Trust in the Police: The Climb of Private Security Services

(1)

2018

TRUST IN THE POLICE: THE CLIMB OF PRIVATE

SECURITY SERVICES

MASTER THESIS CRISIS & SECURITY MANAGEMENT

M.A. VAN DEN BROEK (S1510711)

(2)

Author:

M.A. (Michel) van den Broek, BSc (s1510711)

University: Leiden University

Faculty:

Faculty of Governance and Global Affairs

Institute:

Institute of Security and Global Affairs

Study programme:

Crisis and Security Management

Supervisor:

Dr. L. (Ludo) Block

Second reader: Dr. J. (Joery) Matthys

(3)

Abstract

The private security services sector has been on a steady climb during the last few decades. Research into this phenomenon has been overwhelmingly focused on theoretical implications, while very little attention has been paid to its more practical consequences.

The body of knowledge indicates that individuals have a greater tendency to trust the police than to trust those working in the private security services sector, whilst also at the same time indicating that the means to differentiate between these two sectors are becoming increasingly blurry. This research hypothesises that relative growth by the private security services sector might be reflected negatively in the number of individuals who tend to trust the police, as the relatively negative views of the private security services sector are theorized to be reflected on the police as a consequence of blurring of boundaries between the two fields.

This research tests this hypothesis in addition to several preestablished theoretical notions concerning other factors of influence on individuals their tendency to trust the police using quantitative research methods.

Using existing data, both descriptive statistics as well as multilevel regression analysis techniques are used to conduct this research. Both the country level and individual level variances are studied.

Taking some nuances and limitations into account, the result of this research indeed indicates that there is a negative relationship between the relative size of the private security services sector and the police on the one hand and the number of individuals who tend to trust the police on the other hand.

(5)

Chapter 1: Introduction

In recent decades the private security services sector has been on a steady climb (among others: Bovens, 2007; Davis et al., 2012; White, 2011). In the context of this climb, especially as compared to the public-sector, a significant amount of research has been conducted already on ethical concerns that might arise in regard to among other things; accountability, transparency and privacy. However, the subject of this research is to study whether or not the aforementioned climb of the private security services sector has had an effect (and whether or not it is plausible that it will have an effect in the future if the trend were to be continued) on individual trust, not in those working in the private security services sector, but in the police. In other words, does society tend to trust the police more or less as a consequence of the size of the private security services sector relative to the size of the police – assuming that there is an effect at all.

I hypothesize that there might be an effect primarily based on two theoretical assumptions derived from prior research. First, there is what one might refer to as blurring of boundaries between the private security services sector on the one hand and policing efforts by the state on the other hand. This phenomenon is widely described, for instance by (among others) Button (2007) and Manzo (2010). Second, generally speaking the public is found to hold higher levels of trust in regard to the police than that they do in regard to private security service providers (Saarikkomäki, 2017).

It goes without say that I will further elaborate on these two points in chapter 2, the theoretical framework. Nonetheless, may it suffice for now that it is plausible that the lower levels of trust in private security services relative to the police could in fact be reflected in lower levels of trust in the police, as a consequence of relative growth of the private security services sector and due to blurring of boundaries between these two fields. Also, I mention relative size very specifically, as size or growth could simply infer a general change in demand for security services indifferent of the question whether they are public or private.

What is severely lacking is prior research is empirical research into practical consequences of the climb of the private security services sector, as the existing literature on this subject is overwhelmingly theoretical is nature. This knowledge gap was already explicitly pointed out by Milward & Provan (2000), then again by the 2009 Inex report from the Peace Research

(6)

research is positioned within that knowledge gap it can be considered as being scientifically relevant.

When we talk about the societal relevance of this research, note that it is primarily derived from the part of this research that is about trust. General trust is believed to be beneficial to society at large. It brings with it a whole slew of positive consequences (Saarikkomäki, 2017) (see the theoretical framework, chapter 2, for further elaboration) and can be seen as a countermeasure of the negative side-effects caused by the risk of having to deal with strangers introduced due to the time-asymmetry of those interactions (Kramer, 2006). Additionally, police trust if found to be a particular good proxy of general trust due to the frontline nature of the work performed by those people working in that field (Saarikkomäki, 2017).

Having said that, we can now formulate the central research question that this research will attempt to answer. It is formulated as follows:

‘To what extend does the size of the private security services sector relative to the police influence individuals to tend to trust the police?’

As I mentioned before, the answer to that question is not clear-cut due to empirical study into practical consequences of the climb of the private security services sector severely lacking. There is no way of getting around the fact that this knowledge gap exists, which is why any results derived from this research should also be viewed keeping that in mind. This research is exploratory in nature and it is therefore inevitable that is has some limitations, I will discuss these in chapter 3, the methodology.

In addition to the quest of answering the aforementioned research question, which can be seen as the primary objective of this research, this research has a secondary objective also. As will become clear from chapter 2, the theoretical framework, trust is a multifaceted phenomenon and many factors are of influence on it. Those factors being most importantly signalled as being of influence on trust in the police by the body of knowledge will be included into this research, so that their assumptions can be tested. How this will be done precisely will be elaborated on in chapter 3, the methodology.

(7)

Lastly, for clarity sake I will provide a brief rundown of the structure and build-up of this research in the form of a reading guide. In chapter 2, the theoretical framework, hereafter I will cover the body of knowledge deemed relevant for this research. Among other things, I will define concepts such as trust and policing, explicate on how they are connected and at the same time differentiate between these concepts both in terms of their public-sector context as well as their private-sector context. I wrap up the theoretical framework by formulating a working hypothesis in order to answer the research question as well as by providing a brief summary of theoretical assumptions about other factors influencing trust derived from the body of knowledge. Following the theoretical framework in chapter 3, the methodology, I will explain how precisely this research is conducted. Quantitative research methods are used in the form of multilevel regression modelling. The methodology chapter will provide an explanation of the exact design of this model, the variables being used, the origin of the data and the different considerations as well as limitations going into this design. The results and analysis chapter, chapter 4, is divided into two parts. The first part analysis the data used for this research using descriptive statistic techniques, while the second part covers the results of the regression analysis. Finally, in chapter 5, the conclusion, I will answer the research question by either accepting or rejecting the working hypothesis that was formulated at the end of the theoretical framework. These results will then be discussed and reflected upon in the discussion paragraph of chapter 5.

(8)

Chapter 2: Theoretical framework

In this chapter I will position this research in the relevant body of knowledge, which I will present and structure as a funnel. Working from the general to the specific, I will elaborate on the knowledge gap I want to address and thereby clarify the potentially added value of this research. To this end I will conceptualize relevant concepts, lay bare the relationship between them and ultimately formulate a working hypothesis in order to answer the research question. The ultimate aim of this chapter is to gain a thorough understanding of the relationship between the concept of trust and that of policing, paying special attention to the public/private dichotomy in the process.

To structure the aforementioned funnel, I will first conceptualize the concept of trust in a general sense. Subsequently I will explain in an equally general manner the empirical importance of the concept of trust. I then move on to the conceptualization of the concept of policing. In this conceptualization I will go over the differences between public and private policing, paying special attention crucially to the blurring of the two fields – as has been signalled by prior research (Button, 2007; Manzo, 2010). Transitioning to the realm of specifics I will explicate on trust in the context of policing generally, laying bare the relationship between these two concepts, before subsequently differentiating between trust in the context of the ‘public’ police on the one hand and private security services on the other hand. Lastly, I address other factors being most importantly signalled by the body of knowledge as being related to the concept of trust. Based on this knowledge the multilevel nature of the concept of trust will become evidently clear, i.e. both individual level influences as well as country level influences are to be considered. Crucially, this is taken into account in formulating the working hypothesis in order to answer the research question finally.

2.1: Defining the concept of trust

It is important to point out that there are two types of trust (Bjørnskov, 2006). On the one hand there is particularized trust and on the other hand there is generalized trust. Particularized trust concerns the degree to which we are willing to trust and consequently interact with those individuals that we already know (Bjørnskov, 2006). These individuals might for example include friends, family, colleagues, etc. Generalized trust on the other hand concerns our willingness to put our trust into those who we do not know, who are strangers to us (Bjørnskov, 2006). This kind of trust is of particular importance in the modern societies in which we live

(9)

today, as our everyday lives force us to have to interact with many strangers from day to day (Saarikkomäki, 2017). This is opposed to the more traditional societies where people tended to know each other to a higher degree. Generalized trust is especially important to this research as interacting with the police or those working in the private security industry generally means an interaction with a stranger. Generalized trust is found to be particularly stable, mainly as a consequence of the fact that for many individuals it is a derivative of their view of institutions – which are generally quite stable (Saarikkomäki, 2017). This does not however mean that generalized trust is unchangeable. Both individual level influences as well as broader country level influences can have an effect on generalized trust over time, I will however address the specifics of those factors at a later stage in this theoretical framework.

For analytical proposes we may want to differentiate between the concepts of trust and confidence (Luhmann, 2000). Trust primarily manifests itself at the individual level, it is active and is therefore predominately determined by our personal interactions with others (Luhmann, 2000). Meaning that a negative interaction that one had personally with someone would decrease that person its trust or vice versa. Confidence as a concept, although closely related to that of trust, is more so a broader, more general term (i.e. country level) used to reflect our views or feelings about an individual or institution in a passive manner (Luhmann, 2000). We might for example have never had any personal interactions with the police and therefor, in the theoretical sense anyway, be unable to say whether or not we trust the police as there would have been no personal interaction that we could base that decision on. Nonetheless we would be able to decide whether we had confidence in the police, as this is based on our broader passive views of the police as an institution and it does not necessarily require active involvement from the individual. As was inexplicitly pointed out above already the concepts of trust and confidence are inadvertently muddled. Trust is at least in part also derived from confidence, also in a non-theoretical setting it is unlikely for an individual to differentiate between these terms (Luhmann, 2000). It is therefore near impossible to differentiate between these in an empirical setting. It is therefore only for analytical purposes that we might want to do this.

(10)

2.2: The empirical importance of trust

The concept of trust is also closely related to that of risk (Jøsang & Presti, 2004). In our interactions with others there is always some degree of risk involved, we can never be completely confident about the other party its intensions, motives or thoughts.

This risk in interactions with others is caused by the time asymmetry of the interaction between parties, i.e. the return on the investment is not immediate (Kramer, 2006). Although trust cannot mitigate the risk itself, it can mitigate the negative side-effects of risk. Which is why these two concepts are closely related and important to one another. The importance of trust can subsequently be elaborated on in a variety of different ways. As mentioned, there is no getting around interactions with others in modern societies. Because of the risks introduced by these interactions, trust can be seen as a bet about future actions of others (Sztompka, 1999). The higher the trust, the more positive our bet about the future actions of others can be. We might therefor want to define trust as a social mechanism for dealing with risk (Jøsang & Presti, 2004). Trust in the police can be seen as especially important, as higher degrees of trust in the police is found to be closely linked to a higher level of trust generally speaking (Saarikkomäki, 2017). This introduces a whole slew of positive consequences (Saarikkomäki, 2017). Just one example is that school achievements and career perspectives are for instance associated with higher levels of general trust (and vice versa the opposite is also true). Trust helps us cooperate with each other better, enables interactions with strangers and it keeps society together generally.

2.3: Defining the concept of policing

For analytical purposes I opt in this research to use the term of policing to refer to security services. The term policing encapsulates a variety of tasks. Saarikkomäki (2017) defines it in the most rudimentary way as control. That definition can be further defined as social control on the one hand and crime control on the other hand. Both forms of control can be exercised by both private as well as public security guards in public, private and quasi-public places. I will elaborate further on the definition of both public and private policing specifically in the following sup paragraphs, after having elaborated on the general definition of policing first.

The fact that social control tasks can be seen as one part of policing does not necessarily mean that the concept of policing and social control are one and the same. When we refer to social

(11)

control in and on itself that really is a much broader term than policing. As it encapsulates not only formal control, but also many informal control mechanisms (Button, 2002). Cohen (1985) defines it as society its responses to deviance. This definition leads us to the more specific definition of policing. According to Saarikkomäki (2017) policing is a combination of surveillance combined with the possibility of sanctions. Which is just one manifestation of social control and is why the term is narrower. Those involved in policing activities can be seen as the gatekeepers of the criminal justice system (Feinstein, 2015). This is both true for those working in the public as well as the private sectors, as those working in the private sector often interact closely with their public police counterparts (Saarikkomäki, 2017).

2.3.1: Public policing

Public policing can of course simply be defined as policing duties performed by actors working for the state. At least traditionally they held the monopoly of force (Westley, 1953), however this notion has been challenged by the rise of the private security industry – discussed in the following paragraph. Nonetheless public policing is distinct is at least a couple of different ways, Saarikkomäki (2017) explicitly mentions four. First, it operates predominately in public space, relatively less in quasi-public space and only as an exception in private space. Second, public policing has the function to serve the common interest. That is in opposition to economic interest or being market driven. Third, they have considerably more legal powers than any other actors. Finally, they are often, although not always, better educated and have to meet higher admission standards than their private counterparts.

2.3.2: Private policing

The industry of private security services in EU countries has grown rapidly during the last two decades. In most EU countries, it has grown much faster relatively than its public counterpart during the same time span (Moreira, Cardoso & Nalla, 2015). Although it must be noted that in most EU countries the number of public security guards still outnumber the amount of private security guards. These statements are reflected also in the statistics provided by the Confederation of European Security Services (CoESS, 2011; 2013; 2015). Which will be consulted and used in this research also, how exactly will be elaborated on in the methodological chapter following this chapter.

(12)

As was mentioned already, policing can be most rudimentary defined as the keeping of order and the prevention of crime (Button, 2002). However, it is important to differentiate between the police and policing. The former carries out the latter, but by no means exclusively. This is a common misunderstanding according to Button (2002) and therefor important to point out. Those concerned with the duty of policing might in fact include a variety of different public institutions, individual citizens, but contemporarily now vitally also the private security industry which encompasses a wide variety of different functions (Button, 2002). When we talk about physical private guarding that is just one dimension of the private security industry. Another dimension of the private security industry is that of investigation and intelligence. Which might encompass all activities related to the gathering, circulation and distribution of intelligence. Information being the keyword of this dimension (Hoogenboom, 2006). The final dimension not to be overlooked is that of the private military companies. Any company providing or enabling services for the potential use of lethal force outside its own country can be seen as a private military company. Which might therefor also potentially include training and advice (Mahairas, 2014)

2.4: Trust in the context of policing

Similarly, to trust in a general sense, trust in the police improves our willingness to interact and cooperate with the police (Warner, 2007). In that sense trust in the police and general trust are not different from one another, which might not be that surprising as the police operate the frontline of the state that operates behind it. We might however want to elaborate further on the practical implications of having varying degrees of trust in the police and thus showing why it is important also. Trust in the police is found to improve the police its legitimacy in society (Nix, Wolfe, Rojek & Kaminski, 2015). It is therefore used commonly as a performance indicator for the police its performance (Yang & Holzer, 2006). Not to differently, trust is seen by the private security sector also as a mean to gather more legitimacy. Which is why private security companies like to emphasize the fact that they are trustworthy (Thumala, Goold & Loader, 2011).

Having established higher levels of trust, this might aid policing activities in at least a couple of different ways. First, it is found that individuals are more willing to cooperate with those performing policing duties whenever there are higher levels of trust involved (Nix, Wolfe, Rojek & Kaminski, 2015). For instance, during arrests or routine checks. Second, prior research

(13)

shows that individuals are more willing to report crime or other misdoings when they trust the police more (Warner, 2007). Finally, higher levels of trust might sway us to adhere to rules and regulations to a higher degree. This has to do with our acceptance of the legitimacy of those involved in policing activities. When we trust these actors, we inadvertently accept their legitimacy. This might be especially beneficial to those involved in private policing duties. They have not yet had the equal opportunity to establish the same legitimacy as their public counterpart, but none the less it is in their interest that we adhere to the rules and regulations which it is their task to make sure that they are adhered to (Thumala, Goold & Loader, 2011).

2.5: Public vs private policing in the context of trust

In this paragraph I first compare the differences in trust between public and private policing. Although not a whole lot of research has been conducted comparing these two, key differences seems to exist nonetheless. I than move on to further discuss the blurring of boundaries between the two sectors that I briefly touched upon already. I will further discuss what its implications might be in the light of the differences between the private and public sector when it comes to trust.

2.5.1: Differences

In the research conducted by Saarikkomäki (2017) it is found that people tend to trust the police more than that they trust private actors fulfilling similar rolls. This was found to be true taking into account both personal interactions as well as general views. Following that point, that is both in terms of trust (individual level and active) and confidence (country level and passive). Following the interviews that Saarikkomäki (2017) conducted in Finland, she concluded that nearly all people differentiated between the police and private security actors in their views of interactions with them for reason of politeness, respectfulness and friendliness. However, more interestingly it was found that interviewees almost never brought up the distinction between public and private. Now, according to Saarikkomäki (2017) a possibility exists that this was because the interviewees saw it as self-evident. However, an interesting alternative explanation might be found in what has widely been described as the blurring of boundaries between sectors (Button, 2007; Manzo, 2010). It is found that the private security industry generally tries to deemphasize its private nature (Saarikkomäki, 2017). For example, through the use of uniforms similar to those of the police. Through this notion it is conceivable that individuals find it

(14)

increasingly difficult to differentiate between public and private policing actors. I will further discuss this phenomenon next.

2.5.2: Blurring of boundaries

The functions performed by the police and the private security industry overlap to a large extend in service – prevention and protection (Button, 2007). In fact, Manzo (2010) found that occasionally private security guards themselves do not differentiate between their function and that of their public counterparts and act accordingly because of this. Despite the fact that the private security industry is usually employed by other private actors and not the public sector, they do operate mostly in so called mass private property (or quasi-public places) (Shearing & Stenning, 1981). Which is also a contributing factor to the fact that de facto they are performing similar functions to their public counterparts. A consequence of this similarity is that because of their execution of largely the same functions, whenever there is relative growth of the private security industry compared to the police it will simultaneously result in a shift of social control (Moreira, Cardoso & Nalla, 2015). Which is traditionally held exclusively by the state (monopoly of force) but now suddenly shifting from the public to the private sector. Shearing and Wood (2007) argue that the rise of the private security industry and the blurring of boundaries between public and private have influenced the democratic legitimacy of public governance, because of the so-called hollowing out of state functions which influenced public perception. We must not however limit our view to the mere rise of the private security industry, we must whiteness this rise in the relative perspective of the public police. According to Zedner (2003) the rise of the private security industry has often been met with an extension of the state. Which could point at a generally increased security demand instead of a shift, in which case a shift in public perception can be expected neither.

Because of the differences in trust in the police and private security services combined with the afore mentioned blurring of fields and hollowing out of state functions, it is not unthinkable that a relative rise of the private security industry could affect trust in the police in a negative way. I will return to this point in formulating the working hypothesis in order to answer the research question after having discussed other influences of trust now first.

(15)

2.6: Other factors influencing trust

The public its perception of the police is of course influenced by other factors (both at the individual and country level) as well. These should critically not be overlooked when studying this subject. First, on the individual level, age (younger equalling less trust) and socioeconomic status (higher equalling more trust) are found by Skogan (2006) to be good predictors of trust in policing efforts, because people with these characteristics (socioeconomic status negatively) are found to be more likely to have a negative encounter with the police. In addition, better education (Van Steden and Nalla, 2010) and being female (Nalla & Heraux, 2003) are found to play a moderate role in expressing a more positive attitude towards policing efforts. The degree of urbanization is found to be a negative influence towards public policing perception also (Moreira, Cardoso & Nalla, 2015). In addition, at the country level the ‘economic prosperity’ already mentioned at the individual level can be translated into GDP (per capita). Also, the perception of corruption if found to be an important predictor of trust in the police (Thomassen, 2013). Being seen as an indispensable institution for social order by many, at the same time corruption is seen as antithetical to that mission. Finally, there is the variation through time. Time is also a key explanatory variable in understanding the relationship between relative growth of the private security industry and police perception. As Livingstone and Hart (2003) explain, the relative perception of private security guards compared to the police is often influenced by stereotypes, for example the ‘failed police officer’ stereotype. It is only as time passes that the private security industry can really establish itself and make a name.

2.7: Hypothesis and assumptions

Following the findings derived from the body of knowledge a working hypothesis in order to answer the research question can now be properly formulated. First, the working hypothesis to answer the research question will be given. Second, a number of theoretical assumptions following the knowledge about other factors of influence on trust will be briefly summarized in order to be able to test these assumptions.

The working hypothesis in order to answer the research question is formulated as follows: ‘The higher the ratio of private security services compared to the police, the lower the number of individuals who tend to trust the police.’ Generally speaking, although the research conducted to this end is limited, it was found by the body of knowledge that trust in private security

(16)

that an increase in private security services might therefor be negatively reflected in lower levels of trust in the police. In addition, because of hollowing out of state functions as a consequence of the increasing loss of the monopoly of force, the state might lose some of it legitimacy which is hypothesised to be reflected in the level of trust in the police also.

Several theoretical assumptions about other factors of influence on trust in the police can also be derived from the previous paragraph of the theoretical framework specifically. First, at the individual level age, being female, socioeconomic status and level of education are assumed to be positively related to the trust in the police. Also, at the individual level the degree of rural living is assumed to be negatively related to trust in the police. At the country level GDP (per capita) and positive perception of corruption are assumed to be positively related to the trust in the police. Finally, time is also assumed to be positively related to the trust in the police.

(17)

Chapter 3: Methodology

In this chapter I will explicate on the general research design of this research first. Many considerations and methodological decisions were made following the knowledge gathered in the theoretical framework in the previous chapter, as well as due to data availability and limitations. I will elaborate how exactly this is the case throughout this chapter. After having discussed the general research design of this research I will cover the different variables used in this research in order to operationalize the different concepts following the knowledge of the theoretical framework from the previous chapter. Finally, I will cover the method of analysis used in this research in order to bring all this together.

3.1: Research design

This research will take the form of a deductive quantitative analysis in a multilevel regression model, based on existing data – primarily the Eurobarometer (European Commission, 2018). This dataset contains extensive survey data from EU member states and candidate member states and is run multiple times per year all the way back since the mid-1970s. The EU is selected as case for this research for two reasons; (1) the availability of data (i.e. the Eurobarometer) and (2) the maximization of external validity through generalizability as the extensiveness of this dataset will allow for a solid comparison between relatively comparable countries. The unit of observation in this research is EU citizens between 2003 and 2015. The Eurobarometer roughly totals around 1.000 respondents in each partaking country every time it is conducted. This large number of observations will also contribute to the external validity through generalizability of this research. Additionally, the extensive sampling will help minimize any potential bias, presenting a good representation of EU citizens, enhancing the external validity/reliability of this research also. This sampling quantity should meet the minimum threshold to avoid any possible bias according to Stegmueller (2013). The time period ranging from 2003 to 2015 is selected because of data availability in regard to the main independent variable used in this research (elaborated upon in the next paragraph). Note however that the following years are omitted because of missing data: 2009, 2011, 2012 and 2014. For reasons of validity I opt in this research to omit al observation for which data from at least 1 variable is missing. Because of missing or partially missing data I therefor opt not to include all 28 current EU member states but to limit this research to 16 different countries spread across the EU. These are following countries: Austria, Belgium, Bulgaria, Croatia,

(18)

France, Germany, Greece, Ireland, Italy, Netherlands, Poland, Portugal, Romania, Spain, Sweden and United Kingdom.

3.2: Variables

A table with an overview of all variables used in this research is presented towards the end of this paragraph (table 1). A brief description of each variable is given in this table also, the sources for each of the variables are given in the right-hand column of the table.

The primary use of the Eurobarometer is its data on EU citizens their trust in the police, which is the dependent variable in this research. The dependent variable is binary in nature (tend to trust or tend not to trust). The Eurobarometer survey is performed face-to-face and in case of this question simply asks whether the person being interviewed would say he/she tends to trust or tends not to trust the police. The question is not further elaborated upon. The research has to rely on a binary scale because of the lack of alternative data availability. Importantly however, Ortiz-Ospina and Roser (n.d.) determined that when comparing trustworthiness (in general) on a binary scale as opposed to a 11-point scale (0-10) the results do not significantly differ on a country level, although of course on an individual level it is more nuanced.

In addition to the Eurobarometer, data from the Confederation of European Security Services (CoESS. 2011; 2013; 2015), covering years back to 2003 on turnover by the private security industry in EU countries, is combined with data from Eurostat (2018) on annual general government spending on police services using the standardized OECD classification of the functions of government (COFOG) published by the UN. Through this a variable is constructed which indicates the ratio between ‘public’ police and private security services. Both parts of the construct are also included as country level controls. Note that both are adjusted to a per capita measurement using data from The World Bank (2018) for intercountry comparability sake. The definitions of both the CoESS as well as the COFOG measurement are quite broadly defined. Which means they might include elements of the police and private security industry which are irrelevant to this research. For example, police training, private investigations, etc. Which are generally speaking nonvisible to the general public and therefor unlikely to influence trust in the police. Although this must be noted as one of the limitations of this research, this data is nonetheless the best that is openly available for the use of this research and fits this research its purpose. Turnover and spending are deemed to be appropriate operationalisation of

(19)

measuring the police and private security industry their size, as according to Kääriäinen (2007) government spending on the police and trust in the police are strongly related.

Table 1 - variables

Variable Description Source

Dependent variable

Trust Trust in the police. Binary variable. 0=tend not to trust the

police. 1=tend to trust the police. Eurobarometer (2018) Independent variable

Ratio

Ratio between 'COFOG (PC)' and 'COESS (PC)'. 0-100 scale. 0=0% ‘COESS (PC)' and 100% ‘COFOG (PC)’. 100=100% ‘COESS (PC)’ and 0% COFOG (PC)'.

Eurostat (2018); CoESS (2011; 2013; 2015); The World Bank (2018) Country level control variables

COFOG (PC)

General government expenditure on police services (according to the Classification of the Functions of Government). Measurement per capita.

Eurostat (2018); The World Bank (2018)

COESS (PC)

Yearly turnover of the private security industry (according to the Confederation of European Security Services).

Measurement per capita.

CoESS (2011; 2013; 2015); The World Bank (2018) CPI Corruption perception index. 0-10 scale. 0=high corruption.

10=low corruption.

Transparency International (2018) GDP (PC) Gross domestic product. Measurement per capita. The World Bank

(2018) Individual level control variables

Age Age. Measurement in years. Eurobarometer (2018)

Female Gender. Binary variable. 0=male. 1=female. Eurobarometer (2018)

Education

Age at which full-time education was stopped. 1-5 scale. 1=no education. 2=still studying. 3=15 or younger. 4=16-19. 5=20 or older.

Eurobarometer (2018)

Rural Rurality of living. 1-3 scale. 1=rural area or village. 2=small

or middle-sized town. 3=large town. Eurobarometer (2018) Note: Whenever Eurobarometer (2018) is given as a source it refers to the following standard issues of the Eurobarometer: 59.1 (2003), 60.1 (2003), 61 (2004), 62.0 (2004), 64.2 (2005), 66.1 (2006), 68.1 (2007), 69.2 (2008), 74.2 (2010), 82.3 (2013) and 83.3 (2015).

(20)

According to the insights gained from the body of knowledge several other influences of trust are not to be overlooked when studying this subject matter. The theoretical assumptions about these other influences were already formulated at the end of the theoretical framework chapter previously. In this research these other influences are operationalised through the use of control variables. One limitations of the use of the Eurobarometer is the inability to properly operationalise socioeconomic status at the individual level across all years studied, therefor this control variable is omitted from the research. GDP (PC) can substitute this measurement somewhat at the country level but is of course unlikely to explain any significant individual level variation. The descriptions in table 1 of the other control variables used to operationalize and test the theoretical assumptions provide ample explanation and should not require further elaboration. Time is not included in table 1 but is also used in this research as a control variable (both at the county as well as the individual level), it is measured by year.

3.3: Method of analysis

The method of analysis of this research consists of two consecutive parts; descriptive statistics first and regression analysis second. As many quantitative analyses do, a healthy dose of descriptive statistics is used first in order to thoroughly dissect the dataset being used (Healey, 2012, pp. 21-140). The goal of this part of the analysis is to identify any anomalies, outliers and patterns in the dataset. This will be achieved by using a combination of graphs and figures, both cross-sectional as well as longitudinal (depending on appropriateness and relevance). Any findings following this part of the analysis will be explicated on to the best possible length, keeping relevance to the primary objective of this research and research question in mind.

In order to conduct the regression part of the analysis of this research, multilevel regression analysis will be used as method to analyse the dataset. This will be achieved using Stata its clustering functionality. This method can produce similar results to a traditional multilevel regression model (Serricchio, Tsakatika & Quaglia, 2013). The individual level observations of the dataset will be clustered into country level observations. The individual level observations combined with the country level observations will result into two distinct observation levels. Failing to acknowledge the multilevel nature of the dataset being used, as demonstrated by the elaboration on the different contextual (control) variables at the end of the theoretical framework chapter of this research, would violate the assumption that errors are independent

(21)

(Solt, 2008). Which could lead to underestimation of standard errors and would derogate the internal validity of this research.

The secondary objective of this research is theory testing of the various control variables being used. In order to achieve this secondary objective, the control variables used in this research will not be blindly integrated into the wider regression analysis used for the primary objective of this research and answering of the research question. As that would also be to the detriment of the internal validity of this research, as there would be no way of knowing whether any of the control variables in the dataset being used specifically are in line with the theoretical assumption being tested. In order to test this individual correlation coefficients (the statistical measure that indicates the strength of the relationship between the relative movements of the two variables ranging between 1 for perfect positive correlation and -1 for perfect negative correlation (Bryman, 2012, p. 348)) between the dependent variable ‘Trust’ and the control variable in question are studied, after which their appropriateness for inclusion in the regression analysis is evaluated.

Regression analysis between the dependent variable ‘Trust’ and independent variable ‘Ratio’ will first be performed at the country level, including all country level control variables deemed appropriates for inclusion into the regression analysis following the testing of theoretical assumptions. Each observation of ‘Trust’ at the country level will represent the mean of ‘Trust’ for a particular country in a particular year. Because of this conversion, at the country level, ‘Trust’ is not a binominal measurement and an ordinary least square (OLS) regression model is used for the country level regression analysis (Allison, 1999, p. 2). The model will be estimated once with robust standard errors (Allison, 1999, p. 127) and once with standard errors adjusted for cluster/country (Serricchio, Tsakatika & Quaglia, 2013). Using both of these techniques will further increase internal validity. Robust standard errors can deal with heteroscedasticity (which could cause inefficiency of the standard errors) if it was to occur (Allison, 1999, p. 127) and clustered standard errors adjust for the hierarchal nature of the dataset being used (Serricchio, Tsakatika & Quaglia, 2013) (at the country level countries are included multiple times because of multiple years, which is cluster is used not only at the individual level but at the country level also).

(22)

individual level similarly to the country level regression analysis. Using the same mathematical technique would result in what is known as a linear probability model (LPM) (Allison, 1999, p. 153), which is among other reasons flawed in at least one major way. In order to calculate significance, the standard error is assumed to be normally distributed (Allison, 1999, pp. 122-123). With LPM it is binomially distributed, and significance cannot be confidently assumed. Which is why a Probit regression model is employed instead (Allison, 1999, p. 187). Specifically, the marginal effects of this model will be calculated (at means for accuracy). The meaning of which is that the coefficients of this model represent the percent point probability by why the chance of the value of the dependent variable being one changes for each one unit change of the independent variable in question (Healey, 2012, p. 273). The use of a Probit regression model enables a more accurate level of significance estimation over the LPM model and can additionally deal with heteroscedasticity and non-linearity (Allison, 1999, p. 184), if they were to occur. As with the country level regression analysis two models will be calculated; one with robust standard errors and one with standard errors adjusted for cluster/country.

(23)

Chapter 4: Results and analysis

In this chapter I will report the results of the data analysis that I have conducted as described in the methodological chapter previously. The aim is to do this in the most accurate fashion possible, i.e. complete objective/unbiased, verifiable and systematic. To which end the results will be presented in various tables and figures. Each of which I will subsequently elaborate on and interpret in my analysis. Reflections will be made in relationship to the theoretical framework whenever deemed appropriate. However, conclusions will only be drawn and discussed in the final chapter of this research hereafter.

The structure of this chapter is that I will first discuss the general descriptive statistics part of my data analysis. I then move on to regression analysis, which I have opted to divide into two parts. First, regression analysis at the country level. Second, regression analysis at the individual level. Although as stated in the methodological chapter, because of the multilevel nature of this research, i.e. the hierarchical organised data, adjustments to the standard errors are made based on cluster (country) variance. In both parts of the regression analysis correlation matrixes (table 3 and 5) are used to test theoretical assumptions and to evaluate the appropriateness of the inclusion of control variables. The results of the regression analyses will be reported in results tables after each correlation matrix (table 4 and 7). Additionally, the individual level regression analysis uses a classification table (table 6) to appropriate the goodness of fit of the employed model.

4.1: Descriptive statistics

The table proving a general overview of the descriptive statistics (table 2) can be found on the page hereafter. The different variables used in this research are presented on the x-axis of the table. From left to right these are; the dependent variable (trust in the police), the independent variable (public policing/private security services ratio) and the various control variables used. The only variable omitted from this table is the variable ‘Year’. Note however that the dataset used covers the period 2003 – 2015, with years 2009, 2011, 2012 and 2014 missing. The reported values are averages for the sum of the time period covered by the dataset.

(24)

Table 2 – descriptive statistics (note: N=88.683)

Trust Ratio COFOG (PC) COESS (PC) CPI GDP (PC) Age Female Education Rural

Austria 80% 13% € 232 € 36 8 € 42.551 46 50% 3,8 1,8 Belgium 66% 14% € 323 € 50 7,4 € 39.692 48 51% 4 1,7 Bulgaria 43% 29% € 56 € 25 3,8 € 5.231 48 51% 4 2,1 Croatia 50% 25% € 129 € 41 4,6 € 13.617 46 56% 3,9 1,8 France 69% 23% € 276 € 81 7,1 € 39.837 49 54% 4 1,7 Germany 79% 19% € 226 € 54 7,9 € 37.324 49 49% 3,8 1,9 Greece 61% 16% € 202 € 37 4,3 € 25.096 48 53% 3,7 2,2 Ireland 75% 26% € 372 € 132 7,6 € 53.847 47 51% 4 2 Italy 62% 12% € 322 € 45 4,6 € 38.007 45 54% 3,8 2 Netherlands 77% 21% € 315 € 81 8,6 € 50.992 51 50% 4,2 1,8 Poland 57% 30% € 102 € 42 4,6 € 11.197 47 54% 4 1,9 Portugal 63% 26% € 177 € 61 6,3 € 19.158 48 55% 3,3 1,8 Romania 44% 24% € 86 € 28 4,1 € 8.641 46 48% 4 1,8 Spain 66% 24% € 269 € 83 6,6 € 27.375 46 53% 3,5 1,8 Sweden 77% 25% € 219 € 72 9,1 € 46.497 49 48% 4,1 1,9 United Kingdom 74% 14% € 415 € 68 8,1 € 43.150 50 52% 3,8 2 Min 0 9% € 45 € 11 3,5 € 3.381 15 0 1 1 Max 1 39% € 443 € 144 9,3 € 61.257 99 1 5 3 Average 69% 21% € 244 € 60 7 € 34.064 48 51% 3,9 1,9

(25)

2007 has the lowest amount of observations (N=4975) and 2013 has the highest amount of observations (N=13.610). Averagely there are 9.890 observations per year. The overall amount of observations for the dataset comes in at N=88.683. The overall amount of observations is consistent for all variables used. As discussed in the methodological chapter, all observation with missing values for any of the used variables were omitted. From top to bottom; the 16 countries part of the dataset are listed alphabetically on the y-axis of table 2, followed by minimum, maximum and average values also. In case of individual level variables country means are used (trust in the police, age, gender, education and degree of rural living) and are reported as percentages in case of binominal variables (trust in the police and gender). The minimum and maximum values are observed in a singular year.

I will now go over each variable, pointing out and analysing the noteworthy observations. That is mainly the significant outliers as well as any generalizations to be made. In case that there are no noteworthy observations to be made this will also be mentioned briefly. Trust in the police (‘Trust’) is the first variable shown in table 2. It is the independent variable of this research and I will therefor describe it more in depth than the other variables. The descriptive statistics on ‘Trust’ shown in table 2 are also visualized in figure 1 (shown below).

Figure 1 – trust, by country

(26)

scale only 43% of people in Bulgaria tend to trust the police. That is just under 40% of variation, which is considerable. The percentages of people in other countries that tend to trust the police are gradually distributed somewhere between these two countries, there are no significant outliers (see: figure 1). What is noteworthy however is that we may derive a distinct geographical distribution of trust in the police from figure 1. Using the definition by EuroVoc (2018), people in Northern and Western European countries tend to trust the police the most, followed by people in Southern European countries and finally people in the Central and Eastern European Countries. The only outlier is this regard is Belgium, in which one would expect more people to tend to trust the police based on the geographical classification given above. A possible explanation for this is possibly found in the somewhat tainted Belgian police performance history (i.e. the Dutroux case), however further investigation into this anomaly is somewhat irrelevant to this research. When we chart the progression of trust over time, differentiating between the three geographical areas mentioned, we can also see different trends of evolution for these (see: figure 2, shown below).

Figure 2 – trust, by year

What we may derive from figure 2 is that trust increased in both Northern and Western European countries, as well as Central and Eastern European. Although the progression is more linear in Northern and Western European countries than Central and Eastern European countries. The significant increase in trust in Central and Eastern European courtiers between

(27)

2008 and 2013 is predominately due to an increase in trust in Poland during that period, although it must be noted that it is not necessarily clear why. Over time, the amount of people that tend to trust the police in Southern European countries is more stable compared to the other two geographical areas, although some short-term variation between years can be observed (figure 2).

Next is the variable portraying the ratio between public policing and private security services (table 2: ‘Ratio’), it is the independent variable in this research. It is measured in percentages, the meaning of 100% being that there are only private security services and no public policing. Italy has the lowest ratio (12%) and Poland has the highest ratio (30%). The average ratio of the 16 countries in this dataset is 21%. Most countries seem to have either a ratio a couple of percent below or a couple of percent above the average ratio. The Netherlands (21%) and Germany (19%) are closest to average. At the maximum Poland (30%) and Bulgaria (29%) are somewhat outliers. There are no noteworthy outliers at the minimum, the distribution is more gradual towards the bottom.

That leaves the different control variables used in this research to be discussed. These will be discussed somewhat more briefly than the dependent and independent variables discussed previously. ‘COFOG (PC)’ is the per capita measurement of annual government spending on police services using the standardised COFOG classification. It is lowest in Bulgaria (€56) and highest in the United Kingdom (€415). These two countries are also the most significant outliers. At the minimum Romania is also somewhat of an outlier (€86). In contrast, ‘COESS (PC)’ is the per capita measurement of annual turnover by the private security sector. It is lowest in Bulgaria (€25) and highest in Ireland (€132). Ireland can be seen as a significant outlier at the maximum, there are no significant outliers at the minimum.

‘CPI’ measures the corruption perception index on a 1-10 scale (general perception of corruption). Higher equalling less corruption. The Northern and Western European countries seem to score best, Sweden being the highest (9,1). Whilst the Central and Eastern European countries seem to score worst, Bulgaria being the lowest (3,8). There are no significant outliers, the distribution is quite gradual. GDP per capita (‘GDP (PC)’) is lowest in Bulgaria (€5,231), followed quite closely by Romania (€8.641). It is highest in Ireland (€53.847). At this point it

(28)

dependent variable in this research (trust in the police). This will however be further demonstrated and elaborated on in the regression analysis paragraph of this chapter, later on.

The final few control variables are all individual level controls. The reported values in table 2 are country means. The variation between countries is minimal. The Eurobarometer has a minimum respondent age of 15. This is also the minimum observed age in the dataset, the highest is 99 and the average is 48. As said, the variation between countries is minimal. The respondents from Italy have the lowest average age (45) and the respondents from the Netherlands have the highest average age (51). In regard to gender we can see an almost fifty-fifty distribution between males and females in de dataset, at 51% percent of observations being female. The respondents in this regard are most skewed in Croatia, with 56% of respondents being female. At the other side of the scale we can see Romania and Sweden having the least number of female respondents (48%). Education is measured by number of years of education on a 1-5 scale. On average the respondents in the Netherlands have enjoyed the most years of education, the respondents in Portugal have enjoyed the least. Geographically we might additionally want to point out that the respondents in Spain have enjoyed the second least years of education. The degree of rural living is measured on a 1-3 scale (1=rural area, 2=town or village and 3=city). The respondents from Belgium and France are reported as having the highest degree of rural living, while Greece its respondents are reported as having the lowest degree of rural living. However, again the variation between countries is minimal.

4.2: Regression analysis

To reiterate, the first part of the regression analysis will purely concern the country level. This will allow to us draw conclusions based on variance between the 16 countries included in the dataset. The latter part of this paragraph will look at individual level variance. In both instances the results of two regression analyses will be reported. In case of the first standard errors are robust, whilst in case of the second the standard errors are adjusted for cluster/country (due to the hierarchical organised data). Doing this will result in similar results as compared to a more traditionally modelled multi-level regression analysis (Serricchio, Tsakatika & Quaglia, 2013). In regard to the country level analysis, the results of the regression analysis with standard errors adjusted for cluster/country are also reported. This is done because each observation in the country level analysis represents a country in a single year, as mentioned the dataset used covers

(29)

multiple. Therefore, by adjusting the standard errors for cluster/country at the country level analysis the results will have greater validity by doing this.

I will now cover the country level regression analysis. Table 3 (below) reports the results of the correlation matrix made between the dependent, independent and control variables measured at the country level. Note that N=93 for this part of the regression analysis, despite that fact that the dataset used includes only 16 different countries. This is due, as mentioned, to each observation at the country level representing a country in a single year and the dataset used covering multiple. The correlation matrix allows us to test theoretical assumption and subsequently to evaluate the appropriateness to include certain control variables in our regression analysis.

Table 3 – country level correlation matrix (note: N=93)

Trust Ratio COFOG (PC) COESS (PC) CPI GDP (PC)

Ratio -0,376*** COFOG (PC) 0,617*** -0,545*** COESS (PC) 0,470*** 0,257** 0,603*** CPI 0,808*** -0,252** 0,577*** 0,499*** GDP (PC) 0,792*** -0,426*** 0,792*** 0,569*** 0,795*** Year 0,130 0,013 0,204** 0,211** -0,110 0,113 Note: * = p<0,1 ** = p<0,05 *** = p<0,01

First of all, the ratio between public policing and private security services and trust in the police seem to be interlinked. The correlation coefficient is negative at -0,376, which is significant at the p<0,01 level. This easily meets the generally utilised p<0,05 threshold, which means that we can be 95% confident that our sample is an accurate representation of the population as a whole (Bryman, 2012, p. 348). The fact that the relationship seems to be negative is also in line with our theoretical assumptions and hypothesis. However, the strength of the relationship seems to be somewhat less strong then some of our other variables. I will elaborate on this by means of visualization through the use of scatterplot momentarily, I will however cover the correlation coefficients concerning the control variables reported in table 3 first. Both government spending on the police (‘COFOG (PC)’) and turnover by the private security sector (‘COESS (PC)’) seem to be interlinked with trust in the police. Their relationship seems to be positive correlated, with ‘COFOG (PC)’ having a correlation coefficient of 0,617 and ‘COESS

(30)

level. These positive correlation coefficients seem to confirm our theoretical assumption that trust in the police is derived to a great extend from their presence and visibility. The probability of which is greatly expanded as finances become less of an issue. Furthermore, the fact that ‘COFOG (PC)’ seems to have a greater correlation coefficient than ‘COESS (PC)’ can be linked to the theoretical assumption that private security services have traditionally been less visible than the public police (although they are making an effort to improve this) (Thumala, Goold & Loader, 2011). The above stated can subsequently be visualised in a scatterplot (figure 3, below). Each dot in figure 3 represents an observation, each corresponding with a particular value of trust (x-axis) and police expenditure or private security sector turnover (y-axis). The radian of the trendlines can be used to predict the direction and strength of the relationship. The tighter and more evenly distributed the observations are along the trendline, the more confident we can be about our assumption.

Figure 3 – ‘Trust’ vs ‘COFOG (PC)’ and ‘COESS (PC)’

Corruption perception (‘CPI’) and GDP per capita (‘GDP (PC)’) seem to have the strongest correlation coefficients out of all. They are both positively interlinked with trust in the police. ‘CPI’ has a correlation coefficient of 0,808 and ‘GDP (PC)’ has a correlation coefficient of 0,792. ‘CPI’ and ‘Trust’ are both about perception, therefore it is unsurprising that these are so strongly correlated. Both are significant at the p<0,001 confidence level. These results are in line with the theoretical assumption that the perception that corruption is not so much a problem

(31)

in one its country and higher welfare both contribute to higher levels of trust in the police (respectively: Thomassen, 2013; Skogan, 2006). These relatively high correlation coefficients can again both be visualized in scatterplots for extra elaboration. Below, ‘CPI’ is visualized in figure 4 and ‘GDP (PC)’ is visualized in figure 5. In both instances it can be observed that all observations sit relatively close to the trendline and the radian of the trendline in relatively big.

Figure 4 – ‘trust’ vs ‘CPI’

(32)

‘Year’ does not seem to be interlinked with trust in the police, at least not at the country level that is (more about that in the individual level regression analysis part of this paragraph). The theoretical assumption was that time increased trust, as time was needed to get rid of negative stereotypes (Livingstone and Hart, 2003). However, this correlation matrix does not confirm that assumption as the correlation coefficient of ‘Year’ is not significant at the p<0,05 confidence level threshold, nor is it significant at the p<0,10 confidence level threshold.

As I mentioned I would elaborate on the strength of the main independent variable in this research (‘Ratio’) as compared to the various control variables used. To this end we can compare the visualization of ‘Ratio’ in the scatterplot depicted in figure 6 (below) to figure 4 (‘CPI’), 5 (‘GDP (PC)’) and to a lesser extend figure 3 (‘COFOG (PC)’ and ‘COESS (PC)’). Looking at the trendline in figure 6 we can definitely see the negative correlation coefficient found in the correlation matrix being reflected. However, the visualization is less obvious than in the other figures mentioned previously as the observations are scattered to a greater extend. Which in turn reflects the correlation coefficient of ‘Ratio’ being less strong as compared to the other correlation coefficients.

(33)

What will be discussed next are the results from the regression analysis performed at the country level. The results of which are presented in table 4 on the next page. The function of the discussion of the correlation matrix (table 3) was in part to evaluate the appropriateness of the inclusion of different control variables, which were selected based on theoretical assumptions. As has been discussed the effects and levels of significance of all these variables seemed to confirm our theoretical assumptions, except for our time variable ‘Year’. Because of this all control variables except for ‘Year’ were included in the regression analysis.

Table 4 – country level regression analysis Dependent variable: trust

Independent variables OLS

(SE: robust) OLS (SE: cluster) Ratio -0,008** -0,008** (0,003) (0,003) COFOG (PC) <-0,000** <-0,000** (<0,000) (<0,000) COESS (PC) 0,002** 0,002** (0,001) (0,001) CPI 0,033*** 0,033*** (0,006) (0,006) GDP (PC) <0,000** <0,000** (<0,000) (<0,000) Constant 0,545*** 0,545** (0,081) (0,067) N 93 93 R2 _0,74 _0,74 Note: * = p<0,1 ** = p<0,05 *** = p<0,01

The first result to point out is the minimal difference between the regression analysis with robust standard errors and the regression analysis with standards errors adjusted for cluster/country. As was explained in chapter 3, the methodology, these two techniques control for different potential problems and/or inefficiencies. The explanatory power of this model is reasonably strong with R2_{=0,74. I will not explicitly go over each result, as the results are presented in} table 4 and are for the most part self-explanatory. Nonetheless I will point at some key results. ‘Ratio’ has a negative correlation coefficient of -0,008 (p<0,005 in both instances). As expected

(34)

true for ‘CPI’ which has the biggest correlation coefficient out of all variables at 0,033 (p<0,001 in both instances). The effects of the other variables seem to be largely mitigated when controlled for by other variables. That is not to say the effects are not significant as they all meet the p<0,05 confidence threshold, they are just relatively small. The only result contradicting the correlation matrix in that sense is ‘COFOG (PC)’ which here has a negative correlation coefficient <-0,000 (p<0,005 in both instances). In the correlation matrix the relationship was found to be seemingly positively correlated, the opposite seems to be true when controlled for by these other variables. This is not entirely surprising as previous research showed that higher public police spending could indeed cause lower levels of trust by function of corruption, but only if more money is spent on it as compared to other means (mainly: Kääriäinen, 2017, p. 429). The exact reasoning for this finding will however not be further explored in this research as it is not its primary objective.

Before I move on to the results and analysis of the individual level regression analysis, we must first briefly go over the appropriateness of inclusion of individual level control variables. These were pointed out by the body of knowledge and by inserting them in a correlation matrix with ‘Trust’ (table 5, below) we may again test theoretical assumptions.

Table 5 – individual level correlation matrix (note: N=88.683)

Trust Year Age Female Education

Year 0,028*** Age 0,065*** 0,076*** Female 0,015*** -0,002 -0,003 Education 0,031*** 0,089*** 0,024*** -0,022*** Rural -0,026*** 0,002 -0,060*** 0,008*** 0,058*** Note: * = p<0,1 ** = p<0,05 *** = p<0,01

The correlation coefficients of the relationships of variables depicted at the individual level and ‘Trust’ are all smaller compared to those depicted at the country level in table 3. This is almost inevitable as the individual level at N=88.683 is much more likely to show a greater degree of variance than the country level at N=93. Nonetheless they are all significant at the p<0,01 confidence level. All exceeding the p<0,05 threshold. The relationship between ‘Year’ and ‘Trust’ is positively correlated at 0,028 in line with theoretical assumptions, but as mentioned it is also significant (p<0,01) which is in contrast to the result from the country level correlation matrix (table 3) in regard to this variable. Therefor ‘Year’ is included in the individual level

(35)

regression analysis, whilst it was not included in the country level regression analysis. ‘Age’, ‘Female’ and ‘Education’ (correlation coefficients of 0,065, 0,015 and 0,031 respectively) are all positively related to ‘Trust’, whilst ‘Rural’ (correlation coefficient of -0,026) is negatively related to ‘Trust’. These findings are all in line with theoretical assumptions. All variables from the correlation matrix (table 5) are deemed appropriate for inclusion in the regression analysis at the individual level based on the argumentation above.

The results from the individual level regression analysis are presented in table 7 on the next page. This model is reasonably able to predict ‘Trust’ at 69,46% of observations being correctly classified. The classification table (table 6) is presented below. As this table shows the model (table 7) is primarily limited by a relatively high number of false positives.

Table 6 – classification table of individual level regression analysis Actual y=1 Actual y=0

Predicted yhat=1 59.309 25.206 84.515

Predcited yhat=0 1.877 2.291 4.168

61.186 27.497 88.683

The reported coefficients in table 7 are marginal effects (calculated at means). The reported results for each independent variable therefor depict the number of percent points by which the probability of ‘Trust’ having a value of 1 (=tend to trust the police) changes for each 1 unit increase of the respective independent variable. Similarly, to the results of the country level regression analysis, the results of the individual level regression analysis (table 7, next page) are deemed self-explanatory for the most part. However, again I will elaborate on key results. The main independent variable, ‘Ratio’, has a negative marginal effect of -0,010 on the probability of an individual to tend to trust the police and is significant at the p<0,01 confidence level. This is true in both instances of the regression analysis, the model using robust standard errors and the model using standard errors adjusted for cluster/country.

The direction of the marginal effect of ‘COFOG (PC)’ on the probability of an individual tending to trust the police contrasts the correlation matrix (table 5) similarly to how it did so in the country level part of the regression analysis (table 3 vs table 4). ‘GDP (PC)’, ‘Education’ and ‘Rural’ do not meet the p<0,05 confidence level threshold (although in all instances p<0,1)

Trust in the Police: The Climb of Private Security Services