Climate Policy Elites' Twitter Interactions across Nine Countries

Technical Summary: Climate Policy Elites' Twitter Interactions across Nine Countries

Problem Statement
While social media is recognized as a critical platform for political contestation and democratic participation, existing literature on climate politics often fails to explicitly apply policy and governance frameworks to these digital spaces. Previous studies frequently conflate policy actors with the general public, focus on specific subsets of actors (e.g., only politicians or only NGOs), or rely on system-bounding rules that exclude actors based on activity levels. Furthermore, there is a lack of research examining the "sociotechnical" qualities of these platforms—specifically how the affordances of social media (such as the blurring of personal and professional boundaries) shape the organization and interaction of policy elites. This paper addresses the gap in understanding climate policy systems as polycentric governance structures operating within the specific constraints and opportunities of Twitter (now X), where organizational hierarchies may interact differently than in offline contexts.

Methodology
The study employs a multi-stage data collection and analysis protocol across nine countries: Australia, Brazil, Canada, Czechia, Finland, Germany, India, Ireland, and Sweden.

1. Actor Identification: The authors identified 904 climate policy organizations (government ministries, NGOs, businesses, research institutes) based on local expert knowledge. For each organization, they mapped four distinct levels of Twitter accounts:

   • Organization Primary: The official organizational account.
   • Individual Executive: Top-level decision-makers (e.g., CEOs, board chairs).
   • Organization Auxiliary: Accounts for subunits or specific functions.
   • Individual Non-executive: Rank-and-file members and employees.
     This resulted in a dataset of 16,086 unique accounts.

2. Data Collection: The researchers collected 40 million Twitter activities (tweets, replies, retweets) from January 1, 2017, to December 31, 2022.

3. Climate Content Filtering: To isolate relevant interactions, a strict dictionary-based classification pipeline was applied. This involved:

   • Text preprocessing (concatenating quoted tweets, removing URLs, splitting camelCase).
   • Language-aware lemmatization (using spaCy, Simplemma, or manual methods for English, Finnish, Swedish, German, Czech, Portuguese, French, and Hindi).
   • Keyword matching against a multilingual dictionary of climate terms.
   • Validation via manual inspection and Large Language Model (LLM) classification (Llama-3.3-70B-Instruct) to filter false positives.
   • Approximately 6.8% of the total dataset (2.7 million activities) was classified as climate-related.

4. Position Inference: To map policy stances, the authors utilized a two-step approach:

   • Survey Data: A graded response model was applied to survey data from 444 organizations in seven countries to derive latent climate policy positions (support for strong action vs. opposition).
   • Network Propagation: For organizations without survey data (including all accounts in Brazil and Canada), positions were inferred using a "network lasso" algorithm. This method propagates positions through the retweet network, minimizing the weighted variation between connected nodes while treating survey-derived positions as hard constraints.

5. Interaction Analysis: The study analyzed both direct engagement (direct retweets) and indirect engagement (joint retweets of third-party content) to map the structure of the policy system.

Key Results

• System Composition and Activity: There is significant variation in account numbers and activity levels across countries and organizational levels. While rank-and-file members constitute the largest group by number, primary organizational accounts are often as active, if not more so, than executive or auxiliary accounts.
• Temporal and Cross-National Variation: The proportion of climate-related activity varies significantly by country and time. However, all national systems responded similarly to major global events, showing peaks during the 2019 Fridays for Future strikes and COP26 (2021), and a distinct dip during the onset of the COVID-19 pandemic in 2020.
• Isolation of National Systems: Direct interaction networks (retweets) reveal that national policy systems are largely isolated from one another. While internal connectivity is dense, cross-border direct engagement is rare. Exceptions include higher connectivity between Australia and Ireland, and regional ties between Sweden and Finland.
• Indirect Connectivity via Bridging Actors: Despite the lack of direct cross-national interaction, national systems are indirectly connected. They frequently engage with the same content produced by specific "bridging" accounts. These bridging actors are predominantly international organizations (e.g., UNFCCC, UNEP, IPCC), high-profile climate activists, and researchers. In 2020, scientists were notably overrepresented among top bridging actors.
• Organizational Hierarchy: Primary organizational accounts tend to be the most connected nodes within their respective systems. However, non-official accounts (executives and staff) exhibit different interaction patterns, challenging the assumption of a single, coherent organizational voice on social media.

Significance and Claims
The paper claims to advance the understanding of climate governance by explicitly applying the concept of polycentric governance to social media research. It demonstrates that the policy system does not stop at organizational boundaries; rather, the "flat" nature of Twitter allows for complex, multiscalar behaviors where individual employees and executives interact alongside official accounts.

The study highlights the sociotechnical affordances of the platform, specifically how the blurring of personal and professional lives introduces variation within organizations. This challenges traditional views of policy actors as monolithic entities. The authors argue that while national systems rarely interact directly, they form a transnational system through the shared consumption and amplification of content from international organizations and key individuals.

Future Directions
The authors note that this work is an initial iteration. Future analyses will include:

• Statistical modeling of the generative features of these interaction networks.
• Expansion beyond retweets to include direct replies and links.
• Deeper analysis of bridging actors beyond just the top performers.
• Examination of the textual content of tweets to understand semantic differences across contexts and users.

The paper concludes that social media platforms act as unique spaces for climate policy contestation, where the interplay between organizational hierarchies and platform affordances creates a distinct, polycentric governance dynamic that differs from offline policy processes.