Historical Redlining Is Associated with Disparities in Environmental Quality across California

Historical policies have been shown to underpin environmental quality. In the 1930s, the federal Home Owners’ Loan Corporation (HOLC) developed the most comprehensive archive of neighborhoods that would have been redlined by local lenders and the Federal Housing Administration, often applying racist criteria. Our study explored how redlining is associated with environmental quality across eight California cities. We integrated HOLC’s graded maps [grades A (i.e., “best” and “greenlined”), B, C, and D (i.e., “hazardous” and “redlined”)] with 10 environmental hazards using data from 2018 to 2021 to quantify the spatial overlap among redlined neighborhoods and environmental hazards. We found that formerly redlined neighborhoods have poorer environmental quality relative to those of other HOLC grades via higher pollution, more noise, less vegetation, and elevated temperatures. Additionally, we found that intraurban disparities were consistently worse for formerly redlined neighborhoods across environmental hazards, with redlined neighborhoods having higher pollution burdens (77% of redlined neighborhoods vs 18% of greenlined neighborhoods), more noise (72% vs 18%), less vegetation (86% vs 12%), and elevated temperature (72% vs 20%), than their respective city’s average. Our findings highlight that redlining, a policy abolished in 1968, remains an environmental justice concern by shaping the environmental quality of Californian urban neighborhoods.

(1 + 0.5) For Landsat 8 satellite imagery, we selected the year 2020 and 2021 to best align with the most recent data layers of CalEnviroScreen (see paragraph below), and we selected December and January because we wanted to understand disparities in vegetation during the wetter part of the year (i.e., highest vegetation).We downloaded Bands 4 and 5 to calculate Normalized Differentiated Vegetation Index (NDVI) (see Equation 4) and Band 10 to calculate land surface temperature (Equations 5-6).To calculate noise pollution, we extracted data from HowLoud.com, which scales noise pollution from 50, representing high levels of noise, and 100, representing high levels of silence.
For each city, we extracted values for 2,000 random points within that city's HOLC map.After obtaining these values, we rasterized each point dataset using the 'Kriging' function.In short, the kriging function interpolates data to infer values for particular spaces between points where sampling did not occur 3 .After extracting data from HowLoud, we inverted the scale for visualization purposes but retained the original values for use in our models (see S1.3)

S1.3: Data Analysis
To understand the influence of HOLC grade on the spatial distribution of environmental hazards, we ran generalized linear mixed models (GLMMs) with HOLC grade as the fixed effect, the area of a neighborhood as a log-offset variable, and city as a random effect using the glmmTMB package 4 .We repeated this model approach at the city-level, but removed city as a random effect and used a general linear model with the betareg 5 .For all environmental hazards except NDVI, temperature, and noise we used a beta distribution given the data were bounded between 0 and 1.
For NDVI, temperature, and noise, we used a log-linked gaussian distribution.We built two models for environmental hazards: a model containing HOLC grade as an independent variable and a null model where HOLC grade was omitted.

𝐻𝑂𝐿𝐶 𝑀𝑜𝑑𝑒𝑙: 𝐸𝑛𝑣𝑖𝑟𝑜𝑛𝑚𝑒𝑛𝑡𝑎𝑙 𝐻𝑎𝑧𝑎𝑟𝑑
We then used an AICc model-selection approach, selecting the models with the lowest AICc value.When the top-performing model was identified, we tested for significant differences between the top-performing model and the null model using likelihood ratio tests (LRT).If the differences were significant, we extracted the estimated marginal means and performed Tukey-Kramer's post-hoc analyses to determine which specific HOLC grade dyads (e.g., A vs. C, A vs. D, etc.) differed in the focal environmental hazard(s).Model selection results are found in Supporting Information 2.

S2.2: Intraurban Disparities
We found significant differences in intraurban disparities between HOLC grades across the environmental hazards examined.For pollution burden, grade D held the highest disparity (17.4 ± 20.9; Table S3), followed by grades C, B, and A and we found significant differences between all pairwise comparisons (Table S4).For lead, grade D held the highest disparity (5.4 ± 28.4; Table S3), followed by grades C, B, and A. We found significant differences between all pairwise comparisons except B and C and C and D (Table S4).For groundwater threat, grade D held the highest disparity (9.5 ± 24.0; Table S3), followed by grades C, B, and A. We found significant differences between grades A and B as well as A and D (Table S4).For toxic releases, grade D held the highest disparity (3.7 ± 13.8; Table S3), followed by grades C, B, and A and we found significant differences between all grades except A and B as well as B and C (Table S4).For hazardous waste facilities, grade D held the highest disparity (11.3 ± 29.8; Table S3), followed by grades C, B, and A and we found significant differences between all grades except A and B as well as B and C (Table S4).For cleanup sites, grade D held the highest disparity (13.0 ± 28.6; Table S3), followed by grades C, B, and A and we found significant differences between all grades except A and B as well as B and C (Table S4).For diesel PM, grade D held the highest disparity (18.4 ± 23.8; Table S3), followed by grades C, B, and A, with all pair-wise comparisons showing significant differences (Table S4).For PM 2.5 , grade D held the highest disparity (4.2 ± 12.3; Table S3), followed by grades C, B, and A, with all pair-wise comparisons showing significant differences (Table S4).For NDVI, grade D had the lowest disparity (-0.02 ± 0.02; Table S3), followed by grades C, B, and A. We found significant differences between all pairwise comparisons for NDVI (Table S4).For temperature, grade D had the highest disparity (0.4 ± 0.8; Table S3), followed by grades C, B, and A. We found significant differences in thermal intensity between all pairwise comparisons except grades C and D (Table S4).Lastly, for noise pollution, grade D had the highest disparity (1.7 ± 3.2; Table S3), followed by grades C, B, and A. We found significant differences between all pairwise comparisons for noise pollution (Table S4).Table S1.Pair-wise comparisons for environmental hazards across California from generalized linear mixed-models after controlling for the area of a neighborhood and among-city variation.

S3: Supporting Information Tables
We used Tukey-Kramer's post-hoc analyses to determine which specific HOLC grade dyads (e.g., A vs. C, A vs. D, etc.) significantly differed in the focal environmental hazard(s).

Figure S2 .Figure S3 .Figure S4 .Figure S5 .Figure S6 .Figure S7 .211Figure S8 .Figure S9 .Figure S10 .Figure S11 .Figure S12 .Figure S13 .
Figure S1.The relationship between HOLC grade and environmental hazards.We show (A) lead risk from housing, (B) water contamination, (C) hazardous waste facilities, (D) cleanup sites, (E) diesel particulate matter, (F) particulate matter (pm) 2.5 and (G) toxic releases from facilities across formerly graded HOLC neighborhoods.Measurements are shown in box plots where each dot represents a measurement within a neighborhood.The mean is shown as a black diamond, and whiskers represent 95% confidence intervals.

Table S2 .
Pair-wise comparisons for environmental hazards in Californian cities from generalized linear models that control for the area of a neighborhood.Environmental hazards that showed significance was followed by a Tukey-Kramer's post-hoc analyses to determine which specific HOLC grade dyads (e.g., A vs. C, A vs. D, etc.) significantly differed in the focal environmental hazard(s).Significant comparisons for Tukey-Kramer's post-hoc analyses are bolded.Grey rows indicate no significant differences were found between HOLC grades.*PM 2.5