Predicting Suicide and Self-Harm Risk in Linked Administrative Data

Suicide is a leading cause of death worldwide. It is associated with a range of serious psychiatric disorders, and a substantial portion of individuals who attempt suicide have received a mental-health diagnosis, obtained treatment, or appeared in emergency rooms. Many of these individuals have also received prior clinical care for other types of self-harm behavior. However, despite these prevention opportunities, traditional clinical methods to predict suicide risk perform little better than chance. Barriers include small sample sizes; the low prevalence of suicide attempt; reliance on retrospective self-reports of suicidal behaviors; inability to test many risk factors within individual studies; and short follow-up periods. Researchers and clinicians are now looking to big data and innovative analytic approaches to address these barriers. I propose to derive data-driven risk profiles for suicide attempt and self-harm in nationwide administrative records. 

I will use the Statistics New Zealand Integrated Data Infrastructure (IDI). This is a newly-established collection of whole-of-population administrative data sources that are linked at the individual level by a common spine (N ~ 3.5 million). This comprehensive government register collects information across multiple administrative databases, from birth to death. I have previously established a collaboration with IDI researchers, and I am among the first overseas researchers to work with the data. In the IDI, collaborators and I will (1) link data from a wide range of administrative sectors (sociodemographic, medical, pharmaceutical, criminal-justice, social-service, and education) to data from outpatient and inpatient mental-health visits; and (2) extract information about future suicide attempt and self-harm from hospital records. Using these data, we will develop and validate models of suicide-attempt and self-harm risk following mental-health visits. We will test the performance of these models across different lengths of follow-up, across men and women, across different age groups, and across different levels of self-harm lethality. I will follow up these analyses in the Dunedin Multidisciplinary Health and Development Study, a population-representative birth cohort followed into their fifth decade of life with 94% retention (N = 1,037). I will test the performance of models derived using hospital-register measures of suicide attempt and self-harm in the Dunedin Study, where these outcomes have been assessed using self-reports and clinical interviews. The proposed work will advance computational-modeling approaches for mental-health data and has the potential to identify novel explanatory factors for suicidal behavior.