Background:
Neighborhoods with dense residential development, access to public transportation, and a mix of land-uses may encourage walking for transportation and thus help residents to avoid excess weight and subsequent poor health. However, studies seeking to capture the “walkability” of the environment have faced several problems. First, using several separate built environment characteristics is unsatisfactory in most applications because these measures tend to be collinear, causing problems when introduced together into analytic models. Second, measures that best describe the walkability of the built environment may depend on regional context, density of development, and spatial scale of measurement. Thus measures developed for one context may not be valid in other locations or at other spatial scales. Finally, composite walkability measures may be dominated by a single item, in which case the increased difficulty of measure construction and interpretation may not be justified.
Objectives:
We sought to describe and demonstrate a process for creating a walkability summary measure capturing the variability within an urban environment, and to evaluate whether this summary measure predicts variation in body mass index (BMI) better than population density, which we had previously found to be the most predictive single characteristic related to walkability in New York City (NYC). We also assessed the extent to which neighborhood definitions influence both the development of summary measures, and the predictive validity of our summary measures for three different neighborhood definitions - census tracts, zip codes, and one kilometer buffers surrounding residents’ homes.
Methods:
To capture the variation in the walkability of NYC neighborhoods, we used geographic information systems (GIS) to create variables measuring theoretically relevant aspects of the built environment for each of our three neighborhood definitions. These variables included population density, density of residential units (both the number of residential units over land area and the number units over the residential floor area), land-use mix, the amount of retail floor area, the number of bus and subway stops, and intersection density. Using principal components analysis (PCA) we assessed the degree to which these variables indicate underlying, or latent, characteristics of the built environment. From the PCA results, we constructed factors using the component weights for each of the components revealed in the PCA.
To determine the predictive validity of PCA-derived factors, we used objectively measured BMI data on a diverse sample of 13,102 NYC residents. We used r-squared values to compare the explanatory power of regression models using PCA-derived factors or population density to predict BMI. All generalized estimating equation models controlled for individual and area-based demographic and socioeconomic characteristics, and robust standard errors were used to account for clustering within United Hospital Fund areas (health reporting districts in NYC).
Results:
In PCA analyses of walkability-related built environment characteristics, the first factor was highly correlated with population density regardless of the neighborhood definition used; correlations were 0.92, 0.97, and 0.94 for census tracts, zip codes, and one kilometer buffers, respectively. In each case, the first factor was a significant predictor of BMI (p < 0.05) after adjusting for the first factor; at this level, population density was able to explain more of the variance in BMI than all three PCA factors. For both the zip code and one kilometer buffer analyses, a second factor that appeared to represent the density of commercial destinations was significantly associated with BMI (p < 0.01); at these levels population density explained more BMI variance than the first component, but models with all PCA factors had the most explanatory power.
Conclusions:
Our data suggest that population density predicts BMI as well as any single walkability composite factor within the context of NYC. However, additional walkability composite factors may contribute further explanatory power at some scales of measurement. Efforts to derive a walkability summary measure appear to be sensitive to the scale of measurement. This instance of the modifiable areal unit problem may be due to the clustering of commercial land uses and transit access points along the arterial roads that frequently form census tract boundaries.
Composite or summary measures walkability would be useful to researchers who wish to control for walkability in their investigations of other built environment characteristics, or to track changes in walkability over time. However, development of such measures should be clearly described in the interest of increasing transparency and consistency across contexts, and summary measures should be critically evaluated to consider whether they add value beyond that of their strongest single component.
Support:
This research was supported by a grant (# 5R01ES014229) from the National Institute for Environmental Health Science and by the Robert Wood Johnson Foundation’s Health & Society Scholars Program.
