This project contributes to the much needed evidence base on domestic violence in an attempt to improve prevention of this social problem. This is a research collaboration between University of Manchester, University of Birmingham and researchers affiliated with Cheshire Constabulary. The project has also secured participation from three other forces. The team is particularly interested in developing knowledge to improve the current risk assessment tools used by the police. Although risk assessment in domestic violence has generated a wealth of research in North America, police forces in the UK still rely on a risk assessment tool (DASH) of mostly unknown properties, which was developed using fairly rudimentary methods. It is not known whether the classifications of victims as high-risk made with DASH are good enough. There are good reasons to suspect the accuracy of DASH can be improved. Equally, there is limited evidence about the effectiveness of the responses provided to high-risk victims. This project, therefore, aims to meet the following objectives:
This project will use state of the art techniques and procedures in machine learning and data mining to build appropriate predictive models to address these questions. The study will use a wider set of features than those identified by DASH for predicting future harm. Unlike most of the work in this area, which has relied on in traditional classifiers such as logistic regression, we will use more sophisticated algorithms to try to develop models with less classification error. The models will be fine-tuned via resampling methods and the final model tested in a holdout sample. The proposed work will involve working with a large sample of cases for the models using easily extractable data from police computers and a smaller random sample (stratified by outcome within a year) for a set of analysis that will require manual extraction of information from case files. We will also conduct a number of interviews with practitioners and victims of domestic abuse with the goal of beginning to explore implementation challenges in order to scale up the proposed predictive model.