Assessing the health estimation capacity of air pollution exposure prediction models

Background: The era of big data has enabled sophisticated models to predict air pollution concentrations over space and time. Historically these models have been evaluated using overall metrics that measure how close predictions are to monitoring data. However, overall methods are not designed to distinguish error at timescales most relevant for epidemiologic studies, such as day-to-day errors that impact studies of short-term health associations. Methods: We introduce frequency band model performance, which quantifies health estimation capacity of air quality prediction models for time series studies of air pollution and health. Frequency band model performance uses a discrete Fourier transform to evaluate prediction models at timescales of interest. We simulated fine particulate matter (PM_2.5), with errors at timescales varying from acute to seasonal, and health time series data. To compare evaluation approaches, we use correlations and root mean squared error (RMSE). Additionally, we assess health estimation capacity through bias and RMSE in estimated health associations. We apply frequency band model performance to PM_2.5 predictions at 17 monitors in 8 US cities. Results: In simulations, frequency band model performance rates predictions better (lower RMSE, higher correlation) when there is no error at a particular timescale (e.g., acute) and worse when error is added to that timescale, compared to overall approaches. Further, frequency band model performance is more strongly associated (R² = 0.95) with health association bias compared to overall approaches (R² = 0.57). For PM_2.5 predictions in Salt Lake City, UT, frequency band model performance better identifies acute error that may impact estimated short-term health associations. Conclusions: For epidemiologic studies, frequency band model performance provides an improvement over existing approaches because it evaluates models at the timescale of interest and is more strongly associated with bias in estimated health associations. Evaluating prediction models at timescales relevant for health studies is critical to determining whether model error will impact estimated health associations.