Assessment of Shared and Unshared Exposure Measurement Error in Ensemble Learning Estimates of Nitrogen Oxides and Its Implications on Epidemiological Findings in Air Pollution Studies

Abstract

To improve exposure spatial and temporal resolution, researchers are using machine learning spatiotemporal air pollution models for large cohort studies. We aim to (1) measure shared, unshared, multiplicative, and additive (SUMA) measurement error in a three-stage (mixed effect ensemble learning with constrained optimization) spatiotemporal nitrogen oxides (NOx) model and (2) assess the impact of shared error and advanced exposure algorithms on epidemiological results. By treating NOx ensembles as realizations from an external dosimetry system, we quantified SUMA measurement error by extracting variance and covariance elements across realizations. To identify geographic locations with significantly elevated error, we used generalized additive models with a smooth term for location. We iteratively analyzed the risk of recent wheeze and NOx exposure among children using predictions from each stage of the NOx model to assess incremental influences of modeling stages on epidemiological conclusions and adjusted for shared error. We found evidence of both shared and unshared multiplicative error (p<0.01) in our spatiotemporal NOx predictions. Findings indicate that unshared multiplicative error is 25.8 times larger than the shared multiplicative error. Significant geographic variation of shared multiplicative error was observed (p=0.0004) and the majority (41%) of all predictions with high shared multiplicative error were observed in the earliest prediction period, 1992-2000. Depending on the exposure output used, the wheeze odds ratio for an interquartile range increase in NOx exposure ranged from 1.16-1.29. The standard error increased from 0.0049 to 0.0053 when accounting for shared multiplicative error. Spatial and temporal patterns of shared multiplicative error were mostly observed in densely populated urban regions with complex air pollution sources. Epidemiological conclusions were minimally affected by shared multiplicative exposure measurement error.