Introduction

Redlining is a discriminatory practice that arose out of the Great Depression with the formation of the Home Owners Loan Corporation (HOLC) in 1933. In the following years, the HOLC would produce its infamous security maps to guide decisions concerning who among those facing foreclosure could receive HOLC “rescue” mortgages to avoid default. Guided by racist housing practices and public opinion at the time, contractors responsible for creating these maps attributed excessive risk to neighborhoods that housed people of color [39]. The security maps they produced classified neighborhoods as being in one of four different risk categories: grade A, labeled “Best”; grade B, labeled “Still Desirable”; grade C; labeled “Still Declining”; and grade D, labeled “Hazardous”. Areas categorized as grade D were colored red on the maps, giving rise to the term “redlining”. In recent years, these maps have been digitized and georeferenced by researchers at the University of Richmond, providing researchers with the ability to investigate these inequities using GIS [36].

Though the HOLC’s security maps have long since been abolished, their effects continue to persist today, especially in neighborhoods that have been redlined. Researchers have hypothesized various mechanisms through which the HOLC’s security maps continue to influence the present-day makeup of cities, such as through the persistence of differential patterns of housing stock between neighborhoods that were attributed different levels of risk [2]. Studies have found that neighborhoods that were attributed higher risk, such as those that were redlined, were less likely targets for the construction of new homes on account of their residents’ decreased ability to receive credit; today, these same areas are more likely to have older housing stock, fewer housing units, higher proportions of multifamily homes with rented units, and lower housing values [1, 2, 36]. Because HOLC geographies do not perfectly align with Census block groups, we assigned block groups to HOLC grades based on the proportion of encompassed population at the block level, which is the finest resolution for which national population data is available from the Census. First, the population of each Census block was binned into the grade of the encompassing HOLC area, if any, via point-in-polygon spatial join on block centroids. Block groups were then assigned a grade by selecting the bin that contained greater than some percentage threshold of the block group’s total population, if any (Fig. 2). Unlike other methods that assign HOLC grades based on the proportion of intersecting area, this method does not make an assumption of uniform population density across block groups and can more accurately capture what historically-assigned HOLC grade encompasses the most people in present day. In our main analysis, we used a 90% population threshold. Lower thresholds make the apportionment of HOLC grades to present-day block groups more ambiguous, but higher thresholds exclude more block groups due to no HOLC grade bin sizes meeting the population threshold.

Fig. 2
figure 2

Figure demonstrating (A) the size comparison and spatial misalignment between HOLC-graded areas (outlined in black) and Census block groups (2010, outlined in gray), and (B-D) block groups being graded according to population cutoffs of 50%, 90%, and 99%, respectively. Blue areas are grade A, green areas are grade B, yellow areas are grade C, red areas are grade D (“redlined”), dark grey areas are marked as ambiguous (i.e. no grade bin exceeded the threshold), and light grey areas are marked as unclassified (i.e. the unclassified bin exceeded the threshold). White areas are either bodies of water or block groups that did not intersect with HOLC-graded areas

Full size image

Use of mortality records was permitted by the California, Florida, Georgia, Illinois, Indiana, Kansas, Massachusetts, Michigan, Missouri, New Hampshire, New Jersey, Ohio, and Texas departments of public health and approved by the IRBs of each. This study was reviewed by the IRB of the Harvard School of Public Health and classified as not human research.

Statistical analysis

Prior to statistical analyses, deaths were restricted to only those from internal causes (ICD codes A00-R99) involving individuals 18 years of age or older. A case-crossover analysis was used to control for confounding by all fixed or slow-varying factors by design, such as sex, race and ethnicity, smoking history, etc. Specifically, for each case, we sampled control days bidirectionally from the days within the same month of the case that were on the same day of the week as the case to control for all factors that are either fixed on the scale of a month or vary on a cyclic weekly basis. This includes slowly varying individual or neighborhood predictors. Time-varying exposures were then reassessed for each case on each of these control days. Since each individual is compared to themselves at a different point in time, all fixed and cyclic weekly confounding factors are controlled for by design. Additionally, since we sample controls from days occurring both before and after the case, we are able to control for bias arising from time trends in the exposures [35]. In exchange for controlling for these time trends by design by sampling bidirectionally, a small bias is induced by sampling from controls post-death, but the size of this bias is very small due to the low daily risk of death at baseline [29].

In our primary analysis, we fit two sets of conditional logistic regressions within strata of each individual. For the redlining-PM2.5 interaction, we fit the following:

$$clogit\left(Pr\left(\mathrm{death}\right)\right)=\beta_0+\beta_1{\mathrm{PM}}_{2.5}+\left(\sum\limits_{i=2}^5\beta_1\right)ns\left(\mathrm{TMEAN},\;4\right)+\left(\sum\limits_{j=6}^9\beta_j\right)ns\left(\mathrm{VP},4\right)+\beta_{10}redlined\times{\mathrm{PM}}_{2.5}$$

where redlined is an indicator variable for HOLC grade D, PM2.5 is either the ambient concentration of PM2.5 on the day of the death or a moving average up to 4 days before the death (i.e. the 5-day moving average), TMEAN and VP are mean temperature and vapor pressure, respectively, and β10 is the primary estimand of interest. For the redlining-extreme heat interaction, we instead fit the following:

$$clogit\left(Pr\left(\mathrm{death}\right)\right)=\beta_0+\beta_1\mathrm{extreme}+\left(\sum\limits_{i=2}^5\beta_i\right)ns\left(\mathrm{VP},4\right)+\beta_6redlined\times\mathrm{extreme}$$

where extreme is an indicator variable either for extreme heat or for the 1st, 2nd, 3rd, or 4th day of extreme heat occurring on the day of the death and β6 is the primary estimand of interest.

To investigate the robustness of our findings, we carried out several sensitivity analyses. Firstly, we alternatively considered different block group-level HOLC grade apportionments based on cutoffs of 50% and 99% of the block group-level population. Secondly, we also alternatively considered 85th and 99th percentile cutoffs of minimum temperature in our definition of extreme heat. Thirdly, we carried out subgroup analyses within Black individuals, Black neighborhoods, White individuals, and White neighborhoods, where Black and White neighborhoods were defined as block groups where the proportion of residents that identified as Black or White was 50% or greater to investigate variation in vulnerability across individual- and area-level demographics. Lastly, we carried out subgroup analyses by year and by state to investigate spatial and temporal variation in vulnerability.

Results are presented for each 10 µg m−3 change in PM2.5 concentration. R version 4.1.0 was used for all analyses [38].

Results

We obtained 11,115,380 mortality records from the twelve state departments of public health. From these records, we sequentially excluded 466,874 deaths involving external causes; 139,908 deaths involving individuals younger than 18 years old; 196,558 deaths with geocodes that were missing or coarser than block group-level; 331 deaths involving individuals whose home locations were outside of the state that reported their death; 1,392,423 deaths before January 5th, 2001 or after December 31st, 2016 and 537 deaths whose home block groups had a population of zero according to the preceding Decennial Census (for which 4-day moving averages of population-weighted PM2.5 could not be calculated); and 34,016 deaths with lag days from 0 to 4 that included December 31st on leap years (for which Daymet predictions are not available; Fig. 3). This resulted in a final data set of 8,884,733 records.

Fig. 3
figure 3

Subject restriction flowchart

Full size image

Baseline characteristics are shown in Table 1. In the full data set, cases mostly involved individuals who were white (85.72%), had high school education alone (41.15%), and lived in areas not classified by the HOLC (84.35%, generally established post-1935). Of those that lived in areas that were classified, grade C was the most frequent, followed by grades D and B; deaths in areas graded as A represented less than 1% of records. Of those that lived in areas that were not classified, the vast majority lived in block groups that did not touch the areas assessed by the HOLC (99.24); the remainder were assigned as unclassified using the apportionment algorithm. There were also 53 deaths that occurred in areas classified as E which were left as-is since there is no indication as to what this classification could signify. The mean age of cases was 75.96 years with a standard deviation of 14.68 years.

Table 1 Baseline characteristics of mortality cases, showing mean ± SD or N (%). Deaths that occurred in block groups
Full size table

Cases involving individuals who lived in historically-redlined areas (HOLC class D, “Hazardous”) comprised 2.13% of all observed cases. This subpopulation had a higher proportion of people of color (45.95%, compared to 13.91% in the full population) and individuals who had an education level of high school or less (73.74%, compared to 63.27%). The average age of individuals at the time of death was also lower (72.55 years old, compared to 75.96). Higher proportions of cases in this subpopulation came from the states of California, Illinois, Indiana, Kansas, Massachusetts, Michigan, and New Jersey.

Assigned environmental exposures at the time that each case was reported are shown in Table 2. In the full data set, the mean ambient concentration of PM2.5 was 9.61 µg m−3 (SD 5.72 µg m−3) on the day of the case and 9.59 µg m−3 (SD 4.51 µg m−3) for the moving average comprising the day of the case and the 4 preceding days. 9.57% of cases occurred on extreme heat days, of which 32.63% occurred on the first day of extreme heat. Ambient PM2.5 both on the day of the case and in the 5-day moving average were higher for individuals who lived in historically-redlined areas (10.95 µg m−3; SD 6.38 µg m−3 and 10.91 µg m−3; SD 4.67 µg m−3, respectively). Additionally, a slightly larger proportion of deaths from historically-redlined areas occurred on days of extreme heat (9.71%). Both ambient temperature and vapor pressure were higher among all included individuals vs. cases among individuals from historically-redlined areas.

Table 2 Assessed exposures, showing mean ± SD or N (%)
Full size table

The estimated effects of exposure to extreme heat on mortality within and outside of historically-redlined neighborhoods are shown in Fig. 4 and Supplement A. In general, results suggested that living in a historically-redlined neighborhood increases susceptibility to death by exposure to extreme heat. We found a significant interaction with exposure to any extreme heat (interaction odds ratio 1.0218; 95% CI 1.0031, 1.0408) while we did not observe significant interactions for singleton heat events or when looking at length-specific exposures. In absolute terms, this amounts to a 2.157% (95% CI 0.307%, 4.036%) increase in the daily risk of death death from non-external causes by exposure to any extreme heat in historically-redlined neighborhoods compared to other neighborhoods. The highest overall effects were observed for exposure to any extreme heat, followed by 3, 1, and 2 consecutive days of extreme heat, respectively.

Fig. 4
figure 4

Estimated odds ratios of mortality from non-external causes due to exposure to any extreme heat or the nth consecutive day of extreme heat, within and outside of historically-redlined areas

Full size image

The estimated effects of exposure to ambient PM2.5 on mortality within and outside of historically-redlined neighborhoods are shown in Fig. 5 and Supplement B. As with extreme heat, we found that living in a historically-redlined neighborhood increases susceptibility to death by exposure to ambient PM2.5. We found a significant interaction with same-day ambient PM2.5 (interaction odds ratio for each 10 µg/m−3 increase: 1.0093; 95% CI 1.0084, 1.0101) while we did not observe interactions for different moving averages of ambient PM2.5. In absolute terms, this amounts to a 0.930% (95% CI 0.831%, 1.000%) increase in the daily risk of death from non-external causes for each 10 µg/m−3 increase in ambient PM2.5 in historically-redlined neighborhoods compared to other neighborhoods. However, the point estimates for interactions with 2- to 5-day moving averages of ambient PM2.5 were similar. The highest overall effect was observed for same-day exposure and the lowest overall effect was observed for the 5-day moving average.

Fig. 5
figure 5

Estimated odds ratios of mortality from non-external causes for each 10 µg/m−3 increase in ambient PM2.5, or the average ambient PM2.5 across multiple days, within and outside of historically-redlined areas

Full size image

Results from our sensitivity analyses considering different cutoffs of population for the apportionment between HOLC geographies and Census block groups are shown in Figs. 6 and 7 and Supplements A and B. Among each combination of exposure and exposure window, we did not observe significant differences between the different cutoffs. However, for PM2.5, we did observe that the interaction with same-day ambient PM2.5 was not significant for population cutoffs of 50% and 99%. We also observed that, for PM2.5, estimates using a population cutoff of 50% were smaller.

Fig. 6
figure 6

Estimated interaction odds ratios of mortality from non-external causes due to exposure to any extreme heat or the nth consecutive day of extreme heat, using different cutoffs of population for HOLC grade apportionment

Full size image
Fig. 7
figure 7

Estimated interaction odds ratios of mortality from non-external causes for each 10 µg/m−3 increase in ambient PM2.5, or the average ambient PM2.5 across multiple days, using different cutoffs of population for HOLC grade apportionment

Full size image

Results from our sensitivity analyses considering different cutoffs of minimum temperature for the determination of what constitutes extreme heat are shown in Fig. 8 and Supplement C. We observed that, for the most part, interactions were similar across the different cutoffs. We also observed that the 85th and 95th percentile cutoffs of minimum temperature had higher interactions than the 90th percentile cutoff, with the 85th percentile being the highest for any exposure and the 99th percentile being the highest for the 1st day of extreme heat.

Fig. 8
figure 8

Estimated interaction odds ratios of mortality from non-external causes due to exposure to any extreme heat or the nth consecutive day of extreme heat, using different cutoffs of minimum temperature

Full size image

Results from our sensitivity analyses by individual- and area-level demographics are shown in Figs. 9 and 10 and Supplements D and E. For extreme heat, we found that Black individuals and individuals from Black neighborhoods tended to be less susceptible while White individuals and individuals from White neighborhoods were more susceptible. In particular, we found significant interactions between exposure to extreme heat and both self-identification as White and living in a White neighborhood while the corresponding interactions for self-identification as Black and living in a Black neighborhood were close to null. For PM2.5, interactions were more similar among the different subgroups.

Fig. 9
figure 9

Estimated interaction odds ratios of mortality from non-external causes due to exposure to any extreme heat or the nth consecutive day of extreme heat, restricted to different subsets of the population by individual- and area-level demographics

Full size image
Fig. 10
figure 10

Estimated interaction odds ratios for mortality from non-external causes for each 10 µg/m−3 increase in ambient PM2.5, or the average ambient PM2.5 across multiple days, restricted to different subsets of the population by individual- and area-level demographics

Full size image

Results from our sensitivity analyses by state are shown in Figs. 11 and 12 and Supplements F and G. In general, we did not observe significant heterogeneity in the interactions between living in a historically-redlined neighborhood and exposure either extreme heat or ambient PM2.5, though the interactions between historical redlining and ambient PM2.5 tended to be stronger in Indiana.

Fig. 11
figure 11

Estimated interaction odds ratios of mortality from non-external causes due to exposure to any extreme heat or the nth consecutive day of extreme heat by state

Full size image
Fig. 12
figure 12

Estimated interaction odds ratios of mortality from non-external causes for each 10 µg/m−3 increase in ambient PM2.5, or the average ambient PM2.5 across multiple days by state

Full size image

Results from our sensitivity analyses by year are shown in Figs. 13 and 14 and Supplements H and I. As with our sensitivity analyses by state, there was no clear heterogeneity in the interactions between living in a historically-redlined neighborhood and exposure to either extreme heat or ambient PM2.5 by year. There also did not appear to be any consistent time trend in those interactions, though both exposures exhibited suggestive cyclic patterns with multi-year periods.

Fig. 13
figure 13

Estimated interaction odds ratios of mortality from non-external causes due to exposure to any extreme heat or the nth consecutive day of extreme heat by year

Full size image
Fig. 14
figure 14

Estimated interaction odds ratios of mortality from non-external causes due for each 10 µg/m−3 increase in ambient PM2.5, or the average ambient PM2.5 across multiple days by year

Full size image

Discussion

In this study, we found that some disparities in mortality risks due to exposures to ambient air pollution and extreme heat experienced by individuals living in previously-redlined areas persist nearly a century after the initial creation of the HOLC security maps, and that the injustices fostered by these maps go beyond the quantity of exposure itself and include differential susceptibility. Using a case-crossover design, we were able to control for all time-invariant and individual-level confounders and demonstrated that living in a previously-redlined area has synergistic effects with both ambient PM2.5 and exposure to extreme heat on mortality by non-external causes. These findings have implications for policy going forward, as these results suggest that redlined communities will experience air pollution- and extreme heat-related health disparities even after local air pollution levels are brought to levels comparable to not-redlined communities or interventions to reduce heat through green space and reflective sidewalk installations have been implemented.

Notably, 53.72% of the deaths in historically-redlined communities involved White individuals, so these findings are not simply an effect of larger effect sizes in minority population. This was further confirmed by our sensitivity analyses – we did not find any significant differences between the response of Black or White individuals, individuals from present-day majority Black or White neighborhoods, and the population as a whole. Additionally, we found that our results were consistent for different definitions of what Census block groups count as having been redlined and for different definitions of extreme heat.

Interestingly, we observed that the interaction between living in a historically-redlined neighborhood and exposure to extreme heat was stronger in White individuals. It’s unclear what could explain this finding and more work is needed to investigate its robustness and determine potential mechanisms. Previous studies have shown that Black and White individuals spend similar amounts of time outdoors and that Black neighborhoods may have more outdoor amenities, though these amenities tend to be of worse quality, so this may not be an effect of behavioral differences between these subpopulations [5, 11, 18, 23].

Previous studies also identified acute effects of extreme heat on mortality ([3, 22, 30], p. 50). Simultaneously, previous studies have also identified acute effects of PM2.5 [17, 19, 24, 42], including studies looking at lags up to 30 or 40 days [47, 48]. The present study complements these prior analyses by contributing evidence that historical redlining modifies those effects, which is an important finding for environmental justice concerns. Moreover, previous literature has found that the effects of both extreme heat and PM2.5 exposures are more severe in Black individuals [6, 13, 30, 40, 41, 45], though findings are conflicted. In our sensitivity analysis, we find that the effect modification persists in individual- and neighborhood-level demographic subsets, suggesting that the effect is not simply one of neighborhood composition but rather represents lasting, structural impacts of historic redlining.

The present study has a few major strengths, the most significant being the use of a case-crossover analysis to control for confounding, and the focus on block-groups instead of the more common city level models. By using a case-crossover analysis rather than more traditional epidemiologic methods of confounding control, we were able to significantly limit the number of potential uncontrolled confounders by design. Previous case-crossover studies conducted at a city level used citywide means of PM2.5 or extreme heat, introducing substantial exposure error. Using block-group level exposure reduces this error, better captures urban heat islands, and local adaptation to prevailing temperature. This study also benefits from the generalizability of the data source used – rather than looking at a subset of the population, the death records used encompass the entire population of people who have died in each state during the years collected. Another strength of our study is that our findings remained robust across extensive sensitivity analyses.

This study also has a few limitations. Chiefly, air pollution and meteorology were assessed at the home block group of each individual, which may not have captured their true exposures if they regularly commuted to a work far from home. However, because we looked at acute effects, and the mean age of our population was 76 years old, this misclassification issue is less of a concern. It is also possible that additional factors that influence both air pollution and redlining or both extreme heat and redlining that were not adjusted for may have confounded these findings and biased these results. Additionally, though our data includes deaths from several states across the country with different physical and sociopolitical environments, they are not all-encompassing and it is possible that estimates could be different in areas that we missed, such as in the Pacific Northwest and Great Plains regions. Lastly, these findings also do not identify any single causative agent, as redlining can influence mortality and chronic psychosocial stress in a variety of different ways. This is an important area of future research, as identifying these causative agents will be crucial to designing effective interventions to reduce disparities.

Conclusions

We have observed that the actions of the HOLC nearly a century ago that upheld structural discrimination through the creation and distribution of its racist security maps are still felt today, and individuals living in affected areas will continue to experience extreme heat- and air pollution-related health disparities even after these adverse environmental agents are mitigated. This study highlights the urgent need for future investigation into the specific causal agents driving this health disparity in order to design specific, targeted interventions that can address both extreme heat and air pollution as well as socioeconomic inequalities present in disadvantaged neighborhoods. In the meantime, interventions to reverse the impact of redlining in general, such as efforts to reduce local PM2.5 by increasing access to alternative forms of transportation or plant trees to reduce the effect of urban heat, are a good start. In general, our findings suggest that interventions that focus on the environment alone may not be able to fully achieve environmental health equity and a more holistic approach may be more well-suited to achieving these goals.