Assessing Citizen Participation in Amsterdam Through Municipal Requests

Assessing Citizen Participation in Amsterdam Through Municipal

Requests

SUBMITTED IN PARTIAL FULLFILLMENT FOR THE DEGREE OF MASTER OF SCIENCE

M

ELISSA

R

OFMAN

11362456

M

ASTER

I

NFORMATION

S

TUDIES

H

UMAN-

C

ENTERED

M

ULTIMEDIA

F

ACULTY OF

S

CIENCE

U

NIVERSITY OF

A

MSTERDAM

Assessing Citizen Participation in Amsterdam Through

Municipal Requests

Melissa Rofman

University of Amsterdam,

Amsterdam, The Netherlands

meli.rofman@gmail.com

ABSTRACT

The development of Open Data initiatives has allowed for the surge of new means for communication between government and their citizens. These initiatives often lead to a vast amount of citizen generated data that can be used for new and innovative purposes. The development of such a system in the city of Amsterdam, in The Netherlands, allows its residents to directly communicate requests to the municipality authorities. This lead to the creation of a database (MORA) of 300 thousand records, over the course of approximately 3 years (2014-2017). The purpose of this thesis is to investigate whether the data in MORA can be used as a proxy for citizen participation, and what demographic characteristics of the residents affect the levels of participation. The results of the research will show that, with limitations, patterns of participation can be identified across the various areas of the city, and that there is a relationship between these patterns and voter turnout rates.

Author keywords

Open Data; Citizen Participation; Urban Computing

INTRODUCTION

Open data initiatives and the adoption of digital technologies by governments has allowed for new forms of communication to develop between the different government levels and their citizens [6]. In this regard, open data movements not only entail the opening of administrative data for public use, but also the creation of channels which the citizens can use as a mean to relay their thoughts and expectations regarding various public policies [20].

These open data platforms also provide the government new means for gathering information about the citizens or residents of an area: instead of having to wait for periodic population measurements (censuses, electoral votes, etc.) or having to resort to polling data, these platforms provide user generated data in a constant manner [22]. In this regard, they are very valuable sources of information for policy making and planning.

In this thesis, such a platform will be analyzed. The MORA database is the result of a citizen generated requests system, where issues reported by residents of the

city of Amsterdam have been registered from 2014 until 2017. This database will be used as input to consider two main research questions: is it possible to use a citizen

generated requests database as a means for assessing citizen participation? And if so, what demographics and other neighborhood characteristics influence variations in the levels of participation of the citizens throughout different administrative areas?

In the remainder of this thesis, related work on the topic will be presented. Thereafter, citizen participation will be defined both theoretically and operationally, in order to develop a quantitative analysis. Finally, the results for both research questions will be presented, along with the challenges that the research process entailed, and potential future lines of investigation.

RELATED WORKS

Urban computing for pattern detection

Urban computing is an interdisciplinary field that combines computational sciences with other areas of research, such as sociology, ecology and economy. It aims to generate insights into the workings of a city from the analysis of big and heterogeneous data generated by a diversity of sources in urban contexts, that can help improve the environment, the quality of life and the operational systems of a city. It also provides the means to analyze and understand urban phenomena as well as predict the future development of cities [25].

Aside from government generated datasets, one of the ways to collect the data used in urban computing is through data generated proactively by users. Human crowdsourcing, such as citizen requests and social media posts, when combined with geolocation can help bridge the gap between the physical world and online human activity [25]. Therefore, user generated data can be a good proxy for identifying patterns in citizen behavior and analyzing it based on geographic location.

Open data and citizen participation

Citizen participation, in the context of this thesis, will be understood as “attempts to influence the formulation

[and implementation] of public policy” [21] or, more broadly, as “the process by which members of a society (those not holding office or administrative positions in government) share power with public officials in making substantive decisions and in taking actions related to the community” [18]. In this regard, we will be focusing on direct citizen participation: “when the citizens are personally involved and actively engaged” [18].

This process can be described as governance, or the process through which the government-the authority to execute and implement activities-seeks to share their power in decision making, and promote citizen independence, providing the means for actively pursuing the common good through civic engagement [3].

This objective leads to initiatives such as the development of information technology based platforms that allow citizens to directly interact with their government. In these platforms, a multitude of participation and engagement objectives can be accomplished: reporting local issues affecting the citizens’ neighborhoods, communities, etcetera; local problem solving regarding service provision in the neighborhoods; redefining and amplifying the spaces for direct participation in policy decisions; increasing debate and involvement in legislation and regulation [2].

These types of initiatives take the name of Open Data. The most common form they adopt is the open publication of official information, generally through digital platforms. However, these initiatives can take multiple forms, depending on the breadth of government-population interaction that they permit.

In short, these initiatives can be characterized as data publishing, code exchange, civic issues tracker, and participatory Open Data [20]. Data publishing, which consists of governmental datasets made available for public download and use, represents the lowest level of government-population interaction. On the other extreme, participatory Open Data entails a change of paradigm in the way governments and population interact, in which they both abandon their positions as services providers and recipients, in order to form a “partnership that solves mutually identified problems” [20].

In this context, a municipal requests system such as MORA falls within the category of civic issues tracker, where the government is open to citizen generated contributions in a structured manner. These type of open data initiatives function like crowdsourcing systems: the government opens a two-way channel of communication, restricting the form the messages may take [20].

This level of interaction between government and citizens was made possible by the advent of the Web 2.0, which permitted the creation of e-government systems where citizens can “initiate the process of data creation, catalogue, and report issues within the community” [6]. The potential of adopting a role as producers of information, therefore, increases the level of communication between the population and the government, admitting higher levels of citizen involvement in local political processes [6, 9].

The success of open data applications such as MORA, a citizen generated requests system, lie on the number of citizens who use the application, the number of information feeds that those users generate, and how efficiently that information is communicated to government levels [7].

Citizen request systems

Citizen generated requests systems have existed for a number of years. The system that has most often been the subject of analysis is the ‘311 hotline’, which was originated in 1996 in Baltimore, Maryland with the goal of more closely connecting the government and the population. The ‘311 hotline’ not only provided a free telephone number that the residents could use to communicate their needs, but it also became a source of information for the authorities regarding government operations [24]. Since its first appearance, this system has expanded its reach within the United States, covering a significant number of cities and regions, and has also incorporated new channels of communication, such as online forms on websites and mobile applications.

The ‘311 hotline’ and other similar requests systems have been broadly investigated. Researchers have used public requests data as a means to assess the conditions of neighborhoods and civic response rates [14] as indicators for conflict [11], or as a measure of political engagement [12]. This type of data has also been used as input for visual analysis tools [10].

The data that can be obtained from requests such as the ‘311 system’ or MORA carry a series of advantages and disadvantages when used for assessing levels of participation [22].

First, the continuous occurrence of requests (and their subsequent storage in databases) is an advantage over other participation measures, such as voter turnout, which are periodical by nature.

Second, the voluntary nature of the calls shows that citizens must make the decision to reach out to government, as nothing is forcing them to do so.

Third, the fact that most of the requests are filed anonymously makes impossible to connect an individual to their filed requests. A research where individuals can be identified and their personal characteristics analyzed in relation to their propensity to participate [14] was not the goal of this thesis, due to both ethical and privacy issues. However, the lack of information in this regard poses two risk: the unidentifiable existence of super-users (single users who file a large number of requests), and a potentially uneven distribution of reportable issues through the neighbourhoods [22]. However, the neighborhood aggregated data still provides a suitable proxy for analyzing the levels of demand that different locations in the city have for the government [23].

METHODS

Measures and assumptions

The goal of this thesis is to discover whether it is possible to analyze participation levels and patterns in the different regions of Amsterdam with the MORA database. For this purpose, it was not possible to use the data in its original state, as it only showed the absolute number of requests per region, and not its relationship to the population and their levels of activity in each region.

In this regard, and keeping in mind that the requests refer to issues that occur on public spaces, it is logical to assume that any passerby could be their author. Ideally a participation measure would be the average number of requests per passerby per year. However, that requires knowing exactly how many people circulated through an area of the city during a certain period of time, information that was not available at the time of writing this thesis.

Instead an alternative indicator was designed. Starting from a simple assumption, that individuals who file requests in an administrative area tend to the be residents of that area, participation levels can be defined as the average number of requests per resident of a neighborhood or area per year.

This assumption is partially supported by the findings of [8], who find that residents of Amsterdam do not cover great distances while traveling on a day. Their study, concerning travel behavior and residential context, found that the residents of Amsterdam chose walking, cycling and public transform as their form of transport in 60% of all travels. Additionally, these trips generally cover short distances: in average, residents of Amsterdam covered 1.2 kilometers per day by walking, 2.7 kilometers cycling and 7.5 kilometers per day on public transport. Considering

that these are the aggregated averages for the whole day, it can be said that the general distance of each trip is, at the very least, small, further suggesting that residents of Amsterdam do not tend to stray far from their homes.

A more recent study by [19] who, through the use user mobility data from social media, managed to identify and differentiate areas of the city, supports the trends identified by [8]. Even though they did not focus on the social media users’ place of residence, their findings show that, normally, people circulate through certain areas of the city and not others. Therefore, it is likely that the areas most traveled by residents are those close to their home address.

However, for small areas, or areas with very few residents and/or high numbers of people circulating, the validity of this assumption becomes weaker. This thesis’ findings will show that as the geographic units of analysis increase in size, the consistency of the participation indicator will increase, but at the cost of lower specificity regarding the area characteristics’ influence.

A yearly measure of participation was adopted for this analysis based on the fact that population information in the BBGA database is registered and stored into annual records.

Finally, in order to unequivocally determine whether the measure used throughout this thesis to measure citizen participation is valid or not, it is interesting to compare its performance to that of more established and existent participation measurements, i.e. electoral turnout rates, or perception of community involvement. For this reason, data regarding turnout rates in the 2014 municipal elections was used to measure the reliability of MORA as a proxy for political participation. This data was provided by the Department of Information and Statistics of the Municipality of Amsterdam (OIS Amsterdam).

While it would be interesting to assess how well MORA performs as a proxy for other types of citizen participation as well, only the electoral data was available when this thesis was being written1.

Datasets and Preprocessing

The data in the MORA database is a collection of citizen requests filed in the city of Amsterdam between January 2014 and March 2017. The requests are submitted 1 _{It is important to note that the BBGA database (see section}

Datasets and Preprocessing) contains information regarding community involvement rates. However, the corresponding variables presented too much sparsity and were therefore not considered for this analysis.

through multiple channels, such as Twitter, email, phone calls, a form that can be completed in the municipality of Amsterdam’s webpage2, or a mobile application called Verbeter-de-buurt3.

The requests are organized into a single dataset and classified into one of the following six categories, depending on the topic they concern: waste; roads, traffic, street furniture; ambience in public space; public green and water; nuisance and hospitality; nuisance of animals; and others.

The database stores information about the requests, covering aspects such as the location of the issues being reported (Latitude and Longitude, and Administrative Neighborhood Code), a user generated description of the problem, a descriptive picture, the aforementioned main categories and multiple user generated subcategories, such as the date of the request, whether it has been attended to by the municipality, etcetera.

For this thesis, only the metadata (location, date of the request, and the main categories) was analyzed.

BBGA

The Basic Files for the Amsterdam Areas (Basisbestand Gebieden Amsterdam, or BBGA) [16] contain information about the most used levels of administrative areas of the city from 2005 until 2017. Each area is identified with a unique code, which is systematically used across databases. This thesis focuses on the areas of three out of the four most commonly used administrative levels4: 481 neighborhoods (“buurten” in Dutch), 99 neighborhood combinations (“buurtcombinaties”), 22 work-areas (“gebieden”) and 7 city districts (“stadsdelen”).

The database contains 660 indicators which are classified into thirteen themes: Population, Age, Living, Public space, Traffic, Viability, Safety, Activity, Sports and recreation, Welfare and care, Education, Income, and Participation.

BBGA is generated and maintained by the department of Research, Information and Statistics of the municipality of Amsterdam as part of their open data initiative. Each variable is measured yearly or once every two years.

2 _{Further detail available at}

https://www.amsterdam.nl/wonen-leefomgeving/melding-openbare/ (Date of consult: 15/04/2017)

3 _{Further detail available at}

https://www.verbeterdebuurt.nl/gemeente/amsterdam/ (Date of consult: 15/04/2017)

4 _{The districts were not included in the analysis due to the}

small number of occurrences.

Figure 1. Administrative area divisions. Top left: neighborhoods. Top right: Neighborhood-combinations.

Bottom left: administrative work-areas. Bottom right: districts [1].

Electoral turnout rates

The electoral turnout rates were provided by OIS Amsterdam. The data is organized by administrative area, specifically at the neighborhood-combination level. The rates are calculated by dividing the absolute number of votes cast in a certain booth by the the number of potential voters that reside close by. This data was only available for the 2014 municipal election, and for 93 of the neighborhood-combinations.

Administrative areas

The geographical information of the administrative

areas was obtained from the municipality’s Open Geo Data portal [15]. The portal provides a file for each administrative level, containing the coordinates for the geographical limits and the center of each area, the Administrative Area Code, and the name of the area.

Data reorganization

The datasets used in this thesis were not ready, in their original state, to be used for analyzing levels of participation in the city of Amsterdam. The most common issues were data sparsity and errors in the values for some variables. In this section, a detailed description of how the data was cleaned and adapted for this research can be found.

Requests

Adapting MORA for analysis entailed three steps: removing test cases from the dataset, dealing with sparse data, and aggregating the requests according to the administrative area they occurred in.

First, a low number of the requests registered in MORA did not correspond to citizen filed requests, but were instead generated by testing procedures carried out by the municipality. These were identified in MORA as such, and were excluded from any analysis.

Second, not all the requests in MORA had complete information regarding their geolocation. An important number of cases had no latitude and/or longitude registered, making it impossible to pin-point the exact location of the issue being reported. These were excluded from the analysis.

After these two steps were taken, the total number of registered requests in MORA fell from 372,355 to 333,429.

Additionally, problems were found with the geolocation of some requests: the numbers indicating the latitude and longitude had a wrongly placed decimal point. For example, instead of placing a request at approximately 52o Latitude, and 4.9o Longitude (the general location of Amsterdam), they placed it at 5200000o _{Latitude, and} 490000o _{Longitude. These cases were fixed by the} researcher.

Once MORA had been cleaned, a GeoJson file per year was generated, containing information about the location and type of each request.

Third, as the participation levels analysis would take place at a neighborhood level, the data had to be aggregated into a single case per neighborhood, for each administrative level. This meant generating a new dataset for each administrative level, where each recorded observation corresponded to a particular neighborhood. Through the use of the administrative codes for the neighborhoods, the cases corresponding to each area were identified and grouped. The resulting databases contained information about the total number of requests filed per year, the total of requests per category, the most popular request, and the administrative code for each neighborhood.

Finally, due to the yearly nature of the participation measure generated for this thesis, the year 2017 was not considered in the analysis, since the data for that period was not complete. Therefore, the following results focus on the requests filed between January 2014 and December 2016.

Demographic information

As mentioned previously, the participation indicator was defined as the the average number of requests per

resident per year. Therefore, the total number of requests

in an area had to be divided by its total number or residents.

In order to do this, the cleaned up and aggregated MORA files and BBGA were combined, using the neighborhood administrative codes as the means for matching both sources, for each administrative level and year.

However, the BBGA database also presented problems regarding data sparsity. For each administrative level, several variables with missing data were found.

For the analysis, only the variables that had less than 10% missing observations were taken into consideration. These missing observations were filled with the mean of each variable for each of the years.

RESULTS

Participation patterns

Representing data in a visual form allows researchers to identify patters and regularities, as well as the outliers in data [10]. When it is presented through spatial or graphical representations, it simplifies comparison, facilitates change detection in the data and its interactive qualities allow users to manipulate the visualization according to their objectives [13].

For this reason, the initial approach to finding patterns in the data was through a geo-visualization of MORA requests (Figure 3). The map shows the participation levels for each area for all the administrative levels (choropleth map) and the total number and most frequent type of requests for each neighborhood (colored markers). This information was displayed separately for 2014, 2015, and 2016, allowing for a geo-temporal analysis of the data.

The administrative neighborhoods boundaries are drawn with white dotted lines, and each neighborhood is colored with a low opacity blue, corresponding to the choropleth scale: the lighter the blue, the lower the level of participation for the residents of that region. Neighborhoods with no registered residents are colored with black, and neighborhoods with participation levels of 1 or higher are colored with purple.

The colored circular markers are styled according to the number of requests (size) they represent and the most frequent type of request (color) for that neighborhood. When hovering over a neighborhood, the map displays its name, its total population and the total number of requests filed in the area.

The map only displays information for a single year, and a single administrative level at a time. To switch between years, a toggle menu was created, which allows the user to travel between the data for 2014, 2015 and 2016. Finally, with the zoom, it is possible to either aggregate or divide the data into higher or lower levels of administrative areas, i.e. from neighborhoods to neighborhood-combinations, and back (Figure 3).

Temporal analysis

Over the course of the three analyzed years, the levels of participation show a clear tendency to rise, regardless of the level of administrative area that is being considered. At the administrative work-areas level, all the sectors of the city showed an increment in the levels of participation of their residents (Figure 2). However, the neighborhood and neighborhood-combination levels show more variation than the administrative work-areas. In the smaller administrative area levels, it is not only possible to find areas that increased their levels of participation, but also sectors where the average number of requests per resident fell from 2014 to 2016. However, due to the far superior number of neighborhoods where increased participation levels were registered for this period, it is still possible to state that the general trend is for an increment in the levels of participation.

Figure 2. Variation in levels of participation between 2014 and 2016, at the administrative work-area level. Percentages.

It is also interesting to point out that the levels of participation did not grow evenly across all neighborhoods, nor did they evolve with the same intensity for the periods 2014-2015, and 2015-2016. The levels of participation increased a significantly higher rate during the second period. Both the neighborhood combinations and work-areas showed similar growth rates: 8.5% and 8.3% respectively for 2014-2015, and 32.7% and 33.3%, respectively for 2015-2016. For the neighborhood level, the difference between the rates of

growth for both periods was lower (2015-2016 showed growth at 2.5 times the rate of 2014-2015), but the growth for each period separately surpassed that of the higher levels (18.6% for 2014-2015, and 45.6% for 2015-2016) (see appendix for detailed graphs and visualizations).

Geographic analysis Neighborhoods

Figure 3. Participation levels for 2014 (top), 2015 (bottom-left) and 2016 (bottom-right) for the neighborhood level (see

appendix for more details).

The map provided some insights into MORA (Figure 3). The requests in the city of Amsterdam are mostly related to garbage issues, and the predominance of garbage as a topic has grown over the course of the three analyzed years.

The neighborhoods in city center present some of the highest levels of activity in terms of total number of requests, but do not represent the areas with the highest levels of participation. Instead, higher participation levels seem to occur in neighborhoods located near the outer edges of the city, even though the total number of requests filed in those areas is lower.

The map immediately shows that a small number of neighborhoods have no residents (approximately 20 neighborhoods in each of the observed years. For a detailed list, see the appendix). These correspond to parks (Beatrixpark, Vondelpark West, Sloterpark), industrial areas (Sloterdijk III Oost) or islands (Buiteneiland). The map also shows that while there is no population living there, they are still the location of a significant number of requests. For example, Westergasfabriek, a popular site for public events and festivals, was the site of 162 requests in 2014, 152 in 2015, and 194 in 2016. However,

no residents were registered there for any of the three years in question.

Additionally, it is possible to identify a significant number of neighborhoods where the participation level is exceptionally high: the purple colored areas. These neighborhoods seem to be areas that have either high numbers of people circulating through them (Stationplein and Zuiderhof), or extremely low population numbers (Vondelpark Oost, Amstelpark, Museumplein, Sloterdijk II).

Upon closer inspection, a statistical analysis of the data indicates that all the neighborhoods that have participation rates of 1 or higher, are also those with the lowest numbers of registered residents (lowest 15th percentile in the distribution, according to number of residents-Table 1,

Table 2). Additionally, Figure 4 shows that the neighborhoods with the lowest numbers of residents skew the data distribution significantly.

Table 1. Number of residents in Amsterdam neighborhoods where level of participation is equal to or higher than 1

request per citizen in average per year. All years.

Population of the neighborhoods with more than 1 request per resident– All years

Count 106 Mean 16.59 Standard deviation 17.83 Minimum 1 25% 5 50% 10 75% 22 Maximum 87

Table 2. Number of residents in all Amsterdam Neighborhoods, by percentile. All years.

Population of the neighborhoods– All years

Percentile Number of residents

0 1 10 39 15 114 20 269 25 423 30 725 40 1141 50 1554 60 2096 70 2524 80 3025 90 3755 95 4313 100 6175

Number of residents in the neighborhood

Figure 4. Scatterplot displaying the distribution of the neighborhoods according to the participation level and the

number of residents. All neighborhoods, for all years.

The neighborhoods with exceptionally low numbers of residents and/or located in highly circulated areas pose a challenge for analyzing citizen levels of participation, as it is less reliable to assume that the people filing requests at those locations are residents in those specific areas. For example, in the Square in front of the Central Station in the city center of Amsterdam, 57 requests were filed in 2014, and only one person was registered as living there for that year. It is highly unlikely that a single person filed those requests, when none of the neighborhoods surrounding it surpass a single request in average per resident. In other words, the assumption proposed earlier (that individuals who file requests in an administrative area tend to be residents of that area) does not seem to hold when the area considered have very little or no registered population.,

Therefore, both the exceptional neighborhoods and the resident-less neighborhoods were excluded from the analysis concerning citizen participation levels. Instead, a separate analysis was conducted to identify which factors can best explain the difference between these ‘atypical’ neighborhoods, and the typically populated ones.

Regarding variations in participation levels, the typical neighborhoods showed smaller but relevant changes. The lowest 75% of the neighborhoods have levels of participation between 0.01 and 0.1 per year, showing a tenfold increase between the least active neighborhood and the 75th percentile. Furthermore, the neighborhoods in the highest quartile of the distribution show an even sharper increase in participation levels, climbing up to a maximum of 0.9 per year.

Table 3. Distribution of the levels of participation for the neighborhoods that have participation rates below and up to

1 request per resident in average per year. All years.

Distribution of the levels of participation of all the ‘typical’ neighborhoods

Count 1178 Mean 0.099 Standard deviation 0.097 Minimum 0.008 25% 0.051 50% 0.072 75% 0.105 Maximum 0.92

Out of these neighborhoods, it is interesting to point out that the most active ones correspond to neighborhoods in city center (Kalverdriehoek, Rembrandtpleinbuurt, Lastage) and in the northern district (Papaverweg e.o.). The neighborhoods with the quietest levels of participation are mostly located in the southeastern area and eastern islands of the city (for example, Middeneiland Zuidwest. Nelson Mandelapark, Sportpark Voorland).

Atypical neighborhoods

Even though the atypical neighborhoods are excluded from the participation level analysis, it is interesting to identify some common characteristics that might explain the number of filed requests in these areas.

In order to do this, the neighborhoods were classified into two categories: ‘atypical’ (1 or more requests filed on average per resident per year), and ‘typical’ (less than 1 request filed on average per resident per year). Based on this classification, two methods were used to identify the factors that separate one group from the other, and might be related to the variation in number of requests per citizen: a simple comparison of the average of the variables for both groups, and a the Scikit-Learn [17] Python library implementation of a Random Forest classifier [5].

These approaches present advantages and disadvantages: comparing the averages of the variables entails summarizing what could be high variation rates inside a variable into a single number. This can lead to overlooking subtleties, as well as having a skewed mean that does not represent the general distribution of the neighborhoods in each category. However, it also allows for a simple and transparent comparison of the variables that characterize each group. This approach also shows which group of neighborhoods has higher numbers for a certain variable, as well as how steep the difference is.

The Random Forest classifier, on the other hand, is a machine learning method that is specifically created to find the divisive factors between categories. Therefore, it is a more reliable method for assessing the most important factors when it comes to predicting which neighborhoods could fall in the ‘atypical’ category. However, it does not explicitly show how each variable was used to divide the data. Finally, due to the skewed nature of the data used for this analysis (only 15% of the cases belong to the ‘atypical’ group), and the scarcity of training data, the model can be prone to overfitting. Therefore, a combination of both strategies can lead to better insights.

Figure 5. Difference in the (average of the) characteristics of the ‘atypical’ and ‘typical’ neighborhoods, in percentage of

variation. Top 20 variables, for all years.

According to the first method (Figure 5), the neighborhood characteristic that differs the most between the two groups of neighborhoods is the number of employers without employees: the ‘atypical’ group has a significantly lower number of them. On the other hand, these neighborhoods have higher numbers of students in mandatory education who have exceeded the expected number of studying years, houses with no specified surface size, and houses rented from particulars, leading to no conclusive insights at the moment of writing this thesis.

Secondly, for the Random Forest classifier, the data was divided into training and test data following an 80/20 split. The target classes were two: ‘typical’ and ‘atypical’ neighborhoods.

The model, then, was trained only on 80% of the data, using five-fold cross validation. When applied to the test set, the Random Forest model correctly classified 96.9%. of the test instances. It also yielded information regarding

the most important features for classifying the neighborhoods. In this regard, the five most important features for predicting which neighborhoods would fall in the ‘atypical’ category, and which ones would not, were the size of the potential labor force of a neighborhood (residents with ages between 15-64 years of age, and 15 and 74), the number of autochthonous residents (persons with both parents from Dutch origin, regardless of their own country of birth), the total population of the neighborhoods, and the number of residents between 18 and 54 years of age who are either from western origin or born to Dutch parents, and registered as residents of Amsterdam after turning 18 years of age.

Neighborhood-combinations

A potential explanation as to why the model for assessing the levels of participation does not perform in a reliable manner for 15% of the cases at the neighborhood level is that, for the exceptional cases, the neighborhood size is too small. In other words, assuming that the people who file requests in an area are the people that reside in that area, could be a complicated assumption to make for city divisions as small as the neighborhoods. If this is the case, aggregating the data into a higher level of administrative areas (i.e. neighborhood-combinations) should improve the reliability of the results.

Effectively so, only one neighborhood-combination shows levels of participation that surpass the average of one request per resident per year: IJburg Oost. Additionally, the map shows a more homogeneous distribution of participation levels throughout the neighborhood-combinations.

A more detailed look at the statistical distribution of the neighborhood-combinations based on participation levels confirms this increased homogeneity: only 1% of the neighborhood-combinations can be characterized as ‘atypical’.

These three instances, along with three others which had no registered residents, were excluded from further participation level analysis.

Finally, for all three years, the two most active neighborhoods are Bedrijventerrein Sloterdijk and Noordelijke IJ-oevers Oost, which show opposing trends: while Bedrijventerrein Sloterdijk’s participation levels more than doubled between 2014 and 2016, Noordelijke IJ-oevers Oost is slowly but steadily showing a decline in the participation numbers, leading to an increased distance between the rates of the Bedrijventerrein Sloterdijk and all the other neighborhoods. These neighborhoods mostly filed requests concerning garbage

and roads, traffic and street furniture. Only in 2014 another category of requests was featured in the top ten most active neighborhoods: Nuisance Companies and Hospitality.

The least active neighborhoods, however, show slightly more variation. Osdorp Midden went from being the second quietest neighborhood in 2014 to the fourth in 2016; Bijlmer-Centrum (D, F, H) is one of the two quietest neighborhoods for all three years, and Ijburg Zuid and West systematically appear in the five least active neighborhoods. All of these neighborhoods have garbage as their most common request.

Administrative work-areas

The visualization of the 23 administrative work-areas shows a different scenario from the lower two administrative levels (see Appendix for the visualizations).

First of all, while there is still variation across the administrative work-areas, all of them have participation levels below 1. The most active area is Westelijk Havengebied, which has a very low number of residents in comparison to the other administrative work-areas (approximately 200 residents each year): the second smallest administrative work-areas in terms of population is near 20 times bigger (between 21 and 22 thousand residents each year).

Secondly, three administrative work-areas appear in the top five for all the years: Centrum Oost, Centrum West and Oud Noord.

The administrative work-areas with the lowest rates of participation, on the other hand, are mostly located on the eastern and southeastern parts of the city. While, both 2014 and 2015 show the same ranking of administrative work-areas (Bijlmer Oost, Bijlmer Centrum, IJburg/Zeeburgereiland, Gaasperdam/Driemond, and Bos en Lommer), the list varies slightly for 2016, when Bos en Lommer is replaced by De Aker, Sloten en Nieuw Sloten

In this level, it is clear that the administrative work-areas with the highest levels of participation are located in the most central areas of the city, while the quiet administrative work-areas are located on the city limits.

Neighborhood characteristics

All the neighborhood views in the map clearly showed that the levels of participation vary across the different areas of the city for all administrative levels. The population characteristics that show significant relationships to the participation rates, however, differ between the levels.

In order to analyze the relationship between the levels of participation and the demographic characteristics of the areas of each administrative level, two strategies were adopted: a linear correlation analysis, and a Random Forest forest classifier. The linear correlation analysis was performed between each neighborhood characteristic and the participation levels for each neighborhood, for all years.

The Random Forest model was used to assess the most relevant features for classifying according to their participation levels. In order to do this, the areas had to be labeled as ‘High Activity’, ‘Low Activity’ and ‘Neutral Activity’. Due to the arbitrary nature of this classification (no objective measure of ‘high’ or ‘low’ levels of participation exist for this analysis), three different models with their respective thresholds were implemented:

1. Tertiles: the 33% with the lowest levels of activity were classified as ‘Low Participation’, the neighborhoods in the top 33% levels of activity were classified as ‘High Participation’, and the remaining 33% were classified as ‘Neutral’

2. Quartiles: the 25% with the lowest levels of activity were classified as ‘Low Participation’, the neighborhoods in the top 25% levels of activity were classified as ‘High Participation’, and the remaining 50% were classified as ‘Neutral’

3. Halves: the neighborhoods were divided into only two equal categories (‘Low Participation’ and ‘High Participation’). The median of the distribution functioned as the threshold for the split.

For these analysis, the temporal dimension of the participation levels was disregarded: all the variables considered were scaled per year, to fit between a range of 0 and 1. Additionally, all the atypical neighborhoods were disregarded and the statistical outliers (all areas with participation levels above or below two standard deviations from the mean) were excluded.

Finally, the neighborhood and population characteristics were obtained from the BBGA database, which presented an important degree of data sparsity. As was mentioned previously, only the variables that had 10% or less missing observations were considered. This means that 517, 572 and 610 variables were considered for the neighborhood, neighborhood-combination and administrative work-areas levels analysis, respectively.

In the following sections, the results for these two approaches are presented, organized according to neighborhood level.

Neighborhoods

Correlations

At the neighborhood level, the results show that the correlations are statistically significant, but have weak to moderate strength (see Appendix).

The correlation coefficients show that while none of the analyzed variables can explain all of the variation in the participation levels, two main groups of variables can be identified within the top 15 correlations.

The five most significant positive correlation coefficients correspond to variables that are related to the number of services (catering locations, hotels, cafes, tourism) available in the neighborhood. As was mentioned previously, in these neighborhoods the requests are mostly related to garbage and roads, traffic and street furniture, and (only for 2014) nuisance companies and hospitality.

Inversely, the strongest negative correlations are mostly related to family composition characteristics: the presence of kids between the ages of 0 and 19 years of age, as well as the presence of families with children. Additionally, the share of non-western immigrants is also negatively correlated with the rates of participation, as well as the non-western population between the ages of 0 and 17.

Random Forest classifier

Another approach to identifying which neighborhood characteristics are related to the residents’ level of participation is classifying the neighborhoods into categories based on activity levels, as described before.

As with the ‘typical’ and ‘atypical’ neighborhoods, the data for the Random Forest classifier model was split into training (80% of the data) and test (20% of the data) sets, randomly, for each of the models (Tertiles, Quartiles and Halves). The Random Forests were trained on the first 80% of the data, using five-fold cross validation, and 500 tree estimators. The resulting models were evaluated against the test sets, reaching 69.1%, 80% and 78.9% accuracy for the Tertiles, Quartiles and Halves respectively.

The most important features mostly coincide: All three classifications have the number of first generation non-western residents, the number of residents of Surinamese origin, and the share of residents from non-western origins, within the five most important features for classifying the neighborhood participation rates.

Correlations

The correlation coefficients at the neighborhood-combination level show a predominance of the influence of neighborhood services. The 15 variables with the strongest relationships to the levels of participation of the neighborhood-combinations are related to tourism industry related services (Hotels, Restaurants, Tourism), and the number of employees in the workplaces.

These correlations, while stronger than at the neighborhoods level, are still only moderately strong (between 0.46 and 0.61 coefficients).

Random Forest classifier

The Random Forest classifier models, however, show a completely different story. With accuracy levels of 67.8%, 72.9% and 86.4% for the Tertiles, Quartiles and Halves, respectively, they show that the most influential variables for classifying the neighborhoods are consistently related to the origin of the population younger than 17 years old.

Administrative work-areas

Correlations

Finally, the administrative work-areas level analysis shows the strongest relationships between the neighborhood demographics and the levels of participation of the population.

At this administrative level, the top two variables correspond to workplace characteristics, i.e. the higher the number of places with between 5 and 19 employees, the higher the participation levels. Twelve of the top 15 variables correspond to neighborhood services (restaurants, tourism locations, hotel bed, etcetera). All of these variables are positively correlated with the participation levels of the neighborhoods.

Finally, it is interesting to point out that out of the top 15 variables, the only one that shows a negative relationship with the levels of participation is the average size of the household: the higher the average number of residents in the neighborhood households, the lower the participation rates.

Random Forest classifier

Due to the small number of occurrences at the administrative work-areas level (23 per year, leading to a total of 66 instances for analysis), a Random Forest forest model was not implemented for this administrative level.

Electoral turnout

The analysis of whether MORA can properly function as a proxy for electoral turnout rates has an important limitation: the data that was used for the analysis corresponded only to the year 2014. Therefore, the number of instances that were considered are limited.

The results of the analysis show that there is a significant correlation (p-value of 0.00007) between the participation levels and the electoral turnout rates, but that the relationship between both variables is moderate at best (correlation coefficient of approximately 0.41).

Therefore, it is possible to say that, while the participation levels obtained from MORA do not explain all the variation in the electoral turnout rates, they can potentially function as a proxy for years in which there are no electoral processes. It will be up to future researchers to determine the limitations of this possibility.

FUTURE WORK

This thesis is a first step into researching the use of the MORA database as a measure of participation level. For now, the analysis has been limited to data that was not created or stored with this specific purpose in mind. Therefore, an interesting scenario would emerge if in the future other types of data were stored with MORA analysis possibilities in mind.

For example, it would be interesting to analyze the relationship between public transport routes and participation patterns. It would also be interesting to consider external factors that might be influencing participation levels, such as public’s knowledge of the system. If propaganda campaigns were presented in certain neighborhoods and not in others, or at certain times, they could be influencing the levels of participation. Additionally, an interesting angle would also be to track the resolution of the issues. If the system functions to its full potential, it might be that the requests disappear as time goes by, not due to a drop in participation levels, but to an increased efficiency in governmental policy making and planning.

Furthermore, it would be interesting to see what conclusions Sociologists and city administrators might gather from this data. From an information science perspective, a few lines in terms of interpretation have been suggested. Still, further studying the capabilities of MORA from a sociological level could also provide new grounds for further research in the information science field.

Third, designing and implementing a system that would aid policy makers in their planning and decision making processes would be extremely interesting. The data in MORA, if correctly visualized and presented, can give access to a live feed into the residents’ concerns regarding their surroundings. For this reason, designing and developing a system where MORA is used as daily input for policy making and fast problem solving by the government would be a perfect fit for further Human Centered Multimedia studies.

Fourth, most of the Open Data initiatives that allow citizens to communicate with their government use (or are including) digital means of communication (email, websites with forms, mobile applications). In this regard, it would be interesting to research how digital literacy and the digital gap affect the possibilities the users have to access the system, and therefore the levels of participation that can be deduced from it.

Finally, working with data stored over a longer period of time would prove insightful. Considering the data sparsity issues, more instances of data available for analysis, and more complete and correct data, would prove a valuable source of information.

DISCUSSION AND CONCLUSION

One of the main challenges of this thesis was the need to anchor population information to the geographic location where events occurred. The format used by the MORA database, made impossible to connect a request with its author. While this can be a desirable quality with regards to privacy and identity protection of the residents of the city, it also generates a few challenges in an analysis that involves population characteristics. For this reason, one of the weakest points of this thesis is also one of its central arguments: the assumption that residents file requests in the areas surrounding their place of residence. Its risks are clear throughout the results: 15% of the neighborhoods and 1% of the neighborhood-combinations have to be excluded from analysis for the assumption to hold up, and there is no specific evidence supporting this statement for the smallest of administrative levels. However, the analysis becomes more robust (but less sensitive to region variation) as the administrative levels grow in size, and the related literature shows evidence supporting this assumption for higher levels of aggregation.

From a different perspective, one of the main challenges of this research was dealing and overcoming issues regarding the raw state of the data: all the databases

that were used for the analyses presented data sparsity and inconsistencies that had to be dealt with. The alternatives were to disregard all observations with incomplete information (but this would have resulted in a more serious data sparsity issue) or to define a threshold to accept incomplete data and use some reasonable assumption to complete it. This second option was preferred for this analysis, with the unavoidable consequences in terms of consistency.

Finally, only simple linear correlations were computed for the correlation analysis. In this regard, it might have been interesting to utilize multiple correlations or other more complex approaches. However, for the purposes of this thesis, we believe that the analysis is sufficient to conclude that the MORA database can be used for assessing citizen participation, and for identifying the most influential characteristics that drive the participation levels.

REFERENCES

[1] Amsterdam, City of. 2017. Administrative Areas. 04 15. http://maps.amsterdam.nl/gebiedsindeling/.

[2] Bertot, J. C., P. T. Jaeger, S. Munson, and T Glaisyer. 2010. "Engaging the Public in Open Government: Social Media Technology and Policy for Government Transparency." IEEE Xplore DOI: http://doi.org/10.1109/MC.2010.325.

[3] Bingham, L. B., T. Nabatchi, and R. O'Leary. 2005. "The New Governance: Practices and Processes for Stakeholder and Citizen Participation in the Work of Government." Public

Administration Review 547-558 DOI:

http://dx.doi.org/10.1111/j.1540-6210.2005.00482.x.

[4] Boonzajer Flaes, Joost, Stevan Rudinac, and Marcel Worring. 2016. "What Multimedia Sentiment Analysis Says About City Liveability." Lecture Notes in Computer Science 824-829.

[5] Breiman, Leo. 2001. "Random Forests." Machine Learning

(Springer) 45 (1): 5-32 DOI:

https://doi.org/10.1023/A:1010933404324.

[6] Cavallo, Sara, Joann Lynch, and Peter Scull. 2014. "The digital divide in citizen-initiated government contacts: A GIS approach." Journal of Urban Technology 77-93.

[7] Desouza, Kevin C, and Akshay Bhagwatwar. 2012. "Citizen apps to solve complex urban problems." Journal of Urban Technology 107-136.

[8] Dieleman, Frans M., Martin Dijst, and Guillaume Burghouwt. 2002. "Urban form and travel behaviour: micro-level household attributes and residential context." Urban Studies 507-527.

[9] Hampton, Keith, and Barry Wellman. 2003. "Neighboring in Netville: How the Internet supports community and social capital in a wired suburb." City & Community 277-311. [10] Hubert, Rocio B, A. G. Maguitman, Carlos I. Chesñevar, and Marcos A. Malamud. 2017. "CitymisVis: a Tool for the Visual Analysis and Exploration of Citizen Requests and

Complaints." Proceedings of the 10th International Conference on Theory and Practice of Electronic Governance (ACM) 22-25. [11] Lacoe, Johanna, and Ingrid Gould Ellen. 2012. "Mortgage Foreclosures and the Shifting Context of Crime in Micro-Neighborhoods."

[12] Levine, Jeremy R., and Carl Gershenson. 2014. "From political to material inequality: Race, immigration, and requests for public goods." Sociological Forum 29 (3).

[13] Mazza, Ricardo. 2009. "Introduction to Information Visualization." Springer Science & Business Media.

[14] O'Brien, Daniel Tumminelli, Robert J. Sampson, and Christopher Winship. 2015. "Ecometrics in the age of big data: Measuring and assessing "broken windows" using large-scale administrative recoreds"." Soicological Methodology (45.1): 101-147.

[15] Onderzoek Onderzoek, Informatie en Statistiek Amsterdam. 2017. Open GeoData Portal. 04 15. http://maps.amsterdam.nl/open_geodata/.

[16] Onderzoek, Informatie en Statistiek Amsterdam. 2017. Basisbestand Gebieden Amsterdam (BBGA). 04 15. https://www.ois.amsterdam.nl/online-producten/basisbestand-gebieden-amsterdam.

[17] Pedregosa, F., G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, et al. 2011. "Scikit-learn: Machine Learning in Python." Journal of Machine Learning Research 2825-2830.

[18] Roberts, Nancy. 2004. "Public Deliberation in an Age of Direct Citizen Participation." The American Review of Public Administration (SAGE Publications) 34 (3): 315-353 DOI: https://doi.org/10.1177/0275074004269288.

[19] Rudinac, Stevan, Jan Zahálka, and Marcel Worring. 2017. "Discovering Geographic Regions in the City Using Social Multimedia and Open Data." nternational Conference on Multimedia Modeling (Springer).

[20] Sieber, Renee E., and Peter A. Johnson. 2015. "Civic open data at a crossroads: Dominant models and current challenges." Government information quarterly 308-315.

[21] Whitaker, Gordon P. 1980. "Coproduction: Citizen participation in service delivery." Public administration review 240-246.

[22] White, Ariel, and Kris-Stella Trum. 2016. "The promises and pitfalls of 311 data." Urban Affairs Review.

[23] Wiseman, Jane. 2015. "Innovations in Public Service Delivery: Issue No 01: Can 311 Call Centers Improve Service Delivery? Lessons from New York and Chicago." (Inter-American Development Bank).

[24] Wood, Colin. 2016. "What is 311." Goverment Tech, August 4.

[25] Zheng, Y, L Capra, O Wolfson, and H Yang. 2014. "Urban computing: concepts, methodologies, and applications." Transactions on Intelligent Systems and Technologies (ACM).

Appendix

Participation levels, type and number of requests for 2014 for the Neighborhood level

Participation levels, type and number of requests for 2016 for the Neighborhood level

Participation levels, type and number of requests for 2015 for the Neighborhood-combination level

Participation levels, type and number of requests for 2014 for the Administrative work-areas level

Variation in participation levels for the Neighborhood level, between 2014 and 2015. Percentages.

Variation in participation levels for the Neighborhood level, between 2014 and 2016. Percentages.

Variation in participation levels for the Neighborhood-combination level, between 2014 and 2015.

Percentages.

Variation in participation levels for the Neighborhood-combination level, between 2015 and 2016.

Percentages.

Variation in participation levels for the Neighborhood level-combination, between 2014 and 2016.

Percentages.

Variation in participation levels for the Administrative work-areas level, between 2014 and 2015.

Percentages.

Variation in participation levels for the Administrative work-areas level, between 2015 and 2016.

Percentages.

Variation in participation levels for the Administrative work-areas level, between 2014 and 2016.

Percentages.

Variation in participation levels for the Neighborhood level, between 2014 and 2016. Bar chart,

Percentages.

Neighborhoods of Amsterdam

Variation in participation levels for the Neighborhood combination level, between 2014 and 2016.

Bar chart, Percentages.

Correlation results for top 15 variables for Neighborhoods level

Variable name Variable explanation correlation pValue

BHVEST_HO Hotel locations 0.4334827 7.83E-55

BHWP_TOER Tourism Employment: Number of people employed (at least 12 hours per week) in the tourism sector. 0.420590618 2.12E-51 BHVEST_5TM9WP Number of establishments with 5 to 9 employees with a min. of 12 hours work per week. 0.413441139 1.46E-49

BHVESTI Catering locations 0.406656043 7.42E-48

BHVEST_HORECA Absolute number of hospitality locations (hotels, restaurants, cafes, etc.) 0.406575833 7.77E-48 BEVNW_P Share of residents who are native of non-western countries (Morocco, Turkey, Surinam, Antilles, countries in Africa, Latin America, and Asia – excluding Indonesia and Japan) -0.3874501 3.08E-43 BEV15_19 Absolute number of residents between the ages of 15 and 19 -0.386469275 5.20E-43 BEV0_18 Absolute number of residents between the ages of 0 and 18 -0.371729841 1.11E-39 BEVHHMKIND Absolute number of households with children -0.371172826 1.47E-39

BHVEST_RE Restaurant locations 0.370918594 1.67E-39

BEV17NW_P Share of residents from ages 0 to 17 from non-western origins -0.368872163 4.69E-39 BEV0_17 Absolute number of residents between the ages of 0 and 17 -0.367381294 9.90E-39 BEVEENOUDERHH Absolute number of households with single parents. -0.366127111 1.85E-38

BHVEST_CA Cafe locations 0.365142596 3.01E-38

Correlation results for top 15 variables for Neighborhood-combination level

Variable name Variable explanation correlation pValue

BHWP_HORECA Hospitality Employment: Number of people employed (at least 12 hours per week) in the hospitality sector. 0.611952093 2.30E-30 BHWPI Catering Employment: Number of people (at least 12 hours per week) in the catering sector. 0.605961324 1.18E-29 BHWP_RE Restaurant employment: number of people employed in the restaurant sector. 0.591862766 4.82E-28 BHWP_TOER Tourism Employment: Number of people employed (at least 12 hours per week) in the tourism sector. 0.565090005 3.42E-25

BHVEST_HO Hotel locations 0.564119588 4.29E-25

BHHOTBED Number of hotel beds 0.547431836 1.89E-23

employees with a min. of 12 hours work per week.

BHWINK_1000INW Shops per 1000 residents 0.536538794 2.01E-22

BHVEST_CA Cafe locations 0.532119664 5.10E-22

BHWPI_P Catering Employment: Share of people (at least 12 hours per week) in the catering sector.

0.531919393 5.32E-22

BHHOTKAM Number of hotel rooms 0.530614751 6.99E-22

BHVEST_10TM19WP Number of establishments with 10 to 19 employees with a min of 12 hours work per week. 0.521580819 4.48E-21 BHWP_CA Cafe Employment: Number of people employed (at least 12 hours per week) in the cafe sector. 0.520871435 5.17E-21

BHHOT Number of hotels 0.520541125 5.53E-21

Correlation results for top 15 variables for Administrative work-areas level

Variable name Variable explanation correlation pValue

BHVEST_5TM9WP Number of establishments with 5 to 9 employees with a min. of 12 hours work per week. 0.786266913 5.20E-15 BHVEST_10TM19WP Number of establishments with 10 to 19 employees with a min. of 12 hours work per week. 0.77952495 1.25E-14

BHWINKDG_1000INW Number of food stores per 1000 inhabitants 0.770690306 3.80E-14

BHWP_TOER Tourism Employment: Number of people employed (at least 12 hours per week) in the tourism sector. 0.760493004 1.29E-13 BHWP_RE Restaurant employment: number of people employed in the restaurant sector. 0.749042596 4.73E-13 BHVEST_2TM4WP Number of establishments with 2 to 4 employees with a min. of 12 hours work per week. 0.741020841 1.13E-12

BHVESTI_P Share of catering locations 0.737781846 1.59E-12

BHVESTI Number of catering locations 0.731460002 3.05E-12

BHVEST_HORECA Absolute number of hospitality locations (hotels, restaurants, cafes, etc.) 0.731386386 3.08E-12 BHVEST_TOER Absolute number of tourism locations (accommodation and lodging, other restaurants, recreation, marinas, sailing schools) 0.728392882 4.17E-12

BHVEST_RE Absolute number of restaurant locations 0.727545272 4.54E-12

BHVEST_CA Absolute number of cafe locations 0.724289712 6.27E-12

WBEZET Average household house -0.723313829 6.91E-12