Understanding the role of media in the formation of public sentiment towards the police

The study investigates whether and how media coverage shapes public sentiment toward the police over time in the United States. Motivated by heightened public discourse and polarization following high-profile incidents of police brutality (e.g., George Floyd), the authors posit that sentiment is influenced not only by local crime conditions but also by national media narratives. They operationalize public sentiment via sentiment analysis of geolocated Twitter posts and frame two hypotheses: H1) media coverage of local crime influences positive (H1a) and negative (H1b) sentiment toward the police; H2) national media coverage of police brutality increases negative sentiment toward the police nationwide. The work aims to overcome limitations of snapshot surveys by assembling a decade-long, daily-resolved, multi-source dataset and employing causal discovery techniques beyond simple correlation.

Prior research indicates that high-profile police misconduct undermines public perceptions of implicated departments and increases legal cynicism, particularly in disadvantaged and racially isolated communities. Media framing effects can bias risk perceptions and attitudes; controlled studies show exposure to information about police brutality (statistics, articles, videos) increases negative perceptions of police. While local crime can affect satisfaction with police in survey-based studies, media coverage of crime may be sensationalized and only weakly correlated with actual crime rates, potentially skewing perceptions. Police organizations recognize the critical role of media relations in shaping legitimacy and public trust, engaging in information sharing to influence portrayal. Social media, especially Twitter, provides high-frequency, large-scale data for tracking public sentiment during major events, and has been successfully used to measure reactions to crises, health trends, and financial markets. Prior work has examined sentiment shifts after specific police incidents and compared public and police-projected images online, but comprehensive, time-resolved causal analyses linking media coverage to sentiment across multiple metropolitan areas have been limited.

Ethics: The study was administratively reviewed by NYU’s IRB and determined not to involve human participants per 45CFR46.102. Data collected (May 2022) contain no personally identifying information. Data and variables (10/01/2010–12/31/2020, daily): Four time series were constructed: (1) media coverage of police brutality (MPB; national), (2) media coverage of local crime (MLC; metro-specific), (3) positive tweets about police (PT; metro-specific), (4) negative tweets about police (NT; metro-specific). Two metropolitan areas (Seattle, San Diego) were later excluded due to data limitations. Media coverage of police brutality (MPB): Using ProQuest, the authors counted daily printed articles in the 20 most-circulated US newspapers that contained the exact phrase “police brutality” in text or headline. Because print coverage typically lags events by one day, the time series was shifted backward by one day. Media coverage of local crime (MLC): For the top 20 most-populated US metropolitan areas, the authors searched ProQuest for printed articles in up to the 20 largest newspapers by circulation in the relevant states, matching the term “crime” together with any city (population >100,000) within each metro (detailed queries in Supplementary Note 1). This yields 18 usable metro time series after exclusions. Twitter data and sentiment (PT, NT): Using the snscrape Python library (respecting Twitter’s ToS at the time), the authors scraped geolocated tweets within circular approximations of each metro’s boundary, matching keywords “police,” “cop,” and the main city’s police department abbreviation (e.g., NYPD). Aspect-based sentiment analysis was applied to classify tweets as positive or negative toward police using a DeBERTa model (PyTorch). Resulting daily counts form PT and NT per metro. High-resolution MPB on Twitter: For additional analyses during high attention, the authors scraped minute-level tweets containing “police brutality” from the 10 most-followed US newspaper accounts for two weeks after George Floyd’s murder (00:00 May 29–23:59 June 13, 2020), and aligned these with minute-level NT. Preprocessing: To meet stationarity assumptions for information-theoretic analysis, daily time series were seasonally adjusted and detrended via STL decomposition with Loess (weekly, monthly, yearly components) using statsmodels; trend and seasonal components were subtracted. Transfer entropy (TE): Symbolic transfer entropy with conditional terms was used to infer directional associations in a Wiener-Granger sense. Time series were symbolized into three quantile bins (low/medium/high). Conditional TE was computed to mitigate common-driver and cascade effects, with surrogate significance testing via 20,000 permutations (shuffling the source within joint realizations of target and conditioning variables); significance at α=0.05 if TE exceeded the 95th percentile of surrogate distribution. Tested links: T_MLC→NT|MPB (H1b), T_MLC→PT|MPB (H1a), T_MPB→NT|MLC (H2). For significant links, the sign of association was assessed via Spearman partial correlation between past source and present target, controlling for past target and past conditioning variable (pingouin; Kim method). Convergent Cross Mapping (CCM): As an independent causal inference (not reliant on Wiener-Granger causality), CCM reconstructed one variable from the time-delay embedded manifold of another to test coupling and direction. Parameters included embedding dimension E (selected via simplex projection), delay r, and nearest neighbors n=E+1, with performance measured by Pearson correlation between reconstructed and observed series. Implemented with publicly available Python code. Latent variable detection (LPCMCI): To probe for hidden confounders between MPB and NT, the Latent PCMCI algorithm (Tigramite) performed conditional independence testing on time series with a maximum lag of one (per Supplementary Note 3), classifying edges and indicating potential latent confounding. Significance α=0.05 with Pearson partial correlation tests. High-resolution before/after analysis (George Floyd period): For each newspaper tweet mentioning “police brutality,” the authors computed average public NT per minute in symmetric windows before and after the media post, across multiple semi-widths (30 to 720 minutes, in 30-minute increments). A one-tailed Wilcoxon signed-rank test (n=205 media tweets) assessed whether post-tweet NT rates exceeded pre-tweet rates across metros. Exclusions: Seattle excluded due to concentrated MLC in a short period; San Diego excluded due to insufficient MLC observations (<0.8/day).

- Dataset: 10 years of daily data (m=3,745 days) across 18 US metropolitan areas, encompassing over 2.5 million tweets and ~200,000 newspaper articles. MPB exhibited a national peak of 113 printed articles on June 13, 2020 (19 days after George Floyd’s death). - Transfer entropy (TE): Strong and consistent evidence for H2: conditional TE from MPB to NT was significant (p<0.0001) for all 18 metros, with TE magnitudes typically ~0.0189–0.0391 bits (e.g., Phoenix 0.0388, Tampa 0.0391). H1a (MLC→PT|MPB) showed significance only in Washington, D.C. (p=0.0624 not significant) and Baltimore (p=0.0273 significant for PT) with limited support overall; H1b (MLC→NT|MPB) was significant only in Dallas (p=0.0084) and Washington, D.C. (p=0.0148), with null results elsewhere, indicating weak or absent effects of local crime media coverage on sentiment. - Direction/sign of effects: For all metros, partial correlations for MPB→NT were positive and significant (Spearman partial r ≈ 0.13–0.28; all p<0.0001). Example: Tampa r=0.2843 [0.25,0.31]; Boston r=0.2139 [0.18,0.24]. This indicates increases in national media coverage of police brutality are associated with increases in negative tweets about police. - CCM results: Convergent cross mapping corroborated coupling between MPB and NT across metros, with strong convergence patterns. Reconstruction of MLC from sentiment series was weak in most metros (exception: some signal in Houston), aligning with TE results that local crime media coverage is not a robust driver of sentiment. - Latent confounders (LPCMCI): With maximum lag=1 focusing on contemporaneous links, 9 of 18 metros showed a causal link MPB→NT without evidence of latent confounding; 7 were inconclusive; Los Angeles suggested unlikely link; Minneapolis indicated a latent confounder potentially affecting both MPB and NT. Specific latent variables were not identified by the algorithm. - High-resolution George Floyd analysis: Across symmetric windows from 60 to 1440 minutes, one-tailed Wilcoxon tests (n=205 media tweets) showed that in most metros (≥12 of 18 for windows 240–1080 minutes), public NT per minute increased after newspaper tweets mentioning “police brutality.” This suggests a public response time scale of several hours to nearly half a day. - Illustrative peaks (New York City): MLC daily peak 44 articles (May 14, 2016); PT daily peak 104 (Apr 5, 2013); NT daily peak 1896 (May 31, 2020).

The findings address the central question by demonstrating that national media coverage of police brutality is a robust, directional driver of increases in negative public sentiment toward the police on Twitter. This effect appears nationwide, regardless of where incidents occur, consistent with prior observations that high-profile misconduct reverberates beyond local contexts. Conversely, media coverage of local crime shows little to no causal influence on either positive or negative sentiment toward police in most metropolitan areas, suggesting that perceptions of police legitimacy and sentiment are more sensitive to reports of misconduct than to crime reporting per se. The predominance of negativity aligns with negativity bias and prior evidence that negative experiences and negative news content propagate more strongly on social media. Triangulation via transfer entropy, CCM, latent variable analysis, and high-resolution before/after testing strengthens the causal interpretation while acknowledging observational constraints. The consistent pattern across diverse metros indicates common societal mechanisms in appraising law enforcement through media, and underscores the role of media narratives in shaping public discourse.

This work advances the field by constructing a decade-long, multi-source time series across 18 US metropolitan areas and applying model-free, causal discovery methods to uncover directional interdependencies. It provides converging evidence that media coverage of police brutality substantially drives negative public sentiment toward police, while media coverage of local crime and local crime levels themselves are not salient drivers. The results encourage consideration of more balanced media reporting alongside efforts toward police reform and community trust-building. Future research should: (1) incorporate police behavior and policy response data to probe potential indirect pathways; (2) extend analyses to smaller and more diverse communities; (3) refine sentiment measures for sarcasm, emojis, and stance; and (4) assess heterogeneity by demographic, political affiliation, and media consumption patterns to understand subgroup dynamics in sentiment formation.

- Representativeness and sentiment measurement: Twitter users are not fully representative of the public, and many users may not express views on the platform. Geolocation constraints may miss relevant tweets. Despite state-of-the-art models, sentiment analysis can misclassify slang, emojis, and sarcasm. - Generalizability: The sample focuses on large metropolitan areas with potentially different policing exposure and sociodemographic contexts; results may not generalize to smaller or rural communities. Individual experiences and neighborhood characteristics may shape attitudes independently of media coverage. - Local brutality events: The study cannot isolate effects of rare, local police brutality incidents due to their infrequency and low information content for causal inference, risking false inferences with TE on sparse signals. - Causal inference assumptions: Observational analyses rely on assumptions (e.g., stationarity, faithfulness, no selection bias, acyclicity). Methods like LPCMCI do not rule out indirect links; police behavioral responses to media could mediate effects over longer horizons. High-resolution evidence suggests direct effects but cannot exclude alternative mechanisms generally. - Political composition and bias: Twitter’s partisan composition may vary by time and metro, potentially biasing sentiment measures. Differences in how political groups express negativity/positivity on Twitter may further skew sentiment estimates.