4. Prior Research
Violent victimization: Predicted by arrest with other victims
‣ Exposure to similar risk factors
Gunshot and homicide victimization
‣ Generalizability to more common types of violence?
Chicago, IL; Boston, MA; Newark, NJ
‣ Generalizability to different geographic locations?
5. Our Goal: Network Approach to Violent Victimization
Primary data
‣ Law enforcement
Supplementary data
‣ Criminal courts
‣ State criminal history databases
‣ Emergency room visits
‣ U.S. Census
6. Pilot Project
Violent victimization in criminal networks
‣ Approximately three years of crime-related data
‣ Information on both offenders and victims
‣ Albuquerque, New Mexico
7. Current Goals
1) What can we learn about violence in criminal networks?
‣ Do offending networks link violent offenders and victims?
2) Assess feasibility of a larger project
‣ Do these data support a network approach?
8. Data: Overview
‣ Criminal event data
‣ Case numbers
‣ Suspects, arrestees, and victims: Unique identifiers
‣ Approximately three years of coverage
‣ 2011 into 2014
‣ Albuquerque, New Mexico
12. Data at a Glance: Cases
Records (case/person combination): 155,906
‣ Cases: 120,675
‣ Single offender: 45,348
‣ Multiple offenders: 4,150
13. Data at a Glance: People
Unique person identifiers: 120,345
‣ Offenders (suspects/arrestees): 40,858
‣ Multiple cases: 8,441
‣ Victims: 86,677
‣ Both offender and victim: 7,190
14. Violent Offenses: Examples
Battery/Aggravated battery/Aggravated assault
‣ 8,946 cases
Robbery/Commercial armed robbery/Auto car jacking
‣ 552 cases
Forcible rape/Attempted rape/Sexual assault with an object
‣ 490 cases
Murder and nonnegligent manslaughter/Justifiable homicide
‣ 83 cases
15. Building the Network
Offenders: Linked through crimes
‣ One-mode projection from two-mode data
‣ Assumption: Co-offending signifies social connection
‣ Case removal: > 10 offenders
‣ Case removal: Individual is both offender and victim
‣ Time dimension collapsed: Cross-sectional
20. How are different offender-types connected?
Do victims of violence have ties to violent offenders?
‣ Do links within offending network suggest “spread”?
21. How are different offender-types connected?
Non-victim Victim
Non-violent
offender
Non-violent offender,
non-victim1
Non-violent offender,
victim3
Violent offender Violent offender,
non-victim2
Violent offender,
victim4
26. Exponential Random Graph Models
Y: Random variable for state of network
g(y): Vector of network statistics for model
𝜃: Vector of parameters
k: Normalizing constant
Conditional log-odds of a single tie between two actors:
𝑝 𝑌 = 𝑦 =
𝑒𝑥𝑝(𝜃′
𝑔 𝑦 )
𝑘 𝜃
𝑙𝑜𝑔𝑖𝑡(𝑌𝑖𝑗 = 1|𝑦𝑖𝑗
𝑐
) = 𝜃′𝛿(𝑦𝑖𝑗)
27. Exponential Random Graph Models
Parameter Estimate Std. Error
Edges -9.891*** 0.032
Isolates 1.695*** 0.027
Actor covariate: Number offenses 0.159*** 0.004
Attributes mixes: Offender type
Non-violent offender, victim3
x Violent offender, victim4 0.205 0.147
Violent offender, non-victim2
x Non-violent offender, victim3 -0.626*** 0.111
Violent offender, non-victim2
x Violent offender, victim4 0.617*** 0.070
Violent offender, victim4
x Violent offender, victim4 1.138*** 0.146
29. Conclusions
1) Presence of offender network
‣ Ties: Sparse but not insignificant
2) Clustering of violence within network
‣ Violent offenders and victim ties: Victims are violent
‣ “Selection” vs. “Spread”
30. Next Steps
1) Continuing work: ERGMs
‣ Model specification & Goodness of fit
2) Pilot project: Future directions
‣ Projection of crime network
3) Larger project feasibility
‣ Current mood: “Cautiously optimistic”