Historical red-lining is associated with fossil fuel power plant siting and present-day inequalities in air pollutant emissions

Stationary sources of air pollution are disproportionately located in communities of colour, but the causes for this disparity are unclear. Here we assess whether racialized appraisals of investment risk (‘red-lining’) undertaken by the US federal Home Owners’ Loan Corporation in the 1930s influenced the subsequent siting of fossil fuel power plants. Across 8,871 neighbourhoods in 196 US urban areas, we observed a stepwise correlation between risk grade, number of power plants and cumulative quantity of power plant emissions upwind and within 5 km. Controlling for pre-existing power plants, neighbourhoods deemed ‘hazardous’ (D grade, ‘red-lined’) had a higher likelihood of a fossil fuel power plant being sited between 1940 and 1969 (72%), 1970 and 1999 (20%) and 2000 and 2019 (31%), and higher average present-day emissions of nitrous oxides (82%), sulfur dioxide (38%) and fine particulate matter (63%) compared with ‘declining’ (C-graded) neighbourhoods. Our results suggest racism in the housing market contributed to inequalities in present-day power plant emissions burdens. Historical appraisals of US neighbourhoods are thought to have led to present-day disparities in homeownership, home values and environmental and health outcomes. Cushing et al. examine the links between this historic red-lining, the siting of fossil fuel power plants and the burden of emissions.

Stationary sources of air pollution are disproportionately located in communities of colour, but the causes for this disparity are unclear. Here we assess whether racialized appraisals of investment risk ('red-lining') undertaken by the US federal Home Owners' Loan Corporation in the 1930s influenced the subsequent siting of fossil fuel power plants. Across 8,871 neighbourhoods in 196 US urban areas, we observed a stepwise correlation between risk grade, number of power plants and cumulative quantity of power plant emissions upwind and within 5 km. Controlling for pre-existing power plants, neighbourhoods deemed 'hazardous' (D grade, 'red-lined') had a higher likelihood of a fossil fuel power plant being sited between 1940 and 1969 (72%), 1970 and 1999 (20%) and 2000 and 2019 (31%), and higher average present-day emissions of nitrous oxides (82%), sulfur dioxide (38%) and fine particulate matter (63%) compared with 'declining' ( C-gr ad ed) n ei ghbourhoods. Our results suggest racism in the housing market contributed to inequalities in present-day power plant emissions burdens. Ambient air pollution is estimated to cause over 100,000 premature deaths in the United States each year 1 . Black, Hispanic and Asian communities face higher exposures to hazardous and criteria ambient air pollutants [2][3][4][5] and are more likely to live near polluting industries, even when accounting for differences in income and educational attainment 6 . Fossil fuel power plants contribute to local and regional air pollution through the emission of particulate matter (PM), nitrogen oxides (NO x ) and-in the case of coal-fired power plants-sulfur oxides and mercury. The burden of PM emissions from coal-fired power plants is greater for Americans of colour than White Americans, and emissions reductions due to recent coal-fired power plant retirements appear to have disproportionately benefited White Americans 7 . However, the processes and historical factors that led to these present-day inequalities are poorly understood, and the forms of structural racism contributing to racial disparities in air pollution exposures remain under-explored.
In the wake of the Great Depression, the newly formed US federal Home Owners' Loan Corporation (HOLC) and Federal Housing Administration undertook widespread neighbourhood appraisals as part of an effort to facilitate access to affordable home mortgages and prevent foreclosures. HOLC Residential Security Maps created in the 1930s in consultation with the local developers, financial and real estate professionals ranked neighbourhoods according to perceived lending risk and assigned grades that typically ranged from A ('best') to D ('hazardous') 8 . The practice became known as 'red-lining' because of the red colour given to D-graded neighbourhoods with the highest perceived investment risk. The maps reflected prevailing racist beliefs in the market value of integrated neighbourhoods such that communities with Black, East Asian, Filipino and foreign-born residents received poorer grades, while A-grade neighbourhoods appear to have been almost exclusively made up of White, US-born residents 9 . Through their codification of racist ideologies about value in the residential housing market, the HOLC maps are thought to have deepened racial residential segregation already enforced via racially restrictive deed covenants, zoning, block busting and violence and contributed to the exclusion of Article https://doi.org/10.1038/s41560-022-01162-y the construction of the power plant dataset. We separately examined coal-or oil-fired and peaker plants (that is, plants that turn on only during high-demand days) because they typically have higher emissions rates per unit of energy produced than baseload-generating natural gas plants. All plants that came online in or after 1940 are included in our analysis of power plant siting, including those that have since been retired.
We additionally considered cumulative annual emissions from all upwind power plants within 5 km (emissions burden). Because our primary interest is in how red-lining influenced siting, we did not attempt to estimate pollution dispersion or population exposures to ambient concentrations of these pollutants. Rather, we chose the metric of emissions burden because it reflects both the number of upwind power plants and their 'dirtiness' (for example, generation capacity, technology, fuel type and emissions control measures, which all influence the magnitude of emissions). Because historical emissions data were not available for all time periods, we focused the emissions analysis only on plants still operable in 2019.
Because neighbourhoods graded A and D already differed substantially at the time HOLC grades were assigned, we took the approach of estimating differences associated with adjacent grades that were more similar to each other at baseline (that is, D versus C, C versus B and B versus A) 16 . We controlled for US census region and the presence of power plants before 1940. In a sensitivity analysis, we additionally controlled for 1940 population demographic and socioeconomic characteristics to better isolate the effect of red-lining independent of neighbourhood differences that pre-dated the HOLC maps.

Power plant siting and HOLC grade
HOLC grade, power plant locations and present-day racial/ethnic composition from the US 2020 Census are illustrated across four cities in Fig. 1. On the basis of the availability of digitized HOLC maps, we include 8,871 HOLC-graded areas (hereafter 'neighbourhoods') from 196 urban areas and 3,284 fossil fuel power plants in our final analysis (Table 1). Compared with other grades, a slightly higher proportion of D-graded neighbourhoods (1.9% versus 1.1%, 1.2% and 0.9% for C, B and A, respectively) already had an upwind power plant within 5 km before 1940 (Fig. 2a). The disparity grew as additional power plants came online, with the proportion of neighbourhoods with a plant nearby and average number of upwind power plants within 5 km increasing from HOLC grade A to D during subsequent periods (Fig. 2a,b).
We also observed increased risks of power plant siting when comparing other adjacent grades (C versus B and B versus A) (Fig. 3 and  Supplementary Tables 2-5). In general, effect estimates were smaller in magnitude and less precise, and associations were less consistently observed in every period. Unlike the D versus C comparison, the largest differences in risk when comparing other adjacent grades did not generally occur during the post-war period. For example, the strongest association between HOLC grade and power plant siting when comparing C versus B was during the 2000-2019 period. people of colour from homeownership 10,11 . Recent scholarship suggests historical red-lining-and the racism in the housing market it reflected and codified-contributed to present-day geographic disparities in homeownership, home values, rents and income [12][13][14] .
The recent digitization of HOLC Residential Security Maps has enabled a growing number of studies to document a relationship between historically red-lined neighbourhoods and present-day environmental and health outcomes 15 . Red-lining has been associated with less urban vegetation 16 , higher surface temperatures 17 , higher levels of air pollution 18,19 , the siting of oil and gas wells 20 and higher rates of preterm birth 21,22 , emergency department visits for asthma 23 and late-stage cancer diagnoses 24 . However, studies to date have been almost exclusively cross-sectional due to a lack of historical data on these outcomes, giving little insight into the influence of red-lining on the trajectories shaping neighbourhood health and environmental inequalities over time. Also, few have attempted to control for pre-existing differences in neighbourhoods (for example, in the age of housing and socioeconomic status of residents) that pre-dated the creation of the Residential Security Maps and may partially explain differences in present-day conditions.
We investigate the relationship between HOLC grades assigned in the 1930s and the subsequent siting of fossil fuel (coal, oil or natural gas) power plants and resulting differences in the burden of present-day (2018 and 2019) power plant emissions of NO x , sulfur dioxide (SO 2 ) and fine PM of 2.5 microns or less in diameter (PM 2.5 ). Our primary objective was to understand whether racism in the housing market as reflected in the assignment of investment risk grades on federal HOLC Security Maps resulted in disparities in the likelihood of a power plant being sited nearby. As a second objective, we sought to assess whether these historical siting decisions led to differences in power plant emissions burdens that persist into the present day, indicating the continued relevance of historical red-lining in shaping current neighbourhood-level inequalities in proximity to environmental hazards.
When comparing D-('hazardous' or 'red-lined') to C-graded ('declining') neighbourhoods, we found red-lining was associated with an increased risk of a fossil fuel plant being sited upwind and within 5 km and a higher present-day emissions burden of NO x , SO 2 and PM 2.5 , even when accounting for pre-existing differences in proximity to power plants, socioeconomic factors and city population size. Associations with siting weakened over time but persisted into the twenty-first century, over 70 years after the first red-lining maps were drawn. Our findings suggest racism in the residential housing market has had an enduring impact on the subsequent siting of fossil fuel power plants and distribution of present-day air pollutant emissions.

Approach
We used HOLC boundaries as our spatial unit of analysis and consider them 'exposed' if they were downwind of a power plant within 5 km of the HOLC boundary based on annual average wind direction. We incorporated both distance and wind direction into our metric to focus on power plants that are most likely to impact the air quality of nearby populations and because we hypothesized that wind direction may have influenced siting decisions. Sensitivity analyses in which we defined exposure only on the basis of proximity irrespective of wind direction confirmed this hypothesis in that it resulted in largely consistent but weaker effect estimates (Supplementary Table 1 provides results based on proximity, and Supplementary Table 2 provides results based on proximity and wind direction).
We When we additionally controlled for baseline 1940 city population size (a potential predictor of demand for new power generation) and socioeconomic variables derived from data sheets collected by HOLC appraisers (a potential predictor of subsequent power plant siting), our confidence intervals widened with the reduction in sample size due to missing data for these variables ('Model 2' in Supplementary Tables 2-5). Effect estimates comparing D to C grades generally nevertheless suggested similar or stronger associations between red-lining and the presence, type and number of power plants sited upwind, with the exception of the 1970-1999 period for which effect estimates were weaker (Supplementary Tables 2-5). Effect estimates were consistently attenuated and less precise for C versus B and B versus A comparisons when controlling for these baseline differences (Supplementary  Tables 2-5). Effect estimates were also attenuated when we used 10 km rather than 5 km to define nearby power plants ( Supplementary Fig.  2). For example, when comparing D versus C, the relationship between red-lining and the presence and number of power plants upwind and within 10 km was only apparent during the 1940-1969 period, rather than all three periods as with the 5 km analysis.

HOLC grade and present-day power plant emissions
We then analysed the outcomes of present-day emissions for the 2,878 power plants still operational in 2019. We defined 'emissions burden' as the sum of pollutant emissions from all power plants located upwind and within a certain distance of the border of a HOLC neighbourhood; the main analysis considers a 5 km distance, while we also considered 10 km in a sensitivity analysis. Emissions burden generally followed a gradient, with A-graded neighbourhoods having the lowest on average, followed by B-, C-and D-graded neighbourhoods (Fig. 4). The trend was less pronounced for SO 2 than it was for NO x and PM 2.5 .
Among neighbourhoods located downwind and within 5 km of at least one power plant, multivariate log-linear models comparing present-day emissions by grade suggested that red-lining was associated with an increase in average annual NO x (82%), SO 2 (38%) and PM 2.5 (63%) emissions when comparing D-versus C-graded neighbourhoods (geometric mean ratio (GMR) and 95% CI for NO x = 1. When we additionally controlled for baseline 1940 city population size and socioeconomic characteristics, patterns remained consistent while precision decreased with the reduction in sample size due to missing data ('Model 2' in Supplementary Tables 6-8).The sensitivity analysis suggested a slightly (8-14%) stronger association between red-lining and emissions of all three pollutants when comparing D versus C neighbourhoods (GMR (95% CI) of 1  Tables  6-8). Effect estimates were attenuated for C versus B and B versus A comparisons and remained statistically insignificant when controlling for population size and demographics. Effect estimates were also attenuated in a separate sensitivity analysis utilizing a 10 km rather than 5 km distance to define upwind power plants, while overarching patterns remained consistent ( Supplementary Fig. 3).

Discussion
Our findings suggest racism in the residential housing market as reflected in red-lining maps created by the US federal government in the 1930s had an enduring impact on the subsequent siting of fossil fuel power plants and distribution of present-day power plant emissions. Worse-graded neighbourhoods had increased risk of a fossil fuel plant being sited upwind and within 5 km, with the strongest associations between D-('hazardous' or 'red-lined') to C-graded ('declining') neighbourhoods. While associations between red-lining and plant siting were strongest during the 1940-1969 period, HOLC grades assigned in the 1930s remained significantly associated with the likelihood of a plant being sited upwind over more than seven decades later (between 2000 and 2019). A 'D' grade was also associated with higher emissions burdens of NO x , SO 2 and PM 2.5 from fossil fuel power plants in the present Article https://doi.org/10.1038/s41560-022-01162-y day. Our findings are unlikely to be explained by pre-existing differences between neighbourhoods at the time the HOLC maps were created because we compared adjacent grades that have been shown to have been more similar socioeconomically based on 1940 census data 16 , controlled for differences in the presence of power plants at baseline and-in our sensitivity analyses-differences in baseline neighbourhood socioeconomic and demographic characteristics. Residence near a power plant has been associated with numerous adverse health outcomes, including increased emergency room visits 25 , asthma exacerbations 26 , respiratory-related hospitalizations 27 , preterm births 28,29 and reduced fertility 30 , underscoring the public health importance of our findings. Our analyses do not provide evidence regarding the specific mechanisms by which red-lining may have influenced power plant siting. Decisions about where to site new power plants are influenced by many factors including the cost of land, wages, available labour force, access to transportation and water, local and state regulations and community and political resistance 31 . The associations between red-lining and power plant siting that we observed may have arisen from depressed land values, reduced homeownership rates or reduced household income in neighbourhoods that received poorer scores 13,14 . Red-lining also deepened racial residential segregation and contributed to disinvestment in communities of colour 10,13 that may have made power plant employment opportunities attractive in those communities and/or undermined the political influence of red-lined communities to effectively resist the nearby siting of a power plant that was not locally supported. Prior empirical work has shown that the presence of people of colour and in some cases socioeconomic status predicted subsequent siting of hazardous waste facilities and coal-fired power plants [32][33][34][35] . Evidence from California also shows industry and government officials were highly cognizant of public opposition and considered the demographic and socioeconomic characteristics of communities when deciding where to site waste-to-energy facilities to avoid local resistance 36,37 . Importantly, red-lining maps codified prevailing racism in the perception of what a home was worth based on who lived in it 38 . Thus the associations we observed were not necessarily driven by the maps themselves and how they were used, which has been debated 8,39 , but rather the racism deeply rooted in US society and market institutions of which the maps were a reflection. Future research is needed to understand the specific mechanisms through which this deep-seated racism shaped the geography of electricity generation and continues to influence decisions regarding the siting of pollution sources. Because mechanisms are likely to vary by city and region, such work would benefit from local historical investigations that were beyond the scope of the current analysis.
Present-day emissions burdens reflect siting decisions made decades earlier due to the long lifetime of power plants (median retirement age in our sample was 49 years) and the fact that older plants generally have higher emissions rates ( Table 1). The disparities in emissions burden we observed are thus likely influenced by past siting decisions and variations in emissions due to factors such as size, fuel type, technology and emissions controls. Present-day community socioeconomic status and racial composition are related to the location and emissions of power plants generally and peaker plants specifically 7,40,41 . Peaker plants tend to emit higher rates of NO x and other pollutants per unit energy produced compared with baseload-generating natural gas power plants 42 . We also found increased prevalence of peaker plants upwind of worse-graded neighbourhoods, with the largest disparity between grades D and C and plants sited during the 1940-1969 period (Fig. 3). Although we did not assess this in the current study, differences in emissions burdens could also be driven by recent power plant retirements. Nearly 100 plants have retired at least one coal-fired generator since 2014 7 . Emissions from electricity generation have declined substantially over the last several decades due to more stringent regulations and changing fuel prices that have spurred a transition from coal to natural gas. As a result, the racial gap in ambient concentrations of fine particulate matter attributable to power plants has narrowed significantly since 2000 (although a disparity remains) 43 . Historical inequalities in emissions burden across HOLC neighbourhoods are therefore likely to have been higher than what we observed in the present day.
Our findings are consistent with two prior national studies assessing the relationship between HOLC grade and present-day ambient air pollution. Namin et al. examined respiratory hazard and cancer risks associated with ambient air toxics using data from the 2014 EPA National Air Toxic Assessment (NATA). The authors observed a similar stepwise increase across HOLC grade, with air toxics risks being lowest in A-graded neighbourhoods and highest in D-graded neighbourhoods 18 . Lane et al. assessed year-2010 census block-level estimates of ambient NO 2 and PM 2.5 and similarly found a monotonic association with HOLC grade 19 . However, the pollutant sources considered in NATA and responsible for ambient NO 2 and PM 2.5 are diverse, whereas our study exclusively examined fossil fuel power plants. In a third study in California, Nardone et al. found limited differences in ambient PM 2.5 levels by HOLC grade but much higher diesel PM emissions in worse-graded communities 23 . The lack of substantial differences in ambient PM 2.5 in that analysis likely reflect the fact that PM 2.5 is a regional air pollutant, while, as in our study, differences in emissions-specific sources were apparent. In contrast to these previous studies, we examined emissions burden as opposed to ambient pollutant concentrations. Ambient air pollution attributable to a power plant will vary with distance from the plant, weather, stack height and other factors. In particular, the distribution of secondary PM formed from power plant emissions via chemical reactions in the atmosphere is likely to differ from that of primary PM emissions. SO 2 and NO x can both react with ammonia to create secondary PM. We also did not examine whether the differences in emissions burden that we observed were meaningful in an absolute sense for community health. It is possible that although we found HOLC grade-related differences in emissions burdens, the true health impacts of emissions may follow a different spatial pattern, including that the air quality impacts are experienced much farther away than the distances we considered in our analysis 44 . Our findings should therefore be interpreted not as measures of exposure or health impact but rather magnitude of power plant emissions as driven by the number, type and size of power plants and stringency of emissions controls.
Strengths of this study include its longitudinal nature and national scope. However, we were unable to fully isolate the causal relationship between red-lining and power plant siting given the lack of data on all possible confounders. We did control for the presence of power plants already in operation or likely in the planning stages before the creation of the HOLC maps and, in sensitivity analyses, 1940 population characteristics that may have influenced power plant siting above and beyond red-lining. This decreases, but does not rule out, the possibility that the associations we observed could have been caused by pre-existing differences between HOLC-graded areas in conditions or trends. Although we controlled for the presence of pre-existing power plants, we did not have information on other types of industry present in the 1930s. Prior work suggests the presence of industry or industrially zoned areas influenced HOLC appraisals 45,46 . It is therefore possible that correlations between already-industrialized land and worse HOLC grades in the 1930s may partially explain our results. In addition, other forms of racism in the housing market including racially restrictive covenants, predatory lending, block busting, gentrification and government programs such as Urban Renewal and construction of the Interstate Highway System also reinforced segregation 47,48 and are likely to have shaped the geographic distribution of power generation in ways that we were not able to examine in the present analysis. We also did not assess the siting of other infrastructure projects that may be equally or more consequential for air quality and health than power plants. Construction of the Interstate Highway System in the 1950s, for example, has been shown to have contributed to further racial residential segregation and had a discriminatory impact on communities of colour, particularly urban Black neighbourhoods 49 .
The US Energy Information Administration and US Environmental Protection Agency datasets we relied on in our analysis also contained missing data that may have resulted in misclassification of some power plants and underestimation of present-day emissions. Due to the highly skewed nature of the emissions variables, model residuals were not always normally distributed even after log transformation of the outcome variables. Finally, we did not look at the racial or ethnic composition of HOLC neighbourhoods to assess the extent to which red-lining and/ or power plant siting contributed to racial/ethnic disparities in power plant proximity over time. Future research could assess this question.
In conclusion, racism as codified in historic red-lining maps in the United States appears to have contributed to disparities in proximity to fossil fuel-burning power plants over time and present-day X axes are log scaled. PR is prevalence ratio from Poisson regression models; IRR is incidence rate ratio from negative binomial models. Dotted lines denote the null of no association. Grey shading distinguishes the outcomes.
Article https://doi.org/10.1038/s41560-022-01162-y emissions burdens. While we focus on red-lining maps produced by the federal government, those maps reflected rather than created racist notions of residential property values that were widely held and deeply entrenched-indeed, capitalized upon-by the real estate industry 50 . Our findings highlight the role of racism in the residential housing market as a contributor to environmental disparities and can inform regulatory and political efforts to remedy environmental racism.

HOLC Security Maps
Digitized HOLC neighbourhood boundaries were downloaded as shapefiles from the University of Richmond's Mapping Inequality Project on 20 August 2021 51 . We excluded a small number of E-graded neighbourhoods (n = 4) from further analysis because E was an uncommon grade.

Power plants
Power plants are made up of one or more generators. We constructed our power plant dataset from generator-level information on operable and retired generators and plant-level locations from the 2019 US Energy Information Administration (EIA) Form 860 (https://www.eia. gov/electricity/data/eia860/) (Supplementary Fig. 1). Plants in Alaska and Hawaii were excluded because these areas were not part of the United States at the time the HOLC Security Maps were created; plants in Puerto Rico were excluded because no HOLC maps were available for Puerto Rico. We derived the following plant-level variables: first date of operation, date of retirement, primary fuel source and nameplate capacity from the generator-level datasets. The plant's first date of operation was defined as the first operation date of the earliest operating generator at each plant. A plant was considered retired if all of its generators had retired and the retirement date was the date of the last generator to retire.
A plant's primary energy source was defined as the fuel source reported for the majority of its nameplate capacity on the basis of year 2019 generator-level information (for operable plants) or retirement year (for retired plants). For example, if a plant had three generators of equal nameplate capacity and two had a primary energy source of coal and one of petroleum, the plant's primary energy source was defined as coal. Seven plants had an equal proportion (50%) of nameplate capacity coming from two energy sources. These cases were classified in the following order of priority: coal, petroleum products, natural gas and any other energy sources.
Power plants with a primary energy source other than coal, petroleum products or 'natural gas and other gases' were excluded from analysis. These included plants utilizing nuclear energy or renewable fuels (for example, biomass, solar, wind, geothermal, hydroelectric). For the 1940 baseline, we included a small number of plants (n = 16) whose primary energy source was liquid or solid renewable (biomass) fuel because fossil fuel use was less common during this time.
We used the EIA Data Atlas to identify peaker plants in our dataset (https://atlas.eia.gov/datasets/power-plants-1). We defined peaker plants as plants that currently primarily burn oil and combustion turbines primarily burning natural gas 41 .
We excluded plants that retired before 1939 from the pre-1940 baseline dataset. For all other periods (1940-1969, 1970-1999 and 2000-2019), we included all plants first operational during the period irrespective of retirement date.

Emissions
Data on emissions of NO x and SO 2 were obtained from the US Environmental Protection Agency Emissions & Generation Resource Integrated Database (eGRID) for the most recent year available at the time we conducted the analysis (2019) (https://www.epa.gov/egrid). PM emissions are not currently incorporated into eGRID releases; we used draft eGRID PM 2.5 emissions estimates for 2018 also available from the EPA. eGRID provides unadjusted estimates and estimates adjusted to reflect non-biomass derived emissions associated with electricity generation only (and not heat production). We utilized unadjusted estimates because our interest was in the health burden of emissions rather than their biogenic or non-biogenic source or purpose. Emissions were matched to our power plant dataset using the EIA plant code available in the eGRID datasets. Article https://doi.org/10.1038/s41560-022-01162-y

Population and socioeconomic characteristics
City population size estimates for 1940 were obtained from the Mapping Inequality Project. Baseline socioeconomic and demographic characteristics of HOLC neighbourhoods were obtained from Markley, who extracted and digitized them from HOLC area description sheets 9 . These included the following variables: percent foreign born, percent Black, median household income (in dollars), median building age (in years) and three five-category ordinal variables: occupation, repair status of properties and mortgage availability. City population was available for 8,302 (94%) of HOLC neighbourhoods in our analysis, while all socioeconomic variables were available for 6,220 (70%) of neighbourhoods across 128 cities. We elected to use information from the area description sheets rather than census data as has been used in other red-lining studies because census tract boundaries do not align with HOLC neighbourhood boundaries, and only 60 US cities had census tract boundaries assigned at the time of the 1940 census. Due to the prevalence of missing data, we included these variables only in sensitivity analyses.

Wind direction
We obtained primary monthly forcing data for the years 1980 and 2020 from the North American Land Data Assimilation System (NLDAS-2) (https://daac.gsfc.nasa.gov). Estimates are provided in 0.125° grid spacing and available from 1979 to present. Monthly zonal (u) and meridional (v) winds were used to determine the absolute wind speed (ws) by taking the square root of the summed squared u and v components (√u 2 + v 2 = ws). The absolute wind speed, zonal and meridional wind components were then used to obtain the leeward angle in radians (direction the wind is blowing towards) as follows: ). The result was converted to degrees by multiplying by 180/π and then categorized into four cardinal directions corresponding to N (NW to NE), E (NE to SE), S (SE to SW) and W (SW to NW). We used the mode of monthly observations to define the predominant wind direction at each power plant location and define downwind HOLC polygons. Because (1) data before 1979 were not available, (2) 1980 was the midpoint of our study period (1940-2020) and (3) primary wind directions in 1980 were very consistent with year 2020, we used 1980 to define primary wind direction for all power plants in our analysis.

Analytic strategy
Spatial data were projected using the North American Datum 1983 Contiguous USA Albers coordinate system for the purposes of calculating Euclidean distances between power plants and HOLC polygon boundaries using ArcGIS version 10.5.1 (ESRI). We generated a 5 km buffer around each power plant point location and sliced the buffer into four °90 pieces corresponding to the cardinal directions. The slice corresponding to the downwind wind direction at each power plant's location was then selected to generate the following outcomes based on intersection between the slice and the HOLC polygon boundary: presence or absence of a plant, number of plants and sum of emissions from plants. We additionally calculated the presence of a plant whose primary energy source in 2019 was coal or oil given the higher rate of criteria air pollutant emissions associated with these power plants. We also considered the presence of a peaker plant independently because when they operate, they are disproportionately polluting. We then examined descriptive statistics for each outcome stratified by HOLC grade and period.
We next conducted a series of multivariate regression models with the unit of analysis being HOLC neighbourhood using R version 4.2.0 (https://cran.r-project.org/). We took the conservative approach of comparing adjacent neighbourhood grades rather than using one grade as a reference group to better isolate the impact of red-lining, because A-and D-graded neighbourhoods likely differed more substantially from each other than C and D grades in terms of housing age, socioeconomic status and other factors that could have predicted power plant siting and for which we had limited information. Relative risks of the presence of fossil fuel power plants, coal or oil plants and peaker plants (prevalence ratios) were estimated using Poisson models stratified by time period with robust standard errors. Negative binomial models stratified by time period were used to estimate associations with the number of plants nearby (incidence rate ratios) because of over-dispersion. We used linear regression to estimate associations between HOLC grade the continuous outcomes (NO x , SO 2 and PM 2.5 emissions) after first log-transforming these outcomes to improve the normality of model residuals. The analysis of present-day emissions excluded HOLC neighbourhoods without a power plant that was still operable in 2019 within 5 km.
In all models, we controlled for the presence of a power plant at baseline (before 1940) and US census region (Northeast, Midwest, Article https://doi.org/10.1038/s41560-022-01162-y South or West) as potential confounders. We then undertook a sensitivity analysis in which we additionally controlled for baseline city population size and HOLC neighbourhood socioeconomic variables for the subset of HOLC neighbourhoods for which this information was available. We hypothesized that city size may confound associations with HOLC grade because larger cities were more densely populated, possibly resulting in greater demand for electricity and more densely populated neighbourhoods were more likely to be red-lined. We thought baseline socioeconomic status and demographics might confound our associations of interest because these factors were associated with HOLC grade and may have influenced subsequent siting decisions and factors such as zoning designations. Finally, we conducted a second sensitivity analysis utilizing 10 km rather than 5 km to define nearby upwind power plants.

Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability
All data used in the analysis and to create Figs
Corresponding author(s): Lara J. Cushing Last updated by author(s): Oct 10, 2022 Repor�ng Summary Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency in repor�ng. For further informa�on on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Sta�s�cs
For all sta�s�cal analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods sec�on.

n/a Confirmed
The exact sample size (n) for each experimental group/condi�on, given as a discrete number and unit of measurement A statement on whether measurements were taken from dis�nct samples or whether the same sample was measured repeatedly The sta�s�cal test(s) used AND whether they are one-or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section.
A descrip�on of all covariates tested A descrip�on of any assump�ons or correc�ons, such as tests of normality and adjustment for mul�ple comparisons A full descrip�on of the sta�s�cal parameters including central tendency (e.g. means) or other basic es�mates (e.g. regression coefficient) AND varia�on (e.g. standard devia�on) or associated es�mates of uncertainty (e.g. confidence intervals) For null hypothesis tes�ng, the test sta�s�c (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted

So�ware and code
Policy informa�on about availability of computer code Data collec�on The study relied on secondary data and no so�ware was used to collect data.
For manuscripts u�lizing custom algorithms or so�ware that are central to the research but not yet described in published literature, so�ware must be made available to editors and reviewers. We strongly encourage code deposi�on in a community repository (e.g. GitHub). See the Nature Research guidelines for submi�ng code & so�ware for further informa�on.

Data
Policy informa�on about availability of data All manuscripts must include a data availability statement. This statement should provide the following informa�on, where applicable: -Accession codes, unique iden�fiers, or web links for publicly available datasets -A list of figures that have associated raw data -A descrip�on of any restric�ons on data availability All data used in the analysis and to create Figures 1-5  Field-specific reporting Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Behavioural & social sciences study design
All studies must disclose on these points even when the disclosure is negative.

Study description
Quantitative observational study

Research sample
The study includes (1)

Sampling strategy
Four primary datasets were deemed necessary and sufficient to test the study hypothesis based on their vintage, comprehensiveness, completeness, quality control, geographic and temporal resolution.

Data collection
The study relies on secondary data and no primary data were collected. Researchers were not blinded to the study hypothesis.

Timing
The study relies on secondary data and no primary data were collected.

Data exclusions
Power plants excluded from the analysis based on a-priori inclusion criteria are given in Supplemental Figure 1. Plants in states and territories without redlining maps (Alaska, Hawaii, and Puerto Rico, n=223), plants retired prior to 1940 (n=255), and plants without a fossil fuel as the primary fuel source (n=6,875) were excluded. Non-participation