You are here

Assessing adverse health effects of long-term exposure to low levels of ambient pollution

Principal Investigators: 

Harvard T.H. Chan School of Public Health


Harvard T.H. Chan School of Public Health

This study will examine health effects of low levels of air pollution in the US using data from about ~56 million people enrolled in Medicare and Medicaid. In addition, they will develop new causal modeling methods to characterize the shape of the exposure-response function. See also this Program Summary of HEI's research program on low levels of air pollution. 

Funded under

Abstracts for the 2018 HEI Annual Conference. Please scroll down to view three abstracts and posters. 

Sensitivity of the Association of Long-term PM2.5 and Mortality to Modeling Choices

Danielle Braun1, Marianthi-Anna Kioumourtzoglou2, Xiao Wu1, Christine Choirat1, Qian Di1, Francesca Dominici1

1 Harvard T.H. Chan School of Public Health, Boston, MA, USA; 2 Columbia University Mailman School of Public Health, New York, NY, USA

Background. To date, multiple studies have used spatio-temporal prediction models to assign long-term fine particle (PM2.5) exposure to study participants and investigate the association with mortality. All these studies, nonetheless, have used different exposure models, ways to assign exposure, health models, and confounding adjustment approaches. To our knowledge, few studies have rigorously assessed the sensitivity of the reported results to the above modeling choices.

Methods. We used data from Medicare enrollees from six New England states between 2003 and 2012 (N > 3 million), for whom residential information is available at zip-code level. There is spatial-misalignment between outcomes that are available at the zip-code level, confounders that are measured at the grid cell-, zip code- or ZCTA-level, and PM2.5 concentrations that are predicted at 1 km x 1 km grid cells. Previous work has assigned exposures by estimating zip-code level PM2.5 exposure using the average of the four grid points nearest to the zip-code centroid. With the assistance of GIS experts, we extend this approach to develop novel methodology and software that uses as inputs gridded air pollution predictions and aggregates those in pre-defined spatial polygons, e.g. zip-codes at which the Medicare data are available. The new method uses zonal statistics by performing a spatial merge to aggregate the data at the zip-code level with the option of either area- or population-weighted weights. We conducted health analyses using: (1) three different validated and widely-used exposure models, (2) two different ways to assign zip-code-level exposures (the average of the four grids nearest to the zip-code centroid vs. using zonal statistics and area weights), (3) two different health model parameterizations (Cox vs. log-linear models), (4) use of categorical vs. continuous exposures, and (5) confounding adjustment by inclusion of potential confounders in the health model vs. using generalized propensity scores (GPS).

Results. Overall, we estimated significantly harmful PM2.5 effects on mortality under all scenarios. However, the estimated effects varied across the different exposure models, with significant effect modification by population density observed for only one model. We additionally observed higher effect estimates for area-weighted zip-code averaged exposures, but no significant differences between Cox vs. log-linear models. Finally, when using categorical exposures (defined by policy-relevant cut-offs) we obtained slightly different health effect estimates adjusting for confounding using GPS vs. simply including the potential confounders in the health model.

Conclusions. Although the overall conclusion would not change depending on the modeling choices in this very large study, these differences might be important for smaller sample sizes. Moreover, obtaining accurate estimates would greatly inform risk assessments and cost-benefit analyses, thus impacting regulatory actions.


Poster by Braun, Dominici et al, 2018 HEI Annual Conference

An Ensemble Model-based Approach for Spatially and Temporally Resolved NO2 Exposures in the Continental United States

Qian Di1, Petros Koutrakis1, Christine Choirat2, Joel Schwartz1, Francesca Dominici2

1 Department of Environmental Health, Harvard T.H. Chan School of Public Heath, Boston, Massachusetts 02115, United States; 2 Department of Biostatistics, Harvard T.H. Chan School of Public Heath, Boston, Massachusetts 02115, United States

Background. Nitrogen dioxide (NO2) is a criteria pollutant, commonly used as a traffic emissions tracer, and has been linked to multiple adverse health effects. Unlike particulate matter and ozone, NO2 modeling has received less attention. Previous studies used satellite data and kriging methods to estimate annual NO2. However, few studies attempted to model NO2 with both high spatial and temporal resolutions.

Methods. We used multiple predictor variables and several machine learning algorithms to model daily NO2. More specifically, we used satellite data, chemical transport model outputs, land-use variables, and meteorological variables as predictors to train the model using ground-level NO2 measurements from EPA monitoring stations. We used machine learning models including a neural network, gradient boosting, and random forests as three separate approaches to model NO2. We then used an ensemble model to aggregate these machine learning algorithms, since they are complementary to each other. We used the ensemble model to predict daily NO2 at 1 km x 1km grids in the continental United States from 2000 to 2016. We validated the model using ten-fold cross-validation.

Results. Ten-fold cross-validation indicated a good performance of our ensemble model-based predictions with daily R2 = 0.78 and MSE = 2.74 ppb. The model predicted high NO2 concentrations in urban areas, as expected.

Conclusions. We developed a new machine learning ensemble model to predict daily NO2 concentrations with high accuracy. We are able to predict daily NO2 even at locations without ground monitoring stations. These predictions will allow for the estimation of health effects of both short-term and long-term NO2 exposure across the United States.


Poster by Di, Dominici et al, 2018 HEI Annual Conference

Nationwide Studies of Short- and Long-term Effects of PM2.5 and O3 on Mortality

Francesca Dominici, Qian Di, Yan Wang, Antonella Zanobetti, Lingzhen Dai, Yun Wang, Petros Koutrakis, Christine Choirat, Joel Schwartz

Harvard T.H. Chan School of Public Health, Boston, MA, USA

Background. Although strong links for both short- and long-term fine particle (PM2.5) and ozone (O3) exposures and mortality have been reported, evidence is limited for air pollution levels below the National Ambient Air Quality Standards. Furthermore, previous studies have predominantly focused on urban populations, due to lack of exposure information availability and statistical power at rural areas. We present our findings on two nationwide studies investigating the impacts of short- and long-term PM2.5 and O3 exposure on mortality, and assess if effect estimates differ among populations living below the national standards.

Methods. We used data from Medicare beneficiaries in the continental United States between 2000 and 2012. We used validated prediction models that are highly resolved in time and space to estimate zip-code level daily (for short-term) and annual (for long-term) PM2.5 and O3 exposures. For short-term effects, we employed a time-stratified bidirectional case-crossover design and examined the association between mortality and the average PM2.5 and O3 exposures at the day of the event and the previous day. For the short-term O3 effects we assessed warm-season (April-September) exposures. For long-term effects, we used Cox proportional hazards models with time-varying exposures and adjusted for demographic characteristics, Medicaid eligibility, and area-level covariates. For both short- and long-term analyses we assessed exposure to PM2.5 and O3simultaneously in the model, to account for potential co-pollutant confounding.

Results. We estimated highly statistically harmful effects in all cases. Specifically, for short-term effects, we observed a 1.05% (95%CI: 0.95-1.15%) and 0.51% (95%CI: 0.41-0.61%) increase in mortality per 10 μg/m3 increase in PM2.5 and 10 ppb increase in warm-season O3 respectively. There was no evidence of a threshold in either exposure-response relationships. For long-term effects, we observed a 7.3% (95%CI: 7.1-7.5%) and 1.1% (95%CI: 1.0-1.2%) increase in mortality per 10 μg/m3 increase in PM2.5 and 10 ppb increase in O3 respectively. When the analysis was restricted to person-years with exposure to PM2.5 of less than 12 μg/m3 and O3 of less than 50 ppb, the same increases in PM2.5 and O3 were associated with increases in the risk of death of 13.6% (95%CI: 13.1-14.1) and 1.0% (95%CI: 0.9-1.1), respectively. 

Conclusions. In the entire Medicare population there was significant evidence of adverse effects related both to short- and long-term exposure to PM2.5 and O3. These effects persisted even at concentrations below the current national standards. Our findings suggest that these standards may need to be re-evaluated.


Poster by Dominici, Zanobetti et al, 2018 HEI Annual Conference


Di Q, Wang Y, Zanobetti A, Wang Y, Koutrakis P, Choirat C, Dominici F, Schwartz JD. Air Pollution and Mortality in the Medicare Population. N Engl J Med. 2017 Jun 29;376(26):2513-2522. doi: 10.1056/NEJMoa1702747.

Di Q, Dai L, Wang Y, Zanobetti A, Choirat C, Schwartz JD, Dominici F. Association of Short-term Exposure to Air Pollution With Mortality in Older Adults. JAMA. 2017 Dec 26;318(24):2446-2456. doi: 10.1001/jama.2017.17923.

Dominici F, Zigler C. Best Practices for Gauging Evidence of Causality in Air Pollution Epidemiology. Am J Epidemiol. 2017 Dec 15;186(12):1303-1309. doi: 10.1093/aje/kwx307.

Makar M1, Antonelli J, Di Q, Cutler D, Schwartz J, Dominici F. Estimating the Causal Effect of Low Levels of Fine Particulate Matter on Hospitalization. Epidemiology. 2017 Sep;28(5):627-634. doi: 10.1097/EDE.0000000000000690.

Xiao Wu, Danielle Braun, Marianthi-Anna Kioumourtzoglou, Christine Choirat, Qian Di, Francesca Dominici. Causal inference in the context of an error prone exposure: air pollution and mortality. Submitted to Annals of Applied statistics. Version June 28, 2018.