Participation and Representation in Local Government Speech

1. Problem Statement

Local government meetings (e.g., city council meetings) are the primary formal channel for residents to directly influence policy and shape local agendas. However, empirical research on these institutions has been severely limited by data constraints. Existing studies typically rely on:

• Agendas or minutes: Which fail to capture the actual text of discussions.
• Small, unrepresentative samples: Often limited to single cities or short time horizons.
• Lack of speaker-level data: Inability to link speakers to demographic records (voter registration, property ownership).

Consequently, there is a significant knowledge gap regarding who participates in local democracy, what issues drive participation, and how institutional design choices (specifically access costs) affect the volume and composition of public input. This is critical because if the participant pool is unrepresentative, local governments cannot effectively aggregate local preferences.

2. Methodology

The authors construct a massive, novel dataset and employ advanced computational techniques to overcome previous data limitations.

A. Data Construction

• Scope: 25,000+ city council meetings across 115 California municipalities (2015–2024), covering ~20 million residents (52% of the state's population).
• Audio/Video Collection: Scraped from diverse municipal sources (websites, YouTube, official channels) to avoid selection bias inherent in YouTube-only datasets.
• Processing Pipeline:
  1. Transcription: Used OpenAI's Whisper (large-vocabulary speech-recognition model) to convert 75,000+ hours of audio to text.
  2. Diarization: Used pyannote (neural network-based) to identify distinct speakers and assign unique IDs.
  3. Issue Extraction: Used LLMs (Gemini 3 Flash) to segment meetings into discrete "issues," extract summaries, identify vote outcomes, and classify topics into 10 thematic categories (e.g., Land Use, Public Safety, Governance).
  4. Speaker Identification: Used LLMs to extract speaker names and roles from transcripts, validated against human coding (87% agreement on government vs. public status).
• Linking:
  • Voter Records: Matched speaker names to L2 voter registration data using a fuzzy-matching algorithm (~70% match rate).
  • Property Records: Matched voter addresses to CoreLogic property tax assessor files to determine homeownership status (~81% match rate of matched voters).
• Remote Access Data: Surveyed city clerks and manually reviewed meeting archives to create a fine-grained panel of "in-person," "hybrid," and "fully remote" participation modes, specifically tracking the staggered removal of remote options post-pandemic.

B. Analytical Framework

• Descriptive Analysis: Characterized meeting structure, topic distribution, and the demographic gaps between participants and the registered voter population.
• Causal Inference (Difference-in-Differences):
  • Natural Experiment: Exploited the staggered withdrawal of remote public participation options across California cities after the COVID-19 pandemic.
  • Estimator: Used the Callaway and Sant'Anna (2021) estimator to handle staggered treatment timing and avoid biases associated with standard two-way fixed effects (TWFE).
  • Identification: Compared cities that removed remote access at different times against "not-yet-treated" controls, assuming parallel trends in participation trajectories absent the policy change.

3. Key Contributions

1. Scale and Granularity: The largest dataset of local government speech to date, moving from minutes/agendas to full transcripts with speaker-level identification.
2. Demographic Linkage: Successfully links over 100,000 public speakers to voter and property records, allowing for rigorous comparison against the registered electorate.
3. Computational Social Science: Demonstrates a robust pipeline for using Large Language Models (LLMs) to structure messy institutional speech data (topic classification, stance detection, name extraction) and validates these outputs against human coding.
4. Policy Evaluation: Provides the first causal evidence on how meeting access costs (remote vs. in-person) impact both the volume and demographic composition of local political participation.

4. Key Results

A. Meeting Structure and Content

• Frequency/Length: Cities hold ~24 meetings/year; meetings last ~3 hours with ~30 speakers (nearly half are public).
• Unanimity: 87% of council votes are unanimous, a rate far higher than the U.S. Congress (10%). This is not an artifact of consent calendars or procedural motions but a central feature of local decision-making.
• Topics: Dominated by Governance & Administration (27%) and Infrastructure (15%). "Land Use & Zoning" issues generate the most close votes and the highest public engagement.

B. The Representation Gap

Participants are systematically unrepresentative of the registered voter population:

• Demographics: Participants are significantly older (+4 years), less female (-5 pp), more White non-Hispanic (+16 pp), more liberal (Democrat +6 pp), and more likely to be homeowners (+4 pp).
• Repeat Participation: While most individuals attend only once, the majority of speaking instances come from a small minority of "repeaters," who share the same demographic skew (older, whiter, male, liberal).
• Topic Sensitivity: Participation surges for specific topics. "Land Use & Zoning," "Social Equity," and "Housing & Homelessness" drive attendance. Notably, meetings focused on Social Equity attract younger and more female participants.

C. Impact of Access Costs (Remote Participation)

Using the staggered removal of remote options as a treatment:

• Volume Effect: Eliminating remote access reduces the number of public speakers per meeting (by ~1.8 to 2.4 speakers), though the effect is not always statistically significant in the aggregate.
• Composition Effect (Aggregate): In the pooled sample, removing remote access does not significantly change the overall demographic composition (age, race, gender) of the speaker pool.
• Heterogeneous Effects:
  • Age: In cities with older speaker pools, removing remote access significantly lowers the average age of speakers (older residents drop out). In cities with younger pools, the average age rises (younger residents drop out). These opposing effects cancel out in the aggregate.
  • Race: The reduction in participation volume is concentrated in whiter cities (where the representation gap is widest).
• Conclusion on Policy: Lowering access costs (via remote options) increases the volume of participation but does not clearly correct the underlying demographic skew. The marginal participants added by remote access do not necessarily fill the specific representation gaps (e.g., adding more young or minority voices) in a way that balances the pool.

5. Significance

• Theoretical: Challenges the assumption that simply lowering participation costs (e.g., going remote) will automatically lead to a more representative democracy. The skew in local participation appears driven by deep-seated structural factors (interests, civic norms, stakes) rather than just logistical barriers.
• Policy: While remote access increases total participation, local governments cannot rely on it alone to fix representation gaps. The findings suggest that the "who" of local democracy is resilient to changes in the "how" of access.
• Methodological: Sets a new standard for studying local politics by combining massive-scale data collection with LLM-driven text analysis and causal inference, offering a replicable framework for future research on institutional speech.