An ensemble modeling approach for estimating expected patterns in mortality in the United States.

Abstract

Background: Excess mortality statistics are critical for quantifying the impact of various public health disasters and epidemics, particularly when deaths directly attributable to a given event may be underestimated. The COVID-19 pandemic created challenges for monitoring excess mortality due to the prolonged elevations in mortality throughout the pandemic, and the difficulties in accurately forecasting expected mortality over long time horizons. Additionally, prior research has found that excess mortality estimates can vary depending on the model used and various analytic decisions. Ensemble modeling approaches can mitigate these limitations by incorporating information from multiple models and better accounting for the uncertainty around analytic choices.

Objective: To evaluate the performance of an ensemble modeling approach to estimating excess mortality relative to methods that rely on single models.

Methods: This study assesses an ensemble modeling approach to estimate excess mortality in the United States, based on weekly mortality data from the National Vital Statistics System from 2008 through 2024. We evaluated 76 candidate models with various long-term and seasonal trend specifications along with ensemble models.

Results: Results indicate that an ensemble modeling approach offers more consistent and robust performance, with improved accuracy and coverage compared to individual models.

Conclusions: This analysis highlights the utility of ensemble modeling in providing robust estimates of expected mortality, a critical resource for public health preparedness and response in the face of disasters or epidemics.