Statistical Methods for Exposure Measurement Error in Built Environment Studies

Abstract

Collecting information about the built environment is a challenging process and often leads to measures of environmental exposures that are error-prone and thus result in biased inference. While the use of GIS (Geographic Information System) databases is growing exponentially in health research, there is little guidance on strategies to handle erroneous exposure measures specific to built environment data. Our motivating examples focus on the environmental exposure that is expressed as the number of food outlets (e.g., fast food restaurants) within a circular buffer around the subjects' locations. Measurement error in the exposure can mislead inferences about the impact of the availability of environmental features on health, which is the key factor in place-based strategies or policies to improve public health. This dissertation focuses on developing methods to address measurement challenges in studies of built environment health effects, namely measurement error and data integration.

In Chapter II, we propose a split-combine simulation extrapolation (SC-SIMEX) method that accommodates non-classical measurement error due to incorrect geocodes without requiring external data. The standard SIMEX citep{cook1994simulation} is widely used to correct the effect of measurement error on regression problems under distributional assumptions on errors. However, the exposure measurement errors that arise due to geocode coarsening have a novel distribution compared to the commonly used distributions in existing measurement error literature. The proposed SC-SIMEX relaxes the measurement error assumptions required for the standard SIMEX. The utility of SC-SIMEX is demonstrated in a study of children’s obesity in California in relation to the junk food environment.

In Chapter III, we develop a multi-source measurement error model to effectively integrate different data sources using external knowledge of source credibility. Secondary commercial lists often provide conflicting claims about the presence and location of businesses. Consequently, disagreement among exposures collected from different sources gives different inferences. Our work shows that bias in the naïve regression model is dependent upon the dispersion of true exposure and the source credibility. Knowing the reliability of databases from field validation studies, the proposed measurement error model can derive complete information about latent true exposures using incomplete data tables with partially known margins. Our method uses a Bayesian nonparametric model that makes few distributional assumptions about the counts of businesses in a region and handles the overdispersion in exposure counts. The application assesses the association between children’s BMI in urban schools and exposure to nearby convenience and grocery stores.

Chapter IV extends the data integration methodology in Chapter III to the time-varying setting. For example, commercial business lists often provide annual business listings, which enables us to compute time-varying exposures. We extend the method described in Chapter III to flexibly model the latent time-dependences in true exposures using integer-valued autoregressive process. We aim to estimate the dynamic latent exposures and reduce bias in longitudinal analyses of the exposure health effect.

The dissertation concludes with a discussion of future work and the wider implications of the proposed methods. Our methods help future built-environment studies where measurement errors are inevitable and their impacts are little known. This dissertation contributes to improved understanding of measurement error properties that need to be addressed when using built-environment databases and provides novel methods to overcome the bias in epidemiologic analyses.