Primary Submission Category: Sensitivity Analysis
Inferential impacts of spatial confounding in national-scale air pollution health analyses
Authors: James Celi Kitch, Sophie Woodward, Danielle Braun, Michelle Bell, Francesca Dominici, Daniel Mork,
Presenting Author: James Celi Kitch*
Objective: The health burden attributable to Alzheimer’s disease and related dementias (ADRD) is expected to increase substantially in the coming decades. Mitigating this increase requires understanding factors that influence ADRD progression, and recent national-scale studies have begun to address this. However, existing approaches do not sufficiently incorporate spatial information to address unobserved spatial confounding. Difficulties in developing realistic simulation studies further complicate evaluation of inferential impacts from unobserved spatial confounding.
Methods: We leverage a novel causal inference benchmarking tool to generate semi-synthetic datasets from real, nationwide health data. On this data, we evaluate seven statistical models in their ability to estimate the Exposure-Response Curve (ERC) between fine particulate matter (PM2.5) exposure and ADRD-related hospitalizations, measuring model bias and coverage of the true ERC. We assess the impact of spatial features by masking spatially autocorrelated variables, emulating settings of unobserved spatial confounding.
Results: Models that consider spatial proximity are more robust to unobserved spatial confounding, supporting increased use in national-scale epidemiology studies. We also find that commonly used inferential procedures, such as an m-out-of-n block bootstrap, are insufficient for spatial confounding. Finally, we apply these methods to estimate the PM2.5–ADRD relationship in Medicare data.