Abstract
In 2020, American cities experienced mass protest after fatal police violence against Black Americans. Did police address these protests differently after the murder of George Floyd or the protests in their communities? Did that response vary depending on the racial or ideological makeup of their constituents? We link policing protests and Facebook posts from about 4,000 American police departments to investigate differences in agency responses to protest. Departments posted more about protest after Floyd's death, especially in those communities that also had a policing protest. Republican cities were less likely to post about protests, but racial demographics are not associated with more or less protest posting. Our results are consistent with past research on the durable influence of partisanship on policing attitudes and suggest that protests are a valuable way to alter rhetoric in local communities, but only once there is sufficient national pressure and scrutiny to do so.
Keywords
Chauvin immediately stood and calmly placed his hands behind his back. Imagine where we’d be had George done the same. —City of Fall River Police Department Facebook post, April 21, 2021
1
Hours after a jury found Derek Chauvin guilty of the murder of George Floyd, one police department’s Facebook post was the source of much controversy as it shared a screenshot of a tweet that seemed to blame Floyd for his death. Though the Fall River Police Department quickly took the post down—the officer who posted it apparently intended to share it on his personal page—and apologized to the community, it highlights the power and speed through which social media can facilitate discussion (and controversy) around policing topics. Though police departments have official communication channels to disseminate information to the public and the media, they have also increasingly used another resource in their communication toolkit: social media. In this article, we examine the public and strategic behavior of American police agencies using their Facebook posts. Such an approach offers wide coverage of agencies, both large and small, from across the country—the first study of its kind to do so.
Social media offers police departments a tool to promote a positive image of their officers (Cartwright and Shaw 2020; Dong and Wu 2022; Edwards et al. 2021; Jeanis 2020; Mayes 2021; Rossler 2019). As of 2015, 94% of law enforcement agencies used Facebook, primarily to inform the public of safety concerns and to facilitate citizen engagement (Kim et al. 2017; Tiry, Oglesby-Neal, and Kim 2019). Further, social media platforms like Facebook offer departments a way to influence the way the media covers them (Grygiel and Lysak 2021).
How police use social media has been studied in several contexts. Some studies focus on agencies within one state, region, or city (Beshears 2017; Dewald 2023; Edwards et al. 2021; Hu and Lovrich 2021; Jeanis, Muniz, and Molbert 2021; Livingstone 2022; Patrick and Rollins 2022), and others examine the behavior of agencies from outside the United States (Cartwright and Shaw 2020; Czudnochowski and Ludewig 2023; Fielding 2023; Kane 2024; Ralph et al. 2022; Wood 2020). Although most of these studies focus on Twitter (now called “X”) or Facebook, others examine police use of Instagram (Sjoberg, Cassinger, and Gambarato 2024), YouTube (Walby and Joshua 2021), and TikTok (Babic and Simpson 2025). Here, we contribute to a growing subset of the literature that examines the behavior of a large number of agencies simultaneously (see e.g., Dong and Wu 2022, who examined the 115 largest agencies on Twitter). Our article, to the authors’ knowledge, is the most comprehensive study of police social media behavior to date.
We focus on one aspect of police behavior on Facebook: whether and how agencies use it to discuss the 2020 policing protests. This is especially important in light of the 2020 mass protests in response to the police killings of George Floyd and Breonna Taylor that had a seismic impact on Americans across the country: these killings and their aftermath caused liberalization of racial attitudes, prompted resignations of police officers, and worsened emotional and mental health, especially for Black Americans (Eichstaedt et al. 2021; Mourtgos, Adams, and Nix 2022; Reny and Newman 2021), providing more evidence that widespread protests can influence political outcomes like voting behavior (Wasow 2020). The wide scope of these protests has prompted a flurry of interest in the effects of lethal violence against Americans of color—and in particular, Black Americans—and the response of politicians and bureaucrats to these concerns. According to the Crowd Counting Consortium, there were 14,000 policing protests across the continental U.S. in 2020. 2 As much as protesters and activists were making arguments to “defund the police,” others were defending police officers. And of course, this was all happening in the midst of a global pandemic.
We argue that generally, and especially in this context, Facebook posts are consequential because they represent public and strategic choices by the police—street-level bureaucrats—with the ability and power to use lethal force. We think of police departments like a firm or congressional office (see Davis and Russell 2025; Loomis 1979). No matter who does the posting (whether an officer, chief, or communications staff member), we assume that the agency shares similar goals of providing information and responding to their communities. Though all police agencies experienced the same national trends during the summer of 2020, and thus may feel heightened scrutiny or pressure to respond, we develop theoretical expectations about how features of their communities shape the varied communication decisions that they made following these events. An agency's Facebook posts are publicly available for those inside and outside of its community who choose to look for them. However, for most of these agencies, it is likely that their intended audience is their constituents (Dewald 2023; Finklea 2022), though as strategic actors, police departments should consider that what they post will spread beyond their community. Are these agencies responsive to the concerns of their communities? To which parts? We consider how the aftermath of Floyd's murder and the presence of local policing protests influences the rhetoric of these agencies. We also specifically examine how an area's racial and partisan makeup affects posting behavior. Our analysis looks at police activity on one of the world's most used social media platforms, Facebook (Gottfried 2024).
We collect all Facebook posts from about 4,000 municipal police agencies nationwide in 2020 and connect that data to the Crowd Counting Consortium dataset on policing protests in that year. We explore the use of different frames around a protest and the use of positive and negative sentiment in each post's text. First, we find that George Floyd's murder, especially in those places that also had a policing protest, prompted police to post more about protest. We also find that police do respond to constituent pressure, but they do not respond to all members of their communities equally. Rather, we find that they are responsive to the partisanship of their communities but not the racial breakdown. Our results here suggest that protests can indeed be an effective force at altering the rhetoric of the police, though perhaps only once a significant national event provides additional pressure. We provide evidence of these dynamics in the context of the unprecedented scope and spread of protests after George Floyd's death (Reny and Newman 2021), though we expect these dynamics to influence police behavior in the face of other seismic events like the ones that happened in 2020.
Government Responsiveness and the Bureaucracy
One of the canonical questions in studies of democracies considers the responsiveness of government to the will of the public. Primarily, this perspective investigates who gets what and how much in the context of policy outputs, or how policy follows the preferences of the public (Caughey and Warshaw 2018; Stimson, MacKuen, and Erikson 1995). Elites seem to respond to public opinion in the creation of policies, but this responsiveness is conditioned by a variety of other factors, including institutional constraints, partisanship, policy areas, and the kind of government (Einstein and Kogan 2016; Lax and Phillips 2012; Sances 2021; Tausanovitch and Warshaw 2013). Moreover, these perspectives tend to describe the relationship between public opinion and policy change as an endogenous one, in which one responds to the other and vice versa. A politician can adjust a policy change in response to an assumed or known set of citizen preferences, and citizens can respond by signaling “more” or “less” of that policy until the public preference is met (Wlezien 1995).
These studies, however, are mostly characterized by the responsiveness of elected officials to public opinion. Elections are presumed to motivate officials to listen to the public in an effort to secure reelection (Mayhew 2004). Though the strength of this electoral connection is contested (see Bernardi 2020), the nature of election structures the interactions between the public and the government.
We expect bureaucrats to engage differently with the public. They are not elected, nor do they necessarily rely on public support to retain their jobs or climb the career ladder. Though a robust literature discusses the interrelationships between the bureaucracy and its principal, the executive, here we consider another essential actor: the street-level bureaucrat. Street-level bureaucrats are those lower-level workers who are in charge of the day-to-day implementation of the policy set by the legislature or the upper levels in the bureaucracy (Lipsky 1980). Rather than a passive actor who merely enforces their principal's wishes, bureaucrats exert considerable control over public policy. As a result, bureaucrats’ preferences and policy choices then contribute to variation in government responsiveness and policy choice across political contexts (Boldt and Boyd 2018; Einstein and Glick 2017; Percival, Johnson, and Neiman 2009). In particular, discretion is identified as one of the key determinants of whether and to what degree bureaucrats respond to the public (Riccucci, Van Ryzin, and Lavena 2014), and in this case, police officers have the discretion and the means to use social media as a communicative tool in the way that they wish. This makes it an ideal venue to study these dynamics.
We contribute to a growing literature on how police agencies, consequential bureaucratic institutions, use social media to build and maintain a positive image with the public and gain legitimacy (Cheng 2021; Edwards et al. 2021; Mayes 2021; Ralph 2022). A police officer is a canonical example of a street-level bureaucrat, a person who makes public policy with their day-to-day interactions with the public, with much discretion and autonomy. This immense power has serious, if not lethal, consequences for the race-class subjugated neighborhoods and communities most subject to police oversight (Lerman and Weaver 2014; Soss and Weaver 2017).
Police on Social Media
Changes in technology—the fast-paced nature of social media and the Internet, alongside the desire of the public to see immediate responses to questions or crises—places pressure on police departments to respond to the public (Goldsmith 2010; Mawby 1999). The potential advantages of using these new platforms are significant. Social media is associated with higher trust and efforts to increase transparency of these agencies, and can be used as a tool to improve the public image of that agency (Cartwright and Shaw 2020; Hu, Rodgers, and Lovrich 2020; Song and Lee 2016). For which audiences are these public activities intended to signal legitimacy or diffuse difficult situations, if at all? What are the department's incentives for doing so?
Social media, such as Facebook, Twitter/X, or YouTube, is widely used by political actors and individuals. These platforms offer relatively inexpensive tools that political actors can use to benefit them, by achieving their goals and reaching relevant audiences. About 69% of American adults use Facebook, and about 70% of those use it daily (Gottfried 2024). This presents an extremely wide audience for political actors like police departments to engage with and provide information to the public.
Despite its cheapness and ease of use, it is not necessarily the case that elite messages on social media are simply cheap talk. Interviews with police officers and communications staff from multiple countries find that police are strategic about their use of social media (Sjoberg, Cassinger, and Gambarato 2024) and recognize that it is not costless (Dewald 2023). Here, cost does not mean the cost of creating an account or posting, but of the time it takes to successfully use the platform, such as by providing information or responding to events effectively. We expect this to be particularly true during our time period of interest following the murder of George Floyd, when police departments were under additional scrutiny by the public and the media (Dong and Wu 2022).
Communication to constituents and the public is an essential part of government responsiveness, as officials are able to convey their efforts and successes. In direct, or unmediated, forms of communication, the actor has control over the content and timing of their message (Golbeck, Grimes, and Rogers 2010). Examples of direct communication are television and online advertisements, emails, and speeches. Indirect or mediated, communication is that which the actor does not have such control—for example, media coverage of a speech or an event that they organized. Direct communication can be beneficial because it provides the figure or group the ability to strategically craft their message. As they control what they say and when it is shared, they can work to present themselves in the most favorable light possible.
For police departments, social media is a tool to cheaply and directly reach the people that they support and protect. Through their use of social media, police departments bypass the traditional news media and give the agencies more control over their message (Cheng 2021; Colbran 2020; Edwards et al. 2021; Finklea 2022; Jungblut and Jungblut 2022). Despite the potential use of social media as a way to have two-way communication with the communities that they represent, most of what police departments post is one-way communication (Jungblut and Jungblut 2024). Nevertheless, interviews with community members found that they want two-way communication with the police departments that serve them (Beshears, Beshears, and Bond 2019).
Of course, not all of a city's Facebook users will like and follow their local police department's page; it is still in the department's interest to be able to communicate to those in their area who do follow them, to signal that they are protecting them and acting on their behalf. The International Association of Chiefs of Police, for example, has social media fact sheets for agencies and writes that departments can use social media for a “variety of purposes, to include as an investigative tool, for community outreach and engagement, to make time-sensitive notifications, to inform the media, and as a recruitment tool.” 3
As a direct form of communication, researchers have found that police departments use social media in several ways. Through their posts, they give information to their communities (Edwards et al. 2021; Ferguson and Soave 2021; Melekain and Wexler 2013), debunk misinformation (Melekain and Wexler 2013), curate a positive image (Cartwright and Shaw 2020; Czudnochowski and Ludewig 2023; Edwards et al. 2021; Hu, Dong, and Lovrich 2022; Ralph et al. 2022; Rossler 2019), legitimize (Kudla and Parnaby 2018; Ralph 2022), and even humanize police officers (Mayes 2021). Furthermore, they use their accounts to investigate crimes and gather information (Ferguson and Soave 2021; Finklea 2022; Fortin et al. 2024; Kane 2024) and to recruit officers (Walby and Joshua 2021).
In other words, social media has largely been used to reinforce the existing public-facing strategies of the police (Crump 2011). Despite this relative consistency, though, there are documented differences in social media posts across different agencies (Edwards et al. 2021), and these differences could be for a variety of strategic reasons (Meijer and Thaens 2013). Who a police department is speaking to and responsive to in their social media posts matters. We thus expect differences in behavior across agencies.
Responding to George Floyd's Murder and Local Protests
We draw on previous research that shapes our expectations on how both the murder of George Floyd and local protests themselves influence posting behavior. Dong and Wu (2022) compared trends by 115 large American police departments on Twitter immediately before and after George Floyd's murder in 2020 and found that after his death, the departments tweeted more overall, more about unrest in particular, and received greater engagement, but also found variation across agencies. We use Facebook as a lens to evaluate how (if at all) the police respond to this event publicly, and by additionally considering whether or not a city had a policing protest and demographic factors of race and partisanship. Social media is especially useful at managing and addressing the public during emergencies like mass shootings (Fowler 2017) or ongoing crises like the COVID-19 pandemic (Hu, Dong, and Lovrich 2022). It is also used as a tool during crises that explicitly involve the police, like the fatal shooting of a civilian (Hand and Ching 2020).
Following existing research that finds that Floyd's murder did deepen convictions about police and policing (though, perhaps only among liberals; Reny and Newman 2021), our first hypothesis states that, in order to increase transparency, accountability, and goodwill among the public:
We next explore the influence of local policing protests on the posting behavior of American police agencies. Similar to the officer-involved shooting context by Hand and Ching (2020), we argue that protests provide a unique opportunity for the police to manage public responses and concerns about policing practices. Their posts on Facebook provide us with a way to study police behavior. We contribute to scholarship that focuses on how police departments respond to crises in their communities using social media (Cheng 2021; Dong and Wu 2022; Jungblut, Kumpel, and Steer 2024; Steele and Blau 2023). Further, these protests occurred during a global pandemic when there was fewer in-person interactions and social media became a particularly important tool for police agencies to use to reach their communities (Hu, Dong, and Lovrich 2022; Ralph et al. 2022).
These protests did not happen in isolation; namely, the massive Black Lives Matter (BLM) movement spurred political action among Americans and pushed politicians to alter their rhetoric and policy. Extant research suggests that protests may alter political agendas (Vliegenthart et al. 2016; Wasow 2020) and prompt the dissemination of the protest's demands into popular discourse (Dunivin et al. 2022), though the extent of city responsiveness to BLM demands about defunding the police are disputed (Hoang and Benjamin 2024). Nevertheless, we maintain that rhetoric, as we emphasize here in the context of Facebook posts, and policy responsiveness are two distinct outcomes; though policy outcomes are ultimately the most important indicator of political priorities, we argue that rhetoric is an important first step at integrating the concerns of a community into a political agency. Moreover, the opaque (or completely nonexistent) nature of much data on policing (Clark, Glynn, and Owens 2025; Cook and Fortunato 2023; Schiff et al. 2024) means that comprehensive data on rhetoric, like we have here, may be our best option at comparing a wide set of diverse police agencies.
In sum, we expect there to be a temporal jump in protest posts for those agencies that experienced local protest events in 2020, but only after the murder of George Floyd prompted widespread attention and concern to the problems of police brutality. This leads us to our second hypothesis:
Responding to Constituent Pressure: Who Do Departments Prioritize?
American police departments in the summer of 2020 were all exposed to the murder of George Floyd and the corresponding national discussion about policing. However, we expect there to be variation in whether departments respond to Floyd's murder, whether they respond to policing protests in their jurisdiction, and how they respond if they do, based on the characteristics of their cities. The growth of e-government and social media has increased trust and confidence in the government (Tolbert and Mossberger 2006). Therefore, a straightforward explanation of why police agencies differ in their content is because of their constituencies. Social media is then a tool for “interpersonal, participatory, and interactive communications” between the police and the public (Heverin and Zach 2010, 1). We propose and test two possible, perhaps competing, constituencies which police may prioritize in their Facebook strategy: Black Americans and conservatives. Given the topic of racial justice and policing, these are two groups that vary geographically and whose presence in an area may shape police behavior. 4
Responding to the Concerns of Black Americans
We start from the fact that research shows there is a link between police departments and the racial composition of their communities. Black Americans are disproportionately targeted by the police (Baumgartner, Epp, and Shoub 2018; Soss and Weaver 2017), an empirical reality that has translated into substantive differences in public opinion across racial groups regarding the police (Davenport, Franco, and Iyengar 2022). In 2021, 27% of Black adults said they had a great deal or quite a lot of confidence in the police, compared to 56% of white adults. 5 This support (or lack thereof) has also influenced attitudes about BLM and efforts to address racial inequities. In 2021, 47% of whites expressed strong or some support for BLM, compared to 83% of Black Americans. 6 While Black public opinion is far from monolithic on this issue (Bunyasi and Smith 2019), we expect that in aggregate, Black Americans will be more concerned with issues of racial equality and policing and thus demand their political representatives (including non-elected actors like the police) to address these issues. If the police are responsive to the demands and interests of the public, then that would mean an agency in a more racially diverse area would be more likely to post about these concerns.
Not only that, but we expect that the ways in which agencies will respond to the concerns of Black Americans will be distinct. We expect that agencies in areas that are more racially diverse will discuss protests in a more charitable or positive way to respond to the concerns of that community. This posting behavior could be an effort to increase transparency, accountability, and goodwill among Black Americans in their city, we argue. This follows from existing research that suggests that widespread protests can influence societal frames on certain political issues, as well as on voting behavior (Wasow 2020).
This leads to the next set of hypotheses:
We do acknowledge, however, that the relationship between Black residents and protest posting may be negative; extant research suggests that higher racial diversity may foster racial resentment and thus differences in policy outcomes that are not favorable to communities of color (Blalock 1967). Though we do not discount this possibility, especially as the policy being considered here is a racialized one, much like any criminal justice issue (Baumgartner, Caron, and Duxbury 2023; Wilson, Owens, and Davis 2015), for this examination, we test the story of responsiveness first; that if their community is more diverse, a police agency will have incentives to engage with concerns the of that group.
Responding to the Concerns of Conservative Constituents
A competing story of constituency responsiveness is that police respond selectively to certain constituents. In particular, we might expect police to be especially responsive to more Republican or conservative citizens. Descriptive evidence of the police suggests they themselves are whiter, more conservative, and higher-income than their jurisdictions (Ba et al. 2025). As a result, rather than listening to their constituents, we might expect the police to follow their own preferences and craft their posts to be more conservative than their broader jurisdiction may be. We consider each of these possibilities in turn below.
Americans’ ideology and party identification have become increasingly correlated over time, as have their party loyalty (Barber and Pope 2019). We consider partisanship as an additional variable in predicting variation in responses to protests. Policing in general and BLM in particular are politically divisive issues. Just 19% of Republicans strongly or somewhat support BLM, compared to 85% of Democrats, 7 and in 2021, only 24% of Republicans supported nonviolent protests in response to the fatal deaths of Black Americans at the hands of police. 8 Nationwide protests in response to the killing of George Floyd may have changed and liberalized attitudes among liberal Americans, but did not substantively change those of conservative Americans (Reny and Newman 2021). These negative views are likely enhanced by media coverage, as news media across the political spectrum framed BLM in a negative light, emphasizing the protests’ negative disruptions to their communities (Brown and Mourao 2022).
The higher level of conservatism among police and the polarization of BLM and policing issues suggest that partisanship will be an important contributor to posts about protests. Republican areas are less supportive of these protests, and therefore, we should expect less attention paid to these issues among the police in more conservative cities as they respond to the interests of their communities. We also expect that when agencies in conservative areas choose to post about protests, they will do so in a negative manner. This follows existing research that finds, for example, more conservative media like Fox News mostly described the protests as lawless and violent, negative frames of the 2020 policing protests (Brown and Mourao 2022). This leads us to our final set of hypotheses:
Police Facebook Data and Protests
We rely on CrowdTangle as our source of posts from police Facebook pages. CrowdTangle allowed researchers and journalists access to all Facebook posts from public pages and groups on Facebook (CrowdTangle Team 2019). We used the CrowdTangle tool to gather every post from police departments across the country from 2020. CrowdTangle independently managed a list of these local city departments, about 6,000 unique agencies. 9 Through CrowdTangle, we were able to query for all posts made by each of these agencies. Unfortunately, CrowdTangle was shuttered in 2024. It has been replaced by the Meta Content Library, which researchers can apply for access to.
Notably, this data source allows us to look at the posting behavior of all types of local municipal police agencies—small and large, urban and rural—in comparison to protests in their communities. Next, we subset this list to local, city police departments only—excluding sheriffs’ offices, pages that are primarily for recruiting, for airport police or K-9 units, or school (university, college, or K-12) police. We do this to examine the unique nature of local police departments since the other agency types, like sheriffs, are distinct political offices with different political cultures (Farris and Holman 2024). Then, we pre-processed the data by dropping posts that were just images with no associated text or those that were only URLs. Initially, though we had nearly 800,000 Facebook posts, once we pre-processed and only used those posts with text, that number dropped to about 640,000.
We were left with about 4,000 agencies and about 640,000 posts in all. Figure 1 shows the location of all police departments in our CrowdTangle data. The darker shade of the crosses corresponds to more posts in 2020. While there are some departments that vastly out-post others—like the New York Police Department—overall, there is a fair and consistent amount of posting in agencies nationwide.
Map of police departments in the CrowdTangle data, alongside the logged sum of posts each agency made in 2020. Darker crosses indicate more posts.
Our first hypothesis explores how posting about protests changed after George Floyd's murder on May 25, 2020. For the second hypothesis, we link our CrowdTangle data to the Crowd Counting Consortium (CCC) to gather information on all protests in 2020. 10 We combine the CrowdTangle data with the CCC by merging on that agency's city and state, and create a binary indicator for whether that city ever had a policing protest in 2020 (though, see the appendix using the raw number of all policing protests in 2020 as an alternative independent variable). Figure 2 shows the location of about 14,000 policing protests that took place in 2020. 11 They are widespread across the country, with the largest (over 1,000) taking place in New York City and almost 2,000 cities across the country that only experienced one—places like Bullhead City, Arizona, Hinesville, Georgia, and Tiffin, Ohio.
Map of the logged number of all policing protests in 2020. Data from the Crowd Counting Consortium (CCC).
The size of these protests varied widely; of the about 7,200 protests where we have information on the size of the crowd, the mean number of protesters was more than 400, with protests as small as one person (Troy, Alabama, as one example) and as large as more than 100,000 (Washington, DC). Of the over 2,200 unique sets of actors identified in the CCC as the organizers of the protests, only about 9% were identified as explicit BLM groups. The vast majority instead were labeled simply as general protesters who may have been motivated by BLM but were not organized specifically by that group. Of the more than 4,000 agencies in our CrowdTangle data, about 40% of them experienced policing protests at some point in 2020. Finally, 70% of these protests occurred in May, June, or July.
We finally link this universe of police departments and their posting behavior in 2020 to information on local demographics of those agencies, as well as local policing protests. We first aggregate the data to the department-level (though the results are similar at the post level, see the Appendix). We next connect the agency-level posting data to Census Bureau data on the percent of the city's residents that are Black (Hypotheses 3a and 3b) and the county vote share of Donald Trump in the 2016 election from the MIT Election Lab 12 as a proxy for partisanship (see Hirano and Snyder 2019) (Hypotheses 4a and 4b).
Labeling Protest Posts
To identify which of the posts from these police Facebook groups were related to protests, we created a case-insensitive dictionary to label posts: protest, mob, demonstration, demonstrator, riot, loot, unlawful assembly, unlawful assemblies, disperse, marches, marcher, marching, civil unrest, miscreant, hooligan, rally, rallies, rallied, mob, revolution, rebellion, and clash. 13 A post was labeled as about a protest if it included at least one of these protest words. We created these dictionaries from two rounds of content analysis of random samples of posts by the authors, alongside other scholarly articles with listed dictionaries around protests, derived from reading posts about protests and machine learning (e.g., Hoffmann et al. 2022; Oglesby-Neal, Tiry, and Kim 2019; Spaiser et al. 2017). Doing so allowed us to leverage our and other scholars’ expert knowledge (Muddiman, McGregor, and Stroud 2019).
We validated these dictionaries for both false positives (i.e., to check if our dictionaries labeled posts as about protests which are not) and false negatives (i.e., to check if our dictionaries systematically missed protest posts). To validate, each co-author and a research assistant hand-labeled random samples of 400 posts (1,200 posts in total). For each validation set, 250 were posts that the dictionaries labeled as about protests, and 150 were posts that the dictionaries labeled as not about protests. The accuracy was 81% across the three labelers. Greater information about our validation and dictionary accuracy metrics are included in the appendix. In total, the dictionary labeled 6,113 posts by police departments as about protests.
In our data, the dictionary identified that about 1% of posts were about protests. 14 Moreover, only 57% of departments that experienced policing protests posted about it. And, most of the posts about protests occurred on the day a policing protest occurred—see Appendix Figure A3 for the distribution of protest posts after the city's first post-George Floyd policing protest. We see a significant spike in protest posts after May 25, 2020. Indeed, the protest posts mirror the actual incidence of policing protests, as shown in Figure 3.
All local policing protests and all posts labeled as protest-related by date. The dashed line represents the date that George Floyd was murdered.
Once we labeled the protest posts, our next task was to identify the sentiment of those posts, whether they were positive or negative, to explore Hypotheses 3b and 4b. To do this, we use a variation of the NRC Word-Emotion Association Lexicon from Mohammad and Turney (2013). This method takes a count of words related to positive or negative sentiment, respectively, that are present in each Facebook post. If a word is used multiple times in a post, it is counted each time. Most of the posts had a mix of negative and positive words; for our purposes, we labeled the posts by their dominant emotion. For instance, if 20% of a post's words were negative and 15% were positive, we labeled that post as negative, given that it was the dominant emotion. 15 Of the more than 6,000 posts about protests, about 87% of those had a dominant positive emotion.
Difference-in-Differences Results
To estimate the relationship between local protests, the post-George Floyd environment, and a variety of internal political factors, we run a difference-in-differences regression using ordinary least squares (OLS) at the agency level. Our independent variables, described above, include a binary indicator for after May 25, 2020, George Floyd's murder, a binary indicator for whether that city experienced any policing protest in 2020, an interaction term between the dummy variables for both Floyd's murder and the presence of a local policing protest, the percent of the city's population that is Black, and the percentage of the county that voted for Trump in 2016. Our dependent variables are the number of posts identified as about protest, the number that are about protests with dominant negative sentiment, and the number that are about protests with dominant positive sentiment. 16 We additionally control for the logged city population and the logged page likes of the department. 17 We cluster our standard errors by city, as some places (like Chicago and Illinois) have multiple agencies in the CrowdTangle data.
Table 1 shows the results from this analysis. We see results consistent with Hypothesis 1. The period after George Floyd's death is associated with more protest posts. We also see evidence for Hypothesis 2, as the interaction term between the post-George Floyd variable and any policing protest is significant and positive. Interestingly, however, the coefficient on the policing protest variable alone is negative, suggesting that prior to May 25, 2020, places with local policing protests were less likely to post about protests. These results lend credence to our argument that both national and local pressures are important in conditioning responsiveness from police—that police agencies respond, at least in messaging, to local policing protests pushing for reform, but only after a significant national event pushes them to do so.
Ordinary Least Squares (OLS) Estimation of the Likelihood of Posting About Protests Aggregated by Agency.
Note: * p < .1; ** p < .05; *** p < .01.
SEs clustered by city.
We find little evidence for Hypothesis 3a or 3b. Cities with a higher proportion of Black residents are not associated with increases or decreases in posting about protest—and are also no more or less likely to post positively about the protests. We do see some evidence for Hypothesis 4a, that more Republican counties are associated with fewer protest posts and fewer protest posts with either negative or positive sentiment—though note that the magnitude is larger for positive protest posts, not consistent with Hypothesis 4b.
We do not see much differences in the sentiment of these posts, contra to some of our expectations. However, we still expect the ways in which these departments are posting about protest to be different, perhaps in ways that identifying positive or negative sentiment alone cannot fully capture. The next sections explore possible differences in more detail.
Qualitative Coding of Posts
Police departments discussed policing protests in many ways. The previous section considers posting behavior, but did not distinguish between the context of posts beyond sentiment. While we may expect that all police departments will have a more conservative ideology that is observable in their posts, given that a department has posted about a protest in their area, there are several different ways that they could use to discuss or frame the issue (or even avoid). For example, consider the posts shown in Figure 4. The post from the New York Police Department references an individual who had an illegal firearm and burglarized a business during the “unrest” and “looting.” The post from the Miami-Dade Police Department informs their followers that the city's protests on June 1, 2020, did not lead to any arrests or damages.
Two separate examples of Facebook posts labeled as about policing protests, made by departments who had a protest in their city.
We took a random sample of 190 posts that our dictionary labeled as about protests, and a co-author validated as about protests. We began with an open coding approach, looking for recurring categories or concepts in the posts (LaRossa 2005). After going through one time and coming up with our list of possible labels, we went through the sample a second time to ensure that all labels were applied to all relevant posts. These qualitative labels are not mutually exclusive—many posts have more than one. Most posts seem to be meant for the agency's community, whether providing specific information or directly speaking to the community or town by name.
The most common topic in the random sample of protest posts was providing logistical information to the community about an upcoming or ongoing protest. This was present in 37% of posts. Examples described protest start times and road closures along the route, such as the first post listed in Table 2. The second most common topic (36%) explicitly stated that the policing protest was peaceful. Many of these posts thanked the community and participants for a peaceful event. The next two most common topics (22% each) mention criminal or disobedient behaviors of participants that do not rise to the level of violence or contrast peaceful and violent protesters. The first, disobedient behavior, includes actions like blocking traffic. The contrast between peace and violence compares the behavior of most protesters, who are peaceful, with bad actors who engage in violence either as a part of or adjacent to the protests. An example is the second post in Table 2. Other notable topics include mentioning George Floyd, BLM, or racial injustice (21%), stating that they protect and respect the First Amendment right to protest (see the third post in Table 2), engaging with their community (17%), and describing violent actions of protesters (13%). The latter can be seen in the fourth post in Table 2. The agencies acknowledge that they are watching social media and speaking to community members who are watching social media. A small number of posts (6% of the random sample we coded) explicitly caution their community about misinformation online, such as the fifth post in Table 2.
Examples of Police Department Facebook Posts With Qualitative Labels, Page, and Date Posted.
Keyword-Assisted Topic Models
Not all posts about police protests are the same, as we showed in Table 2. Some are more supportive, others are more critical. Here, we examine all protest posts in more detail. As an exploration of the variation within these types of posts, we use keyword-assisted topic models to label topics of interest (Eshima, Imai, and Sasaki 2024). This is a method that allows researchers to incorporate their own context-specific knowledge into topic models to increase the accuracy of label predictions and help topics to be more interpretable and relevant to the quantities of interest. Researchers can specify the number of topics for the model and choose to provide relevant keywords for any of those topics. Doing so can help the model pick up on topics of interest to researchers, as long as the keywords effectively distinguish between the different topics.
The selection of keywords here is also informed by our content analysis described above, in which several trends emerged. Some posts discussed the protests but also mentioned that the protest was “peaceful” or “lawful” and that the police department supports the “constitutional rights” of people to demonstrate. Another group of posts mentions the protests as well as any crimes and vandalism that occurred—these posts also described “unlawful” assemblies and “civil unrest.”
The trends outlined above shape the two broad topics that we provided the model keywords for, as shown in Table 3. We call them “Peaceful” and “Violent.” We estimated the model with different number of topics without keywords, and the model performed best with one additional topic, which we call “Other.” Top words for the “Other” estimated topic include “police,” “officers,” “department,” “units,” and “town.” The predicted labels from the keyword automated topic model are mutually exclusive. We apply the label for the most likely topic as predicted by the model. Table 3 shows the frequency with which posts were labeled as about each topic. About 80% of the posts were labeled as “Peaceful,” about 15% as “Violent,” and the remaining 5% as “Other.”
Keywords Used to Label Sub-Topics Within the Posts About Police Protests, Using Keyword-Assisted Topic Models. A Post Is Assigned to Its Most Likely Topic as Predicted by the Classifier.
What is the content of these posts? Table 3 provides some indication of the keywords used, but do not give us a full sense of the precise content of these messages. We look into positive and negative sentiment of these posts by topic, following other research that finds emotional responses to police shootings on social media (Oglesby-Neal, Tiry, and Kim 2019), especially by the race of protesters (Gause, Moore, and Ostfeld 2023). In Table 4, we categorize posts into one of three mutually exclusive categories using the counts of emotive language: the majority sentiment is positive (i.e., more positive words than negative words), negative (i.e., more negative than positive words), or has an equal number of positive and negative words. Here, we use the same sentiment codings as used in our regression analysis above.
The Percentage of Each Type of Post That Was Labeled as Containing a Majority Positive or Negative Sentiment, or an Equal Number of Words of Both Sentiments.
Note: Percentages may not total 100 due to rounding.
Perhaps unsurprisingly, the vast majority of posts in the “Peaceful” topic are either positive or equally positive and negative. The “Violent” topic, on the other hand, is split between positive, negative, or equal emotive sentiments (though most posts are of negative sentiment). This initial examination suggests that there are distinct emotional tenors between posts about police protests and that police are likely eliciting different responses to these posts. Next, we turn to the public response to these messages.
Table 5 shows the engagement that the protest posts received: likes, shares, and comments. We see differences across topics. Posts using the “Peaceful” frame received more likes than posts with the “Violent” frame, but received fewer shares and comments. The “Other” category received much lower levels of engagement across metrics. The average engagement for the “Peaceful” and “Violent” frames is higher than for the whole sample of police department posts from 2020, where the average number of likes is 79, shares is 46, and comments is 18.
The Average Number of Engagement on Facebook (Likes, Comments, and Shares) That Posts of Each Type Received.
Social media engagement is not valueless. As we now know from the 2021 “Facebook Files,” the company had an internal ranking system of different engagement types designed to increase interactions. Of likes, shares, and comments, likes were worth 1 point, shares 5, and comments 15 or 30, depending on their “significance” (Hagey and Horwitz 2021). When a post receives more points, the further it spreads, receiving more engagement (and more points) in turn. If engagement, particularly shares and comments, are favorable to an organization on Facebook, then posting about the protests, and particularly highlighting the violence sometimes associated with them, rather than peaceful protesters, may be useful for police departments to get a wider audience.
Discussion and Conclusion
This study provides insights into how police use their public Facebook pages and, in particular, how they use those pages around protest events. We expected agencies to post more about protest following George Floyd's murder, generally, and especially in those places that ever had a policing protest in 2020. We also expected variation in responsiveness to policing protests for two additional competing reasons. First, we expected areas with more Black residents to post more about protests and more in a positive light, following the assumption that the police would want to respond to that constituent group's pressures and demands. Second, a competing assumption argues that because police officers tend to be more conservative, that partisanship in their jurisdictions would be a strong predictor of the protest posts. We find mixed support for our hypotheses and assumptions. We see strong evidence that George Floyd's death and after his death in those places that experienced policing protests were associated with more protest posts. Partisanship appears to drive posting about protest, but the racial demographics of the city does not, and only partisanship substantively relates to differences in the positive or negative sentiment in those posts. Our keyword-assisted topic models find that the most common frame used by police in their protest posts was “Peaceful”—highlighting information and the right to protest. However, it is posts using the “Violent” frame that receive higher levels of engagement, something that is often desirable for those online. Taken together, our results suggest that local protest dynamics may prompt a change in public-facing rhetoric from the police, but only once broader national forces add an additional layer of pressure to do so. And that partisanship appears to drive posting behavior of local agencies rather than racial dynamics, is consistent with extant research that finds partisanship to be the strongest influence on policing attitudes (Reny and Newman 2021; Schiff et al. 2024).
We may see engagement with these events on individual officers’ Facebook pages and not necessarily on public ones, however. Departments do not consistently have written social media policies: they often employ different positions to manage their social media (from public information officers to civilian employees to police chiefs, etc.), and some require approval of posts by a central body, and some do not (Kim et al. 2017). Nevertheless, we might expect more control over the official language and posts on an official agency Facebook page than an individual officer's, for example. Okegbe (2021) details several instances of inflammatory posts by officers about George Floyd or BLM that prompted their agency to place them on leave and issue official statements distancing themselves from those officers. We may also expect more robust responses on police department union pages, for example, that are not governed by official department guidelines.
These results are important as they point to the essential nature and character of bureaucratic responsiveness to the public. Though bureaucrats are not elected, the fact that we see variation in responses suggests that police agencies are making strategic choices about what to post. In particular, racial demographics do not seem to influence posting behavior, but party does, similar to research in the Congressional context finding variation by party (Russell 2018). We hope that future research will continue to excavate the content of these posts in even more detail—and extend the time frame of the data. Prior to the 2020 mass protests, did the police exhibit more (or less) responsiveness to issues of racial justice in policing? How has this shaped police department behavior and rhetoric moving forwards? How does this vary across communication channels?
We argue that analyzing how police use social media as a communication tool, particularly in times of crisis, represents an opportunity for those police agencies to communicate their preferences and values to the public. Those messages, we contend, are representative of a broader orientation of these agencies to their communities. Moreover, by directly posting to their constituents, police agencies are able to bypass traditional media (Colbran 2020, Grygiel and Lysak 2021). In effect, posting on social media sites like Facebook can help police “control their message and provide clear, concise information to correct misinformation and combat the problem of negative public image” (Jeanis, Muniz, and Molbert 2021, 572). Yet, our results suggest that not all agencies are taking advantage of this opportunity and will only do so once both broad national pressure and local dynamics prompt them to. And, these dynamics influence the sentiment of the posts to the public. We argue that this has important implications for the role of both public and local pressure to prompt responsiveness from the police—that for a shift in how police communicate about issues, there must be both local and national pressure to do so.
There are several additional avenues for future research that we have not yet explored here. First, we can further investigate the public reaction to posts about protests. How does the public respond to these posts (Hu, Rodgers, and Lovrich 2020)? How does liking or sharing behavior change as cities experience a protest? Or, does the public response depend on the partisanship of the area? We also did not investigate the possibility of posting about counter-protests or movements that push back against BLM, for example. Perhaps agencies are not engaging with our protest terms but are instead posting about #AllLivesMatter or #BlueLivesMatter (Carney 2016; Gallagher et al. 2018). Another (less common) trend that we noticed when hand-coding police protests posts was that police departments explicitly mentioned and sought to debunk misinformation in their communities around protests. This is not something we anticipated, but it could be an additional avenue of study.
Another open question is who is posting or directing the posting behavior of these agencies. Though we cannot test that directly here, future research ought to examine variation in social media policy—a survey of police departments, for example, would highlight who their assumed audience is via different communication methods, if anyone in the department can post content, and what they hope to convey with that information. Social media could be a broader tool for police departments to engage in productive conversations with the public. Is that the case? Or, do other considerations or the audience shape police social media strategy? It is also possible that other police agencies with distinct politics, like sheriffs, may use their social media differently—future research into these other contexts would help illuminate whether and to what degree the patterns we find here are common to all policing agencies or to only those local, municipal agencies.
Third, local police agencies are just one key component of local politics. The variation in the institutions in municipalities means that city councils and mayors are additional factors to consider in the behavior of police (Hoang and Benjamin 2024). Mayors, for example, may be more or less effective at applying pressure on the police agencies under their jurisdiction. Are these results conditional on the partisanship of these other political actors? Or, perhaps the race of these actors? Other research could take up this potential interactive effect to include both public-facing communication from the police and substantive policy impacts in local politics, from budgets to law changes, and others.
Finally, we look at one particular instance of responsiveness as the unprecedented nature of the policing protests in 2020 offers a unique case to test whether and how police respond to both national and local concerns about racism and policing. However, there are other events like officer-involved shootings, establishment of police reform like civilian review boards, mass casualty events, or others (Olzak 2021) that likely influence posting strategies. Since protests are more likely in places that experience police shootings (Williamson, Trump, and Einstein 2018), we might expect the differences we theorized here to be especially present in those areas. A full analysis of that type is beyond the scope of this paper, but it would be interesting to see whether and how communication changes with a variety of policing events or controversies.
The results here are consistent with other findings that social media primarily serves to reinforce existing public platforms and statements from the police (Crump 2011). The promise of social media as a tool to engage more with the community and their concerns about policing may be overly optimistic. More simply, technology may be a necessary, but not sufficient, condition to facilitate increasing community engagement (Bullock 2018; Crump 2011). Rather, pressures from both local and national events may be needed to prompt changes in the rhetoric and communication strategies of the police.
Supplemental Material
sj-pdf-1-uar-10.1177_10780874251414036 - Supplemental material for Bureaucratic Responsiveness in Times of Crisis: The 2020 Mass Protests and Police Department Social Media *
Supplemental material, sj-pdf-1-uar-10.1177_10780874251414036 for Bureaucratic Responsiveness in Times of Crisis: The 2020 Mass Protests and Police Department Social Media * by Anna Gunderson and Maggie Macdonald in Urban Affairs Review
Footnotes
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
The data that support the findings of this study were available from the CrowdTangle tool, which has since been discontinued. Restrictions apply to the availability of these data, which were used under license for this study. Data were available from CrowdTangle with the permission of Facebook, but could now be collected using the Meta Content Library. The authors will make the cleaned data and code used for this article available on Dataverse.
Supplemental Material
Supplemental material for this article is available online.