Targeting domestic abuse by mining police records

This dissertation presents findings from analyses of three large datasets of domestic abuse records sourced from multiple police forces in England and Wales. It seeks to address research questions in relation to the extent of repeat and serial abuse, concentration and escalation of harm, and the forecasting of future serious crimes. Using a variety of statistics, it shows that most victims and offenders report domestic abuse to the police forces just once in a multi-year period. Among these cases however, are many of the individuals who comprise very small 'power few' groups that account for most of total crime harm. Using the Cambridge Crime Harm Index as the instrument of measurement, analysis shows that 80% of cumulative harm is attributable to fewer than 3% of victims and offenders, and almost half of these most harmed victims or harmful offenders have only one record of domestic abuse in police databases. Police forces are therefore presented with a substantial challenge when it comes to preventing serious harm from domestic abuse, because in more than 40% of the most harmful cases they have not dealt with the victims or offenders of domestic abuse before. Furthermore, among the victims and offenders who are linked to multiple records of domestic abuse, analysis detects no pattern of escalating severity. In fact, the first crime reported is, on average, the most harmful domestic crime reported to the police. This runs contrary to popular theories of escalation and further illustrates the forecasting challenge facing police agencies. Contemporary harm reduction strategies have placed some emphasis on the management of serial perpetrators of abuse, but analysis shows that these do not offer a complete solution for harm reduction either. These analyses show that serial offenders account for only between 10% and 15% of all domestic offenders and contribute no more to the 'power few' than repeat offenders who have just one victim. However, analysis of the non-domestic crime records of domestic offenders does show that serial perpetrators are less specialised in domestic abuse than repeat or single-time offenders - they commit more non-domestic types of crime and account for more total crime harm overall. Serial and repeat offenders with the greatest generalist tendencies were shown to be attributed to the most domestic abuse harm of any domestic offender types, indicating potential relevance to non-domestic offending records in the pursuit of predicting serious domestic crimes. How then, might police agencies seek to identify the most harmful cases before they occur? This research explores a large bank of records relating to arrests for any type of crime, using a statistical model created by a supervised machine learning algorithm. This model processes each arrestee every time they enter custody and using predictors from 35 pieces of information primarily concerning the prior offending history of the arrestee, generates a forecast of future arrest for a domestic crime within two years. The forecast has three classifications - no arrest, an arrest for a 'less serious' domestic crime, or an arrest for a serious domestic crime. In this fashion, the model could, at best, predict 49% of future serious arrests. The other 51% of serious domestic crime arrestees have no prior arrest record in the two years preceding their serious domestic crime and so there is no opportunity to forecast their crime based on police records alone. However, within the 49% of serious domestic crimes that are possible to forecast, the model accurately predicts more than three quarters. Hypothetically then, using this method of forecasting the police could identify 37% of all future serious domestic arrests up to two years before they occur. This presents the foundation of a major opportunity to reduce the prevalence and harm of domestic abuse. Taken together, these findings illustrate the scale of the challenge the police and other agencies face with reducing domestic abuse. A small proportion of individuals generate the majority of harm, but there are very limited opportunities to identify these individuals before the harm occurs. Yet, modern statistical techniques such as machine learning can help to target harm reduction strategies more precisely and even identify a sizeable proportion of serious cases before they occur. This dissertation presents with a series of ideas about how these objectives may be achieved and how the research can be developed further still.