Spatial measurement error methods in air pollution epidemiology
MetadataShow full item record
Air pollution epidemiology cohort studies often implement a two-stage approach to estimating associations of continuous health outcomes with one or more exposures. An inherent problem in these studies is that the exposures of interest are usually unobserved. Instead observations are available at misaligned monitoring locations. The first stage entails building exposure models with the monitoring data and predicting at subject locations; the second stage uses the predictions to estimate health effects. This induces measurement error that can induce bias and affect the standard error of resulting estimates. Berkson-like error arises from smoothing the exposure surface, while classical-like error comes from estimating the exposure model parameters. Accurately characterizing and correcting for both types of measurement error depends on assumptions made about the spatial surface and exposure model used to derive predictions. This dissertation addresses spatial measurement error in air pollution epidemiology. We first describe and apply parametric measurement error methodology when assuming the exposure surface is a stochastic Gaussian process. We extend these parametric approaches by deriving P-SIMEX, which yields more flexible bias correction. We then motivate a semi-parametric framework wherein the exposure surface is viewed as fixed and modeled with penalized regression splines. We discuss the resulting measurement error, describe how the exposure model penalty regulates measurement error, and derive an analytic bias correction. Finally we extend the semi-parametric methodology to the multi-pollutant setting. We show the direction of the biases are unpredictable, and the magnitude of the biases are much larger than those in single-pollutant studies. We derive a multi-pollutant bias correction that can be combined with a simple non-parametric bootstrap to achieve accurate 95% confidence interval coverage. Throughout we apply our methods to analyzing associations of continuous health outcomes with predicted exposures in the Multi-Ethnic Study of Atherosclerosis and the Sister Study of the National Institute of Environmental Health Sciences.
- Biostatistics