Using machine learning to forecast domestic homicide via police data and super learning.

We explore the feasibility of using machine learning on a police dataset to forecast domestic homicides. Existing forecasting instruments based on ordinary statistical instruments focus on non-fatal revictimization, produce outputs with limited predictive validity, or both. We implement a "super learner," a machine learning paradigm that incorporates roughly a dozen machine learning models to increase the recall and AUC of forecasting using any one model. We purposely incorporate police records only, rather than multiple data sources, to illustrate the practice utility of the super learner, as additional datasets are often unavailable due to confidentiality considerations. Using London Metropolitan Police Service data, our model outperforms all extant domestic homicide forecasting tools: the super learner detects 77.64% of homicides, with a precision score of 18.61% and a 71.04% Area Under the Curve (AUC), which, collectively and severely, are assessed as "excellent." Implications for theory, research, and practice are discussed.

Comparing panic alarm systems for high-risk domestic abuse victims: a randomised controlled trial on prevention and criminal justice system outcomes.

Background: The use of panic alarm systems for victims of domestic abuse is becoming increasingly popular. However, tests of these devices are rare. Consequently, it is presently unknown whether domestic abuse offenders are deterred by warning stickers informing them that a panic alarm system is installed on the premises, or whether alarm systems reduce domestic abuse recidivism. There is also a lack of data regarding whether adding an audio-recording feature to the panic alarm results in more prosecutions of domestic abuse offenders compared to standard panic alarm systems. Measuring the efficacy of warning stickers and audio recordings will enhance understanding of the overall effectiveness of panic alarm systems for domestic abuse. Methods: This study used a pre-test-post-test, control group design, in which 300 eligible high-risk domestic abuse victims in London, UK, were randomly allocated to either a standard panic alarm system or a panic alarm system with audio-recording capabilities and a red warning sticker on a durable, A6-size sign displayed at eye level at the entrance to the premises. Each sticker was well lit to ensure maximum visibility. The gain scores of multiple measures at 6 months prior and 6 months post-randomisation were used to assess the treatment effects (including the number of calls for service, recorded crimes, and harm score), and a negative binomial generalised linear model was utilised to estimate the likelihood of criminal charges for domestic abuse offenders in the two systems. Outcomes: Pre-post comparisons of recidivism suggested an overall reduction in both treatment arms, but there were no statistically significant differences between the two types of alarm systems across these crime measures. Nevertheless, the estimation model indicated a significant 57% increase in charges using the audio-recording alarm relative to the standard panic alarm system. Conclusions: Using deterrent stickers to warn domestic abuse offenders of panic alarm systems does not lead to a reduction in subsequent harm to victims. Compared to ordinary panic alarms for high-risk domestic abuse victims, audio-recording systems provide valuable evidence that increases subsequent charges, and thus, these systems should be explored further. Supplementary Information: The online version contains supplementary material available at 10.1007/s11292-022-09505-1.

Victims, offenders and victim-offender overlaps of knife crime: A social network analysis approach using police records.

Knife crime is a source of concern for the police in England and Wales, however little published research exists on this crime type. Who are the offenders who use knives to commit crime, when and why? Who are their victims, and is there a victim-offender overlap? What is the social network formation for people who are exposed to knife crime? Using a multidimensional approach, our aim is to answer these questions about one of England and Wales' largest jurisdictions: Thames Valley. We first provide a state-of-the-art narrative review of the knife crime literature, followed by an analysis of population-level data on central tendency and dispersion of knife crimes reported to the police (2015-2019), on offences, offenders, victims, victim-offender overlaps and gang-related assaults. Social network analysis was used to explore the formations of offender-victim networks. Our findings show that knife crime represents a small proportion of crime (1.86%) and is associated largely with violence offenses. 16-34 year-old white males are at greatest risk of being the victims, offenders or victim-offenders of knife crime, with similar relative risks between these three categories. Both knife offenders and victims are likely to have a criminal record. Knife crimes are usually not gang-related (less than 20%), and experienced mostly between strangers, with the altercation often a non-retaliatory 'one-off event'. Even gang-related knife crimes do not follow 'tit-for-tat' relationships-except when the individuals involved have extensive offending histories and then are likely to retaliate instantaneously. We conclude that while rare, an incident of knife crime remains predicable, as a substantial ratio of offenders and victims of future knife crime can be found in police records. Prevention strategies should not be focused on gang-related criminals, but on either prolific violent offenders or repeat victims who are known to the police-and therefore more susceptible to knife crime exposure.