Causal Concentration-Response Modeling with Continuous Curves and Exposure Error Correction: PM2.5 and Mortality in...

Many studies have reported associations of fine particulate matter with aerodynamic diameter $\le 2.5\mu m$ (${\mathrm{PM}}_{2.5}$)with mortality but fewer at low concentrations and even fewer using causal modeling or correcting for exposure error bias. None have corrected for the nonrepresentativeness of monitoring locations.

We examined the association of ${\mathrm{PM}}_{2.5}$ with all-cause mortality in the Medicare cohort using a combination of causal modeling, flexible concentration-response modeling, and bias correction for exposure error, while controlling for ${\mathrm{NO}}_2$ and $O_3$ as well as standard confounders.

Using monitors not used to fit our ${\mathrm{PM}}_{2.5}$ model, we fitted 72 regression calibration models stratified by season, region, and elevation in the US. We fitted a B-spline with 4 degrees of freedom to the calibrated ${\mathrm{PM}}_{2.5}$ and fitted separate generalized propensity score models for each spline component using gradient boosting. We also used inverse probability weights to account for the nonrepresentativeness of monitoring locations. Using the generalized propensity scores and the B-splines, we fitted quasi-Poisson models to counts of deaths in each ZIP code-year stratified by race, Medicaid status, and gender. Separate models were fit for participants identifying as black and as white and for ZIP codes with higher and lower poverty rates. We fit a model using the original exposure to estimate the extent of exposure error bias.

The propensity score analysis achieved good balance for all covariates. Controlling for the propensity scores, we found a concentration-response curve with no evidence of a threshold and whose confidence interval did not include the null from $4{\mu g/m}^3$ and upward. There were 223,666,531 person-years of follow-up between the current US Environmental Protection Agency (EPA) standard of $9{\mu g/m}^3$ and the World Health Organization (WHO) guideline of $5{\mu g/m}^3$, and the rate ratio between them was 1.088 [95% confidence interval (CI): 1.064, 1.113]. Using the original exposure, the rate ratio was 1.076 (95% CI: 1.070, 1.083). Hence, effects continue below the EPA standard, and calibrated estimates of effect were 16% higher. Effects were larger from $8{\mu g/m}^3$ among participants identifying as black.

The concentration-response curve between air pollution and mortality remains after adjustment for exposure error and using causal models and continues to concentrations below current US EPA and EU standards and even below WHO guidelines. Exposure error in the original exposure resulted in noticeable downward bias at low concentrations. Persons identifying as black are more susceptible. https://doi.org/10.1289/EHP15238.