Psychological Medicine

Original Articles

Predicting non-familial major physical violent crime perpetration in the US Army from administrative data

A. J. Rosellinia1, J. Monahana2, A. E. Streeta3a4, S. G. Heeringaa5, E. D. Hilla1, M. Petukhovaa1, B. Y. Reisa6, N. A. Sampsona1, P. Bliesea7, M. Schoenbauma8, M. B. Steina9a10, R. J. Ursanoa11 and R. C. Kesslera1 c1

a1 Department of Health Care Policy, Harvard Medical School, Boston, MA, USA

a2 School of Law, University of Virginia, Charlottesville, VA, USA

a3 National Center for PTSD, VA Boston Healthcare System, Boston, MA, USA

a4 Department of Psychiatry, Boston University School of Medicine, Boston, MA, USA

a5 Institute for Social Research, University of Michigan, Ann Arbor, MI, USA

a6 Predictive Medicine Group, Boston Children's Hospital and Harvard Medical School, Boston, MA, USA

a7 Darla Moore School of Business, University of South Carolina, Columbia, South Carolina, USA

a8 Office of Science Policy, Planning and Communications, National Institute of Mental Health, Bethesda, MD, USA

a9 Departments of Psychiatry and Family Medicine & Public Health, University of California San Diego, La Jolla, CA, USA

a10 VA San Diego Healthcare System, San Diego, CA, USA

a11 Department of Psychiatry, Center for the Study of Traumatic Stress, Uniformed Services University School of Medicine, Bethesda, MD, USA

Abstract

Background. Although interventions exist to reduce violent crime, optimal implementation requires accurate targeting. We report the results of an attempt to develop an actuarial model using machine learning methods to predict future violent crimes among US Army soldiers.

Method. A consolidated administrative database for all 975 057 soldiers in the US Army in 2004–2009 was created in the Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS). Of these soldiers, 5771 committed a first founded major physical violent crime (murder-manslaughter, kidnapping, aggravated arson, aggravated assault, robbery) over that time period. Temporally prior administrative records measuring socio-demographic, Army career, criminal justice, medical/pharmacy, and contextual variables were used to build an actuarial model for these crimes separately among men and women using machine learning methods (cross-validated stepwise regression, random forests, penalized regressions). The model was then validated in an independent 2011–2013 sample.

Results. Key predictors were indicators of disadvantaged social/socioeconomic status, early career stage, prior crime, and mental disorder treatment. Area under the receiver-operating characteristic curve was 0.80–0.82 in 2004–2009 and 0.77 in the 2011–2013 validation sample. Of all administratively recorded crimes, 36.2–33.1% (male-female) were committed by the 5% of soldiers having the highest predicted risk in 2004–2009 and an even higher proportion (50.5%) in the 2011–2013 validation sample.

Conclusions. Although these results suggest that the models could be used to target soldiers at high risk of violent crime perpetration for preventive interventions, final implementation decisions would require further validation and weighing of predicted effectiveness against intervention costs and competing risks.

(Received June 02 2015)

(Revised August 11 2015)

(Accepted August 12 2015)

(Online publication October 06 2015)

Key words

  • Actuarial model;
  • crime perpetration;
  • machine learning;
  • military violence;
  • physical violence;
  • risk model

Correspondence

c1 Address for correspondence: R. C. Kessler, Ph.D., Department of Health Care Policy, Harvard Medical School, 180 Longwood Avenue, Boston, MA 02115, USA. (Email: Kessler@hcp.med.harvard.edu)

Metrics