Abstract for the 2017 HEI Annual Conference
Assessing Adverse Health Effects of Long-Term Exposure to Low Levels of Ambient Pollution
Francesca Dominici and Antonella Zanobetti (PI and co-PI), Brent Coull, Joel Schwartz,
Petros Koutrakis, Cory Zigler, and Christine Choirat
Harvard TH Chan School of Public Health, Boston, MA, USA
As air pollution levels continue to decrease and regulatory actions become more costly, steps taken to quantify the public health benefits of cleaner air will be subject to intense scrutiny. Previous epidemiological analyses of claims data have provided strong evidence of the adverse health effects of air pollution. Yet, significant gaps in knowledge remain, particularly with regard to the health effects of long-term exposure to lower levels of air pollution.
Our project will address current gaps in knowledge through the following specific aims:
Aim 1. No large study to date has investigated the health effects of long-term air pollution in areas with sparse monitoring. We will apply and extend already developed hybrid prediction models to estimate long-term exposure to low levels of air pollution for the continental US during the period of 2000-2014 and link these predictions to health data. We will then link the exposure, health, and confounder data at the ZIP code level.
Aim 2. Measuring the health effects associated with long-term exposure to low levels of air pollution presents a number of methodological challenges. We will develop methods for new casual inference to estimate exposure response that adjusts for confounding factors and accounts for exposure error.
Aim 3. Little is known about the health effects of low pollution levels on mortality and morbidity outcomes, disease progression, or its effects on highly susceptible populations including children, pregnant women, low-income adults, the elderly and the disabled. Using data from Medicare, Medicaid and Medicare Current Beneficiary Survey enrollees and applying the new methods developed in Aim 2, we will estimate the health effects of long-term exposure to low levels of ambient air pollution in children, low-income adults, and the elderly.
Aim 4: Methods for data sharing and reproducibility in air pollution epidemiology are of paramount importance, yet the scientific community lacks tools to make this possible. We will provide new tools for data access and reproducibility, including statistical software to implement the methods developed in Aim 2 and specific instructions on how to reproduce our analyses.
No other cohort has ever had access to data with this level of spatio-temporal coverage, resolution, and accuracy, and no other study will have the capability of estimating health effects of low exposure within a causal inference framework. A unique feature of these analyses is that they can be conducted routinely every few years as new claims data become available and can be used to track effectiveness of regulatory actions and mitigation strategies over time. These contributions will yield groundbreaking evidence essential for supporting cost-effective regulations.
Poster by Dominici et al, 2017 HEI Annual Conference
A Neural Network-based Model for Spatially and Temporally Resolved PM2.5 Exposures in the Continental United States
Qian Di1, Itai Kloog1,2, Petros Koutrakis1, Alexei Lyapustin3, Yujie Wang4, Joel Schwartz1, and Francesca Dominici1
1Department of Environmental Health, Harvard T.H. Chan School of Public Heath, Boston, Massachusetts 02115, United States
2Department of Geography and Environmental Development, Ben-Gurion University of the Negev, Beer Sheva, P.O.B. 653, Israel
3National Aeronautics and Space Administration (NASA) Goddard Space Flight Center (GSFC), Code 613, Greenbelt, Maryland 20771, United States
4University of Maryland, Baltimore County, Baltimore, Maryland 21250, United States
Background Fine particulate matter (PM2.5) is a major public health concern and accurate exposure assessment of PM2.5 is essential to investigate its adverse health effect. Previous studies have used a number of models to estimate PM2.5 exposure, including interpolation, satellite-based aerosol optical depth (AOD) models, land-use regression, or chemical transport model simulation. All approaches have both strengths and weaknesses. Besides, other variables, such as normalized difference vegetation index (NDVI), surface reflectance, absorbing aerosol index, and meteorological fields, are also informative to PM2.5 modeling.
Methods We used multiple variables to model PM2.5 with a neural network, for its capacity of handling complex relationships and interactions between variables. We used AOD data, chemical transport model outputs, land-use variables, meteorological variables, NDVI, surface reflectance, absorbing aerosol index as predictors to model ground-level PM2.5 from EPA monitoring stations. We used convolutional layers to aggregate nearby information into neural network to account for spatial and temporal autocorrelation. We validated the model by ten-fold cross-validation. After model training, the trained neural network predicts daily PM2.5 at 1 km × 1km grids in the continental United States from 2000 to 2012.
Results Ten-fold cross-validation indicated a good performance of our neural network approach with daily R2 = 0.84 and MSE = 2.94 µg/m3. Model performance also exhibited regional variations with higher model performance in the Eastern and Central U.S. than the Western U.S. The Model still performed well at low PM2.5 levels (<12 µg/m3). Prediction results indicated higher PM2.5 concentrations in the Eastern and Central U.S. Summer time had higher PM2.5 levels than other seasons.
Conclusions This study explored a data-intensive approach with novel modeling technique to achieve PM2.5 prediction with high accuracy. Our results provide exposure assessment for PM2.5 in places without ground monitoring stations, which allows epidemiologists to access PM2.5 exposure in both the short-term and long-term.
Poster by Qian et al, 2017 HEI Annual Conference
Methods to Estimate the Effect of Long-term PM2.5 Exposure on Health Outcomes when the Exposure is Mis-measured
Danielle Braun1, Marianthi-Anna Kioumourtzoglou2, Xiao Wu1, Francesca Dominici1
1Harvard T.H. Chan School of Public Health, Boston, MA, USA; 2Columbia University Mailman School of Public Health, New York, NY, USA
Background Long-term PM2.5 exposure has consistently been associated with adverse health outcomes. Estimating the effect of long-term PM2.5 exposure on health outcomes, such as number of deaths, is an important yet challenging task. Most previous studies treat PM2.5 exposure is error-free, inducing bias in the estimated effects by ignoring the exposure measurement error. Inadequate adjustment for confounding might induce further bias in the estimated effects. These two limitations restrict the ability to obtain accurate estimates of air pollution health effects. Although some studies have addressed the issue of exposure measurement error, no study to our knowledge has done so while adjusting for confounding in a causal framework. To address this important knowledge gap we propose a new method that has a wide range of applications.
Methods Using validation data, we developed a new method to address exposure measurement error in a causal framework and ran extensive simulations to evaluate its performance. Our interest is in estimating the exposure effects in the main study (five New England states, consisting of 94,814 1km×1km grid cells), for which PM2.5 daily error-prone concentrations were predicted at grid-cells at a high resolution from a spatiotemporal model. For a subset of those grid-cells (validation study consisting of 116 1km×1km grid cells) we have PM2.5 daily error-free concentrations measured at monitoring stations. Using this data, we developed a regression calibration (RC)-based adjustment using generalized propensity scores (GPS) to adjust for confounding (RC-GPS). The advantage of this approach is that it allows for confounding adjustment using GPS, yet doing so requires an ordinal exposure. We fit a RC model based on PM2.5 as a continuous exposure, and then transformed the continuous exposure into an ordinal exposure to estimate the exposure effects. Outcome analysis is then conducted by using novel methods to aggregate PM2.5 exposure at the grid cells to zip-codes. We assessed the performance of the RC-GPS method using both sub-classification and inverse probability treatment weighting (IPTW) approaches to adjust for confounding.
Results Our simulations show that the proposed method is able to fully adjust for both the mis-measured exposure as well as confounding bias. When implementing GPS with sub-classification, the bias in the exposure effects improved from -22.5% and -21.6% when using the error-prone exposure without any adjustment to 0.02% and 0.03% using the RC-GPS approach (assuming an ordinal exposure with three categories, the two numbers correspond to the exposure effect between the first and second category and second and third category respectively). Similarly, when implementing GPS with IPTW, the bias improved from -22.9% and -22.1% to -0.50% and -0.57%. We plan to apply this method to investigate the effect between long-term PM2.5 exposure and mortality in New England, using zip-code aggregated mortality in Medicare enrollees.
Conclusions We propose an innovative approach to adjust for mis-measured exposures while using generalized propensity scores to adjust for confounding bias. Our simulations show that our approach results in more accurate estimations of the exposure effect, and has the potential of impacting health policy.
Poster by Braun et al, 2017 HEI Annual Conference