Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
89,004 characters · 21 sections · 44 citation commands
Participation and Representation in Local Government Speech
\pagenumbering{arabic} \setcounter{page}{1}
City governments are economically important institutions. They employ millions of workers, spend hundreds of billions of dollars each year, and make decisions over land use, public safety, infrastructure, parks, and other place-based services that shape household welfare and local economic activity us_census_bureau_2022_2024. In the classic fiscal federalism view oates_essay_1999, these are precisely the kinds of policy domains in which decentralized decision-making can be valuable, because local officials are better positioned to respond to heterogeneous local conditions and preferences. The efficacy of these institutions is predicated on the ability of local governments to effectively aggregate preferences of the local population. In this paper, we focus on the most direct and persistent formal forum for citizen participation in local government: city council meetings.
Under California’s Brown Act and analogous open-meeting laws nationwide, local elected bodies must deliberate in public, post agendas in advance, and provide opportunities for residents to speak. Yet despite the scale of the decisions made in these settings, data constraints have resulted in typically limited empirical evidence on the dynamics and characteristics of local government meetings, with existing datasets either relying on agendas or minutes and therefore failing to capture the actual text of discussion (e.g., einstein_who_2019) or collecting full transcripts but restricting attention to a potentially unrepresentative sample of cities (e.g., barari_localview_2023). As a consequence, we know surprisingly little about who actually participates, what kinds of issues draw them in, and how institutional design affects the costs of participation, factors which are central to local governments' ability to effectively aggregate local preferences.
To answer these questions, we construct a new dataset of over 25,000 city council meetings across 115 California municipalities from 2015–2024, comprising over 75,000 hours of audio generated by over 150,000 meeting participants that we transcribe, diarize, and annotate using state-of-the-art language models. We further augment these data via linking to voter registration and property ownership records. With this newly constructed dataset, we are able to provide the largest, most comprehensive characterization of meeting structure, participation, and content to date.
Our analysis proceeds in three steps. First, we provide a comprehensive characterization of city council meetings in California. Meetings are long, frequent, and have many participants. From the text of the meeting transcripts, we further identify issues of discussion and associated vote outcomes, thereby recovering meeting agendas. We then classify these issues into distinct meta-topics of interest to city councils. We find that most council votes are unanimous and that cities exhibit modest variation in topics of discussion.
Second, we conduct a comprehensive study of meeting participation, and find significant representation gaps between participants and registered voters: participants tend to be older, less likely to be female, more likely to be Democrats, more likely to be of White non-Hispanic origin, and more likely to be a homeowner than the registered voter population. These same demographic characteristics also drive selection into repeat participation. Participants also exhibit differential tendencies to appear in a meeting based on the topic mix of issues listed on the meeting agenda, with meetings emphasizing “land use and zoning" seeing substantially more public participation. We further find that while city-level characteristics impact participation rates, representation gaps in participants persist. As a consequence of these gaps, the participant pool is systematically different from the registered voter pool. This speaks to a core concern regarding the efficacy of local governance: if the participant pool is unrepresentative, then city governments are unable to effectively aggregate local preferences.
Third, we present a simple framework for an individual’s participation decision, in which a registered voter enters the participant pool if the benefits of participation (in terms of discussion of relevant content) outweighs the costs. Under this framework, there are two potential policy levers available to local governments to increase representativeness: they may either alter the topical content of their meetings, or they may alter meeting access costs. Since local governments are typically constrained in what topics they must discuss, we focus on changing access costs as the key policy lever available to local governments to improve representation. By leveraging quasi-experimental variation in meeting access costs during and after the COVID-19 pandemic, we are able to estimate causal effects of changing meeting access costs on participant representativeness. We find that increasing meeting access costs leads to lower participation rates and alters the age makeup of the participant pool.
Taken together, our results constitute the most comprehensive empirical study to date of the mechanics of local democracy, allowing us to provide novel characterizations of how meetings are structured, what is discussed, who participates, and how both topics of discussion and variable costs of meeting access drive trends in participation.
This paper contributes to several literatures on public economics, political economy, political science, and computational social science.
First, our paper relates to work in public economics, urban economics, and local political economy on decentralized governance and the politics of local public goods. In classic fiscal-federalism accounts, decentralization can improve welfare when local governments are better able to respond to heterogeneous local preferences and conditions oates_essay_1999, besley_centralized_2003. A large related literature emphasizes that local institutions are shaped by unequal political participation, differential stakes, and the distribution of property ownership fischel_homevoter_2009, tausanovitch_representation_2014, warshaw_local_2019. These concerns are especially salient in land use and housing, where local political voice can be highly asymmetric and where homeowners often have both stronger incentives and greater capacity to participate fischel_homevoter_2009, yoder_does_2020, einstein_who_2019, hankinson_when_2018, marble_where_2021. Our paper speaks to this literature by documenting participation gaps in one central local institution at scale and investigating the efficacy of a convenient policy lever available to local governments to reduce these gaps.
Second, our paper contributes to a growing empirical literature in political science on public meetings, hearings, and descriptive representation in local politics. Scholars have emphasized that municipal institutions are understudied relative to their importance in structuring daily life and public service provision trounstine_representation_2010, warshaw_local_2019. Related work has also documented representational skew in other local participatory settings, including land-use hearings and planning processes, and has begun to connect commenter preferences to policy outcomes in particular agencies and cities sahn_public_2025, einstein_still_2023.
The closest substantive antecedent to our work in this strand of the literature is einstein_who_2019, who show that participants in suburban planning and zoning meetings are disproportionately older, whiter, male, and homeowner-heavy relative to the broader public. We extend on their analysis in several ways. First, our dataset exhibits far greater geographic and temporal breadth, covering 115 diverse cities in California over the last decade, compared to 97 towns in metropolitan Boston in a three-year span. Consequently, we observe a far greater number of public participants who we are able match to voter records (over 100,000, compared to approximately 3,000), and collect full transcripts rather than minutes. Additionally, we collect data on city council meetings, a materially different venue than planning and zoning boards which covers a far wider range of topics. Finally, by collecting data before, during, and after the COVID-19 pandemic, we are able to speak to the variable effect of remote participation options on meeting structure, content, and participation.
Third, our paper also relates to work on participation costs, institutional design, and the accessibility of democratic fora. A long tradition in political economy treats participation as sensitive to the costs of time, travel, scheduling, and information downs_economic_1957, riker_theory_1968. Those cost considerations are especially important in local settings, where meetings are often held on weekday evenings, at fixed physical locations, and on agendas that vary substantially in salience. Recent work on local meetings during and after the pandemic suggests that remote and hybrid formats can alter who appears in these fora, though the effects on representation are not straightforward einstein_still_2023. Our contribution is to bring a larger panel and a cleaner institutional design to this question. Leveraging staggered withdrawals of remote public participation across cities, we study how lowering or raising the cost of attendance affects both the extensive margin of participation and the demographic composition of speakers, investigating a key policy lever available to local governments to address documented representation gaps.
Finally, our paper contributes to the computational social science and text-as-data literature by showing how modern language models can convert a large, messy corpus of institutional speech into analyzable data. Existing work in computational social science has developed tools for measuring meaning and group differences in political and administrative text gentzkow_measuring_2019, card_computational_2022. More recent work has used large language models for classification, extraction, and annotation tasks in economics and related fields, including the analysis of regulatory and legal texts bartik_costs_2023, modarressi_causal_2025, durvasula_counting_2025. The closest large-scale data precedent in the local-government context is the LocalView project, which assembled a large corpus of local government meeting transcripts from YouTube and demonstrated the promise of these materials for social-scientific research barari_localview_2023. We build on that agenda in several ways: by expanding coverage beyond YouTube-based meetings, by recovering speaker-level participation and issue-level structure from raw transcripts, by linking speakers to voter and property records, and by validating each major stage of the extraction pipeline against human-coded labels. In that sense, the paper contributes both new substantive evidence on local participation and a reusable measurement framework for studying local political speech at scale.
A core contribution of this paper is the construction of a comprehensive, novel dataset of city council meeting transcripts from 115 cities in California. The dataset covers over 25,000 city council meetings between 2015 and 2024, representing more than 75,000 hours of recorded deliberation. These cities range from large metropolitan areas such as Los Angeles and San Francisco\footnote{As San Francisco is both a city and a county, its meetings are technically those of the County Board of Supervisors rather than a city council. They are included regardless.} to mid-sized and smaller jurisdictions like Palo Alto and Los Gatos, together accounting for over 20 million residents or about 52% of the state’s population. We augment these data by matching our compiled transcripts with L2 voter record data labels__lists_inc_l2_l2_2025, CoreLogic property ownership corelogic_nz_limited_and_cotality_cotality_2024, and a survey of city clerks we conducted to determine remote meeting availability during and after the COVID-19 pandemic. An overview of our data pipeline can be found in (ref). This section describes each component of this pipeline.
We begin by scraping archived meeting audio and video files from municipal websites and official YouTube channels. While towns are required to make recordings of their city council meetings available to the public, there is significant heterogeneity in their choice of channel. Past work has scraped meetings from YouTube (most notably the LocalView dataset, compiled by barari_localview_2023), but we found that this significantly limits coverage, both within and across towns. By expanding our dataset to include additional channels, we remove potential sources of confounding via selection into certain meeting channels and dramatically expand coverage (see (ref) in (ref) for details). Our resulting dataset covers a wide array of cities, with significant heterogeneity across size (in terms of both geography and population), age, racial characteristics, income, and property values. (ref) summarizes these characteristics, with the final column displaying statewide values for reference. For a view on the geographic spread of our collected data, see (ref) in (ref). Since our selected sample contains all of California's largest cities and misses the long tail of small municipalities in the state, cities in our sample exhibit somewhat higher population densities, percentages of Black residents, incomes, and home values, and correspondingly exhibits somewhat lower percentages of Hispanic residents.
We then generate transcripts using OpenAI's large-vocabulary speech-recognition model, Whisper radford_robust_2022, providing a mapping from timestamps to text. We simultaneously identify distinct speakers in the recorded audio via neural network-based diarization framework, pyannote bredin_pyannoteaudio_2019, which detects distinct speaker voiceprints and assigns each detected speaker a unique meeting-level ID. This provides a mapping from timestamps to speaker, which we then merge with the transcript data to provide a speaker-level text record. We then implement filters to remove extremely short ($<$15 minutes) and extremely long ($>$15 hours) meetings that are unlikely to provide informative content.
To further structure our data, we segment meetings into discrete “issues”: continuous stretches of discussion focused on a specific policy topic or agenda item. We prompt an LLM (Gemini 3 Flash) to extract, for each issue: (1) a short title and summary; (2) whether the issue was agendized (i.e. was included on the published meeting agenda) or was raised in unstructured public comment periods; (3) whether the city council and/or members of the public engaged with it; (4) whether a vote was held on the issue; (5) the vote outcome (e.g., number of votes for vs. against); and (6) the stage at which the vote occurred (e.g. final passage, a motion to table discussion to the next meeting, etc.). See (ref) for prompting details. Across all meetings, we identify roughly 320,000 issues (about twelve per meeting).
We then classify each issue into one of ten thematic categories via LLM discovery modarressi_causal_2025, liu_llm-guided_2025. We identify these topics through a two-stage procedure. First, because the full set of issues summaries would exceed a single context window, we randomly shuffle all issues and their summaries and pass them to an LLM into eight context-window sized chunks. For each chunk, we prompt the LLM to inductively identify five to ten recurring city-governance topics, including names, descriptions and representative examples. We then provide the chunk-level topic lists to a second LLM prompt, which consolidates overlapping categories into a fixed taxonomy of ten topics. We then use an LLM to classify issues into each of these ten topics. The resulting ten topics capture the dominant domains of city‐council activity—from budgets and land use to policing and infrastructure—with stable frequencies across cities and years. (ref) illustrates the learned topics, along with the number of issues classified as belonging to that topic. (ref) in (ref) provides examples of common issues within each topic.
We validate this procedure against hand-labeled ground truth from seven randomly sampled meetings, observing high levels of agreement between human labelers and the LLM on issue agendized status, vote occurrence, vote tallies, and issue-level precision and recall. See (ref) for details.
To augment our speaker-level text records, we employ the same LLM to infer speaker names and roles from linguistic cues and context contained in the produced transcripts. This task is made feasible by the fact that council meetings are highly formalistic, and almost always require public speakers to identify themselves at the beginning of their speech. We pass a meeting transcript to the LLM and prompt it to return a mapping from speaker ID (output by the diarization process) to the speaker's full name, whether or not they are a member of the government, and whether or not they speak on behalf of a group. See (ref) for prompting details. Manual validation against 1,300 hand-labeled speakers shows 87% agreement on government/member of public status of the speaker and relatively high accuracy in retrieving the names of public speakers in a given meeting. Among speakers where both the LLM and human labelers identified a name, last-name exact match rates reach 74%, with most disagreements reflecting misattribution across speaker segments rather than fabricated names.\footnote{We focus on last name match accuracy because our fuzzy matching algorithm prioritizes last name similarity.} Fewer than 2 percent of LLM-only names ultimately matched a voter record, suggesting that potentially hallucinated names rarely propagate into downstream analyses. Additional details on the hand-validation exercise can be found in (ref).
We supplement our speaker-level speech data with demographic information about speakers drawing from two datasets. First, we link identified public speakers to the L2 voter registration data labels__lists_inc_l2_l2_2025 using a custom fuzzy-matching algorithm which takes into account possible errors in spelling from the transcription process. The algorithm prioritizes exact last-name and locality matches, returning a unique candidate when confidence exceeds a threshold. See (ref) for further details on the matching algorithm. We achieve an overall match rate of approximately 70%.\footnote{This is a relatively high matching rate given the messy nature of transcripts and the scale of the data. For example, sahn_public_2025 achieves a match rate of 48% when matching San Francisco Planning Commission public commenters to the voter file; donahue_politics_2023 achieves a match rate of 68% matching police officers to a voter file using name and date of birth; einstein_who_2019 achieve a match rate of 83% matching Boston metropolitan area meeting participants to the voting file.} To investigate the possibility of poorer matching on non-white names, we run all participant names through a transformer-based model for predicting race and ethnicity (ethnicolr, chintalapati_predicting_2023)
In (ref), we check balance between matched and unmatched names by calculating a standardized difference. We observe reasonable balance across broad origin groups, though our sample size is so large that even small differences are statistically significant. The worst match is among African-origin names (e.g., “Salah El-Bakri” or “Zuzka Ejena”), which compose 3.7% of the matched sample and 6.2% of the unmatched sample. Otherwise, the matching algorithm performs quite consistently across names by predicted racial/ethnic origin. Matched voter records provide age, gender, party registration, and race/ethnicity estimates, allowing us to benchmark meeting participants against the registered electorate in each city-year.
We further match speakers to CoreLogic property-tax assessor files corelogic_nz_limited_and_cotality_cotality_2024 by the address reflected in their voter registration obtained from L2. This allows us to identify homeownership status of public participants. We are able to match 81% of matched L2 speakers to CoreLogic addresses. As Panel B of (ref) demonstrates (using L2's demographic data), we are able to achieve an overall match rate of 58% between the registered voter populations of the 115 towns in our sample (as given by L2) and CoreLogic property record data. Most demographics appear well-balanced, with the exceptions of “Age" and “Big City" - our matched sample is biased slightly towards younger residents and residents from outside of the six largest cities in the state.
Ordinarily, city council meetings take place in person, with residents delivering remarks at a podium in council chambers. The COVID-19 pandemic abruptly disrupted this routine. In March 2020, California implemented statewide emergency orders suspending in-person quorum and posting requirements, allowing city councils to convene and accept public comment remotely for the first time. Within weeks, most cities in our sample adopted some form of remote public participation, such as Zoom or call-in phone lines. In this paper, however, we define remote public participation narrowly to mean live remote participation: residents must be able to participate synchronously as part of the meeting itself, whether by video or live telephone. Cities that accepted only written emails, e-comments, or prerecorded voicemail submissions are therefore not coded as offering remote public participation. While the onset of live remote access was nearly simultaneous across jurisdictions, the return to in-person-only participation was highly staggered. Some cities reinstated in-person-only public comment by mid-2021, while others continued to allow live remote participation in hybrid or fully remote formats for years.
We systematically document this variation using two complementary sources. First, we conducted an email-based survey of city clerks across all 115 municipalities in our sample, asking when remote participation was introduced, modified, and discontinued for both the public and councilmembers. Second, we manually reviewed hundreds of meeting videos, agendas, and city websites to verify whether each meeting offered a live remote option for public comment. These two sources together yield a fine-grained panel of meeting-level participation modes, categorizing each meeting as “in-person only", “hybrid", or “fully remote" for both the council and the public.
Summary statistics of our compiled remote public participation data are found in (ref). Approximately 83% of towns in our sample offer live remote public participation at some point during the pandemic, of which roughly two thirds have since eliminated that option. On average, towns which eliminated previously offered live remote public participation do so after approximately 30 months, or 2.5 years, yet we observe substantial heterogeneity in exact timing of this removal.
A summary of how our sample size changes throughout each step of this data augmentation process can be found in (ref). Note that we omit the survey of city clerks since, unlike the other filters, this operates at the town level and not the meeting level; since we learn remote access dates for all cities in our sample, this does not lead to a reduction in sample size. Note that we filter to meetings occurring after Jan 1, 2018 due to data limitations for the L2 voter record, for which we have access to annual data beginning in 2018.
We proceed to leverage this data to characterize the structure and content of city council meetings, analyze the participant pool, and assess the efficacy of remote public participation options as a policy lever to improve representation in city council meetings.
In this section, we present a descriptive analysis of the structure and content of city council meetings in California. We begin by presenting summary statistics, displayed in (ref). Cities hold an average of 24 meetings a year, with meetings lasting approximately 3 hours on average. We observe approximately 30 speakers on average per meeting, with nearly half of all speakers being members of the public. Meetings cover an average of 15 issues, and roughly half of those issues receive public comment.
A particularly striking pattern is what happens once issues reach a formal vote. Roughly 43% of raised issues include a vote by the council, and virtually all votes pass: only about 1% fail. More surprising, however, is the degree of consensus: 87% of votes are unanimous. This is far above the rate observed in Congress: only 10% of votes in the modern Congress have no recorded nays.\footnote{Using Voteview's historical roll-call database, we calculate that only 6,330 of 111,979 House and Senate roll calls in completed Congresses through the 118th had no recorded nays, or 5.65%. Even restricting the comparison to the modern Congresses for which official online roll-call records are most complete, the rate remains low: among the 101st through 118th Congresses, 3,504 of 34,041 roll calls, or 10.29%, had no recorded nays. Voteview's roll-call export excludes quorum calls and vacated votes. See lewis_voteview_2026.
State legislatures provide a more mixed comparison and data is more sparse. jewell_party_1955, discussing keefe_parties_1954, reports that in the 1951 Pennsylvania legislature, 82% of Senate roll calls and 70% of House roll calls were unanimous; he also notes that unanimous votes averaged roughly two-thirds of roll calls in Keefe's study of the 1949 and 1951 Illinois sessions.}
Because this is an unusually high level of agreement, we audited whether it was an artifact of how votes were extracted. One concern was that consent calendars might mechanically inflate unanimity, since a council often casts a single vote on a batch of consent items while our issue-level data can record each item separately. Carefully ensuring to collapse consent-calendar items to one row only brings the unanimous vote share to 85.5%. A second concern was that unanimity might be driven mainly by purely procedural motions, such as votes to continue an item rather than approve or reject a substantive policy. We therefore use an LLM to separate final votes from procedural votes. This does not explain away the pattern: final votes are actually more likely to be unanimous at 88% compared to procedural votes which are only unanimous 78% of the time. Finally, we also consider whether the result is driven by ceremonial business, such as proclamations or certificates of recognition.\footnote{We classify an item as ceremonial if the concatenated issue, summary, and vote-outcome text matches the following case-insensitive regular expression: \textbackslash bproclamations?\textbackslash b|\textbackslash bcommendations?\textbackslash b|\textbackslash bcertificates? of (recognition|appreciation)\textbackslash b|\textbackslash btributes?\textbackslash b|\textbackslash bhonoring\textbackslash b|\textbackslash bin honor of\textbackslash b|\textbackslash bin memoriam\textbackslash b|\textbackslash badjourn(ed|ment)? in memory\textbackslash b|\textbackslash b(awareness|heritage|history|appreciation|prevention) month\textbackslash b|\textbackslash bday of remembrance\textbackslash b, excluding matches that also contain \textbackslash b(covid|coronavirus|emergency proclamation|local emergency|state of emergency|disaster|disasters)\textbackslash b.} Votes on these topics are more likely to be unanimous--95%--but there are so few that removing these votes leaves the unanimity count at 87%. Thus, the high rate of unanimity is not merely a consent-calendar artifact or a byproduct of procedural housekeeping. It appears to be a central feature of city council decision-making in our sample.
In (ref) we examine the varying content of local government meetings based on topics of discussion. We find that most issues focus on matters related to Governance & Administration, while Land Use & Zoning sees a markedly higher incidence of close votes (defined as a vote outcome in which one vote flipping would tie or change the outcome) than other categories.
We can further analyze the variable attention paid to these classes of issues by the city council versus members of the public. We accomplish this by examining the share of speaking time dedicated to a given class of issues within each meeting, separately for issues that are included on the meeting agenda and those that are raised in unstructured public comment sessions. This study reveals notable heterogeneity, with certain cities seeing marked divergence in attention between city councils and members of the public. Two examples are presented in (ref). In (ref) and (ref), we see that members of the public place greater emphasis (as measured by speaking time allocation) on Housing & Homelessness and Environmental Sustainability than the city councils of San Diego and Palo Alto, respectively. By contrast, members of the public in Eastvale under-emphasize discussion of city council operations and administration, as evidenced by (ref).
Finally, we can examine how the attention paid to various classes evolves over time, separately for agendized issues and unstructured public comment. (ref) presents these time series for three selected topics. (ref) shows a sharp increase in attention paid to matters related to Public Health with the onset of the COVID-19 pandemic. (ref) shows a similar sharp increase for issues related to Public Safety, but notably the increase is much larger in magnitude for unagendized public comment than agendized discussion. (ref) shows a pattern of cyclical increases and decreases in discussion related to Social Equity over time, with clear spikes and regions of elevated focus following major events related to civil rights and racial or religious violence.
In this section, we characterize the pool of public participants in city council meetings. We find that participants tend to be older, whiter, more liberal, more male, and more likely to be a homeowner than the full population of registered voters. We further find significant heterogeneity in the participation patterns across cities, with increased public participation correlated with city-level renter share, Gini coefficient, level of education, median income, and share of White residents. Cities with larger populations and larger shares of Black residents, meanwhile, see lower rates of public participation. We continue to characterize the subpopulation of repeat participants, individuals who attend multiple city council meetings in the same city. We find that most participants in city council meetings are in fact repeat participants, and that repeat participants are similarly more likely to be older, whiter, more liberal, more male, and less likely to be a homeowner than the participant pool as a whole. Finally, we show that participants select into discussion based on the topics of meeting agendas, with discussion around Land Use & Zoning, Social Equity, and Housing & Homelessness being significant drivers of attendance across demographic groups. Collectively, these findings provide a comprehensive overview of who participates in city council meetings in California.
A central question in assessing city council meetings as venues for deliberative democracy is whether the individuals who participate resemble the broader public they ostensibly represent. Prior research (e.g., fowler_what_2025) has shown that the voting population skews older and whiter relative to the general population. Yet far less is known about the population who takes the additional step beyond voting to actually show up to speak in local government meetings.
The most systematic prior evidence comes from einstein_who_2019, who hand-coded participants in planning and zoning board meetings in 97 cities and towns in metropolitan Boston between 2015 and 2017. They successfully matched 83 percent of 3,123 speakers to voter records and found that meeting participants were, on average, eight years older and more likely to be male than the general voting population. In a detailed case study of one town, they further linked 85 participants to property records and found that homeowners comprised 78 percent of attendees, compared to 61 percent of the town’s population.
In addition to adding geographic and temporal diversity, our dataset scales that approach by nearly two orders of magnitude: we match over 100,000 participants across 115 California municipalities to voter registration and property data for the same cities. Approximately 1 in 1,000 registered voters participate in their local city council meeting in a given year. (ref) reports simple differences in means between participants and non-participants. As in the Boston sample, meeting attendees in our data are older (+4 years), less likely to be female (–5 percentage points), more likely to be Democrats (+6 p.p.), and substantially more likely to be of White non-Hispanic origin (+16 p.p.). Participants are also somewhat more likely to be homeowners (+4 p.p.), although as noted earlier, our matched ownership sample is biased toward smaller towns, providing somewhat high baseline ownership rates.\footnote{For example, the Census Bureau estimates the homeownership rate in California to be around 55% bureau_housing_nodate. However, the Public Policy Institute of California finds the rate to be closer to 73% among likely voters, suggesting that our overall mean of 70% may not be a significant overstatement baldassare_californias_2025.}
These gaps are substantively large even relative to familiar forms of political participation. One benchmark is voting itself. Although this comparison is imperfect---our benchmark is the registered voter population, whereas Census benchmarks compare voters to the citizen voting-age population (CVAP)---the racial skew in meeting participation is considerably larger. Our White non-Hispanic gap is 16.4 percentage points, compared with 6.9 percentage points among 2022 voters and 4.2 percentage points among 2020 voters.\footnote{These Census benchmarks compare the White non-Hispanic share of voters to the White non-Hispanic share of the citizen voting-age population (CVAP), rather than to the registered voter population. In the CPS Voting and Registration reports, the corresponding White non-Hispanic overrepresentation is 4.2 percentage points in 2020 and 6.9 percentage points in 2022. See fabina_voting_2022 and fabina_voting_2024.} A second benchmark is campaign giving. Here, if anything, the donor class appears even more selective: in linked donor-voter data, Democratic donors are 19.0 percentage points whiter than non-donors hill_representativeness_2017. In this sense, city council speakers appear less representative than the electorate and closer to the donor class than to the voting public.\footnote{For scale, about 66.8% of adult citizens voted in 2020, whereas roughly 8.5% appeared as federal donors in the 2020 cycle. Because these are national benchmarks and the donor figure is measured over a presidential cycle, we treat them as rough reference points rather than directly comparable participation rates.}
To account for local composition and year effects, (ref) reports results from an individual-year-level logistic regression with city and year fixed effects. Each coefficient represents the marginal association between a demographic trait and the log-odds of that person participating in a council meeting. The results confirm the descriptive patterns: older, Democratic, and White non-Hispanic-origin residents are significantly more likely to participate, while women remain less likely to do so even after controlling for location and time. Homeownership is also positively associated with participation, though its effect is smaller in magnitude once other demographics are held constant. Together, these estimates suggest that the demographic biases in participation are systematic and persistent across cities and years.
Another way to visualize this, at least for age, is to estimate the probability of participation by age among registered voters. As (ref) shows, those around 65-75 years old are the most likely to participate in meetings, with a noticeable bump for 18-year-olds, which is likely driven by the frequency of high school students appearing at meetings and the relatively low denominator of registered voters aged 18.
Having documented extensive individual-level differences between participants and non-participants, we now identify city-level characteristics that are predictive of these differences. (ref) plots per-capita participation against six Census-derived covariates. The share of adults holding a bachelor's degree or higher is the strongest single predictor: a one-standard-deviation increase in educational attainment is associated with roughly a doubling of per-capita participation. Median household income and the Gini coefficient show similar positive associations, consistent with participation being concentrated in affluent communities with high human capital. City population exhibits a strong negative relationship---larger cities have substantially lower per-capita rates---which likely reflects both the mechanical denominator effect and the greater anonymity and lower perceived influence of any individual speaker in a large jurisdiction. Racial diversity (measured as $1 - \text{HHI}$) is negatively associated with participation.
Importantly, the demographic biases documented in (ref) are not artifacts of a few outlier cities. Across nearly every city in our sample, the age gap between participants and non-participants is positive, the White non-Hispanic-origin gap is positive, and the female gap is negative, as reflected on the vertical axes of (ref) below. The universality of these patterns suggests that the forces generating unrepresentative participation---whether differential costs of attendance, differential stakes, or differential civic norms---operate similarly across very different local contexts.
A natural question is whether cities that attract more participation also attract more representative participation. (ref) plots each city's representation gap against its per-capita participation rate. For age, the relationship is positive: cities with higher participation tend to have larger, not smaller, age gaps between participants and non-participants. The White non-Hispanic-origin gap is roughly flat across participation levels, and the female gap, on the other hand, improves in high-participation cities. These patterns suggest that simply increasing the volume of participation may not, on its own, correct representational distortions.
Among those speakers for whom we are able to obtain a match to the L2 voter record, we can further analyze the frequency with which a given speaker appears repeatedly at meetings within a given town. We do so by classifying participants into one of two categories: “one-timers" (individuals who participate in exactly one meeting in a given city) and “repeaters" (individuals who participate in more than one meeting in a given city). We begin by presenting summary statistics of the repeater status of the participant pool. (ref) demonstrates that while a significant majority of participants are one-timers (Panel A), most participation instances are driven by repeaters (Panel B), with the median repeater participating in six separate meetings and a long tail of participants appearing many more times than that (Panel C). These facts collectively demonstrate that most individuals who participate in a meeting participate only once, but most participation is perpetuated by a relatively small minority of repeat participants.
Additionally, we observe significant heterogeneity in the proportion of participants who are one-timers across cities. (ref) shows the distribution of this proportion, averaged across all meetings in a given city and demonstrates that these cities behave very differently in their norms regarding repeat participation, with towns ranging from one-time participant rates of 16% (Baldwin Park) to 65% (San Diego).
We can further characterize the difference between repeaters and one-timers by regressing whether a participant is a repeater (i.e. they have either participated previously or will appear again in the future) on their observed demographics, via logistic regression. The results of this regression are shown in (ref).
These coefficients suggest that repeat participants are more likely to be older, more liberal, have White non-Hispanic ancestry, and be male, findings which are qualitatively similar to those presented in (ref), suggesting that selection on observables into repeat attendance, conditional on attending once, behaves similarly to selection into attendance in the first place. A key difference, however, is that conditional on attending at least once, homeowner status is (weakly) associated with a lower likelihood of repeat attendance ($p=0.25$). This is consistent with a paradigm in which homeowners are more motivated to attend to speak on an issue directly relating to their property ownership, while non-homeowners who attend meetings do so out of concerns for the broader community and so are more likely to attend repeatedly. However, this effect is not statistically significant and so may be the result of random noise, in which case the selection patterns into repeat participation are identical to those into participation in the first place.
We can additionally study how varying concentrations of topics in meeting agendas result in different participant pools. That is, by merging our L2-matched speakers to our extracted issue identification and topic classifications, we can determine how meetings with various emphasis on different topics draw different types of attendees. In order to reconstruct a proxy for the meeting agenda (which potential participants may use when choosing whether or not to participate) we calculate the percentage of issues in a given meeting belonging to each topic class, among those issues which were determined to be agendized via our LLM-based issue identification task. We then merge these proportions with participant demographics for the given meeting and regress each of our demographics variables on these topic proportions. The resulting coefficient estimates can be found in (ref). Notably, we observe that dicsussions around Land Use & Zoning and Social Equity see high attendance; meetings with greater emphasis on Social Equity and Economic Development see younger participants; and meetings with greater emphasis on Community Services, Social Equity, and Public Health see more female participants.
One advantage of full transcript data is that it allows us to move beyond asking which issues draw public participation and instead examine how residents frame those disputes once they arrive. We illustrate that added leverage in the domain that generates the highest levels of public engagement in our data: land use and zoning. Across our full sample, 96,604 individual speaker-issue comments—spanning 7,502 distinct agenda items in 6,335 meetings across 115 California cities—were made on issues classified as “Land Use, Zoning, and Urban Development.” To study how supporters and opponents of development differ in their expressed views and rhetorical framing, we construct a topic-specific stance measure: for each comment, we provide an LLM with both the issue description and the speaker's text and ask whether the speaker supports or opposes a more development-oriented outcome on a continuous scale from --1 (clearly opposed) to 1 (clearly supportive).
We aggregate this stance measure at the city level to characterize the public mood toward development among a city's most politically active residents. On average, public commenters lean modestly pro-development (mean score = 0.14),\footnote{This estimate likely overstates pro-development sentiment among ordinary residents because our speaker classification distinguishes only between city officials and non-city government speakers. As a result, developers, applicants, and their representatives are included in the “public speaker” category rather than separated out from other members of the public.} though with substantial variation (standard deviation = 0.61). Pro-development sentiment was relatively stable from 2015 through 2021, hovering around 0.17--0.19, before declining modestly in 2022 and 2023. This slight softening in expressed pro-development sentiment in the most recent years of our sample may reflect the intensification of community opposition to state-mandated housing policies, though interpretation requires caution given the changing composition of issues that come before councils over time.
To understand how supporters and opponents differ in their rhetoric, we examine which topics are most frequently invoked by each side ((ref)) through keyword frequency.\footnote{Keyword frequencies were computed by searching for regular‐expression matches in the transcript text. For instance, we counted mentions of “parking” (\textbackslash bparking?\textbackslash b); “traffic” (any of traffic|congestion|cars?|gridlock); “affordability” (affordable|affordability|income); “crime” (crime|safety|police|violence|theft|assault|criminal); “schools” (school|student|classroom|enrollment); “infrastructure” (infrastructure|sewer|water|drainage|utility|utilities|pipes?|electric|power); “aesthetics” (\texttt{aesthetic|appearance|beauty|character|historic|preserv|style|design|architecture|shadow|tall}); and “environment” (\texttt{environment|wildlife|animal|bird|tree|pollution|toxic|habitat|greenhouse|climate|hazard}). All expressions are case‐insensitive and bounded by word boundaries (\texttt{\textbackslash b}).} Pro-development speakers are defined as those with a stance score of at least 0.7 (24,516 comments); anti-development speakers are those with a score of --0.7 or lower (13,145 comments).
The results confirm a clear rhetorical divide. Pro-development speakers disproportionately reference affordability and economic opportunity, while opponents focus on traffic, crime, schools, environmental impacts, and building aesthetics. These word-level differences suggest that each side situates development debates in distinct moral and practical frames---economic necessity and housing need versus neighborhood preservation and quality-of-life concerns.
Having observed substantial representation gaps in (ref), we now turn to a key policy lever available to local governments to improve representativeness of public participation: meeting access costs. We introduce a simple framework for meeting participation in which access costs enter into the individual's participation decision. We then leverage the introduction and subsequent staggered removal of remote public participation options during and after the COVID-19 pandemic as a natural experiment to estimate causal effects. We find that the elimination of remote participation options, corresponding to an increase in participation costs, results in fewer participants per meeting.
Why does anyone show up to a city council meeting? Adapting the canonical calculus-of-voting framework presented in riker_theory_1968, suppose citizen $i$ attends a meeting when $B_i + D_i > C_i$, where $B_i$ captures the expected policy benefit of participation--likely a function of how much is at stake for i on the agenda and how likely i's comment is to influence the outcome, $D_i$ captures expressive returns from showing up and feeling heard (or, in the authors' conception, “civic duty"), and $C_i$ is the cost of attendance: travel time, opportunity cost of an evening, childcare, mobility constraints and so on. Note that $C_i$ likely varies systematically with demographics: older residents with health limitations may face physical costs of attending in person while working-age residents with young children or inflexible jobs may face high opportunity costs concentrated in the evening hours when meetings occur. Homeowners with property values directly at stake may have higher $B_i$ on land use items--but renters with the same $C_i$ may have weaker perceived $B_i$ if they believe the council is less responsive to their preferences einstein_who_2019, fischel_homevoter_2009.
Under this framework, a municipal government faced with unrepresentative public participation has two primary policy levers at its disposal: it can alter the topical content of meetings (impacting $B_i$) or it can alter meeting access costs (impacting $C_i$). City councils, however, are largely constrained in what topics of discussion they must cover in a meeting, with procedural matters, budget and finance, and discussion of anodyne issues of governance dominating meetings, as shown in (ref). Consequently, the chief discretionary policy lever available to local governments is control over meeting access costs $C_i$, which they can reduce/increase via introduction/removal of a remote participation option. Given the aforementioned heterogeneity in $C_i$ across demographics, the effect on such a change on the composition of the participant pool will depend on who the marginal participant is.
As described in (ref), California's emergency orders moved nearly all city council meetings online in March 2020. Of the 101 cities with usable participation data, 96 adopted live remote public comment during the pandemic. The return to in-person formats was highly staggered: 61 of those 96 eventually eliminated remote access — at dates ranging from early 2021 through late 2024 — while 35 continued to offer it through the end of our sample ((ref)). The remaining 19 cities never offered live remote access at all, accepting only email or voicemail submissions, or permitting in-person attendance.
(ref) plots average public speakers per meeting separately for cities that ever adopted remote public access and those that did not.\footnote{The “Never Added Zoom” group consists of the 19 cities in our sample that accepted only email or voicemail submissions and never offered live remote participation. These cities serve as a useful visual benchmark but are excluded from the DiD analysis, which relies on within-Zoom-adopter variation in the timing of shutoffs.} Both groups track closely in the pre-period — averaging roughly 15–18 speakers per meeting through early 2020 — and both collapse to around 5–7 speakers in March 2020. The recovery, however, diverges slightly. By mid-2020, cities with remote public participation rebounded to approximately 14 speakers per meeting and sustained that level through 2022. Cities without live remote access recovered slightly more slowly.
(ref) decomposes the participation time series by age group, plotting average voter record-matched speakers per meeting for residents below 40 and above 65. Before the pandemic, the over-65 group contributed substantially more speakers per meeting — roughly 3.0–4.0 compared to 2.0–2.5 for the under-40 group, a gap consistent with the age skew documented in (ref). Both groups collapse sharply in March 2020. But the recovery paths differ in a revealing way. The under-40 group recovers quickly and essentially reaches its pre-pandemic level by mid-2020. The over-65 group, by contrast, recovers more slowly and never fully returns to its pre-pandemic baseline — settling around 2.5 speakers per meeting, compared to roughly 3.5 before March 2020. The net result is a persistent narrowing of the age gap in participation.
This convergence is consistent with two (non-mutually-exclusive) interpretations. One is that the initial shift to remote access disproportionately benefited younger, working-age residents — reducing their opportunity costs enough to close the gap with retirees who had always found it relatively easy to attend. The other is that the pandemic itself discouraged older residents from attending, whether due to health risk aversion that persisted beyond the acute phase or a loss of the in-person social routines that had sustained their participation. The DiD design presented in (ref) below — which isolates the effect of removing remote access, conditional on the pandemic having already occurred — helps separate these channels.
Due to the lack of variation and clear confounding in the initial introduction of live remote public participation options at the start of the pandemic, we instead exploit the staggered withdrawal of remote public comment across California cities to estimate the causal effect of meeting format on participation. Let $Y_{it}$ be an outcome---total public speakers, mean age, share female, share Democrat---for city $i$ in month $t$. Define $D_{it} = \mathbf{1}\{t \geq G_i\}$, where $G_i$ is the month city $i$ first eliminates live remote public comment. Treatment is absorbing: no city in our sample reintroduces remote access after discontinuing it. We estimate:
\[ Y_{it} \;=\; \alpha_i + \lambda_t + \beta\, D_{it} + \varepsilon_{it}, \]
with city fixed effects $(\alpha_i)$ and month fixed effects $(\lambda_t)$. Under the parallel trends assumption---that, absent a shutoff, outcomes in treated and control cities would have followed similar trajectories---$\beta$ identifies the average treatment effect on the treated (ATT). With staggered treatment timing, standard two-way fixed effects can produce biased estimates due to negative weighting of already-treated units goodman-bacon_difference--differences_2021, borusyak_revisiting_2024. We therefore implement the callaway_difference--differences_2021 estimator, which avoids this issue by restricting comparisons to not-yet-treated and never-treated controls and aggregating group-time ATTs with non-negative weights. Standard errors are clustered at the city level throughout. In order to remove potential confounding due to systematic differences between cities that never introduce remote participation options and those that do, we restrict ourselves to a control set of not-yet-treated units.
The key threat to identification is that cities which eliminated remote access earlier may have been on different participation trajectories than those which kept it. We do not think this is likely. In our survey of city clerks, respondents frequently cited idiosyncratic, administrative reasons for ending remote comment---the expiration of a software contract, changes in how the city attorney interpreted emergency authorizations, or, in several cases, a particularly disruptive “Zoom bombing” incident that prompted an immediate return to in-person formats. These accounts suggest that shutoff timing was driven more by logistical happenstance than by strategic considerations about who was participating or what was being discussed.
We corroborate this qualitative evidence with three statistical checks. First, we test whether shutoff timing is predicted by observable city characteristics. A regression of remote access duration on pre-pandemic ACS covariates---population, median income, racial composition, educational attainment, renter share, and income inequality---yields an $R^2$ of 0.03, indicating that the decision to withdraw remote comment was largely orthogonal to city demographics. A discrete-time hazard model predicting month-to-month shutoff probability from lagged county COVID case counts and vaccination rates yields a pseudo-$R^2$ of 0.12, with neither predictor statistically significant after controlling for city-level demographic characteristics. The exact timing of shutoffs therefore does not appear to have been driven by local pandemic conditions.
Second, we inspect pre-treatment dynamics directly. The event study in (ref) plots dynamic ATTs by event time relative to the shutoff month. Pre-treatment coefficients are flat and jointly indistinguishable from zero, consistent with the parallel trends assumption.
Third, we report results both with and without residualizing on meeting-level topic composition shares. If shutoff timing happened to coincide with shifts in agenda content---which could independently affect who shows up---controlling for topic shares should change the estimates. As (ref) shows, it does not: the adjusted and unadjusted estimates are nearly identical.
We begin with participation counts, the outcome most directly tied to access costs. The event study in (ref) plots dynamic ATTs relative to the shutoff month (event time $e=0$). Pre-treatment coefficients are flat and jointly indistinguishable from zero, supporting the parallel-trends assumption. We find an overall ATT of a reduction of 1.8 speakers ($p=0.33$) following the removal of remote participation options, an effect which increases in magnitude to a reduction of 2.4 speakers ($p=0.21$) when controlling for meeting topical content.
We extend this analysis to the various demographic variables we obtain from matching to the voter record. (ref) summarizes the overall ATT for each outcome. We detect no significant changes the demographic makeup of the participant pool.
One possible interpretation is that meeting format simply does not matter much for participation, but the framework in (ref) also motivates a different reading. If remote access lowers attendance costs for different demographic groups in different cities---older residents facing mobility barriers in some places, younger workers facing scheduling constraints in others---then compositional effects could run in opposite directions across cities and wash out in the pooled sample. We investigate this possibility directly, focusing on age and race as the dimensions along which heterogeneity may be present, motivated by our descriptive analysis in (ref). We split treated cities at the median of (i) pre-treatment mean speaker age and (ii) city-level share white (from the 2019 ACS), and re-estimate the ATT separately for each subsample. All never-treated cities serve as controls in both splits. (ref) reports the results.
The age split is striking. In cities with older speaker pools (mean speaker age $\geq 53$), the speaker count does not change significantly ($-0.9$, $p = 0.57$), but mean age falls by 3.6 years ($p < 0.001$). Remote access appears to have drawn in older residents---plausibly those facing mobility or health barriers to in-person attendance---who exit when that option is removed, leaving a younger residual pool. In cities with younger speaker pools (mean speaker age $< 53$), the pattern reverses: mean age rises by 5.9 years ($p = 0.06$). This is consistent with younger, working-age residents---who may have valued the scheduling flexibility of remote comment---dropping out when in-person attendance is required. These opposing compositional effects cancel in the pooled sample, explaining the null on age in (ref). The share-white split tells a complementary story about the extensive margin. Eliminating remote access reduces the speaker count by about 5.7 speakers ($p = 0.09$) in whiter cities but has no detectable effect in less-white cities ($-0.3$, $p = 0.91$). The participation decline concentrates entirely in cities whose existing commenter pool most over-represents white residents.
Taken together, these results suggest that the pooled null on participation and demographics conceals meaningful heterogeneity along the very dimensions where public comment is least representative. Remote access expands participation along different margins depending on the city's demographic baseline---but the quantity effect concentrates in whiter cities, precisely those where the existing representation gap is widest. At the same time, even where remote access increases the volume of participation, it does not obviously correct the underlying compositional skew documented in (ref). Lowering the cost of attendance brings in more speakers, but not necessarily more representative ones.
Local government meetings remain the most durable and visible arena of participatory democracy in the United States. Yet, they have remained chronically under-studied, with most existing evidence suffering from small sample sizes and inconsistent coverage. In this paper, we compile what is to our knowledge the largest, most comprehensive dataset of city council meeting transcripts in a given region of the United States, transcribing and analyzing over 25,000 meetings across 115 cities in California over the last decade. By linking this data to voter and property ownership records, we are able to provide a highly granular and detailed look at the structure and format of these meetings, characterize who participates and how participation changes with meeting content, and assess the causal effect of changing barriers to meeting access (in the form of remote participation options) on participation rates and demographics.
Our evidence shows that meetings are long, frequent, have many participants, and cover a variety of topics; participants tend to be older, whiter, more male, more liberal, and more likely to be homeowners than the full registered voter population; participation patterns vary across cities depending on city size, median income, level of education, renter share, development, and racial diversity; participants are drawn to meetings that focus more on Land Use & Zoning, Housing & Homelessness, and Social Equity; and increasing costs of attendance by removing a remote participation option leads to a short-term reduction in the number of public participants but does not clearly change the demographic composition of speakers.
This project also opens up compelling avenues for future work. This includes examining how measures discussed in meetings evolve through the meeting process; how public comment sways future vote outcomes; and how comments and votes by councilmembers are correlated with electoral outcomes. More broadly, this line of research shows significant promise for deepening our understanding of local government dynamics and informing the design of more effective, representative democratic institutions.