Space-time modeling of health effects while controlling for spatially varying exposure surfaces

In fields such as environmental science and epidemiology, data are collected over space and time. In many cases, the spatial information in the data is aggregated. The aggregation reduces the information available for the analysis. However, the goal of the study remains the same, spatial risk assessment. In this thesis, we work with a dataset where the risk is varying relative to a constant exposure, race, and a dataset where we consider the risk to be constant relative to a varying exposure.;Racial and ethnic characteristics usually do not change over time, but environmental conditions may vary dramatically across space. In epidemiological studies, estimating the association between individual level covariates and adverse outcomes, the association is believed to be dependent on the environmental conditions. A common method to incorporate spatial information in the analysis is to use neighborhood level variables. They are usually considered sufficient to explain any spatial variation in risk estimates of an adverse birth outcome. We demonstrate that neighborhood level variables may not contain all relevant information to estimate how risk changes across space. By using a spatially-varying coefficients model for race, we show there is a significant amount of information available which may not be captured in models using only individual level covariates and neighborhood covariates.;In contrast, air pollution exposure changes in space and time. The difficulty in estimating the risk associated with air pollution exposure is that the health data are typically aggregated to some geographic unit. It is well known that this aggregation introduces ecological bias in the estimates. We demonstrate how the bias can affect risk estimates at different levels of aggregation using a dataset in North Carolina for the years 2001 and 2002. Also, there has been extensive research to reduce the bias by relating the aggregated outcomes to a continuously estimated surface for the exposure. However, due to the additional uncertainty introduced, the bias correction framework usually significantly increases the standard error of the risk estimates. We take a different approach where, instead of using a continuously approximated surface, we estimate the pollution at geographic units which are smaller than the level at which the data are aggregated. The new geographic unit is chosen where we can estimate the exposure within the region without drastically increasing the uncertainty. We demonstrate, using the same North Carolina data, that our method reduces the bias induced through aggregation as if the data were actually aggregated to the smaller geographic units.