Association of historic redlining and present-day health in Baltimore · Huang SJ, Sehgal NJ

Materials and methods

CSA neighborhood maps were taken from the Baltimore Neighborhood Indicators Alliance (BNIA) [48]. This is a common set of CSA polygons that are each composed of several masked census tracts. These shapefiles include neighborhood characteristics such as median household income uptake and racial composition for each CSA in 2013. They also include various neighborhood health outcomes including 2013 life expectancy at birth (calculated using prevailing rates of death through the life span), teen pregnancy (per 1000 female residents aged 15–19), percentage of births at term, percentage of births at satisfactory birth weight, cases of elevated childhood lead levels per 1000 children 0–6, and mortality per 10,000 deaths in a five-year period at various age categories. Additionally, there are health process outcomes such as the uptake of prenatal care, as well as neighborhood features such as rate of liquor stores and fast-food restaurants. Shapefiles that include this data in a map were downloaded from BNIA.

The redlining map was taken from the Mapping Inequality [16] project on HOLC security maps. HOLC categories are assigned to areas within the current city of Baltimore geography and extend beyond the city of Baltimore’s boundaries. Given the age of the maps, not all of present-day Baltimore is categorized. Shapefiles that include this data in a map were downloaded from the University of Richmond’s Digital Scholarship Lab.

Methods

Transformations and overlay

All mapping analysis was done in QGIS 2.18.13. BNIA’s CSA shapefiles were projected using WGS84, these were converted into WGS84 pseudo-Mercator projection to be compatible with the Mapping Inequality HOLC shapefiles so that intersect overlays could be performed.

In order to determine which map sections lined up with each other, the intersection was taken of HOLC categories and Baltimore CSAs. The intersection of two fields in two maps returns all sub areas.

In line with methods for determining the classification of minimum mapping areas [42,43], the largest HOLC category by area within a CSA was then taken to represent the CSA’s HOLC categorization. This method is used across several disciplines [44–46]. We do not follow Hillier’s [47] usage of interval HOLC categories: they were trying to find determinants of HOLC categories and had noted the problematic limitations with treating HOLC categories as interval rather than ordinal. Aaronson et al. [12] were looking specifically at boundary blocks and tracts, which is beyond the scope of this analysis.

Statistical analysis

All statistical analysis was done in StataMP 16. The main independent variable of interest was HOLC category. Summary statistics were taken of HOLC category, life expectancy, median household income, and proportion of African American residents.

Next, our main dependent variable of life expectancy was regressed with ordinary least squares (OLS) against the main independent variable. Red and yellow lined CSAs were analyzed against a reference category of combined green/blue CSAs (for sample size reasons). A second OLS analysis was run with the control variable of median household income—to account for the independent non-HOLC effects of CSA income on health—and the control variable of proportion of African American residents—to account for other forms of anti-Black structural racism and racial composition change on health [2,29]. These measures cannot be fully disentangled from redlining, so we test for multicollinearity using variance inflation factor. We define our threshold of statistical significance to be p < 0.05.

Various other health outcomes across the life course were also regressed on HOLC category. These outcomes measures included mortality in BNIA-defined age bins (infancy, 1–14, 15–24, 25–44, 45–64, 65–84, and 85+), teen birthrate, percentage of births at term, percentage of babies born with satisfactory weight, and percentage of children testing positive for elevated blood lead levels. Process measures that were also regressed as dependent variables against HOLC category included percentage of mothers receiving prenatal care. Neighborhood factors that might impact health behaviors were regressed including the per population rate of liquor stores and fast-food restaurants in a CSA.

Finally, we ran a sensitivity analysis by redoing our first OLS analysis while excluding CSAs that have less than 25% HOLC coverage to see if our analysis is sensitive to CSAs that have low HOLC coverage.