Public opinion regarding the police in the United States has become increasingly polarized, fueled by widely publicized incidents of police misconduct and the rise of social movements like Black Lives Matter and Blue Lives Matter. These contrasting viewpoints highlight a significant societal divide, particularly evident in reactions to trials of police officers accused of excessive force. This division is especially pronounced in large metropolitan areas where police actions are subject to more scrutiny. Previous research shows that high-profile incidents of police misconduct negatively impact public perception of involved departments. Legal cynicism, or a distrust in the legitimacy of laws and social norms, is exacerbated by such events, leading to heightened distrust of the police, especially in marginalized communities. The public's perception is shaped not only by objective facts but also by the way these events are presented in the media, a phenomenon known as the "framing effect." The choice of words used in reporting, for instance, can significantly influence public perceptions of risk and create cognitive biases. Studies have shown that exposure to information about police brutality, regardless of format (statistics, articles, videos), fosters negative perceptions of the police. While media coverage of police misconduct is a significant factor in shaping public opinion, media coverage of local crime presents another potential driver. Frequent reports of crime can negatively affect the public's view of the local police, as police departments are responsible for maintaining safety and order. However, the relationship between media coverage of local crime and public perception of the police is less studied and often complex, affected by media's tendency to sensationalize events to increase viewership. News reports may exaggerate crime stories, highlighting blameworthiness of offenders, or emphasizing police ineffectiveness. Police agencies themselves recognize the importance of positive police-community relations for effective partnerships, information exchange, and resource mobilization. They often strategically engage with media outlets to manage their public image. This study aims to use social media data to investigate the dynamic relationship between media coverage (of both police brutality and local crime) and public sentiment towards the police, overcoming limitations of traditional survey-based research which often suffer from lack of continuity and sampling challenges.

Existing research highlights the significant impact of media coverage, particularly regarding police misconduct, on public sentiment. Studies using surveys have demonstrated the negative correlation between high-profile incidents of police brutality and public trust in law enforcement. The framing of such events in media reports, emphasizing certain aspects while downplaying others, influences public perception and contributes to the formation of biases. While the effect of local crime on public opinion is acknowledged, the impact of media portrayal of local crime on this opinion is less clear. There is some evidence to suggest that frequent crime reports, even when not directly correlated with actual crime rates, contribute negatively to public confidence in the police. The role of media in shaping public opinion is not unidirectional. Police agencies are active participants, recognizing the importance of media relations for managing their public image and building trust with communities. Strategic collaborations with media can promote transparency, but also risk allowing narratives to be shaped by selective information release. The study uses social media data to overcome limitations of traditional survey methods, allowing for longitudinal analysis and exploration of the intricate relationships between media representations and public sentiment.

This study utilized a novel dataset spanning ten years (2010-2020) of daily-resolved data for four variables: media coverage of police brutality, media coverage of local crime, positive Twitter posts about the police, and negative Twitter posts about the police. Data on media coverage was collected from ProQuest, focusing on print media articles due to their reliability compared to other sources. Media coverage of police brutality was considered a national variable, as such incidents tend to receive national attention, unlike local crime. For media coverage of local crime, the study focused on local newspapers within 18 of the most populous metropolitan areas in the US. Daily counts of articles were recorded, considering articles mentioning "crime" and the name of the main city in the metropolitan area. Twitter data was collected using the Python library snscrape, focusing on geo-located tweets containing keywords like "police," "cop," and police department abbreviations. The limited geographic scope of Twitter data, along with the use of aspect-based sentiment analysis using the DeBERTa model, enabled the distinction between positive and negative sentiments. To account for events around George Floyd's death, high-resolution, minute-by-minute data on Twitter posts regarding police brutality from major newspapers' accounts was also collected. This allowed for a more nuanced understanding of the real-time response of the public on Twitter. All time series were detrended and seasonally adjusted to ensure stationarity before further analysis. The core analytical method was transfer entropy, an information-theoretic approach for detecting causal relationships in time series data. Transfer entropy measures the reduction in uncertainty in the prediction of a target variable's future from its past, given the past of a source variable. This methodology addresses the limitations of linear correlation analyses in identifying causal associations. Conditional transfer entropy was used to control for potential confounding variables (e.g., media coverage of police brutality influencing both media coverage of local crime and public sentiment). To validate the findings, three additional analyses were conducted: 1) Statistical analysis of high-resolution data around George Floyd's murder, tracking chains of posting events between newspapers and users on Twitter, 2) Convergent cross mapping (CCM), a dynamical systems method for detecting causality, and 3) The Latent Peter-Clark Momentary Conditional Independence (LPCMCI) framework for identifying latent confounders affecting the causal relationship. Partial correlation was employed to determine the sign of the associations.

The transfer entropy analysis provided strong support for hypothesis H2 (media coverage of police brutality influences negative public sentiment towards the police). This effect was consistent across all 18 metropolitan areas studied. The partial correlation analysis confirmed a positive association between increased media coverage of police brutality and a rise in negative tweets. Conversely, the study found little to no support for hypothesis H1 (media coverage of local crime influences public sentiment). Transfer entropy analysis, along with a supplementary analysis using objective crime data, indicated that neither media coverage of local crime nor the objective local crime statistics themselves substantially influenced public sentiment. The convergent cross-mapping analysis further supported the finding that media coverage of police brutality is a key driver of negative public sentiment, while the influence of local crime was weak. This consistency across multiple analytical methods strengthens the robustness of the findings. The latent variable analysis (LPCMCI) indicated that for most metropolitan areas, there were no confounding variables masking the direct causal relationship between media coverage of police brutality and negative tweets about the police. This strengthens the conclusion that the observed relationship is not spurious but rather indicative of a direct causal effect. The high-resolution analysis focusing on the period following George Floyd's murder corroborated the main finding. A significant increase in negative tweets about the police was consistently observed across most metropolitan areas following tweets from major newspapers covering police brutality. This immediate reaction reinforced the causal relationship and highlighted a relatively short response time (on the order of hours) in the immediate aftermath of highly publicized events. The consistency of this pattern suggests similarities in how the public processes information related to law enforcement, regardless of city-specific contexts.

The findings suggest a clear and consistent influence of national media coverage of police brutality on negative public sentiment towards the police, while the effect of local crime reporting or actual local crime statistics was far less pronounced. This asymmetry underscores the disproportionate impact of negative news related to police misconduct. The public's response appears driven primarily by national media narratives of police brutality, irrespective of local contexts or the actual crime rates in a given area. This pattern aligns with negativity bias theory which posits that negative information holds more weight in shaping judgments and perceptions. The robust consistency of these findings across multiple analytical techniques and datasets, and the immediate response observed in the high-resolution analysis, indicate a strong causal effect. While acknowledging limitations in data sources and the complexities of public opinion, the study provides compelling evidence for a direct causal relationship between media coverage of police brutality and negative public sentiment. This suggests a need for a more balanced approach to reporting, highlighting not just instances of misconduct but also the positive contributions of law enforcement.

This study demonstrates a strong causal link between media coverage of police brutality and negative public sentiment towards the police. This relationship was consistent across various metropolitan areas and analytical methods. In contrast, media coverage of local crime and objective local crime data did not show a strong effect. Future research should focus on understanding the specific mechanisms that mediate this influence, examining the impact of different types of media coverage, and exploring strategies for fostering more balanced and nuanced reporting of police-related issues.

The study acknowledges several limitations. Firstly, sentiment analysis of Twitter data may not perfectly reflect the sentiment of the entire population. Secondly, the conclusions may not generalize to smaller urban communities or those with different police-community dynamics. Thirdly, the study could not fully isolate the role of local police brutality due to data limitations. Fourthly, causal inference, even with robust methods, remains subject to assumptions and potential indirect effects. Finally, the study notes potential biases in the sampling of Twitter users across the political spectrum.