Did the public attribute the Flint Water Crisis to racism as it was happening? Text analysis of Twitter data to examine causal attributions to racism during a public health crisis

Abstract

The Flint Water Crisis (FWC) was an avoidable public health disaster that has profoundly affected the city’s residents, a majority of whom are Black. Although many scholars and journalists have called attention to the role of racism in the water crisis, little is known about the extent to which the public attributed the FWC to racism as it was unfolding. In this study, we used natural language processing to analyze nearly six million Flint-related tweets posted between April 1, 2014, and June 1, 2016. We found that key developments in the FWC corresponded to increases in the number and percentage of tweets that mentioned terms related to race and racism. Similar patterns were found for other topics hypothesized to be related to the water crisis, including water and politics. Using sentiment analysis, we found that tweets with a negative polarity score were more common in the subset of tweets that mentioned terms related to race and racism when compared to the full set of tweets. Next, we found that word pairs that included terms related to race and racism first appeared after the January 2016 state and federal emergency declarations and a corresponding increase in media coverage of the FWC. We conclude that many Twitter users connected the events of the water crisis to race and racism in real-time. Given growing evidence of negative health effects of second-hand exposure to racism, this may have implications for understanding minority health and health disparities in the US.

In April 2014, the state-appointed emergency manager of Flint, Michigan switched the city’s source of drinking water from the Detroit water system to the Flint River to save money ^[1]. The failure to properly treat water from the Flint River resulted in corrosion of pipes and solder in the city’s water distribution system, causing lead and other metals to leach into the water supply ^[2]. Despite complaints from residents about the appearance and taste of the water, the detection of fecal coliform bacteria in the water, and evidence of high lead levels, authorities continued to insist that water from the Flint River was safe to drink and only agreed to reconnect to the Detroit water system after a report documenting high blood lead levels in Flint children was released in September 2015 ^[3,4]. Nearly two years after the Flint water crisis (FWC) began, Michigan governor Rick Snyder declared a state of emergency in Flint, and less than two weeks later, on January 16, 2016, President Barack Obama declared a federal state of emergency ^[3,4].

Coverage of the FWC became prominent in regional and national outlets after the state and federal emergency declarations ^[5]. Noting that Flint is a majority-Black city, many media reports suggested that racism played a role in the events that led to the crisis, as well as the government’s failure to adequately respond to it ^[6,7]. The Michigan Civil Rights Commission substantiated these claims, concluding that racist policies and practices in the areas of employment, housing, and education, as well as racially disparate effects of the state’s emergency manager law, contributed to the water crisis ^[8]. In the aftermath of the crisis, a growing body of academic research has attributed the FWC to structural racism ^[9–13], and Flint residents themselves have described the water crisis as an act of genocide targeting Black residents ^[14]. To date, however, little is known about the extent to which the public attributed the FWC to racism during the months in which it was first unfolding.

This is important to understand because of the role of social media in public opinion formation in the US. Social media has been likened to a guiding force on public opinion around certain events and issues such as the FWC ^[15]. Social media has been found to influence national conversations around issues of social change, for example ^[16]. In addition, it has been used to gauge public opinion while mediating access to the news ^[17]. One of the most salient issues in social media has been recurring events that highlight inequality in the US. Specifically, police-involved killings such as that of Michael Brown in Ferguson, Missouri, which happened the same year as the switch to the Flint River, increased discussions around race and racism. Much of this discussion happened on Twitter ^[18,19]. The association of these shootings with racism has been clear in this research, but it has not been explored in the context of the FWC.

Another reason why this is important to understand is because there is mounting evidence that exposure to vicarious structural racism–defined as witnessing the effects of racist structural conditions or practices on members of one’s own racial or ethnic group ^[20,21] –may negatively impact the health of minoritized people ^[22–24]. Much of the research in this area has used a quasi-experimental design comparing health indicators before and after a racialized event, such as an immigration raid ^[22] or police shootings of unarmed Black people ^[23]. Applying this study design to the FWC is challenging for two reasons. First, although Flint is a majority-Black city, residents of all racial and ethnic groups were negatively impacted by the water crisis, which is not the case in an immigration raid or a police shooting. For this reason, it is possible that the public did not attribute the FWC to racism. Second, unlike an immigration raid or police shooting, which are discrete events, the FWC unfolded slowly over time. Thus, even if the public did ultimately attribute the water crisis to racism, it is not clear when this connection was made, which is important for establishing timing of exposure in a quasi-experimental study. An analysis of Twitter data can facilitate future research on the health effects of indirect exposure to the FWC by establishing whether and when the public attributed the water crisis to racism.

Previous research has shown that data from Twitter and other social media platforms can be used to track public discourse about events like the FWC close to real time ^[25–27]. In this study, we used natural language processing tools to analyze nearly six million Flint-related tweets posted between April 2014, the month in which Flint’s source of drinking water was switched to the Flint River ^[1], and June 2016, the month in which the Environmental Protection Agency declared that the water in Flint was safe for everyone to drink ^[28]. We examined the following research questions:

• 1Did key developments in the FWC correspond to quantitative changes in the conversation about race and racism on Twitter?

We hypothesized that the total number and percentage of tweets that included terms related to race and racism increased immediately following key developments in the FWC, including the boil water advisories in August and September 2014, the lead warnings in September and October 2015, and the emergency declarations in January 2016, which made the FWC part of the national news cycle (see Table 1 for a timeline of events in the FWC). To the extent that the public attributed these key developments in the water crisis to racism, we would expect to see an increase in the number and percentage of tweets mentioning terms related to race or racism following these developments.

• 2How did changes in the conversation about race and racism on Twitter compare to changes in the conversation about other topics hypothesized to be related (i.e., water and politics) and unrelated (i.e., weather) to the FWC?

Table 1.

Timeline of key events in the Flint Water Crisis, April 2014 - June 2016

Date	Event
April 25, 2014	The city of Flint switches the source of its drinking water from the Detroit Water and Sewerage Department to the Flint River
April 2014	Flint residents raise complaints about water from the Flint River, noting rashes, hair loss, and other health concerns
August 14, 2014	A boil water advisory is issued for a neighborhood on the west side of Flint after fecal coliform bacteria is detected in the water supply.
September 5, 2014	A second boil water advisory is issued after fecal coliform bacteria is again detected in the water supply
October 1, 2014	The Michigan Department of Environmental Quality (MDEQ) issues a report discussing potential causes of contamination in Flint’s water
October 2014	The General Motors plant in Flint stops using city water after noting corrosion of engine parts
January 2, 2015	Flint residents are warned by city officials that disinfectant byproducts in the water may cause health issues; elderly residents and those with young children are told to consult with their doctors about the risks
January 12, 2015	City officials decline an offer from the Detroit Water and Sewerage Department to waive the fee to reconnect the city with water from Lake Huron
January 21, 2015	Flint residents bring containers of discolored water to a community forum, and the Detroit Free Press reports on health problems thought to be caused by the water
February 2015	Governor Rick Snyder announces $2 million in funding to fix problems in Flint’s pipes and sewers
February 26, 2015	The MDEQ is notified by the Environmental Protection Agency (EPA) that dangerous levels of lead were detected in the water at the home of Flint resident Lee-Anne Walters
March 18, 2015	Lee-Anne Waters contacts the EPA after a test reveals lead levels in the water at her home that are more than 25 times higher than the EPA limit
March 23, 2015	State-appointed emergency manager Jerry Ambrose overrules the Flint city council’s 7-1 vote to reconnect to the Detroit Water and Sewerage Department
June 5, 2015	Clergy and activists file a lawsuit against the city of Flint, charging that the water presents a risk to residents’ health
June 24, 2015	An EPA memo citing results of lead testing by scientists from Virginia Tech University warns that the city is not properly treating water from the Flint River
July 9, 2015	A video about the lead in Lee-Ann Walters’s water is posted by the American Civil Liberties Union (ACLU)
July 13, 2015	A spokesperson from the MDEQ tells Michigan Public Radio that testing of 170 homes in Flint revealed that lead-contaminated drinking water is not a widespread problem
July 22, 2015	Dennis Muchmore, chief of staff to Governor Snyder, emails the Department of Community Health to inquire about reports of elevated lead levels in Flint’s water
August 17, 2015	State testing reveals elevated lead levels in Flint’s water supply during the first half of 2015, leading the MDEQ to order Flint officials to provide proper corrosion control treatment
September 8, 2015	A preliminary report showing that 40% of Flint homes have elevated lead levels is issued by a team of researchers from Virginia Tech University
September 9, 2015	The EPA says it will help develop a corrosion control treatment for water from the Flint River
September 11, 2015	Virginia Tech University researchers recommend that the state declare that Flint’s drinking water is not safe for drinking or cooking
September 24, 2015	Flint pediatrician Dr. Mona Hanna-Attisha and colleagues release a study showing that the number of children with elevated blood lead levels nearly doubled after the city switched the source of drinking water; state regulators claim that the water is safe
September 29, 2015	Governor Snyder acknowledges that lead-contaminated drinking water in Flint is a problem and promises to take action
October 2, 2015	The Michigan Department of Health and Human Services (MDHHS) verifies findings of elevated blood lead levels in Flint children, and the state begins testing drinking water in schools and providing free water filters
October 8, 2015	Governor Snyder announces that Flint will stop using water from the Flint River after three schools in Flint test positive for elevated lead levels in the water
October 15, 2015	Governor Snyder signs a bill to provide more than $9 million to reconnect Flint to the Detroit Water and Sewerage Department and to provide health services to city residents
October 16, 2015	Flint switches the source of its drinking water back to the Detroit Water and Sewerage Department, and the EPA establishes the Flint Safe Water Task Force
November 4, 2015	A redacted version of the EPA’s final report on high lead levels in three Flint homes is published
November 13, 2015	A class action lawsuit is filed by Flint residents against state and city officials for knowingly exposing them to unsafe water
December 14, 2015	The city of Flint declares a state of emergency
December 29, 2015	MDEQ director Dan Wyant resigns, and Governor Snyder apologizes for the water crisis in Flint
January 5, 2016	Governor Snyder declares a state of emergency in Genesee County
January 12, 2016	The Michigan National Guard begins to distribute bottled water in Flint
January 13, 2016	State health officials report an increase over the past two years in cases of Legionnaires’ disease in Genesee County
January 14, 2016	Governor Snyder asks President Barack Obama to declare an expedited major disaster in Flint
January 15, 2016	An independent review into the water crisis in Flint is launched by Michigan Attorney General Bill Schuette
January 16, 2016	President Obama declares a state of emergency in Flint, allowing the Federal Emergency Management Agency (FEMA) to provide assistance
January 21, 2016	The EPA raises concerns about the construction of the new pipeline to Lake Huron and criticizes the state’s response to the water crisis in Flint
January 27, 2016	A federal lawsuit alleging violation of the Safe Water Drinking Act is filed against the state in Detroit
March 17, 2016	Governor Snyder provides testimony before the House Committee on Oversight and Government Reform
March 23, 2016	A panel appointed by Governor Snyder concludes that the state is responsible for the water crisis due to decisions made by environmental regulators
March 31, 2016	Attorneys from the NAACP and other organizations file a class action lawsuit seeking damages for Flint residents who were affected by the water crisis
April 20, 2016	Criminal charges are filed against Mike Glasgow, Stephen Busch, and Mike Prysby, government employees involved in the water crisis
April 25, 2016	Current and former Flint residents file a class action lawsuit against the EPA alleging negligence
May 4, 2016	President Obama visits Flint to meet with residents
May 4, 2016	Mike Glasgow agrees to cooperate as a witness in the investigation of the water crisis and is released on personal bond following the plea agreement

Sources: https://www.cnn.com/2016/03/04/us/flint-water-crisis-fast-facts/index.html and https://apnews.com/article/1176657a4b0d468c8f35ddbb07f12bec

We hypothesized that the total number and percentage of tweets that included terms related to water and politics–issues that are central to a discussion of the switch to the Flint River and the series of decisions by government officials that exacerbated the crisis–also increased immediately following key developments in the FWC, whereas the total number and percentage of tweets that included terms related to weather did not change following key developments in the water crisis. Examining a topic hypothesized to be unrelated to the water crisis (i.e., weather) allows us to ensure that changes in the volume and proportion of tweets in topics hypothesized to be related to the water crisis (i.e., race and racism, water, and politics) are not due to unobserved time-varying confounders. To the extent that trends in the number and percentage of tweets related to race and racism were similar to trends for tweets related to water and politics but different from trends for tweets related to weather, results would suggest that the public connected key developments in the FWC to racism.

• 3How did the distribution of positive and negative sentiment in conversations about race and racism compare to sentiment in conversations about water, politics, and weather (control group)?

We hypothesized that sentiment scores in tweets including terms related to race and racism, water, and politics would be more negative than sentiment scores in tweets including terms related to weather. While we expected sentiment scores to be negative for all topics related to the FWC, we hypothesized that sentiment scores would be the most negative for tweets that included terms related to race and racism, reflecting negative emotional reactions to a preventable public health crisis in a majority-Black city with a long history of segregation and disinvestment.

• 4Did the most commonly used words in Flint-related tweets refer to race and racism?

We hypothesized that terms related to race and racism would appear among the 50 most common words (i.e., unigrams) and word pairs (i.e., bigrams) in our data set. Whereas the first two research questions seek to establish whether there is any evidence that the public attributed the water crisis to race and racism, the fourth question seeks to understand the centrality of racism in discussions about the water crisis. To the extent that the most commonly used words and word pairs in Flint-related tweets referred to race and racism, this would suggest that racism was central to discussions about the water crisis on Twitter.

• 5Did the most commonly used words and word pairs change over time as traditional media coverage of the FWC increased?

We hypothesized that words and word pairs including terms related to race and racism became more common over time. While early developments in the water crisis, such as the 2014 boil water advisories, received relatively little traditional media coverage, later developments, including the 2015 lead warning and the 2016 emergency declarations, received considerable news coverage ^[5]. As previously noted, a number of media reports in 2016 suggested that racism played a role in the events that led to the crisis, as well as the government’s failure to adequately respond to it ^[6,7]. For this reason, we expected to find that words and word pairs including terms related to race and racism increased over time, particularly after the 2016 emergency declarations.

• 6Did latent topics derived from tweets about Flint and the FWC include terms related to race and racism?

Finally, we hypothesized that terms related to race and racism would be evident in the latent (i.e., hidden, or not directly observable) topics in our data set. In addition to the analysis of unigrams and bigrams, topic modeling provides another way to assess the centrality of race and racism in discussions about the water crisis on Twitter.

Background

Despite the fact that lead has been recognized as a neurotoxin for thousands of years ^[29], it was an integral part of the US economy for many decades. In addition to being used as an additive in paint and gasoline, lead was used for much of the water piping laid in US cities prior to 1986, when Congress banned the use of lead pipes ^[30]. Legacy water systems in older cities still have many lead pipes delivering water to residents. As a result of demographic changes that took place in these cities following World War II, racial minorities and the poor have borne a disproportionate burden of lead exposure ^[31].

The links between race, class, and exposure to lead were so strong that combating lead in the environment was seen as part of the Civil Rights Movement and the War on Poverty ^[31]. The Clean Air Act of 1970 made regulating lead the responsibility of the newly formed Environmental Protection Agency (EPA). Not only did poor people and racial minorities tend to live in homes with lead paint and lead pipes, but they were also more likely to live next to highways built during the urban renewal era ^[31]. Cars traveling on these highways burned leaded gas, sending lead particles into the air. As the urban environment became less healthy, neighborhood racial composition became a reliable predictor of the presence of pollutants such as lead ^[32].

The contamination of Flint’s water supply with high levels of lead between April 2014 and January 2016 exposed these longstanding racial inequities in exposure to pollutants in US cities. Political scientist Kemi Fuentes-George argued that despite the poor economy contributing to the events that led to the FWC, “…we cannot meaningfully discuss what is happening with water in Flint without discussing racism. Simply put, the economic situation in Flint and the logic that made providing poisoned water seem like a reasonable idea would not have occurred without racism” ^[33].

Cities like Flint have struggled for more than half a century due to shifts in the US and global economies. As a result, states including Michigan installed emergency managers who were empowered to oversee failing cities and overrule elected local leaders. Many times these managers have one objective, which is to balance the budget, and all other considerations are secondary ^[34]. This was the case in Flint, Michigan more than one time. The last time led to the decision on April 25, 2014, to switch the source of the city’s drinking water from the Detroit water system to the Flint River to save money ^[32,35].

Problems with the switch developed almost immediately, especially since those managing the water treatment plant decided not to put an anti-corrosive agent in the water ^[2]. This would have prevented the already very corrosive Flint River water from corroding the pipes. The pipes, predominantly made from lead, corroded in the water and leached lead into the water supply. The only thing that reversed the decision was the revelations of a pediatrician, Dr. Mona Hanna-Attisha, who reported alarmingly high blood-lead levels of children in Flint ^[36]. The resulting national public outcry led state officials to approve a switch back to the Detroit water system on October 16, 2015. Between the initial switch to the Flint River and the switch back to the Detroit water system 18 months had passed. On February 17, 2017, one year after the emergency declaration in Flint, the Michigan Civil Rights Commission concluded that the FWC was the result of systemic racism ^[8].

MATERIALS AND METHODS

Data

Data for this study were obtained from Twitter, Inc. using a detailed search of tweets related to the Flint Water Crisis. The full data set includes 5,727,977 tweets posted between April 1, 2014, and June 1, 2016, using the following hashtags and keywords related to Flint and the FWC: (1) #flint, flint; (2) #flintlivesmatter, flintlivesmatter, flint lives matter; (3) #istandwithflint, istandwithflint, i stand with flint; (4) #helpflint, helpflint, help flint; (5) #prayforflint, prayforflint, pray for flint; (6) #flintwatercrisis, flintwatercrisis, flint water crisis; (7) #fwc, fwc; (8) #poisoningflint, poisoningflint, poisoning flint (9) #flintwater, flintwater, flint water. We excluded 397,666 tweets that were not in English, leaving 5,330,311 tweets. The number of unique tweets, excluding retweets, was 2,808,185. This data set is not publicly available as it is based on Enterprise data acquired through an agreement with Twitter, Inc. Information on how to obtain the Tweet IDs and reproduce the analysis is available from the corresponding author upon request.

Data Preprocessing

All tweets were preprocessed using Python. Specifically, URLs, numbers, punctuation marks, and emojis were removed, as were duplicate whitespaces and any words shorter than three characters. Finally, all text was converted to lower-case for uniformity.

Data Analysis

Did key developments in the FWC correspond to increases in the number and percentage of tweets about selected topics?

Our first research question asked whether responses to key developments during the first two years of the FWC were reflected in quantitative changes in the conversation about race and racism on Twitter. To address this question, we began by selecting a set of terms related to race and racism, as shown in Table 2. Next, we created word filters to identify tweets that included any of these terms. We then used the date information provided in the tweets to track daily counts and percentages of tweets that included terms related to race and racism.

Table 2.

Total and unique tweet counts by topic

Topic	Number of all tweets (unique tweets)	Filter terms
Race and Racism	138,382 (62,474)	white people, black people, african american people, people of color, poc, bipoc, white kids, black kids, african american kids, white children, black children, african american children, white folks, black folks, african american folks, whites, blacks, african americans, ethnicity, race, racism, racist, racial, discriminate, discrimination, genocide, bigotry, bias, segregation, illiberality, partiality, hate crime, supremacist, supremacy, apartheid, xenophobia, xenophobe, nationalism, nationalist, fascism, fascist, blacktwitter, blacklivesmatter, blm, blklivesmatter, black lives matter, black twitter, flint lives matter, flintlivesmatter
Water	2,358,564 (987,996)	water, drink, crisis, lead, bath, shower, detroit river, flint river
Politics	1,344,102 (531,485)	gov, governor, mayor, republican, democrat, snyder, rick, republicans, democrats, politics, political, deq, department of environmental quality, emergency manager, city manager, michigan house, michigan senate, legislature, legislative, epa, environmental protection agency, mdhhs, health department, michigan department of health and human services, genesee county, state of emergency, obama, president, earley, kurtz, ambrose, lyon, debate, clinton, sanders
Weather	76,730 (43,142)	weather, forecast, sunny, rain, tornado, storm, snow, humid, cloud, balmy, hail, temperature, dewpoint, precipitation, barometric, windy, freezing

For our second research question, we asked how changes in the conversation about race and racism on Twitter compared to changes in the conversation about other topics hypothesized to be related (i.e., water and politics) and unrelated (i.e., weather) to the FWC. Search terms for each of these topics are shown in Table 2, which also includes the number of tweets (total and unique) containing terms related to all four topics. A single tweet was counted under multiple topics if it contained terms included in more than one filter. For example, a tweet that included the terms “lead” and “governor” was counted under both the water and politics topics.

Do FWC-related tweets focused on race and racism exhibit different emotional valence than others?

Our third research question asked how the sentiment in conversations about race and racism compared to the sentiment in conversations about water, politics, and weather. To examine this question, we calculated sentiment scores using Textblob, a Python library for processing textual data ^[37].

How common were word groupings related to race and racism among FWC-related tweets?

Our fourth research question asked whether the most commonly used words and word pairs in tweets about Flint and the FWC included terms related to race and racism. To examine this question, we used the N-gram method, which considers a sliding window of size n in a given text to reveal patterns in the text and to extract features for classification tasks ^[38]. For different values of n, unigrams (n=1), bigrams (n=2), trigrams (n=3), and longer co-occurring adjacent words can be found and analyzed.

To identify the most common unigrams and bigrams in our data set, we started by tokenizing the cleaned text and removing stopwords. Stopwords are commonly used words such as “the”, “yes”, “what”, and “you” that do not add significant value to the contextual meaning of a given document ^[39]. We used a list of predefined stopwords ^[40] from the NLTK library of Python ^[41] and added a number of additional stopwords that were found to be highly prevalent in our dataset (see Table 3 for the extended list of stop words). After removing the stopwords, we identified the 50 most common unigrams and bigrams in the set of all tweets. We used the built-in function “nltk.bigrams” from Python’s NLTK library to identify the bigrams. Next, we determined the number of unigrams that included a filter term from the “race and racism” topic (as shown in Table 2) and the number of bigrams that included at least one filter term from the “race and racism” topic.

Table 3.

List of extended stopwords

job, amp, rt, im, via, https, mi, lol, omg, haha, hey, ad, http, can, gt, rap, be, to, smh, know, everyone, everybody, lot, wow, something, dont, pic, repost, profile, twitter, pm, am, na, ta, rd, ave, iot, news, post, youtube, video, watch, follow, like, retweet, tweet, following, away, back, full, always, never, first, second, today, tonight, year, day, tomorrow, weekend, time, get, new, big, small, good, bad, well, best, many, already, top, yet, could, let, please, they, i, even, would, will, might, may, next, guy, lady, hot, back, way, soon, journal, detroitnews, yeah, yes, no, nothing, one, two, three, four, really, people, real, saginaw, baycity, abcwjrt, nbc, detroit, live, genesee, fwc, flint, nice, must, much, bitch, town, great, far, half, hard, true, great, man, woman, wife, husband, city, male, cool, happy, female, date, month, group, ever, still, single, captain, absolutely, ng, aint, lmao, about, bout, yall, thing, night, right, city, state, thanks, feel, think, look, say, make, see, talk, write, ask, call, take, thank, hear, tell, say, put, work, go, listen, use, come, show

Our fifth research question asked whether the most commonly used words and word pairs changed over time as traditional media coverage of the FWC increased. To examine this question, we extracted the date information from each tweet and created three-month intervals to identify the 50 most common unigrams and bigrams in each time interval. We then determined whether the number of unigrams and bigrams that included one or more filter terms from the “race and racism” topic increased over time.

What were the key topics of Twitter conversations around the FWC?

Our sixth question asked whether latent topics derived from tweets about Flint and the FWC included terms related to race and racism. We hypothesized that terms related to race and racism were evident in the latent topics in our data set. To test this hypothesis, we performed topic modeling, an unsupervised machine learning algorithm that can reveal topics from unstructured text ^[42]. For this analysis, we started with the tokenized data set that excluded stop words, as described above in the N-gram analysis. Additionally, we removed collection words (i.e., words used when collecting data from Twitter), including “flint”, “water”, “crisis”, “flintwatercrisis”, “watercrisis”, “michigan”, “flintmichigan”, “flintwater”, “flint water crisis”, and “water crisis”, since they would dominate the topics to be discovered otherwise. Next, we applied part-of-speech tagging, lemmatized the text, and converted the lemmatized data. To identify topics in the list of lemmatized tweets, we used BERTopic, which relies on pretrained sentence transformers to create dense clusters and offers contextual information, allowing for easily interpretable topics, as opposed to conventional topic modeling techniques such as Latent Dirichlet Allocation (LCA), which uses bag-of-word representation^[43]. There are four key steps in BERTopic: converting each document to its embedding representation using pre-trained embedding models, reducing dimensionality using UMAP algorithm to optimize clustering ^[44], clustering texts in groups that have similar meaning using HDBSCAN ^[45] and extracting topic representations using a custom class-based variation of TF-IDF ^[46]. Each topic modeling technique has its strengths and weaknesses. We preferred BERTopic due to its increased performance on short and unstructured text such as Twitter data as shown in earlier studies ^[47,48].

RESULTS

Daily Tweet Counts and Percentages

First, we examined daily counts of tweets that included terms related to race and racism, water, and politics in order to determine whether key developments in the FWC corresponded to quantitative changes in the conversation about topics hypothesized to be related to the water crisis. We also examined daily counts of tweets that included terms related to weather, a topic hypothesized to be unrelated to the water crisis. As shown in Figure 1, the frequency of tweets was greatest for terms related to water, followed by terms related to politics, race and racism, and weather. A small peak in water-related tweets appeared around October 2015 after three key developments in the FWC timeline: (1) the Michigan Department of Health and Human Services (MDHHS) confirmed Dr. Mona Hanna-Attisha’s report of high blood lead levels in Flint children; (2) the Michigan Department of Environmental Quality (DEQ) announced that three Flint schools had high lead levels in their water; and (3) Flint switched back to the Detroit water system. Another small peak in water-related tweets appeared after the city of Flint declared a state of emergency in December 2015. For all topics except weather, a dramatic peak in the number of tweets appeared in January 2016, following the state and federal emergency declarations in Flint, which garnered substantial national media coverage ^[5]. Other significant peaks in tweets related to race and racism, water, and politics appeared in March 2016, when Governor Snyder testified before the House Committee on Oversight and Government Reform, and in April 2016, after attorneys from the NAACP and other organizations filed a class action lawsuit on behalf of those affected by the water crisis and criminal charges were filed against state officials. In contrast to these results, tweets that included terms related to weather appeared to follow a seasonal pattern unrelated to the water crisis. A pair-wise analysis of daily tweet counts also showed that weather-related tweets had little correlation with other topics, whereas remaining topics exhibited a higher collinearity (Supplementary Figure 1).

Figure 1.

Daily count of tweets by topic area during the period from April 1, 2014 - June 1, 2016 with the enlarged time period starting from January 1, 2016

Next, we examined the daily percentage of tweets that included terms related to our topics of interest. In contrast to daily counts, daily percentages account for differences in the total number of tweets at different stages of the water crisis. Consistent with the findings for daily volume of tweets described above, the percentage of tweets that included terms related to water increased sharply around October 2015, then declined somewhat over the next two months, and increased even more dramatically in January 2016 (see Figure 2).

Figure 2.

Daily percentage of tweets by topic area during the period from April 1, 2014 - June 1, 2016 with the enlarged time period starting from September 1, 2015

Between January and June of 2016, water-related tweets accounted for 53% of daily tweets on average. Tweets related to politics followed a similar pattern, accounting for an average of 28% of daily tweets in the six months following the January 2016 emergency declarations. Although tweets related to race and racism accounted for a smaller percentage of tweets compared to those related to water and politics, we observed significant increases following key developments in the water crisis. For example, the highest peak, accounting for around 26% of daily tweets, occurred in May 2016, after President Obama visited Flint to meet with residents. Four other peaks, accounting for around 10% of daily tweets, occurred in February 2015, after the EPA found dangerous levels of lead in the water at the home of a Flint resident and Governor Snyder announced a $2 million grant to address problems in the pipes and sewers; in March 2015, after Flint’s emergency manager overruled the city council’s decision to stop using water from the Flint River; in December 2015, after the city of Flint declared a state of emergency; and in April 2016, after criminal charges were filed against state officials. The percentage of weather-related tweets showed random fluctuations across the 27-month time period, ranging from around 0–25% of daily tweets.

Sentiment Analysis

Using sentiment analysis, we compared the mean polarity of tweets related to race and racism, water, politics, and weather (see Figure 3). We found that mean polarity was positive (i.e., greater than 0) for each topic, with an overall mean polarity for all tweets of 0.14. Weather-related tweets had the highest mean polarity (0.17), followed by water-related tweets (0.08), tweets related to politics (0.07), and tweets related to race and racism (0.04). The mean polarity for tweets related to race and racism was significantly lower than the mean polarity for all tweets (t=44.98, p<0.0001, df=29925).

Figure 3.

Polarity score by topic

Unigrams

The 50 most commonly used words in the set of all tweets related to Flint and the FWC are shown in Figure 4. Filter terms from the water (e.g., water, crisis, lead) and politics (e.g., governor, obama, debate) topics appeared in this list, but terms from the race and racism topic did not. In examining this list, we observed words such as jobs, truck, company, and driver that are unrelated to the water crisis. To eliminate this source of noise, we excluded “flint” as a keyword and created a restricted dataset containing tweets directly related to the water crisis. We then identified the 50 most commonly used words in this restricted set of tweets. As shown in Figure 5, terms related to water (e.g., water, michigan, crisis, flintwatercrisis, lead) and politics (e.g., snyder, gov, governor, obama, emergency, onetoughnerd [Twitter handle of Governor Snyder], rick, epa, officials, health, manager, mayor) continued to dominate the list, but two filter terms from the race and racism topic (flintlivesmatter, black) also appeared.

Figure 4.

The 50 most common words in the full set of tweets related to Flint and the Flint Water Crisis

Figure 5.

The 50 most common words in the restricted set of tweets directly related to the Flint Water Crisis

In order to see the level of association of certain words with the selected subsets of the data, we compared the frequency of common words between different tweet groups in further exploration of unigrams. Namely, we first compared the tweets that included terms related to race and racism to tweets that included terms related to politics and measured how often they mentioned a set of unigrams found in both groups. Then we performed the same analysis for words used in the tweets related to race and racism and the tweets related to water. While unigrams with similar frequencies are expected to appear close to the 45-degree line in Figures 6 and 7, words that appear above and below the line are prominent for their respective group. In both comparisons, we observed that “flintwatercrisis” appeared closer to the race and racism group due to its higher relative frequency, which suggests that “flintwatercrisis” had a stronger association with the discussion around racism versus politics (see Figure 6) or water (see Figure 7).

Figure 6.

Comparison of the frequency of unigrams in tweets related to race and racism versus politics

Figure 7.

Comparison of the frequency of unigrams in tweets related to race and racism versus water

Bigrams

Using the restricted set of tweets directly related to the water crisis, we identified the 50 mostly commonly used word pairs (bigrams). As shown in Figure 8, bigrams that included terms related to water (e.g., “clean, water”, “lead, water”) and politics (e.g., “gov, accountable”, “arrestgovsnyder, flintwatercrisis”) were most common, but two of the commonly used word pairs (i.e., “black, poor”, “residents, black”) included terms related to race and racism.

Figure 8.

50 most common bigrams in the restricted set of tweets directly related to the Flint Water Crisis

Change in N-gram Frequencies over Time

To determine whether the most commonly used words and word pairs changed over time as media coverage of the FWC increased, we examined change in unigrams/bigrams in three-month intervals. The unigrams related to race and racism did not change significantly over time. Focusing on bigrams that included terms related to race and racism, we found that the bigrams “black, poor” and “residents, black” did not emerge until the interval starting in December 2015 and ending in February 2016. As previously noted, the city, state, and federal emergency declarations were announced during this period, resulting in a dramatic increase in media coverage of the water crisis ^[5].

Topic Modeling

When initiating the BERTopic model, we set the number of topics variable to 50, a decently high value, to prevent topics being merged that should not ^[43]. Then, by inspecting the interactive intertopic distance map while iteratively reducing the number of topics, we concluded that 12 topics provided a good balance between maximizing intertopic distance and maximizing the similarity within each topic. Topics are represented by a set of words, and the 12 topics identified by BERTopic in the FWC tweet corpus did not include keywords directly related to race or racism, as shown in Table 4. Keywords related to water (e.g., showering, drink, lead) and politics (e.g., republican, obama, manager) were common and also reflected some of the key developments in the FWC including the federal emergency declaration by President Obama. Overall, these results suggest that rather than defining a distinct topic within the set of FWC-related tweets, references to race and racism were common across all topic areas.

Table 4.

Latent topics and associated keywords from the restricted set of tweets directly related to the Flint Water Crisis

Topic	Keywords
1	‘showering’, ‘bathing’, ‘quality’, ‘glass’, ‘obama’, ‘republican’, ‘ripped’, ‘party’, ‘improve’, ‘drinking’
2	‘bankruptcy’, ‘dp’, ‘conference’, ‘mackinac’, ‘report’, ‘abc’, ‘replacing’, ‘estimate’, ‘double’, ‘culture’
3	‘flintwatercrisis’, ‘crisis’, ‘life’, ‘story’, ‘iunewstalk’, ‘documentary’, ‘death’, ‘happened’, ‘lawsuit’, ‘tragedy’
4	‘flintwatercrisis’, ‘help’, ‘crisis’, ‘problem’, ‘safe’, ‘chlorine’, ‘byproduct’, ‘disinfecting’, ‘drink’, ‘obama’
5	‘drink’, ‘president’, ‘obama’, ‘safe’, ‘filtered’, ‘scientist’, ‘work’, ‘ackn’, ‘flesher’, ‘writer’
6	‘foundation’, ‘million’, ‘crisis’, ‘ten’, ‘pledge’, ‘garden’, ‘waste’, ‘worry’, ‘showering’, ‘president’
7	‘researcher’, ‘improving’, ‘reassure’, ‘filtered’, ‘report’, ‘drink’, ‘president’, ‘led’, ‘cost’, ‘energy’
8	‘flintwatercrisis’, ‘bonnie’, ‘reimburse’, ‘crisis’, ‘finding’, ‘handling’, ‘agency’, ‘lead’, ‘katrina’, ‘raise’
9	‘report’, estimate’, ‘double’, ‘cost’, ‘replacing’, ‘pipe’, ‘minimum’, ‘per’, ‘increased’, ‘wage’, ‘grumpymichael’
10	‘agree’, ‘flintwatercrisis’, ‘actudave’, ‘stamp’, ‘rubber’, ‘party’, ‘check’, ‘thing’, ‘go’, ‘ethandolan’
11	‘hearing’, ‘committee’, ‘flintwatercrisis’, ‘dem’, ‘interview’, ‘jimananich’, ‘senate’, ‘manager’, ‘crisis’, ‘emergency’
12	‘monahannaa’, ‘talk’, ‘dare’, ‘badsummercampnames’, ‘flintwatercrisis’, ‘random’, ‘request’, ‘love’, ‘hug’

DISCUSSION

The Flint Water Crisis was an avoidable public health disaster that has profoundly affected the city’s nearly 100,000 residents, a majority of whom are Black ^[49]. Although the role of racism in the events leading up to the crisis, as well as the government’s failure to adequately respond to the crisis, has been acknowledged by the news media ^[6,7], the Michigan Civil Rights Commission ^[8], academic researchers ^[9–13], and Flint residents ^[14], little is known about the extent to which the general public attributed the water crisis to racism. Building on previous research, which has shown that data from social media platforms can be used to track public discourse about events like the FWC in real time ^[25–27], we used natural language processing tools to analyze nearly six million Flint-related tweets posted between April 1, 2014, and June 1, 2016. Overall, there is some evidence that Twitter users connected the events of the water crisis to race and racism. However, this was not a dominant theme when compared to other potentially relevant topics, including water and politics.

First, we found that key developments in the FWC corresponded to increases in the number and percentage of tweets that included terms related to race and racism. The largest increase in the number of tweets occurred in January 2016, following widespread media coverage of the emergency declarations in Flint. Meaningful increases in the number of tweets that included terms related to race and racism were also seen in March 2016, when Governor Snyder testified before the House Committee on Oversight and Government Reform, and in April 2016, after attorneys from the NAACP and other organizations filed a class action lawsuit on behalf of those affected by the water crisis and criminal charges were filed against state officials. While we found that the percentage of tweets that included terms related to race and racism also increased following key developments in the water crisis, the specific events differed somewhat. For example, meaningful increases in the percentage of tweets were seen in March 2015, when Flint’s emergency manager overruled the city council’s decision to stop using water from the Flint River and in May 2016, when President Obama visited Flint to meet with residents.

When comparing changes in the conversation about race and racism on Twitter to changes in the conversation about other topics hypothesized to be related to the water crisis, including water and politics, we found similar patterns, with spikes after key developments in the water crisis. In contrast, we found that changes in the conversation about weather, a topic hypothesized to be unrelated to the FWC, followed a different pattern. Based on these results, we conclude that some Twitter users attributed the water crisis to racism. We note, however, that the number and percentage of tweets that included terms related to water and politics were much greater than the number and percentage of tweets that included terms related to race and racism. This suggests that racism was a relatively minor theme in Twitter conversations about the water crisis.

Sentiment analysis revealed that mean polarity was positive in the full set of tweets and in the subsets of tweets related to race and racism, water, politics, and weather. This was somewhat surprising given the negative context of the water crisis, particularly as it relates to politics. Consistent with expectations, however, mean polarity was lowest for tweets related to race and racism, and tweets with a negative polarity score were more common in this subset of tweets when compared to the full set of tweets.

When examining the full set of tweets, terms related to race and racism were not among the most commonly used words or word pairs. After restricting the data set to tweets that were directly related to the water crisis, we found that “flintlivesmatter” and “black” were among the 50 most commonly used words, while “black, poor” and “residents, black” were among the 50 most commonly used word pairs. Interestingly, word pairs including terms related to race and racism did not appear until after the emergency declarations and the corresponding increase in media coverage of the water crisis ^[5]. This may suggest that public perceptions about the role of racism in the water crisis were shaped by media reports.

Independent from the timeline, unigram analyses also revealed that people were more likely to describe the situation in Flint as a crisis when also commenting on race and racism versus politics or water. Another intriguing finding that emerged from the analysis of unigrams was that “poisoned” and “poisoning” were among the most commonly used words. The use of these words suggest that the public may have perceived that the water crisis was an act of intentional harm. In support of this interpretation, a recent study found that over 50% of a sample of Michigan women who were not directly affected by the water crisis agreed or strongly agreed that the water crisis happened because government officials wanted to hurt people in Flint ^[50].

An analysis of change in the most common bigrams over time revealed that, in the first interval (April 1, 2014 – June 30, 2014), “howard, croft” (former director of the Flint Department of Public Works) was used in 5% of tweets. This word pair almost disappeared in the second interval (July 1, 2014 – September 28, 2014), while the word pairs “crowdfunding, helpflint” and “helpflint, pls” reached 8% each. This suggests that people started to organize social media campaigns to ask for help early in the water crisis. Interestingly, from interval 5 (March 28, 2015 – June 25, 2015) to interval 6 (June 26, 2015 – September 23, 2015), the word pair “courtroom, flintwater” decreased from 2% to almost 0%, while word pairs “waterjusticejourney, wjjstart” and “waterjusticejourney, detroitflintho” increased from almost 0% to 8% and 5% respectively. This is a change showing how the public started to ask for justice collectively after June 2015. Between interval 7 (September 24, 2015 – December 22, 2015) and interval 8 (December 23, 2015 – March 22, 2016), the word pairs “black, poor” and “residents, black” became dominant. The word pairs “sign, petition”, “flintwatercrisis, arrestgovsnyder”, “hold, gov”, and “accountable, poisoning” reached 7%, 4%, 3% and 3% respectively. The conversation moved to signing a petition and holding the governor accountable for poisoning Flint residents in this interval. This change corresponded to emergency declarations and national media coverage around January 2016.

Finally, our topic modeling analysis did not identify any latent topics that included terms directly related to race or racism. One topic did, however, include the term “katrina”, which might be a reference to Hurricane Katrina. Although Hurricane Katrina was a natural disaster and the FWC was a human-made disaster, some have suggested that the inadequate government response to each was due, in part, to the racial composition of the affected populations ^[51,52].

Strengths, Limitations, and Directions for Future Research

A key strength of this study was the use of Twitter data, which allowed us to examine what people were saying about the water crisis in real time. This is advantageous because people may not accurately recall how they felt about an event when asked months or years later, as is often the case in surveys. Another limitation of surveys is that the wording of survey questions may influence responses. For example, respondents may agree that racism played a role in the water crisis when directly asked, despite not having previously thought about the water crisis in the context of racism. The analysis of Twitter data allowed us to avoid this problem, and results may facilitate future research on the health effects of indirect exposure to the FWC by establishing when the public attributed the water crisis to racism.

The decision to examine all tweets that mentioned Flint, rather than examining tweets directly related to the water crisis, had advantages and disadvantages. The advantage of this approach is that it allowed us to compare trends in tweets related to topics that were hypothesized to be related to the water crisis (i.e., race and racism, politics, water) to trends in tweets related to a topic that was hypothesized to be unrelated to the water crisis (i.e., weather). We were able to demonstrate that tweets related to race and racism followed the same pattern as tweets related to politics and water, while tweets related to weather followed a different pattern that did not correspond to key developments in the water crisis. The disadvantage of this approach is that it introduced a lot of noise, which negatively impacted some of the planned analyses. To address this problem, we restricted the dataset to tweets that were directly related to the water crisis for the N-gram and topic modeling analyses.

In addition to examining trends over time in the number of tweets related to our topics of interest, we also examined trends in the percentage of tweets related to each topic. This approach allowed us to account for changes over time in the total volume of tweets and revealed that although the number and percentage of tweets that included terms related to race and racism both increased following key developments in the water crisis, the specific events that preceded increases in the number of tweets differed somewhat from the events that preceded increases in the percentage of tweets. The analysis of percentages allowed us to identify specific events in the water crisis that appear to have triggered an increase in the conversation about race and racism on Twitter.

An important limitation of this study is that hashtags or keywords related to the water crisis that we were unaware of are missing from our query, resulting in an incomplete and potentially biased sample of tweets. Similarly, bias may have been introduced if we missed terms related to our four topics of interest: race and racism, water, politics, and weather. However, we believe that our approach to selecting search terms, involving extensive exploratory analysis of publicly available tweets, should mitigate these risks. Results may be biased if the quantity and/or quality of filter terms varied significantly across topics. For example, if our list of filter terms for the race and racism topic was less complete than our list of filter terms for the politics topic, then we likely underestimated the number and percentage of tweets related to race and racism relative to the number and percentage of tweets related to politics. We found it much more challenging to identify terms related to race and racism than to identify terms related to the other topics. One reason for this is that people often talk about racism indirectly. For example, someone may be more likely to say “The water crisis wouldn’t have happened in Grosse Pointe” than to say “The water crisis wouldn’t have happened in a city with more white people.” The two statements express the same sentiment, but only the second statement includes a term from our race and racism filter.

In addition to expanding the list of filter terms related to race and racism, future studies should examine retweets to determine whether tweets related to race and racism were more or less likely to be retweeted than tweets related to other topics. An analysis of retweets could also help identify events that preceded increases in information sharing. Another direction for future research is to examine potential geographic differences in tweets. For example, researchers could compare the number and percentage of tweets related to race and racism in Michigan versus other states or in majority-Black cities, like Detroit, versus majority-White cities, like Grand Rapids. Previous research has shown that Black women in Michigan communities outside of Flint were more likely than White women to attribute the water crisis to anti-Black racism ^[50]. For this reason, we might hypothesize that the percentage of Flint-related tweets that included terms related to race and racism was greater in areas where a larger share of the population was Black. Although it is beyond the scope of the current study, future studies should also examine temporal trends in polarity scores to determine whether sentiment in tweets related to race and racism, water, and politics became more negative as news of the water crisis accumulated.

Further, Twitter users tend to be male while living in urban areas with extreme ideological viewpoints ^[17]. This means that those using Twitter may not be representative of the population. This is especially the case since about 70% of social media users report never or rarely posting on the platform ^[17]. Discussions on Twitter have also been associated with creating a ‘buzz’ around an issue ^[53] in which it is more reactive, especially during ‘online firestorms’ in which people only share opinions that align with the majority ^[54]. Despite these caveats, Twitter is still seen as an influencing factor on public opinion and through online discussions can influence voter engagement ^[53]. One might assume that it was these discussions in Flint that also helped frame the current discussion around another water crisis in Jackson, Mississippi, another majority-Black city. Future research may discern these patterns.

Conclusions

The results of this study suggest that Twitter users were aware of the role of structural racism in the FWC. Given the growing body of research demonstrating that vicarious, or second-hand, exposure to racism can negatively impact health ^[20,22–24], our findings could have implications for minority health and health disparities. More work is needed to understand the public health impacts of the FWC, including impacts on those who weren’t directly exposed to the lead-contaminated water in Flint.