The Lancet Public Health (Aug 2023)

Insights from linking police domestic abuse data and health data in South Wales, UK: a linked routine data analysis using decision tree classification

  • Natasha Kennedy, PhD,
  • Tint Lwin Win, MSc,
  • Amrita Bandyopadhyay, MSc,
  • Jonathan Kennedy, EngD,
  • Benjamin Rowe, BSc,
  • Cynthia McNerney, BBus,
  • Julie Evans, MSc,
  • Karen Hughes, ProfPhD,
  • Mark A Bellis, ProfDSc,
  • Angela Jones, MPH,
  • Karen Harrington, MSc,
  • Simon Moore, ProfPhD,
  • Sinead Brophy, ProfPhD

Journal volume & issue
Vol. 8, no. 8
pp. e629 – e638

Abstract

Read online

Summary: Background: Exposure to domestic abuse can lead to long-term negative impacts on the victim's physical and psychological wellbeing. The 1998 Crime and Disorder Act requires agencies to collaborate on crime reduction strategies, including data sharing. Although data sharing is feasible for individuals, rarely are whole-agency data linked. This study aimed to examine the knowledge obtained by integrating information from police and health-care datasets through data linkage and analyse associated risk factor clusters. Methods: This retrospective cohort study analyses data from residents of South Wales who were victims of domestic abuse resulting in a Public Protection Notification (PPN) submission between Aug 12, 2015 and March 31, 2020. The study links these data with the victims’ health records, collated within the Secure Anonymised Information Linkage databank, to examine factors associated with the outcome of an Emergency Department attendance, emergency hospital admission, or death within 12 months of the PPN submission. To assess the time to outcome for domestic abuse victims after the index PPN submission, we used Kaplan-Meier survival analysis. We used multivariable Cox regression models to identify which factors contributed the highest risk of experiencing an outcome after the index PPN submission. Finally, we created decision trees to describe specific groups of individuals who are at risk of experiencing a domestic abuse incident and subsequent outcome. Findings: After excluding individuals with multiple PPN records, duplicates, and records with a poor matching score or missing fields, the resulting clean dataset consisted of 8709 domestic abuse victims, of whom 6257 (71·8%) were female. Within a year of a domestic abuse incident, 3650 (41·9%) individuals had an outcome. Factors associated with experiencing an outcome within 12 months of the PPN included younger victim age (hazard ratio 1·183 [95% CI 1·053–1·329], p=0·0048), further PPN submissions after the initial referral (1·383 [1·295–1·476]; p<0·0001), injury at the scene (1·484 [1·368–1·609]; p<0·0001), assessed high risk (1·600 [1·444–1·773]; p<0·0001), referral to other agencies (1·518 [1·358–1·697]; p<0·0001), history of violence (1·229 [1·134–1·333]; p<0·0001), attempted strangulation (1·311 [1·148–1·497]; p<0·0001), and pregnancy (1·372 [1·142–1·648]; p=0·0007). Health-care data before the index PPN established that previous Emergency Department and hospital admissions, smoking, smoking cessation advice, obstetric codes, and prescription of antidepressants and antibiotics were associated with having a future outcome following a domestic abuse incident. Interpretation: The results indicate that vulnerable individuals are detectable in multiple datasets before and after involvement of the police. Operationalising these findings could reduce police callouts and future Emergency Department or hospital admissions, and improve outcomes for those who are vulnerable. Strategies include querying previous Emergency Department and hospital admissions, giving a high-risk assessment for a pregnant victim, and facilitating data linkage to identify vulnerable individuals. Funding: National Institute for Health Research.