Learning through online participation: A longitudinal analysis of participatory budgeting using Big Data indicators

2.1Big Data analytics

The term Big Data is often characterized by the three Vs (Connelly et al., 2016; Höchtl et al., 2016; Laney, 2001): high volume (e.g., data size), variety (e.g., data forms or features), and velocity (e.g., the speed at which data are generated). For instance, digital platforms such as Twitter produce a large amount of data (e.g., millions of tweets) in real time (e.g., seconds) consisting of various data forms (e.g., strings, numbers, media) and data features (e.g., ID, URL, language, words). Big Data includes machine-generated data, such as sensor networks from the ground, air, and space, and user-generated data, including Web access logs, texts, media, transactions, clickstreams, timestamps, threads, locations, voting, and Facebook Likes collected from web applications, the Internet of Things, and mobile devices (Krishnan, 2013). This dataset is highly unstructured (non-machine-readable formats) and high dimensional (the number of features is higher than the number of observations), making conventional statistical approaches challenging to process.

Big Data analytics refers to techniques applied to collecting, managing, and analyzing Big Data (Elgendy & Elragal, 2014; Shu, 2016; Vassakis et al., 2018). The term analytics is used to highlight multiple methods involved in multiple stages from data mining to final applications (Shu, 2016). In contrast to standard statistics based on data sampling and inference, Big Data analytics usually do not involve sampling but rather combining multiple and unstructured datasets to identify latent patterns and theoretical insights using, for instance, a mix of machine-learning techniques, social network analysis, and time-series analysis (Chang, 2018; Elgendy & Elragal, 2014; Shu, 2016; Wang et al., 2020). These methods rely on diverse algorithms to automatically classify, cluster, and summarize large datasets for fast and better decision-making.

Big Data analytics has its flaws, though. Bigger data does not mean representative data, unless such data have been properly processed (Boyd & Crawford, 2012). For instance, neither Twitter or Facebook nor public platform data alone fully represent citizens’ voices online. Therefore, Big Data needs to be synthesized and contextualized to ensure its quality and to answer “how” and “why” questions. Researchers have recently started combining Big Data analytics with surveys and qualitative methods to improve data quality (Davies et al., 2021; Eck et al., 2019). Moreover, many studies have pointed out that Big Data suffers from low quality due to a lack of collaboration across the various public silos in producing, managing, and sharing such data (Connelly et al., 2016; Ruijer & Meijer, 2020), calling for a new data collaborative governance (Ruijer, 2021).

2.2Big Data analysis for online deliberation research

This section now narrows its focus to Big Data analytics for online deliberation research. Deliberation refers to a process of weighing competing considerations regarding specific issues (Fishkin, 2009; Fishkin & Luskin, 2005). Many scholars distinguish deliberation from other similar forms of political communication, such as debate, dialogue, and consultation, by claiming that deliberation focuses on addressing collective problems or making decisions through a fact-regarding, other-regarding and future-regarding exchange of arguments (Cohen, 2003; Fishkin, 2009; Fishkin & Luskin, 2005; Gastil & Black, 2007; Gutmann & Thompson, 2004; Manin, 1987; Offe & Preuss, 1991). For instance, political dialogue involves exchanging views on collective issues and discovering a new viewpoint, but decisions are not the primary focus (Escobar, 2011). While it might be almost impossible to reach a complete consensus among stakeholders (Gastil & Black, 2007; Thompson, 2008), deliberation is essentially about “the process that precedes decision” (Manin, 1987, p. 345), more specifically, a binding decision that influences stakeholders for a certain period (Gutmann & Thompson, 2004). Therefore, deliberation research involves normative debates on the principles underlying public engagement in “justifiable” collective decision-making regarding equality, respect, and accessibility (Dahlberg, 2001; Fishkin, 2009; Gastil & Black, 2007; Löfgren & Agger, 2021; Steenbergen et al., 2003).

Online deliberation is a young subfield of deliberation literature that studies the recent phenomena of digital-technology-enabled deliberation processes (Beauchamp, 2018; Beierle, 2004; Friess & Eilders, 2015; Jonsson & Åström, 2014; Strandberg & Grönlund, 2018). Digital technologies help overcome the limitations of face-to-face engagements in terms of time and space, enabling a many-to-many deliberation at a large scale. In addition, online deliberation can provide resilient public spheres during a crisis, such as a pandemic (Elstub et al., 2021). However, scholars have also found negative aspects of online deliberation and communication in terms of new digital divides, low effectiveness, cyberbullying, disinformation, and radical communities (Lim & Oh, 2016; Lowry et al., 2017; Norris, 2001; Vaccari & Chadwick, 2020). Therefore, online deliberation does not automatically enhance democratic values and requires further assessment.

Many studies have classified deliberation research into an input (e.g., who participates and under what conditions), throughput (e.g., how people deliberate and its democratic quality), and output dimension (e.g., results and their distribution in society) with distinctive research design and methods (Beauchamp, 2018; Beste & Wyss, 2019; Friess & Eilders, 2015; Geissel & Newton, 2012; Strandberg & Grönlund, 2018). To our knowledge, Big Data analytics has mostly been used to explore the black box of digital platforms, particularly discussion quality and communicative processes. For instance, recent studies have applied automated computational methods, including unsupervised (e.g., topic modeling) or supervised machine learning techniques (e.g., random forests), within a broad family of natural language processing to assess the content quality of large-scale online discussion data (Arana-Catania et al., 2021; Beauchamp, 2018; Fournier-Tombs & di Marzo Serugendo, 2019; Fournier-Tombs & MacKenzie, 2021; Parkinson et al., 2020) or in combination with network analysis to assess its network structures (Aragón et al., 2017; Choi, 2014; Gonzalez-Bailon et al., 2010). A fundamental set of questions raised by the above articles is based on what transpires on the digital platform. With respect to registered users, just who engages in online deliberation? What topics do they discuss, and with whom? How do they deliberate, and what is the quality?

Given that emphasis has been placed on content quality, our contribution is to assess dynamic interactions in online deliberation (Shin & Rask, 2021). Our indicators are based on Fishkin’s (2009) two democratic dimensions – participation and deliberation – where equality serves as an overarching value. Instead of focusing on a certain sampled population, digital platforms can theoretically invite “everyone” to a virtual roundtable and give them an equal chance to obtain information, exchange voices, and vote. Indeed, it has great potential precisely because mass participation is a cornerstone of democracy (Fishkin, 2009). Based on the participation dimension, we developed participation rate (how many participate), activeness (how actively they participate), and continuity (how consistently they participate) indicators to measure the volume of online deliberation over time (Shin & Rask, 2021).

The deliberation dimension focused on how different viewpoints are equally considered during deliberation based on five criteria (Fishkin, 2009; Fishkin & Luskin, 2005): information (equal access to factual claims), substantive balance (of competing sides through reciprocal communications), diversity (of viewpoints), conscientiousness (sincere weight and respect), and equal consideration (regardless of the speakers). As Fishkin and Luskin (2005) highlighted, these criteria promote mutual, diverse, and sincere interactions between different social groups engaged in deliberation. Based on the deliberation dimension, we developed responsiveness (degree of reciprocal communications), inter-linkedness (connections between diverse actors and proposals), and commitment (degree of equal commitment) indicators by considering various digital environments. We operationalize each indicator later in detail.

4.1Data collection

The primary data source was the OmaStadi platform (https://omastadi.hel.fi), which had adopted the Decidim system (https://decidim.org/). While Decidim offers a comprehensive GraphQL API schema for public access to the database, the second-round data were not available during the data collection period. Therefore, we programmed a custom web scraper in a Python environment and collected first-round data in May 2020 and second-round data in May 2021 that cover the first three stages – proposal, screening, and co-creation – in the two rounds (first round: until April 21, 2019; second round: until May 26, 2021). Since Shin and Rask (2021) found almost no online activity after the co-creation stage in the first round, primarily due to the prolonged summer break in Finland, this article focuses on the earlier deliberation processes. The parsed data contain the proposal ID, proposal title, proposal area (n= 8), proposal status (impossible or possible), proposal description, type of post (initial comments or replies), comments, author ID (anonymized), and date.

The data collection process was based on a research agreement with the City of Helsinki, and we followed Internet Research: Ethical Guideline 3.0 (Franzke et al., 2020), which provides the updated guideline for Big Data analytics. Based on the guideline, the collected data do not contain any personally identifiable information, with pseudonymized IDs used to prevent personal tracking. Figure 1 provides a graphical representation of the online discussion data we collected. We count one incident of engagement in online deliberation when a person comments on a citizen-generated PR proposal. For instance, assume that Actor #1 posted the following comment for Proposal #1: “I like this idea, but …”. We graphically express the actor and proposal as a circle and square, called nodes, and the comment as an edge. We call a set of actors and proposals connected via comments an actor-proposal network, technically a two-mode or bipartite network. While we did not collect any personal information (e.g., age and gender), we collected metadata on public comments and reconstructed them as network and time-series data (Shin & Rask, 2021).

4.2Deliberative quality indicators and methods

Table 1 summarizes six democratic indicators proposed by Shin and Rask (2021). First, the participation dimension focuses on the volume of online participation. The participation rate indicator measures what percentage of the population participated in online deliberation, calculated by the number of commentator IDs divided by the total population of Helsinki. While Shin and Rask (2021) measured a pool of registered IDs, this article more narrowly focuses on commentator IDs to capture actual engagement in online deliberation in a comparative setting. The activeness indicator measures a longitudinal trend of participation in deliberation. We traced daily and longitudinal trends using moving average and linear regression models (Hamilton, 1994). The continuity indicator measures the durability of online engagement without significant interruptions. We created a binary variable (1: if daily comments exceed 5; 0, otherwise) to measure the consistency of daily activities. Using this variable, we calculated continuity by counting the number of active days divided by the number of days in the investigation period.

Figure 1.

Graphical representation of online discussion data.

Second, the deliberation dimension focuses on interactions in online deliberation. The responsiveness indicator measures the degree to which comments are reciprocated, reflecting back-and-forth conversations. It is calculated by the number of replies divided by the total number of comments. The inter-linkedness indicator examines who communicates with whom and how proposals are related to each other through a three-step process. The first step is to investigate interaction patterns by visually inspecting actor-proposal networks to check for an abnormal network structure, as shown in Fig. 2a. For instance, Actor 1 commented on three proposals, whereas Actor 2 commented on two proposals. The second step is to use this information to create a new proposal-proposal network for examining indirect interactions between proposals, called a one-mode projection of the two-mode network (Zweig, 2016). A few active participants often engage in many proposals in online deliberation settings, creating challenges for observing proposal-proposal interactions. Following Newman (2001), we addressed this problem by inversely assigning weights to comment ties with a simple equation, 1/(n-1), where n denotes the number of commentators’ comments. For instance, in Fig. 2b the strength of the tie between proposals 1 and 3 is 0.5=1/(3-1) because of the number of comments made by Actor 1. If Actor 1 comments on 100 proposals, his/her contribution to the proposal-proposal relationship will be reduced to 1/99. The third step is to use the Louvain algorithm (Blondel et al., 2008), a family of community detection methods, to identify clustered proposals. Overall, Fig. 2b is an abstract version of Fig. 2a using the three-step process. Lastly, the commitment indicator measures the variability in the degree of engagement in online deliberation to identify the extent to which commentators and proposals equally engage in deliberation.

Figure 2.

Graphical representation of a network projection.

We combined network and time-series analyses to calculate the six indicators. A time-series analysis refers to the process of analyzing time-series data, which records data points indexed within specified periods of time to analyze the structural patterns of the observed past data or forecast future trends (Hamilton, 1994). In this article, we used a time-series analysis to measure participation rate, activeness, and continuity in the participation dimension. A network analysis refers to the process of analyzing network data, which records relations between nodes (e.g., actor, protein, and computer) within specified periods of time (Zweig, 2016). We conducted a network analysis to measure responsiveness, inter-linkedness, and commitment in the deliberation dimension.

5.1Descriptive statistics

The first case of COVID-19 in Finland was discovered in January 2020, when the second round of OmaStadi was beginning the co-creation stage. The daily confirmed cases rapidly increased to 500 until March 2020, when the Finnish government declared the first state of emergency since World War II (Kuitunen et al., 2020). Lockdowns followed, and all public facilities and shops were closed and gatherings of more than ten people were prohibited. Helsinki, the capital city of Finland, has a dense population, meaning it was gravely affected by the worldwide disaster. It canceled all offline meetings and workshops in the second round. Instead, OmaStadi shifted the whole PB process online and extended the co-creation stage to almost five months, as compared to two months in the first round. The city’s quick response mitigated the shock, a fact reflected in the descriptive statistics.

Table 2 summarizes the two rounds of the OmaStadi project. The first noticeable change in the second round is that the budget doubled in size. It was a welcoming change since only 6.5% of 1,273 proposals had ultimately been selected in the first round due to the initial low budget, making the process highly competitive (Rask et al., 2021; Shin & Rask, 2021). The number of proposals and plans also increased. In the second round, citizens submitted 1,456 proposals, an increase of 14.4 percent compared with the previous round. During the co-creation stage, citizens and experts combined the proposals into 468 plans (the final number was changed to 466) based on their content. Citizens voted on the various proposals in the final stage. On average, 3.1 proposals were transformed into a single plan.

The number of comments significantly increased from 2,699 to 5,740 in the second round during the investigation period, partly due to the fact that city officials provided the rationale for screening the results or practical guidelines through comments. We believe that such content is also part of the online deliberation process of conveying the voices of those in the city. However, since this article focuses on the interaction (network) and dynamic processes (time series) centered around residents, we excluded comments made by city offices and borough liaisons. Borough liaisons are experts employed by the municipality who help citizens to prepare proposals, facilitate collaborations, and coordinate the implementation of approved PB plans (Rask et al., 2021). Overall, descriptive statistics show that the second round substantially improved the deliberative process despite the pandemic.

In the next subsections, we investigate online deliberation using six democratic indicators.

5.2Participation rate

This article counts participation as occurring whenever people commented on a proposal, a fundamental step in engaging in online deliberation. Therefore, the participation rate indicator measures the number of commentators as a percentage of the Helsinki population. Figure 3 shows that the numbers of commentators were 1,058 and 1,091 in the first and second rounds, respectively. Based on population data for Helsinki (Helsinki Region Infoshare, n.d.), we obtained a 0.16 percent participation rate in the first round and a 0.17 percent participation rate in the second round. This result was somewhat unexpected because the number of comments doubled in the second round (Table 2). It indicates that a few individuals posted many comments in the second round, providing the rationale for excluding city officials’ comment data from the analysis.

Figure 3.

Participation rate.

Overall, we found that the participation rate was low and not substantially enhanced in the second round. This is striking, since in general OmaStadi has received significant attention from residents. For instance, the voter turnout in the first round was 8.6% of the population, with 80% of the voting occurring online (Rask et al., 2021, p. 44), a remarkable result because it was the piloted project for distributing 0.1% of the city budget.

The low participation rate is not a problem unique to OmaStadi. Borge et al. (2022) found low online participation rates based on a survey of 34 Catalan municipalities that adopted Decidim, pointing out that it might lead to biased citizen input and a loss of trust and legitimacy. It is difficult to pinpoint a root cause for the problem because many intrinsic and extrinsic reasons influence people’s motivation to participate (Brabham, 2012; Li et al., 2020; Royo et al., 2020). While this calls for a comprehensive evaluation to explore barriers in more detail, it is worthwhile to consider the digital platform (Decidim) itself since the platform design can facilitate or restrict online deliberation (Esau et al., 2017, 2020).

Deseriis (2021) reviewed six existing digital platforms and characterized the Decidim as a “hybrid” model. The platform enables citizens to endorse proposals and comment online, resulting in 3,533 pages of proposals and plans accumulated in the last two OmaStadi rounds (see Table 2). Therefore, it serves as a grand reservoir of citizens’ voices without any flow, requiring more in-depth discussions on synthesizing diverse proposals for allocating the budget. The Decidim platform is hybrid in the sense that it makes it possible to share contact information and event calendars for organizing face-to-face meetings to weigh proposals outside the platform, while the OmaStadi case did not use this function. In other words, Decidim provides deliberation spaces for developing individual PB proposals but relies on other means, including offline meetings, citizen assemblies, Facebook, and Zoom meetings, for the synthesis of proposals. Without the latter function, citizens face challenges in navigating and contributing to fragmented proposals.

Previous studies found that this hybrid model did not work properly since online and offline activities were not seamlessly integrated (Borge et al., 2022; Rask & Ertiö, 2019). Moreover, the COVID-19 pandemic impacted the second round by forcing officials to restrict all offline meetings, exacerbating the condition of participation. Therefore, we underline that the early history of OmaStadi might not be the sole cause of the low online participation rate.

5.3Activeness

The activeness indicator measures the number of daily engagements in online deliberation. Following Shin and Rask (2021), we applied two time-series techniques to monitor activeness trends: moving average and linear regression models (Hamilton, 1994). First, we applied a linear regression model to draw a long trend line that fits the data. An upward trend is a positive signal for activeness. In Fig. 4, the grey lines (with a dark grey area) show daily engagements, whereas the red lines show the linear trend. The linear lines show that neither round had any significant upward or downward trend in the long run, despite drastic spikes when residents submitted their proposals. The spike shows the maximum level of activeness in the year-long deliberation process because public interest generally fades over time.

Figure 4.

Activeness. Note: The grey line shows the number of daily engagements in online discussions. The red line shows the linear trend; the blue line shows the moving average. The shaded areas distinguish different stages: proposal, screen, and co-creation.

Second, a (simple) moving average was used to smooth out short-term fluctuations with a constantly updated moving average. In Fig. 4, the blue lines show the smoothed trend. We found potential seasonality (a recurring pattern at a fixed period, such as a week or year) of activeness by pairing the two rounds. Activeness as a measure of participation reached its lowest level in both rounds when city officials filtered citizen-generated proposals by labeling them as either “possible” or “impossible” in the screening stage.

Activeness provides three pieces of useful information for monitoring online deliberation. First, it quantifies whether online participation shows an uptrend or downtrend in the long run. Second, it helps researchers detect abnormal daily spikes due to various reasons (e.g., topicality, spam, events) that require a prompt response to facilitate resilient deliberation. Third, it helps actors design deliberation processes for the next round. Based on the two-round data, we learned that the digital platform can temporarily become overloaded because so many people participated in discussions at the proposal deadline, raising the following questions: Is the proposal stage too short for citizens to brainstorm and submit their ideas? Is there a way to disperse the over-concentration caused by specific events? Given the rapid cooling-down period in participation during the screening stage, is there a way to maintain public attention online?

5.4Continuity

The continuity indicator measures the extent to which the digital agora is consistently alive with public engagements. Based on our observations of a low participation rate and fluctuating activeness in both rounds, we investigated system-wide continuity by imposing a threshold. That is, we created a dichotomic variable of continuity: yes, if the number of daily comments was > 5; no, otherwise. We then recorded the values in both rounds. We could customize the threshold depending on different contexts and platforms. Figure 5 provides a graphical representation of the results, where the white areas show inactive participation on a corresponding day.

Figure 5.

Continuity. Note: If the number of daily comments exceeds five in the OmaStadi platform, the corresponding day is filled in with the color black; otherwise, it is filled in with the color white. Therefore, white areas show the days when online participation was inactive.

We found that 59.1% of the investigation period demonstrated continuity in the first round and 62.5% in the second round. For instance, if the entire period was ten days, then the first round shows 5.9 days of continuity and the second round 6.3 days, constituting an improvement. It is also insightful to note the high level of durability during the co-creation stage when COVID-19 severely disrupted the local community. It implies that online deliberation can still function quite well even when face-to-face interactions are restricted. We also found significant discontinuation around and during the screening stage, supporting the results regarding activeness.

The results show that continuity improved slightly in the second round. Gutmann and Thompson (2004) have argued that the continuation of deliberation is crucial because it keeps the process dynamic, reflecting citizens’ diverse viewpoints in changing contexts. When the participatory platform is constantly alive, it becomes much easier to motivate inflows of new ideas, actors, and information against dominant perspectives; thus, continuity is a convenient way to measure the prosperity of digital agora. Although the second round showed a slight improvement in continuity, the result was not fully satisfactory because we had set a substantially low threshold of five comments across all proposals per day.

Overall, all three indicators measuring participation rate, activeness, and continuity point to the conclusion that the volume of online deliberation was still low despite minor improvements in the second round of OmaStadi. The results reveal that the significantly increased statistics, such as the number of proposals and comments (Table 2), merely scratch the surface of online participation. We need to be cautious when interpreting the quantified information and triangulate data as much as possible. The rest of the sections thus focus on interactions in online deliberation.

5.5Responsiveness

The fourth indicator, responsiveness, measures the proportion of replies that demonstrate the extent to which online deliberation is reciprocated. In Fig. 6, the light grey area shows the number of total comments, whereas the dark grey area represents the number of replies. We can visually identify two prominent patterns. On the one hand, the second round showed a higher responsiveness rate, specifically 24.7%, as compared to 14.7% in the first round, meaning that citizens created a higher degree of back-and-forth discussions than in the first round. While deliberation “cannot make incompatible values compatible” (Gutmann & Thompson, 2004, p. 11), reciprocal communication creates opportunities to consider the multiple viewpoints of others, which can help clarify or alter their preferences when those arguments have merit (Fishkin, 2009). Therefore, responsiveness is a fundamental trait of deliberation.

Figure 6.

Responsiveness.

It is also significant that different stages shape different levels of responsiveness. In the first round, responsiveness was only 6.6% in the proposal stage but increased to 22.6% in the co-creation stage. Similarly, the second round showed a responsiveness rate of 8.9% in the proposal stage and 38% in the co-creation stage. This implies that people tend to express simple opinions regarding newly proposed ideas during the proposal stage (e.g., I like this proposal!); they gradually engage in more reciprocal discussions during the co-creation stage.

The increased level of responsiveness in the second round could partly be due to conversations between city officials and proposal organizers on technical matters (e.g., submission guidelines, administrative matters, and meetings). Borge et al. (2022) found a tendency for managerial use of the Decidim platform in their Catalonian case. While such a tendency can improve the transparency and traceability of the participatory process between citizens and public officials, Borge et al. (2022) highlighted that a large proportion of technical communication can also restrict use of the platform to merely collecting proposals and voting on them rather than fostering citizen debate. It is a crucial argument to be tested in OmaStadi. Overall, the results show that responsiveness was substantially increased in the second round, but we acknowledge the need for content analysis to classify discussion contents and measure their democratic quality.

5.6Inter-linkedness

The inter-linkedness indicator explores interactive patterns among commentators and proposals. OmaStadi classifies proposals according to eight categories (east, west, north, south, southeastern, northeastern, central, and the entire City of Helsinki), among which this article focused on the “entire” category, a collection of city-wide proposals. Proposals for the entire/whole city category have received substantial attention from citizens in terms of the proportion of proposals (first round: 20.7%; second round: 27.5%), comments (17.5%; 25.2%), and commentators (16.7%; 24.6%).

Figure 7 visually represents the discussion networks in the entire category. As explained in Fig. 1, circles denote actors, squares denote proposals or plans (red: “impossible” proposals; green: “possible” proposals; yellow: plans), and the ties that connect nodes represent online comments. Here, we have discussed two important aspects of these network graphs. First, the screening stage substantially influenced online deliberation. At the screening stage, city experts labeled 838 out of 1,273 proposals (65.8%) as “possible” in the first round and 1,146 out of 1,456 proposals (78.7%) as “possible” in the second round. While the screening intervention had become more generous (from a 65.8% to a 78.7% acceptance rate), Fig. 7 shows that online comments for “impossible” proposals (red squares) quickly faded from public interest. This transition indicates that residents acknowledge the fundamental mechanism of OmaStadi and strategically shift their focus. Second, we did not find a divided pattern of interactions, as found in certain areas in the first round (Shin & Rask, 2021). Identifying polarized discussion patterns will help resolve conflicts in PB processes.

Figure 7.

Actor-proposal networks in the entire category.

Now, the actor-proposal network during the co-creation stage (Fig. 7) can be converted into a proposal-proposal network (Fig. 8) to explore how proposals (squares) cluster with each other via comments. During the co-creation stage, 94 commentators engaged in 87 proposals and plans with 150 comments in the first round, while 133 commentators engaged in 168 proposals and plans with 280 comments in the second round. Therefore, the second round involved a broader range of commentators and proposals, reflected in a denser proposal-proposal network in Fig. 8b.

Figure 8.

Proposal-proposal networks in the entire category. Note: Squares represent proposals and ties represent comments for the Entire area category. For instance, if Actor A commented on both on Proposals A and B, we drew a tie between the proposals, showing an inter-linkedness. Different colors were assigned based on the Louvain community detection algorithm.

The central proposal groups in both rounds were marked red in Fig. 8. Note that the designation “central” does not mean that they were the “most influential” or “most popular” proposal groups, as measured by the number of comments. Instead, they served as a hub connecting other proposals via comments. For instance, some red-colored proposals received few comments, but they were linked with other popular proposals. In the first round, this group consisted of 13 proposals concerned with the management of urban forests and parks or environmental campaigns; the second-round proposals also shared similar concerns, including a carbon-neutral community, renewables, and wildlife campaigns, with a total of 33 proposals. This result provides crucial information on how proposals interact and create a central PB theme for a locality. By exploring proposal-proposal networks in two rounds, we found that residents raised their voices with respect to environmental concerns regarding parks, outdoor and cultural activities, and campaigns viewed as a city-wide problem. Content analysis of proposals in the first round reflects this finding by showing a large proportion of environmental, equity, and development of communication issues in the entire/whole city category (Rask et al., 2021, p. 11).

5.7Commitment

The commitment indicator measures the varying degrees of participation of commentators and proposals in online deliberation. Figure 9 shows that the distribution curves commonly leaned to the left with long right tails, called a right-skewed (positive skew) distribution. This means that only a few actors were substantially active, and a few proposals received many comments while most of them remained inactive in both rounds. The silent majority on a digital platform could be problematic in the online deliberation context because only a few actors and proposals could dominate the process. While Barabási and Albert (1999) have claimed that high skewness is a universal phenomenon characteristic of many social networks that follow power-law degree distribution, recent studies have found that such empirical networks are rare and that many cases fit traditional log-normal distributions (a bell-shaped distribution) better (Broido & Clauset, 2019). Since there is no clear consensus, we did not test a power-law degree distribution in this article.

Instead, we used skewness to measure varying degrees of participation. Skewness is a straightforward way to measure the extent to which values are equally distributed around the mean (Franceschet, 2011). If it is right-skewed, skewness then becomes positive, indicating many inactive participants, and vice versa. Therefore, a value of skewness close to zero indicates a higher level of equal commitment (data is symmetrical). According to these criteria, we found that residents were more unequally engaged in online deliberation (skewness: 7.3 to 10.3) despite the increasing volume of participation (average: 1.8 to 2.1) in the second round. In terms of proposals, we found that residents were more equally engaged in proposals (skewness: 9.6 to 4.5) with a similar number of comments (average: 2.3 to 2.2). Overall, we found that residents engaged in proposals more equally, yet the gap between inactive and active residents widened in the second round. Conversely, proposals received more equally distributed comments in the second round.

Figure 9.

Commitment.

Since Fig. 9 shows the varying degrees of commitment, it helps to introduce follow-up questions. For instance, one could explore the primary backgrounds of the inactive majority and how to promote active engagement in subsequent rounds. In addition, one could focus on why a few proposals attracted substantially more public attention or why some actors stridently voiced their opinions. These questions will contextualize the results, providing fruitful insights into promoting equal engagement.