Leveraging police incident data for intelligence-led policing

Introduction

Police often claim that they, and their investigations, are intelligence led. By and large, this is hyperbole. Operationally, police investigations remain investigator-led: the investigator decides what to investigate and when, then provides direction for what information to collect in support of an investigation. This is especially tine of investigations once a crime has occurred where the focus is on the evidentiary value of information, rather than its intelligence value.

Strategically, police tend to have an impoverished understanding of the threat environment as it emerges and evolves because they do not have access to the requisite data, then data are unsystematic and/or incomplete, or they have limited capacity to analyse the data at their disposal to conduct a robust and reliable environmental and horizon scan. The objective of this article is to demonstrate the empirical value of applying advanced methodological techniques to a data set of incident records that virtually all police forces collect and retain: then- Records Management System (RMS). However, the findings also lead us to formulate a new theory built on an old idea: co-offenders and the ‘social networks’ induced from co-offending are well established in the literature (e.g., Papachristos, Braga, Piza, & Grossman, 2015); however, the findings here suggest an analogous theory of co-presence networks - including people who interact regularly with police, who do not necessarily end up being charged (witnesses and victims), but who nonetheless occupy positional advantage within the structure of the RMS incident data network.

We present an exploratory analysis of incident Record Management System data provided by Kingston Police. Kingston is a medium-sized city, with a population of about 125,000 and covering an area of 91 square kilometres, in Eastern Ontario, Canada. Kingston Police incident reports for 2015 were pulled from the RMS database where they are recorded. An incident is any interaction between the general population and police: this includes police attendance at crime scenes, but also civilian interactions at the police station (for example, to report a car accident) or by phone or online. In 2015, there were 46,668 incident records, and 188 attributes potentially associated with each record. Records contain information about the date, time, and place of each incident, the individuals concerned, and

Leveraging police incident data 55

specific aspects relevant to each kind of incident. Some attributes are numeric, some are times and dates, some are identifiers, and some are free-form text (addresses, officer comments, data about resolution).

At present, such data are commonly used in two ways. First, they may be queried in a straightforward way: have we had one of these previously? How many like this have happened? Second, they are analysed statistically: is there more of this kind of crime this year than last year? Are our response times acceptable, and are they improving? More subtle analysis has not, in general, been attempted because of the complexity of analysing semi-structured data where the attributes are so heterogeneous. The thrust of this chapter is that the inductive use of such data is much more powerful than conventional approaches. Instead of querying the data, clustering makes it possible to extract information from the data that indicates what general structures are present, and perhaps where there are unexpected local structures that suggest investigative targets. Many police forces are making their RMS data available on public sites (wwrv.policedatamitiative.org/ datasets/), which suggests that they believe that these data sets contain value. At the same time, making the data available implicitly conveys that they do not believe that they have the expertise to analyse them and extract then value inhouse. The data sets available online are also much less rich than the one on which this chapter is based, perhaps because police departments are concerned about privacy issues and de-anonymisation.

This article begins by surveying the state of the art for leveraging data to enhance policing. The second section shows how the order-preserving hashing technique can be applied to a data set. The results show that this technique makes it possible to ascertain, empirically, responses to key questions that are commonly raised about policing, such as law enforcement performance along dimensions such as lack of social bias, lack of regional bias, and expected and appropriate responses to the problems with which it is presented. One emergent property that may repay further study is the role of apartment buildings in the landscape of certain types of crime. There was also interest in whether calls could be categorised as mental health related in advance, so that appropriate dispatch responses could be deployed, and the results suggest that this is possible.

The subsequent section analyses social networks of co-presence at the same incident. The qualitative view of the social network of criminals - at least those who come to the attention of law enforcement - that emerges suggests an absence of highly organised groups with a strategic approach to their criminality. This too is a significant finding in terms of appropriate intervention and prevention strategies. The conclusion then sums up these findings as a function of the value-added by these two techniques to analysing police RMS incident data and gauges the broader validity of the findings.