Does spatial information improve forecasting of influenza-like illness?

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Accurate forecasting of seasonal influenza-like illness (ILI) is crucial for public health preparedness and individual decision-making.
• Traditional forecasting models often rely solely on historical data from the target region.

Purpose of the Study:

• To evaluate if incorporating influenza activity data from neighboring US states improves the accuracy of ILI forecasting.
• To compare the predictive performance of statistical models using different combinations of spatial and temporal ILI data.

Main Methods:

• Utilized CDC FluView ILI data from 2010-2019 for weekly forecasting in each US state.
• Employed quantile, linear, and Poisson autoregressive models.
• Compared models using target state ILI, neighboring state ILI, and US population-weighted average ILI data.

Main Results:

• Models incorporating data from neighboring states and/or the US average ILI demonstrated slightly higher predictive accuracy compared to models using only lagged ILI data from the target state.
• The performance improvement from including neighboring state data was comparable to including the US average, suggesting geographic proximity is not the primary driver of accuracy gains.
• Significant within-season and between-season variability was observed in the impact of spatial information on prediction accuracy.

Conclusions:

• Including spatial information, such as data from neighboring states or a national average, offers a modest improvement in ILI forecasting accuracy.
• The benefits of spatial data in ILI forecasting are not solely driven by geographic proximity.
• Forecasting models should account for the dynamic and variable influence of spatial data across different time periods.