Ability of crime, demographic and business data to forecast areas of increased violence

• Type of Material: Article
• Authors: Daniel A. Bowen, Centers for Disease Control and Prevention; Laura M. Mercer Kollar, Centers for Disease Control and Prevention; Daniel Wu, Emory University; David A. Fraser, DeKalb City Police Department; Charles E. Flood, DeKalb City Police Department; Jasmine C. Moore, Grady Health System; Elizabeth W. Mays, Grady Health System; Steven A. Sumner, Centers for Disease Control and Prevention
• Language: English
• Date: 2018-01-01
• Publisher: Taylor and Francis, Ltd.
• Publication Version: Author Accepted Manuscript (After Peer Review)
• Copyright Statement: © 2018, © 2018 Informa UK Limited, trading as Taylor & Francis Group.
• Final Published Version (URL): http://dx.doi.org/10.1080/17457300.2018.1467461
• Title of Journal or Parent Work: International Journal of Injury Control and Safety Promotion
• Volume: 25
• Issue: 4
• Start Page: 443
• End Page: 448
• Grant/Funding Information: This research was partially supported by a grant from the Robert Wood Johnson Foundation: “Piloting the Cardiff Model for Violence Prevention.”
• Keywords: Life Sciences & Biomedicine; Public, Environmental & Occupational Health; Science & Technology; machine learning; Violence; DENSITY; violent injuries; NEIGHBORHOODS; INTIMATE PARTNER VIOLENCE; UNITED-STATES; crime forecasting; MULTILEVEL
• Research Categories: Health Sciences, Public Health

Abstract

Identifying geographic areas and time periods of increased violence is of considerable importance in prevention planning. This study compared the performance of multiple data sources to prospectively forecast areas of increased interpersonal violence. We used 2011–2014 data from a large metropolitan county on interpersonal violence (homicide, assault, rape and robbery) and forecasted violence at the level of census block-groups and over a one-month moving time window. Inputs to a Random Forest model included historical crime records from the police department, demographic data from the US Census Bureau, and administrative data on licensed businesses. Among 279 block groups, a model utilizing all data sources was found to prospectively improve the identification of the top 5% most violent block-group months (positive predictive value = 52.1%; negative predictive value = 97.5%; sensitivity = 43.4%; specificity = 98.2%). Predictive modelling with simple inputs can help communities more efficiently focus violence prevention resources geographically.