Eight years of homicide evolution in Monterrey, Mexico: a network approach

The paper addresses how homicidal violence emerged, evolved, and spread in time and space across the Monterrey Metropolitan Area (MMA) during and after Mexico’s drug war period. Following the 2007 militarized strategy against drug cartels, homicide rates surged nationally but exhibited heterogeneous patterns across states and cities. In Nuevo León and the MMA, violence increased dramatically around 2009–2012 due to conflicts between the Cartel del Golfo and Los Zetas, then declined around 2013–2015, and rose again after 2016. The study’s core questions are whether homicide violence in MMA spreads temporally and spatially, whether municipalities and neighborhoods show correlated patterns, if simultaneous bursts occur within specific time windows, and whether high-speed roads and highways relate to homicide locations. The purpose is to leverage detailed geo-located data to construct temporal and spatial networks that reveal correlation structures, clusters, and urban barriers, thereby informing understanding and potential policy responses in a critical economic region of Mexico.

The paper surveys approaches to the spatial-temporal analysis of crime and conflict. Past works range from literature-based analyses and network inference to reaction–diffusion and mixed models. Many urban crime studies emphasize burglary and robbery due to data availability; some show property crime correlates with human mobility and floating populations, and scale-adjusted metrics reveal crime–property relationships. Gang-related violence has been modeled via networks (e.g., gang nodes and directed homicide links) and exponential random graph models including proximity, retaliation, and reciprocity; models indicate violent crime often decreases with distance and can predict territorial boundaries. Spatial correlations and percolation-like transitions in homicides have been documented at broader scales. However, homicide data with precise geolocation and high temporal resolution are rare; available datasets are typically aggregated at municipality, state, or national levels, limiting fine-grained analysis. The authors leverage a uniquely detailed, geo-located homicide dataset for the MMA to address this gap.

Data: Daily homicide events (2011–Feb 2018) were obtained from El Norte newspaper’s Mapa del Crimen, assembled into the El Norte Data Base (ENDB) with date, latitude, longitude, casualties, title, and URL (total 2264 observations). Events were mapped to neighborhoods using shapefiles (DataMX) for Nuevo León; only entries falling within neighborhood polygons were retained (2114 events). The shapefile contained 2691 neighborhood polygons with names, municipality, and geospatial attributes. Official homicide counts (2000–2018) for Mexico, Nuevo León, and MMA were taken from INEGI for contextual time series. Temporal aggregation and correlation: Weekly casualties per municipality were computed. Pearson correlation coefficients (PCC) between municipalities were calculated over the full period. Significance was assessed via a null model: 1000 reshuffles of weekly casualties per municipality to build a null distribution, assigning Z-scores to observed PCCs. Spatial correlation: For each year t and distance interval r (20 intervals defined by 0.05 quantiles of inter-neighborhood distances), spatial correlation G(t,r) followed Alves et al. (2015): G(t,r) = sum over neighborhood pairs within r of (h_i − μ)(h_j − μ) divided by Nσ^2, where h_i is homicides per capita for neighborhood i at time t, μ and σ^2 are the mean and variance across neighborhoods, and N is number of pairs. Neighborhood-level populations were unavailable, so a proxy A_i P_m was used, with A_i the neighborhood area and P_m the municipality population, to compute per-capita rates. Network construction:

• Yearly co-occurrence networks: Nodes are neighborhoods; an edge connects two neighborhoods if at least one homicide occurred in both within the same week during a given year. Edge weights equal the number of weekly co-occurrences; visualization highlighted edges with 1–2 (grey), 3–7 (blue), and 8–12 (red) co-occurrences.
• Whole-period adjacency network: Nodes are neighborhoods; edges connect adjacent polygons (geospatial neighbors) that each had at least one homicide at any time during the 86 months, removing temporal ordering to reveal spatial clustering and barriers. Adjacency between neighborhoods was computed using GeoPandas; networks were built with NetworkX and visualized/analyzed with Cytoscape. Highways and urban features: OpenStreetMap data for highways were retrieved via the Overpass API (bbox covering MMA). Ways were parsed with pyosmium to extract node coordinates. Distances from homicide locations to the nearest highway were computed by year; the 99th percentile distance summarized yearly proximity (to reduce outlier influence). The path of Federal Highway 85 (85th freeway) was overlaid to assess alignment with co-occurrence links and crime clusters. Granularity and coverage: MMA comprises 19 municipalities and 2691 neighborhoods; 769 neighborhoods (about 28%) had at least one homicide during 2011–2018.

• Temporal patterns: National homicide rates rose from 2007–2012 and again from 2016–2018. In Nuevo León/MMA, a sharp rise occurred in 2009–2012, a decline around 2013–2015, with increases after 2016. Santa Catarina showed relatively steady homicide levels throughout, contrasting with other municipalities.
• Municipality-level correlations: Significant positive weekly correlations were found primarily among geographically close municipalities. Examples include Monterrey–San Nicolás de los Garza (PCC 0.3581, Z=6.4388), Guadalupe–Monterrey (0.3557, Z=6.3556), Apodaca–Monterrey (0.3170, Z=5.6961). Distant pairs showed weaker or negative correlations (e.g., García–Pesquería anti-correlated). This suggested distance-related correlation at the city level.
• Neighborhood-level spatial correlation: Spatial correlation G(t,r) showed little to no dependence on distance in early years; a modest positive correlation among close neighborhoods emerged in 2014–2016. Overall, distance did not strongly drive co-occurrence at neighborhood scale during most years.
• Co-occurrence network structure: Yearly networks (2011–2013) exhibited many edges reflecting concurrent weekly events; edges declined after 2014. High-frequency co-occurrences linked peripheral neighborhoods to a downtown polygon in Monterrey (e.g., Centro de Monterrey, Independencia, Moderna, Arturo B de la Garza, Nueva Res Española, La Campana, Burócratas Municipales). Star-like patterns indicated strongest co-occurrences between distant neighborhoods and this central polygon rather than between immediate neighbors. Notable pairs included Cadereyta Jiménez Centro–Centro de Monterrey (12 co-occurrences in 2012; 7 in 2011), and Centro de Monterrey–Independencia (10 in 2012).
• Whole-period adjacency network: Main connected components aligned strongly with municipal boundaries, indicating geopolitical segregation of homicide regions. For example, a component with main municipality Monterrey contained 220 of 305 nodes (72%); others were dominated by Guadalupe (entire components with 55 and 34 nodes), Apodaca (65, 100%), Santa Catarina (48 of 50), Cadereyta (23 of 23), Juárez (19 of 19), and General Escobedo (15 of 16). Safe neighborhoods often lay along municipality borders, disconnecting large violent components; the Santa Catarina–San Pedro Garza García border showed sharp contrasts in homicide presence.
• Highways as backbones/barriers: Many homicides lay near highways. The 99th percentile distance from homicides to highways was shortest during 2011–2012 (4379 m and 4475 m), longer in 2013–2016, decreasing again in 2017, indicating closer crime–highway proximity during peak cartel conflict. The 85th freeway (Mexico City to Nuevo Laredo border) intersected 11 of the 25 highly correlated neighborhoods and crossed the downtown polygon, suggesting a strategic corridor. Highways and municipal borders appeared to delineate violent vs. non-violent areas.
• Socioeconomic barriers and transit municipalities: Sharp contrasts existed across single avenues (e.g., San Pedro Garza García vs. Santa Catarina), consistent with socioeconomic segregation. San Nicolás de los Garza and Guadalupe showed patterns consistent with transit corridors (fewer central violent clusters; peripheral incidents).

The findings indicate that while municipality-level homicide trends exhibit distance-related correlations (nearby cities more synchronized), neighborhood-level dynamics during the drug war period did not generally follow simple spatial decay. Instead, violence displayed a hub-and-spoke pattern centered on a downtown polygon in Monterrey, with frequent co-occurrences between distant peripheral neighborhoods and central areas, consistent with conflict over strategic urban nodes and movement corridors. The proximity of high-frequency co-occurrence links and homicide concentrations to Federal Highway 85 supports the hypothesis that control of this corridor—and, by extension, the downtown polygon it traverses—was a key driver of violent events. Whole-period adjacency networks revealed that homicide clusters are strongly segregated by municipal boundaries, and that safe neighborhoods often sit at interfaces, acting as buffers that disrupt percolation of violence. Socioeconomic disparities across adjacent municipalities (e.g., San Pedro Garza García vs. Santa Catarina) likely contribute to these boundaries’ effectiveness, with implications for targeted interventions. The temporal proximity of homicide locations to highways during peak years underscores the role of transport infrastructure as both a facilitator of criminal activity and an urban boundary shaping crime diffusion. Overall, the network framework clarifies how temporal co-occurrence and spatial adjacency interact with geopolitical and urban structures to shape the spread and containment of homicide in the MMA.

The study introduces a network-based, multi-scale framework leveraging precise geo-located homicide data and urban infrastructure maps to characterize the spatiotemporal evolution of violence in the Monterrey Metropolitan Area (2011–2018). Key contributions include (i) evidence of municipality-level temporal correlations that decay with distance, (ii) discovery that neighborhood-level co-occurrences largely connect distant peripheries to a central downtown polygon, (iii) identification of highways—especially Federal Highway 85—as backbones associated with homicide concentrations, and (iv) demonstration that municipal borders and socioeconomic differences are strong separators of homicide clusters. The analysis suggests that strategic control over the downtown polygon and key corridors likely underpinned violent dynamics. A decline in homicides after 2012 may be linked to state-level security reforms (Fuerza Civil) and leadership disruptions in cartels, though levels never returned to pre-2007 rates. Future work should formalize null models to rigorously test the segregating effects of geo-socio-political divisions, incorporate richer covariates (e.g., socioeconomic, policing deployment), and continue efforts in high-resolution data collection and curation to refine causal inference and intervention design.

• Victim/offender and gang affiliation data were unavailable; analyses relied solely on event location and date, preventing directionality or perpetrator–victim linkage and limiting causal interpretation of co-occurrences.
• Neighborhood-level population counts were not available; a proxy (neighborhood area × municipality population) was used to estimate per-capita rates, potentially introducing bias in spatial correlation estimates.
• Events were retained only when lying within mapped neighborhood polygons, reducing the dataset from 2264 to 2114 entries; this urban-area filter may exclude relevant peri-urban incidents.
• The whole-period adjacency network intentionally removes temporal ordering, aiding spatial clustering detection but precluding temporal sequence inference.
• Reliance on a newspaper-based database, while highly curated and consistent with official trends, may still contain reporting biases or omissions compared with official or forensic records.
• Choice of a 1-week co-occurrence window captures short-term synchrony but may miss other temporal scales of diffusion; alternative windows could yield different network structures.