Introduction
In hazard contexts, forecasts and warnings are designed to inform people about a threat and support effective responses. As described by Mileti and Sorensen (
1990), hazard warning response involves multiple interconnected subprocesses, including hearing (or seeing) the warning; understanding, believing, personalizing, and confirming the information; and deciding on and taking action. Multiple factors, including recipient attributes and warning message source, content, and style, influence these processes (
Mileti and Sorensen 1990;
Mileti 1995).
Research on a variety of hazards utilizes this framework developed by Mileti and colleagues, which is referred to as the Warning Response Model (e.g.,
Mileti and Sorensen 1990;
Sorensen 2000;
Bean et al. 2015;
Wood et al. 2018;
NASEM 2018;
Sutton et al. 2018,
2023;
Kuligowski et al. 2023). However, warning creation, communication, and response has changed significantly over the last few decades. For weather-related hazards, scientific and technological advances have markedly improved forecasts (
Bauer et al. 2015;
Cangialosi et al. 2020). At the same time, advances in information and communication technology, such as the Internet, mobile phones, and social media, have transformed how people access, share, and interact with hazard information (
Gladwin et al. 2007;
Sutton et al. 2008;
Fraustino et al. 2012;
Bean et al. 2015;
Morss et al. 2017). These advances enable more skillful forecasts and warnings for weather-related hazards, issued further in advance, as well as more rapid and widespread forecast and warning dissemination. They also enable extensive use of visuals in weather and other hazard communication (
Lindell 2020;
Millet et al. 2020;
Prestley and Morss 2023;
Wilhelmi et al. 2024). Along with such opportunities for hazard risk communication, these changes bring challenges, including the potential for rapidly spreading false information and amplification of low-credibility sources (
Starbird et al. 2014;
Shao et al. 2018;
Vosoughi et al. 2018;
Swire-Thompson and Lazer 2020).
The research presented here aims to advance empirical and theoretical understanding about how modern forecast and warning systems function. This includes updating Mileti’s and others’ work on warnings to reflect the dynamic modern information environment, improved forecast and warning skill, and new forms of communication described above (
Mileti and Sorensen 1990;
Sorensen 2000;
Basher 2006;
Gladwin et al. 2007;
NASEM 2011,
2018;
Lindell and Perry 2012;
Morss et al. 2017;
Anderson-Berry et al. 2018). To do so, we investigate how visual forecast and warning communication evolved in real time leading up to and during a complex disaster with multiple types of embedded hazards, viewed through the lens of data from social media posts.
The study focuses on Hurricane Harvey, which caused more than 100 deaths and $125 billion in damage in the U.S. in August 2017 (
National Oceanic and Atmospheric Administration 2018). As illustrated by the timeline in Fig.
1, Harvey was a dynamic situation that presented multiple interconnected hazard threats—including strong winds, storm surge inundation, heavy rainfall and associated flooding, and tornadoes—in different areas of south Texas and Louisiana over the course of more than a week. By studying Hurricane Harvey from this perspective, we seek to elucidate communication and response processes for a variety of types of hazards that pose spatially and temporally varying risks.
As a tropical cyclone such as Harvey evolves, weather forecasters, media personnel, and public officials generate and communicate updated information about the threat multiple times each day (
Mileti and Sorensen 1990;
Lindell et al. 2007;
Demuth et al. 2012;
Bostrom et al. 2016). This evolving information propagates through social and information networks and is interpreted and used in a variety of ways (
Morss et al. 2017). Visuals can play important roles in these processes, for example, by influencing attention, message passing, risk personalization, and response (
Keib et al. 2018;
Sutton et al. 2019;
Clive et al. 2021;
Wilhelmi et al. 2023,
2024;
Prestley and Morss 2023). Thus, we use hurricane risk imagery conveyed by these types of sources as an entry point for analysis. Although these sources communicate across multiple media, the Internet—especially social media—provides a rich view of the visual information landscape at different times. Twitter, in particular, provides near-real-time, temporally detailed data about the online content that emerges during hazards, as well as about how Twitter users engage with information in this multidimensional communication space (
Starbird and Palen 2010;
Wu et al. 2011;
Sutton et al. 2014;
Anderson et al. 2016;
Veltri and Atanasova 2017;
Fellenor et al. 2018;
Reuter et al. 2018;
Silver and Andrey 2019;
Netzel et al. 2021).
Mileti and colleagues noted that hurricanes present complex risk communication situations. However, much of their work focuses on simpler textual warning messages designed to notify populations at risk about a specific hazard and motivate protective actions (
Mileti and Sorensen 1990;
NASEM 2011,
2018;
Bean et al. 2015). Here, we extend Mileti’s work by focusing on visual risk communication. This is important because a variety of maps and other forms of imagery are now commonly used to convey complex, geospatially detailed information about hazard threats, but visual warning messages have been less frequently studied (
Liu et al. 2017;
Bica et al. 2019;
Clive et al. 2021;
Sutton et al. 2023).
We also expand prior work by investigating individual warning messages in the context of an evolving multimessage, multihazard forecast and warning situation. The Warning Response Model and some other warning systems research conceptualize predictions as precursors to warnings. However, forecast and warning information is now regularly updated and communicated during hurricane threats, beginning days before impacts reach land (
Gladwin et al. 2007;
Demuth et al. 2012;
Bostrom et al. 2016;
Morss et al. 2017). We integrate this reality into our analysis by examining warning messages together with the forecast information that is communicated preceding and concurrent to warnings. In addition, research shows that many people now obtain hurricane forecasts from multiple sources and use that information to assess risk and make protective decisions—along with traditional warning messages such as evacuation orders (
Dow and Cutter 1998,
2000;
Gladwin et al. 2001;
Dash and Gladwin 2007;
Zhang et al. 2007;
Morss and Hayden 2010;
Demuth et al. 2023). Bringing these concepts together, we contribute to the hazards literature by taking a multisource, integrated forecast and warning approach, which incorporates the broader context of modern hurricane forecast and warning, risk communication, and response processes.
Building on the body of research studying how authoritative sources use Twitter to communicate and engage with different audiences (e.g.,
Neiger et al. 2013;
Hughes et al. 2014;
Sutton et al. 2014,
2020;
Eriksson 2018;
Vos et al. 2018;
Olson et al. 2019;
Rufai and Bunce 2020), this study explores what hurricane-related forecast and warning information was disseminated during Harvey’s threat, by whom, when, and how. More specifically, we combine qualitative and quantitative analysis of Twitter data to address three research questions:
1.
How did authoritative sources communicating with populations in areas at risk convey visual forecast and warning information on Twitter during Harvey?
2.
To what extent were different types of visual forecast and warning information, disseminated by different sources, diffused on Twitter?
3.
How did these processes change over Harvey’s lifetime, as the hazard threats posed by the storm and associated forecast and warning content evolved?
We study dissemination and diffusion because both influence who is exposed to information. These processes therefore provide a critical bridge between the creation of information and subsequent hearing or seeing, understanding, personalizing, believing, and confirming that information as described in the Warning Response Model (
Mileti and Sorensen 1990).
To investigate dissemination, we analyze what types of forecast and warning information are tweeted by which types of authoritative sources as the threat progresses. To investigate diffusion—in other words, whether and how information passes through formal and informal networks and is potentially amplified to reach different audiences—we analyze patterns of retweets of authoritative sources’ original tweets (
Toriumi et al. 2013;
Kogan et al. 2015;
Vos et al. 2018;
Bica et al. 2019;
Sutton et al. 2020). In addition, we examine how diffusion occurs through information originated by one source propagating into others’ communications. For example, during weather threats, the U.S. National Weather Service (NWS) issues a variety of meteorological forecast and warning products, in textual and graphical formats. As NWS creates this information, it disseminates both the products and underlying data outside of Twitter for others to use. We incorporate this into our analysis by investigating how other authoritative sources further diffuse NWS-generated information by including it in their own tweets. We also use retweets as markers of the salience of different information from different sources and of people’s attention to and amplification of that information (
Ripberger et al. 2014;
Vos et al. 2018;
Silver and Andrey 2019;
Sutton et al. 2019). Finally, we use the content of replies to forecast and warning tweets to explore how people in areas at risk interpret, make sense of, and respond to Harvey’s evolving threat. We examine all of these processes at multiple times during Harvey, taking a longitudinal approach (
Siegrist 2014;
Demuth et al. 2023,
Forthcoming).
When analyzing the types of forecast and warning information communicated during Harvey, we incorporate Mileti’s findings on five topics important to include in warning messages—hazard or risk, guidance, location, time, and source—and associated stylistic aspects such as specificity and clarity (
Mileti and Sorensen 1990;
Bean et al. 2015;
Wood et al. 2018;
Kuligowski et al. 2023;
Sutton et al. 2023). We also explore how our results intersect with the confirmation and milling processes discussed in the warning response literature, where people seek additional information from others to verify warning messages (
Mileti and Sorensen 1990;
NASEM 2011,
2018;
Sutton et al. 2023) and make sense of emerging, uncertain situations (
Wood et al. 2018;
Carlson and Barbour 2023). Social media data enables us to see multiple aspects of these complex processes, which are well known in the hazards community but difficult to observe and analyze as they unfold (
Chung 2011;
Lachlan et al. 2014;
Sutton et al. 2014;
Spence et al. 2015;
Veltri and Atanasova 2017;
Reuter et al. 2018). By integrating concepts and knowledge from hazard warning, weather prediction and predictability, and risk and crisis communication and studying how they manifested in a real tropical cyclone event, we aim to advance understanding about public warnings and alerts from an interdisciplinary perspective (
Gall et al. 2015;
Peek et al. 2020;
Morss et al. 2021;
Sherman-Morris et al. 2021).
We start by describing the research methods, which use Twitter data collected for an earlier study of hurricane risk imagery by Bica et al. (
2019), but here with a different goal: studying forecast and warning communication with people in areas at risk from a specific hurricane threat. Thus, prior to analysis, we filtered the Bica et al. data set to focus on image tweets that convey forecast and warning information for Hurricane Harvey, posted by authoritative sources communicating with populations in areas at risk from this storm. Next, we present results, starting with the overall roles of different types of imagery and authoritative sources in Twitter forecast and warning communication during Harvey. This is followed by analysis of how hazard communication evolved throughout Harvey’s lifetime. To explore this evolution in greater depth, we then examine dissemination, diffusion, and response in three 3-h periods during Harvey’s threat, analyzing changes within and across these periods. We close with a summary of key findings, areas for further research, and suggestions for improving hazardous weather risk communication.
Methods
This section presents key aspects of the research methods important for understanding the study results. To help advance the design of rigorous, replicable methodologies for using social media data to study forecast and warning communication, we provide additional details about the study’s methods and the reasons underlying different methodological choices in Supplemental Materials. The coding schemes used in the study are available in Prestley and Morss (
2024).
Data Collection and Original Image Tweet and Time Filtering
The study uses data collected by a team at University of Colorado Boulder in collaboration with researchers at the National Center for Atmospheric Research, as part of a larger project studying dynamic hazard prediction, risk communication, and decision making in the modern information environment (
Morss et al. 2017). Bica et al. (
2019) previously used these data to investigate global diffusion of and reactions to hurricane risk images that authoritative sources of hurricane risk information tweeted during the 2017 Atlantic hurricane season. To identify relevant Twitter accounts for data collection, Bica and colleagues started with two public lists of Twitter accounts providing official information for Hurricanes Harvey and Irma [see Bica et al. (
2019) for details]. Members of the research team then manually added national, regional, and local NWS, other government, weather media, and news media Twitter accounts providing reliable hurricane information during four of the season’s major hurricane threats to the U.S.: Harvey, Irma, Maria, and Nate. These 796 Twitter accounts provided the starting point for data collection. However, as subsequently described, we later narrowed the data set to tweets from a smaller, more curated set of sources.
Bica and colleagues collected data that included all tweets posted by these 796 accounts from August 17, 2017 to October 10, 2017. Here, we start with the set of original tweets posted by these accounts, filtered to tweets containing at least one still image, video, or animated GIF (collectively referred to as
image or
imagery). We also use retweet counts that Bica et al. (
2019) calculated for these tweets, overall and during specific time periods after posting, and the content of replies to these tweets. Additional details about the data collection and image filtering are provided in Bica et al. (
2019) and section
S1.
Given this study’s focus on Hurricane Harvey, we filtered these original image tweets to those posted during the period when Harvey was a potential or active threat: 00:00 UTC on August 17, 2017 (19:00 CDT on August 16) to 15:00 UTC on September 2, 2017 (10:00 CDT on September 2); all subsequent times are provided in local time (Central Daylight Time, CDT). The resulting
time-filtered data set contains 47,342 original image tweets (Fig.
2).
Authoritative Source Coding, Categorization, and Filtering
We coded authoritative source accounts for two purposes: (1) to filter the data set to tweets from sources communicating about Harvey with populations in areas at risk from the storm, and (2) to categorize source accounts for analysis. During the coding process, we confirmed that all accounts included in the data set were credible information sources. We coded the accounts along three dimensions: geographic area of responsibility, representation as an individual or organizational account, and type of professional role. Two researchers (R. Prestley and R. Morss) independently coded each account and then adjudicated differences to develop a consensus coding.
For geographic area, we categorized sources according to whether their area of responsibility or primary coverage was national or international (National); local or regional to the primary areas threatened or affected by Harvey (south, southeast, or central Texas or southwest Louisiana; Local); or sub-national, not well defined, or local or regional to a different area (Other). We coded accounts as organizational if their username or profile indicated an official organizational affiliation and individual if these indicated a personal account (see section S2 for details).
For professional role, we categorized accounts associated with media organizations (television, radio, print, and/or online) as Weather Media if they focused on producing or providing information about the weather and News Media if they focused primarily on providing other news content. We categorized accounts associated with government organizations as either NWS or Non-NWS Government, and the remaining accounts as Other. We then combined the geographic area and professional role codes into a single set of source type categories (see section S2).
Because the Bica et al. (
2019) Twitter data collection included a variety of sources posting throughout the 2017 hurricane season, the time-filtered data set contains many sources local to areas not at risk from Harvey, and many tweets unrelated to Harvey (or to weather at all). Therefore, to help narrow the data to tweets of interest for this study, we filtered the data set to include only tweets from seven types of National and Local Harvey sources, shown in Table
1. We developed this source filtering approach using knowledge about hurricane risk and social media communication, along with results from preliminary analysis as described in section
S2. The resulting
source-filtered data set contains 14,284 original image tweets (Fig.
2).
Hurricane Risk Image, Harvey Relevance, and Forecast and Warning Coding and Filtering
Although the time and source filtering processes described previously removed many tweets not relevant to this study, the source-filtered data set still contains many tweets not relevant to hurricane forecasts and warnings or to Harvey. Thus, we used tweet content—both text and imagery—to further filter the data to tweets of interest for our research questions.
First, as described in Bica et al. (
2019) and section
S3, we filtered the data set to include only tweets with hurricane risk images (Fig.
2). Next, we coded tweets in this
hurricane risk image data set according to whether they were
relevant to Harvey. Relevance was defined as either the tweet text or imagery referencing Harvey by name or mentioning or visually representing the storm’s threat, its impact, or related protective action information. Using this coding, we filtered the data to contain only Harvey-relevant tweets, generating a
Harvey-relevant data set (Fig.
2). We then coded these tweets according to whether they contained
Harvey forecast or warning information, defined as either the tweet text or imagery conveying Harvey-relevant threat, impact, or protective action information in future terms. Tweets that included only observational data, such as depictions or descriptions of recent storm evolution, were not coded as forecast or warning information (
Rosen 2019).
We tested the Harvey relevance and forecast/warning code definitions as described in section
S3; intercoder reliability was high for both Harvey relevance (Krippendorff’s
) and forecast or warning information (
). The two coders discussed and adjudicated differences, and one coded the remaining tweets. After coding, we filtered the data to retain only tweets that contained Harvey forecast or warning information. The resulting
forecast and warning data set contains 3,441 tweets, as shown in Fig.
2.
Image Categorization
After the filtering steps described previously, we categorized each tweet in the Harvey forecast and warning data set by image type and branding, for use in the analysis. We developed the initial set of image type codes (Table
S1) based on the analysis in Bica et al. (
2019), images observed in earlier rounds of coding, and the research team’s knowledge about commonly used hurricane forecast and warning visuals. It includes 14 codes representing common visualizations of hurricane risk information, along with two Other codes. Two researchers tested and revised this coding scheme through several initial rounds of cross-coding, including calculating intercoder reliability as described in section
S4. One researcher then coded the remaining data.
During the coding process, we identified three additional image types in the data and added them to the coding scheme (Table
S2). In addition, as discussed further in the results, during coding we found that the majority of images in the data set depicting NWS Watches or Warnings were in the format of the example in Fig.
3(a), an experimental NWS product referred to as Severe Weather Impact Graphics (
National Weather Service 2016;
Walawender et al. 2017). Given the prevalence of these images, we separated them out of the
Watch/Warning code into their own category, called
NWS Impact Watch/Warning (or
NWS Impact W/W). We then revised the image type codes into a more compact, mutually exclusive categorization to combine infrequently used codes and account for common overlaps, as described in section
S4 and shown in Table
S2.
The resulting image type categorization used in the analysis contains 13 categories, described in Table
2 with examples shown in Figs.
3–6. This categorization includes image types that NWS, media, and other authoritative sources frequently use to convey hurricane forecast and warning information across multiple communication platforms, including television, the Internet, and social media [see, e.g., Morss et al. (
2022b) and Bostrom et al. (
2022)]; here, we observe them on Twitter. In two of these categories—
NWS Impact Watch/Warning and
Key Messages—all images are in a similar NWS-generated format [Figs.
3(a) and
5(d)]. Many of the other categories include images in both NWS and non-NWS formats, often with the non-NWS formats generated using the data underlying a corresponding NWS graphical product [see, e.g., Figs.
4(a and b)]. The
Other Forecast group contains images that included Harvey forecast or warning information, but that were not prevalent in the data set or for which we could not reliably define distinguishing criteria (section
S4). The
Other Non-Forecast group contains tweets that include Harvey forecast and warning information, but not in the imagery. The
Multiple group contains tweets that included two or more distinct types of imagery, either combined into one media attachment [as in Fig.
6(d)] or in multiple attachments.
The image branding coding scheme included two codes (section S4). We tested the code definitions as described in section S4 and obtained excellent intercoder reliability (). We then categorized each tweet using a binary scheme: NWS-branded if any image has an NWS logo, symbol, or name or is in an NWS format, and Non-NWS-branded otherwise.
Data Analysis
As shown in Fig. S1 and discussed further in section S5, the data set described above contains two outlier tweets with anomalous retweet and reply behavior. Together, these two tweets account for 19% of retweets and 34% of replies in the data set. In addition, many of the replies to these tweets focused on climate change or U.S. politics. This indicates that much of people’s interactions with them on Twitter is not related to communication about Harvey’s threat with populations potentially exposed to the storm. We therefore removed these two tweets from the data set used in the analysis, leaving 3,439 tweets.
Our integrated forecast and warning approach requires simultaneously understanding individual pieces of information and processes and examining how these evolve and intersect. To do so, we combined multiple forms of qualitative and quantitative analysis. To understand dissemination, we conducted an in-depth investigation of what types of information different sources tweeted as Harvey’s threat evolved. To understand diffusion and response, we coupled this investigation with in-depth analysis of retweets of and replies to different types of information tweeted by different sources at different times. We then used these in-depth analyses to inform how we structured the higher-level integrated analyses presented in this article. We also used the in-depth analyses to help us interpret the results of higher-level analyses.
In the results presented in this article, we use counts of original tweets as the primary quantitative metric of dissemination, and we use counts of retweets as the primary quantitative metric of diffusion. We also examine diffusion by tracking whether and how imagery and other information content generated by one authoritative source propagated into other sources’ tweets, through online and offline mechanisms. In addition, we use retweets as a measure of how Twitter users responded to authoritative sources’ tweets, including the salience of the information and their attention to it.
Finally, to explore additional aspects of people’s responses to the evolving forecast and warning information that authoritative sources tweeted during Harvey, we incorporated reply content into the analysis. Given our research questions, we focused on replies from people whose Twitter content and/or location in their Twitter profile at the time of Harvey indicated that they are members of the public either located in or communicating with people in an area at risk. Although the replies provide some data of interest, the data set contains a limited number of interpretable, relevant replies. Moreover, the reply data are from Twitter users engaging with authoritative sources, which is a limited sample. Thus, we only incorporated brief analyses of reply content into this article, using that content to reveal some of the different ways that people responded to Harvey’s evolving forecasts and warnings.
Summary and Conclusions
This study developed a mixed-method approach for using near-real time data from online social media posts to investigate evolving hazard forecast and warning communication with populations in areas at risk. We implemented this approach to study forecast and warning dissemination, diffusion, and response leading up to and during Hurricane Harvey, which posed a complex set of interconnected, spatially and temporally varying risks. The entry point for analysis was original image tweets posted by authoritative sources of weather risk information. We focused in particular on dissemination and diffusion because being exposed to, attending to, and engaging with information about a threat are important aspects of risk communication and decision making (e.g.,
Mileti and Sorensen 1990;
Griffin et al. 1999;
Lindell and Perry 2012;
Silver 2019;
Kasperson et al. 1988,
2022).
Our analysis found that along with warning messages of the type examined in Mileti and colleagues’ Warning Response Model, authoritative sources tweeted a variety of types of forecast information during Harvey, beginning days before warnings. On average, forecast tweets were retweeted more than warning tweets and for a longer period, suggesting greater attention to and diffusion of the information. Our analysis of replies also found that, consistent with other research, members of the public were interpreting and responding to this forecast information (
Dow and Cutter 1998,
2000;
Zhang et al. 2007;
Morss and Hayden 2010;
Morss et al. 2017;
Demuth et al. 2018). These results indicate that for hazards that can be predicted hours or days in advance, including hurricanes, the forecast information communicated about a threat provides important context for people’s interpretations of and responses to warning messages. Understanding how people attend to, perceive, and respond in these types of hazardous situations therefore requires reconceptualizing warning response—expanding it to encompass a broader forecast and warning perspective. As the meteorological, earth system, and other communities continue to improve prediction capabilities, updating hazard and disaster research models to incorporate these new developments will grow even more important.
The study’s findings also demonstrate the large volume and variety of frequently updated forecast and warning information that was available to members of the public during Harvey, as the risks evolved in space and time. This illustrates how, in complex, dynamic hazardous situations such as hurricanes, warning communication and response are part of a larger risk information ecosystem. In this ecosystem, hearing or seeing, understanding, personalizing, believing, and confirming information are still important, as are milling processes. However, our analysis suggests that these may not be distinct processes that people proceed through in response to specific forecast and warning information. Instead, these processes can occur simultaneously and continually, as people access the evolving collection of information about a threat, perceive risks, make sense of the situation, and decide what to do.
Another key finding from this study is how different types of authoritative sources played complementary, changing roles in communication as Harvey evolved. From August 18–20, when Harvey was not an active tropical cyclone, NWS sources tweeted little forecast information about the potential threat, and Weather Media sources filled this gap. As the threat to the U.S. increased and then Harvey redeveloped, National NWS sources led the generation and dissemination of new forecast and warning content, with other sources diffusing this content and providing additional interpretations. Then, as the situation transitioned to ongoing impacts and warnings for imminent hazards after landfall, Local NWS and other Local Government sources led the generation of new locally relevant forecast and warning content, with National NWS and Media sources using Twitter to help this information reach a larger audience. In these different phases, we observed how purposeful social media posting by multiple types of authoritative sources, using similar or different imagery, helped gain attention for, amplify, and augment forecast and warning messages. This provides evidence of how, as discussed in Demuth et al. (
2012), the different types of authoritative sources studied here work together to collectively create and communicate hurricane forecast and warning information. Although these sources operate on an increasingly crowded and fragmented information landscape, they continue to serve important, complementary roles in providing reliable hazard information.
Our analysis also reveals the important roles that information originated by NWS plays in other authoritative sources’ hurricane forecast and warning communication. One example is how, during Harvey, both NWS and non-NWS sources tweeted many images conveying NWS-issued Tornado and Flash Flood Warnings. Social media can help rapidly disseminate such warnings—if the information gains traction. However, many of the NWS tweets with this content disseminated NWS Impact Watch/Warning images, most of which had few retweets. Another example of NWS’s role in driving forecast and warning communication is our finding that despite the Cone’s well-known limitations, a variety of sources tweeted Cone images after NHC releases of new forecast packages. These examples underscore the importance of NWS providing visuals and data that help other communicators effectively convey the most important aspects of a threat at any given time.
Regarding forecast and warning messaging, we found that many of the tweets in this data set do not contain all five types of content that Mileti and Sorensen (
1990) recommended including in warning messages: hazard or risk, guidance, location, time, and source. In particular, many of the common visuals found in this data set contain little or no guidance about recommended protective actions. This contravenes Wood et al.’s (
2012) findings on communicating actionable risk, as well as work on the importance of efficacy in risk communication and decision making (e.g.,
Bourque et al. 2013;
Ruiter et al. 2014;
Demuth et al. 2016;
Morss et al. 2016). Those creating and communicating warning messages may therefore benefit from using Mileti and Sorensen’s (
1990) recommendations about message content and style. However, much of the information studied here is communicated when key aspects of the threat are uncertain. Moreover, a growing body of research indicates that rather than obtaining comprehensive warning messages from one authoritative source, many people access multiple types of hurricane risk information from a variety of sources and synthesize it to make sense of the situation and decide what to do (
Zhang et al. 2007;
Morss and Hayden 2010;
Demuth et al. 2018,
2023;
Lazrus et al. 2020). People can therefore access important content collectively, from multiple sources and messages. Thus, it is important to update Mileti and Sorensen’s recommendations to reflect the larger volume and longer lead times of information now available for weather-related hazards and the greater complexity of modern hazard risk communication and response.
One goal of this research was advancing methods for studying the complex, evolving dynamics of modern forecast and warning communication and response. Online posts produce vast amounts of potentially informative data, but we found that choices about structuring data collection, filtering, and analysis influence what is learned. In order to obtain meaningful results about our research questions, we needed to sample, categorize, filter, and analyze the data carefully (
Palen and Anderson 2016). This included conducting in-depth qualitative analysis to support and help guide quantitative analysis. In addition, to develop robust results relevant to the study domain, we needed to analyze the data from multiple perspectives, informed by knowledge of the information content and communication context. This type of deep content-based knowledge can help researchers build up to analyses with larger data sets involving more quantitative and automated methods.
Along with the theoretical and methodological contributions discussed above, our analysis leads to several practical recommendations:
•
The gap in NWS Twitter communication from August 18–20 provides further support for recent recommendations that NWS develop new products and strategies for communicating tropical cyclone scenarios at longer lead times, before NHC provides the track and intensity forecasts used in Cone images (
Morss et al. 2022a).
•
Our results on how different sources play complementary roles in Twitter forecast and warning communication suggest the potential for NWS, public officials, media sources, and others to develop new ways of coordinating online to quickly communicate critical information to a variety of audiences.
•
The prominence of Cone image tweets during Harvey provides additional support for recent recommendations that NWS update the Track Forecast Cone graphic or develop an improved tropical cyclone threat summary product that more effectively communicates hurricane risks to different populations (
Evans et al. 2022;
Henson 2022;
Morss et al. 2022a).
•
Our related finding that some authoritative sources emphasized newly released information, e.g., about Harvey’s rapid intensification on August 24th, suggests that NWS develop improved strategies to highlight the most important information at any given time, across their product suite.
•
Our findings on Watch/Warning tweets, including the lack of protective action guidance in many Watch/Warning tweets and the few Watch/Warning tweets on the morning of August 24, suggest that NWS may want to develop new approaches for communicating Watch and Warning messages—approaches that prominently feature the key information NWS is trying to convey and facilitate other sources diffusing and amplifying this information.
•
These Watch/Warning results, together with our results on NWS Impact Watch/Warning tweets, suggest that NWS develop revised visuals for communicating longer- and shorter-term Watches and Warnings. To be effective, these visuals should be developed in collaboration with other weather information communicators, so that the imagery can readily be diffused or adapted for others’ communications.
These recommendations were developed by interpreting this study’s findings in the context of current knowledge about hurricane forecast and warning communication research and practice.
One limitation of this study relates to our aim of studying forecast and warning communication with people in areas at risk. Although we structured our methods with this aim in mind, the retweets in the data set are from a mix of populations. Moreover, many people are not on or not active on Twitter, and some populations are not engaged with the types of authoritative source communications that we analyzed. Thus, it is important to complement this type of research with work that uses other approaches to understand how people interact with and use hazard information, including sociodemographically diverse populations and those who may be more susceptible to harm (e.g.,
Dash and Gladwin 2007;
Lazrus et al. 2012,
2020;
Anderson et al. 2016;
Huang et al. 2016;
Wilhelmi et al. 2023,
2024).
Another limitation is that we only analyzed tweets that included hurricane risk imagery, based on our use of Bica and colleagues’ data set and our interest in visual communication. Future work could extend these methods to analyze forecast and warning content more broadly and to investigate the interplay between text and imagery. In addition, Harvey was a unique tropical cyclone with an atypical evolution; to develop more generalizable findings, it is important to study other events. To begin addressing these limitations, two members of our research team conducted a follow-on study of tweets with and without imagery posted by professional sources during Hurricane Irma (
Prestley and Morss 2023). Compared to Harvey, Irma had a longer predictive lead time. In addition, Irma and its impacts evolved differently from Harvey and affected a region with different geography, which led to a Twitter data set less dominated by short-fused, post-landfall warnings in a major metropolitan area. This allowed the Irma study to build on the analysis presented here in several ways, including identifying forecast and warning sub-phases and using regression analysis to understand how tweet timing, content, and other factors affect retweets.
In addition, after we conducted our analysis, Twitter went through major changes, leaving its future role in hazard forecast and warning communication uncertain. We do not yet know whether Twitter will continue to be a robust, trusted platform for conveying risk messages, or whether Twitter will continue to provide valuable data for research. Nevertheless, we anticipate that aspects of our findings are relevant beyond Twitter specifically, and social media more broadly, to modern multisource, multiplatform, and multimessage communication.
Overall, the processes we see in these data connect to those in Mileti’s and others’ work on hazards and warnings. Yet during the last few decades, the Internet, social media, and other advances have dramatically increased the volume, heterogeneity, and complexity of information available during times of threat. They have also opened up new opportunities to observe in depth how information communication and response processes evolve and interact. It is challenging to simultaneously investigate multiple aspects of these processes during dynamic, uncertain situations, as information about a threat evolves. Yet approaching data collection and analysis from this perspective, guided by relevant knowledge from multiple fields, enabled this study to update current understanding about real-world hazard forecast and warning systems. The findings can be used to contextualize other research that focuses on subsets of the topics studied here, guide additional work, and improve forecast and warning communication.