Agent-Based Modelling for Criminological Theory Testing and Development

Charlotte Gerritsen, Henk Elffers, Henk Elffers, Charlotte Gerritsen

• Publisher: Routledge
• Year: 2020
• ISBN:
  9781000287059
• Edition: 1
• Topic: Law
• Subtopic: Criminal Law
• Index: Law

Agent-Based Modelling for Criminological Theory Testing and Development addresses the question whether and how we can use simulation methods in order to test criminological theories, and if they fail to be corroborated, how we can use simulation to mend and further develop theories.

It is by no means immediately obvious how results being observed in an artificial environment have any relevance for what is going on in the real world. By using the concept of a "stylized fact, " the contributors bridge the gap between artificial and real world. With backgrounds in criminology or artificial intelligence (AI), these contributors present agent-based model studies that test aspects of various theories, including crime pattern theory, guardianship in action theory, near repeat theory, routine activity theory, and general deterrence theory. All six simulation models presented have been specially developed for the book. Contributors have specified the theory, identified stylized facts, developed an agent-based simulation model, let it run, and interpreted whether the chosen stylized fact is occurring in their model, and what we should conclude from congruence or incongruence between simulation and expectations based on the theory under scrutiny. The final chapter discusses what can be learnt from these six enterprises.

The book will be of great interest to scholars of criminology (in particular computational criminologists and theoretical criminologists) and AI (with an emphasis on AI for generative social processes), and more widely researchers in social science in general. It will also be valuable for master's courses in quantitative criminology.

1Agent-based modeling for criminological theory testing and development

Charlotte Gerritsen and Henk Elffers

Introduction

Within the field of criminology, research is typically conducted using well-established methods. Information is gathered by extracting insights from existing law enforcement data, conducting field research and interviews, through the use of vignette studies, or by performing controlled experiments. The outcomes of this research are most often analyzed manually or by the use of statistical methods. While these types of research have all proven successful, they do not cover the entire field of possibilities. What if data sets become too large to study by hand? Or if the number of relevant concepts increases and the possible interactions grow beyond limits? Such data cannot be analyzed in the traditional manner. While the experimental method is a crafty approach for research, sometimes real-life experiments are impossible; for instance, in cases where we would like to experiment with alternatives to the present legal system of law enforcement. It is not possible to treat people in a way that the law prohibits, but it would be very useful to explore what the results of such alternative measures, now contra legem, would be. This is where innovative research methods come into play. New methods are being developed every day with an emphasis on computational data science analysis.

One research method that can extend the possibilities of the current spectrum of research is called agent-based modeling (ABM). This method, which is often used to perform dynamical social simulations, can be used to test what implications certain strategies with respect to crime might have before these strategies actually become part of daily practice (Brantingham & Brantingham, 2004; Gerritsen & Klein, 2014; Gerritsen, 2015; Gerritsen & Elffers, 2016). As well as testing informal theories in a formal manner, ABM may also be used to make predictions. In the next section we will go into the method of ABM in more depth.

Agent-based modeling

ABM is a computational method that enables the user to create, analyze, and experiment with models composed of agents (for a detailed explanation of the modeling and simulation cycle, see Gerritsen, 2015). Here, agents are autonomous entities that interact with each other and with their environment in some artificial simulated world (Gilbert, 2008). A model is a representation of an object, system, or idea other than that of the entity itself, according to Shannon (1975). In a model the most important concepts of the relevant real-world system are described, as well as the relationships between them. The subject of the model may be very complex and often not all concepts or relationships between the concepts are fully known. This means that it is usually not possible to describe all aspects and relations completely and unambiguously. Hence, a model is typically a simplification of reality and is not presumed to be a complete representation. When drawing conclusions based on the model it is important to realize that the model is the outcome of some assumptions and decisions made by the developer. So it might not represent reality as you know it.

Once a model has been developed one thing to do with it is to imitate the dynamics of a process over time, which is called simulation. Doing so, a simulation model helps to clarify the interaction between different aspects in order to accurately study a model. The output of the simulation model can be evaluated by comparing it to real-world situations. If there is a mismatch then something is wrong, either with the model or with the underlying theory.

Constructing a model of a real-world process has a number of advantages. First of all, it can be less expensive, less time consuming, and more feasible to perform experiments using a model than in the real world. Besides that there are some more profound advantages. Using a model makes it possible to study a process that cannot be studied directly – for example, processes that occur inside the human brain, e.g., cognitive processes. Although a lot is known about cognitive processes, it is not always trivial to get insight into mechanisms that are going on. Another example is when a researcher wants to study processes that occur over long time periods, so long that repeated observations are difficult or impossible.

Models also have great potential when you want to study a process that does not yet exist in real life. An example is the impact of installing safety measures in a public location. To determine where and how much equipment needs to be installed it is useful to test the effect in advance by using hypothetical situations.

In these examples the researcher may get more insight into the process under investigation although the model may not be identical to the real-world process.

Agent-based modeling in criminological research

A nice introduction of the use of ABM in environmental criminology has been given in Birks (2018). Reasons for building agent-based models may differ, from practical predictions to theory development. When focusing on practical predictions, authors want to relate real-world data directly to their simulation model, for example for studying crime displacement in existing cities (e.g. Liu et al., 2005), with a goal to make actual predictions, e.g., on how the crime pattern would change if a certain bridge was closed for traffic. Theory-oriented scientists may deliberately abstract from empirical information, and use their simulation environment as an analytical tool for researching the consequences of theoretical assumptions, within a given or newly proposed theory (e.g., Bosse & Gerritsen, 2008). In this second perspective, researchers and policy makers use simulations as formalized thought experiments, to shed more light on the process under investigation. In this line of thinking, simulation is used as a method for investigating the structure of a theory, though it is of course possible, and to be hoped for, that the resulting insights may, sooner or later, be helpful in developing and improving existing policies (e.g., for surveillance) (Elffers & Van Baal, 2008). Some authors take an intermediate position (e.g. Bosse et al., 2010; Malleson & Brantingham, 2009) in which they initially build their simulation model to study the structure of a phenomenon, but they define their basic concepts in such a way that it can be connected to empirical information, if that becomes available.

So while more and more researchers within the criminological domain find their way to the use of ABM, with lots of beautiful research as a result (e.g., Birks, 2017; Groff et al., 2018; Liu & Eck, 2008; Malleson & Evans, 2013), ABM is currently mainly used for applied research. The field of testing and developing theories is still highly underdeveloped, with only a few publications illustrating the potential (e.g. Birks et al., 2012; Birks & Elffers, 2014). In the present volume all authors focus on this niche, highlighting the use of ABM for criminological theory testing and theory development.

In the next sections we will first discuss theory testing, and then enlarge the field towards theory development.

Criminological theory testing

How exactly can ABM help theory testing? Within empirical sciences the standard way of theory testing is the empirical cycle:

1.Formulate a theory covering a certain field.

2.Derive expectations from that theory: if the theory holds, what will be the case in such-and-such circumstances?

3.Design and execute a study that observes what is actually occurring in the specified circumstances.

4.Compare the observations with the expectations. If they are congruent, the theory is corroborated; if they are incompatible, the theory is rejected.

5.Then:

a. If a theory is corroborated, we can always go back to step 2 in order to test the theory again in a different set of circumstances.

b. If a theory has been rejected, we have to go back to step 1 in the cycle, and formulate an adapted version of the theory, or indeed replace the original one with a completely new one. With the adapted or new theory we can go through the cycle again.

We like to stress that, in a strict sense, theory testing comprises the first four steps of the cycle. Step 5a is in fact testing the same theory all over again, which presumably is most fruitful by deriving expectations in step 2 for a different set of circumstances. Step 5b, however, may be seen as theory development, informed by the results of a test performed in steps 1–4.

Deriving expectations

Deriving expectations from a theory is not always easy, due to the complexity of the theory and the complexity of the “circumstances” in which the results of that theory are expected to materialize.

ABM is useful when the complexity of theoretically founded expectations defies our analytical shrewdness. This is often the case when the theory is formulated on the level of actions of individual agents, while these actions are governed by the state of the world in its totality (i.e., the combination of the states of all other agents), which then is the aggregated result of all actions, aggregated over all agents. It is exactly this aggregation over sometimes many interacting agents that constitutes the complexity of the enterprise.

Let us present an example: imagine a criminal decision-making theory (within the routine activity paradigm) that claims: potential offenders aim to victimize only attractive targets and do so only when they will not be seen by passers-by (guardians).

Following the empirical cycle discussed above, let us try to derive expectations from that theory, in the special case of offenders aiming at home burglary. We try to establish what spatial distribution of burglaries we may expect according to the theory. First of all, we have to be more explicit on the content of the theory, in order to make progress (specification of formulated theory). In its present form, it is too gross to be able to derive what actions offenders take and how they exactly react to guardians. (Notice that such specification of a formulated theory is just as necessary when we try to test the theory by methods other than by ABM.) In our example, a possible specification of the theory may be as follows. Imagine a prospective burglar walking down a street, looking for an opportunity to burgle an attractive target; if the target he meets is not surpassing a minimum attraction threshold, he moves on to the next possible target, but as soon as he finds a target that surpasses a minimal threshold of attraction value, he looks around to check whether somebody might see him. If he perceives somebody in that street closer than a given distance, he decides not to burgle the given attractive house, as he is afraid of being seen by such a guardian. If, however, there is no guardian available within the given distance, he continues and burgles the selected target. We then need to be more explicit on the whereabouts of the guardians, e.g., like this: let people pass through the streets at a given pace, starting at arbitrary points, and at each crossroad taking an arbitrary direction.

Though the original theory was rather simple, the still rather straightforward specification makes the situation already pretty complicated; indeed, we think it is already prohibitively difficult to derive what spatial burglary pattern we may expect. That pattern will be dependent on the number of active burglars, how they move through space, their preference for attractive targets, the distribution of attraction values over space, the number of passers-by, their movement patterns, and most complexifying, on the dynamics of how burglars and guardians meet in the presence of attractive targets. We challenge the reader to bring forward an analytical solution to the question of what spatial pattern will be produced by the (specified) theory.

Many theories about human behavior meet the same difficulty, viz. that deriving exact expectations in given circumstances is very hard and often impossible. It is here that ABM simulation comes in handy. In fact, our specification above is already almost enough to construct an agent-based model, with offender agents and guardian agents, moving around on a street network of our choice, where the streets have houses (house agents) of various attraction levels. Indeed, such ABMs, in rather more complex situations, have been published (e.g., Birks & Davies, 2017; Bosse et al., 2010). Running such an ABM will indeed produce a spatial pattern, and given that the ABM has a number of non-deterministic processes on board, repeatedly running the model will provide us with a distribution of expected spatial patterns. Moreover, these results can be parameterized by some of the parameters of the specified theory, such as the minimal attraction level, the distance at which guardians are a threat, the number of offenders and guardians, the pace of guardians, the distribution of attraction values, and so on. All in all, using an ABM is a very powerful approach to deriving expectations from a theory, not analytically but constructively. As a side remark, building an ABM as an expression of a theory is also very helpful in the specification phase: how do we specify how the various agents react in various circumstances? Sometimes the theory is already explicit, but quite often we need to make a choice, which comes down to further specification of a theory. The necessity to be explicit on how agents behave, as a function of the state of other agents, helps in identifying where a theory needs specification.

Observation on what actually occurs in given circumstances

We have argued that ABM methodology can help in mastering the second step of the empirical cycle, producing expectations. Indeed, this is ABM’s strong point. But what about the third step, the observation phase, which is constituted by doing an empirical study and producing an empirical realization of the theory? Here we meet the weak point of ABM for theory testing. The observation stage should produce empirical results in the same circumstances that were present when the expectations were derived. But circumstances used in ABMs are generally of a highly abstract type, and reality, in which the empirical study can be situated, is not abstract, but real. For example, the study of Bosse et al. (2010), cited above, has an ABM in which burglar agents move over a chessboard, and guardians (police officer agents in this case) too. (This study was not a theory-testing study, but investigated, given routine activity theory, which police surveillance strategies were more or less effective in curbing crime.) We cannot observe “real behavior” of “real burglars” and “real police offers” on a chessboard. They, obviously, live and act in the real world. So, does the whole ABM approach to theory testing then run aground? We...

Table of contents

• Cover
• Endorsements
• Half Title
• Series Information
• Title Page
• Copyright Page
• Contents
• List of contributors
• Acknowledgments
• 1 Agent-based modeling for criminological theory testing and development
• 2 Generating crime generators
• 3 Using agent-based models to investigate the presence of edge effects around crime generators and attractors
• 4 Examining guardianship against theft
• 5 A simulation study into the generation of near repeat victimizations
• 6 Creating a temporal pattern for street robberies using ABM and data from a small city in South East Brazil
• 7 Corruption and the shadow of the future: A generalization of an ABM with repeated interactions
• 8 Agent-based modeling for testing and developing theories: What did we learn?
• Index