How citizens engage with the social media presence of climate authorities: the case of five Brazilian cities

The study investigates how local climate authorities in five major Brazilian cities use social media to communicate about weather and climate, and how citizens engage in response. Guided by Social Media Presence and Human Engagement frameworks, it asks: (1) What publication patterns emerge when local-government authorities communicate through social media? (2) How do citizens respond to authorities' publications? Motivated by the importance of preparedness and adaptation at the city level, and the growing use of social media for information dissemination, the study emphasizes practice-based evidence from routine, long-term interactions rather than short-lived events or hashtag-focused analyses. The work seeks to inform communication strategies that can improve preparedness, raise awareness of climate risks, and support adaptation actions.

Background covers preparedness, adaptation, and resilience at the local level, where authorities disseminate alerts and guidance and citizens actively seek information online. Social media presence is framed as designing content and posting strategies (single- vs double-loop; active vs passive posts) to reach audiences, build credibility, and potentially foster two-way exchanges. Human Engagement theory conceptualizes engagement cycles (engagement, sustained engagement, disengagement, re-engagement) and distinguishes transient versus regular participation, typically showing long-tail distributions in online systems. Prior climate-related social media studies often focus on single events or term/hashtag-based samples, finding roles for authorities in risk communication and public discussion of climate change. However, gaps remain in understanding routine, long-term citizen-authority communication, especially in under-studied contexts like Brazilian cities. This study addresses methodological (beyond events/hashtags), population (Brazilian cities), and practical-knowledge gaps (deriving lessons from practice).

Design: Observational study of all public posts by local climate authorities and citizen replies on the X platform across five Brazilian capitals (São Paulo, Rio de Janeiro, Belo Horizonte, Porto Alegre, Belém) over one year (17 Jul 2021–16 Jul 2022). Data include 10,229 authority tweets and 5,471 citizen replies from 2,600+ citizens; meteorological warnings from the national meteorological institute (INMET) contextualize climate events. City selection: From all Brazilian state capitals, retained cities with an active, dedicated weather/climate account and regular posting prior to July 2021, yielding five cities with diverse climates (Af, Am/Aw, Cfa, Cfb, Cwb) and large populations. Social media presence analysis: Focus on posting periodicity and content. Characterized communication as single- vs double-loop (replying behavior), active vs passive posts, use of mentions (mention network), hashtags, and publication patterns via text clustering. - Text preprocessing: remove duplicates, numbers, stop-words. - Clustering structure assessment: vary k=1..20; evaluate Average Silhouette Width (s) and Within-Cluster Sum of Squares (elbow criterion). Interpret s: strong (0.71–1.00), reasonable (0.51–0.70), weak (0.26–0.50), none (≤0.25). - Clustering: K-means++ initialization; K-means with Jaccard similarity; interpret cluster centers as publication patterns/templates. Citizen engagement analysis: Quantitative and qualitative. - Quantitative: reaction time (hours between authority post and reply); classify engagement as transient (replies only within first 24 h, no further activity) vs regular (activity beyond 24 h); activity ratio (days replied / days authority posted); relative activity duration a/b (a: time between first and last reply; b: time between first reply and authority’s last post); Kaplan–Meier survival analysis for probability of continued activity after first engaged day. - Qualitative: Topic analysis of reply content using Latent Dirichlet Allocation (LDA). • Preprocessing: stemming/lemmatization; remove stop-words, numbers, hashtags, links, emojis; keep unique words per reply. • Documents: aggregate all replies to a given authority post into one document to preserve context and ensure adequate length. • Topic quality and model selection: run LDA with topics=2..10; choose by coherence score (threshold ~0.5) and elbow criterion. • Interpretation: Reflexive Thematic Analysis; emphasize discriminative and stable topic words; merge semantically equivalent distinctions when appropriate. Also searched replies for predefined climate terms (e.g., climate change/global warming and Portuguese variants) to detect explicit linkage to climate change. Ethics/data: Only public tweets/replies via API; no sensitive personal data collected. Tools included TwitterAPI/Tweepy for collection; scikit-learn/SciPy for clustering; gensim for LDA; igraph for networks; R/ggplot2 for stats/plots.

Context and volume: - Over one year, five authorities posted 10,229 tweets; citizens posted 5,471 replies from ~2,600 people. Authorities rarely retweeted (78 in one year). Cities had differing meteorological warning profiles (e.g., Belém heavy rain 97.26% of alerts; Porto Alegre storms 56.52%). São Paulo posted with a predictable cadence (~3 posts/day, few or none on weekends). Communication strategy (largely single-loop, passive): - Citizen textual replies received: Belo Horizonte 2,656; São Paulo 62; Rio de Janeiro 2,602; Porto Alegre 89; Belém 62. - Authority replies to citizens were rare: São Paulo 0; Belo Horizonte 10; Rio de Janeiro 305; Porto Alegre 6; Belém 6. - Mentions: São Paulo used none. Others mostly mentioned other authorities, rarely citizens. Share of posts with mentions: Belo Horizonte 25.75%; Rio de Janeiro 5.61%; Porto Alegre 8%; Belém 6.14%. - Hashtags uncommon: Rio de Janeiro highest at 5.61% of posts; others <1%. Publication patterns (clustering evidence in BH, RJ, Belém): - Silhouette >0.52 and elbow improvements indicated reasonable/strong clustering in Belo Horizonte, Rio de Janeiro, Belém; not evident in São Paulo, Porto Alegre. - Dominant passive, informative templates: • Belo Horizonte: Rain record updates (64.29%), weather forecasts (12.70%), current weather (15.92%), geological risk alerts (1.05%), long-lived alerts (5.13%), upcoming event alerts (0.91%). • Rio de Janeiro: Rain record updates (46.81%), weather forecasts (39.13%), wind record updates (14.06%). • Belém: Tide table information (40.14%), weather forecasts (32.01%), technical inspection info (12.65%), phone/newsletter info (11.52%), current weather (3.68%). Citizen engagement dynamics: - Reaction time: Replies typically within minutes to hours of an authority post; rarely to older posts, even for long-lived alerts. - Engagement duration: Majority transient (single, short-lived cycle). Among regulars, activity ratio shows heavy-tailed inequality (few very active, many minimally active). Probability of remaining engaged beyond 200 days <0.25 across cities. Reply content (topics/themes): - LDA coherence ≥0.5 in 4 of 5 cities (BH 0.65 with 3 topics; SP 0.57 with 3 topics; RJ 0.54 with 3 topics; Belém 0.51 with 6 topics). Themes identified: • Disagreeing with information about their locations (e.g., reporting inaccuracies). • Complementing information with local observations (e.g., neighborhood-level updates). • Updating with current conditions (e.g., now/moment changes). • Informing about donation actions (observed in Belém). • Expressing thanks. - Citizens did not explicitly link local weather events to climate change in replies (predefined term search found minimal/none).

The study demonstrates that, despite the inherently two-way design of social platforms, local climate authorities primarily adopt a one-way, passive information dissemination approach centered on standardized templates (alerts, forecasts, records). This likely reflects resource and operational constraints and the need for clarity and consistency during weather events. Citizens respond rapidly but mostly transiently, correcting or complementing local details and offering real-time updates rather than engaging in sustained dialogue or broader climate change discourse. The heavy-tailed participation pattern mirrors known online participation inequalities. Engagement similarity across cities suggests that citizen behavior is driven less by specific authority posting styles and more by the nature of information needs during weather events and general online behavior. Practical implications include: (1) social media is effective for quick awareness and situational updates but may be less optimal for managing high-frequency updates during prolonged events; (2) geographic specificity matters—platforms lack fine-grained targeting by neighborhood, leading to information overload; (3) citizen feedback can support co-production and improve information accuracy, potentially influencing trust dynamics. The absence of explicit climate change framing in replies may relate to psychological distance, knowledge gaps, or reluctance to enter perceived controversial debates, indicating opportunities for better climate risk framing and education. Overall, social media can aid preparedness and adaptation by rapidly disseminating localized risk information and enabling citizen-sourced corrections. However, sustained engagement and deeper climate discourse likely require complementary channels, better geographic targeting, and strategies to foster trust and ongoing participation.

The work fills methodological, population, and practice-knowledge gaps by analyzing a full year of routine citizen–authority interactions on social media across five under-studied Brazilian cities. It shows that authorities build a structured, largely one-way social media presence with passive, periodic posts focused on alerts, forecasts, and records. Citizens engage quickly but transiently, primarily correcting, complementing, and updating local information, with little explicit linkage of weather events to climate change. Contributions include a combined behavioral modeling, clustering, and topic modeling approach to characterize publication patterns and engagement modes over time, and practice-oriented insights into how social media can support awareness, timely updates, and trust-building in climate event contexts. Future research should explore mechanisms to: (1) better integrate citizen-sourced local observations into official updates; (2) improve geographic targeting and information management during prolonged events; (3) understand and strengthen information credibility and trust dynamics; and (4) evaluate evolving platform policies and alternative platforms for public-interest climate communications.

- Representativeness: Social media users are not representative of city populations; active participants tend to be younger and more tech-savvy. - Scope: Focuses on X (formerly Twitter) only, excludes other platforms and citizen–citizen interactions (only authority–citizen threads analyzed). - Generalizability: Five Brazilian capitals with dedicated, active accounts; findings may not generalize to smaller cities or different governance contexts. - Observational design: One-year window; behaviors may vary in other periods or under exceptional events. Limited inference on causality. - Data granularity: Platform lacks neighborhood-level targeting; analysis relies on text content without geolocation of replies; predefined term search may miss nuanced climate discourse. - Clustering/topic limits: Publication pattern clustering evident only in some cities; LDA coherence below threshold in one city, and topic interpretation is qualitative.