Copresence networks: criminals in their environment

Abstract

Cooffending networks, whose nodes are criminals and whose edges reflect when sets of criminals were involved in the same crime, have been well-studied. They inform the motivations for cooffending, for example, when criminal collaborators have particular skills; and the way in which criminality and risk spread. However, little is known about how cooffender networks intersect with the ordinary social networks that connect criminals to those around them: families, neighbors, and acquaintances. Such networks are important for law enforcement because these individuals are at risk because their associations with criminals, and because they may, perhaps unwittingly, play a role in criminal activities. Such networks are difficult to leverage because the data the require is hard to collect. Work by, among others, Papachristos has shown that the risk of gun violence is better understood from social-network position than from more conventional properties such as race, demographics, or socioeconomics. Given this, using social-network analysis provides a path away from broad-brush profiling based on such properties to a finer-grained assessment of risk based on properties that are demonstrably relevant. The innovation in this article is to define an alternative to laborious hand-gathered data about the social networks in which criminals are embedded: copresence networks whose nodes are those present at incidents attended by police, and whose edges count the number of times each pair of individuals are at the same incident. Within such networks, nodes can be divided between those who are charged (criminals), and those who are not (victims or witnesses). At first glance, victims and witnesses might seem to be a random subset of ordinary citizens whose presence at an incident with a perpetrator is happenstance. However, spectral embedding of copresence networks shows that the presence of victims and witnesses at incidents is not random; their rates of involvement are highly skewed. This suggests that they do not become involved in incidents by being at the wrong place at the wrong time, but because they are already at such a place and time. In other words, victims and witnesses are often already in the milieu of criminals. Thus, copresence networks approximate the intersection of criminal social networks with the wider networks in which criminals live, and they can be constructed from ordinary record management system data rather than in more intrusive ways. The contributions of this paper to applied policing are: (a) we show that cooffending networks are not freestanding and separate from the wider social connections of civilians, but rather are intertwined with them in complex ways; and (b) we show that the skew in activity already known for cooffending networks also applies for copresence networks. Hot spot policing techniques used for crime should leverage the extra connection information from copresence networks to, for example, understand connections between apparently unconnected criminal groups, and develop models for victim choice.