Is Residential Exposure to Oil Refineries a Novel Contextual Risk Factor for Coronary Heart Disease?

Abstract

Despite a multi-decade decrease in cardiovascular disease, geographic disparities have widened, with excess mortality concentrated within the United States (U.S.) South. Petroleum production and refining, a major contributor to climate change, is concentrated within the U.S. South and emits multiple classes of atherogenic pollutants. We investigated whether residential exposure to oil refineries could explain variation in self-reported coronary heart disease (CHD) prevalence among adults in southern states for the year 2018, where the majority of oil refinery activity occurs (Alabama, Mississippi, Louisiana, Arkansas, Texas, New Mexico, and Oklahoma). We examined census tract-level association between oil refineries and CHD prevalence. We used a double matching method to adjust for measured and unmeasured spatial confounders: one-to-n distance matching and one-to-one generalized propensity score matching. Exposure metrics were constructed based on proximity to refineries, activities of refineries, and wind speed/direction. For all census tracts within 10km of refineries, self-reported CHD prevalence ranged from 1.2% to 17.6%. Compared to census tracts located at ≥5km and <10km, one standard deviation increase in the exposure within 5km of refineries was associated with a 0.33 (95% confidence interval: 0.04, 0.63) percentage point increase in the prevalence. A total of 1119.0 (123.5, 2114.2) prevalent cases or 1.6% (0.2, 3.1) of CHD prevalence in areas within 5km from refineries were potentially explained by exposure to oil refineries. At the census tract-level, the prevalence of CHD explained by exposure to oil refineries ranged from 0.02% (0.00, 0.05) to 47.4% (5.2, 89.5). Thus, although we cannot rule out potential confounding by other personal risk factors, CHD prevalence was found to be higher in populations living nearer to oil refineries, which may suggest that exposure to oil refineries can increase CHD risk, warranting further investigation.

1. Introduction

Despite a sustained decline in cardiovascular disease over the past several decades, geographic disparities have widened as excess mortality has shifted from the United States Northeast and Mid-Atlantic to the South (Casper et al. 2016; Singh et al. 2015). As geographic disparities have increased, the rate of decline in cardiovascular mortality has slowed and may be reversing in some geographic regions. Indeed, the rates of decline in cardiovascular mortality (9–50%) within southern counties have lagged their geographic counterparts (64–83%) between the late-twentieth and early-twenty-first centuries (Casper et al. 2016).

The mechanisms underlying geographic disparities in cardiovascular outcomes are not fully understood. Although traditional vascular risk factors and socioeconomic conditions explain much of cardiovascular disease prevalence, these characteristics do not fully explain geographic variability (Gebreab et al. 2015; Glynn et al. 2021). Within the United States (U.S.) South, Mississippi has experienced a disproportionate share of age-adjusted mortality due to heart failure and among the slowest rates of decline in all-cause cardiovascular mortality (Casper et al. 2016; Glynn et al. 2021). The spatial patterns of cardiovascular outcomes in the U.S. South resemble those observed for stroke outcomes, often referred to as the Stroke Belt that are not fully elucidated (Cushman et al. 2008; Glymour et al. 2007; Howard 1999; Howard and Howard 2020; Howard et al. 2016; Karp et al. 2016; Liao et al. 2009). Within the South, sizable disparities in county-level cardiovascular mortality have prompted recommendations for small-area surveillance of cardiovascular disease to elucidate the potential explanatory mechanisms (Casper et al. 2016).

A body of evidence shows that environmental exposures such as particulate air pollution have strong sociodemographic and geographic patterning (Bell and Ebisu 2012; Bravo et al. 2016; Hajat et al. 2013; Jbaily et al. 2022; Jones et al. 2014; Liu et al. 2021) and that they are associated with increased cardiovascular risk (Brook et al. 2010; Rajagopalan et al. 2018; Rajagopalan and Landrigan 2021). Together this evidence suggests that environmental exposures may contribute to geographic disparities in cardiovascular outcomes.

To our knowledge, residential exposure to oil refineries has not previously been evaluated as a contextual risk factor for coronary heart disease (CHD). Roughly, two-thirds of petroleum production and refining in the U.S. occur within the third Petroleum Administration for Defense District (PADD-3). PADD-3 consists of six southern states (Alabama, Arkansas, Louisiana, Mississippi, New Mexico, and Texas) (US EIA, 2023). While the sector of oil refineries is the second highest ranked sector regarding greenhouse gas emissions per facility in the U.S. (1.15 million metric tons of carbon dioxide equivalent for the year 2020) (US EPA 2021), oil refineries also emit multiple airborne pollutants that have been implicated in cardiovascular pathogenesis, including sulfur dioxide (SO₂), particulate matter (PM), nitrogen oxides (NO_x), volatile organic compounds (VOCs), and polycyclic aromatic hydrocarbons (PAH)(Adebiyi 2022). Byproducts of petroleum production and refining may also impact adjacent soil, and potable water in residential areas (Damian 2013; Lynch et al. 2004), which may independently increase the risk of cardiovascular diseases. Thus, a mixture of environmental hazards from oil refineries may be related to CHD pathogenesis.

Therefore, we evaluated whether residential exposure to oil refineries could explain geographic variation in CHD within the U.S. South.

2. Materials and Methods

Health outcome

Our study population includes adults (≥18 years) living in seven states (Alabama, Mississippi, Louisiana, Arkansas, Texas, New Mexico, Oklahoma), where CHD is disproportionately concentrated (Casper et al. 2016; Singh et al. 2015), residing within approximately 10km from refineries. We restricted our study population as residents of the census tracts within approximately 10km from refineries because residents living in close proximity to refineries may be exposed to refinery-related pollution. Additionally, we needed to account for unmeasured spatial confounders, as outlined in Statistical Analysis subsection in this section. The health outcome of interest in this study is census tract-level self-reported CHD prevalence for the year 2018. We obtained this from the Centers for Disease Control and Prevention (CDC)’s Population Level Analysis and Community Estimates (PLACES) (Greenlund et al. 2022). Self-reported CHD prevalence was defined as the percentage of respondents aged ≥18 years who reported ever having been told by a doctor, nurse, or other health professional that they have had angina or CHD. The CDC PLACES estimates are small area estimates constructed by using statistical approaches and the Behavioral Risk Factor Surveillance System (BRFSS) that samples from a noninstitutionalized population. BRFSS is a telephone-based survey and samples were collected separately by each state. Except for Guam, Puerto Rico, and the U.S. Virgin Islands, the sampling design was a disproportionate stratified sample design, which used random digit dialed probability samples for those aged 18 years or older. The population weighted average of the prevalence in the seven states was 7.0% while that of the prevalence in the U.S. was 6.4%.

Exposure classification

Exposure of interest 1: the proximity to refineries

The geocoded locations for 59 oil refineries for the period from 2015–2017 were obtained from the U.S. Energy Information Administration (EIA). Because residents residing nearer to oil refineries may be exposed to higher levels of environmental pollution, we considered distance from refineries to be an important determinant of exposure. We assumed that proximity to refineries might be related to many plausible exposure pathways (Figure 1) (Adebiyi 2022; Damian 2013; Lynch et al. 2004), including airborne pollutant mixtures majorly.

Figure 1.

A conceptual framework for potential relationship between oil refineries and cardiovascular risk (US EIA 2023 ; US EPA 2015; Garcia-Gonzales et al. 2019; CEIP 2023). The hypotheses of this study are presented as dashed lines. Elements considered in this study are presented as red bolded texts in the boxes.

We calculated the Euclidian distance to the closest oil refinery from the centroid of each census tract. We dichotomized the proximity of refineries using three a priori selected distance thresholds (<2.5km, 2.5–5km, or <5km) from the centroid of a census tract to an oil refinery. We defined oil refineries-exposed census tracts as those with at least one refinery within a radius of <2.5km or <5km from their centroid. Controls were defined as census tracts located at ≥5km and <10km (approximately one degree) from a refinery; tracts ≥10km from a refinery were not included in analysis. The three categorizations (<2.5km, 2.5–5km, or <5km) of exposed census tracts were selected a priori, considering that air pollutants decay with distance from smokestacks of oil refineries (Chen et al. 2012) and the conventions applied in our previous study (Kim et al. 2022). Although some air pollutants may travel farther than 5km, we did not consider census tracts located ≥5km from refineries as exposed, due to the need to adjust for unmeasured spatial confounding. The use of census tracts located at both ≥5km and <10km from a refinery as the control group and the exclusion of tracts 10km or further from a refinery enabled us to adjust for unmeasured spatial confounding (Kim and Bell 2022; Kim et al. 2022). Thus, there was a trade-off between increasing the distance thresholds used to define exposure and choosing appropriate control groups to adjust for unmeasured spatial confounding.

Exposure of interest 2: wind-integrated inverse squared distance-weighted sum of actual petroleum production

In addition to proximity, we considered the productivity of a refinery as a second exposure domain because higher levels of petroleum production may correspond to a higher burden of polluting byproducts. Also, if census tracts are situated in close proximity to refineries in the downwind position, residents in those census tracts may have a higher likelihood of exposure to air pollutants emitted from those refineries. Additionally, even in the downwind position, higher wind speeds may reduce the likelihood of people being exposed to high levels of air pollutants compared to lower wind speeds, as pollutants can more easily disperse with stronger winds.

We obtained the average petroleum production capacities of refineries for the years 2015–2017. We defined production capacity using the reported atmospheric distillation capacity of each facility. We obtained average annual petroleum production for each PADD corresponding to states within our study area (PADDs 2 and 3). We estimated actual petroleum production (APP) for each refinery by approximating its fraction of annual oil production as a share of the total production capacity within each PADD. We utilized facility-reported atmospheric distillation capacity to estimate their share of production. We obtained ERA5 monthly averaged data for wind information at a 0.25 degree grid for the years 2015–2017. We calculated the average of the 10m u and v components of the wind speed. The u component is parallel to longitude and the v component is parallel to latitude. We propose a novel continuous exposure metric, a wind-integrated inverse squared distance-weighted sum of APPs for each census tract as follows:

$$X_c=\sum _{r\in R_b}AP{P}_r\times \frac{1}{\delta \text{exp}(\sqrt{\left[{\left(W{S}_{u,c,r}W{D}_{u,c,r}\right)}^2+{\left(W{S}_{v,c,r}W{D}_{v,c,r}\right)}^2\right]})}\times \frac{1}{{\left(\frac{d_{u,c,r}}{\text{exp}\left(\gamma W{D}_{u,c,r}\right)}\right)}^2+{\left(\frac{d_{v,c,r}}{\text{exp}\left(\gamma W{D}_v,c,r\right)}\right)}^2}$$

X_c denotes a wind-integrated inverse squared distance-weighted sum APP at a census tract c. d_u,c,r and d_v,c,r denote u and v-components of the distance vector between the centroid of census tract c and a refinery r, respectively. R_b is a set of refineries within a pre-specified buffer distance, b, from the centroid of census tract c. b was 2.5km or 5km. WD_u,c,r and WD_v,c,r are binary indicators of the u and v-components of the wind direction between c and r. If the average of each component of the wind direction over the period of 2015–2017 is downwind, WD=1; otherwise WD=0. If both components are upwind, the term, $\frac{1}{{\left(\frac{d_{u,c,r}}{\text{exp}\left(\gamma W{D}_{u,c,r}\right)}\right)}^2+{\left(\frac{d_{v,c,r}}{\text{exp}\left(\gamma W{D}_{v,c,r}\right)}\right)}^2}$ becomes the inverse squared distance. If at least each component of the wind is downwind, the value of the term would increase, meaning that people may be more likely exposed to the pollution byproducts of the refinery r.

The term $\frac{1}{\delta \text{\hspace{0.17em}exp}\left(\sqrt{\left[{\left(W{S}_{u,c,r}W{D}_{u,c,r}\right)}^2+{\left(W{S}_{v,c,r}W{D}_{v,c,r}\right)}^2\right]}\right)}$, considers that air pollutants emitted from the refinery r may disperse less with lower wind speed, implying that people may be more likely exposed to air pollutants generated from that refinery. WS_u,c,r and WS_v,c,r are the u and v-components of the wind speed between c and r. This term would increase with the lower wind speed. Even in the downwind position, high wind speeds can easily disperse pollution, causing it to linger for only a short period of time. The term would decrease with the higher wind speeds.

The product of these two terms would equal the inverse squared distance factored by $\frac{1}{\delta }$ only if the average wind is upwind, which is $\frac{1}{\delta }\frac{1}{{d_{u,c,r}}^2+{d_{v,c,r}}^2}$. If the wind is downwind (at least one component) and wind speed is low, the product may be higher than what would be if it was $\frac{1}{\delta }\frac{1}{{d_{u,c,r}}^2+{d_{v,c,r}}^2}$. This implies that people may be more exposed to higher levels of air pollutants emitted from a refinery during downwind conditions with low wind speeds than they would be during upwind conditions. Since wind varies daily, and monthly, we assume that people may still be exposed to air pollutants even if the average wind over 2015–2017 is upwind, but their likelihood of exposure to high levels of air pollutants may be elevated only if the distance between c and r is short. This corresponds to a higher value of $\frac{1}{{d_{u,c,r}}^2+{d_{v,c,r}}^2}$.

We considered different values of the two scale factors, δ and γ: 0.5, 1, and 1.5. Among the combination of these values, we found that X_cs are highly correlated, approximately 0.99. Therefore, we used 1 of δ and γ.

Potential confounders

We obtained information on potentially confounding sociodemographic factors at the census tract level from the American Community Survey (ACS, 2014–2018), including age group, sex, race, and ethnicity, as well as indicators of household and community socioeconomic conditions (percentage of the population living below the federal poverty line, with less than high-school educational attainment, and median household income). We obtained health and behavioral vascular risk factors, including the prevalence of tobacco use, hypertension, hypercholesterolemia, diabetes mellitus, and obesity, from CDC PLACES (Greenlund et al. 2022).

Statistical analysis

We conducted census tract-level cross-sectional analyses.

Binary exposure variable analysis

For the proximity to refineries, analyses were separately performed for the three dichotomous, distance-based classifications of exposed census tracts (<2.5km, 2.5–5km, and <5km). Controls were defined as census tracts located at ≥5km and <10km from a refinery. The use of these controls was based on our previous studies to adjust for unmeasured spatial confounding (Kim and Bell 2022; Kim et al. 2022). The choice of 10km was informed by Figure S1. Figure S1 includes semi-variogram for spatial correlation of CHD prevalence before and after adjustment for several risk factors (i.e., sociodemographic variables, and smoking). As noted in the introduction section, the spatial patterns of cardiovascular outcomes in the U.S. South are not fully elucidated. There remain the spatial patterns of the CHD prevalence in our study population after the risk factors were controlled, implying potential unmeasured spatial confounding.

Specifically, we conducted one-to-n distance-matching with replacement. This technique matched one exposed census tract to one or more control census tract(s), depending on the Euclidean distance between the exposure and control census tract(s). Matching was performed only if census tracts were within 10km of each another. After this one-to-n distance-matching, we conducted one-to-one nearest neighbor matching with replacement by propensity score (PS). PS models were fit using generalized additive models (GAM). When fitting PS models, we included several covariates to adjust for measured confounding. Covariates were manually selected by checking standardized mean difference (SMD) ≥±0.25 (Stuart et al. 2013). Additionally, a spatial smoother (i.e., thin-plane spline) was added to augment adjustment for unmeasured spatial confounding in addition to one-to-n distance-matching. The two rounds of matching produced one-to-one matched pairs of exposed and control (unexposed) census tracts.

After matching was completed, we performed linear regression to evaluate the association between proximity to refineries as a binary exposure variable and CHD prevalence. Dummy variables that indicate each of the distance-matched strata were added for matched analysis. We conducted bootstrapping with a distance- and PS-matched dataset to estimate a standard error of the association estimate.

Continuous exposure variable analysis

In addition to a binary exposure variable, we conducted separate analyses with a continuous variable, X_c. We conducted separate analyses for each of X_c within 2.5km and within 5km. Similar to analyses with a binary exposure variable, we conducted one-to-n distance-matching with replacement. One exposed census tract was matched to one or more controls (i.e., census tracts located at both ≥5km and <10km from a refinery).

We then conducted one-to-one nearest neighbor matching with replacement (NNWR) by GPS. GPS was for X_c=w (w>0) as a continuous exposure variable. The GPS estimation factored in two dimensions of X_c with bimodal distributions: 1) binary dimension (i.e., whether there is a refinery within a prespecified buffer distance); and 2) continuous dimension (i.e., the quantity of petroleum production by refineries). We estimated GPS for X_c conditional on the proximity to refineries as the binary exposure variable, referred to as conditional GPS (CGPS) (Kim and Bell 2022), and then multiplied CGPS and the above-described PS (See Binary exposure variable analysis subsection). CGPS models were fit using generalized additive models (GAM). When fitting CGPS models, we included several covariates to adjust for measured confounding. Covariates were manually selected by checking SMD ≥±0.25 and their correlations with X_c (≥±0.25) in the exposed census tracts (Kim and Bell 2022). Additionally, a spatial smoother (i.e., thin-plane spline) was added to augment adjustment for unmeasured spatial confounding in addition to one-to-n distance-matching. We estimated GPS for X_c=w>0 in each distance-matched stratum.

After matching was completed, we performed linear regression to evaluate the association between X_c and CHD prevalence. Dummy variables that indicate each of the distance- and GPS-matched strata were added for this matched analysis. To estimate a standard error of the association estimate, we conducted bootstrapping with the distance- and GPS-matched dataset.

Others

We quantified the extent to which exposure to oil refineries might explain geographic variation in CHD prevalence. We used $\widehat{\beta }{X}_c\times Po{p}_c/100$ to estimate the number of cases that are potentially explained by exposure to oil refineries. $\widehat{\beta }$ is a coefficient estimate indicating a percentage point increase in self-reported CHD prevalence (%) per one unit increase in X_c, and Pop_c is the population age 18 years or older at census tract c. We used $\widehat{\beta }{X}_c/\mathit{\text{Pre}}{v}_c\times 100$ to estimate the percentage of cases that are potentially explained by exposure to oil refineries, where Prev_c is the CHD prevalence (%) at census tract c.

As sensitivity analyses, we used one-to-one nearest neighbor matching without replacement (NNWoR) and one-to-one nearest neighbor caliper matching with/without replacement (NNCW(o)R) instead of NNWR. We tested non-linear associations using a thin-plane spline in GAM.

Figures S2 and S3 show SMD for 18 potential confounders in our main analysis and sensitivity analyses. (Stuart et al. 2013).

We conducted all statistical analyses using R software 3.5.3 with mgcv, and CGPSspatialmatch packages (Kim and Bell 2022).

3. Results

Table 1 presents descriptive statistics of potential confounders and self-reported CHD prevalence by proximity to refineries (<2.5km, 2.5km–5km, <5km, as exposed groups and 5km–10km as the control group). The respective percentages of population with low socioeconomic status and Hispanic ethnicity were higher in census tracts nearer to refineries. The percentage of non-Hispanic Black and White populations were lower nearer to refineries. Census tracts nearer to refineries had higher prevalence of current smoking and obesity. The average of self-reported CHD prevalence (one standard deviation (SD)) was 8.0% (1.8%), 8.0% (1.9%), 8.0% (2.2%), and 7.8% (2.2%) for census tracts within 2.5km, between 2.5km and 5km, within 5km, and between 5km and 10km from refineries, respectively. For all census tracts within 10km from refineries, self-reported CHD prevalence ranged from 1.2% to 23.1% and its SD was 2.3%.

Table 1.

Distribution of census tract-level coronary heart disease prevalence and potential confounders by proximity to oil refineries.

	<2.5kma	P-valueb	2.5 – 5kma	P-valueb	<5km	P-valueb	5km – 10kma
Number of census tractsc	79		237		314		541
Sex, %
Males	48.4 (4.6)	0.344	48.1 (5.5)	0.470	48.2 (5.3)	0.310	47.8 (5.9)
Females	51.6 (4.6)	0.694	51.9 (5.5)	0.470	51.8 (5.3)	0.310	52.2 (5.9)
Age, %
18–19 years	4.0 (2.7)	0.694	3.6 (3.2)	0.582	3.7 (3.1)	0.776	3.8 (5.4)
20–24 years	10.1 (3.2)	0.772	10.1 (4.8)	0.611	10.1 (4.5)	0.572	10.3 (6.4)
25–44 years	37.4 (6.5)	0.122	35.8 (6.5)	0.826	36.2 (6.5)	0.613	36.0 (7.9)
45–64 years	32.1 (4.4)	0.610	32.8 (5.4)	0.023	32.6 (5.2)	0.032	31.7 (6.6)
65–84 years	14.4 (4.9)	0.028	15.6 (5.3)	0.513	15.3 (5.3)	0.134	15.9 (5.7)
85+ years	2.0 (1.5)	0.114	2.1 (1.9)	0.104	2.1 (1.8)	0.041	2.4 (2.0)
Race/ethnicity, %
Hispanic	39.2 (36.0)	0.124	35.7 (35.8)	0.273	36.6 (35.8)	0.120	32.7 (34.7)
Non-Hispanic White	28.6 (26.6)	0.260	31.2 (28.1)	0.573	30.6 (27.7)	0.342	32.5 (28.8)
Non-Hispanic Black	28.5 (32.6)	0.655	29.0 (33.0)	0.615	28.8 (32.9)	0.546	30.2 (32.3)
Socioeconomic conditions
Educational attainment of less than high school, %	25.5 (10.7)	<0.001	22.4 (12.8)	0.007	23.2 (12.4)	<0.001	19.8 (11.9)
Under the federal poverty line, %	22.1 (9.9)	0.224	22.4 (10.1)	0.886	23.2 (10.1)	0.676	19.8 (11.7)
Have health insurance, %	73.1 (11.9)	<0.001	76.7 (11.8)	0.020	75.8 (11.9)	<0.001	78.7 (10.9)
Median Household Income, $	39882.7 (15095.5)	0.074	44153.7 (22190.2)	0.935	43079.1 (20696.7)	0.508	44022.3 (19710.5)
Cardiovascular risk factors, %
Current smoking	23.2 (5.6)	<0.001	21.1 (4.9)	0.090	21.7 (5.2)	0.001	20.4 (5.3)
Obesity	41.6 (7.2)	0.021	40.8 (6.8)	0.037	41.0 (6.9)	0.007	39.7 (6.6)
High cholesterol	35.8 (2.8)	0.395	36.1 (3.4)	0.031	36.0 (3.2)	0.028	35.4 (4.1)
Diabetes	15.9 (4.3)	0.376	15.8 (4.7)	0.286	15.8 (4.6)	0.202	15.4 (5.1)
High blood pressure	38.3 (7.8)	0.7995	38.8 (8.2)	0.467	38.7 (8.1)	0.544	38.3 (8.6)
Coronary heart disease prevalence, %	8.0 (1.8)	0.462	8.0 (1.9)	0.180	8.0 (1.9)	0.148	7.8 (2.2)

^aMean (Standard deviation).

^bP-value for the difference with respect to 5km–10.km

^cDistribution of variables are shown regardless of whether they were used in matched analyses.

Distributions of variables used in matched analyses are presented in Figures S2 and S3 as standardized mean differences.

Table 2 presents the associations between proximity to refineries and self-reported CHD prevalence. We did not find statistically significant associations for three distance-based, binary exposure thresholds (<2.5km, 2.5km–5km, and <5km). The tested associations were also nonsignificant in sensitivity analyses, except for one association (Table S1). In contrast, we found a statistically significant association between X_c for 2.5km and for 5km and self-reported CHD prevalence. A one-SD increase in X_c for 2.5km and 5km was associated with a 0.54 (95% confidence interval: 0.01, 1.06) and a 0.33 (0.04, 0.63) percentage point increase in CHD prevalence, respectively. Sensitivity analyses demonstrated consistent results, with point estimates ranging from 0.17 (−0.21, 0.54) to 0.34 (−0.22, 0.91) (Table S2).

Table 2.

Association between proximity to oil refineries and self-reported coronary heart disease prevalence

Exposurea	Percentage point increase (95% confidence interval)b
Proximity to refineries
<2.5km	0.04 (−0.51, 0.59)
2.5–5km	0.12 (−0.19, 0.43)
<5km	0.07 (−0.16, 0.29)
The weighted average of actual petroleum production (Xc)
<2.5km	0.54 (0.01, 1.06)
<5km	0.33 (0.04, 0.63)

^aControls were defined as census tracts located at ≥5km and <10km from a refinery.

^bFor proximity to refineries, this increase is an increase compared to the CHD prevalence (%) of controls (binary variable). For X_c, this increase is an increase per one standard deviation (SD) increase in X_c compared to the CHD prevalence (%) of controls (X_c=0).

Table 3 presents the fraction of variation in self-reported CHD prevalence that is explained by X_c within 5km across seven states. For each of the seven states, the percent of explained geographic variation in CHD prevalence was 1.6 (0.2, 3.1), and the number of cases that are potentially explained by exposure to refineries was 1119.0 (123.5, 2114.2). This percent differed by state: 3.0% (0.3, 5.7) in Arkansas, 2.5% (0.3, 4.8) in Mississippi, 2.0% (0.2, 3.7) in Texas, 1.5% (0.2, 2.8) in Louisiana, 0.5% (0.1, 0.9) in New Mexico, 0.4% (0.0, 0.7) in Oklahoma, and 0.2% (0.0, 0.3) in Alabama. The number of cases explained by exposure to oil refineries was the highest in Texas (791.7 (87.5, 1495.8)), followed by 240.1 (26.5, 453.7) in Louisiana, 41.8 (4.6, 78.9) in Arkansas, 25.4 (2.8, 48.0) in Oklahoma, 12.9 (1.4, 24.4) in Mississippi, 4.1 (0.4, 7.7) in Alabama, and 3.0 (0.3, 5.7) in New Mexico. For a total of 316 census tracts within 5km from refineries, CHD prevalence explained by X_c ranged from 0.02% (0.00, 0.05) to 47.4% (5.2, 89.5) (Figure 2).

Table 3.

Self-reported coronary heart disease prevalence and number of cases that are explained by exposure to oil refineries within 5km distance from refineries by seven states

State	% of prevalence explained (95% confidence interval)	Prevalent cases explained (95% confidence interval)
Alabama	0.2 (0, 0.3)	4.1 (0.4, 7.7)
Arkansas	3.0 (0.3, 5.7)	41.8 (4.6, 78.9)
Louisiana	1.5 (0.2, 2.8)	240.1 (26.5, 453.7)
Mississippi	2.5 (0.3, 4.8)	12.9 (1.4, 24.4)
New Mexico	0.5 (0.1, 0.9)	3.0 (0.3, 5.7)
Oklahoma	0.4 (0.0, 0.7)	25.4 (2.8, 48.0)
Texas	2.0 (0.2, 3.7)	791.7 (87.5, 1495.8)
All seven states	1.6 (0.2, 3.1)	1119.0 (123.5, 2114.2)

Figure 2.

The percentage of coronary heart disease prevalence that is explained by exposure to oil refineries. Each point indicates the percentage at an exposed census tract within a state. Each horizontal line indicates its 95% confidence interval. Census tracts were sorted by the percentage explained for each state (vertical axis)

4. Discussion

The U.S. Environmental Protection Agency (EPA) regulates several classes of air pollutants emitted from oil refineries and other byproducts that may pollute surrounding soil and water under the Clean Air Act., the Clean Water Act., and the Petroleum Refining Effluent Guidelines and Standards. Our findings suggest cardiovascular health impacts of pollution mixtures from oil refineries, meaning disproportionate concentration and small-area variation of CHD in the southern U.S. Petroleum production and refining may become a stronger explanatory factor of geographic differences in CHD prevalence when the petroleum production and refining level is higher and the level of the other contributing factors to the prevalence is lower.

Our exposure metric reflected the number, location, capacity of oil refineries, and wind. Similar approaches have been used in studies of the association between oil and gas wells and various health outcomes (McKenzie et al. 2014; McKenzie et al. 2017; Rasmussen et al. 2016). Such approaches can provide insight regarding the contribution of pollution point sources to regional and small-area variation in CHD. In addition, we proposed integrating wind speeds and directions into the exposure metric, enhancing its accuracy in representing potential exposure to pollution emitted from refineries compared to previous metrics based solely on only the number, location, and capacity of emission sources. Although our exposure metric has the advantage of accommodating geographic heterogeneity in exposure mechanisms and the composition of pollutant mixtures, it is not designed to isolate the health effects of single pollutant classes or discern among the various pathways (e.g., air, water) through which these activities impact health. To inform environmental policy, future research should disentangle the contribution of each pollutant class within byproducts of petroleum production and refining to CHD risk.

We found a significant positive association between residential exposure to oil refineries and CHD prevalence using wind-integrated inverse distance squared-weighted sum of APP (X_c), which was robust to different analytical choices, as confirmed by sensitivity analyses. We did not find consistent and significant associations using solely distance-based binary exposure definitions. This suggests that X_c, which considers count, proximity, the expected productivity of refineries, and wind speeds and directions may be a better surrogate for exposure to pollutant mixtures than distance alone. Nonetheless, the degree of individual exposure to pollutants depends on multiple factors that could not be considered within this study. These include the source and quality of fuels input into the oil refining process, pollution control measures, water treatment systems, chemical compositions of crude oils, and other meteorological factors, among others. Nevertheless, inverse distance squared weighted average of APP for 2.5km and 5km that is similar to X_c, were positively associated with measured levels of SO₂ in other study (Kim et al. 2022). SO₂ is a reasonable approximate pollutant indicator for local concentrations of byproducts because crude petroleum contains a large amount of sulfur and may be coupled with other pollutants emitted from refineries, including fine particles, black carbon, VOCs, NO_x, and PAH (Adebiyi 2022). SO₂ emissions are also a relatively unique signature of exposure to oil refineries due to multiple decades of regulation under the Clean Air Act that have reduced SO₂ emission from other sources, including automotive transport. Moreover, one of the EPA’s recommended regulatory dispersion models also confirmed that SO₂ and VOCs generated from oil refineries in Texas City can travel a sufficient distances to contribute to pollutant exposure within surrounding communities (Chen et al. 2012).

The potential contribution of exposure to oil refineries in the exposed census tracts to CHD prevalence (i.e., $\widehat{\beta }{X}_c/\mathit{\text{Pre}}{v}_c\times 100$) varied by state and census tracts. This variation suggests that petroleum production and refining may become a stronger explanatory factor of geographic differences in CHD prevalence when the petroleum production and refining level is higher and the level of the other contributing factors to the prevalence is lower.

We note several limitations. First, our study is ecological and therefore cannot establish temporal or causal relationships, despite our use of a causal inference method ((G)PS matching). Second, we used CHD prevalence. Prevalence of self-reported CHD (i.e., those who reported ever having been told by a doctor, nurse, or other health professional that they have had angina or CHD), by definition, depends on incidence and case-fatality. Our results do not indicate the association for the incidence or the association for the case-fatality. Third, CHD prevalence and health and behavioral risk factors were based on self-report and could not be triangulated or adjudicated using other data sources. Our source of health information, moreover, required us to dichotomize the vascular risk-factors, which may have a dose-response relationship to CHD outcomes, such as tobacco use and obesity. Also, the CDC PLACES estimates are census tract-level (small areas) estimates from statistical approaches, inherently having uncertainties. Fourth, we were unable to account for the route(s), magnitude, or duration of individual exposures or model complex mixtures of different environmental pollutants such as air, noise, soil, and water. Our exposure metric could not capture variation in major determinants of emissions, such as complex refining processes, pollution control measures, the quality of crude oil inputs, and temporal fluctuations in operating and idle capacity, among others. Exposure to air pollution generated from refineries is also determined by multiple factors including smokestack attributes (e.g., height, diameter), topology, and meteorology, all of which were not fully considered in our study. Fifth, it was not feasible to evaluate potential occupational exposures with our data, either. Sixth, we did not investigate buffer distances other than 2.5km and 5km. We categorized census tracts located at >5km and ≤10km from refineries as the control group. There are competing considerations: to expand exposed groups as to include long-distance impacts from a refinery versus to choose appropriate control groups to adjust for unmeasured spatial confounding. If we expand exposed groups, the analysis is less likely to achieve adjustment for unmeasured spatial confounding because census tracts farther from a refinery are less similar to exposed groups regarding unmeasured spatial confounders (Figure S1). If we narrow the exposed groups, the analysis is more likely to achieve adjustment for unmeasured spatial confounding. However, this may result in underestimation if these census tracts categorized as the control group had been actually exposed to oil refineries. Unmeasured spatial confounding is concerning in our study population because entrenched cardiovascular risks have not been fully explained by many known risk factors (Gebreab et al. 2015; Glynn et al. 2021). We adjusted for this confounding by distance-matching and (G)PS-matching integrated with spatial smoothers. Seventh, because the distance between census tracts and refineries was determined using the geographic centroid of census tracts and point coordinates of oil refineries, the exposure status of some residents of census tracts with a large spatial footprint might have been misclassified. Such census tracts are characterized by uneven population dispersion due to topological features, such as mountainous terrain.

5. Conclusion

We evaluated the association between local exposure to oil refineries and small-area variation in the prevalence of CHD. Our cross-sectional ecological analysis is based on a causal inference method that adjusts for measured confounders and unmeasured spatial confounders and suggests a potential relationship between residential exposure to oil refineries and CHD prevalence. Our research suggests that the sector of oil refineries, a major contributing sector to climate change, may also contribute to higher cardiovascular diseases in the southern U.S. Longitudinal studies with more detailed exposure assessments and incidence-based health outcomes will enhance evidence of the relationship between oil refineries and CHD risk, including analysis of the specific pollutants and their putative contributions to atherogenesis. This new area of research will improve understanding of potentially modifiable exposures that may contribute to the disproportionate concentration of cardiovascular diseases in the southern U.S and their small area variation.

Highlights.

• Association of oil refineries with coronary heart disease (CHD) was examined.

• We found positive association in seven states of the southern U.S.

• The prevalence of CHD explained ranged from 0.02% to 47.4%.

• Residents living nearer to oil refineries may have a higher risk of CHD.

• Oil refineries may explain small area variation in the risk of CHD.

Abbreviations

APP

Actual petroleum production

ACS

American Community Survey

CDC

Centers for Disease Control and Prevention

CDC PLACES

CDC Population Level Analysis and Community EStimates

CHD

Coronary Heart Disease

CGPS

Conditional Generalized Propensity Score

EPA

Environmental Protection Agency

GAM

Generalized Additive Model

GPS

Generalized Propensity Score

NNWR

One-to-one Nearest Neighbor matching With Replacement

NNWoR

One-to-one Nearest Neighbor matching With”o”ut Replacement

NNCWR

One-to-one Nearest Neighbor Caliper matching With Replacement

NNCWoR

One-to-one Nearest Neighbor Caliper matching With”o”ut Replacement

NO_x

Nitrogen oxides

O₃

Ozone

PAH

Polycyclic Aromatic Hydrocarbon

PM

Particulate Matter

PS

Propensity Score

SD

Standard Deviation

SMD

Standardized Mean Difference

SO₂

Sulfur dioxide

U.S.

United States

VOCs

Volatile Organic Compounds

Data sharing statement

Census tract-level variables are publicly available from the United States Census Bureau (https://data.census.gov) and CDC PLACES website (https://www.cdc.gov/places/index.html). Information on oil refineries is available in the United States Energy Information Administration (https://www.eia.gov/petroleum/refinerycapacity/). The data that support the findings of this study are openly available at the following URL: https://github.com/HonghyokKim/CHD_PPR.