Forecasting U.S. 12-Month-Ending Overdose Death Counts: Multi-Model National and Regional Projections
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Aims
To assess whether recent declines in U.S. rolling 12-month-ending drug overdose death counts are projected to continue, compare the retrospective performance of short-term forecasting models, and estimate national and regional 12-month changes.
Design
Comparative time-series forecasting study with a retrospective March 2025–February 2026 forecast evaluation and subsequent 12-month-ahead projections through February 2027.
Setting
United States and four U.S. Census regions: Northeast, Midwest, South, and West.
Cases
Aggregate drug overdose deaths reported in the National Center for Health Statistics Vital Statistics Rapid Release system (VSRR) and identified using ICD-10 underlying cause-of-death codes X40–X44, X60–X64, X85, and Y10–Y14.
Measurements
The primary outcome was the monthly series of rolling 12-month-ending overdose deaths. Candidate models included ARIMA, generalized additive models, Prophet, and AICc-ranked n -sub-epidemic models. Models were calibrated using January 2020–February 2025 data and evaluated against March 2025–February 2026 observations using mean absolute error, mean squared error, empirical 95% prediction interval coverage, and weighted interval score (WIS). Individual models were ranked by retrospective WIS, and normalized inverse-WIS weights were used to construct ensembles from the top-ranked models. Final forecasts were generated for March 2026– February 2027 after recalibrating models using January 2020–February 2026 data.
Results
Retrospective WIS performance differed geographically: GAM performed best nationally and in the Midwest and West, the leading n-sub-epidemic model in the Northeast, and ARIMA in the South. Median forecasts from all individual models and ensembles projected declines from February 2026 to February 2027 nationally and in each region, although prediction intervals varied substantially. Individual national median projections ranged from declines of 12.8% to 25.8%, while weighted-ensemble median projections indicated declines of 20.5% to 21.7%. Projected weighted-ensemble declines were larger in the Northeast (25.4%–30.0%) and Midwest (24.3%–26.3%) than in the South (16.6%–19.5%) and West (18.0%–21.8%).
Summary
Ensemble forecasts of the reported provisional VSRR series were consistent with continued declines in rolling annual overdose death counts nationally and across U.S. Census regions. The projected magnitude of decline differed by region and model specification. Because the forecasts used provisional rolling 12-month-ending counts, they should be interpreted as surveillance projections rather than exact monthly mortality predictions.
