The Structure of Trade-type and Governance-type Organized Crime Groups: A Network Study

Abstract

The paper provides a theoretical framework for categorizing organized crime groups based on what they do – whether they produce, trade or govern – as well as their aims. This paper then tests whether the internal structure of a heroin distribution organization in New York City, a Sicilian mafia group and the Provisional Irish Republican Army differ. Applying Exponential Random Graph Models (ERGMs) methods to network data, we find the organizational structure of trade-type organized crime differs markedly from governance-type, as well as between financially-motivated and politically-motivated groups. Trade-type organized crime and financially-motivated groups display a high level of centralization, an even distribution of clustering values, short paths and low homophily. Governance-type organized crime and politically-motivated groups display the opposite features. We conclude that the core activity and aim of the group are crucial in understanding the organizational structure.

The term ‘organized crime’ is often used to refer to a diverse range of illegal organizations, including mafias, ringing operations, insurgencies, pirates, price-fixing conspiracies, drug distribution networks, human smuggling operations and more. For the United Nations, an organized crime group is: ‘a group of three or more persons, existing for a period of time, acting in concert with the aim of committing at least one crime punishable by at least four years’ incarceration in order to obtain, directly or indirectly, a financial or other material benefit’ (UN 2004: art. 2). The British National Crime Agency adopts a definition that stresses the relative sophistication of the organization of the groups: ‘Organized crime can be defined as serious crime planned, coordinated and conducted by people working together on a continuing basis. Their motivation is often, but not always, financial gain’ (cit. in Campana and Varese, 2018: 1381). The US Organized Crime Control Act of 1970 also refers to the high level of disciple and organization: ‘[t]he unlawful activities of […] a highly organized, disciplined association’ (cit. in Varese 2022: 341). Such definitions—and many others—have come under severe criticisms by academics for being rather vague (see, e.g., van Duyne 2000: 369; Paoli 2002: 82; Finckenauer 2005: 64; Varese 2010: 17; for a recent discussion, see Bright and Whelan 2021). Indeed, some authors have concluded that a definition is not necessary (Levi 1998; van Dijk 2007; von Lampe 2015). And yet most designations to be found in statutes and academic writing refer to the ‘organization of criminals’ (Finckenauer 2005), namely the planning, the coordination among lawbreakers and the longevity of the group. In this view, ‘Organized crime is crime that is organized, often spanning several countries’ (Varese 2022: 341). Starting in the 21st century, academic writing on this topic has suggested that the organization has changed, from the (assumed to be) strict hierarchies of traditional mafias of the past to ‘self-organizing’, ‘decentralized,’ ‘flexible’ structural arrangements. These new forms of organized crime are lighter on their feet and quick to adjust to changing situations and opportunities (Morselli 2009: 11).

An alternative way to think about organized crime is to focus on activities. Such a perspective can be traced back to at least the Task Force Report (1967) and Smith (1975), suggesting that organized crime groups are illicit enterprises (Haller 2013). In this paper, we argue that an organized crime group’s structure depends on its aims and its line of business. To give empirical substance to this claim, we analyze a heroin distribution ring, a Cosa Nostra (Sicilian mafia) group and the Provisional Irish Republican Army. Evidence from descriptive measures and hypothesis testing using exponential random graph models (ERGMs) for network data supports this claim, suggesting that the groups engaged in trading differ from those who aspire to govern territories and markets, while groups with financial aims differ from those with political aims. We also posit that governance-type organized crime contains a continuum of cases, including at the extreme insurgencies. The paper offers empirical validation to the framework.

The paper is structured as follows: in Section 1, we lay out the framework of organized crime activities and aims and its theoretical implications, along with testable hypotheses about how activities and aims impact the structure of organized crime groups. Section 2 presents the three organized crime groups under consideration and the associated data sets being analyzed, along with a discussion of our methods and potential limitations. Section 3 describes the structure of the three organized crime groups, while Section 4 uses descriptive measures to investigate homophily within the networks; Section 5 contains the results of the exponential random graph models (ERGMs) for each of the organized crime groups, along with a discussion of the descriptive and model results. Section 6 concludes.

THEORY AND HYPOTHESES

While law enforcement agencies, policymakers, journalists and academics alike routinely apply the label ‘organized crime’ to a wide variety of groups, the concept requires to be specified beyond referring to a group of actors collaborating and organizing for the purpose of committing a crime. To address this, we adopt a framework for categorizing the different types of organized crime groups developed in Campana and Varese (2018) (see also Shortland and Varese 2016 and Varese 2017). The framework introduces three well-established concepts, ‘production’, ‘trade’ and ‘governance’, and applies them to organized crime groups. Some criminal groups are enterprises specialized in the production of illegal goods and services. For instance, some grow and refine drugs in Latin America, Afghanistan, the Golden Triangle (Zaitch 2002; Thoumi 2003; Paoli et al. 2009; Chin and Zhang 2015) and in developed economies such as the United Kingdom (Silverstone and Savage 2010; Decorte et al. 2011; Kirby and Peal 2015: 3). These groups are akin to small agricultural or industrial firms and workshops. Clearly, a number of factors have a significant impact on the structure of production, the (changing) technological requirements and the level of enforcement being surely important. In turn, these factors will affect the size of the firm, the level of capitalization and the size of inventories. We should also expect a degree of hierarchy within each firm and that producers have a tendency to have more ties within the group than across groups. Other criminal groups specialize instead in the transportation and trade of illegal goods. Illegal trades include human trafficking, human smuggling, drugs trafficking, trafficking in animals and the buying and selling of stolen data. The trade can take place in physical or virtual marketplaces. It is likely that these groups exhibit a low hierarchy and that traders have more ties with people performing other tasks than with other traders. Thirdly, there are organizations that provide extralegal governance, such as mafias and insurgencies. Governance is the ‘set of rules and norms that regulate exchange’ (Shortland and Varese 2016: 812). Governance-type organized crime groups seek to establish authority over a territory and to govern interactions and exchanges within that territory, in ways that are analytically similar to legitimate states. Governance activities by, e.g., mafias include employee and union intimidation, protection of property rights, loan retrieval, cartel agreement enforcement and dispute settlement.¹ Typically, mafias and insurgencies are organized as local hubs coordinated by a committee such as the mafia Commission. This network form of coordination is normally called ‘distributed’ (Baran 1964). Each component acts locally while being connected to a larger entity with a collective goal. The lower components (families or cells) are aware of each other yet operate independently.

One implication of this conceptual framework is that governance-type organized crime and states belong to the same category, which include some hybrids, such as insurgencies and paramilitary groups. In certain settings, paramilitary groups and insurgencies are the main players in the protection market: they levy taxes on the population living in their territories, carry out regular censuses and impose a specific ‘tax’ on producers of illegal goods such as drugs. Some insurgencies have even established a rudimentary judicial system and are involved in policing, infrastructure maintenance and the provision of public services like utilities and schools. Examples include the (now disbanded) Colombian FARC, the Burmese Wa Army (Varese 2017: 178–182), the Indonesian Free Aceh Movement, the Provisional Irish Republican Army in the 1970s (Gill et al. 2014) and the Islamic State in the cities of Mosul and Raqqa in 2013 and 2014 (Robinson et al. 2017). Governance implies attempts at monopolistic control of territories and markets, turf wars, control of local communities, engagement with politics and even having a state-building project. A key difference between these governance-type groups is their aim – while production, trade and governance groups can have any number of potential motives, we focus here on groups with financial aims and groups with political or ideological aims (Morselli et al. 2007). Financially-focused OCGs conduct their illegal operations with profit as their motivation, while politically-focused groups seek to enact their political or ideological goals. We posit that the aim of the group and the activity undertaken – production, trade or governance – determines the organizational structure of the group, hence the starting point of analytical definitions should be the former rather than the latter.

In this paper, for reasons of space, we focus on two types of groups, trade and governance. Surely, all organized crime groups operate in an environment of risk and uncertainty arising from their illegality. Yet they also differ in key ways. We conceive of trade-type organized crime, like drug trafficking and human smuggling, as typically transactional, with a commodity changing hands in a chain of exchanges where each actor only possesses the commodity for a brief period of time, is only activated momentarily and primarily receives payment as a one-time payout from involvement in a particular exchange (Morselli and Roy 2008). Furthermore, trade-type organized crime groups may be understood as action sets, ‘temporary coalitions formed to achieve specific goals which disperse once those goals are reached’ (Paoli 2004: 199). Instead of hierarchical, stable organizations, these action sets are more akin to flexible supply chains populated by criminal entrepreneurs (Zhang and Chin 2003). As these organized crime groups are not trying to exclude others and to set themselves up as monopolists in the supply of a given commodity, we expect that competition is great and that there are low barriers to entry for new criminal entrepreneurs. Since payoffs are short-term and the startup cost of involvement is low, we expect that criminals will be less worried about the longevity of the business, opting instead to structure their personal networks in a way that prioritizes short-term profits and efficiency (Morselli et al. 2007). This is especially true given the risk of arrest faced by the members of the criminal enterprise, incentivizing participants to secure their profit in the present since future payoffs may not materialize (Paoli 2004; Morselli and Roy 2008). The above characterization can be tested empirically through the lenses of Social Network Analysis measures. We can thus formulate the following hypotheses. We expect trade-type organized crime to exhibit:

Trade1: dense, centralized, clustered network structures, for the purpose of more effective, short-term coordination (Morselli et al. 2007; Hofmann and Gallupe 2015);

Trade2: short paths between nodes to increase the ease of information transmission through the network, and a tendency for clustering that further shortens path lengths (Morselli et al. 2007; Calderoni 2012; Hofmann and Gallupe 2015);

Trade3: low levels of demographic homophily, i.e. a tendency for nodes with different characteristics to form ties with one another (McPherson et al. 2001).

Governance-type organized crime tries to rule over portions of the underworld (Schelling 1971). This attempt requires high startup costs and assumes a longer time horizon than trade-type organized crime. A key investment is in violence. The group must be ready to face down challenges from other groups that question its authority, hence the concrete possibility of turf wars. In practice, violence is rarely used, since governance-type groups invest significantly in their reputation for violence, which allows them to economize on its actual use. Reputation building is a process that requires high startup costs, in the form of actually committing violence and a long time horizon, as the capacity for and willingness to commit violence must be demonstrated persistently for a reputation to form. Gathering relevant information on who does what in a given territory or market is also a crucial asset, and yet it takes time to develop information networks. Ultimately, the payout from governance-type organized crime typically materializes over a longer horizon than it does for trade-type organized crime. While trade-type organized crime provides returns as a one-time prize after an exchange, governance-type organized crime payoffs come in the form of multiple smaller prizes over a period of time, such as regular payments to mafia protection rackets or incremental political and social control of communities by insurgent groups. The higher startup costs and longer time horizon of governance-based business, then, encourage OCGs involved in these activities to prioritize security over efficiency, as detection would entail the end of operations before most of the rewards could be realized, wasting the initial investment of time and resources. Homophilic selection based on demographic characteristics like ethnicity, gender or social group membership act as a trustworthiness heuristic and so are expected to be more likely to occur in security-oriented networks: ‘[W]hen there is ambiguity and uncertainty in the environment, individuals seek to ameliorate it through the shortcut of relationships with similar others with the goal of creating homogenous, easily predicted networks’ (Gill et al. 2014: 56). Thus, we expect Governance-type organized crime to exhibit:

Gov1: decentralized network structures with low clustering;

Gov2: long path lengths between nodes that reduce the visibility that a given node has of other actors in the network (Baker and Faulkner 1993; Krebs 2001);

Gov3: high level of homophily based on demographic attributes.

A similar divide occurs for groups’ aims. Financial aims require a shorter time horizon for operations, as participants expect a monetary payout in a reasonable amount of time for their involvement and fear that longer-term profits might not materialize given the risk of the enterprise being shut down by law enforcement.

These pressures force OCGs to generate profits quickly and frequently, reducing their ability to prioritize far-off profits by emphasizing security in lieu of bolstering efficiency. Political or ideological aims, on the other hand, afford a longer time horizon, as participants are more likely to be satisfied with patiently waiting for the right moment to act rather than requiring immediate rewards, meaning that security is more important than efficiency (Morselli et al. 2007). Thus, we expect organized crime groups engaged in the same activity but with differing aims to exhibit:

Aims1: more centralized structures with more clustering for groups with financial aims;

Aims2: shorter path lengths between nodes for groups with financial aims;

Aims3: lower levels of homophily based on demographic attributes for groups with financial aims.

In the next section, we describe the methods we use to test the above hypotheses.

DATA AND METHODS

To test the hypotheses, we consider three different networks. First, we analyze a trade-focused organized crime group, a heroin distribution network of criminal entrepreneurs in New York City prosecuted between 1991 and 1993. The data set was collected by Natarajan (2006) by coding wiretap transcripts that were included in the prosecution files. Nodes represent actors, while edges represent calls between actors. The network consists of the group’s core members who had two or more contacts and were involved in five or more conversations (Natarajan 2006: 179). Attributes describing nodes’ gender and role were reconstructed from Natarajan (2006). Role attributes were determined through content analysis of the transcripts; the four roles identified were Seller, Retailer, Broker and Secretary.

Second, we consider a financially-focused enterprise run by Cosa Nostra, a governance-type organized crime group, involved in the rigging of public procurement contract auctions, a form of cartel agreement enforcement, from 2003 to 2007. The data set was collected by Cavallaro et al. (2020) from a pre-trial detention order from the Court of Messina’s preliminary investigation judge in 2007, a result of the anti-mafia ‘Montagna operation’ conducted by the Special Operations Unit of the Italian Police. The network is comprised of mafia members and mafia-affiliated entrepreneurs, with ties representing phone calls between criminals. The data set also includes node attribute data on role and mafia clan affiliation. Nodes were classified either as being affiliated with the Mistretta mafia family, the Batanesi mafia family, or another tangentially-involved mafia families or as being unaffiliated entrepreneurs who were not members of any mafia clan; nodes were further classified by roles as either bosses or foot-soldiers (or picciotti). Node attributes data were reconstructed from Ficara et al. (2020), which used the same data set.

Finally, we consider a politically-focused governance-type organized crime group, a network of Provisional Irish Republican Army (PIRA) members between 1977 and 1980. This period was chosen because it represents a time when the PIRA shifted its focus towards governance operations, as evinced by the creation of an auxiliary unit for policing in Catholic strongholds, the creation of new roles in the organization like ‘Education Officer’ and the promotion of the PIRA’s political party Sinn Fein (Gill et al. 2014: 56). The data set was compiled by Gill et al. (2014) using longitudinal and cross-sectional data collected by the International Center for the Study of Terrorism at Pennsylvania State University. Nodes in the network represent PIRA members with at least five connections to capture the core of the network, while edges represent at least one of the following relationships: (1) involvement in PIRA activities together, (2) friends before joining the PIRA, (3) blood relatives and (4) related through marriage. Although the relationships are of different types, the network only includes active PIRA members as nodes, excluding former members and IRA sympathizers who were not active in the group. The data set also includes node attribute data on demographic characteristics like gender, age at recruitment, marital status and university attendance, as well as data on node affiliation with PIRA brigades, operational units responsible for specific geographical combat areas. Further node attributes include whether an actor was involved in violent activities, what roles an actor held in the organization and what tasks an actor completed. Roles included Senior, Gunman, IED Constructor and IED Planter; tasks include Involvement in Foreign Operations and Involvement in Criminal Operations, such bank robbery, kidnapping, hijacking and drug smuggling. Missing data for the age at recruitment variable were imputed using the mean recruitment age for that node’s marital status.

In order to increase the comparability, we coded the edges as either 1 or 0 to denote the presence or absence of a tie (binary) and consider them undirected where applicable, meaning that if there is an edge from i to j there is also an edge from j to i. While each of these networks has been analyzed individually in their respective studies before, we apply new descriptive and statistical methods and approach the analysis from an explicitly comparative perspective. The Cavallaro et al. (2020) data set was provided by the authors in a Github repository (Cavallaro 2020), while the Natarajan (2006) and Gill et al. (2014) data sets were made available by the Covert Network repository compiled by the Mitchell Centre for Social Network Analysis at the University of Manchester (Mitchell Centre for Social Network Analysis 2016a, 2016b).

Social network data pose particular validity and reliability issues, especially data on illegal or secret organizations like organized crime groups. Two common issues are missing data and boundary specification. First, missing data are particularly troublesome for social networks research, even more so than for other quantitative methods. A small number of missing nodes or edges can substantially affect how a network is interpreted (Campana 2016; Borgatti et al. 2018); this issue is exacerbated when investigating organized crime groups because the secret and hidden nature of these groups increases the likelihood that data will be missing, as researchers seldom have prior knowledge about all of the relevant actors (Campana and Varese 2022). A second, related issue is that of boundary specification: when conducting social networks research, researchers must make choices about which nodes and ties to include in the data set and which to exclude. The aim is to define a boundary that matches the real-world structure under investigation; an excluded node or edge that ought to be included has the same effect as missing data, while the inclusion of nodes or edges that ought to be excluded can cause the network to fail to represent the social structure it seeks to reflect (Campana 2016; Borgatti et al. 2018). As a heuristic for network data on organized crime groups collected by law enforcement investigations, ‘one can accept the boundaries of an investigation as a stopping rule’ as long as the investigation has provided adequate coverage of the group (Campana and Varese 2022). The network data sets from Natarajan (2006), Cavallaro et al. (2020) and Gill et al. (2014) address these potential issues. All three networks were created using both primary and supplementary data sources to corroborate the data and reduce the likelihood of missing nodes or edges. Further, the boundaries set by the initial investigations from which the network data were derived, either law enforcement operations or academic research, should suffice, as initial data collectors ensured proper group coverage.

Other limitations should be mentioned and kept in mind when comparing the three networks. First, the edges in each network correspond to different types of social relationships. The PIRA network employs edges that represent social relationships, while in the other two networks they represent phone calls between actors. Second, the size of the heroin network is smaller than the that of the two other networks. Third, the PIRA network covers a greater geographic area than the other two networks under consideration. Finally, given the availability of different attributes that could be included in each ERGM model, we could include gender in the heroin network and PIRA network but not in the Cosa Nostra network, and the PIRA network contains many more attributes than available for the other two networks. Thus, some caution is warranted when assessing the importance of homophilous ties in each network.

To analyze the network structures of these three organized crime groups, we use a combination of descriptive methods and hypothesis testing. We present descriptive measures similar to those used by other studies, such as degree centrality, assortativity coefficients and minimum geodesic distance distributions (see Calderoni 2012; Krebs 2001; Morselli 2003; Morselli et al. 2007). While useful, they have limitations. First, while they can illuminate how a network is structured, descriptive measures do not allow researchers to determine which of multiple competing alternative explanations of those observed structures is at play in the network. Second, descriptive measures alone do not allow for inferences about whether certain network configurations or structures are more common in the observed network than would be expected by chance (Robins et al. 2007). More generally, network data pose a third challenge for analysis: since network data are inherently relational, such data violate the assumption of independence of observations required by standard statistical models (Campana 2016; Levy 2016a). As such, traditional statistical methods like OLS regression cannot be used to test network hypotheses. Instead, exponential random graph models (ERGMs) can be used to supplement descriptive analysis of networks and test hypotheses. The purpose of ERGMs is to ‘describe parsimoniously the local selection forces that shape the global structure of a network’ (Hunter et al. 2008b: 2). These models provide a means for investigating the tie formation process within a network, uncovering the mechanisms and preferences that lead nodes to form ties with each other. ERGMs are generative models that simulate networks and use Markov Chain Monte Carlo maximum likelihood estimation to estimate parameters, allowing for hypothesis testing and the derivation of significance values (see Robins et al. 2007; Hunter et al. 2008a; 2008b; Lusher, Koskinen, and Robins 2012 for technical discussions). Our approach, then, is to investigate the hypotheses using descriptive statistics, as well as to use descriptive statistics to guide predictions about the selection mechanisms at play in the tie formation processes of the three networks under consideration, which are then tested using ERGMs. It is important to note that the magnitudes of ERGM coefficients are not easily comparable across models based on different networks; as such, the discussion focuses on the sign and significance of the estimated parameters to establish the presence or absence of certain social processes within each network, rather than the size of their effect.

NETWORK FEATURES OF TRADE- AND GOVERNANCE-TYPE ORGANIZED CRIME GROUPS: CENTRALIZATION, CLUSTERING AND PATH LENGTH

This section considers the level of centralization, clustering and path length of the three networks, in order to test Hypotheses Trade1-2, Gov1-2 and Aims1-2. Visualizations of the heroin distribution, Cosa Nostra and PIRA networks are shown in Figure 1. Summary statistics for the three networks are presented in Table 1.

We first consider three common descriptive structural metrics. Degree centrality, as defined by Freeman (1979), measures how many edges a given node is a part of, based on the intuition that a node’s degree acts as an index of its popularity and importance in the network. Using degree centrality, Freeman (1979) further created a measure for centralization C that captures how equally degree is distributed among nodes in the network. C = 1 suggests that one node has the maximum degree possible while all other nodes have only one edge, meaning the network is maximally centralized; C = 0 suggests that all nodes in the network have the same degree, meaning the network is fully decentralized. Centralization measures for the three networks are presented in Table 2 below. In line with Hypotheses Trade1, Gov1 and Aims1 laid out in Section 1, C_HD is the highest, followed by C_CN, then C_PIRA. This suggests that the heroin network is the most centralized, with nodes’ degrees differing substantially and the PIRA network is the most decentralized, with nodes’ degrees being relatively similar. The financially-focused trade network exhibits the most centralized structure, while the politically-focused governance network exhibits the least centralized structure.

A node’s local clustering coefficient C(p), as defined by Watts and Strogatz (1998), measures the extent to which a node’s neighbors—those nodes with which it shares an edge—are also connected. C(p) = 1 means that all of node p’s neighbors are also neighbors, while C(p) = 0 means that none of node p’s neighbors share ties. The average local clustering coefficient $C¯$ measures the degree to which nodes in a network tend to cluster together. The average local clustering coefficients for the three networks are reported in Table 3 below. $C¯HD$ is the highest, followed by $C¯PIRA$⁠, then $C¯CN$⁠. These results confirm the expectation that the two governance networks would be relatively less clustered compared to the trade network; the average local clustering coefficient values, however, seem to suggest that there is greater clustering in the PIRA network than the Cosa Nostra network and that the two governance networks are dissimilar from each other.

Considering the distribution of C(p) for the three networks in Figure 2 provides a better understanding of these results. While $C¯CosaNostra$ and $C¯PIRA$ suggest different levels of clustering, the distributions for these networks are actually similar, with a primary peak around C(p) = 0 and a smaller secondary peak around C(p) = 1 and relatively few nodes in between; the difference between the average values is a result of the fact that there is a greater share of nodes near C(p) = 1 in the PIRA network than in the Cosa Nostra network. While this somewhat contradicts Hypothesis Aims1, the difference is potentially attributable to the differences in the underlying data sets: the PIRA network’s tie definition includes a number of social associations, while the Cosa Nostra network’s ties represent phone calls between actors. A more permissive tie definition is likely to tend towards greater clustering, as more edges are likely to be included; thus, the difference in clustering metrics could be attributable to differences in the data collection methods rather than differences in the networks. As such, the clustering piece of hypothesis Aims1 cannot be validated with descriptive statistics alone. The distribution for the heroin network, on the other hand, shows a more even distribution across the range of values. Thus, the two governance networks share similar distributions of local clustering coefficient values, suggesting a similar clustering trend in the two networks, while the trade network’s distribution is dissimilar, consistent with Hypotheses Trade1 and Gov1.

Lastly, we consider the geodesic distance distribution. A geodesic is the shortest path between two nodes in a network, meaning it contains the fewest edges of all possible paths between those two nodes. For example, a geodesic path consisting of two edges is said to have a length of two. The maximum geodesic length in a network is the network’s diameter. Geodesic path lengths and network diameters are used in Social Network Analysis to understand the ease or difficulty with which information or goods travel through the network; typically, a longer geodesic between two nodes will increase the difficulty with which information flows between those nodes. Further, Friedkin (1983) found that a geodesic distance of two represented a ‘horizon of observability’ in communication networks. Nodes for which the shortest path between them was two edges or fewer had visibility of each other’s attributes and actions, while nodes outside of this distance did not (Friedkin 1983; Krebs 2001). Figure 3 shows the distributions of geodesics lengths in the three organized crime group networks.

The PIRA network has the greatest diameter of length 16 and the lowest proportion of geodesics within the 2-length horizon of visibility at 18 per cent; the Cosa Nostra network has intermediate values for diameter (7) and proportion of geodesics within the horizon of visibility (23 per cent); the heroin network, on the other hand, has the shortest diameter of length 4—meaning the furthest-apart nodes in the network are only separated by 4 edges—and the largest proportion of geodesics within the horizon of visibility at 64 per cent. These structural findings support Hypotheses Gov2 and Aims2: the governance-focused PIRA network has long minimum path lengths. The trade-focused heroin network, on the other hand, has short minimum path lengths. The Cosa Nostra network has moderately-long geodesic lengths that lay between the values for the PIRA and heroin networks, as expected given the Cosa Nostra network’s governance activities and financial aim. Overall, these three structural descriptive measures support Hypotheses Trade1-2 and Gov1-2, as governance activities correspond with low centralization, low clustering and long path lengths, while trade activities correspond with high centralization, high clustering and short path lengths; further, Hypotheses Aims1-2 are largely supported, with financial aims corresponding with higher centralization and shorter path lengths than political aims.

HOMOPHILY IN TRADE- AND GOVERNANCE-TYPE ORGANIZED CRIME NETWORKS

In this section, we now test Hypotheses Trade3, Gov3 and Aims3. Similar to our investigation of structural metrics, we consider homophily in two ways: we first investigate the presence or absence of homophilous ties between nodes of the same attribute value within the observed networks and then use these descriptive results to provide predictions for the presence of homophily as a selection mechanism at play in the tie formation processes of the three networks under consideration; these predictions are tested using ERGMs in the next section alongside terms used to represent structural factors. This two-step approach is especially important when investigating homophily because the presence of homophilous ties in a network can arise even if the social mechanism of homophily—a preference for nodes to form ties with other nodes with similar attributes—were not at play in the tie formation process, instead resulting from another mechanism like triadic closure or reciprocity. Thus, we must investigate homophily as both a trend in observed edges and a tie formation mechanism. The assortativity coefficient for a given node attribute provides a means to investigate homophilous edges in a network and is interpreted like Pearson’s correlation coefficient. A value closer to 1 suggests assortative mixing, meaning that nodes of the same attribute value tend to form ties with one another, while a value closer to −1 suggests disassortative mixing, meaning that nodes of different attribute values tend to form ties. Table 4 shows the assortativity coefficients for node attributes in the three networks.

Overall, the heroin network shows slight heterophily by both gender and role, meaning that nodes tend to form ties with nodes that have different attributes, as expected by Hypotheses Trade3 and Aims3. An investigation of the counts of ties between the various roles shows that the only homophilous ties by role in the network are between Sellers; there are no ties between two Retailers, Brokers or Secretaries. We expect that the overall role heterophily and the Seller homophily will disappear in the ERGM once other structural and attribute control variables are added; further, we expect that gender will not play a role in the tie formation process. In line with Hypotheses Trade3 and Aims3, we expect that trade networks and financially-focused networks will exhibit less homophily than other networks. In line with Hypotheses Gov3 and Aims3, the Cosa Nostra network shows moderate heterophily by role and negligible homophily by clan, meaning that nodes with the same role tend not to form ties with one another. Finally, in line with Hypotheses Gov3 and Aims3, the PIRA network shows strong assortative mixing by brigade, participation in violence, participation in foreign operations, participation in violent foreign operations and the IED Planter role, along with moderate homophily by gender, marital status, participation in non-violence, participation in criminal operations and the Gunman role. As a decentralized network, these trends are expected to be confirmed by the ERGM analysis; in particular, tasks such as foreign and criminal operations and roles such as Gunman and IED Planter are expected to be one basis of homophily in the network, as these tasks are more complex and so are more likely to require collaboration between individuals with the same role. We expect that brigade membership-based homophily will persist in the model except for individuals who are unaffiliated with a brigade, as non-membership in a brigade is less likely to be a positive individual attribute over which actors could feel camaraderie and similarity; the same is expected for participation in nonviolence.

TIE FORMATION AMONG ACTORS IN TRADE- AND GOVERNANCE-TYPE ORGANIZED CRIME GROUPS: EXPONENTIAL RANDOM GRAPH MODEL RESULTS

Using the results presented above, we now specify a series of exponential random graph models. We present the results of the three best-fitting models for the heroin network, the Cosa Nostra network and the PIRA network in Tables 5–7 respectively. Four types of variables were included in these models. First, node factor (NF) and node covariate (NC) terms are used to include categorical and continuous node attributes like gender and age, capturing the likelihood of a node of a particular attribute forming ties in general; for example, women in a given network may tend to form more ties than men. Since the theory presented in Section 1 does not specify hypotheses about the likelihood of tie formation based upon node attributes, these are included as control variables. Second, node match (NM) and absolute difference (Absdiff) terms capture homophily for categorical and continuous variables. Finally, two structural terms are considered. GWDEGREE captures the level of degree centralization in the network, where negative coefficients suggest centralization and positive coefficients suggest decentralization (Levy 2016b). GWESP captures the tendency for the formation of closed triads, meaning that if A and B share a tie and B and C share a tie, then A and C share a tie as well. A positive GWESP coefficient suggests the presence of triadic closure processes in the network, while a negative coefficient suggests that there is less triadic closure and clustering than would be expected at random.² An important note is that, where possible, degree centralization (GWDEGREE) and triadic closure (GWESP) should be included together so that the similar processes of centralization and clustering can be disentangled, such that the centralization effect can be investigated while controlling for clustering and vice versa (Levy 2016b). Finally, an edges term is included in each model to capture the baseline likelihood of tie formation and acts similarly to the intercept term in linear and logistic regression models. Simulated network visualizations, Markov Chain Monte Carlo diagnostics and goodness-of-fit indicators are presented in the Supplementary Appendix ; all three tests suggest that the models presented in this section are well-behaved, non-degenerate and tended to capture the structural characteristics of the observed networks well.

Table 5 shows the results of the ERGM for the heroin network. This model includes node factor (NF) terms for the different roles in the network, as well as a node match (NM) term for the Seller role; the node match (NM) terms for the other roles were excluded as their coefficients were set to -Inf, as there were no same-role ties for Brokers, Retailers and Secretaries in the observed network. The node factor (NF) and node match (NM) terms for gender were excluded, as model fit increased when they were removed and neither term was significant when included, suggesting that gender had neither baseline nor homophilous effects. Looking at the significant parameters, the Broker, Retailer and Secretary node factor (NF) terms were both significant and negative when the Seller role was used as the base, suggesting that all roles were less likely to form ties than Sellers. The Seller node match (NM) term was significant and negative, suggesting that Sellers were less likely to form ties with each other than would be expected at random when controlling for the other included variables, as expected by Hypothesis Trade3. Finally, while both degree centralization (GWDEGREE) and triadic closure (GWESP) terms were included in the model, only the GWESP term was significant. The positive and significant triadic closure (GWESP) coefficient, along with the non-significant degree centralization (GWDEGREE) term, suggests the presence of a clustering mechanism within the network, even when controlling for centralization.

Table 6 shows the results of the ERGM for the Cosa Nostra network. This model includes node factor (NF) terms for both role and mafia clan affiliation. Looking at the significant coefficients, members of both the Batanesi and Mistretta families were significantly more likely to form ties than Unaffiliated nodes in the network. The clan node match (NM) terms were excluded from the model, as model fit increased when they were removed and neither coefficient was significant when included. The negative and significant coefficient of the role node match (NM) term suggests that members were less likely to form ties with members of the same role, in line with the descriptive statistics. Finally, both the degree centralization (GWDEGREE) and triadic closure (GWESP) coefficients were positive and significant, suggesting that the observed network exhibits greater clustering and less centralization than would be expected at random, even while controlling for node attributes and homophily; this network structure supports both the efficiency and the security of the group’s operations.

Table 7 shows the results of the ERGM for the PIRA network. This model includes node factor (NF) and node match (NM) terms for gender, university attendance, marital status, brigade membership, participation in violence and nonviolence, participation in violent foreign operations, role, involvement in foreign operations and involvement in criminal operations, as well as node covariate (NC) and absolute difference (Absdiff) terms for the continuous recruitment age variable. While few of the node factor (NF) and node covariate (NC) terms were significant, the model performed worse when they were excluded; as such, these variables were included in the final model. Considering homophily in the network, the descriptive results suggested that brigade membership, violence participation, marital status and involvement in collaboration-requiring tasks and roles like foreign operations, criminal operations, shooting and IED planting would be attributes that ordered homophilous tie formation. The node match (NM) term for marital status and the absolute difference (Absdiff) term for recruitment age were both significant and their signs suggest homophily on these attributes in the tie formation process. The coefficients for all of the brigade membership node match (NM) terms except for Unaffiliated nodes were positive and significant as expected; the Unaffiliated node match (NM) term was non-significant, as expected. The node match (NM) terms for violence participation, participation in violent foreign operations, involvement in foreign operations, involvement in criminal operations and the Gunman and IED Planter roles were all significant and positive as expected, while the node match (NM) term for participation in nonviolence was non-significant in line with expectations as well. Finally, both the degree centralization (GWDEGREE) and triadic closure (GWESP) coefficients were positive and significant; this suggests that the network tended towards both decentralization, which mitigates the risk of disruption and clustering, which enables more efficient communications and operations.

The results from the descriptive analysis and the ERGM hypothesis testing confirm the Hypotheses laid out in Section 1. While all three models had positive and significant triadic closure (GWESP) coefficients, the tendency for triadic closure captured by this term is common across many social networks. Since the magnitudes of the coefficients cannot be compared across models, the average local clustering coefficients $C¯$ and the distribution of local clustering coefficients C(p) provide a better metric for comparison, showing that the PIRA and Cosa Nostra networks had similar local clustering coefficient distributions with large numbers of highly- and lowly-clustered nodes, while the heroin network had a more even distribution of clustering values. The results from the investigation of the networks’ centralization values were confirmed in the models, as the Cosa Nostra and PIRA models both included significant degree centralization (GWDEGREE) terms and had positive coefficients, suggesting a decentralizing mechanism as expected based on the networks’ governance business (Levy 2016b). The descriptive results concerning homophily were also largely confirmed by the models: the PIRA network was the only one to see substantial amounts of homophily as a tie formation mechanism, while the Cosa Nostra network displayed heterophily by role. Further, the more complex tasks and roles in the PIRA network in particular exhibited homophily, with actors of the same role or task more likely to form ties with one another to complete operations that would be difficult to conduct alone. The homophily results for the PIRA network also help explain how decentralized networks can still conduct governance business and enforce a set of rules: nodes tended to cluster with nodes in the same brigade, creating a ‘distributed’ network as described by Baran (1964) consisting of local intra-connected cells that independently control a given territory, but are connected to the central organization such that they can receive overarching norms and rules. Finally, for the PIRA network, homophily on traits like non-affiliation with a clan or participation in nonviolence was not present as a tie formation mechanism, as expected. The results from the prior two sections provide support for the idea that the activities and aims of organized crime groups affect their network structures in predictable ways, and support Hypotheses Trade1-3, Gov1-3 and Aims1-3.

CONCLUSIONS

Developing an understanding of how organized crime groups are structured is a crucial undertaking. The literature on the determinants of crime network structures has focused on how the conditions of risk and uncertainty that all crime groups face as a result of their illicit nature impact their structures (Morselli et al. 2007). The literature has focused on the efficiency/security trade off as the only factor affecting the structure of the network (for a review, see Bright and Whelan 2021: 60–62). In this paper, we consider the different activities and aims undertaken by organized crime, using a theoretical framework for categorizing different organized crime groups and novel statistical methods to expand this literature. In particular, we focus on how the aims of and the type of activities in which an organized crime group is engaged influence its network structure. The transactional nature of trade means that organized crime groups engaged in this business will face low startup costs and a short time horizon for return-on-investment, while the fact that governance organized crime groups’ chief assets are their reputation for violence entails high startup costs and longer-term thinking. Organizations operating for profit will be pressured towards shorter-term thinking by their members who expect a timely payout, while the members of groups focusing on political gains or ideological aims are more likely to accept a longer-term view of success and payoff. Organized crime groups with shorter time horizons as a result of their business or aim are more likely to emphasize quick coordination among diverse people in their network structures; for governance-type organized crime groups that seek to control communities or politically-motivated groups that seek wider-scale change, security is paramount for ensuring that the long-term goal is reached before the group is detected and disrupted. Considering the networks of a heroin distribution ring, a mafia-led public contract bid rigging enterprise and a governance-focused insurgent group, both descriptive and statistical results suggest that the theoretical framework is valid and the Hypotheses Trade1-3, Gov1-3 and Aims1-3 supported. Governance organizations and groups with political aims form sparse, decentralized networks with homophilous ties and long path lengths between actors. Trade organizations and groups with financial aims, on the other hand, show denser, more centralized network structures where ties are less determined by trustworthiness heuristics like homophily and are more dependent on creating network configurations that allow for quick coordination and rapid spread of information through the network.

While promising, there are several limitations to the research conducted in this paper. First, the three networks considered all consist of different types of social relations: edges in the heroin distribution and Cosa Nostra networks encode calls between actors, while ties in the PIRA network consist of a number of social associations. There is a possibility that a portion of the findings could be attributable to differences in how these types of networks are structured, not differences between the organizations themselves. The likely impact of this difference is limited, however. As mentioned in Section 3, the more permissive edge definition in the PIRA network is likely to tend towards producing a denser, more connected and more transitive network than a more stringent definition, as more edges would be included; but to the contrary, the PIRA network is sparser than the HD and CN networks and less transitive than the HD network, while exhibiting similar clustering trends to the CN network. Second, each of the specific organized crime groups considered operated at different times, in different places and within different cultural contexts, all of which could potentially impact network structure in ways not attributable to the groups’ businesses and aims. Further, the data sources were of varying richness, with the PIRA network data set containing many more node attributes than the heroin distribution and Cosa Nostra data sets, meaning that the evidence of homophily in the PIRA network but not in the others could be attributable to the differing levels of data available. Finally, while rich and representative of the group during its shift towards governance activities, the PIRA network data set may not be an ideal match for the research question given its emphasis on crime-focused roles like Gunman and IED Planter, rather than governance-focused roles like community policing. Despite these limitations, the results of the analysis conducted on these networks still provide a promising start to the study of how organized crime groups’ businesses and aims affect how they respond to risk and uncertainty in their network structures.

The results in this paper are significant in that they provide support for the theoretical framework presented above (Shortland and Varese 2016; Campana and Varese 2018), as well as the importance of an organized crime group’s aims and motivations (Morselli et al. 2007). Rather than conceiving of organized crime as a homogenous category, groups are better understood as distinct organizations engaged in a variety of markets with a variety of goals that impact groups’ structures, operations and methods. While organized crime groups face risk and uncertainty as a result of the environment in which they operate, how they respond to these conditions will be shaped by their core activity and goals.

ACKNOWLEDGEMENTS

We are grateful to Paolo Campana and Heather Hamill for their comments on an earlier version of this paper. The results of this paper were presented during a seminar for the Extra-Legal Governance Institute based at the Department of Sociology, University of Oxford. The authors are listed in alphabetic order.

FUNDING

This work was partly supported by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant agreement No. 101020598 – CRIMGOV).

REFERENCES

Footnotes