Simulating desegregation through affordable housing development: an environmental health impact assessment of Connecticut zoning law

Abstract

Residential segregation drives exposure and health inequities. We projected the mortality impacts among low-income residents of leveraging an existing 10% affordable housing target as a case study of desegregation policy. We simulated movement into newly allocated housing, quantified changes in six ambient environmental exposures, and used exposure-response functions to estimate deaths averted. Across 1,000 simulations, in one year, we found on average 169 (95% CI: 84, 255) deaths averted from changes in greenness, 71 (49, 94) deaths averted from NO₂, 9 (4, 14) deaths averted from noise, 1 (1, 2) excess death from O₃, and 2 (1, 2) excess deaths from PM_2.5, with rates of deaths averted highest among non-Hispanic Black and non-Hispanic White residents. Strengthening desegregation policy may advance environmental health equity.

1. Introduction

Residential segregation has a profound impact on health and well-being in the U.S. Notably, segregation is associated with inequities in preterm birth, infectious disease, cancer, asthma, mental health, and other health outcomes (Anderson et al., 2021; Krieger et al., 2020; Nardone et al., 2020). In the United States, residential segregation is rooted in a history of exclusionary zoning and lending practices. This is most emblematic via “redlining,” a practice in 1930-1934 through which the Home Owners’ Loan Corporation designated predominantly Black or low-wealth neighborhoods as “hazardous” for mortgage investment while grading predominantly White or affluent neighborhoods as “best”, reinforcing racialized economic segregation (Krieger et al., 2020; Nardone et al., 2020; Nardone et al., 2021). Despite the 1968 Fair Housing Act, segregation is far from declining - a recent report from Menendian et al. (2021) found that 54% of U.S. metropolitan regions were more highly segregated in 2020 than in 1990 (Steil et al., 2021).

Residential segregation in large part persists due to exclusionary zoning mechanisms. Presently, local zoning ordinances that reserve residential land exclusively for single-family detached housing perpetuate residential segregation by blocking the development of more affordable, high-density units, such as duplexes and apartments, in areas of accumulated wealth and social capital (Girouard, 2023; Lens, 2022; Steil & Lens, 2023; Whittemore, 2020). Given the low supply of housing nationwide, these exclusionary ordinances push a disproportionate burden of much-needed development onto neighborhoods which tend to be those most impacted by disinvestment and thereby made vulnerable to gentrification (Whittemore et al., 2020). Through these mechanisms, residential segregation continues to shape access to housing, wealth, education, nutrition, and other place-based resources integral to long-term health (Appel & Nickerson, 2016; Eaton, 2020; Havewala, 2021; Menendian et al., 2021).

These exposure inequities are further compounded by environmental racism, which describes the disproportionate exploitation of communities who are predominantly Black, Indigenous, Hispanic or Latino (herein referred to as Latino), and other People of Color through planned allocation of environmental hazards (e.g., pollution, extreme heat, noise) and environmental goods (e.g., greenspace) (Carrión et al., 2022; Seamster & Purifoy, 2020). As a result, segregation concentrates adverse social and environmental exposures in neighborhoods with predominantly low-wealth, Black, or Latino residents, while concentrating health-promoting resources in neighborhoods with predominantly White or affluent residents (Carrión et al., 2022; Morello-Frosch, 2011; Seamster & Purifoy, 2020). Policies that promote integration are therefore necessary to advance environmental justice and health equity.

Policy mechanisms for integration can generally be categorized as people-based or place-based approaches (Steil & Lens, 2023). People-based mechanisms, such as the Housing Choice Voucher program established by federal Section 8, typically aim to reduce individual barriers to mobility among low-wealth and minoritized residents (Steil & Lens, 2023). In contrast, place-based desegregation mechanisms generally seek to redistribute access to place-based resources by, for instance, increasing development of high-density, affordable housing in areas of concentrated wealth (Steil & Lens, 2023). Place-based approaches to desegregation may advance environmental and health justice through multiple concurrent pathways. Expanding the supply of affordable housing in high-resource neighborhoods is not only critical to create opportunities for integrative movement, but is also necessary to control housing costs and in turn reduce rent burden and housing insecurity, both of which disproportionately impact low-income and minoritized residents (Flint, 2022; Harvard Law Review, 2022; Lens, 2022; Swope & Hernandez, 2019). Densification, alongside other zoning interventions such as transit-oriented development, also has promising environmental co-benefits, in the form of reduced vehicle and building emissions (Harvard Law Review, 2022). Yet, despite growing recognition that increased development of affordable, densified housing is necessary to expand access to health-promoting environments, curb inflated housing costs, and mitigate emissions, several recent state-level attempts to legislate sweeping zoning reforms have been cut short or diluted by opponents of state zoning preemption (Harvard Law Review, 2022; Sisson, 2023).

One longstanding state zoning law, Section 8-30g of the Connecticut General Statutes, provides a case study of place-based policy that aims to equitably increase affordable housing supply (CT Gen Stat, 2022; Desegregate Connecticut, n.d.; Steil & Lens, 2023). Connecticut is highly segregated - two-thirds of Black and Latino residents lived in only 15 out of 169 municipalities in 2010 (Eaton, 2020). Section 8-30g, passed in 1989, established the Affordable Housing Land Use Appeals Procedure, through which developers can appeal local zoning decisions that deny or significantly restrict affordable housing proposals (CT Gen Stat, 2022; Desegregate Connecticut, n.d.). Contrary to the status quo, Section 8-30g shifts the burden of proof in these appeals off developers and onto municipalities, which must prove that the rejection is grounded in public health or other valid considerations and that these issues outweigh the need for affordable housing (CT Gen Stat, 2022; Desegregate Connecticut, n.d.). This procedure presents a valuable incentive for developers to pursue affordable housing development, as well as a valuable disincentive for municipalities to wield frivolous restrictions to deny these proposals, as municipalities have won only about one-third of Section 8-30g appeals (Desegregate Connecticut, n.d.). Additionally, Section 8-30g further incentivizes development by exempting municipalities if at least 10% of their housing stock is affordable, or if they qualify for a moratorium by demonstrating adequate progress (CT Gen Stat, 2022; Desegregate Connecticut, n.d.). Yet, despite this incentive, only 29 out of 169 Connecticut municipalities met the 10% benchmark in 2022 (Zaldonis & Connecticut Department of Housing, 2023). Therefore, an important policy question remains as to whether full realization of the 10% target codified by Section 8-30g has the potential to promote desegregation and advance environmental health equity.

The aim of this study was to assess potential mortality reductions among low-income Connecticut residents from strengthening desegregation policy to enforce this 10% benchmark across all townships. We quantified all-cause deaths averted from six ambient environmental exposures, stratified by ethnoracial group. We hypothesized that desegregation policy implementation would avert mortality among low-income households across all ethnoracial groups from changes in each of the following ambient exposures: fine particulate matter (PM_2.5), ozone (O₃), nitrogen dioxide (NO₂), summertime daily maximum heat index, greenness, and road traffic noise. We also hypothesized that the rate of all-cause deaths averted would be higher among non-Hispanic Black and Latino low-income households compared to non-Hispanic White and Asian low-income households.

2. Methods

2.1. Summary

We allocated new affordable housing units to achieve a housing supply within each township in which 10% of total units were designated as affordable. We then simulated random movement of low-income households between townships, based on availability and proximity of new housing, and quantified changes in each ambient exposure, averaged at the township level. Next, we used published exposure-response functions to estimate deaths averted over one year associated with changes in each exposure, stratified by ethnoracial group. Additionally, we assessed changes in racialized segregation by calculating a multi-group dissimilarity index by county, and we tested the sensitivity of our results to our mobility model and year of analysis (2018 versus 2019).

2.2. Environmental data

Our analysis focused solely on ambient ground-level averages rather than indoor environmental parameters because we did not have access to indoor environmental exposure models. The spatial unit of analysis was the township, since our aim was to model enforcement of the town-level 10% benchmark established by Connecticut’s Section 8-30g. Townships are also meaningful to how individuals identify their place of residence and are therefore advantageous for marking relocations. We retrieved point estimates of daily average PM_2.5 levels and daily 8-hour maximum O₃ levels by census tract centroid in 2019 from the U.S. Environmental Protection Agency’s Fused Air Quality Surface Using Downscaling datasets (U.S. Environmental Protection Agency, 2023). Daily values were averaged across the year, and tracts were assigned to townships based on geometric centroid locations. Tract-level PM_2.5 and O₃ annual estimates were then averaged across each township and population-weighted using American Community Survey (ACS) 5-year population estimates (Figure A.1.a,b) (U.S. Census Bureau, 2019i).

NO₂ annual averages were retrieved from NASA’s Nitrogen Dioxide Surface-Level Annual Average Concentrations V1 dataset in 0.0083 degree gridded estimates (Figure A.1.c) (Anenberg, 2023). Gridded 1 x 1-km daily maximum air temperature and vapor pressure from NASA’s Daymet datasets were used to calculate daily maximum heat index (Figure A.1.d) (Anderson et al., 2013; Thornton et al., 2022; Thornton et al., 2000; Thornton & Running, 1999; Thornton et al., 1997; Thornton et al., 2021). Additionally, we retrieved 1 x 1-km gridded estimates of monthly average normalized difference vegetation index (NDVI) from NASA’s Terra MODIS Vegetation Indices (MOD13A3.061) (Figure A.1.e) (Didan, 2021). Finally, 30 x 30-m gridded road traffic noise estimates in 2018 and 2020 were retrieved from the Bureau of Transportation Statistics National Transportation Noise Map, and linear imputation was used to estimate 2019 values (Figure A.1.f) (U.S. Department of Transportation, 2022). Population-weighted annual or daily averages were derived for all gridded exposure estimates (NO₂, heat, NDVI, noise) by township using the Columbia University Center for International Earth Science Information Network’s (CIESIN) 2020 Gridded Population of the World (CIESIN, 2018).

2.3. Housing, demographic, and health data

We retrieved counts of affordable and total housing units by township from the Affordable Housing Land Use Appeals List through Connecticut Open Data (Figure A.2.a,b) (Zaldonis & Connecticut Department of Housing, 2023). Households by income bracket, town, and ethnoracial group (Asian, Latino, non-Hispanic Black, and non-Hispanic White) and average household size by town were retrieved from 2020 Decennial Census and 2019 ACS 5-year estimates (U.S. Census Bureau, 2019a–e, 2020). Low-income status was assigned using Department of Housing and Urban Development (HUD) fiscal year 2019 three-person household income limits, based on statewide average household size (U.S. Census Bureau, 2019a; HUD, 2023). Finally, we retrieved statewide crude all-cause mortality rates per 100,000 people by ethnoracial group, year, and month from the Centers for Disease Control and Prevention Wide-ranging Online Data for Epidemiologic Research (CDC WONDER) database (National Center for Health Statistics, 2022).

2.4. Allocating new affordable housing units

We first allocated new affordable housing units to each township until 10% of total units were affordable. Affordable housing, as defined by the Department of Housing, includes: government-assisted housing, housing with tenants receiving rental assistance, housing financed by single-family Connecticut Housing Finance Authority and/or U.S. Department of Agriculture mortgages, and housing with deed restrictions pricing them as affordable for households earning 80% or less of the area median income (AMI) (Zaldonis & Connecticut Department of Housing, 2023). Townships where less than 10% of housing units were affordable at baseline were allocated the minimum integer number of affordable units necessary so that the final proportion of affordable units was no less than 10%. Townships with 10% or more of their housing stock designated as affordable at baseline were allocated 0 additional units, i.e. their baseline housing stock was kept.

2.5. Simulating population movement and mortality impacts

We simulated random movement of low-income households in response to this new affordable housing. We then estimated all-cause deaths averted from subsequent changes in each environmental exposure, using randomly drawn exposure-response functions derived from published effect estimates. The simulation was repeated 1,000 times; all steps are summarized in Figure 1, and detailed explanations are provided in Section 2.5.1 and Section 2.5.2.

Figure 1.

Flow chart of steps and inputs used in each simulation of residential movement and estimation of mortality impacts, repeated 1,000 times.

Parallelograms represent inputs, and rectangles represent processes. Acronyms and symbols are defined as follows: HUD, Department of Housing and Urban Development. HMFA, HUD Metropolitan Fair Market Rent/Income Limits Area. AMI, area median income. AF, attributable fraction. HR, hazard ratio. RR, rate ratio, μ, mean, σ, standard deviation. SE, standard error.

Sizing: Single-column.

In print: Black and white requested.

2.5.1. Simulating movement of low-income households

We simulated random movement or non-movement of low-income households by ethnoracial group across towns into hypothetical new affordable units. Low-income households were defined as households earning an annual income at or below 80% AMI in fiscal year 2019 in the HUD Metropolitan Fair Market Rent/Income Limits Area (HMFA) of residence (HUD, 2023).

Step 1:

We first listed out every possible pair of pre-move and post-move towns, each representing a potential move (e.g. Town 1 to Town 2) or, in the case of same-town pairs, non-move (e.g. Town 1 to Town 1). Using population estimates described in Section 2.3., counts of low-income households at baseline by ethnoracial group and income bracket were assigned by pre-move town as the initial starting population for each potential move.

Step 2:

Households were then restricted from moving to towns contained in HMFAs where their estimated income would no longer fall at or below the low-income limit (80% AMI) post-move, under the assumption that these households would not be eligible for affordable housing in towns where they are not considered low-income. Therefore, for a given potential move, households within an income bracket with an upper bound exceeding 80% AMI in the post-move town’s HMFA were considered ineligible for that move, and the count of households in that bracket eligible for that move was set to 0. Households in any other income bracket were considered eligible for that move and their baseline count was kept.

Step 3:

Counts of eligible low-income households were then summed across all income brackets by ethnoracial group for each potential move. Each resulting count represented the final pool of eligible pre-move households by ethnoracial group from which households were selected to make a given move or non-move.

Step 4:

For each ethnoracial group in a given pre-move town, we simulated movement into each potential post-move town sequentially. For the first potential move, we drew a random sample from the binomial distribution with a sample size equal to the count of pre-move town households (n_initial) in that ethnoracial group and a success probability (p) weighted by availability of and proximity to new housing in the post-move town. The number of successes (n_selected) drawn represented the number of households selected to make this move. Availability was measured using the proportion of new units allocated in the post-move town out of all units allocated. Proximity was represented by the distance between population-weighted town centroids, and an inverse distance weighting penalty was used to model the assumption that households were more likely to relocate to shorter rather than longer distances, given potential community and workplace ties (Gillespie, 2017, pp. 89-125; Moeckel, 2016). More detailed weighting methods are described in Text A.1. These assumptions align with parameters discussed by Moeckel (2016) and operationalized in the Simple Integrated Land Use Orchestrator (SILO) model of land use and household relocation (Moeckel et al., 2018). This model, for instance, considers the affordability of a new housing unit, given a household’s budget, as one necessary condition for movement to that unit, and similarly our model was premised on availability of new affordable units (Moeckel, 2016; Moeckel et al., 2018). Our method of penalizing moves over longer distances also aligns with SILO’s commute time constraint as it operationalized individuals’ ties to their starting locations (Moeckel, 2016; Moeckel et al., 2018). We further assessed this constraint in our sensitivity analysis, discussed in Section 2.7.

Step 5:

For each subsequent post-move town (town_j) from a given pre-move town and ethnoracial group, we again drew a random sample from the binomial distribution with the success probability described above and a sample size equal to the remaining pool of pre-move households not yet selected across previous moves in that ethnoracial group and town $({\text{n}}_{\text{initial}}-{\sum }_{i=1}^{j-1}{\text{n}}_{\text{selected}})$.

Step 6:

Households not selected for any of the 169 potential moves for a given pre-move town and ethnoracial group $({\text{n}}_{\text{initial}}-{\sum }_{i=1}^{169}{\text{n}}_{\text{selected}})$ were assigned to stay in their pre-move town.

After Steps 4-6 were completed for all ethnoracial groups in all pre-move towns, counts of households selected for each move were used to estimate deaths averted in one run of the simulation, as described in Section 2.5.2, to be averaged across 1,000 runs.

2.5.2. Assessing mortality impacts

For each household, we quantified changes in exposure to summertime daily maximum heat index and annual average PM_2.5, O₃, NO₂, NDVI, and road noise at the township level. We used the following exposure-response functions to project changes in all-cause mortality associated with each exposure:

• Hazard ratio of 1.12 (95% confidence interval: 1.08, 1.15) per 10 μg/m³ increase in long-term PM_2.5 (Pope et al., 2019)

• Hazard ratio of 1.02 (1.01, 1.04) per 10 ppb increase in long-term O₃ (Turner et al., 2016)

• Hazard ratio of 1.06 (1.04, 1.08) per 10 ppb increase in long-term NO₂ (Huang et al., 2021)

• Rate ratio of 1.014 (1.004, 1.024) per 5° F increase in daily maximum heat index, between 75° F and 105° F (Wellenius et al., 2017)

• Hazard ratio of 0.96 (0.94, 0.97) per 0.1 increase in NDVI within 500 m (Rojas-Rueda et al., 2019)

• Risk ratio of 1.05 (1.02, 1.07) per 10 dB increase in road traffic noise (Hao et al., 2022)

In each simulation repeat, we randomly drew an effect estimate from a normal distribution centered at the rate ratio for daily maximum heat index or the natural log of the hazard or risk ratio for PM_2.5, O₃, NO₂, NDVI, and noise, with a standard deviation equal to the published margin of error divided by 1.96.

We calculated the attributable fraction for each annual exposure and for each pre-move (i) and post-move (j) town pair as follows, based on the pre/post change in exposure (exposure_j – exposure_i), divided by the exposure change (x) corresponding to the function (10 units for PM_2.5, O₃, NO₂, and noise; 0.1 unit for NDVI), and the randomly drawn coefficient for a given run (k) of the simulation (log(HR)_k):

$$1-e^{({\mathit{exposure}}_j-{\mathit{exposure}}_i)/x\ast (\log {(HR)}_k)}$$ (1)

For daily maximum heat index, we calculated the attributable fraction for each day as follows, using the randomly drawn rate ratio for a given run of the simulation (RR_k) corresponding to a 5° F increase:

$$({\mathit{exposure}}_j-{\mathit{exposure}}_i)/5\ast (1-R{R}_k)$$ (2)

We then multiplied the attributable fraction by the number of households experiencing a given move, average household size in the pre-move town, and crude all-cause annual or daily mortality rate by ethnoracial group to estimate deaths averted in each potential move (Zocchetti, 2022). These were summed across all moves and, for daily deaths averted from heat, across all days in May 1-September 30, to estimate total deaths averted in one year from each exposure. We repeated this simulation 1,000 times to derive average estimates and 95% confidence intervals for deaths averted and rates of deaths averted per 100,000 low-income residents.

2.6. Segregation indices

We computed multi-group Duncan’s dissimilarity index at baseline and in the counterfactual population distribution to assess changes in the magnitude of racialized segregation (Sakoda, 1981; Tivadar, 2019). We estimated this index at the county level, since dissimilarity index relies on summing population proportions in smaller administrative units (townships) across larger units (counties) (Sakoda, 1981; Tivadar, 2019).

2.7. Sensitivity analyses

First, we tested the sensitivity of our results to our method of inverse distance weighting. Given that all eight Connecticut counties are considered one commuting zone, it is plausible that intra-state distances would not restrict movement to the extent assumed in the primary analysis (Fowler & Jensen, 2020; U.S. Department of Agriculture, 2012). We thus reproduced the analysis treating all relocation distances within Connecticut as equally probable by removing our inverse distance weighting.

Additionally, we tested an alternative mobility model by penalizing movement probabilities based on drive time instead of distance, under the assumption that households are likely to stay within a reasonable commute time of their existing workplaces (Moeckel, 2016). Based on the 2019 statewide average travel time to work of 26.6 minutes, isochrones delineating a 26.6-minute drive time radius starting at each population-weighted town centroid were retrieved via Open Source Routing Machine (OSRM) (Giraud, 2022; Luxen & Vetter, 2011; Project OSRM, n.d.; U.S. Census Bureau, 2019f). A second isochrone at each weighted centroid delineated a 60-minute radius. For each pre-move town, the base probability of movement to a post-move town whose weighted centroid lay within a 26.6-minute drive of the pre-move weighted centroid was set to approximately 2.44 times the base probability of movement to a post-move town whose weighted centroid lay between a 26.6-minute and 60-minute drive. This reflected the ratio of the proportion of reported commute times below 30 minutes to the proportion of commute times between 30 and 60 minutes, derived from 2019 statewide ACS 5-year estimates (U.S. Census Bureau, 2019h). These probabilities were again weighted by new housing availability, as described in Text A.1. The probability of movement to a town whose population-weighted centroid lay outside a 60-minute drive was set to 0, thereby imposing a stronger mobility restriction than that of the primary analysis. This 60-minute limit also reflects a more conservative boundary than that included in the SILO housing relocation parameters, which limit households to relocating within a 200-minute drive (Moeckel, 2016).

Finally, we tested the sensitivity of our results to the year of analysis by running the simulation using exposure, population, housing, income limit, and mortality data from 2018 instead of 2019.

2.8. Computation

All analyses were conducted using R version 4.3.2. R packages used for data retrieval, cleaning, and analysis are specified in Text A.2 (Anderson et al., 2013; Giraud, 2022; Hufkens et al., 2018; Richardson, 2022; Tivadar, 2019; Walker et al., 2024; Walker & Rudis, 2024).

3. Results

3.1. Sample Characteristics

In 2019, we estimated there were 1,345,504 households in Connecticut with Asian, Latino, non-Hispanic Black, or non-Hispanic White householders. Of these households, 40.0% (n = 537,763) were designated as low-income and therefore included in the simulation. Across 1,000 simulations, on average, 39.6% of low-income households stayed in their town of residence, while 60.4% (approximately 324,823 households) moved to a different town. Total low-income households at baseline and post-simulation by township are shown in Figure 2, while low-income households by ethnoracial group at baseline and post-simulation are shown in Figures A.3–A.6.

Figure 2.

Map of low-income households by township A) at baseline in 2019 and B) post-move in 2019, averaged over 1,000 simulations.

Sizing: 1.5-column.

In print: Color requested.

3.2. Index of dissimilarity

The simulation dispersed an average of 324,823 low-income households, most notably from towns with the highest concentrations of low-income households at baseline (e.g., Hartford, New Haven, Bridgeport) (Figure 2). The percent of total low-income households simulated to move in or out, summarized by county, is shown in Table A.1, and was larger than the corresponding percent of total Connecticut residents who moved into or out of each county according to 2019 ACS migration flows (U.S. Census Bureau, 2019i; U.S. Census Bureau, 2023). Multi-group index of dissimilarity declined in all eight Connecticut counties after simulating desegregation, with the largest reduction in Middlesex County (Table A.2, Figure A.7).

3.3. Environmental exposures

Average town-level annual exposure to PM_2.5 remained approximately constant, while exposure to O3 increased slightly post-simulation (Figures A.8.a–b, Table A.3). We observed reductions in annual average NO₂ exposure, summertime heat index, and road noise as well as an increase in annual average NDVI exposure (Figures A.8.c–f, Table A.3).

3.4. Averted all-cause mortality

Deaths averted within each ethnoracial group were primarily driven by changes in exposure to greenness, followed by NO₂ and noise (Table 1). The highest numbers of deaths averted from NDVI, NO₂, and noise were among non-Hispanic White low-income residents. Asian low-income residents experienced marginal reductions in deaths from NDVI and NO₂ and no change in deaths from noise, heat, O₃, or PM_2.5.

Table 1.

Average deaths averted (95% confidence intervals) in Connecticut in 2019 across 1,000 simulated moves and corresponding exposure contrasts.

Ethnoracial group	Heat	NDVI	NO2	Noise	O3	PM2.5
Asian	0 (0, 0)	1 (1, 2)	1 (0, 1)	0 (0, 0)	0 (0, 0)	0 (0, 0)
Hispanic/Latino	0 (0, 0)	25 (12, 38)	10 (7, 14)	1 (0, 2)	0 (0, 0)	0 (0, 1)
Non-Hispanic Black	0 (0, 1)	40 (20, 60)	17 (12, 23)	1 (0, 1)	0 (0, 0)	1 (1, 1)
Non-Hispanic White	1 (0, 1)	103 (51, 155)	43 (30, 57)	7 (3, 11)	−1 (−1, 0)	−3 (−4, −2)
Total	1 (0, 2)	169 (84, 255)	71 (49, 94)	9 (4, 14)	−1 (−2, −1)	−2 (−2, −1)

NDVI, normalized difference vegetation index.

Changes in mortality from PM_2.5, O₃, and heat were marginal; we observed on average two excess deaths from PM_2.5, one excess death from O₃, and one averted death from heat in total, but no change among Asian or Latino low-income households from these exposures and no change among non-Hispanic Black low-income households from O₃ or heat (Table 1).

Average rates of deaths averted from NDVI and NO₂ per 100,000 low-income residents were highest among non-Hispanic Black residents, while the rate of deaths averted from noise was highest among non-Hispanic White residents (Table 2). Rates of averted or excess deaths were lowest in magnitude among Asian residents across all exposures (Table 2).

Table 2.

Rates of average deaths averted (95% confidence intervals) per 100,000 low-income residents in Connecticut in 2019 across 1,000 simulated moves and corresponding exposure contrasts.

Ethnoracial group	Heat	NDVI	NO2	Noise	O3	PM2.5
Asian	0.03 (0.01, 0.06)	3.54 (1.76, 5.32)	1.55 (1.06, 2.05)	0.21 (0.09, 0.33)	−0.02 (−0.03, −0.01)	0.03 (0.02, 0.05)
Hispanic/Latino	0.07 (0.02, 0.13)	9.22 (4.59, 13.85)	3.81 (2.60, 5.02)	0.35 (0.15, 0.56)	−0.03 (−0.05, −0.02)	0.15 (0.10, 0.20)
Non-Hispanic Black	0.19 (0.06, 0.33)	21.26 (10.59, 31.93)	9.13 (6.23, 12.02)	0.47 (0.20, 0.74)	−0.07 (−0.11, −0.04)	0.40 (0.27, 0.53)
Non-Hispanic White	0.09 (0.03, 0.16)	12.12 (6.02, 18.22)	5.10 (3.48, 6.71)	0.86 (0.38, 1.35)	−0.10 (−0.15, −0.05)	−0.35 (−0.46, −0.23)
Total	0.10 (0.03, 0.17)	12.56 (6.24, 18.87)	5.29 (3.62, 6.97)	0.69 (0.30, 1.07)	−0.08 (−0.12, −0.04)	−0.13 (−0.18, −0.08)

NDVI, normalized difference vegetation index.

3.5. Sensitivity analyses

Our first sensitivity analysis removed the inverse distance weighting penalty (Table A.4) and found that total deaths averted from NDVI, NO₂, noise, and heat were slightly larger compared to deaths averted when movement was restricted according to distance (Table 1). Marginal averted deaths rather than excess deaths were associated with PM_2.5, while excess deaths from O₃ increased slightly (Table A.4).

Deaths averted using drive time weighting were again slightly larger for NDVI, NO₂, noise, and heat (Table A.5), compared to deaths averted using inverse distance weighting (Table 1). Excess deaths from O3 remained, and excess deaths from PM_2.5 were slightly attenuated (Table A.5).

Meanwhile, average deaths averted from NDVI in 2018 (Table A.6) were higher than 2019 estimates (Table 1). Deaths averted from noise were slightly higher and excess deaths from PM_2.5 were slightly lower in 2018 simulations than in 2019, while impacts from other exposures remained constant (Table A.6).

4. Discussion

Our results suggest that full implementation of Section 8-30g’s 10% affordable housing target would reduce segregation in Connecticut and reduce mortality among low-income Connecticut residents, primarily from increased exposure to greenness and reduced exposure to ambient NO₂ and road noise. Benefits were most pronounced among non-Hispanic Black residents for NDVI and NO₂, in line with our hypothesis that ethnoracial groups disproportionately impacted by environmental racism would experience the highest rates of averted deaths. Contrary to our hypothesis, however, the highest rate of averted deaths from noise, as well as the second highest rates of averted deaths from NDVI and NO₂, were observed among non-Hispanic White residents, rather than non-Hispanic Black or Latino residents. We observed marginal changes in mortality from PM_2.5, O₃, and heat, suggesting that these exposure contrasts were not large enough to impact all-cause mortality. Our sensitivity analyses meanwhile suggest that mortality impacts remained consistent regardless of the year and thus may persist.

Strengthening Section 8-30g to more heavily incentivize or require the 10% benchmark presents an opportunity to advance environmental health justice by redistributing access to health-promoting environments while addressing the urgent need for housing. There are only an estimated 37 affordable and available housing units for every 100 extremely low-income renter households in Connecticut (National Low Income Housing Coalition, 2023). Meanwhile, recent polling reveals that 70 percent or more of Americans support policies to allow increased development of apartments (Horowitz & Kansal, 2023). Stronger state laws like Massachusetts’s Chapter 40B and California’s Senate Bill 9 may serve as models for strong zoning reform in Connecticut, which to date has been stymied in the legislature (California Senate Democratic Caucus, 2021; Flint, 2022; Harvard Law Review, 2022; Reid etal., 2017). For example, Connecticut’s recent Public Act 21-29 allows accessory dwelling units as-of-right so that households are automatically permitted to convert property space into an extra housing unit in single-family residential zones, reduces parking space minimums, and mandates municipalities to submit plans to further affordable housing development, but was arguably the weakest and the only to pass among three Connecticut zoning reform bills proposed in 2021 (Harvard Law Review, 2022; Public Act 21-29, 2022). While this legislation does reduce some barriers to densification, stronger reforms remain necessary to expand housing development. Meanwhile, our finding of significant mortality impacts associated with exposure to NO₂, a major traffic-related pollutant, and road noise highlights the ongoing racialized harms of highway construction in the U.S. (Archer, 2020; Collins et al., 2020; Willis et al., 2023). Using the Federal Aid Highway Act of 1956, the U.S. government systematically demolished community institutions and displaced residents in predominantly Black neighborhoods, reinforcing residential segregation and exposure inequities (Archer, 2020; Collins et al., 2020; Willis et al., 2023). These policies appear to have reverberating impacts today.

By examining the implicit 10% affordable housing goal established by Section 8-30g, this analysis presents a case study of the potential health impacts of one place-based desegregation policy as well as a template for further state, regional, and national analyses. Strengthening desegregation policies is not zero-sum and has the potential to provide universal benefit. Though some households in our simulation experienced adverse exposure contrasts, the modeled intervention would only increase opportunities for integrative movement; it would not in itself incentivize relocations with an adverse health impact or impose competition for existing residences. In fact, increasing housing supply statewide would likely reduce rent burden and improve housing security overall, while individual preferences for health-promoting amenities, which were not modeled here, may largely inhibit relocations with adverse consequences (Been et al., 2019; Gillespie, 2017, pp. 109-111). It is also important to note that the housing allocation and population movement modeled in this analysis would, in reality, occur over several years, delaying the mortality benefits from this instantaneous simulation by potentially 4-7 years for affordable housing development (Alameda County Housing Development Capacity Building Program, 2019). Our analysis therefore sought to compare mortality in a counterfactual 2019 scenario to mortality in the observed 2019 scenario, rather than mortality pre-intervention versus post-intervention, to estimate impacts that would be observed after long-term exposure redistribution.

To our knowledge, this is the first environmental health analysis to simulate the role of zoning and housing policies toward mitigating segregation and exposure disparities. By using both availability of and proximity to new affordable housing to estimate movement probabilities, we simulated two major drivers of relocation decisions (Moeckel, 2016). Our sensitivity analyses meanwhile suggest that our results are robust to the method of movement weighting and year of analysis and may be slightly conservative. Additional key strengths of this study are its consideration of six environmental exposures, each independently associated with mortality, and its analysis of mortality impacts stratified by ethnoracial group (Hao et al., 2022; Huang et al., 2021; Pope et al., 2019; Rojas-Rueda et al., 2019; Turner et al., 2016; Wellenius et al., 2017). This analysis shows, in line with study hypotheses, that mortality benefits from desegregation may be most prominently felt among those most impacted by residential segregation and environmental racism, i.e., non-Hispanic Black and Latino residents, therefore highlighting the potential for this and similar policy interventions to advance environmental health equity.

However, there are several limitations to this study. Our analysis reflected township-level zoning decisions, but this may have masked within-township exposure disparities. Analyses using tract-level estimates have been shown to reveal exposure disparities of a greater magnitude than those estimated using county- or state-level data (Clark et al., 2022). Likewise, our aggregated town-level spatial resolution may have led to underestimations of true exposure contrasts and deaths averted that would have resulted from equitably dispersed development, including increased development in neighborhoods with concentrated advantage. These underestimations were likely greatest among non-Hispanic Black and Latino residents, as neighborhoods with predominantly Black and Latino residents experience a disproportionate burden of adverse exposures (Clark et al., 2022). Additionally, our ethnoracial categorizations did not include Indigenous households due to low sample sizes and our inadequate power to detect an effect in this group. This remains pervasive and problematic among analyses of health effects across ethnoracial groups (Lett, Asabor, et al., 2022). We also assessed impacts using only one year and thus did not account for longer-term effects and annual variation in between-township exposure contrasts, although our results were robust to the year of choice in sensitivity analyses. Moreover, while we focused on environmental exposures and mortality, these policies would likely impart benefits through changes in other health-relevant social exposures and non-mortality endpoints as well (Anderson et al., 2021; Appel & Nickerson, 2016; Eaton, 2020; Havewala, 2021; Krieger et al., 2020; Menendian et al., 2021; Nardone et al., 2020; Steil et al., 2021).

Relocation from high-density to low-density areas, however, may adversely influence air quality through increased reliance on car travel over public transportation and active transport modes (Scheiner & Holz-Rau, 2013). These transport mode transitions and their downstream effects on pollution, physical activity, and health were not considered in our projections, although median town-level prevalence of car usage for workplace commuting increased only marginally from 89.0% at baseline to 89.1% post-simulation (Le & Poom, 2023; U.S. Census Bureau, 2019g). Similarly, we did not consider potential impacts of densification and relocation on greenspace, which may be reduced from densification processes (Haaland & van den Bosch, 2015). There are also multiple limitations associated with the exposure-response functions used in this analysis. Namely, these effect estimates were not derived from studies specific to low-income Connecticut residents, potentially limiting their generalizability to our study population (Hao et al., 2022; Huang et al., 2021; Pope et al., 2019; Rojas-Rueda et al., 2019; Turner et al., 2016; Wellenius et al., 2017). Furthermore, we used one effect estimate to project mortality impacts from greenspace across towns of varying levels of urbanicity, but some studies suggest the protective effect of greenspace on mortality is stronger in urban versus non-urban areas (Browning et al., 2022; Vienneau et al., 2017). Our analysis therefore may have overestimated deaths averted associated with relocation to greener, less urbanized areas.

Our simulation also relied on several strong assumptions about population mobility. Namely, we maintained state population constant and did not model other important movement drivers, such as relationships, life course stage, length of residence, housing tenure, and household composition (Gillespie, 2017, pp. 89-125). We therefore did not examine effects specific to age groups, but instead used exposure-response functions derived across all ages, potentially flattening variation in impacts between age groups (Hao et al., 2022; Huang et al., 2021; Pope et al., 2019; Rojas-Rueda et al., 2019; Turner et al., 2016; Wellenius et al., 2017). Moreover, our simulation was not weighted by “pull” factors known to influence location choice, including air quality and other amenities, and thus probabilities of movement toward areas with health-promoting amenities may be greater than those modeled (Gillespie, 2017, pp. 109-111). Additionally, by restricting households from relocating to towns in which their income may exceed the HMFA low-income limit, households with incomes at the higher end of the low-income range were made to stay within the same or a wealthier HMFA. This may have led to some underestimation of movement toward towns with lower wealth and subsequently adverse exposure contrasts. However, this underestimation was likely minimal, since this subset, comprising only 1.6% of low-income households, were permitted to move to lower-wealth towns within HMFAs for which they remained eligible, and the majority of Connecticut towns belonged to HMFAs with the same 2019 income limit (Figure A.9.) (HUD, 2023). Finally, these results were premised on a large scale of population movement, with simulated movement exceeding 2019 migration estimates (Table A.1). While this may limit the feasibility of our projections, our goal was to examine the potential impacts of leveraging this policy intervention to its fullest extent and over the long term, as one important piece of evidence for policymakers and as a blueprint for additional analyses. Our results are fully reproducible with available code to facilitate these extensions.

4.1. Conclusions

Our novel simulation-based health impact assessment demonstrates the importance of considering place-based policy approaches to desegregation as one mechanism to advance environmental health equity (Steil & Lens, 2023). Desegregation through increased, equitably dispersed affordable housing development represents a structural remedy to structural injustice. This must be implemented in conjunction with people-based mechanisms and policies that improve environmental quality, particularly in communities impacted by environmental racism (Steil & Lens, 2023). The need for legislative action toward health equity and the popularity of policies to increase apartment production meanwhile underscore the importance of further research to evaluate actionable policy interventions. Persistent reproduction of racialized health inequities, as well as the harm inflicted to date by extractive research practices and misrepresentation of race as a biological rather than social construct, moreover compel us to adopt action-oriented frameworks, center community expertise, and allocate research efforts to drive and strengthen solutions (Carrión et al., 2022; Lett, Adekunle, et al., 2022; Lett, Asabor, et al., 2022). These interventions lie at the nexus of social and environmental equity - further studies assessing the impacts of desegregation policy have the potential to catalyze lasting change.

Highlights.

• Simulating movement into new affordable housing reduced racialized segregation

• Desegregation may reduce deaths from environmental exposures

• Deaths averted were driven by changes in greenness, NO₂, and noise exposure

• Strengthening desegregation policy can advance environmental health equity