Generalization of Constrained Mixed-Effect Modeling Framework with Ensemble Learning to Broader Geographic Areas for Predicting Nitrogen Oxides at High Spatiotemporal Resolution

摘要

Background: Spatiotemporal models developed for a specific region typically provide high spatial and temporal resolution in predicting exposures locally; however, it is challenging to directly apply them to broader regions in large epidemiological studies. Aim: To extend our southern California (CA) nitrogen oxides (NOx) spatiotemporal modeling framework (Li et al., 2017) to the entire state, evaluate its performance, and recommend key parameters to tune for future spatiotemporal model extension applications. Methods: In addition to our southern CA model data, we incorporated data from 105 ambient monitoring stations to cover CA. We conducted sensitivity analyses to determine the optimal number and aggregation distance to use in reconstructing temporal basis functions (temporal variability) and Thiessen polygons (spatial effects), respectively. We conducted ensemble and 10-fold cross validation (CV) to determine model prediction performance against from long-term ambient monitoring data and from short-term independent measurement campaigns. Results: We achieved an ensemble learning CV R2 of 0.88 for both our N02 and NOx CA-wide global models. Without considering regional differences, global models had slightly diminished performance (R2 reduced ～9% for N02 and 4% for NOx) than region-specific models for southern CA. Sensitivity analyses showed that 4-6 temporal basis functions and a smaller aggregation distance (200 m) ensured that the global model captured a wider range of temporal and spatial patterns in NO2 and NOx variability, respectively. Additional temporal basis functions may result in overfitting and smaller aggregation distances severely impact computational time with minor incremental improvement in model performance. Conclusions: This study illustrates the importance of accounting for regional differences for tuning local, region-specific models such that they can be applied to larger areas.