1.1 Marriage as an instrument of organizational rationality in ’Ndrangheta

Some authors argue that marriage is used strategically by mafia organizations [e.g., 18–20] and, indeed, “is always strategic, never the culmination of genuine passion” [21, 125]. In this regard, the ’Ndrangheta represents a distinctive case, due to its unique organizational structure. In the Sicilian Cosa Nostra, for instance, the number of relatives within the same organizational unit (the famiglia) is intentionally limited to prevent the formation of alternative, informal centers of power that might prioritize the interests of a minority over those of the famiglia as a whole [22]. By contrast, the basic organizational unit of the ’Ndrangheta, the ’ndrina, is a genuinely patriarchal, hierarchical, blood-based family, in which the sons of “men of honor” (i.e., members) are considered young men of honor from birth, even though the ritual of “baptism” that formalizes their membership takes place later in their lives [23,24].

Because the ’Ndrangheta recruits primarily through kinship, marriages function as a strategic tool employed by blood families to seal alliances with other ’Ndrangheta families. Such alliances aim to consolidate power within the organization, ease conflicts with rival groups, access new business opportunities, and/or expand territorial control [18].

In these matrimonial strategies, women are treated as objects of exchange and “given away” to cement such agreements [18], a logic that reflects the tendency of mafia organizations to relegate women to passive and subordinate roles [25]. A striking case is that of Giulia Immacolata, a woman born in the ’Ndrangheta family Coluccio. In 2014, at the age of just thirteen, she was forced by her parents to become engaged to Cosimo Commisso, nephew of the powerful boss Vincenzo Macrì, to enable her family to gain access to the Commissos’ drug trafficking business [26, 1115]. This instrumental role of women reflects the structure of the ’Ndrangheta, described as “an authoritarian organization [...] with unwritten laws, traditional behavior, and interpersonal relationships explicitly shaped by powerful patriarchal family traditions” [19,39].

Beyond material considerations, endogamous marriages also serve important cultural and security functions. Matrimonial ties within the ’Ndrangheta preserve the organization’s values across generations, transmitting cultural norms like loyalty, silence (omertà), and honor. Historically, women have been central to this process, safeguarding family reputation and reinforcing the ’Ndrangheta’s identity by instilling core values in younger generations [23,27]. In this respect, marriages also contribute to secrecy and, consequently, enhance organizational security. Indeed, the ’Ndrangheta has the lowest number of collaborators with justice among Italian mafias [22]. These data can be partly explained by the fact that collaboration with law enforcement agencies entails severing ties not only with the organization, but also with one’s own relatives. The cost of such a decision for a member of the ’Ndrangheta is well illustrated by the case of Emanuele Mancuso, a prominent member of a powerful ’Ndrangheta family. To deter him from cooperating with law enforcement, both his mother and his child’s mother used his newborn child as emotional leverage, while his aunt applied psychological pressure by appealing to his attachment to his parents: “Come on, your father and mother are your family... how is your mother doing? She’s not well! She knows she no longer has a son, how do you think she feels?” [28].

In sum, matrimonial unions, by creating blood ties, ultimately function as a mechanism for fostering trust and cooperation through personal relationships [29]. This aspect is particularly important in the case of the ’Ndrangheta because, as an illegal organization, its members cannot rely on third parties or formal institutions to regulate transactions and resolve disputes. Nor can they readily resort to communication and violence as formal control mechanisms, since both increase the risk of exposure to law enforcement agencies [24,30].

Marriage, as an organizational mechanism for cultural reproduction, resource acquisition, and security enhancement, is therefore a key feature of the ’Ndrangheta. We argue that these practices deserve more in-depth scholarly attention, as they may help explain the organizational puzzle posed by the ’Ndrangheta. As discussed, ’Ndrangheta is the only mafia organization that adopts a kinship-based recruitment strategy, a strategy that typically creates an organizational dilemma [22]. On the one hand, recruiting through blood ties helps overcome information asymmetries about candidates, ensuring the entry of well-socialized and trustworthy members, thereby enhancing organizational security. On the other hand, this approach does not necessarily guarantee the recruitment of individuals with the skills needed by the organization. As a result, kinship-based recruitment may lead to a lower level of human capital, potentially constraining the organization’s capacity to manage sophisticated criminal or business activities.

However, this does not appear to be the case for the ’Ndrangheta, which has managed to reconcile a kinship-based recruitment model with organizational and economic success, and is today often described by law enforcement agencies as one of the most dangerous organized crime groups worldwide [31,32]. Organizationally, the ’Ndrangheta emerged in Calabria around the mid-19^th century and has since managed to reproduce its structure across several Italian regions [e.g., 22,20] and abroad [e.g., 33–36]. Economically, according to Italian law enforcement agencies, it exhibits a “persistent business-oriented and entrepreneurial vocation” and an “extraordinary tactical versatility” that have no equivalent among other traditional mafia organizations [10,37].

We explore this apparent paradox by analyzing the organizational logic of marriage in the ’Ndrangheta, with particular attention to the role of women and to differences in marriage strategies between powerful and less powerful families. While previous work highlighted the strategic use of marriage among ’Ndrangheta families, much of the existing discussion remains qualitative and largely reliant on biographical sources [e.g., 38], with large-scale quantitative studies still being rare [e.g., 39]. In this respect, the literature still lacks organization-wide, quantitative evidence linking marriage ties to actual power and organization’s cohesion. Specifically, there is little analysis of whether and how the direction of gender exchange—who “gives” versus who “receives” brides—relates to power and network structure.

We contribute to this body of research by examining matrimonial patterns within the ’Ndrangheta from a social network perspective, an approach that, notwithstanding the methodological challenges [see 40,41], has proven to be fruitful in the study of criminal and mafia organizations [e.g., 41–50].

Accordingly, we test whether reputed powerful families sit at the structural core of the ’Ndrangheta matrimonial network, whether the network would fragment faster if their marriages—or instead peripheral marriages—were removed, and whether a clan’s share of incoming brides predicts its power and centrality. These tests examine both status versus structure and gender exchange versus power in a way that the literature has not yet systematically attempted.

As such, this article provides the first large-scale empirical investigation of marriage strategies within the ’Ndrangheta, drawing on an extensive dataset that covers hundreds of families and marriages across years. Our analysis reveals the structural centrality of powerful families within the matrimonial network, demonstrating their critical role in maintaining cohesion and influence. At the same time, the findings highlight the complexity underlying the relationship between family power, network centrality, and gender dynamics, challenging simplistic narratives about bride exchange.

The remainder of the article is organized as follows. Section 2 sets out data and methods: we describe the judicial sources underlying the 623–family marriage network, define our proxies for clan “power” and detail the permutation tests, edge-removal simulations, and sliding-window analyses of gender bias. Section 3 presents the empirical results in three parts—status versus structure, network resilience, and gendered exchange—supplemented by robustness checks. Section 4 discusses the implications of these findings for understanding how the ’Ndrangheta balances lineage consolidation with alliance-building, and flags limitations and future research directions. Finally, Section 5 concludes by summarising the main contributions that follow from viewing marriage as an organizational technology.

2.1 Sources, data, and proxies

2.1.2 Proxies for power.

Power in kinship-based criminal organizations such as the ’Ndrangheta is neither directly observable nor reducible to a single structural indicator. We therefore adopt a multidimensional approach, combining externally validated measures of clan status with a battery of complementary network-based centrality metrics. Specifically, we rely on an externally sourced classification (described in detail below) identifying powerful clans based on judicial, journalistic, and investigative assessments of their criminal prominence. We then operationalize clan power structurally by computing centrality measures derived from the marriage network, capturing relational dimensions of influence such as direct connectedness, brokerage potential, reachability, and prestige.

Taken together, these measures allow us to assess the degree of alignment between externally perceived status and internal network prominence. While centrality metrics effectively reflect relational aspects of power, we acknowledge that structural position alone does not exhaustively capture all dimensions of clan power—such as economic wealth, coercive capacity, or political influence. Future research could fruitfully integrate additional indicators to enrich our structural perspective.

Centrality Measures To operationalise structural power in the marriage network, we compute four normalised centrality indices, each tapping a distinct facet of prominence (Fig 1). Centrality is a classic proxy for power in network theory, where influence stems from position rather than intrinsic attributes [60–62]. Network topology shapes opportunities for control, brokerage and information access; centrality metrics translate those positional advantages into measurable scores [63,64].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Distribution of normalized centrality scores in the largest component (n = 623).

Ridgeline density plots of Degree, closeness, betweenness and eigenvector centralities (textbook-normalized) reveal starkly different shapes: over 50% of clans have zero betweenness and very low degree (medians at 0.002), closeness is more symmetrically distributed (median = 0.155), and only one clan achieves the maximum eigenvector score of 1. The horizontal axis is zoomed to [0, 0.3] to focus on the bulk of the distribution.

https://doi.org/10.1371/journal.pone.0345859.g001

The first metric, normalised degree centrality, captures a clan’s direct matrimonial ties. A high value signals dense immediate contacts and, hence, rapid access to resources and coordination [61]. Second, normalised closeness centrality reflects how quickly a clan can reach all others in the network. Families with high closeness require few intermediaries and can disseminate or gather information efficiently [65]. Third, normalised betweenness centrality gauges brokerage potential: it is the share of all shortest undirected paths between every pair of clans that pass through the focal clan. Under the standard normalisation for undirected graphs, the denominator (n−1)(n−2)/2 converts raw counts into a 0–1 scale [66,67]. Finally, normalised eigenvector centrality rewards connections to already well-connected families [63]. Let x be the principal eigenvector of the adjacency matrix A such that

(1)

We rescale

(2)

so that the most central clan attains 1. This formulation captures the idea that alliances with prestigious partners are associated with additional standing [68].

Status To complement network-derived measures of power, we incorporate an external, network-independent ranking of the twenty most influential ’Ndrangheta families identified by the investigative programme La C Play [69]. Based on wiretaps, interrogations, video footage, investigative reporting, interviews, and court-anchored inquiries [70], the list reflects qualitative, judicially grounded assessments rather than structural metrics. While institutionally credible, such visibility may also be shaped by media coverage, prosecutorial focus, or clan communication strategies.

We classify every family named in La C Play as a “Powerful Clan.” Two entries in the source combined surnames—Barbaro–Papalia and Farao–Marincola—which we disaggregate, resulting in 22 distinct clans. Matching these against our network yielded 17 present (Crea, Tegano, Gallace, Nirta, Arena, Giampà, Molè, Morabito, Grande Aracri, Alvaro, Commisso, Barbaro, Papalia, Bellocco, Mancuso, De Stefano, Piromalli) and five absent (Lanzino, Farao, Marincola, Libri, Pesce). The absent clans likely lie beyond the empirical boundaries of our dataset. Families that appear in both sources are coded as “Powerful Clans,” while every other observed family is assigned to the “Other Clans” category (Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. ‘Ndrangheta matrimonial alliance network: structure, core, and tie composition.

(A) Undirected network of inter-clan matrimonial alliances (largest connected component); nodes represent families and node size is proportional to eigenvector centrality; red nodes indicate clans classified as powerful (per the La C Play classification), blue nodes other clans; labels show powerful clans. (B) k-core backbone of the same network (upper quartile coreness), using the same encoding for node size and clan status. (C) Composition of unique alliances by clan status: powerful–powerful (red), powerful–other / other–powerful (purple), and other–other (blue).

https://doi.org/10.1371/journal.pone.0345859.g002

To further validate this classification, we cross-referenced each surname from La C Play with the Anti-Mafia Investigation Directorate’s (Direzione Investigativa Antimafia, DIA) Semi-Annual Report (I Semester 2023) [37], analyzing the main ’Ndrangheta chapter, regional breakdowns, and annexes detailing investigations and asset seizures. Nearly all surnames appear with varying degrees of prominence, ranging from extensive coverage (e.g., Mancuso, De Stefano, Piromalli, Libri) to more marginal or incidental mentions (e.g., Crea, Giampà, Papalia). Only Lanzino is not cited. This broad alignment reinforces the list’s consistency with official threat assessments.

2.1.3 Bride–groom exchange directionality.

Gender asymmetry in matrimonial strategies is measured using information from the event-level directed marriage graph, in which each individual marriage constitutes a directed tie from the bride-giving clan to the bride-receiving clan. All other analyses employ a simplified undirected counterpart in which repeated intermarriages between the same clans are collapsed into a single alliance.

Specifically, gender asymmetry in alliance building is captured with one normalized statistic per family, the bride-receiving share. For clan i let in-degree_i be the number of daughters-in-law (women who marry into the lineage) and out-degree_i the number of daughters the clan gives in marriage. We define

(3)

Lower values of bride-receiving share correspond to families that predominantly send out daughters, while higher values indicate a bias toward receiving brides. The measure is bounded in [0,1]. A value of 0.5 denotes gender-balanced exchange, whereas values above (below) 0.5 indicate a bias toward receiving brides—respectively consistent with consolidating patrilineal capital or forging external ties. This normalisation follows established practice in directional network analysis [64] and is computed on the loop-pruned graph, so consanguineous marriages do not influence the ratio.

The distribution of bride-receiving shares across ’Ndrangheta clans reveals a polarized marital economy with peaks at both extremes of the spectrum (Fig 3). The median is 0.50, while the mean is 0.44, indicating a left-skew driven by a substantial mass of clans with low bride-receiving shares.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of bride-receiving shares across ’Ndrangheta families.

Histogram showing counts of 623 families in the largest weakly connected component. Bin ranges: [0.0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), [0.8, 1.0]. Bimodal peaks occur at distribution extremes.

https://doi.org/10.1371/journal.pone.0345859.g003

2.2 Analytical strategy

Our empirical strategy proceeds in four steps. First, we assess whether the clans identified by investigative sources as powerful occupy structurally advantaged positions in the matrimonial alliance network. Second, we examine whether gender-biased patterns of bride exchange are associated with structural influence in the same network.

First, to determine if clan status is linked to network influence, we analyze our four key centrality metrics—degree, eigenvector, betweenness, and closeness—within the network’s largest connected component. We compare these scores between ’powerful’ and ’non-powerful’ clans using non-parametric permutation tests (subsection 3.1). By randomly reshuffling the ’powerful’ labels while keeping the network structure constant, we establish a rigorous benchmark to verify if the observed differences in social prominence are statistically significant or coincidental.

Second, having established whether powerful clans occupy systematically different positions in the alliance structure, we then shift from node-level prominence to edge-level structural contribution. In particular, we assess whether alliances involving powerful clans disproportionately sustain the cohesion of the overall network. If such ties play a central integrative role, removing them should fragment the network more rapidly than removing ties that do not involve powerful clans. To test the network’s resilience, we compare the impact of removing edges connected to powerful clans against removing edges between non-powerful clans. We track the resulting loss of cohesion using several complementary indicators (see Appendix A).

Then, third, to study gendered alliance strategies, we compute each clan’s bride-receiving share, defined as the proportion of its matrimonial ties in which it occupies the bride-receiving role. This measure captures whether a clan tends, on balance, to receive brides into its lineage or to send brides outward through marriage ties. We begin by testing whether clans considered to be powerful display systematically different bride-receiving patterns than other clans (subsection 3.3). Given the modest sample size and non-normal distribution of the data, we opted for non-parametric permutation tests instead of parametric t-tests, as the former rely solely on the exchangeability of labels under the null hypothesis. This approach is suitable for network data, where relational dependencies violate standard assumptions of independence and normality, ensuring robust inference without distributional assumptions [60,71–73].

Fourth, we finally turn to the question of how gendered marriage orientation relates to structural prominence in the alliance network (subsection 3.4). Rather than imposing a single global functional form, we estimate a series of local contrasts along the bride-receiving share spectrum, indexed by c, comparing clans at each value of c to a neutral benchmark centered around parity (c ≈ 0.50). This approach reveals how network centrality—measured via normalized degree, closeness, betweenness, and eigenvector centrality—varies as clans transition from predominantly bride-sending to predominantly bride-receiving roles.

For example, a positive contrast at c = 0.60 indicates that clans with a 60% bride-receiving share are, on average, more central than parity-oriented clans, whereas a negative contrast at c = 0.90 indicates lower centrality relative to that benchmark. To ensure that these patterns do not hinge on how narrowly “parity” is defined or on other arbitrary modelling choices, we compute contrasts using both a strict neutrality band () and a relaxed band (), and replicate the full exercise under a range of alternative specifications and sample restrictions (see Appendix C).

In performing our analyses, for gender-asymmetry indicators such as the bride-receiving share we retain every individual marriage, so the directed graph may contain parallel edges whenever two clans inter-marry more than once. In contrast, all centrality estimates and the resilience simulations operate on a simplified, undirected projection in which at most one edge connects any pair of clans. This distinction ensures that alliance frequency shapes the bride-receiving metric but does not overweight the same clans in the structural-power analyses.

3.1 Status and structural prominence

The clans singled out by journalists and prosecutors as powerful occupy the most advantageous positions in the matrimonial alliance network, as confirmed by 5,000-replicate permutation tests (Table 1).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Network centrality by clan status (5,000 permutation tests).

https://doi.org/10.1371/journal.pone.0345859.t001

Powerful families maintain roughly triple the normalized degree of other clans (, 95% CI [0.006, 0.010], p < 0.001), meaning they are directly connected to 1.2% of other families on average versus 0.4 % for non-powerful clans. Powerful clans also exhibit significantly higher closeness centrality (, 95 % CI [0.012, 0.036], p < 0.001), indicating more efficient network access. Likewise, powerful clans command a larger share of shortest paths via betweenness (, 95 % CI [0.022, 0.043], p < 0.001), underscoring their brokerage role, and enjoy a strong prestige advantage in eigenvector centrality (, 95 % CI [0.075, 0.156], p < 0.001). Together, these permutation-based contrasts demonstrate that external reputational authority aligns with internal structural prominence within the marriage network. While these status-based advantages are clear, we further examine the robustness of this relationship across the full range of matrimonial strategies in Appendix C.

3.2 Network cohesion and load-bearing ties

While centrality metrics identify powerful families as topologically privileged actors, they do not establish whether these clans’ alliances are functionally critical for network cohesion. To evaluate the load-bearing role of powerful family ties, we conducted simulations systematically removing edges under two protocols: proportional and absolute deletion regimes (Fig 4).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Resilience of the ’Ndrangheta matrimonial network under two edge-removal schemes.

Monte-Carlo simulations (100 replicates) in which edges are deleted either by equal-share (left column) or equal-count (right column) schemes, separately for edges involving powerful clans (red) and those not involving powerful clans (blue). Rows: (A, B) Connectivity—size of the largest connected component as a share of all families; (C, D) Reachability density—the proportion of unordered family pairs that remain connected by some undirected path; (E, F) Largest bicomponent—size of the largest biconnected component as a share of all families; (G, H) Average path length (giant)—average shortest-path length within the largest connected component. Columns: (left) Equal-share removal: at each step a fixed fraction of that edge category is removed (x-axis = fraction removed within the category); (right) Equal-count removal: at each step the same absolute number of deletions is applied to each category (x-axis = number of edges removed, up to the maximum number of powerful-involving ties). Shaded ribbons show the 10^th – 90^th percentile range across replicates.

https://doi.org/10.1371/journal.pone.0345859.g004

Edge betweenness further supports this distinction between prominence and load-bearing importance. While alliances involving powerful clans exhibit higher average edge betweenness (mean = 0.0115; median = 0.0065) than those among non-powerful clans (mean = 0.0082; median = 0.0032), the most extreme brokerage ties are not necessarily powerful family-linked (max = 0.0804 among non-powerful family alliances). This pattern suggests that powerful families tend to participate in broadly central connections, yet a smaller subset of peripheral marriages can occupy key bridging positions.

Peripheral alliances emerge as critical cohesion anchors. Under equal-proportion removal, eliminating 50% of ordinary marriages reduces reachability by 40%, nearly triple the 15% decline when removing powerful clan ties. This disproportionate fragility persists under equal-count removal: the first 100 ordinary deletions degrade connectivity twice as severely as deletions of powerful-clan ties.

The gradual decay of powerful clan-linked ties signals structural redundancy: powerful clans cultivate multiple, overlapping alliance channels that cushion the impact of losing any one connection. In contrast, the rapid erosion of non-powerful clan ties underscores their role as singular bridges that knit together otherwise distant parts of the network. This divergence highlights how redundancy protects powerful families. Because powerful clans occupy highly central positions and maintain numerous alliances with non-powerful families, they form a structurally redundant backbone of the matrimonial network, even though direct marriages among powerful clans themselves are rare. On the other hand, non-powerful clan unions often span structural bottlenecks; when alliances between less prominent families dissolve, the broader network splits much more quickly.

Largest bicomponent size and average path length within the giant component refine the interpretation of the removal exercise by moving beyond whether clans remain connected to how they remain connected. When edges not involving powerful clans are removed, the network rapidly loses its biconnected core—indicating that cohesion increasingly depends on single “bridge” ties whose removal would fragment the system—and the surviving giant component becomes progressively less efficient, with longer paths between families. By contrast, removing edges involving powerful clans produces a slower contraction of the biconnected backbone and a more contained increase in path length, consistent with a core that is structurally redundant: powerful clans are embedded in overlapping alliance circuits, so the loss of any single powerful clan-linked tie is more easily absorbed without undermining either robust cohesion or internal reachability.

This result has potential operational implications: while powerful clans occupy structurally prominent positions, network integrity and cohesion depend critically on alliances between non-powerful families. Disruption strategies targeting these inter-peripheral alliances would induce faster organizational fragmentation than direct attacks on core families.

3.3 Gender directionality and status

Powerful clans receive a greater share of brides (mean = 0.559) than other clans (mean = 0.437), a difference of +0.122 (Table 2). Although this suggests a tendency for high-status families to act as net receivers, the effect is not statistically significant (95% CI [–0.076, 0.312]; p = 0.225). The conclusion is unchanged when we additionally control for clans’ overall matrimonial activity by permuting status labels within strata of total marriages (degree-stratified permutation; p = 0.294). Consequently, while the point estimate favors powerful clans, the evidence remains inconclusive.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Permutation test of bride-receiving share by clan status.

https://doi.org/10.1371/journal.pone.0345859.t002

3.4 Gender directionality and network position

Building on this gender-strategy insight, we next examine how variation in bride-receiving share aligns with network position. Across all measures, structural centrality displays a pronounced inverted-U pattern over c, peaking between c ≈ 0.55 and 0.60. This pattern suggests that while parity constitutes a stable baseline, clans that lean modestly toward a bride-receiving orientation occupy the most advantageous positions in the alliance network.

Under the strict neutrality definition (), the strongest and most consistent positive contrasts occur at c = 0.55 for degree, betweenness, and eigenvector centrality (Fig 5). At this point, clans with a modestly bride-receiving orientation exhibit higher normalized degree than parity-oriented clans (, p < 0.001) and substantially higher betweenness (, p < 0.001), indicating enhanced brokerage potential. Eigenvector centrality is likewise elevated at c = 0.55 (, p = 0.0048), suggesting that these clans are disproportionately embedded in the network’s most structurally prominent neighborhoods. Closeness centrality follows a compatible, but noisier, pattern: the contrast is positive at c = 0.55 but only marginal (, p = 0.0718), while the local advantage becomes statistically reliable one step to the right, at c = 0.60 (, p = 0.0108). Under the relaxed neutrality definition (), results mirror the strict benchmark. The alliance network exhibits its strongest structural advantages among clans that lean modestly toward bride-receiving. The stability of this interior peak across both values suggests that the “sweet-spot” at c ≈ 0.55–0.60 is not driven by an arbitrary definition of parity.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Adaptive sliding-window contrasts of structural centrality across the bride-receiving spectrum.

Each panel reports local contrasts in normalized centrality as a function of bride-receiving share c, where c = 0 denotes exclusive bride-sending and c = 1 denotes exclusive bride-receiving. The y-axis shows the difference in average centrality between clans near each c and parity-oriented clans (c ≈ 0.50); values above zero indicate greater centrality than parity, and values below zero indicate a structural penalty. Two specifications assess sensitivity to the parity definition: a strict neutrality band (; thinner/darker intervals) and a relaxed band (; thicker/lighter intervals). Pointwise 95% randomization intervals are obtained from the empirical 2.5% and 97.5% quantiles of across B = 5,000 outcome permutations; points and lines track the strict specification.

https://doi.org/10.1371/journal.pone.0345859.g005

Moving away from this interior interval, advantages attenuate and—especially at the extremes—contrasts turn sharply negative. Under the strict benchmark, heavy bride-senders at c = 0.00 are significantly less central than parity-oriented clans across all four measures (e.g., , p < 0.001; , p < 0.001; , p < 0.001; , p < 0.001). A comparable structural penalty emerges for heavy bride-receivers at the opposite end of the spectrum (e.g., at c = 0.90, , p < 0.001; , p < 0.001; , p < 0.001; , p < 0.001).

Because bride-receiving shares can be discrete for sparsely connected clans and because local contrasts depend on the definition of the parity benchmark, we re-estimated the full sliding-window analysis under a pre-specified battery of alternative specifications. Across these checks, the central substantive pattern is unchanged: the strongest positive contrasts for degree, betweenness, and eigenvector centrality consistently occur around c ≈ 0.55–0.60, and highly asymmetric clans at the extremes remain structurally disadvantaged. These conclusions hold when (i) widening the parity band from ±0.025 to ±0.10, (ii) excluding low-exposure clans (e.g., single-edge families), (iii) tightening window widths and lowering minimum cell sizes, (iv) winsorizing centrality outcomes to dampen outliers, and (v) repeating the analysis with directed prestige outcomes (PageRank and HITS authority) on the directed marriage network. Closeness is comparatively less stable under sparsity—as expected for a path-length measure sensitive to pendant nodes—but the location of the “sweet-spot” remains aligned with the other centrality metrics once low-exposure clans are excluded. Full results are reported in Appendix C (see especially Table 4 and related diagnostics).

4.1 Limitations

Our findings should be interpreted with consideration of several methodological constraints inherent to studying covert organizations, which highlight opportunities for future refinement.

Data derived from judicial sources prioritizing prosecutable crimes, potentially underrepresent peripheral clans. Such “non-random missingness” is a well-known limitation of court-derived networks and can bias path-based statistics, especially closeness and betweenness, if very central actors were omitted [78]. Specifically, biased geographical coverage may result in the underrepresentation of families operating in some areas, causing them to appear peripheral or absent in the observed network despite potentially playing a more central role in the underlying relational structure. In addition, differences in visibility among clans may inflate the observed centrality of more closely scrutinized actors and contribute to a structure that partly reflects investigative attention rather than underlying relational patterns.

To mitigate these risks, we adopted several safeguards aimed at ensuring sufficiently broad coverage in terms of both cases and groups [79,80]. First, we relied on investigations selected through a stratified sampling strategy across the two Calabrian Antimafia Districts Directorates (i.e., Reggio Calabria and Catanzaro). This approach ensures coverage of the entire Calabrian region –the ’Ndrangheta’s historical stronghold– by proportionally representing investigations conducted under both Directorates, thereby reducing the likelihood that central families are omitted for jurisdictional reasons. Second, to limit biases arising from differences in clan visibility, we randomly selected 40 investigations drawn from the full set of the 192 cases reported in the DIA reports over a ten-year period (see the Methodological section). While this strategy does not eliminate the selectivity inherent in judicial data, sampling 40 investigations across a broad temporal window (ten years) helps to reduce the influence of case-specific and time-specific investigative priorities. Third, we cross-checked that information on families identified by journalists and prosecutors as powerful actors is independently confirmed in the DIA’s national threat assessments, suggesting that the central core of the network is effectively captured. Fourth, simulation studies show that degree and betweenness-based rankings remain remarkably stable even when many peripheral nodes are missing, and analyses based on arrest warrants or judgments (the very documents we use) reliably identify organizational leaders [e.g., 78, 80]. Finally, because 90.3% of all recorded exogamous marriages fall inside the largest weak component, the observable network retains sufficient density and reach to yield meaningful structural inferences despite inevitable gaps at the margins.

Matrimonial ties constitute only one dimension of relational capital in organized crime networks, alongside mechanisms such as business partnerships, political alliances, and informal patronage relationships [24,39]. While focusing exclusively on marriage might overlook other influential social ties, the strong correlation between our network centrality measures and independent assessments of clan power (Table 1) indicates that marital alliances indeed capture essential and representative relational dynamics. Furthermore, empirical research on organized crime groups underscores the importance of kinship bonds, particularly marriages, as stable and strategic means of maintaining cohesion, trust, and resource access [21]. Hence, any additional non-marital alliances likely reinforce, rather than contradict, the structural patterns shown by matrimonial networks.

Gender-ratio extremes among marginal clans could reflect sparse data; in these cases, the ratio is discrete, not strategic. To keep such sparse points from dictating the results, we (i) replicated all tests after dropping isolates and single-edge clans, (ii) used sliding windows that automatically widen until they hold at least five—ten in the baseline—observations, and (iii) plotted the pattern with a robust LOESS smoother that down-weights high-leverage cases. These safeguards show that eliminating or winsorizing marginal clans leaves the position and significance of the “sweet-spot” intact; only closeness centrality, which is inherently sensitive to long paths created by sparse nodes, exhibits a modest shift. Residual noise therefore resides on the network’s fringe and does not affect the powerful families that drive our substantive conclusions.

We treat matrimonial ties as contemporaneous by collapsing into a single snapshot unions formed over several decades and reported in the judicial documents spanning the years from 2007 to 2016. While this temporal aggregation smooths short-term fluctuations, it potentially inflates estimates of cohesion by implying simultaneous rather than sequential ties. However, consistent with research in other covert networks, relational ties, especially familial ones, typically represent long-lasting commitments that remain significant even if not continuously active [54]. Moreover, our robustness checks via edge-removal simulations (Fig 4) confirm genuine structural fragility rather than artificial hyper-cohesion. Future studies employing dynamic network approaches could clarify how matrimonial ties evolve over time and further validate the sequencing effects of marriage-based alliances.

The La C Play designation, while cross-validated with DIA reports, offers a discrete snapshot. Gradations within powerful family tiers exist but would likely strengthen our findings: finer status differentiation might reveal stricter marriage-strategy stratification. That 5/22 clans were absent due to data boundaries highlights the need for multi-source validation—a challenge for all criminal network research.

A Analytical strategy: Formal definitions and robustness checks

A.2 Network cohesion and load-bearing ties.

We complement the analysis of the relationship between status and structural prominence by investigating whether alliances that involve powerful clans sustain cohesion disproportionately. We perform edge-removal simulations on the undirected marriage alliance graph G=(V,E), where vertices represent clans and an undirected edge indicates at least one intermarriage between members of the two clans. We focus on the largest connected portion of the alliance network and treat parallel marriages between the same pair of clans as a single alliance tie (i.e., G is simplified to remove loops and multi-edges). Clans are labeled as “powerful” based on external media classifications; these labels are treated as exogenous to the marriage network and are used to partition edges into analytically meaningful categories. Each edge is assigned to E^pow if at least one endpoint is a powerful clan and to E^non otherwise. Thus, and .

At every removal step, we delete edges only—no vertices are pruned—so the vertex set remains fixed; clans whose last alliance tie is removed become isolated nodes. Two complementary deletion regimes are considered. First, an equal-count scheme removes the same number of edges from each subset: at step k, we delete k edges sampled from E^pow and k edges sampled from E^non, for . This design ensures comparability under an identical absolute number of removals across edge categories. Second, an equal-share scheme removes an identical fraction of edges within each subset: for a fraction , we remove edges from E^pow and edges from E^non, thus standardizing edge removal intensity relative to subset size. In both regimes, at each step, edges are chosen uniformly at random without replacement from the relevant subset, and removals are applied sequentially to generate a perturbed graph G_k.

After every removal step, we record multiple complementary cohesion metrics capturing both global connectivity and internal redundancy. First, we measure the relative size of the largest connected component,

(4)

which captures fragmentation of the alliance system into disconnected blocs. Second, we compute the component concentration index,

(5)

where are component sizes in G_k; is the probability that two randomly selected clans belong to the same connected component, and therefore captures the overall reachability density of the network. Third, to assess cohesion under the stricter requirement that connectivity remains robust to the removal of any single clan, we track the relative size of the largest biconnected component in G_k, i.e., the largest subset of vertices that remains connected after deleting any one vertex. Finally, to capture how efficiently alliances sustain within-component reachability even when the network stays connected, we compute the average shortest path length within the largest connected component of G_k. This measure increases as the remaining alliance structure becomes more “stretched” and indirect, even before it fully fragments.

We repeat the entire deletion process 100 independent times for each regime and report the mean curve with 10^th–90^th-percentile envelopes across replicates. In addition, to examine whether alliances involving powerful clans occupy more bridging positions in the baseline network, we compute normalized edge betweenness centrality for all edges in G and compare its distribution between E^pow and E^non. This provides a complementary, non-simulation-based indication of whether ties attached to powerful clans disproportionately lie on shortest paths connecting otherwise distant parts of the alliance network.

A.4 Gender directionality and network position.

Finally, we study how a family’s bride-receiving share (Section 2.1.3) relates to our four measures of structural prominence—degree, eigenvector, betweenness, and closeness. To this aim, we construct local, distribution-free contrasts that are deliberately robust to sparsity and non-normality in network data. Unlike a regression smoother, this approach estimates a sequence of discrete, local mean contrasts across the bride-receiving spectrum, making non-linearities visible without imposing a global functional form. Unlike the status comparisons in Section A.1, where we permute the powerful-clan label, here we construct a null by permuting the outcome vector (centrality scores) across clans while holding bride-receiving shares fixed.

For each centre , we define an adaptive comparison set

(6)

where the half-width starts at h(c)=0.05 and is increased in increments of 0.05 until , subject to a maximum half-width of 0.50. This rule preserves fine resolution in dense regions while automatically widening the window—and thus stabilising estimates—where data are scarce.

For the neutral reference group we begin with a parity band around equal exchange and expand it only if needed to ensure stable cell sizes. Formally, for we define

(7)

where the neutral half-width starts at and is increased in increments of until . The strict setting () targets clans closest to parity, whereas the relaxed setting () increases stability by averaging over a wider set of near-parity clans; concordant patterns across the two baselines indicate that results are not driven by an arbitrary definition of parity. These clans form the benchmark group against which all local contrasts are computed: at each centre c, we compare the mean centrality of clans with bride-receiving share near c to the mean centrality of clans whose exchange is approximately balanced.

For each outcome we compute the local mean difference

(8)

Positive (negative) values indicate that clans with bride-receiving share near c are, on average, more (less) central than near-parity clans. This permutation test does not reshuffle the marriage network. For each c, we hold fixed the memberships of and and generate a null distribution by randomly permuting the centrality values Y_i across clans (i.e., permuting the outcome vector). This breaks any association between bride-receiving share and centrality while preserving the marginal distribution of Y.

Let , where Y^(b) denotes the outcome vector after the b-th random permutation across clans. Sampling variability is quantified via a two-sided permutation procedure with B = 5,000 random permutations:

(9)

We report pointwise 95% randomization intervals by adding the {0.025,0.975} quantiles of the permutation (null) distribution to the observed contrast,

where denotes the empirical -quantile. Inference is distribution-free and does not rely on normality, equal variances, or large-sample approximations. Whenever the 95% pointwise randomization interval lies entirely above zero, we infer that clans with bride-receiving share near c exhibit higher centrality than parity-oriented clans at that location on the spectrum; when the interval lies entirely below zero, they exhibit lower centrality. Substantively, a positive mean difference at c indicates a local structural advantage (e.g., more connections, greater brokerage potential, or more embedded ties to central clans) relative to the near-parity reference group, whereas negative values indicate a structural disadvantage.

Key features of this approach include: (1) adaptive resolution, with window widths expanding only where needed (up to a half-width of 0.50) to stabilise local estimates; (2) local, distribution-free inference via permutation tests induced by permuting Y across clans; (3) retention of comparisons even when and overlap near parity; and (4) diagnostic transparency through exporting the full set of contrasts, confidence intervals, and p-values alongside the figure to facilitate replication and sensitivity checks. Scanning c from 0.00 (strong bride-sender) to 1.00 (strong bride-receiver) thus yields a stable, assumption-lean map of structural prominence across gendered alliance strategies.

B Directed robustness checks: Alliance embeddedness versus exchange hierarchy

Marriage ties can be represented as either undirected or directed relations, depending on the construct of interest. For alliance embeddedness, direction is not substantively relevant: a marriage creates reciprocal, durable inter-family obligations (in-laws, reputational interdependence, and repeated interaction), regardless of whether the bride originates from clan i or clan j. Accordingly, all cohesion, centrality, and resilience analyses in the main text use the undirected alliance projection.

However, exchange hierarchy in the marriage market is inherently asymmetric. To capture this aspect, we replicate the status-centrality comparison using a directed clan network in which each tie is oriented from the bride’s clan to the groom’s clan (bride family → groom family). This allows us to assess both hierarchical prestige and directed brokerage in marriage exchange. We compute directed centrality measures on the largest weakly connected component and remove consanguineous self-loops.

Table 3 reports 5,000-permutation comparisons between powerful clans and other clans on normalized in- and out-degree, PageRank, HITS hub/authority scores, and directed betweenness. The results confirm that powerful clans remain significantly more central under directed prestige and brokerage measures.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Directed robustness checks of status differences.

https://doi.org/10.1371/journal.pone.0345859.t003

C Sensitivity of the bride-receiving share–centrality relationship

The baseline specification keeps all clans in the archive, regardless of how many marriages each records, and applies adaptive sliding-window contrasts that begin with a half-width of h = 0.05, expand in 0.05 increments, and require at least ten clans on either side of every focal midpoint. This design exploits the entire network while avoiding the volatility that arises from extremely narrow windows or tiny comparison cells.

To verify whether the main findings depend on these choices, we reran the procedure under several alternative settings. First, we widened the reference band around the neutral point c = 0.50 from ±2.5 to ±10 percentage points. Second, we excluded pendant clans by retaining only clans with undirected alliance degree ≥ 2 in the collapsed undirected network (i.e., clans connected to at least two distinct partner clans). Third, we combined the degree filter with a narrower sliding-window rule, starting at a half-width of h(c)=0.02 and expanding in increments of 0.02 until at least five clans were available per comparison set. Fourth, we winsorized every centrality score at its 95^th percentile, damping the influence of extreme outliers. Finally, we applied winsorization and the degree filter simultaneously. Each variant swept the sliding window across the full 0–1 range of bride-receiving shares.

We also verified that the “sweet-spot” pattern does not hinge on computing outcomes on an undirected projection of marriages. For this directed robustness check, we retained the very same sliding-window and permutation design, but replaced the undirected centrality outcomes with two directed measures of prestige computed on the directed marriage graph (bride clan → groom clan): PageRank and HITS Authority.

Table 4 reports, for every specification, the bride-receiving share at which the contrast against the neutral group is largest in absolute value, the corresponding mean difference, and its two-sided permutation p-value. Three measures–degree, betweenness and eigenvector prestige–tell a consistent story: their strongest positive contrasts cluster at in most robustness runs and remain significant at the 5 % level or better. Winsorization predictably shrinks the magnitudes, yet the advantages never change sign for , confirming that the headline “sweet-spot” is not an artefact of extreme observations. The directed outcomes align with this pattern, with both PageRank and Authority attaining their strongest positive contrasts at c = 0.55 (p < .001).

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Peak mean differences in network centrality under alternative specifications.

https://doi.org/10.1371/journal.pone.0345859.t004

Closeness centrality is more volatile than the other metrics, but its overall pattern remains compatible with a non-monotonic, inverted-U relationship. In specifications that retain all clans, the most extreme contrast occurs at the bride-sending end of the spectrum (c ≈ 0.00) and is negative, reflecting the influence of peripheral clans with long geodesic distances. Across the full range of c, however, closeness typically increases toward the interior of the distribution and only declines again at the extremes. This sensitivity arises because closeness averages distances to all nodes and is therefore especially affected by low-degree or pendant clans, whereas Degree, Betweenness, and Eigenvector centrality are driven by more local or recursive structures and display more stable profiles.

Finally, Fig 6 provides a visual robustness check that clarifies the intersection of clan status and matrimonial strategy. While both groups follow a similar distributional logic, the structural advantage of the “sweet spot” is significantly more pronounced for powerful families. For powerful clans (red), the transition toward a balanced-to-moderate bride-receiving share (c ≈ 0.55 − 0.70) is associated with a sharp, non-linear surge in betweenness and eigenvector centrality—a “prestige payoff” that is far more muted among other clans (blue). This suggests that while the matrimonial economy has a general structural logic, it is the powerful families who most effectively weaponize these positions to consolidate brokerage and recursive influence. Their dominance is not merely a higher baseline of connectivity, but a superior capacity to translate specific matrimonial configurations into disproportionate network power.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Status-conditioned returns on matrimonial strategy.

Scatterplots show the relationship between bride-receiving share and normalized centrality, faceted by measure and colored by clan status. Solid curves overlay a robust, local-linear LOESS smoother drawn separately for each status group.

https://doi.org/10.1371/journal.pone.0345859.g006