Background: Predicting suicide at population scale is difficult due to low base rates, reporting lags, and rapidly changing local conditions. Expanding access of the linked administrative data in Australia offers a path to more timely, reliable area-level risk signals that can inform upstream prevention and coordinated support.
Objective: Test whether linked, population-scale administrative data can produce calibrated short-horizon forecasts of regional suicide risk suitable for planning and early intervention.
Data: We utilise Australian linked administrative sources from the ABS Person-Level Integrated Data Asset, including Australian Taxation Office (ATO) income and employment information, Medicare Benefits Schedule (MBS) service claims, and Pharmaceutical Benefits Scheme (PBS) prescriptions, alongside other institutional datasets describing demographics and socio-economic conditions. Forecasts are enhanced with NSW admitted-patient and emergency department data, using hospital-treated intentional self-harm as a leading indicator of area-level risk. Analyses occur in secure environments with standard confidentiality protections.
Methods: We engineer quarterly, LGA-level features capturing labour-market dynamics, income distribution, primary and mental-health service use, medication dispensing, demographic structure, remoteness, health-service utilisation, and seasonality. Recent hospital-treated self-harm rates provide near-term signal. We compare elastic-net count models, gradient-boosted trees, and random forests, and combine them via stacked ensembling. Rolling-origin evaluation produces one- to four-quarter-ahead forecasts.
Anticipated results: Models that incorporate self-harm, service, and socio-economic signals are expected to outperform naive baselines at short horizons, yielding better-calibrated alerts and earlier identification of LGAs with rising risk.
Implications: If validated, this approach can provide practical, transparent early warning to guide place-based prevention and coordinated support.