Statistical Methods for Analyzing the Built Environment

Abstract

The built environment refers to the human made space in which humans live, work and recreate on a day-to-day basis. As such, it constrains and enables individual choices and has consequently received increased attention for its potential influence on human health. For example, the availability of junk food outlets near children's schools may influence child obesity and the availability of supermarkets near residential addresses of study participants may influence longitudinal change in body mass index. However, efforts to estimate these influences have met methodological obstacles including the need to address residential self-selection bias and challenges related to defining measurement of environmental attributes. In response to these challenges, this work develops three statistical methods that seek to characterize the health effect of the built environment features that can be thought of as a point pattern – the locations of businesses, community resources or other amenities that provide goods and services that support or discourage an individual's health. To complete this objective, this dissertation adapts a suite of predominantly non-parametric Bayesian modeling techniques including Dirichlet and Gaussian Processes in order to analyze the non-linear relationships between individuals' and their environments across space and time.

In Chapter II we develop the Spatial Temporal Aggregated Predictor (STAP) model framework, to empirically determine the spatial and temporal scales at which built environment features (BEFs) have their greatest impact on human health. The framework also enables the selection of different functions that best describe the spatial-temporal exposure relationship between environment and health. This approach thus removes the unnecessary, though widely used, pre-specification of distances in time or space (e.g. 1 mile buffer) within which to measure environmental features at the population level.

Chapter III extends the work in Chapter II to identify varying forms of the spatial-temporal relationship across the population. There is a prominent interest in these effects because person and place-based characteristics shape how individuals experience and utilize the built environment. Identifying heterogeneous effects such as these hence addresses a critical question in developing place-based interventions: where and for whom are built environment interventions more likely to promote health?

Chapter IV addresses a third issue related to measurement of built environment exposures: namely characterizing the spatial distribution of BEFs around, for example, subjects' residences or places of work by identifying exposure clusters. We go on to show two ways in which these cluster assignments can then be used in a health outcomes model to identify the effect associated with the cluster assignment while still accounting for uncertainty in cluster assignment.

Chapter V presents and briefly illustrates the suite of software packages that have been developed to implement the methods discussed in the previous chapters. The texttt{bentobox} (Built Environment Network Objects Tool Box) texttt{R} package contains custom statistical functions, data structures and visualization functions that assist in the exploration and analysis of built environment data. We illustrate how these tools can be paired with publicly available data to more easily perform built environment analyses.

An improved understanding of the impact of environmental features on health is critically important for developing place-based strategies and policies to improve population health. This dissertation contributes to that understanding by developing methodological approaches and software tools to improve how the scientific community quantifies the influence of built environment on health.