Network Analysis of Water Contamination Discourse on Social Media Platforms

Abstract

This study examines water contamination issues using social media data (n = 420.7 k) to unveil the roles and patterns from the top 10 most influential users in their respective networks determined by their reply, retweet, and mention networks. Findings from RQ1 show diverse entities within this group, encompassing political figures, organizations, cause-related actors (e.g., EPA), educational institutions, and high-activity accounts all contribute to the discourse on water contamination. While RQ2 sought to examine the evolution of discourse regarding water contamination and its related dynamics, instead, communication “shocks”, or deviations from typical discourse that returned to the original state, were identified, rather than revolutionary moments on social media that triggered long-term changes.

1. Introduction

In 2022, 2.2 billion individuals globally lacked access to safely managed drinking water services [1] (World Health Organization, 2023). However, access to safe, reliable drinking water is a priority for any society. The typical procedures for establishing and maintaining safe water management systems in U.S. communities include locating reliable water resources and adequately treating water following standard regulatory allowances; stakeholders in water resource management often include utility operators, local authorities, and elected officials [2]. The Safe Drinking Water Act (SDWA), under the Environmental Protection Agency (EPA), sets standards for water quality and monitors entities, including water suppliers, who must comply with these standards. The EPA also has a website through which water suppliers must submit consumer confidence reports to facilitate transparency in water quality management and information. Although priorities and policies suggest that at least developed nations (e.g., the U.S.) would ensure constituents’ access to reliable, safe drinking water, aging infrastructures, poorly resourced and mismanaged areas, and crises often give way to water contamination threats. These issues are multifaceted and have adverse effects on the environment, ecosystem, and public health [3], negatively impacting human rights and well-being.

Procedures, processes, and concerns related to water contamination inherently involve degrees of scientific and technical uncertainty [4]. This uncertainty is critical information that must be effectively communicated to all stakeholders to facilitate well-informed decisions concerning the environment and public health that affect all people; this need is exacerbated when risks manifest in crises (e.g., water systems become contaminated). However, although transparent communication is vital regarding water treatment and regulatory practices at the local citizen level (such as making decisions regarding filtration based on information received), at the organizational level (such as the EPA; see Ref. [4]), and via media coverage, information dissemination and exchange around water quality risks and uncertainties and the collective management of risk surrounding communal resources (e.g., water systems) invite a host of challenges.

Understanding the information ecosystem related to water contamination on social media aids in understanding where potential opportunities and challenges lie in public discourse and where communication and information interventions may be warranted to realize productive discussions about and management of water resources in the future.

The current research delves into an exploration of social media data to uncover prevalent patterns emerging from interactions between opinion leaders, organizations such as the Environmental Protection Agency, and various media outlets regarding issues concerning water contamination. Through an analysis of tweets related to water contamination, this study accentuates the influence wielded by substantial actors who dominate the discourse in networks, whether through notable moments or provocative and inflammatory content. These occurrences, or “shocks”, provide valuable insights into the intricate dynamics of the conversation.

The primary objective of this study is to gain a more precise understanding of the extent and characteristics of online discourse concerning water contamination by examining communication patterns and information dissemination on social media. As a result, its contribution lies in enriching the discourse on this vital issue broadly, facilitating a better understanding of the communication and information landscape around water contamination on social media, which may inform the development of communication strategies and intervention development in the future.

2. Social Media and Water Contamination Issues

Given the evolving landscape of social media, including information accuracy, trust, and credibility [5], a central concern arises in the cultivation of expertise and reputation within this context relating to the identification of key nodes and sub-networks where discussions related to water contamination evolve over time, including the involvement of laypersons, organizations, and mass media in shaping these discussions. These have substantial ramifications for several areas, from identifying and disseminating vital information—such as fixes for clean public spaces and safe drinking water—to comprehending how they affect public emergencies [5]. Research suggests that relational linkages in social networks seem to help individuals in refining their own risk perceptions [6], which denotes the importance of understanding information flows in social networks related to risks like water contamination [7]. In working to understand and inform such challenges, prior studies have examined social media through several approaches. Local and informal discussions tend to make up much of the microblogging space on platforms such as X (formerly known as Twitter) and can inform citizen science initiatives [8]. Related research suggests that individuals may use social media to seek out environmental health information if they feel uninformed about the issue or depending on informational subjective norms [9].

As social media has become a central location for information sharing and discussion [10], scholars are increasingly analyzing how online conversations play a role in understanding social phenomena [11]. For instance, it has been argued that online spaces are a main locale for sensemaking [12] and can play a significant role in understanding public perceptions of health-related issues [11]. Additionally, Toivonen et al. [13] note that social media platforms are vital for studying interactions between users and their environments.

By engaging with the social landscape, especially through opinion leaders, to address the vital issue of water contamination, this reveals channels for sharing vital information among various stakeholders. This contributes to multiple benefits, including gaining insights into how the citizens shape their perspectives, how organizations act on information, and how they communicate these actions with the public. Water-related public-facing issues and communication about them can have implications for broader public policy and society, often in perhaps unpredictable ways. For example, VanDyke et al. [14] found that communication about COVID-19 was connected to discussions about other science-related issues and risks, such as climate change. These results suggest that online discussion about one issue can have a spill-over effect on others and that discussions evolve and change over time.

The complexity of water contamination issues and the fact that social media users can drive conversation in unique ways present significant challenges. In working to understand and inform such challenges, prior studies have examined social media through several approaches. Local and informal discussions tend to make up much of the microblogging space on platforms such as Twitter/X and can inform citizen science initiatives [15]. Related research suggests that individuals may use social media to seek out environmental health information if they feel uninformed about the issue or depending on informational subjective norms [9].

In addition, formal and informal opinion leaders emerge in social spaces as influential [15]—whether those opinion leaders are politicians, local citizens holding a particularly influential role in their community, influencers on social media, and the like; these are vital voices within these networks. Tweets in response to crises, such as emergency preparedness, also emerge from organizations sharing vital information with citizens, or organizations sending planned campaign messages, which play important roles within communities.

Water Contamination and Challenges

The context surrounding specific water contamination issues and challenges is important for understanding the stakeholders involved and the complexities and nuances surrounding related communication and information dynamics. Mueller and Gasteyer [16] documented plumbing, wastewater, and poor drinking water quality conditions in the U.S. and found that poor water conditions were associated with rurality, poverty, and other factors associated with environmental injustices. Further, in a 2016 study of networked communication on Twitter during the 2014 water crisis in Charleston, West Virginia, Getchell and Sallow [17] found that communication happened quickly, and both local and national agencies were involved. However, the network significantly lacked density and included isolates. The network belied a lack of communication between community and federal sources, which likely hindered the effectiveness of the crisis response.

Prior research [18,19] has emphasized the crucial importance of developing an initial understanding of the intricate foundations related to health and environmental issues. Within the context of communication and information dynamics in these scenarios, studies have shown there are various communication channels stakeholders and organizations may use [20]. These include authorities communicating with citizens (crisis communication), authorities exchanging information with other authorities (interorganizational crisis management), citizens interacting with fellow citizens (self-help communities), or individuals who directly reach out to authorities [21]. Research has also explored how organizations and mainstream media detect and respond to such issues utilizing platforms like X (e.g., [16,21,22], along with other activities regarding water pollution [23,24]. For example, Zhou and Chen [22] proposed a location-time constrained model to detect crisis events over social media, which was found to be highly effective for crises.

Other research examined how individuals use social media in risk and crisis contexts. For example, Yuan and Gasco [8] examined the 2013 Water Contamination Emergency in Shanghai and showed how local citizens used microblogging as a bottom-up channel to engage with governmental organizations. Other research demonstrated that individuals’ emotions in such contexts may drive their social media use. For example, social media use has been found to positively relate to feelings of fear and anger [25], which also positively relate to risk perceptions, and feelings of anxiety may lead to increased social media use for social support and information seeking purposes [26].

Other research examined formal organization and initiative communication responses to related challenges. Research suggests that water utility public communication strategies often are not uniform (e.g., some entities communicate proactively and others do not), and a majority of water utility consumers may be unaware of the actions implemented by their water utilities during public health crises [27]. In a scoping review of surveillance actions and initiatives for safe drinking water quality [28], demonstrated the importance of professional qualifications and routine action with the ability to communicate about risks. Connell et al. [29] analyzed the National Water Initiative (NWI), a program in Australia that aims to achieve sustainable water use, expand trade in water rights, and manage demands, among other factors. At the time, the program had yet to undergo policy testing, and Connell et al. [29] suggested that national programs needed proper communication refinement to be socially and environmentally effective. Timing of messages is critical [30], in combination with practical emergency measures in water supply emergencies. They developed a framework that emergency managers can use when determining the best strategies for activating targeted and citywide emergency alerts during water contamination events. Targeted messaging about how consumers should respond to a water contamination emergency worked, as well as citywide alerts to reach citizens and to reduce disruptions to those unaffected. In another study, Fritsch et al. [31] examined the activities of the European Union (EU) Water Initiative finding evidence that larger social discourses were needed to engage members and stakeholders that are part of different nations.

As these examples in prior research in the literature demonstrate, previous studies examined technical solutions for detecting and managing water contamination issues [22,31] or implemented case study approaches [29,31] to understand water contamination events and how stakeholders responded.

Less research has focused on the broader water contamination discourse that exists (at least observably via social media) and on understanding how individuals, organizations, and language may connect between and among specific cases of water contamination across locations and time. To resolve this gap in the literature, the present study examines the fundamental characteristics of the social structure surrounding water contamination discourse, with a particular emphasis on identifying the leading figures in that social structure, as well as observing its changes over time. To that end, the following research questions were posed:

• RQ1: Who are the most influential users in the relational network surrounding water contamination on social media?
• RQ2: How has the configuration of the relational network evolved over time?

3. Materials and Methods

To assess the relational network and its evolution over time, we collected all U.S.-based posts from X (formerly known as Twitter) on the topic of water contamination over a five-year period from 2018 to 2022. The Sprinklr [32] data warehouse was used to perform data collection using a Boolean search query designed to capture a comprehensive range of relevant discussions. The query included terms like “water contamination”, “contaminated water”, “water pollution”, and “water quality issue”, ensuring broad coverage of general discussions. More specific terms such as “arsenic contamination”, “bacterial water contamination”, “lead in water”, and “waterborne disease” were included to target tweets discussing particular contamination incidents and health concerns. This structure enabled us to gather data that spans both general water quality issues and specific contaminant-related conversations, ensuring a holistic understanding of the discourse on social media.

All data collection was completed from 24–26 October 2023. The resulting data set comprised 420,775 tweets, including 30,782 replies, 297,576 retweets, and 92,417 additional tweets.

Analysis

RQ1 was assessed via network analysis. Three network graphs were constructed using the igraph R package [33] based on three distinct types of user–user interactions within the data set: replies, retweets, and mentions. In the reply network graph, an edge with a weight of 1 was formed from one user to another—or, if the relevant edge already existed, its weight was incremented by 1—whenever the first user posted a reply to a tweet made by the second. The retweet and mention network graphs were built in the same fashion, except that edges in those network graphs were formed or strengthened whenever one user retweeted a fellow user or mentioned a fellow user in an original tweet (i.e., not a mere retweet of text someone else previously posted), respectively.

Once these three network graphs were constructed, the ten users in each network with the largest values of inbound degree centrality, outbound degree centrality, betweenness centrality, and eigenvector centrality were identified using the igraph R package [33]. Following the approach of [34,35], betweenness centrality was computed using an unweighted, undirected representation of each network. Additionally, to reduce its computational complexity, this metric was estimated using paths with lengths of 4 or less. All other centrality measures accounted for edge weights and direction in their computation.

A given user’s inbound degree centrality indicated the extent to which other users were inclined to make connections to that individual (whether via replies, retweets, and/or mentions), while outbound degree centrality demonstrated the extent to which a given user deliberately forged such connections to others in the network. Users with high betweenness centrality stood as bridges between otherwise disparate groups in the network, thereby exerting significant influence over the flow of information between them. Lastly, eigenvector centrality indicated the relative power that a given user held in the network such that users with high eigenvector centrality were those toward whom other users with high eigenvector centrality chose to form connections in the network.

RQ2, in turn, was examined using a combination of network analysis and stepwise segmented regression. To assess changes in the relational network as they manifested over time, network graphs were constructed representing the set of user–user interactions that emerged in each successive week. Specifically, for each week within the sampling frame (1–7 January 2018; 8–14 January 2018, etc.), all tweets in the data set made during that week were compiled. Reply, retweet, and mention network graphs were then constructed in the same fashion as they were for RQ1, but using only tweets made during that week rather than all tweets in the data set.

It should be noted that the network construction approach described above deviates from that previously employed by [35] to analyze collaborative contributions to Wikipedia. In that prior study, each weekly network consisted of all edges formed and/or strengthened prior to or during the week in question, whereas in the present study, only edges formed and/or strengthened during the week in question were included in each weekly network. This modification was made due to the nature of the study site. On Wikipedia, any given article that a user might revise is the aggregation of all prior revisions made to it. As such, anyone editing a Wikipedia article is implicitly interacting with the contributions of all prior editors. Therefore, in assessing the Wikipedia co-editorial network, it is appropriate to treat it as growing over time, retaining all prior vertices and edges indefinitely. On Twitter (now X), on the other hand, topics of discourse rapidly change, and there is minimal, if any, direct connection between a given tweet and other prior tweets on a related topic. In other words, discourse during one period is separable from discourse during prior periods, so in the present study, it was reasonable to treat each successive weekly network as being distinct rather than an ever-growing aggregation of contributions.

Afterward, for each weekly network graph, several network measures (inbound and outbound degree centralization, betweenness centralization, and assortativity) were computed using the igraph R package [34]. Additionally, the social entropy [35] of the distribution of the number of tweets made by each user during that week was computed, with a value of 0 indicating that one user made all of the tweets and a value of 1 indicating a perfectly equal number of tweets posted by all contributing users during that week.

Once this was achieved, the stepseg R package [36] was used to perform multivariate stepwise segmented regression [37] for all five measures evaluated with respect to all three network graphs, for a total of fifteen variables, in order to identify statistically significant breakpoints in their values. Based on existing recommendations [36,37], mean squared error was used to select potential breakpoints to add to each model; p < 0.15 and p > 0.20 were used as thresholds for adding and removing breakpoints, respectively, applying a Bonferroni correction to each value based on the presence of 15 dependent variables; and intercept and slope terms for each potential breakpoint were evaluated in all cases. All dependent variables were standardized prior to analysis.

The revolutionary changes represented by those breakpoints, and the evolutionary trajectories of these metrics as they manifested between breakpoints, were then mapped onto the expected measurement levels for organizational configurational archetypes theorized ([34], p. 16; see also [35], p. 109], thereby demonstrating how the relational networks described by these user–user interactions shifted between configurations over time.

4. Results

4.1. RQ1: Who Are the Most Influential Users in the Relational Network Surrounding Water Contamination on Social Media?

The most central users in each network (reply, retweet, mention) with respect to inbound and outbound degree centrality, betweenness centrality, and eigenvector centrality are detailed in

Table 1. Top 10 users with respect to inbound degree centrality, outbound degree centrality, betweenness centrality, and eigenvector centrality in the reply network.

Inbound Degree Centrality		Outbound Degree Centrality
User	Centrality	User	Centrality
@realDonaldTrump	700	@JingKlaus	1050
@tedcruz	334	@Lamarurquharth1	460
@thehill	126	@Tami1501	266
@WhiteHouse	120	@SafeH2o4Schools	245
@GOP	105	@isawthesethings	147
@FoxNews	82	@BreaveHeart43	87
@CNN	80	@awatarius	67
@POTUS	67	@TheWheltonGroup	51
@SatoMasahisa	54	@luisa_tasayco	46
@nytimes	52	@FieldRoamer	42
Betweenness Centrality		Eigenvector Centrality
User	Centrality	User	Centrality
@realDonaldTrump	12599203	@PhilipCPrice	1.0000000
@JingKlaus	7478517	@ngaphambbc	0.7500000
@isawthesethings	6929866	@LohrThoughts	0.7500000
@SafeH2o4Schools	5931217	@TheWheltonGroup	0.6123724
@WhiteHouse	4116278	@gkygirlengineer	0.5000000
@Tami1501	3771911	@bogdan09261613	0.5000000
@thehill	2809384	@Susan_Masten	0.5000000
@POTUS	2384643	@meredithcolias	0.5000000
@FoxNews	2323644	@loufreshwater	0.5000000
@Lamarurquharth1	2299150	@flintwaterstudy	0.5000000

,

Table 2. Top 10 users with respect to inbound degree centrality, outbound degree centrality, betweenness centrality, and eigenvector centrality in the retweet network.

Inbound Degree Centrality		Outbound Degree Centrality
User	Centrality	User	Centrality
@LilNasX	31711	@imagine_garden	1710
@kylegriffin1	10052	@SafeH2o4Schools	684
@SonnyVermont	9943	@yamatho2	589
@elonmusk	9777	@Bob_Stinson1234	280
@GlumBird	6530	@surfspup	231
@AP	6511	@amyrbrown12_amy	159
@krassenstein	5829	@FrackHazReveal	154
@RedTRaccoon	4016	@dgendvil	132
@charliekirk11	3983	@nutgraham	122
@QasimRashid	3806	@Eco1stArt	120
Betweenness Centrality		Eigenvector Centrality
User	Centrality	User	Centrality
@LilNasX	5563232127	@SafeH2o4Schools	1.00000000
@elonmusk	2031520664	@NRDC	0.30501560
@SonnyVermont	1773605052	@Hydroviv_h2o	0.28393199
@kylegriffin1	1644580076	@toxicreverend	0.26767209
@AP	1456901538	@Fix_Our_Schools	0.18300738
@GlumBird	1149732758	@BeCauseWater	0.15257359
@krassenstein	981795193	@EDFHealth	0.13438709
@charliekirk11	823302514	@nywaterproject	0.10593838
@L0vingnature	633367902	@enpressllc	0.10427548
@RedTRaccoon	626617931	@TheWheltonGroup	0.07600653

and

Table 3. Top 10 users with respect to inbound degree centrality, outbound degree centrality, betweenness centrality, and eigenvector centrality in the mention network.

Inbound Degree Centrality		Outbound Degree Centrality
User	Centrality	User	Centrality
@realDonaldTrump	1334	@JingKlaus	698
@EPA	1000	@IMJUSTTHEMAN1	651
@POTUS	484	@Hydroviv_h2o	382
@YouTube	383	@smartdissent	310
@Change	365	@SafeH2o4Schools	279
@EPAScottPruitt	346	@Lamarurquharth1	237
@NRDC	317	@DavidNoriega81	231
@CREDOMobile	284	@PracticalLefty	197
@GOP	277	@rln_nelson	195
@CNN	227	@ExMissionary	192
Betweenness Centrality		Eigenvector Centrality
User	Centrality	User	Centrality
@realDonaldTrump	159246469	@environmentguru	1.000000
@EPA	132152984	@TysonFoods	2.859874 × 10−16
@POTUS	61415205	@StandMighty	2.789181 × 10−16
@SafeH2o4Schools	43288269	@HealthyGulf	2.226099 × 10−16
@JingKlaus	38314543	@EPA	1.890578 × 10−16
@SunshineTheGrey	36643147	@EPAnewengland	6.836742 × 10−17
@Hydroviv_h2o	25596677	@mpsaz	5.843926 × 10−17
@CNN	25589816	@KateMishkin	3.606897 × 10−17
@GOP	23864343	@dougducey	2.807676 × 10−17
@nytimes	23742705	@tedcruz	1.976746 × 10−17

.

Regarding the findings of RQ1, the top 10 most influential users based on the networks for replies, retweets, and mentions were diverse in nature, and in many cases included politically oriented accounts, environmental advocates, vocal users on social media, educational organizations, news organizations, and other individuals who were simply active. For instance, regarding the reply network (Table 1), @realDonaldTrump is a prominent figure, and betweenness centrality shows @JingKlaus and @realDonaldTrump as having high scores, while eigenvector shows @PhilipCPrice as a central influencer. News organizations such as @thehill, @FOXNEWS, and @CNN had high scores for inbound degree centrality, while @CNN and @nytimes also had scores in the top 10 most influential users for betweenness degree centrality.

Regarding the retweet network, @LilNasX is a central figure in inbound degree centrality (higher levels of influence). Those with higher levels of betweenness centrality included users such as @LilNasX, @elonmusk, and @SonnyVermont. The eigenvector value shows that @SafeH2o4Schools is a central influencer in this particular network. Both news-focused accounts @AP and @krassenstein also had high scores for inbound centrality and betweenness centrality in the retweet network.

Finally, in the mention network, @realDonaldTrump, @EPA, and @POTUS were all accounts that served as those with especially high inbound degree centrality. These were the same three users in the network for betweenness centrality. Eigenvector, on the other hand, shows @environmentguru and @TysonFoods as centrally influential accounts. Further, the news account @CNN appeared in the top 10 accounts for inbound degree centrality and also appeared with @nytimes in the top 10 accounts for betweenness degree centrality. Notably, large news-focused accounts like @AP and @CNN did not appear in the top users for eigenvector values.

4.2. RQ2: How Has the Configuration of the Relational Network Evolved over Time?

The multivariate segmented regression model is presented in

Table 4. Final multivariate segmented regression model. * denotes significance.

	ꞵ	SE	t	p
Retweet Outbound Intercept * Week > 8	46.296	9.808	4.720	3.932 × 10−6
Retweet Entropy Intercept * Week > 8	−32.386	9.964	−3.250	0.001
Retweet Outbound Slope * Week > 8	−4.047	1.035	−3.909	1.196 × 10−4
Retweet Outbound Intercept * Week > 10	−46.002	9.794	−4.697	4.369 × 10−6
Retweet Entropy Intercept * Week > 10	33.119	9.950	3.328	0.001
Retweet Outbound Slope * Week > 10	4.094	1.029	3.978	9.123 × 10−5
Retweet Entropy Intercept * Week > 74	−3.310	0.407	−8.126	2.087 × 10−14
Retweet Entropy Slope * Week > 74	0.058	0.006	10.175	<2 × 10−16
Reply Betweenness Intercept * Week > 173	668.530	39.909	16.751	<2 × 10−16
Mention Outbound Intercept * Week > 173	864.608	219.180	3.945	1.039 × 10−4
Mention Betweenness Intercept * Week > 173	2093.445	140.572	14.892	<2 × 10−16
Mention Entropy Intercept * Week > 173	−996.294	218.208	−4.566	7.826 × 10−6
Reply Betweenness Slope * Week > 173	−3.767	0.229	−16.473	<2 × 10−16
Mention Outbound Slope * Week > 173	−4.927	1.256	−3.923	1.132 × 10−4
Mention Betweenness Slope * Week > 173	−11.955	0.806	−14.840	<2 × 10−16
Mention Entropy Slope * Week > 173	5.687	1.250	4.548	8.462 × 10−6
Reply Betweenness Intercept * Week > 175	−668.779	39.909	−16.758	<2 × 10−16
Mention Outbound Intercept * Week > 175	−864.711	219.182	−3.945	1.037 × 10−4
Mention Betweenness Intercept * Week > 175	−2092.592	140.573	−14.886	<2 × 10−16
Mention Entropy Intercept * Week > 175	995.990	218.209	4.564	7.875 × 10−6
Reply Betweenness Slope * Week > 175	3.769	0.229	16.478	<2 × 10−16
Mention Outbound Slope * Week > 175	4.929	1.256	3.924	1.126 × 10−4
Mention Betweenness Slope * Week > 175	11.951	0.806	14.836	<2 × 10−16
Mention Entropy Slope * Week > 175	−5.688	1.250	−4.549	8.423 × 10−6

. For the sake of brevity, only variables exhibiting statistically significant changes (with α = 0.05, using a Bonferroni correction for the 15 dependent variables) at each breakpoint are listed. Additionally, a plot of all five (unstandardized) variables across all three networks, with statistically significant breakpoints indicated by vertical red lines, is given in Figure 1.

As is evident from Figure 1, all three varieties of centralization were generally low in the reply and mention networks for the majority of the study period, while assortativity was negative (suggesting a disassortative network in which well-connected users were more likely to be connected to poorly connected peers and vice versa) and entropy was close to 1, suggesting highly egalitarian levels of representation (i.e., the sheer number of replies, retweets, or original tweets in the discourse were quite evenly distributed across different authors). These features exemplify an adhocratic organizational structure [34,35], which makes intuitive sense given the free-flowing nature of discourse on Twitter during this period.

In contrast, inbound degree centralization and betweenness centralization were much more volatile in the retweet network, reaching much higher peaks on a routine basis, but still with low outbound degree centralization, negative assortativity, and relatively high entropy. This combination of features is not a particularly good match to any organizational configuration [35], suggesting that this may represent a different sort of structure altogether. Specifically, this bears a closer resemblance to a “star” network.

With that in mind, Table 4 indicates that statistically significant breakpoints were identified after week 8 (ending 25 February 2018), week 10 (ending 11 March 2018), week 74 (ending 2 June 2019), week 76 (ending 16 June 2019), week 173 (ending 25 April 2021), and week 175 (ending 9 May 2021).

Across these cases, the breakpoints largely represented momentary “shocks” that manifested as temporary changes in one or more variables, after which they subsequently returned to their prior levels. For instance, the period around the week 8 and week 10 breakpoints was marked by a spike in outbound degree centralization for the retweet network and a corresponding decrease in social entropy (Figure 1). These changes corresponded to a sharp increase in the number of replies made by @BreaveHeart43 in week 6 and the number of retweets made by @nizmycuba, @ngarainstitute, @amyrbrown12_amy, and @nutgraham in weeks 6, 7, 9, and 10, which effectively shifted the organizational system from a highly adhocratic structure toward a more entrepreneurial configuration [35], if only for a few weeks.

Regarding the week 74 breakpoint, the only statistically significant change around that time was a large increase in entropy for the retweet network. The change in this metric was attributable to a surge in retweets of @BernieSanders and @LilNasX during week 75. In the former case, Sen. Bernie Sanders, who at the time was ramping up a campaign for the 2020 U.S. presidential election, made the following widely retweeted comment (Figure 2):

Rapper Lil Nas X expressed similar sentiments about the political and environmental sphere, but instead invoking one of his own tracks to attract a wave of retweets (Figure 3):

These two users in combination effectively became the center of the social media discourse around water contamination for the week in question, drawing in numerous other users who were otherwise minimally engaged in the water contamination discourse such that the existing structure of the relational network was briefly upended. Again, though, the effect was only temporary.

The short surge in betweenness centralization for the reply network and the shifts in outbound degree centralization, between centralization, and entropy within the mention network around the week 173 and week 175 breakpoints, on the other hand, had a much simpler cause: one user, @JingKlaus, made an extraordinary 541 replies to fellow users in week 175, many of which just replicated the same set of anti-Japan messages (often delivered in response to users whose posts did not originate in the United States), regardless of the contents of the original tweet (Figure 4):

While the 541 replies made by @JingKlaus in week 175 represented a clear high point around this time, such that replying activity was dominated by this one user, the pace was unsustainable in the long term, such that the aforementioned metrics returned to their original levels shortly thereafter.

5. Discussion

The goal of this study was to investigate social media data to identify patterns that arise from interactions between local citizens and organizations to develop a clearer understanding of the scope and nature of water contamination as it is presented online.

RQ1 asked about the role of the top 10 most influential users determined by metrics including betweenness, eigenvector, and inbound and outbound centrality, which encompassed both individuals and organizations. The results demonstrated that these influential accounts included a diverse range of entities, spanning political figures, organizations, cause-related actors (such as the EPA and related accounts), educational institutions, and news-focused accounts. Additionally, several individual accounts exhibited high activity levels, contributing to the prominence of their roles in the network.

The network analysis revealed that mainstream news organizations play a critical role in facilitating discourse about water contamination and should thus be a part of information campaigns and emergency response efforts. These organizations often serve as hubs for conversation, particularly in times of public concern or crisis. Their presence in both the reply and retweet networks indicates their influence in not only generating discussion but also disseminating key information. For example, in the reply network, top-10 accounts @FOXNEWS and @CNN held space for commenting (i.e., as demonstrated with the in-bound degree centrality scores for @FOXNEWS and @CNN). Betweenness centrality scores for the news organizations @CNN and @nytimes suggest they continue to facilitate the spread of information among users in the reply network. In the retweet network, @AP, The Associated Press account, and the political journalist @krassenstein were in the top 10 for betweenness centrality, suggesting these news sources hold information and serve as bridges in the flow of the conversations. This same betweenness was seen in the mention network for @CNN and @newyorktimes, as they served as connection points in the flow of information. The Associated Press, CNN, and The New York Times are all estate newsrooms considered high in factual reporting and lower in opinion content [38]. Notably, none of the traditional news organizations appeared in the top 10 for eigenvector centrality, suggesting that though they serve as essential points in the networks, they were not top influencers. Future research should explore how these influential accounts, beyond traditional media outlets, shape public discourse over time and how their collaboration with mainstream media can further amplify critical environmental messages. Additionally, expanding the analysis to include other social media platforms could provide a more comprehensive understanding of the broader information landscape, offering insights into how cross-platform dynamics influence public engagement with water contamination and other environmental health risks.

With respect to RQ2, the original objective of the analysis was to assess the evolution of social media discourse about water contamination to show how the organizational discourse of those interactions evolved over time. This began with an assessment of how the organizational configuration generally appeared based on the values of relevant metrics across the data set. The reply and mention networks appeared adhocratic in nature, replicating similar findings for other online social structures [36]. However, the retweet network was dissimilar from any typical organizational configuration, appearing closer to a “star” network. This sort of network is less commonly observed in real-world social systems. However, it can manifest as fans or followers coalescing around celebrities or other public figures, such as topically relevant social media influencers. In other words, when a sufficiently prominent individual within this social circle posted attention-grabbing content, the lay users engaged with the discourse to varying degrees rapidly jumped on that bandwagon, retweeting the post en masse and thereby manifesting their social connection to the prominent individual. This, again, is not unlike fans flocking around a celebrity, athlete, politician, or other public figure of note.

With that in mind, unlike past studies of platforms such as Wikipedia [36], the present study did not identify any sustained changes in the organizational configuration, as almost all the individual metrics were largely stable across the five-year study period. Rather than revolutionary moments that triggered long-term changes in the discourse, the conversation instead exhibited occasional “shocks” that manifested as momentary departures from the typical activity on this topic, after which the discourse rapidly returned to its original state. Those “shocks” tended to be cases in which a small number of users ascended to extreme—but brief—dominance over the discourse. In some cases, those users amplified the discourse [30,38], such as when @BernieSanders and @LilNasX expressed outrage over relevant public policy changes and the overall state of the environment. In others, the conversation was essentially hijacked by the sheer volume of inflammatory or entirely off-topic content that rapidly emerged in the conversation. Both varieties of individual domination over the discourse have significant implications for global human rights—the former example quantifies the power that public figures such as politicians and celebrities have to shape public opinion about grave human rights issues, even through as little as a single tweet, while the latter demonstrates the ease with which a single individual can subvert such discourse to serve their own agenda. These momentary disruptions, although brief, highlight the vulnerability of social media discourse to manipulation by a small number of highly active users. This dual nature of shocks—both amplifying legitimate discourse and disrupting it—underscores the complexity of online communication dynamics, where influential voices can both enhance and subvert important conversations.

Limitations and Future Research

This study predominantly focused on X (formerly known as Twitter) and issues relevant to water contamination. An examination of multiple social media sites and platforms may be useful to gain a fully comprehensive understanding of the discourses that emerge, thereby building upon the foundational, X-focused insights offered by this study. Additionally, this study focused on the relationships among individuals and organizations through a social network analysis, so further examining the contents of tweets is critical.

The findings suggest that social media discourse on environmental issues like water contamination remains relatively stable over time, punctuated by bursts of heightened activity driven by influential actors. These bursts do not lead to long-term structural changes in the discourse but rather represent temporary fluctuations in network configurations. Future interventions aiming to enhance public engagement with critical environmental issues should take these dynamics into account, recognizing that sustained engagement may require strategies that can address both the amplifying and disruptive impacts of key influencers. Exploring the types of content that drive sustained engagement, rather than momentary spikes, could provide valuable insights into how to foster more consistent discourse on important public health topics like water contamination.