WP22/13: Predicting suicide and self-harm risk in administrative data

Working Paper 22/13

Leah Richmond-Rakerd
Stephanie D’Souza
Terryann Clarke
Barry Milne

Healthcare providers are most likely to assess risk for suicide and self-harm, but typically only have access to information available in health records. However, risk factors for suicide exist in other domains (e.g. history of violent offending). Further, individuals at risk for suicide may face barriers to presenting for treatment (e.g. stigma), and so may be more likely to have contact with other types of health and social sectors (e.g. social welfare, pharmaceutical systems). These could be important, untapped settings in which to identify individuals at risk, and connect them to appropriate services. As such, conducting suicide risk prediction using data that are routinely collected and available at the national level can inform population-level prevention efforts.

However, risk prediction models raise concerns around bias and fairness, so any model developed should be assessed for measures of bias and fairness, such as:

1. between-group calibration: whether groups have the same likelihood of self-harm given the same predictive risk score;
2. predictive parity: whether groups have the same likelihood of self-harm at different risk score thresholds; and
3. error rate balance: whether groups have the same error rates, such as false positives or negatives at a specified risk score threshold.

This project aims to assess the viability and fairness of using administrative records for suicide and self-harm prevention. We aim to:

1. Test whether integrating information across multiple health and social systems improves prediction of suicide and self-harm risk in Aotearoa New Zealand.
2. Assess model performance, including measures of bias and fairness, across sex, age, ethnicity, and deprivation level.

We will use the following tables from the Integrated Data Infrastructure (IDI):

• Life event data
• Person overseas spell
• Publicly funded hospital discharges
• Pharmaceutical data
• National Non-Admitted Patient Collection (NNAPAC)
• Court charges data
• Benefit dynamics data
• Mortality data

Predicting Suicide and Self-Harm Risk in Linked Administrative Data. American Foundation for Suicide Prevention grant to Assistant Professor Leah Richmond-Rakerd.