Temporal trends in sociodemographic composition and land development within U.S. fenceline communities surrounding hazardous industrial facilities: 2001–2019

We define industrial facilities as those using hazardous substances in 10 high-hazard North American Industry Classification System (NAICS) codes, including pulp mills, petroleum refineries, and chemical manufacturing, reporting to the Toxics Release Inventory (TRI) maintained by the US Environmental Protection Agency (EPA) (figure S1) (www.epa.gov/toxics-release-inventory-tri-program/tri-basic-data-files-calendar-years-1987-present, accessed 24 June 2023). We select this subset of hazardous facilities as facilities in these NAICS with processes using threshold quantities of hazardous substances and are subject to the EPA's risk management plan rule (RMP). Between 2001 and 2019, 2457 unique facilities in our included industries reported to the TRI. The majority of these facilities in 2019 were chemical manufacturing facilities (88%, n = 2167), followed by petroleum refineries (9%, n = 213) and pulp mills (3%, n = 77). The facilities are located in 49 states and two US territories, with 20% of facilities located in Texas (n = 235) and 7% in Louisiana (n = 86) in 2019 (n = 1080) (table S1). We present results at the national level as the EPA's RMP program requirements apply nationally.

We first compare trends in fenceline and neighboring community sociodemographic characteristics to test the hypothesis that the percentage of historically marginalized fenceline populations has increased between 2001 and 2019 compared to neighboring communities with no industrial facilities. A time period of 2001–2019 and the census tract geographic unit was selected as these years are available across all input datasets. We define a fenceline community as within a 5 km circular buffer of included industrial facilities, as buffers greater than 3 km have been demonstrated as necessary in preserving characteristics of the fenceline community beyond the immediate facility fenceline [9]. We select a local comparison group (an additional 5 km buffer beyond the facility (5–10 km)), rather than state or national comparisons, to assess differences between fenceline communities and neighboring communities [21]. We employ a series of aggregation methods to avoid the unit-hazard coincidence and reduce the modifiable areal unit problem (figure S2) [9, 22, 23]. We calculate change as the difference in census tract sociodemographic percentage for 2019 less 2001 (2019–2001), rather than percent change ([(2019–2001)/2001]), as a small increase in population among less-populated subgroups can result in a large percentage change, challenging interpretability. Similarly, we present results as the change in the median census tract percentage rather than the means to avoid the influence of outliers.

We extract census tract level data on race, ethnicity, educational attainment, poverty status, linguistic isolation, and female headed households using both decennial Census data and the American Community Survey (ACS) estimates. We apply 2000 U.S. Census data to the years 2001 and 2004, the 2010–2014 ACS to the years 2006, 2008 and 2013, and the 2015–2019 ACS to the years 2016 and 2019 [24] (table S2). Examination of longitudinal trends between spatially varying census tracts is enabled by the Longitudinal Tract Database cross walk tables [25] which employ population and areal weighting to apportion 2000 census-tract boundaries to the 2010 census boundaries.

For our indicators of privilege, we include White race and educational attainment greater than a high school education. We develop a working definition of 'historically marginalized populations' to include ethnicity of Hispanic or Latino origin (Latinx) within all racial groups, racial groups of Black (Black and African American) and other non-white (American Indian, or Alaska Native, Asian, Native Hawaiian or Other Pacific Islander, two or more races, and some other race), as well as persons with less than or equal to one ratio of income to poverty (poverty threshold) in the last 12 months. We also include in our definition the proportion of female headed households, which was found to be significant in a prior study of air toxics [26].

We assess sociodemographic trends by varying sources of potential exposure. First, we assess the difference in sociodemographic trends between facilities reporting greater than the 75th percentile of annual total onsite releases of hazardous substances (>2300 pound) and less than or equal to the 75th percentile. We next calculate the number of facilities within a 5 km radius of census tract centroids and compare sociodemographic changes within communities with more than one facility as compared to those with one facility. Finally, as a surrogate for 'new' facilities, we assess sociodemographic changes in communities surrounding facilities reporting for the first time after 2001 as compared to facilities reporting in 2001 or earlier. We hypothesize that increases in the percentage of historically marginalized fenceline populations is greater among communities with a greater volume of annual releases, more than one industrial facility, and near facilities reporting for the first time after 2001.

We next examine the association between land development and fenceline community sociodemographic characteristics, assuming a concurrent and bi-directional relationship. We hypothesize that more privileged populations have decreased within fenceline communities and that these decreases are associated with increases in the rate of land development.

We quantify change in land development surrounding included industrial facilities by quantifying the census tract level change (2019–2001) in total square kilometers (km²) of developed land within fenceline communities compared to neighboring communities. We employ the National Land Cover Database (NLCD), a nationwide land cover classification database with a high resolution for the years 2001 and 2019 (figure S3). Developed by the US Geological Survey, the NLCD is available at 30 m × 30 m gridded resolution and applies a decision-tree raster classification of six Landsat bands to classify images into 20 land use types [27]. We create a new binary raster dataset by categorizing all developed types together as well as all undeveloped land types together. We lose some spatial heterogeneity by reprojecting the 30 m² gridded estimates to 1 km² to maximize computational efficiency.

Similar to Clement and Alvarez (2020), we conduct a spatial autoregressive model with autoregressive disturbances (SARAR), modeling the association between the change between 2001 and 2019 in developed land (km²) per census tract (dependent variable) and the change in census tract percentage of included sociodemographic characteristics (independent variables). The SARAR model accounts for spatial autocorrelation in the dependent variable (lag) and error term, where ${y_{it}}$ represents the change in the dependent variable (developed land (km²)) for census tract $i$ between 2001 and 2019 $\left( t \right)$; $\beta {X_{it}}$ represents the change, $t$, in the percentage of the independent variables, $X$, for census tract $i$ and $\beta$ represents the effect of each independent variable $X\,$ (equation (1)). We log-transform and difference our dependent and independent variables to create a change score for each variable accounting for the change between 2001 and 2019, adding a constant of '1' to correct for zero-values when log-transforming and interpreting the slope estimates as elasticities. When using two periods of data, first differencing facilitates longitudinal analysis because it controls for time-invariant characteristics of the unit of analysis, i.e. census tracts.

Owing to multicollinearity among included sociodemographic variables (table S3), we include as our independent variables ($X$) the log-transformed, differenced (i.e. percentage in 2019 less the percentage in 2001) change score in the percentage White, greater than high school education, Latinx, and below the poverty threshold. We also include several control variables including the count of facilities within 5 km ($Facilities$), the log-transformed, change in total census tract population ($Total$), total census tract area (km², $Area$) and total population and total subgroup population ($XTotal\_2001)$ in 2001. Representing the spatial lag, $\rho$ denotes the lag term of the weighted average of the dependent variable for each census tract based on the spatial weight $W$ for that census tract, established using first-order queen contiguity. The error term $\upsilon$ is comprised of spatial error term $\lambda$, also calculated using the weights matrix $W$, as well as the terms $\varepsilon$.

Equation 1. SARAR Model,

$$\begin{align} {y_{it}} & = { }\alpha + { }\rho W{y_{it}} + { }\beta {X_{it}} + \beta {\text{Facilitie}}{{\text{s}}_i} + { }\beta {\text{Tota}}{{\text{l}}_i} \nonumber\\ & \quad + \beta X{\text{Total}}\_{{2001}_i} + \beta {\text{Are}}{{\text{a}}_i} + { }{\upsilon _{it}} \nonumber\\ {\upsilon _i} & = \lambda W{\upsilon _{it}} + { }{\varepsilon _{it}} .\end{align} \tag{ 1 }$$

In addition to our main model described above, we assess the relationship between the change in community sociodemographic characteristics and land development for census tracts with available spatial information on historical segregation, described in the supplemental material.

We find the total population living within fenceline (<5 km) communities decreased 48% between 2001 and 2019 (85 million surrounding 1841 facilities in 2001 versus 44 million surrounding 1080 facilities in 2019); a greater degree than observed in neighboring communities (5–10 km) (33%: 117–78 million) (table S4). Among historically marginalized populations, we observe the largest increase among Latinx fenceline populations (22% in 2001–28% in 2019) (table S4 and figure 2). Developed land increased modestly within industrial facility fenceline communities (59 600 per 1841 facilities to 47 300 km² per 1080 facilities), less than the change in developed land in neighboring communities (69 600–72 900 km²).

Zoom In Zoom Out Reset image size

Comparing median census tract level change (2019–2001) in sociodemographic percentages (table 1 and figure 3), we find that Latinx, below the poverty threshold and female headed household population categories grew more within fenceline communities compared to neighboring communities (Latinx = +6% vs. +4%, respectively; female headed households = +2% vs. 1%; below the poverty threshold = +2% vs. +1%). White populations increased within fenceline compared to a decrease in neighboring communities (+12% and −1%, respectively). The change in median census tract percentages was the same within both fenceline and neighboring communities for Black, other non-white, greater than high school education, below the poverty threshold and linguistically isolated populations. Though the percentage of the Black populations did not change within fenceline communities over this time period, we find a minor sustained higher percentage of Black residents within fenceline communities relative to neighboring communities (20% versus 19%, respectively) (table S4). We note that our results intend only to compare fenceline communities to their immediate neighboring communities and are not necessarily independent of national trends.

Zoom In Zoom Out Reset image size

We observe fluctuation in the total annual releases reported by facilities. In 2001, a total of 543.7 million pounds of hazardous substances were released from 1841 facilities (table S5), while in 2019, 298.8 million pounds were released from 1080 facilities. We next compare fenceline community sociodemographic composition between 2001 and 2019 surrounding facilities reporting greater than the 75th percentile of total annual onsite releases (>2300 pounds), to facilities with less than or equal to the 75th percentile total annual onsite releases (≤2300 pounds). We observe minimal difference (±2%) for greater than high school education, below the poverty threshold, linguistically isolated, female headed households, and other non-white populations (table S6 and figure S4). White populations decreased (−4%) and Black populations increased (+4%) surrounding facilities with greater than the 75th percentile of total onsite releases. Latinx populations increased surrounding facilities with less than the 75th percentile releases as compared to those with greater than the 75th percentile (+9% vs. +3%, respectively), while Black populations decreased surrounding facilities with less than the 75th percentile total onsite releases (−3%). We observe regional differences across US EPA federal regions, including an increase in the percentage of historically marginalized fenceline communities within Regions 3 and 5, and a decrease in Region 7 (figure S5, table S7).

We observe a positive association between increasing facility count and the median census tract change (2019–2001) in percentage of Latinx, female headed households, White, and greater than high school education population subgroups, as well as an inverse association for other non-white populations (figure 4 and table S8). Black, below the poverty threshold, and linguistically isolated population categories had limited or no difference. Increases in Latinx populations from 2001 to 2019 were amplified for communities in the top quartile of the number of nearby facilities (5–32 facilities; +6 percentage points) compared to communities with only one to two facilities nearby (+4 percentage points). Within communities with the third highest quartile (3–4 facilities), the change in Latinx populations was four times higher than White populations (+4% and no change, respectively) (table S8). We observe several non-linear trends between increasing facility count and percentage sociodemographic composition (e.g. the decrease and then subsequent increase for Black populations).

Zoom In Zoom Out Reset image size

Total count of facilities varied over the study period, with 431 (18%) facilities reporting for the first time after 2001 and 1010 (41%) facilities that reported in 2001 discontinuing reported before 2019. We observe minimal change (less then ±2 percentage points) in sociodemographic composition between facilities that first reported after 2001 as compared to those that reported in 2001 or earlier for Black, White, below the poverty threshold, linguistically isolated and female headed households (figure S6 and table S9). The percentage of other non-white populations increased modestly surrounding new facilities as compared to existing ones (+4% and +1% respectively), while the percentage of Latinx populations increased at a rate 1.5 times higher surrounding new facilities as compared to existing facilities (+12% and +7% respectively). We also find that the percentage of populations with greater than a high school education increased at a slower pace surrounding new facilities as compared to existing facilities (+4% as compared to +9% respectively).

We next examine the relationship between the change in developed land (km²) per census tract within fenceline communities (n = 15 991) and the change in percentage of included sociodemographic variables (White, greater than high school education, Latinx ethnicity and below the poverty threshold). We find a negative, insignificant relationship between change in percentage White and below the poverty threshold, a negative significant relationship between change in percentage Latinx populations, as well as a positive, insignificant relationship between change in percentage greater than high school education and change in land development (table 2). In other words, census tracts that experienced an increase in the percentage Latinx populations experienced a lower rate of land development. Results for the analysis by HOLC neighborhoods is presented in table S10.

Table 2. Results for the spatial autoregressive with autoregressive disturbances (SARAR) models of the association between the change between 2001 and 2019 in developed land (km²) per census tract and the change in census tract percentage of included sociodemographic characteristics.

Fenceline census tract sociodemographic variables	Change in developed land per census tract area (km2) for fenceline (<5 km) communities
	(n census tracts = 15 991)
	β	SE
Independent variables (Δ between 2001 and 2019)
Percentage White	−0.063	0.050
Percentage greater than high school education	0.042	0.071
Percentage Latinx	−0.279*	0.063
Percentage below the poverty threshold	−0.109	0.064
Controls (Δ between 2001 and 2019)
Change in total population	0.296	0.011
Controls (time invariant)
Count of facilities within 5 km of census tract centroid in 2001	−0.013*	0.003
Total population in 2001	0.000*	0.000
Total White population in 2001	−0.000*	0.000
Total population with greater than high school education in 2001	0.000*	0.000
Total Latinx population in 2001	−0.000*	0.000
Total below the poverty threshold	−0.000	0.000
Total census tract area (km2)	0.137*	0.003
$\rho$	0.004	0.021
$\lambda$	0.010	0.003

Independent variables have been log-transformed and first-differenced, representing the change score between 2001 and 2019. All control variables aside from number of facilities and total census tract area have been log-transformed. β = beta coefficient; SE = standard error; $\rho$ = spatial lag parameter; $\lambda$ = spatial error parameter; * = p < 0.05.