Statistical Methods To Assess Environmental JusticeEPA Grant Number: U915240
Title: Statistical Methods To Assess Environmental Justice
Investigators: Conlon, Erin M.
Institution: University of Minnesota
EPA Project Officer: Edwards, Jason
Project Period: January 1, 1997 through January 1, 2000
Project Amount: $102,000
RFA: STAR Graduate Fellowships (1997) RFA Text | Recipients Lists
Research Category: Fellowship - Risk Assessment , Academic Fellowships , Ecological Indicators/Assessment/Restoration
The objective of this research project is to develop statistical methods for determining "environmental justice," which refers to areas in which no subpopulation experiences an unduly increased risk of disease due to environmental exposures. Geographic information systems provide tools for displaying geographic data. However, statistical tools are needed to analyze geographic data in order to determine geographic areas with increased exposure, disease incidence, and disease risk.
We use spatial hierarchical models are used to analyze possible associations between population characteristics, exposure, and health status, and to provide an overall assessment of environmental justice. Crude rates of outcomes are often are unstable due to rare outcomes from small areas or a small number of persons at risk. A Bayesian approach is used to "smooth" regional estimates toward a local mean rate of neighboring observations.
The fully hierarchical Bayesian model includes spatial similarity random effects. A conditional autoregressive prior for the spatial random effects induces spatial similarity between neighboring counts. In such a model, regional rate estimates are stabilized by combining information from neighboring regions. In typical applications, the neighborhood definition is fixed as regions sharing a common boundary. We use two methods for generalizing the neighborhood structure that allow the data to inform on the strength and geographic extent of spatial similarity. The first method involves modeling weights defining the contribution of neighboring rates in pooled estimates, and the second method involves weights induced by direct modeling of spatial correlation.
Stabilized rate maps are based on the posterior distributions of the log relative risks resulting from the fixed- and flexible-neighborhood definitions. We compare implementation, interpretation, and model fit are compared for the fixed- and flexible-neighborhood formulations using differing types of regionalization. The first data set involves lip cancer incidence in Scotland, and the second involves lung cancer mortality for one1 year in Ohio. Additional data sets include mortality rates for three diseases in Glasgow, Scotland.
We extend the spatial models are extended to include temporal effects. The spatio-temporal models assume spatial correlation is independent of time, and induce longitudinal correlations through a first-order autoregressive prior structure for an additional set of parameters. We use the data set of 21 years of lung cancer mortality in Ohio.