Hindcasts and forecasts of suicide mortality in US: A modeling study

Area of Science:

• Public Health
• Epidemiology
• Data Science

Background:

• Suicide deaths and ideation have risen in the US over the last 20 years.
• Effective interventions require timely, geographically specific suicide activity estimates.
• Current data release delays hinder real-time response.

Purpose of the Study:

• To assess the feasibility of predicting suicide mortality using a two-step hindcasting and forecasting process.
• To evaluate the utility of proxy data sources (crisis hotline calls, online searches) for improving suicide mortality forecasts.
• To develop and test forecasting models for state-level suicide risk.

Main Methods:

• Generated hindcasts (past mortality estimates) using Autoregressive Integrated Moving Average (ARIMA) models.
• Augmented hindcasts with crisis hotline call rates and Google search trends for suicide-related terms.
• Created 6-month ahead forecasts for all 50 states (2012-2020) and evaluated using Quantile Score (QS).

Main Results:

• The primary hindcast model (auto) showed better performance than the baseline.
• Augmented models, incorporating proxy data, demonstrated improved forecast calibration.
• While augmented models' median QS was not significantly different from the auto model, they showed better calibration.

Conclusions:

• Proxy data sources can effectively address delays in suicide mortality data release.
• The findings support the feasibility of an operational, state-level suicide risk forecasting system.
• Sustained collaboration between modelers and public health departments is essential for refining methods and ensuring forecast accuracy.