several of his tweets during this period suggest concern. For example, immediately after tweeting about his observations of the “waist high n rising” flooding and associated car lights and alarms (M10), he tweets that he should be okay because he lives on the second floor and has food and water (M11). Eight hours pass before he tweets again, after landfall, at which time he tweets about the extent of Sandy’s flooding and impacts (M12, M13, M16, M17). These observations are accompanied by expressions of positive coping, including comments that he and others are okay (M14, M16) and that he is “surviving” (M17).
These examples reveal the ways that the constructs in Fig. 3 evolved and interacted for these at-risk people as the hurricane threat unfolded. The discussion also illustrates how people’s experiences with a threat in the context of their lives can be revealed through social media narratives in ways that go beyond the text of individual posts.
5. Summary and discussion
This paper develops new understanding of people’s thoughts and behaviors about the risks they faced leading up to and during a hurricane landfall through analysis of social media data. Specifically, the findings are based on qualitative analysis of Twitter narratives collected from 53 people who were in Far Rockaway, New York, an area that was under a mandatory evacuation order from Hurricane Sandy. This analysis of people who were at significant risk of harm reveals the different types of risk information that they attended to and their risk perceptions and responses as the threat evolved, viewed through the lens of their tweets. The analysis is informed by the relevant existing literature, but it is grounded in the Twitterers’ own words and images shared, which were posted at times and in ways that are meaningful to them.
Some of the findings echo those from past research, but with an added texture from the context that the Twitter narratives lend. Other findings offer novel perspectives about the risk assessments and decisions people make. We note which findings match those from previous research on weather hazards and which are new in our elaboration below of the key results. Overall, the results contribute both theoretically and practically to our understanding of people who are at risk of an extreme weather event.
Very few of the 53 people in our sample explicitly tweeted about Sandy’s forecast sources or attributes, and the few explicit references to forecasts were composed mostly of retweeted factual information. Most people, however, tweeted before the storm arrived about their neighborhood or their home being threatened, the timing of landfall, or the possible impacts of Sandy. This reveals that most users had received forecast information about the threat through some means. This finding is similar to that from Meyer et al. (2014), who found through phone surveys with people threatened by Sandy that every respondent was aware of the hurricane threat but that many were unaware of specific types of forecast information. People’s implicit forecast references further tended to include some mention of what the threat meant for them, meaning they personalized the risk in some way.
These findings provide a new perspective on how people access and consider forecast information, compared to past studies that focus on characterizing what forecast information people receive (e.g., Lazo et al. 2015; Sherman-Morris 2013; Stein et al. 2010; Taylor et al. 2009; Zhang et al. 2007). They also suggest a possible mismatch between the ways that experts—including weather forecasters and researchers—think about and measure at-risk populations’ attention to a threat versus the ways that people actually are attuned to the threat. Measuring the types of forecast information people received may be less meaningful than measuring what aspects of the forecast threat they interpreted as applicable to them, particularly as it might negatively impact them.
Many people in our sample paid attention to the mandatory evacuation order that included all of Far Rockaway. Moreover, most people who tweeted about the evacuation order also tweeted about what it meant for them, even if they ultimately did not evacuate because of other reasons. This suggests that a mandatory evacuation order serves as an important risk informational cue for many people. This finding corroborates results from other studies that have used surveys and interviews to examine evacuation responses to hurricane risk messages (e.g., Baker 1991; Gladwin et al. 2001; Lindell et al. 2005; Dash and Gladwin 2007; Zhang et al. 2007; Taylor et al. 2009; Morss and Hayden 2010; Huang et al. 2012; Cuite et al. 2017). It further suggests, though, that a mandatory evacuation order has resonance even in the vast constellation of information that is now available during an evolving hurricane threat.
In addition to “official” forecasts and evacuation orders (i.e., risk information issued by public authorities), “unofficial” information in the form of social and environmental cues emerged as particularly salient and influential to people. Businesses closing and certain governmental cues, such as closing roads or suspending public transportation, emerged as types of social cues that tended to heighten people’s risk perceptions. Also, many people tweeted, in some cases frequently, about natural and built environmental cues as Sandy approached and made landfall. These cues often trigged concern or fear, and they motivated—and in some cases changed—protective decisions for some people. Although social and environmental cues are known to influence how people judge and respond to risks (e.g., Lindell and Perry 2012), they are typically not discussed in much detail or depth in the literature on hurricane risk communication and decision making. The social media narratives allow us to “see” the types of information that are salient to people and their breadth, extent, and power in the context of a weather risk in ways that past research utilizing other methods and datasets has not afforded.
The analysis also reveals important aspects of how people perceive their general risk from hurricanes as well as their storm-specific risk. People’s risk perception-related tweets include assessments of their exposure to harm based on where they live geographically (i.e., proximity to the ocean), vertically (i.e., what floor they live on), and structurally (i.e., what kind of building they reside in). Two common ways of parsing and measuring people’s hurricane risk perceptions are as exposure and severity (Lazo et al. 2015; Morss et al. 2016, 2018; Rickard et al. 2017), but these three aspects of perceived exposure revealed by our analysis indicate the nuanced ways that people think about their risk.
The analysis also reveals the ways in which these nuanced risk perceptions can influence people’s protective action decision-making. For instance, people who mentioned they are safe because they live on the second floor are at least implicitly considering their risk of flooding due to rain or surge (even if they do not explicitly mention it). Yet, they may not be accounting for uncertainty in the hazard that puts them at risk—as evidenced by Twitterer34, who expressed surprised when the flood waters reached his doorstep (G4, Fig. 6)—nor for secondary effects (e.g., lack of utilities, being stranded due to damaged transportation infrastructure). Also, people who evacuate to a stronger structure may perceive more of a risk from wind than surge. In order for authorities to create risk communication messages that encourage appropriate protective action for people at risk of an approaching hurricane, they need to understand such perceptions—that is, what people believe they are at risk from—and the Twitter narratives help illustrate this.
The Twitterers we analyzed also conveyed their risk perceptions in terms of the severity of the risk. Earlier in the Sandy timeline, these mentions tended to be related to forecast information and were more specific about the type or magnitude of the projected impacts (e.g., “life-threatening storm surge”). As landfall approached, tweets about the storm severity tended to be tied to environmental cues and were more all-encompassing (e.g., that the storm was “real” or “crazy”). Taken together, the ways that the people in our sample discuss their general and storm-specific cognitive risk perceptions suggests ways that this important construct might be further theorized and measured in future studies, especially across the timeline of an event (see also Lin et al. 2014; Meyer et al. 2014).
Affect is well understood in the risk literature to contribute to people’s risk perceptions and responses (Slovic et al. 2004; Greenberg et al. 2012), but, as Peters et al. (2004, p. 1352) note, “public reaction to hazards can include more complex feelings than good or bad.” Indeed, in addition to positive or negative affect, the Far Rockaway Twitterers expressed specific emotions about the risk of Sandy. The emotions typically were negative, such as fear and anger. In some cases, these emotions represent heightening risk perception, which triggered protective behaviors. However, in other cases, the emotions represent responses themselves.
Some research has examined how specific emotions, like fear and anger, influence people’s risk assessments (Peters et al. 2004; Lerner and Keltner 2001; Lerner et al. 2015; Lindell et al. 2016), but overall, this is an understudied area. For example, Keller et al. (2012) identified the importance of research to understand “what are the relevant, specific emotions” (p. 249) and “what are the consequences of specific emotions for decision-making and behavior” (p. 250) for particular environmental risks, such as hazardous weather. People’s use of social media to express their emotions during disaster events has been noted by some (Houston et al. 2015), but most research on this topic thus far has focused on studying postevent mourning and memorializing. The data presented here reveal that people express a range of different emotions that evolve in complex ways as a threat unfolds. This suggests that further study of this topic is needed, and it suggests that analysis of social media data can help build understanding about the roles that different emotions play in how people perceive and respond to risks.
As discussed in sections 4b and 4c, the analysis illustrates some of the ways that people at risk used evolving information to evaluate the risk that Sandy posed to them personally and to make decisions about evacuating or taking other preparatory actions. In addition to taking protective actions, the data revealed that people engaged in other types of behaviors to help them cope with the threat of the approaching hurricane. Examples include “everyday” distractions (e.g., watching a sporting event, eating, and spending time with family) that took on new meaning during Sandy, as well as behaviors inspired by the threat (e.g., praying, expressing humor, finding fun activities while evacuated or sheltering in place). Moreover, people’s risk perceptions were not heightened and continuously maintained in an elevated state as the hurricane approached. Rather, people found ways to process and manage their concerns about the risk through these coping behaviors (as well as protective behaviors). Thus, another important area for future work is developing deeper understanding of the dimensions and functions of such emotion- and meaning-based coping behaviors (Folkman 1997; Parkhill et al. 2011).
Methodologically, the analysis presented here demonstrates how social media data can be sampled and analyzed to richly investigate how people experience and respond to an evolving weather risk. This is useful given that interest is rapidly growing in more fully leveraging the power of “big data,” such as that from social media platforms like Twitter, for risk-related research (NASEM 2017). Yet, doing so requires careful consideration about who is the population of interest, how to identify and reach them, and how to gather meaningful data from them, all framed within an understanding of the relevant literatures. Designing such studies of social media data are time intensive and come with challenges, however. Most studies that utilize Twitter data rely solely on keyword searches to draw a sample. To address research questions such as those examined here, such types of sampling approaches alone are insufficient because they can miss relevant tweets if people abbreviate or misspell words, tweet phonetically, or tweet without mentioning any of the keywords used. Moreover, interesting and relevant content often lies in nontextual or supplemental content, such as emoticons/emojis, pictures, or links to other social media platforms. Further, most of the interesting constructs that pertain to people’s risk communication, perceptions, and responses are latent, multidimensional, and require context for full understanding. The in-depth analyses like those done here, though, provide essential insight about what can (and cannot) be “found” in the data so that these concepts and processes can be examined at larger scales (e.g., using machine learning and other natural language processing techniques) in order to determine which findings generalize across events, populations, and time. In short, returning to Palen and Anderson’s (2016, p. 225) words, “mak[ing] ‘Big’ data bigger, then smaller” offers potential to make it meaningfully bigger again.
In summary, this research illustrates the complex and dynamic ways that people interact with risk information, think and feel about risks, and respond. It does so from the perspective of individuals at risk as told in their voices through Twitter. Such understanding is vital to developing effective risk communication and preparedness and response policies to reduce harm from future weather threats.
Acknowledgments
The authors thank Jennifer Boehnert for designing Fig. 2. This research is supported by National Science Foundation Award 1331490. The National Center for Atmospheric Research is sponsored by the National Science Foundation.
REFERENCES
Anderson, J., and et al. , 2016: Far far away in Far Rockaway: Responses to risks and impacts during Hurricane Sandy through first-person social media narratives. Proc. 13th Int. Conf. on Information Systems for Crisis Response and Management, Rio de Janeiro, Brazil, ISCRAM, 16 pp., http://idl.iscram.org/files/jenningsanderson/2016/1388_JenningsAnderson_etal2016.pdf.
Anderson, K. M., and A. Schram, 2011: Design and implementation of a data analytics infrastructure in support of crisis informatics research. Proc. 33rd Int. Conf. on Software Engineering, Honolulu, HI, IEEE, 844–847, https://doi.org/10.1145/1985793.1985920.
Anderson, K. M., A. Schram, A. Alzabarah, and L. Palen, 2013: Architectural implications of social media analytics in support of crisis informatics research. IEEE Bulletin of the Technical Committee on Data Engineering, Vol. 36, No. 3, IEEE Computer Society, Washington, DC, 13–20, http://sites.computer.org/debull/A13sept/A13SEP-CD.pdf.
Baker, E. J., 1991: Hurricane evacuation behavior. Int. J. Mass Emerg. Disasters, 9, 287–310, http://www.ijmed.org/articles/412/.
Bica, M., and J. Anderson, 2016: “You are what you tweet!”: The ethics of (re) publishing public data as crafted narratives. Workshop on Ethical Encounters in HCI Research/ACM Conf. on Human Factors in Computing Systems, San Jose, CA, Association for Computing Machinery, 7 pp., https://ethicalencountershci.files.wordpress.com/2016/03/bica-and-anderson.pdf.
Boyd, D., and K. Crawford, 2012: Critical questions for big data: Provocations for a cultural, technological, and scholarly phenomenon. Inf. Commun. Soc., 15, 662–679, https://doi.org/10.1080/1369118X.2012.678878.
Brynielsson, J., F. Johansson, and A. Westling, 2013: Learning to classify emotional content in crisis-related tweets. IEEE Int. Conf. on Intelligence and Security Informatics, Seattle, WA, IEEE, 33–38, https://doi.org/10.1109/ISI.2013.6578782.
Cates, A. L., and et al. , 2013: Impact of dual-polarization radar technology and twitter on the Hattiesburg, Mississippi tornado. Disaster Med. Public Health Prep., 7, 585–592, https://doi.org/10.1017/dmp.2013.113.
Christensen, L., and C. E. Ruch, 1980: The effect of social influence on response to hurricane warnings. Disasters, 4, 205–210, https://doi.org/10.1111/j.1467-7717.1980.tb00273.x.
Creswell, J. W., 2007: Qualitative Inquiry and Research Design: Choosing Among Five Approaches. Sage Publications, 472 pp.
Cuite, C. L., R. L. Shwom, W. K. Hallman, R. E. Morss, and J. L. Demuth, 2017: Improving coastal storm evacuation messages. Wea. Climate Soc., 9, 155–170, https://doi.org/10.1175/WCAS-D-16-0076.1.
Dash, N., and H. Gladwin, 2007: Evacuation decision making and behavioral responses: Individual and household. Nat. Hazards Rev., 8, 69–77, https://doi.org/10.1061/(ASCE)1527-6988(2007)8:3(69).
Demuth, J. L., 2018: Explicating experience: Development of a valid scale of past hazard experience for tornadoes. Risk Anal., https://doi.org/10.1111/risa.12983, in press.
Demuth, J. L., R. E. Morss, J. K. Lazo, and C. Trumbo, 2016: The effects of past hurricane experiences on evacuation intentions through risk perception and efficacy beliefs: A mediation analysis. Wea. Climate Soc., 8, 327–344, https://doi.org/10.1175/WCAS-D-15-0074.1.
Folkman, S., 1997: Positive psychological states and coping with severe stress. Soc. Sci. Med., 45, 1207–1221, https://doi.org/10.1016/S0277-9536(97)00040-3.
Gladwin, C. H., H. Gladwin, and W. G. Peacock, 2001: Modeling hurricane evacuation decisions with ethnographic methods. Int. J. Mass Emerg. Disasters, 19, 117–143, http://www.ijmed.org/articles/479/.
Greenberg, M., C. Haas, A. Cox Jr., K. Lowrie, K. McComas, and W. North, 2012: Ten most important accomplishments in risk analysis, 1980–2010. Risk Anal., 32, 771–781, https://doi.org/10.1111/j.1539-6924.2012.01817.x.
Houston, J. B., and et al. , 2015: Social media and disasters: A functional framework for social media use in disaster planning, response, and research. Disasters, 39, 1–22, https://doi.org/10.1111/disa.12092.
Huang, S.-K., M. K. Lindell, C. S. Prater, H. C. Wu, and L. K. Siebeneck, 2012: Household evacuation decision making in response to Hurricane Ike. Nat. Hazards Rev., 13, 283–296, https://doi.org/10.1061/(ASCE)NH.1527-6996.0000074.
Huang, S.-K., M. K. Lindell, and C. S. Prater, 2016: Who leaves and who stays? A review and statistical meta-analysis of hurricane evacuation studies. Environ. Behav., 48, 991–1029, https://doi.org/10.1177/0013916515578485.
Hughes, A., and L. Palen, 2014: Social media and emergency management: An academic’s perspective. Critical Issues in Disaster Science and Management: A Dialogue between Scientists and Emergency Managers, J. E. Trainor and T. Subbio, Eds., FEMA Higher Education Project, 350–363.
Kasperson, J. X., R. E. Kasperson, N. Pidgeon, and P. Slovic, 2005: The social amplification of risk: Assessing 15 years of research and theory. Publics, Risk Communication, and the Social Amplification of Risk, Vol. I, The Social Contours of Risk, R. E. Kasperson and J. Kasperson, Eds., Routledge, 202–229.
Kasperson, R. E., O. Renn, P. Slovic, H. S. Brown, J. Emel, R. Goble, J. X. Kasperson, and S. Ratick, 1988: The social amplification of risk: A conceptual framework. Risk Anal., 8, 177–187, https://doi.org/10.1111/j.1539-6924.1988.tb01168.x.
Keller, C., A. Bostrom, and M. Kuttschreuter, 2012: Bringing appraisal theory to environmental risk perception: A review of conceptual approaches of the past 40 years and suggestions for future research. J. Risk Res., 15, 237–256, https://doi.org/10.1080/13669877.2011.634523.
Knox, J. A., B. Mazenec, E. Sullivan, S. Hall, and J. A. Rackley, 2016: Analysis of the Twitter response to Superstorm Sandy: Public perceptions, misconceptions, and reconceptions of an extreme atmospheric hazard. Atmospheric Hazards, J. Coleman, Ed., Intech, 21–40, https://doi.org/10.5772/64019.
Kogan, M., L. Palen, and K. Anderson, 2015: Tweet local, retweet global: Retweeting by the geographically-vulnerable during Hurricane Sandy. Proc. 18th Conf. on Computer Supported Cooperative Work and Social Computing, Vancouver, BC, Canada, Association for Computing Machinery, 981–993, https://doi.org/10.1145/2675133.2675218.
Krimsky, S., and D. Golding, 1992: Social Theories of Risk. Praeger, 424 pp.
Kryvasheyeu, Y., H. Chen, N. Obradovich, E. Moro, P. Van Hentenryck, J. Fowler, and M. Cebrian, 2016: Rapid assessment of disaster damage using social media activity. Sci. Adv., 2, e1500779, https://doi.org/10.1126/sciadv.1500779.
Lachlan, K. A., P. R. Spence, X. Lin, and M. D. Greco, 2014a: Screaming into the wind: Examining the volume and content of tweets associated with Hurricane Sandy. Commun. Stud., 65, 500–518, https://doi.org/10.1080/10510974.2014.956941.
Lachlan, K. A., P. R. Spence, X. Lin, K. M. Najarian, and M. D. Greco, 2014b: Twitter use during a weather event: Comparing content associated with localized and nonlocalized hashtags. Commun. Stud., 65, 519–534, https://doi.org/10.1080/10510974.2014.956940.
Lazo, J. K., A. Bostrom, R. E. Morss, J. L. Demuth, and H. Lazrus, 2015: Factors affecting hurricane evacuation intentions. Risk Anal., 35, 1837–1857, https://doi.org/10.1111/risa.12407.
Lazrus, H., R. E. Morss, J. L. Demuth, J. K. Lazo, and A. Bostrom, 2016: “Know what to do if you encounter a flash flood”: Mental models analysis for improving flash flood risk communication and public decision making. Risk Anal., 36, 411–427, https://doi.org/10.1111/risa.12480.
Lazrus, H., O. Wilhelmi, J. Henderson, R. E. Morss, and A. Dietrich, 2017: Information as intervention: How can hurricane risk communication reduce vulnerability? 12th Symp. on Societal Applications: Policy, Research, and Practice, Seattle, WA, Amer. Meteor. Soc., 9B.2, https://ams.confex.com/ams/97Annual/webprogram/Paper315976.html.
Lerner, J. S., and D. Keltner, 2001: Fear, anger, and risk. J. Pers. Soc. Psychol., 81, 146–159, https://doi.org/10.1037/0022-3514.81.1.146.
Lerner, J. S., Y. Li, P. Valdesolo, and K. S. Kassam, 2015: Emotion and decision making. Annu. Rev. Psychol., 66, 799–823, https://doi.org/10.1146/annurev-psych-010213-115043.
Lin, C.-C., L. K. Siebeneck, M. K. Lindell, C. S. Prater, H. C. Wu, and S. K. Huang, 2014: Evacuees’ information sources and reentry decision making in the aftermath of Hurricane Ike. Nat. Hazards, 70, 865–882, https://doi.org/10.1007/s11069-013-0853-1.
Lindell, M. K., 2012: Response to environmental disasters. The Oxford Handbook of Environmental and Conservation Psychology, S. D. Clayton, Ed., Oxford University Press, 391–412.
Lindell, M. K., and R. W. Perry, 2012: The Protective Action Decision Model: Theoretical modifications and additional evidence. Risk Anal., 32, 616–632, https://doi.org/10.1111/j.1539-6924.2011.01647.x.
Lindell, M. K., C. J. Lu, and C. S. Prater, 2005: Household decision making and evacuation in response to Hurricane Lili. Nat. Hazards Rev., 6, 171–179, https://doi.org/10.1061/(ASCE)1527-6988(2005)6:4(171).
Lindell, M. K., C. S. Prater, H.-C. Wu, S.-K. Huang, D. M. Johnston, J. S. Becker, and H. Shiroshita, 2016: Immediate behavioral responses to earthquakes in Christchurch New Zealand and Hitachi Japan. Disasters, 40, 85–111, https://doi.org/10.1111/disa.12133.
Meyer, R. J., K. Broad, B. Orlove, and N. Petrovic, 2013: Dynamic simulation as an approach to understanding hurricane risk response: Insights from Stormview lab. Risk Anal., 33, 1532–1552, https://doi.org/10.1111/j.1539-6924.2012.01935.x.
Meyer, R. J., J. Baker, K. Broad, J. Czajkowski, and B. Orlove, 2014: The dynamics of hurricane risk perception: Real-time evidence from the 2012 Atlantic hurricane season. Bull. Amer. Meteor. Soc., 95, 1389–1404, https://doi.org/10.1175/BAMS-D-12-00218.1.
Mileti, D. S., and J. H. Sorensen, 1990: Communication of emergency public warnings. Rep. ORNL-6609, 162 pp., https://doi.org/10.2172/6137387.
Morss, R. E., and M. H. Hayden, 2010: Storm surge and “certain death”: Interviews with Texas coastal residents following Hurricane Ike. Wea. Climate Soc., 2, 174–189, https://doi.org/10.1175/2010WCAS1041.1.
Morss, R. E., J. L. Demuth, J. K. Lazo, K. Dickinson, H. Lazrus, and B. H. Morrow, 2016: Understanding public hurricane evacuation decisions and responses to hurricane risk messages. Wea. Forecasting, 31, 395–417, https://doi.org/10.1175/WAF-D-15-0066.1.
Morss, R. E., and et al. , 2017: Hazardous weather prediction and communication in the modern information environment. Bull. Amer. Meteor. Soc., 98, 2653–2674, https://doi.org/10.1175/BAMS-D-16-0058.1.
Morss, R. E., C. L. Cuite, J. L. Demuth, W. K. Hallman, and R. L. Showm, 2018: Is storm surge scary? The influence of hazard, impact, and fear-based messages and individual differences on responses to hurricane risks in the USA. Int. J. Disaster Risk Reduct., https://doi.org/10.1016/j.ijdrr.2018.01.023, in press.
NASEM, 2017: Integrating Social and Behavioral Sciences within the Weather Enterprise. National Academies Press, 198 pp., https://doi.org/10.17226/24865.
Neeley, L., 2014: Risk communication in social media. Effective Risk Communication, J. Arvai and L. Rivers III, Eds. Routledge, 143–164.
New York Times, 2012: New York City hurricane evacuation zones. New York Times, 28 October, http://www.nytimes.com/interactive/2012/10/28/nyregion/hurricane-evacuation-zones.html?_r=0.
NOAA, 2013a: Tropical cyclone report: Hurricane Sandy. 157 pp., http://www.nhc.noaa.gov/data/tcr/AL182012_Sandy.pdf.
NOAA, 2013b: Service assessment: Hurricane/post-tropical cyclone Sandy, October 22–29, 2012. 66 pp., http://www.nws.noaa.gov/os/assessments/pdfs/Sandy13.pdf.
Palen, L., and K. M. Anderson, 2016: Crisis informatics—New data for extraordinary times. Science, 353, 224–225, https://doi.org/10.1126/science.aag2579.
Palen, L., K. M. Anderson, G. Mark, J. H. Martin, D. Sicker, M. Palmer, and D. Grunwald, 2010: A vision for technology-mediated support for public participation and assistance in mass emergencies and disasters. Proc. 2010 ACM-BCS Visions of Computer Science Conf., Edinburgh, United Kingdom, Association for Computing Machinery, 12 pp., https://dl.acm.org/citation.cfm?id=1811194.
Parkhill, K. A., K. L. Henwood, N. F. Pidgeon, and P. Simmons, 2011: Laughing it off? Humour, affect, and emotion work in communities living with nuclear risk. Br. J. Sociol., 62, 324–346, https://doi.org/10.1111/j.1468-4446.2011.01367.x.
Peacock, W. G., S. D. Brody, and W. Highfield, 2005: Hurricane risk perceptions among Florida’s single family homeowners. Landscape Urban Plann., 73, 120–135, https://doi.org/10.1016/j.landurbplan.2004.11.004.
Peters, E. M., B. Burraston, and C. K. Mertz, 2004: An emotion-based model of risk perception and stigma susceptibility: Cognitive appraisals of emotion, affective reactivity, worldviews, and risk perceptions in the general of technological stigma. Risk Anal., 24, 1349–1367, https://doi.org/10.1111/j.0272-4332.2004.00531.x.
Renn, O., 2008: Review of psychological, social and cultural factors of risk perception. Risk Governance: Coping with Uncertainty in a Complex World, O. Renn, Ed., Earthscan, 98–148.
Rice, R. G., and P. R. Spence, 2016: Thor visits Lexington: Exploration of the knowledge-sharing gap and risk management learning in social media during multiple winter storms. Comput. Human Behav., 65, 612–618, https://doi.org/10.1016/j.chb.2016.05.088.
Rickard, L. N., Z. J. Yang, J. P. Schuldt, G. M. Eosco, C. W. Scherer, and R. A. Daziano, 2017: Sizing up a superstorm: Exploring the role of recalled experience and attribution of responsibility in judgments of future hurricane risk. Risk Anal., 37, 2334–2349, https://doi.org/10.1111/risa.12779.
Ripberger, J. T., H. C. Jenkins-Smith, C. L. Silva, D. E. Carlson, and M. Henderson, 2014: Social media and severe weather: Do tweets provide a valid indicator of public attention to severe weather risk communication? Wea. Climate Soc., 6, 520–530, https://doi.org/10.1175/WCAS-D-13-00028.1.
Schram, A., and K. M. Anderson, 2012: MySQL to NoSQL: Data modeling challenges in supporting scalability. Proc. Third Annual Conf. on Systems, Programming, and Applications: Software for Humanity, Tucson, AZ, Association for Computing Machinery, 191–202, https://doi.org/10.1145/2384716.2384773.
Shelton, T., A. Poorthuis, M. Graham, and M. Zook, 2014: Mapping the data shadows of Hurricane Sandy: Uncovering the sociospatial dimensions of ‘big data.’ Geoforum, 52, 167–179, https://doi.org/10.1016/j.geoforum.2014.01.006.
Sherman-Morris, K., 2013: The public response to hazardous weather events: 25 years of research. Geogr. Compass, 7, 669–685, https://doi.org/10.1111/gec3.12076.
Slovic, P., M. L. Finucane, E. Peters, and D. G. MacGregor, 2004: Risk as analysis and risk as feelings: Some thoughts about affect, reason, risk, and rationality. Risk Anal., 24, 311–322, https://doi.org/10.1111/j.0272-4332.2004.00433.x.
Spence, P. R., K. A. Lachlan, X. Lin, and M. D. Greco, 2015: Variability in Twitter content across the stages of a natural disaster: Implications for crisis communication. Commun. Quart., 63, 171–186, https://doi.org/10.1080/01463373.2015.1012219.
St. Denis, L., L. Palen, and K. M. Anderson, 2014: Mastering social media: An analysis of Jefferson County’s communications during the 2013 Colorado floods. Proc. Conf. on Information Systems for Crisis Response and Management, University Park, PA, Pennsylvania State University, 10 pp.
Stein, R., L. Dueñas-Osorio, and D. Subramanian, 2010: Who evacuates when hurricanes approach? The role of risk, information, and location. Soc. Sci. Quart., 91, 816–834, https://doi.org/10.1111/j.1540-6237.2010.00721.x.
Superstorm Research Lab, 2013: A tale of two Sandys. Superstorm Research Lab White Paper, 22 pp., https://superstormresearchlab.org/white-paper/.
Sutton, J., C. League, T. Sellnow, and D. D. Sellnow, 2015: Terse messaging and public health in the midst of natural disasters: The case of the Boulder floods. Health Commun., 30, 135–143, https://doi.org/10.1080/10410236.2014.974124.
Taylor, K., S. Priest, H. Fussell Sisco, S. Banning, and K. Campbell, 2009: Reading Hurricane Katrina: Information sources and decision-making in response to a natural disaster. Soc. Epistemology, 23, 361–380, https://doi.org/10.1080/02691720903374034.
Trumbo, C., L. Peek, M. Meyer, H. Marlatt, E. Gruntfest, B. McNoldy, and W. Schubert, 2016: A cognitive-affect scale for hurricane risk perception. Risk Anal., 36, 2233–2246, https://doi.org/10.1111/risa.12575.
Twitter, 2016: Twitter user agreement. Twitter, https://g.twimg.com/policies/TheTwitterUserAgreement_1.pdf.
Wilhelmi, O. V., and M. H. Hayden, 2010: Connecting people and place: A new framework for reducing urban vulnerability to extreme heat. Environ. Res. Lett., 5, 014021, https://doi.org/10.1088/1748-9326/5/1/014021.
Wu, H.-C., M. K. Lindell, and C. S. Prater, 2015a: Process tracing analysis of hurricane information displays. Risk Anal., 35, 2202–2220, https://doi.org/10.1111/risa.12423.
Wu, H.-C., M. K. Lindell, and C. S. Prater, 2015b: Strike probability judgments and protective action recommendations in a dynamic hurricane tracking task. Nat. Hazards, 79, 355–380, https://doi.org/10.1007/s11069-015-1846-z.
Zhang, F., and et al. , 2007: An in-person survey investigating public perceptions of and responses to Hurricane Rita forecasts along the Texas coast. Wea. Forecasting, 22, 1177–1190, https://doi.org/10.1175/2007WAF2006118.1.
Zhang, Y., C. S. Prater, and M. K. Lindell, 2004: Risk area accuracy and hurricane evacuation from Hurricane Bret. Nat. Hazards Rev., 5, 115–120, https://doi.org/10.1061/(ASCE)1527-6988(2004)5:3(115).
Zimmer, M., and N. J. Proferes, 2014: A topology of Twitter research: Disciplines, methods, and ethics. Aslib J. Inf. Manage., 66, 250–261, https://doi.org/10.1108/AJIM-09-2013-0083.
The day that Hurricane Sandy made landfall, it transitioned to a posttropical storm, and thus was no longer referred to as a hurricane by the National Weather Service (NOAA 2013a,b). However, for simplicity, we refer to it as Hurricane Sandy throughout the article.
These authors operationally define useful informational tweets as “those whose primary intent was to provide information concerning the technical aspects of the storm or specific mitigation efforts” (Lachlan et al. 2014b), and that such information “included risk, loss of assets, food/shelter, evacuation, the whereabouts of others, financial assistance, cancellations, and care for the sick and elderly” (Spence et al. 2015).
Full details on the infrastructure supporting the data collection and analytics are in Anderson and Schram (2011), Schram and Anderson (2012), and Anderson et al. (2013).
We added “far rock” and “farrock” to our search after observing in the dataset with mentions of “farrockaway” and “far rockaway” that the neighborhood is often referred to in this way.
