Spatial and Temporal Models for Environmental Health EffectsEPA Grant Number: R828686
Title: Spatial and Temporal Models for Environmental Health Effects
Investigators: Clyde, Merlise
Institution: Duke University
EPA Project Officer: Louie, Nica
Project Period: February 12, 2001 through February 11, 2004 (Extended to February 11, 2006)
Project Amount: $557,859
RFA: Interagency Announcement of Opportunity for Grants in Environmental Statistics (2000) RFA Text | Recipients Lists
Research Category: Environmental Statistics , Health
Based on accumulating evidence linking health effects to ambient airborne particulate matter, the U.S. Environmental Protection Agency in 1997 promulgated revised, stricter air quality standards for particulate matter air pollution. While many epidemiological studies have found consistent relative risks across many areas of the country, there remains considerable uncertainty regarding the health risk associated with particulate matter at levels below the National Ambient Air Quality Standards, and in particular, with the effect of fine particles less than 2.5 micrometers in diameter. In 1998, the National Research Council identified priorities for particulate matter research for setting regulatory standards. Most epidemiological studies are based on Poisson regression models using ecological time series, where health outcomes are based on spatial aggregates and exposures approximated by measured concentrations at monitoring sites. The NRC report raises questions about whether these relatively simple models are adequate for estimating the risks associated with exposure to ambient particulate matter. Important statistical issues concern the effect of measurement error in exposures, the possibilities of false positives due to multiple hypothesis testing, and whether more sophisticated methods are needed to account for spatial variation in the data.
The goals of this project are to address these issues by building a novel hierarchical Bayesian model that integrates health outcomes with spatial-temporal models for pollutant and meteorology fields. The model at the first stage is an extension of the Cox proportional hazards model that utilizes evolutionary covariates to reflect an individual's exposure through space and time. This leads to a non-homogeneous Poisson process for the observed counts of health-related events (e.g., deaths, hospital admissions). By using evolutionary covariates for exposure we avoid many of the problems with model selection in choosing the lag structure for timing/duration of exposure. As the evolutionary covariates reflect an individual's true exposure, they are (generally) unobservable, and so at the next stage of the model are linked to specific spatial-temporal ambient pollution fields. These fields incorporate information from the observed concentration measured at monitoring sites, as well as information about known point sources. A novel feature of the models is that they can easily accommodate information available at different levels of resolution, such as individual covariates, point sources for emissions, lines sources (roads), and information at coarser levels, such as demographic characteristics provided by census data. While more complex than the usual Poisson regression models, Markov chain Monte Carlo algorithms can exploit the conditional independence encoded through the hierarchical model leading to modular updating of parameters in blocks of manageable sizes. Validation studies using simulated data will be used to compare the performance of the proposed methodology and standard Poisson regression approaches for estimation of risk.
We will use mortality and pollution data available from Phoenix, Arizona to motivate model development and provide a case study for the methodology. Phoenix is in a region where particulate pollution is a problem year round, and has monitoring data with both PM10 (particulate matter with diameters less than 10 micrometers), PM2.5 (particulate matter with diameters less than 2.5 micrometers), particle chemical composition data, as well as other criteria pollutants such as ozone and carbon monoxide.
By incorporating measurement error, spatial information, and other available pollution data (point source emissions and 6 day monitors), we envision that the proposed model will lead to more accurate risk assessments of the effects of both fine and coarse particulate matter, as well as health effects of other pollutants.