=Paper=
{{Paper
|id=Vol-2103/paper_6
|storemode=property
|title=The Politics and Biases of the "Crime Anticipation
System" of the Dutch Police
|pdfUrl=https://ceur-ws.org/Vol-2103/paper_6.pdf
|volume=Vol-2103
|authors=Serena Oosterloo,Gerwin van Schie
}}
==The Politics and Biases of the "Crime Anticipation
System" of the Dutch Police==
https://ceur-ws.org/Vol-2103/paper_6.pdf
The Politics and Biases of the “Crime Anticipation
System” of the Dutch Police
Serena Oosterloo1,2 and Gerwin van Schie1,2
1 Utrecht Data School, Drift 13, 3512 BR, Utrecht, The Netherlands
2 Utrecht University, Drift 15, 3512 BR, Utrecht, The Netherlands
Abstract. In line with developments in many areas of business and gov-
ernance, where bureaucracies of all sorts are increasingly datafied for
budgetary reasons and the additional possibilities for automated analysis,
the Dutch Police started with so-called Intelligence-Led Policing. This
development led to the creation of the Crime Anticipation System (CAS).
This data-driven system tries to predict crimes with statistics based on
three data sources: BVI (Central Crime Database), GBA (Municipal Ad-
ministration) and CBS (Demographics from Statistics Netherlands). By
analyzing the used data categories with a critical data studies approach,
we will show that the epistemological question concerning predictive po-
licing systems turns into an ontological one: how are living environments
and police work mutually shaped and determined by data? We will argue
that intelligence-driven policing is not only a qualitative shift, but also
has its continuities, since already existing ideas and biases concerning
suspects and crimes are reproduced in the information and system of
CAS.
Keywords: Predictive policing, intelligence-driven policing, critical data stud-
ies, data visualization, information bias.
1 Introduction
During an interview, Dick Willems, data scientist at the Amsterdam Police Department
and the creative brain behind the Dutch predictive policing system, shared an anecdote
about the negative correlation he found between the occurrence of flashers and the
amount of burglaries in a neighborhood. As a result he once, jokingly, suggested police
officers to sometimes take off their clothes in certain neighborhoods in order to make
the number of burglaries go down. In this case it is easy to spot the confusion of causa-
tion and correlation. Because of the occurrence of flashers, the police decided to patrol
more often in certain areas. This increase in surveillance was the reason for a drop in
the number of burglaries, not the fact that exhibitionists gave in to their urges. As we
will show in this paper, the Dutch Crime Anticipation System (from now on CAS) hosts
a number of other assumptions about the relationship between crime, human character-
istics, locations, and the interpretations police officers might have. Since CAS is devel-
oped in-house (rather than being a commercially available product, like PredPol), the
Dutch police itself is able to shape and tweak the system. In addition, because of the
open attitude of the Dutch police and existing transparency laws, researchers are able
to request information regarding the types of data that are used and the way in which
information is presented.
In this paper we will critically analyze the data categories and data sources that are
employed in CAS. With a critical data studies approach (see Dalton and Thatcher 2014;
Dalton et al. 2016), which helps us discuss the underlying influences from people, the-
ory and organizational structures that frame and color a data-driven system, we try to
reconstruct this assemblage of actors related to CAS. As Andrew Illiadis and Frederica
Russo argue, these data-driven systems are not neutral, but should be seen “as always-
already constituted within wider data assemblages” (Iliadis and Russo 2016, p.1). Our
task, in relation to CDS, is to uncover these ideas that are implemented in a system,
which eventually also influence decisions and situations in the world.
We will perform an infrastructural inversion, a way of “recognizing the depths of
interdependence of technical networks and standards, on the one hand, and the real
work of politics and knowledge production on the other” (Bowker and Star 1999, p.
34). We will show that categorization systems are not free from the history and culture
of the organizations in which they are embedded. CAS will be understood as a data
assemblage, a term introduced by Rob Kitchin and Tracey Lauriault, which they ex-
plain as “a complex socio-technical system, composed of many apparatuses and ele-
ments that are thoroughly entwined, whose central concern is the production of data”
(Kitchin and Lauriault 2014, p. 6). By analyzing these apparatuses one focusses on “the
technological, political, social and economic apparatuses and elements that constitute
and frame the generation, circulation and deployment of data” (Kitchin and Lauriault
2014, p. 1), in order “to track the ways in which data are generated, curated, and how
they permeate and exert power on all manner of forms of life” (Iliadis and Russo 2016,
p. 2). By combining the notion of CAS as a data assemblage with an infrastructural
inversion we intent to reflect both on the socio-technical network of CAS within the
police organization as on its dependence on categorization practices in Dutch society,
science and policy as a whole. Through our analysis we will show that the epistemo-
logical question concerning predictive policing systems turns into an ontological one:
how are living environments and police work mutually shaped and determined by data?
We will first reflect on the (not always voluntary) choices of data categories, which
are not only shaped by availability, but also through the existing culture of policing
which already has ideas about what characteristics are prevalent in potential suspects.
Second, we will show how CAS is in no way exhaustive or all-inclusive. It is tied to
types of crimes that are location and time based. Third, by highlighting the data-visu-
alization in the form of a map and the possibilities and impossibilities for the users of
CAS (information specialists), we will show that the data cannot speak for itself; it
needs an interpreter. Finally we will argue that intelligence-driven policing is not only
a qualitative shift, but also has its continuities, since already existing ideas and biases
concerning suspects and crimes are reproduced in the information and system of CAS.
2 The Crime Anticipation System
In line with developments in many areas of business and governance, where bureaucra-
cies of all sorts are increasingly datafied for budgetary reasons and the additional pos-
sibilities for automated analysis (see Kitchin 2014; Schäfer and van Es 2017), the
Dutch Police started with so-called Intelligence-Led Policing (Kop and Klerks 2009).
The groundwork for this development was laid out by the formation of a national police
force from the formerly locally organized “base teams”, which also resulted in a stand-
ardized central national crime database (Strategische Beleidsgroep Intelligence 2008).
This development led to the creation of the Crime Anticipation System (CAS). This
data-driven system tries to predict crimes through statistics based on three data sources:
BVI (Central Crime Database), GBA (Municipal Administration) and CBS (De-
mographics from Statistics Netherlands). CAS uses different kinds of data, which date
back up till three years, which can be classified in three types (Willems and Doeleman
2014, p. 41). The first type of information is socio-economic data from the Central
Bureau of Statistics (CBS) (Willems 2017c; Willems 2017b). This data focusses on
people’s age, incomes and the amount of social benefits in an area. The second type is
data from the Basisvoorziening Informatie (basic information provision, BVI), this is
the data that is gathered by the police force itself and focusses on previous crimes,
locations and known criminals (Willems 2017b). The third type of data comes from the
Municipal Administration (BAG). The data from this source, consisting of streets and
addresses, is not used as a predictor but as structure for the map whereon predictions
are made.
Fig. 1. Picture of the interface of CAS. Color code for risk factors: red = burglary;
green = street robbery; blue = disturbance by youth; pink = bicycle/scooter theft
CAS can be categorized as a spatiotemporal prediction system, meaning it focuses
on hot spots and hot times in a city, and not on high-risk individuals. Through this data,
CAS constructs heat maps (see Fig. 1), that illustrate which places have a higher risk
for high-impact crimes. These heat maps show blocks which are left blank when the
risk is either low or nonexistent, and the color increases in intensity, when the amount
of risk increases (Mali et al 2017, 91). Different colors correspond to different crimes.
Only the top three percent of high-risk areas will be colored in one of the colors. This
is to relate the amount of colored fields to the capacities of a police force.
The goal of CAS specifically, is to predict more at-risk areas in a city, and improve
efficient distribution of manpower (Mali et al. 2017, p. 91; Willems and Doele-
man 2014). Dick Willems initially constructed CAS for the Police of Amsterdam in
2013 (Mali et al 2017, 18). After testing this system in Amsterdam, a pilot project was
started in the Dutch cities of Enschede, Hoorn, Hoefkade, and Groningen-Noord. De-
spite non-conclusive results with respect to user-experience of police officers in the
four before mentioned cities (see Mali et al. 2017), CAS was made available for all
police teams in The Netherlands in May 2017. Currently CAS is in use at 110 base
teams, out of a total of 167, in 6 out of 10 districts (Willems 2017c). It could therefore
be seen as a major factor in determining where and how the Dutch police do their sur-
veillance and patrols.
3 Location and Time
The CAS maps are created with GBA-data (municipal database with addresses, zip
codes etc.). The map is divided in blocks of 125 x 125 meters, a size which was deter-
mined through a process of trial and error by each time halving squares, starting from
1000 x 1000 meters. It proved to be a good compromise between precision and practi-
cality. Too big of a square would be unworkable for patrolling officers, too small of a
square would have a low hit rate and render it useless in the prediction (Willems 2017c).
After plotting the squares on a map, all “empty” squares, such as squares with no houses
or companies, and squares consisting of only water, forest or farmland are removed.
Then of each square, it is determined to which zip code it belongs, and which addresses
would fall within its borders.
Next, historic crime data from the BVI-database (addresses, locations and time/date
of recent incidents of a certain type) is added. The number of incidents within each
square and within its surrounding squares for several time periods before the prediction
is determined (see Fig. 2).
Week -1 -2 -3 -4 -5 -6 -7 -8 -9 -10 -11 -12 -13 -14 -15 -16 -17 -18 -19 -20 -21 -22 -23 -24 -25 -26
CAS 2015 Regression 1
Regression 2
Total
CAS 2017 Regression
Fig. 2. Conceptualization of time periods before prediction in different versions of CAS
In CAS 2015 the resolution of BVI was two weeks, resulting in crime numbers of
four consecutive two-week-periods, four consecutive four-week-periods, and one total
number of incidents for a period of 26 weeks. Two directional coefficients were calcu-
lated. The first directional coefficient is derived from the number of incidents in the
previous four two-week-periods (see Fig. 2, CAS 2015, Regression 1); the second is
derived from the number of incidents in the previous four four-week-periods (see Fig.
2, CAS 2015, Regression 2).
In CAS 2017, the resolution of BVI has increased to one week. A choice was made
for the calculation with twelve consecutive one-week-periods. In addition, the way in
which crime trends are calculated has changed as well. Only one directional coefficient
based on the twelve one-week-periods was used (see Fig. 2, CAS 2017). The change
from two-week numbers to one-week numbers has increased the resolution of the sys-
tem by a factor of two.
3.1 Limitations and biases of location and time variables.
Because CAS is based on locations, it is only suitable for the prediction of crimes
that can be tied to a very specific place, such as burglary, pickpocketing, mugging, etc.
It is not made to predict or identify incidents of, for example, fraud. It therefore also
only targets a specific demographic amongst criminals, something which will be dis-
cussed in more detail in the next paragraph.
For accurate predictions it is necessary to have a relatively precise time of previous
incidents. For pickpocketing, and mugging the time is usually quite precise (within an
hour or so), for burglary however it is often much harder to determine. People are often
not home or not awake when a burglary happens. Since BVI only works with exact
times, rather than with a time window, the time exactly in between leaving and return-
ing home (or falling asleep and waking up), is registered in the database (Willems
2017c). It is unclear whether or not this affects outcomes of predictions.
Another limitation of CAS is the number of incidents of a given type that happen
within a delimited timeframe. Incidents that do not happen very often, such as murder,
or incidents that are not always reported, such as rape or the sale of fake drugs, cannot
be predicted accurately within a spatiotemporal system.
4 Demographic Data
In order to make predictions about locations, CAS makes use of open demographic data
provided by CBS (see Table 1). After determining the zip code of which a square is
part, aggregated information regarding that zip code area is incorporated in the risk
score of a square. CBS indicators 1 to 11 have to do with the people that live in an area;
indicators 12 to 16 have a more economic nature.
Indicators 1, 2, 3, and 4 refer to the number of inhabitants, men, women and house-
holds in an area. Indicators 5, 7, 8, 9, 10 and 11 refer to the composition of the house-
holds, incorporating the average number of people in a household, whether or not there
are children in the house and if there are one or two parents in the household. Indicators
10 and 11 refer to the property values in an area and how many properties are currently
empty. The next three indicators, 14, 15, and 16, are about the number of people that
receive an income, whether or not this income consists of social benefits and what the
total income is in an area.
Given the size of the squares (125x125 m) it becomes clear that the system is more
suited for an urban area, since in rural areas the number of people and households might
not be statistically valid. In addition a predisposition can be seen towards the shape of
a family (one or two parents) and economic factors. It gives the idea that both criminals
and/or victims are more often part of “broken families”. Because of the indicator re-
garding the recipients of social benefits it seems as if poorer people have higher likeli-
hood of criminal behavior or victimhood, as the other end of the spectrum is not part of
the system (i.e. there is no indicator mentioning the number of millionaires).
Table 1. CBS indicators used for each CAS square.
Nr. Explanation
1 Number of inhabitants in zip code area
2 Number of men in zip code area
3 Number of women in zip code area
4 Number of households in zip code area
5 Average size of households in zip code area
6 Number of non-Western allochthones in zip code area (removed in 2017)
7 Number of one person households in zip code area
8 Number of one parent households in zip code area
9 Number of multiple person households without children in zip code area
10 Number of two parent households in zip code area
11 Average age in zip code area
12 Housing stock in zip code area
13 Average property value in zip code area
14 Number of income recipients in zip code area
15 Number of social benefits recipients in zip code area
16 Fiscal monthly income in zip code area
The final indicator that should be mentioned is number 6, which consists of the num-
ber of so called non-Western allochthones and was removed in CAS 2017. The term
allochthone, together with the term autochthone are used to describe the national ori-
gins of Dutch citizens. The term autochthone is used for people with two parents born
in the Netherlands. Allochthone is used for people with at least one foreign-born parent.
Allochthones are then split up in “Western” and “non-Western” people based on the
“different socio-economic and cultural position of Western and non-Western countries”
in comparison to the Netherlands. This results in the categorization of European coun-
tries (except Turkey), North American countries, Japan and Indonesia as “Western” ,
and South American, African and Asian countries (except Japan and Indonesia) as
“non-Western” (CBS 2000). The inconsistency of classifying some former colonies of
the Netherlands, such as Suriname and the Antilles, as non-Western, and others, such
as Indonesia, as Western, combined with the ways in which the word allochthone is
used in everyday language make the categorization system can be understood as race-
ethnically determined (Yanow and Van der Haar 2013). This part of the Dutch govern-
mental knowledge infrastructure was taken over in the CAS system unchanged.
What is striking in CAS 2015 is not only the presence of an indicator about the num-
ber of non-Western allochthones, but also the absence of an indicator for Western al-
lochthones or autochthones. In CAS 2017 the indicator regarding non-Western Alloch-
thones was removed, because “it didn’t add predictive value” (Willems 2017a). This
remark could be understood in at least two ways: 1) race-ethnic origins are not suitable
as a predictor for crime, or 2) the race-ethnic origins of possible suspects are still im-
plicitly woven in the system through the other indicators, making it unnecessary to have
it in there explicitly. An argument could be made for both statements. A 2014 study on
criminal behavior of youth aged 12-18 in the Dutch city of Rotterdam could not find a
relationship between race-ethnic origins and criminal behavior (see Driessen et al.
2014). In addition, in their research on NEET-youth (UK youth aged 16-24, who are
Not in Employment, Education or Training), Helen Thornham and Edgar Gómez Cruz
(2018) found that, although this category is officially not gendered, the lived reality
proved that mostly women fell into this classification. Similarly, CAS 2017 could,
while no longer explicitly take into account the race-ethnic origins of people living in
a neighborhood, implicitly, through the other demographic indicators, and location
based approach, be race-ethnically biased. In addition to the question of effectivity,
Oscar H. Gandy (2016, p. 62) suggests that some indicators, such as race or ethnicity,
could best not be used in predictive systems, not because they are bad predictors, but
because it would be unethical to use non-causal factors.
4.1 Known Offenders
The final three indicators come from BVI and refer to data about previously convicted
criminals or so called “known offenders”. For each square in CAS, first the distance to
the closest known offender is calculated (see Table 2, indicator 1). The remaining two
indicators refer to the number of recently active (within the past six months), known
offenders that live within 500 and a 1000 meters respectively of a square.
Table 2. BVI indicators used for each CAS square.
Nr. Explanation
Distance in km of the address of the closest known offender (suspect) of
1
an incident who has been active in the past 6 months
Number of known offenders (suspects) of an incident who have been ac-
2
tive in the past 6 months that live within 500 meters
Number of known offenders (suspects) of an incident who have been ac-
3
tive in the past six months that live within 1000 meters
Interestingly enough in CAS 2017 the term “known offenders” has been replaced
with the term “suspects”. Although the label still refers to exactly the same category,
one could argue that this is indicative of the mentality of predictive policing in which
someone is a suspect before a (new) crime has happened. One could even question
whether or not the use of these indicators is in conflict with the principle of letting
someone start over with a clean slate after doing time. In addition, legally speaking, the
word “suspect” is reserved for people of who “a reasonable suspicion of guilt to a crim-
inal offense is assumed through facts and circumstances” (CBS 2018, translation by the
authors). This classification is therefore not valid before a crime has happened.
4.2 Continuities, police culture and feedback loops
The indicators used to predict crime are not new discoveries, but rather datafied conti-
nuities of the indicators that police officers use in their work on the street. In his exten-
sive ethnographical fieldwork, cultural anthropologist Sinan Çankaya (2012; 2015)
joined police officers on patrol and observed their decision making process for several
months. He found out that “street level bureaucrats” use three physical categorizations
to determine whether or not to stop and interrogate someone (see Fig. 3). The charac-
teristics of what Çankaya calls the “biological body” and the “disobedient body” can
be seen as real life parallels of the demographic information of CBS and the “known
offenders” data from the BVI. Police officers match this information with the time and
location to determine if there is a mismatch and a need for action (Çankaya 2012 p. 74-
76). In addition, Çankaya found that race-ethnicity plays an important role both in de-
cision making during patrols (Çankaya 2012, p. 46-51; Çankaya 2015) and in the self-
identification of the Police force as (white) Dutch (Gowricharn and Çankaya 2017).
Person Time/Place
day/night
Cultural Body Biological Body Disobedient Body
hotspots
Clothing Skin color
Known offenders
Jewelry Facial features
Body decoration Physical fitness
Age
Sex
Fig. 3. Part of decision making process of Dutch police officers (Çankaya 2012, p. 78)
Because “street level bureaucrats” and CAS are largely aligned in their decision
making processes, the risk for a feedback loop emerges. Subconscious biases regarding
less economically advantaged people and intended and unintended race-ethnic profiling
can result in biased data concerning certain groups in society. This biased data is then
used to determine where to patrol etc. This process could be seen as a form of datafied
“cumulative disadvantage” (Gandy 2016). Cumulative disadvantage creates and rein-
forces differences is the quality of life of different groups. Discrimination in this sense
automatically creates inequality which builds up on itself, hence cumulative disad-
vantage (see Gandy 2016, p. 55).
5 Visualization and Framing
CAS visualizes results in a standardized way for all police departments throughout The
Netherlands that use this system. This is done in two ways. Firstly, CAS provides a
simple line-graph where the x-ax shows time (of a single day, or week) and the y-ax
shows the likelihood of a high-impact crime happening (Willems 2017a). Secondly, the
system presents its results in a grid map. Both ways of visualizing results can be ques-
tioned critically because of the data that they use and the way results are simplified.
The graph that shows time and risk is very easy to read – which is a good thing.
However, the problem with this type of visualization is not the visualization itself, but
the data behind it. Because of the aforementioned problem to determine the time and
date of a burglary, the police use an average time in their report. Say, one leaves their
home at 8 A.M. and comes home at 6 P.M., the report will state that the crime happened
at 1 P.M. (Willems 2017c). Eventually, the graph in CAS will maybe not show a relia-
ble prediction for when the risk is high for a crime as burglary, but more when the risk
is higher of ‘people being halfway through their work day’. Because CAS is a closed
system information officers using CAS do not have the option to look up the data be-
hind this graph.
The second visualization within CAS is a grid map, which is often used in other
predictive policing systems. As stated before, the squares that are colored red represent
the areas with the highest risk for a certain crime to take place, orange squares have a
medium risk, and yellow squares have a low, but mentionable risk.
The only specifics or filter a user of CAS can add within the grip-map visualization,
is selecting a type of crime. This map will then only show the predicted amount of risk
for that selected type of crime. A user cannot, for example, select a square or a crime
and receive more in-depth information on why the risk is predicted as high or low. In
other words, the user cannot see if a risk is predicted on the basis of a stable variable or
because of a more changeable or unstable variable as the presence of a known-offender.
Especially in the early stages of CAS, often the same three percent of squares would
light up, which gave the impression that some undeterminable factor was causing the
risk to always be high for certain areas.
5.1 The role of the data officer
The role of the data officer is to use the visualizations in CAS as a starting point, and
from there to try to explore and retrieve the background of the risks in certain areas.
Data officers are able to use other databases or sources from within police departments.
At the end of their search, they come up with a possible explanation for the certain
amount of risks. In a way, they enrich the results presented by CAS, because they do
not only use the results from CAS, or information from the BVI, but also use their
contextual knowledge on the area and their expertise on crime to look for possible cau-
sational effects. In addition, they can add unpredictable events that are not part of CAS,
such as football matches of the cities home team or music festivals, and determine by
themselves how much this influences the prediction.
Of course one should add that the function of visualizing a large amount of data is
to make information more accessible and readable for people. This is also the function
of the graph and the grid map within CAS. One could not process and calculate all the
variables and data as quickly as CAS. However, as often is said about visualizing data,
it does mean that information is simplified and some elements are left out. Trying to
understand something as complex as crime, what is influenced by many factors, by
‘reading’ a simple visualization could be challenging and is prone to misinterpretations.
Nonetheless, Dick Willems argues that by giving data officers so little extra information
with the visualization and specifics of the predictions made by CAS, stimulates data
officers to research the risk areas in a more open-minded way, because they have to
rely on their own knowledge instead of being closely guided by the results of CAS
(Willems 2017a).
The visualization of results within CAS is kept simple and the numbers of options
given to users to explore these visualizations are small. One could conclude, through
this brief analysis of the interface and visualizations used in CAS, that the Dutch police
have consciously chosen to keep this system closed for users. This changes the role of
the user from someone that merely looks at visualizations and reproduces these results,
to someone that enriches these results with a kind of qualitative explanation. However,
with this extra layer of interpretations by the data officers, it can be said that there is
another moment where possible (personal) biases can come into the equation. As dis-
cussed before, the production of data, such as mainly arresting people from particular
groups, is biased. In the case of CAS, after the data is processed through the algorithms,
humans can add and contextualize these results through their own frame of refer-
ence. However, there is always the risk of skipping the time-consuming process of data
enrichment, and the temptation of simply taking over the predictions of the system that
keeps their users in the dark on purpose. Given the current discussions on issues of
work pressure and the lack of personnel within the Dutch National Police, this scenario
is not unlikely.
5.2 Epistemology or ontology?
As we have shown the CAS epistemology is deeply rooted in the existing (Dutch) cul-
ture of the Dutch police. As it starts to increasingly determine which places and people
are under surveillance it could be approached as ontology as well. We understand on-
tology as “ a social construction of reality, defined in the context of a specific epistemic
culture as sets of norms, symbols, human interactions, and processes that collectively
facilitate the transformation of data into knowledge” (Kuiler 2014, p. 312). Through
this definition it becomes clear that CAS is not a neutral instrument, but a particular
social construction of reality, shaped by what is seen by the Dutch police as deviant
physical traits, economic situations and behavior. CAS does not merely represent these
characteristics in a certain way, but it actively shapes the living environment of people
that are considered not to be part of the norms set in Dutch society through knowledge
infrastructures.
6 Conclusions
We have shown that a critical data studies approach can give us much needed insights
into systems that are shaping our life world. With this perspective we have shown that
the used data categories in CAS are not always voluntary choices of data categories,
but are often determined by availability, not only in terms of what is measured, but also
how it is measured. The CBS data is shaped by Dutch history, culture and politics re-
garding what to measure in terms of social benefits, household composition and race-
ethnic origins of people. BVI data is a result of the existing Dutch culture of policing
which already has ideas about what characteristics are prevalent in potential suspects
and how to record crime (in terms of time and location). By using GBA data the maps
of CAS are tied to existing formats such as streets and house numbers. Because of these
predispositions, CAS is in no way exhaustive or all-inclusive. It is tied to types of
crimes that are location and time based, and happen often enough to be able to make
statistically valid predictions about.
By highlighting the data-visualization in the form of a map and the possibilities and
impossibilities for the users of CAS (information specialists), we have shown that the
data cannot speak for itself; it needs the interpretation of a data officer. However, the
possibilities for the data officer to find out the origins for a prediction are severely lim-
ited. The aim to create uniform predictions creates the possibility for a less accurate but
time saving approach at the end of the interpreter by only consulting CAS.
Another valid concern that needs much more research is the effects of the combina-
tion of possible system biases with human biases. Possible race-ethnic biases of CAS
could be magnified when combined with police practice. For future research critical
data studies approaches could be combined with ethnographic studies to get more in-
sights in work processes and possible feedback loops.
Acknowledgements
We would like to thank Dick Willems, data scientist at the Amsterdam police depart-
ment, for taking the time to answer our questions and providing us with a 2017 version
of the CAS indicators list.
References
Bowker, G.C., Star, S.L.: Sorting things out: Classification and its consequences. MIT press,
Cambridge, MA (1999)
Çankaya, S.: De controle van marsmannetjes en ander schorriemorrie. Het beslissingsprocess
tijdens proactief politiewerk. Boom Lemma, Amsterdam (2012)
Çankaya, S.: De politiële surveillance van ras en etniciteit. In: Moor, L.G., Janssen, J., Easton,
M., and Verhage, A. (eds.) Ethnic profiling en interne diversiteit bij de politie. pp. 13–
33. Maklu-Uitgevers, Antwerpen (2015)
CBS: Allochtonen in Nederland. CBS, Voorburg (2000)
CBS: Begrippen, http://www.cbs.nl/nl-nl/onze-diensten/methoden/begrippen
Dalton, C.M., Taylor, L., Thatcher, J.: Critical Data Studies: A dialog on data and space. Big
Data Soc. 3, 2053951716648346 (2016). doi:10.1177/2053951716648346
Dalton, C.M., Thatcher, J.: What does a critical data studies look like and why do we care?,
http://societyandspace.org/2014/05/12/what-does-a-critical-data-studies-look-like-and-
why-do-we-care-craig-dalton-and-jim-thatcher/, (2014)
Driessen, F.M.H.M., Duursma, F., Broekhuizen, J.: De ontwikkeling van de criminaliteit va n
Rotterdamse autochtone en allochtone jongeren van 12 to 18 jaar: De rol van achter-
standen, ouders, normen en vrienden. Politie & Wetenschap and Bureau Driessen, Apel-
doorn and Utrecht (2014)
Gandy, O.H.: Coming to Terms with Chance: Engaging Rational Discrimination and Cumulative
Disadvantage. Routledge, London (2016)
Gowricharn, R., Çankaya, S.: Policing the Nation: Acculturation and Street-Level Bureaucrats
in Professional Life. Sociology. 51, 1101–1117 (2017). doi:10.1177/0038038515601781
Iliadis, A., Russo, F.: Critical Data Studies: An Introduction. Big Data Soc. 3, 205395171667423
(2016). doi:10.1177/2053951716674238
Kitchin, R.: The Data Revolution: Big Data, Open Data, Data Infrastructures and their Conse-
quences. Sage, London (2014)
Kitchin, R., Lauriault, T.: Towards Critical Data Studies: Charting and Unpacking Data Assem-
blages and Their Work. Social Science Research Network, Rochester, NY (2014)
Kop, N., Klerks, P.: Doctrine Intelligencegestuurd Politiewerk. Politieacademie, Apeldoorn
(2009)
Kuiler, E.W.: From Big Data to Knowledge: An Ontological Approach to Big Data Analytics.
Rev. Policy Res. 31, 311–318 (2014). doi:10.1111/ropr.12077
Mali, B., Bronkhorst-Giesen, C., Den Hengst, M.: Predictive Policing, Lessen voor de toekomst.
Politieacademie, Apeldoorn (2017)
NOS: Politie gaat misdaad voorspellen met nieuw systeem, https://nos.nl/artikel/2173288-poli-
tie-gaat-misdaad-voorspellen-met-nieuw-systeem.html
Schäfer, M.T., van Es, K.: New Brave World. In: Schäfer, M.T. and van Es, K. (eds.) The Data-
fied Society: Studying Culture Through Data. pp. 13–22. Amsterdam University Press,
Amsterdam (2017)
Strategische Beleidsgroep Intelligence: Waakzaam tussen Wijk en Wereld: Nationaal Intelli-
gence Model Sturen op en met Informatie. Politieacademie (2008)
Thornham, H., Gómez Cruz, E.: Not just a number? NEETs, data and datalogical systems. Inf.
Commun. Soc. 21, 306–321 (2018). doi:10.1080/1369118X.2017.1279204
Willems, D.: Interview, (2017)(a)
Willems, D.: Email Variabelen CAS, (2017)(b)
Willems, D.: Predicting criminal events with Data Science. Data Science Utrecht Meet Up. ,
Utrecht (2017)(c)
Willems, D., Doeleman, R.: Predictive Policing: Wens of Werkelijkheid. Tijdschr. Voor Politie.
4, (2014)
Yanow, D., Van der Haar, M.: People out of place: allochthony and autochthony in the Nether-
lands’ identity discourse—metaphors and categories in action. J. Int. Relat. Dev. 16, 227–
261 (2013)