Sulfur dioxide reduction at coal-fired power plants in North Carolina and associations with preterm birth among surrounding residents : Environmental Epidemiology

Secondary Logo

Journal Logo

Original Research Article

Sulfur dioxide reduction at coal-fired power plants in North Carolina and associations with preterm birth among surrounding residents

Wilkie, Adrien A.a,b; Richardson, David B.b,c; Luben, Thomas J.d; Serre, Marc L.e; Woods, Courtney G.e; Daniels, Julie L.b

Author Information
Environmental Epidemiology 7(2):p e241, April 2023. | DOI: 10.1097/EE9.0000000000000241


What this study adds

No research to-date has evaluated the association between residential proximity to coal-fired power plants (CFPP) and preterm birth (PTB) within North Carolina (NC) although the state has a long history of dependence on coal combustion for electricity. Understanding environmental factors that contribute to PTB is important; PTB is strongly associated with infant mortality and an increased risk for later morbidities. Results suggest a reduction of PTB among NC residents 4–<10 miles of the CFPPs that installed scrubbers—demonstrating a potential impact of this regulation on reducing PTB among those living near CFPPs.


Coal-fired power plants (CFPPs) combust coal to produce electricity while concomitantly releasing emissions into the air, discharges into surface waters, and producing waste byproducts (such as coal ash) stored on- and off-site in landfills and surface impoundments.1–3 CFPPs are major contributors of air pollution, especially the primary gas-phase pollutant sulfur dioxide (SO2).1,2,4 SO2 is a precursor to fine particulate matter (PM2.5) sulfate, and CFFPs are the main source of anthropogenic PM2.5 sulfate.5,6 Between 2000 and 2008, approximately 60% of electric power generation was produced by coal in North Carolina (NC).7 The 2002 NC Clean Smokestack Act (CSA) required NC CFPPs to substantially reduce SO2 emissions by 2009 and 2013: “Utilities also must reduce actual sulfur dioxide (SO2) emissions from 489,000 tons in 1998 to 250,000 tons by 2009 (49% reduction) and 130,000 tons by 2013 (73% reduction).”8 Utilities met the CSA SO2 2009 benchmark by installing flue gas desulfurization equipment (FGD; i.e., scrubbers) on the coal electricity generating units (EGUs) at seven large (>150 megawatts) CFPPs within the state. The utilities then met the CSA SO2 2013 benchmark by retiring the coal EGUs at seven other large CFPPs within the state.9 Large statewide reductions in reported SO2 emissions and ambient SO2, as well as secondary formation PM2.5, were observed during the decade following the passage of the NC CSA.7,10–12 NC’s state-level change in air pollution regulations, and subsequent interventions at CFPPs to reduce SO2 emissions, provides a unique opportunity to investigate the impact of the interventions employed to control SO2 pollution from CFPPs on gestational parent and child health.

A causal relationship is established between ambient SO2 and respiratory effects and ambient PM2.5 and cardiovascular effects and mortality.4,5 However, the evidence is less definitive for reproductive/developmental effects with ambient SO2 being inadequate to infer, and PM2.5 being suggestive of, but not sufficient to infer, a causal relationship.4,5 Evidence suggests the inflammatory pathway may be an important mechanism between air pollution exposure and PTB either via direct stimulation of inflammation or indirect influence on oxidative stress.13–15 The results of a systematic review of 25 studies published between 1980 and 2015 that examined the association between ambient air pollution exposure (SO2, PM, carbon monoxide, nitrogen dioxides, ozone) and adverse pregnancy outcomes in China indicated that there was more consistent evidence of a statistically significant association between SO2 and PTB than the other pollutants evaluated.16 Among a growing body of literature evaluating associations between pregnant people living near power plants and adverse birth outcomes, all nine studies (with seven studies located in the United States: California, Florida, New Jersey, Tennessee, Wisconsin, and eastern region of the United States) suggest adverse associations.15,17–24 However, few of these studies focus specifically on CFPPs and PTB, and none are based in NC.

No research to-date has evaluated the association between residential proximity to CFPPs and PTB within NC although the state has a long history of dependence on coal combustion for electricity (e.g., in 2002, NC ranked seventh in the nation for electricity net generation from coal by electric utilities),7,25 and a consistently higher rate of PTB than the nation (e.g. in 2005, NC ranked 11th in the nation for PTB rate).26,27 NC also shows a persistent disparity in rates of PTB among non-Hispanic Black gestational parents compared with non-Hispanic white gestational parents, with NC PTB rates from 2003 to 2012 ranging from 15.7 to 18.9 per 100 live births for non-Hispanic Black gestational parents and 10.0–12.3 for non-Hispanic white gestational parents).28 Understanding environmental factors that contribute to adverse birth outcomes is important because PTB and reduced birth weight are strongly associated with infant mortality and an increased risk for later morbidities such as neurodevelopment disorders, cardiovascular disease, and endocrine disorders.29,30 In addition, the economic burdens of premature babies are substantial, with the estimated 2006 annual cost of PTB in the United States reaching $26 billion.31

In response to the NC CSA, the two emissions strategies implemented to reduce SO2 at 14 large CFPPs in NC were installment of desulfurization equipment (i.e., scrubbers) on coal EGUs or retirement of coal EGUs. The seven CFPPs that installed scrubbers were NC’s largest CFPPs and remained operating as such through at least 2015; where as, the seven CFPPs that retired coal EGUs either fully retired the power plant facility or transitioned the power plant from coal to natural gas. Using a quasi-experimental study design, we leveraged this policy intervention to investigate source-specific changes in emissions and associations with PTB. Unlike covariate adjusted observational studies that control for measured confounders, this quasi-experimental study design controls for measured and unmeasured confounders that are time-invariant under specified assumptions. The objective of this study was to evaluate the impact of these emissions reduction strategies (i.e., interventions) on the prevalence of PTB among nearby residents.


Study population

From live birth records obtained from the Birth Defects Monitoring Program within the State Center for Health Statistics of the NC Department of Health and Human Services, we constructed an administrative cohort of NC resident live births between 1 January 2003 and 31 December 2015 (N = 1,600,409). We excluded non-singleton births (N = 54,830) from our study population due to being a strong risk factor for PTB.32 We also excluded births with reported clinically estimated gestational age less than 20 weeks (N = 641), greater than 44 weeks (N = 11), or missing (N = 853). We excluded births that completed 20 weeks of estimated gestational age before 1 January 2003 (N = 40,620) and could have completed 44 weeks of estimated gestational age after 31 December 2015 (N = 11,125) to minimize fixed cohort bias.33 We additionally excluded births with reported residences that were unable to be sufficiently geocoded (N = 2,508).

We excluded births ≥ 15 miles from the 14 large NC CFPPs that implemented interventions to reduce SO2 emissions. Because we were interested in the local impact of SO2 emissions reduction strategies at 14 large (>150 MW) CFPPs pre- and postintervention, we restricted and stratified the study population into two groups: (1) births within 15 miles of at least one CFPP that installed scrubbers and had gestational parent with last menstrual period (LMP) that occurred within the pre- or postintervention window (CFPP-scrubber, N = 42,231) and (2) births occurring within 15 miles of at least one CFPP that retired and had gestational parent with LMP that occurred within the pre- or postintervention window (CFPP-retired, N = 41,218). The observations within CFPP-scrubber and CFPP-retired cohorts are not mutually exclusive; however, we additionally analyzed these two cohorts together (CFPP-all) (Supplemental Material, Figure S1;

The University of North Carolina at Chapel Hill’s Office of Human Research Ethics approved this study (#19-0050).

Exposure assessment

We defined the intervention as a CFPP installing scrubbers on existing coal EGUs or retiring coal EGUs. We developed the exposure across two axes of time and space. To assign the temporal aspect of exposure, we accessed United States (US) Environmental Protection Agency (EPA) Air Markets Program Data (AMPD) to identify the SO2 reduction intervention dates at each CFPP (i.e., dates scrubbers were installed or coal units retired), and we validated those dates by evaluating the monthly SO2 emissions reported by each CFPP from 2002 to 2015. We carefully constructed the pre- and postintervention windows of exposure by deriving eligible births from a 1-year period of estimated LMPs (i.e., reported date of birth minus clinically estimated gestational age). We restricted the analyses to those births for which the time from LMP to birth occurred entirely before the intervention (pre-) or entirely after the intervention (post-) dates. Additionally, we standardized the 1-year eligible months of LMP for each CFPP in the pre- and postintervention time periods to control for any potential outcome- or exposure-related temporal/seasonal variation (e.g., monthly trends for when people get pregnancy or monthly trends in electricity generation/emissions).33,34 For example, if the 1-year LMP postintervention period was from May 1 to April 30 of a given 12-month timespan then the preintervention period would similarly be from May 1 to April 30 of a given 12-month timespan.

The postintervention exposure period started 3 months after the final coal EGU intervention date at the CFPP—meaning every eligible birth experienced at least a 3-month preconception washout after the final coal EGU intervention date. (Due to temporal boundary issues, the postintervention period for two sites started 1–2 months after the final intervention.) The start of the preintervention exposure period occurred at least 2 years before the initial coal EGU intervention date at the CFPP. Additionally, the start and end months of the 1-year LMP exposure windows matched across the pre- and postintervention periods for the CFPP. This approach ensured that the entire pregnancy (i.e., from LMP to live birth) for all eligible births within the preintervention period occurred completely before the start of the first coal EGU intervention date and the entire pregnancy for all eligible births within the postintervention period occurred completely after the end of the last coal EGU intervention date. For example, one CFPP (GG Allen) in the CFPP-scrubber group had the initial scrubber installed February 2009 and the final scrubber installed May 2009. The 1-year LMP for the postintervention exposure period began 1 September 2009 and ended 31 August 2010 (making the possible range of live births occurring between February 2010 and July 2011). The 1-year LMP for the preintervention exposure period began 1 September 2006 and ended 31 August 2007 (making the possible range of live births occurring between February 2007 and July 2008).

To account for proximity to the source of exposure, we geocoded residential address at time of birth as recorded on the birth record and calculated the distance from the residence at birth to each of the 14 CFPPs. We categorized births into those occurring 0–<4, 4–<10, or 10–<15 miles from at least one CFPP that installed scrubbers or retired. We used methods employed by Casey et al17 to determine the distance cut-points (buffers) for the CFPP cohorts—visually from mean difference in proportion of PTB by incremental distance from CFPPs. Births in the 10- to <15-mile buffer served as the distal comparison group because we considered this distance to be less impacted by the interventions at the CFPPs than those closer to the power plants. Although research shows that air emissions from CFPPs can travel farther than 30 miles, similar to Casey et al, we were interested in investigating the local impacts surrounding these CFPPs (i.e., within 15 miles). We assumed those living closest to the CFPP prior to the intervention were the most exposed while those living farthest from the CFPP postintervention were the least exposed. Two of the CFPPs that installed scrubbers were within 10.6 miles of each other and had almost entirely overlapping buffers and subsequently eligible births. We combined the births surrounding those two CFPPs together.

Outcome and covariate assessment

We defined PTB as delivery at 20–36 weeks of completed gestation, with term birth defined as 37–44 weeks of completed gestation, based on the clinical estimate of gestation age provided on the birth certificate. We also stratified by timing of PTB: early (less than 32 weeks completed gestation) or moderate-to-late (32–36 weeks completed gestation). Covariates accessed from the birth records included gestational parent’s age (continuous), race/ethnicity (non-Hispanic American Indian, non-Hispanic Asian/Pacific Islander, non-Hispanic Black, Hispanic regardless of race, non-Hispanic white, and non-Hispanic/unknown), education (less than high school, high school and greater), marital status at delivery (yes, no), cigarette smoking during pregnancy (yes, no), Medicaid at delivery (yes, no), and parity (first live birth, second live birth, and third or more live births). We evaluated the distributions of these covariates within pre- and postintervention periods as well as across distance buffers to get a crude understanding of potential temporal trends and the appropriateness of the comparison group.35,36

These covariates in the birth records are derived from the “Mother’s Worksheet for Child’s Birth Certificate” that the gestational parent is supposed to complete at the time of delivery. Although the birth certificate collection form(s) and subsequent birth records database uses gendered language (e.g., “mother,” “maternal”), we purposely use the more inclusive term “gestational parent” to represent the pregnant person who gave birth.37 Information about gestational parent race and ethnicity had dimensions of self-classification at the time of data collection (i.e., the race and ethnicity checked on the birth certificate worksheet had constrained options,38 and were further bridged into combined race/ethnicity categories by the National Center for Health Statistics for consistency and comparability of this variable over time and across data collection systems.39

Statistical analysis

We used a difference-in-differences (DiD) model that leverages a “natural” shift to mimic an experimental research study design for analyses of observational data.40 We estimated the association between CFPP SO2 reduction intervention strategies and the prevalence of PTB surrounding the CFPPs. Our model assumes that any change in secular trends for confounding factors would be captured by reference to the comparison group but changes in outcome due to the intervention (reduced exposure) can be observed within each geographically defined exposure group.17 The resulting DiD estimator is equated to the difference in PTB prevalence attributable to the SO2 reduction interventions if the assumptions hold that (1) the secular trends in the exposed and control groups are parallel and (2) the model is correctly specified. We generated DiD estimates and 95% confidence intervals (CIs) using generalized linear fixed intercept models for the CFPP-scrubber and CFPP-retired cohorts. For example, among the CFPP-scrubber population, and treating those births 10–<15 miles from a CFPP-scrubber as the reference group

Pr(Preterm birth for gestational parentanear CFPPbthat installed scrubbers) =β0+ (β1× 0to4milesab) + (β2× 4to10milesab)+ (β3× postab) + (β4× [0to4milesab× postab]) +(β5× [4to10milesab× postab]) +εab

where 0 to 4 miles is the binary indicator for gestational parent residence at time of delivery within 0–<4 miles of CFPP b at birth; 4 to 10 miles is the binary indicator for gestational parent residence at time of delivery within 4–<10 miles of power plant b at birth; post is a binary indicator of CFPP b intervention status (pre-/postintervention with postintervention equals to 1); β4 and β5 are the DiD estimates of interest; εab is the error term (distributed normal with mean zero).


The DiD estimate =(PTB prevalence0 to 4 miles = 1; postintervention = 1 PTB prevalence0 to 4 miles = 1; postintervention = 0)  (PTB prevalence10 to 15 miles = 1; postintervention = 1 PTB prevalence10 to 15 miles = 1; postintervention = 0)

Thus, interpretation of a negative DiD estimate (i.e., the mean difference in prevalence of PTB) would be a reduction in PTB after the intervention occurred, whereas a positive DiD estimate would be an increase in PTB after the intervention occurred.

Due to the well-established and persistent disparity in prevalence of preterm delivery among non-Hispanic Black compared with non-Hispanic white gestational parents,31 as well as the documented national trend for Black residents to be more likely than white residents to live near industrial facilities, including CFPPs,41–43 we additionally stratified the models by race/ethnicity of the gestational parent. We consider race/ethnicity not as biological in nature but rather a social construct that may be a proxy for historical and current systemic racism.44 We collapsed gestational parent race/ethnicity from a six-category variable to a four-category variable (non-Hispanic Black, Hispanic, non-Hispanic white, and other) to account for sparsity of data within our stratified analyses by race/ethnicity. For comparison to Casey et al,17 we also stratified by timing of PTB (early or moderate-to-late).

Sensitivity analyses

We conducted sensitivity analyses to evaluate the appropriateness of the assignment of buffer distances by evaluating how sensitive our results were to the specification of distance definitions. Similar to Casey et al, we undertook the main analysis using pooled observations of births nearby CFPPs.17 However, to acknowledge suspected heterogeneity in exposure potential and characteristics of eligible births surrounding each CFPP, we additionally ran models for each CFPP individually. Given sample size limitations at many of the individual CFPPs, we further conducted a meta-analysis to estimate the fixed effect and random effects for CFPP-scrubber, CFPP-retired, and CFPP-all. Because our main analysis does not distinguish between births within 15 miles of one CFPP or more than one CFPP, we conducted a sensitivity analysis that removed observations from the comparison group (10–<15 miles) at each CFPP that were within 15 miles of one of the other large CFPP evaluated in this study.

We further evaluated two important assumptions required for a DID analysis related to stable populations and parallel trends.35,36 First, we checked that key sociodemographic characteristics of births nearby the CFPP with the largest sample size (CFPP-GG Allen: N = 21,182) had similar composition in the exposed groups (i.e., 0–<4 and 4–<10 miles) and comparison group (i.e., 10–<15 miles) pre- and postintervention by graphing the annual proportions of gestational parent educational attainment and race/ethnicity within the temporally unrestricted birth cohort from 2004 to 2014. Second, we assessed the parallel trends assumption by graphing the annual PTB prevalence for CFPP-GG Allen for each year in the temporally unrestricted birth cohort by our exposed groups and comparison group before the initiation of the intervention (first scrubber installed February 2009).

The main DiD estimates (from unadjusted models) should be similar to adjusted models if the parallel trends assumption is met. However, to further test that assumption we additionally adjusted for potential confounders that we identified a priori based on substantive knowledge of gestational exposure to CFPP-related air pollution and PTB: gestational parent’s age at delivery, race/ethnicity, Medicaid status at delivery, marital status, and parity. When NC transitioned from the old version of the birth certificate to the new version in 2010, some questions were completely missing from the birth records for that year, such as gestational parent education and cigarette smoking use during pregnancy. Consequently, we were unable to adjust for these variables in the sensitivity analysis.

All analyses were performed using SAS version 9.4 (SAS Institute Inc., Cary, NC), R version 3.6.1, RStudio version 1.2.5001, and ArcGIS version 10.5.


Between 2003 and 2015, 14 large CFPPs (>150 MW) in NC implemented strategies to reduce their SO2 air emissions. Figure 1 shows the geographic distribution of these 14 CFPPs by capacity based on their 31 December 2015 operating status as well as the intervention strategy employed at each plant. Generally, the largest capacity CFPPs installed scrubbers and continued as coal-burning (CFPP-scrubber average capacity = 1,402 MW, median = 1,140 MW, maximum = 2,422 MW, minimum = 376 MW), whereas the smaller of the >150 MW CFPPs retired their coal EGUs by closing the power plant or transitioning to a natural gas-fired power plant (CFPP-retired average capacity = 347 MW, median = 316, maximum = 575 MW, minimum = 171 MW) (Supplemental Material, Figure S2;

Figure 1.:
Location of the 14 North Carolina coal-fired power plants included in this study characterized by their 2003 capacity, intervention strategy, and operating status as of 31 December 2015. CFPP indicates coal-fired power plant; MW, megawatt; NG, natural gas.

As illustrated by Figure 2 (and Supplemental Material, Figure S3; the CFPP-scrubber group had much higher reported SO2 emissions preintervention than the CFPP-retired group although both groups showed large reductions in SO2 emissions postintervention. From 2003 to 2015, the CFPP-scrubber group had collectively much higher reported monthly SO2 emissions than the CFPP-retired group (Figure 2A). The interventions implemented for the CFPP-scrubber group primarily happened before 2010 and the interventions implemented for the CFPP-retired group happened after 2010. Preintervention, the CFPP-scrubber group emitted 4,933 tons of SO2 monthly on average, which reduced to 342 tons in postintervention. Preintervention, the CFPP-retired group emitted 909 tons of SO2 monthly on average, which reduced to <1 ton in postintervention (Figure 2B).

Figure 2.:
Monthly SO2 emissions reported by 14 coal-fired power plants in North Carolina (2002–2015). A, Summed monthly SO2 emissions grouped by CFPP-scrubber and CFPP-retired. B, CFPP-scrubber (left) and CFPP-retired (right) monthly SO2 emissions grouped by 1-year LMP exposure window pre- and postintervention. (Data source: US EPA’s AMPD.) CFPP indicates coal-fired power plant; LMP, last menstrual period; SO2, sulfur dioxide.

The study population comprised 42,231 NC resident live births within 15 miles of at least one of the seven CFPP-scrubber and 41,218 NC resident live births within 15 miles of at least one of the seven CFPP-retired. Overall, PTB was 9.1% and 8.2% for the CFPP-scrubber cohort and CFPP-retired cohort, respectively. Stratified by gestational parent’s race/ethnicity, PTB within CFPP- scrubber and CFPP-retired cohorts for non-Hispanic Black gestational parents (12.9% and 11.2%) and non-Hispanic American Indian gestational parents (10.1% and 10.9%) were higher than the overall cohort PTBs (9.1% and 8.2%) (Table 1). Figure 3 shows the mean difference and 95% CIs in the proportion of PTB pre- and postintervention for the CFPP-scrubber, CFPP-retired, and CFPP-all cohorts by 1-mile increments, which informed the distance buffers of 0–<4, 4–<10, and 10–<15 miles.

Table 1. - Prevalence of preterm birth overall and by gestational parent race/ethnicity within the CFPP-scrubber and CFPP-retired cohorts, North Carolina.
CFPP-scrubber CFPP-retired
PTB, n (%) PTB, n (%)
Overall 42,231 (9.1) 41,218 (8.2)
 American Indian, nH 119 (10.1) 1,086 (10.9)
 Asian, nH 1,398 (6.9) 1,678 (7.0)
 Black, nH 9,749 (12.9) 10,626 (11.2)
 Hispanic 6,975 (7.8) 6,112 (6.7)
 White, nH 23,928 (8.0) 21,618 (7.2)
 Other, nH/unknown 62 (9.7) 98 (9.2)
CFPP indicates coal-fired power plant; nH, non-Hispanic; PTB, preterm birth; n, number of observations; %, percent.

Figure 3.:
Mean difference and 95% confidence intervals in the proportion of preterm birth by distance in miles from coal-fired power plants. Mean difference = postintervention minus preintervention. A) CFPP-scrubber cohort. B) CFPP-retired cohort. C) CFPP-all cohort. *For <1 mile from CFPP-all and CFPP-scrubber, the 95% CIs extend past the shown y axis due to small numbers of observations (−9.4 to 11.3 and −15.3 to 12.4, respectively). For <1 mile from CFPP-retired, the mean difference and 95% CI could not be calculated because no cases of PTB occurred within that distance. CFPP indicates coal-fired power plant; PTB, preterm birth; 95% CI, 95% confidence interval.

Table 2 shows the characteristics of the live births by proximity to CFPP and pre-/postintervention for both CFPP-scrubber and CFPP-retired cohorts. The number of births was similar pre- and postintervention within each buffer group. The fewest number of births occurred closest to the CFPPs (0–<4 miles). We looked at characteristics across exposure buffers within the preintervention group for each cohort to assess the appropriateness of the comparison group (10–<15 miles). When comparing by distance from the CFPPs, the characteristics for the CFPP-scrubber cohort were similar between the 4–<10 and 10–<15 miles within the preintervention groups for all covariates. However, for the 0- to <4-mile group, gestational parent race/ethnicity, education, Medicaid status, and marital status were different from the distributions in the 10- to <15-mile group by having more non-Hispanic white gestational parents (74.3% vs. 54.7%), fewer gestational parents with less than a high school degree (13.9% vs. 23.8%), fewer gestational parents with Medicaid at delivery (31.5% vs. 48.2%), and more married gestational parents (72.91% vs. 58%).

Table 2. - Characteristics of births near CFPP in North Carolina pre- and postintervention for CFPP-scrubber and CFPP-retired cohorts as well as by distance groups (0–<4, 4–<10, and 10–<15 miles)
CFPP-scrubber CFPP-retired
0–<4 miles 4–<10 miles 10–<15 miles 0–<4 miles 4–<10 miles 10–<15 miles
N (%) Pre Post Pre Post Pre Post Pre Post Pre Post Pre Post
Total births 988
PTB 73
GP age (years) at delivery
 <20 71
 20–24 177
 25–29 282
 30–34 282
 35+ 176
 Missinga .
GP race/ethnicity
 American Indian, nH 4
 Asian/Pacific Islander, nH 35
 Black, nH 100
 Hispanic, regardless of race 115
 White, nH 734
 Other, nH/unknown 0
GP education b
 Less than high school 136
 High school and more 845
 Missing 7
GP Medicaid status at delivery
 No 677
 Yes 311
GP marital status
 Married 720
 Unmarried 268
 Missing .
Cigarette smoking use during pregnancy b
 No 891
 Yes 93
 Missing 4
 Primiparous (1st live birth) 422
 Multiparous (2nd live birth) 339
 Multiparous (≥3rd live birth) 226
 Missing 1
aMissing (.) observations were not included in calculations of percent.
bVariable missing for 2010 births.
CFPP indicates coal-fired power plant; GP, gestational parent; N, number of observations; nH, non-Hispanic; PTB, preterm birth; %, percent of observations .

For the CFPP-retired cohort, when comparing the composition of the exposed groups (0–<4 and 4–<10 miles) to the comparison group (10-<15 miles) within the preintervention, the groups were less similar than the CFPP-scrubber cohort. Furthermore, for the CFPP-retired cohort, the 0- to <4-mile buffer was more different to the comparison group than the 4- to <10-mile buffer group to the comparison group for gestational parent race/ethnicity, marital status, Medicaid status, and cigarette smoking during pregnancy. The percentage of non-Hispanic Asian (0–<4: 1.9%, 4–<10: 2.7%, 10–<15: 4.9%) and Hispanic gestational parents (12.5%, 13.6%, 17.0%) increased with increasing buffer distance from CFPP-retired, whereas the percentage of non-Hispanic Black gestational parents (31.0%, 26.9%, 22.5%) decreased. Gestational parent marital status of married increased moving away from the CFPPs (47.0%, 53.2%, 58.3%), whereas those with Medicaid at delivery (60.2%, 54.1%, 49.5%) and cigarette smoking during pregnancy (17.8%, 14.2%, 11.3%) decreased. Gestational parent education was similar between the 4- to <10-mile buffer and the 10- to <15-mile buffer (24.9% vs. 25.1%, respectively), whereas the 0- to <4-mile buffer had a higher percentage of gestational parents with less than a high school degree than those in the 10- to <15-mile buffer (30.9% vs. 25.1%, respectively) (Table 2).

When comparing by pre- and postintervention, among the births closest (0–<4 miles) to the CFPPs, the prevalence of PTB increased from 7.4% to 9.1% in areas nearest CFPPs that installed scrubbers and decreased from 8.7% to 8.4% in areas nearest CFPPs that retired coal EGUs. Among births within 4–<10 miles of CFPPs, the prevalence of PTB decreased from 9.9% to 8.5% after scrubbers were installed and from 9.0% to 7.6% after coal EGUs were retired. For gestational parents within 4–<10 miles from a CFPP, we estimated that the absolute prevalence of PTB decreased by −1.5% (95% CI: −2.6, −0.4) associated with the installation of scrubbers and decreased by −0.5% (95% CI: −1.6, 0.6) associated with coal EGU retirements (Figure 4A and D, and Supplemental Material, Table S1; When combing both SO2 reduction intervention strategies (CFPP-all), we estimated that the absolute prevalence of PTB decreased by −1.0% (95% CI: −1.8, −0.2) (Figure 4G and Supplemental Material, Table S1; [From the row for PTB in Table 2 (and only for the PTB row), the crude DiD estimates for PTB prevalence can be calculated. For example, the crude DiD estimate for PTB prevalence within 4–<10 miles of a CFPP that installed scrubbers is (8.5–9.9)–(9.1–9.0) = −1.5%. This matches the result from Equation 1, which yielded the estimate of −1.5% (95% CI: −2.6, −0.4).] This suggests that the installation of scrubbers decreased the prevalence of PTB by ~150 preterm births per 10,000 live births. However, for those living within 0–<4 miles of CFPP-scrubber, the absolute prevalence of PTB increased by 1.5% (95% CI: −1.0, 4.0), although this estimate is imprecise and crosses the null (i.e., 0.0).

Figure 4.:
Absolute difference in proportions of preterm birth pre-intervention versus post-intervention to reduce sulfur dioxide emissions at coal-fired power plants in North Carolina. Overall results: A, D, G. Stratified by early and moderate-to-late PTB: B, E, H. Stratified by gestational parent race/ethnicity: C, F, I. Top row (A–C) shows CFPP-scrubber cohort results. Middle row (D–F) shows CFPP-retired cohort results. Bottom row (G–I) shows CFPP-all cohort results. No results provided for Asian, nH in (F) and (I) because models did not converge. Distance categories are 0 miles to less than 4 miles (circles) and 4 miles to less than 10 miles (triangles), compared with 10 miles to less than 15 miles. CFPP indicates coal-fired power plant; DiD, difference-in-differences; nH, non-Hispanic; 95% CI, 95% confidence interval.

When stratified by race/ethnicity, the absolute prevalence of PTB decreased for births to non-Hispanic Asian and non-Hispanic white gestational parents within the 4- to <10-mile buffer by −8.2% (95% CI: −14.1, −2.4) and −2.0% (95% CI: −3.4, −0.6), respectively, for the CFPP-scrubber cohort. Among those living within 0–<4 miles of CFPP-scrubber, the absolute prevalence of PTB decreased by −4.0% (95% CI: −13.9, 5.8) for births to non-Hispanic Asian gestational parents, but increased by 7.0% (95% CI: 0.6, 13.4) for births to Hispanic and 1.9% (95% CI: −7.5, 11.2) for births to non-Hispanic Black gestational parents (Figure 4C and F). Estimates stratified by race/ethnicity, except for non-Hispanic white gestational parents, were consistently more imprecise than the overall estimates.

For the CFPP-retired cohort, the DiD estimates were generally imprecise (i.e., CIs were wide relative to the magnitude of the point estimates), null (i.e., the magnitude did not differ substantially from zero), or both for the 0- to <4- and 4- to <10-mile buffers across all race/ethnicity categories. However, non-Hispanic Black and Hispanic residents within 0–<4 miles of CFPP-retired had absolute prevalence of PTB increase by 1.9% (95% CI: −2.7, 6.4) and 1.4% (−3.8, 6.6), respectively, although imprecise. These estimates are in the same direction as the non-Hispanic Black and Hispanic residents within 0–<4 miles of CFPP-scrubber (Figure 4C and F).

When stratified by timing of PTB, most DiD estimates regardless of distance buffers, CFPP cohort, or timing were imprecise and generally null. However, for the CFPP-scrubber cohort within the 4–<10 miles buffer, moderate-to-late PTB had an absolute prevalence of PTB decrease by −1.4% (95% CI: −2.5, −0.4) (Figure 4B).

The results aggregating both CFPP-scrubber and CFPP-retired cohorts—(CFPP-all)—are essentially the average of the two stratified cohorts (Figure 4). The interpretation of the DiD estimates and 95% CIs from our sensitivity analyses (i.e., adjusted analyses; fixed effect and random effects meta-analyses; analysis removing observations from the comparison group at each CFPP that were within 15 miles of one of the other large CFPP evaluated in this study) were consistent with our main analyses for CFPP-scrubber, CFPP-retired, and CFPP-all (Supplemental Material, Table S1; The DiD estimates from the adjusted models demonstrate that the main DiD estimates remain robust after controlling for baseline time-invariant covariates in the model.

Additionally, our assessment of the parallel trends assumption by graphing the annual PTB prevalence for CFPP-GG Allen illustrates PTB was not differentially changing in the comparison group (10–<15 miles from CFPP). PTB annual trends were approximately parallel for the 4- to <10-mile group and comparison group before the installment of scrubbers on CFPP-GG Allen in 2009. We see more year-to-year variation in trend for the 0- to <4-mile group before the installment of scrubbers, possibly driven by small sample size within that buffer (Supplemental Material, Figure S4; Our assessment of stable populations for births nearby CFPP-GG Allen indicate that the annual secular trends from 2004 to 2014 for gestational parent education and race/ethnicity did not change in a meaningful way following the intervention to reduce SO2 emissions at the CFPP in 2009 (Supplemental Material, Figure S5;


Using a DiD design, we evaluated whether interventions to reduce SO2 emissions (installment of FGD scrubbers on coal EGUs or retirement of coal EGUs) at CFPPs in NC were associated with changes in prevalence of PTB. Our findings suggest that SO2 emission reduction interventions were associated with a decrease in PTB for gestational parents living within 4–<10 miles compared with 10–<15 miles away, especially among those living near CFPPs that installed scrubbers on coal EGUs. We saw a similar DiD result when stratified by timing of PTB among gestational parents who experienced moderate-to-late PTB. When stratified by race/ethnicity, this decrease in the prevalence of PTB became stronger among non-Hispanic Asian and non-Hispanic white gestational parents living near CFPP-scrubber sites.

Our findings were imprecise and generally null among those living within 0–<4 miles regardless of the intervention type. However, the pattern of the effect estimates falling above the null suggested an increase in the prevalence of PTB. When stratified by gestational parent’s race/ethnicity, non-Hispanic Asian gestational parents had a decrease in PTB, whereas Hispanic gestational parents had an increase in PTB within the 0-<4 mile buffer for those living near CFPP-scrubber sites. Among non-Hispanic Black gestational parents, PTB increased near CFPP-scrubber and CFPP-retired sites although we were limited in our ability to stratify by race/ethnicity due to power limitations.

The cause of disparity by gestational parent race/ethnicity in the prevalence of PTB is not well understood, and the differences often persist even after adjusting for well-established risk factors, such as smoking, gestational parent education level, prenatal care, and socioeconomic status.45–48 Recent research suggests that stress and chronic worry due to interpersonal racism and structural racism may contribute to the risk of PTB.49–53 For our analyses, we consider race/ethnicity not as biologically deterministic but rather a social construct that may proxy, although insufficiently, for historical and current structural racism that we are otherwise unable to measure.54 The administrative collapsing of the already constrained options for (assumed) self-report of race and ethnicity on the birth certificate limited our ability to disentangle race and ethnicity as well as distinguish differences among subgroups vulnerable to racism in NC. Still, we felt it important to evaluate the association between CFPP interventions and PTB by race/ethnicity because of the known disparities.28,31

NC birth records have been shown to be reliable for gestational parent and infant demographics.55,56 However, due to the change in NC birth certificate form in 2010, information on gestational parent education and cigarette smoking during pregnancy were not recorded for all births during that year. This is reflected in the postperiods for the CFPP-scrubber cohort and pre-periods for the CFPP-retired cohort. Because there is no indication within the 2003–2015 births data that the 2010 characteristics would be markedly different than surrounding years, we have confidence that the proportional distributions would remain the same if those 2010 variables were known. However, in our sensitivity analyses, we were unable to adjust for gestational parent education or smoking due to lack of reporting for the 2010 calendar year. Additionally, the birth certificate form changed how it asked education attainment resulting in inconsistent variable measurement prior to 2010 compared with after 2010. We harmonized the education variable by collapsing high school degree and more than high school degree together for our study period, which minimized the measurement error. However, this reduced resolution may have left us with residual confounding for a key covariate in our sensitivity analysis—although an advantage of the DiD study design, if assumptions are met, does not require an additional model adjustment of potential time-invariant cofounders. Additionally, although the covariates may have differed in some instances between the exposed and comparison groups, the DiD approach does not require exchangeability for time-invariant baseline covariates.

This quasi-experimental study design—DiD—utilized a “natural” shift in SO2 emissions to mimic an experimental research study design for analyses of observational live birth data, which helps with confounding control present in traditional observational study designs.40 We must rely on observational or quasi-experimental study designs when evaluating this type of research question given that the hypothetical experiment should not (and feasibly could not) be undertaken. In this case, we cannot randomize populations (i.e., soon-to-be gestating individuals) to intervention and nonintervention locations (i.e., living nearby CFPPs that installed scrubbers vs. not, or retiring coal EGUs vs. not). Casey et al. similarly leveraged a DiD approach showing that oil (n = 6) and coal (n = 2) power plant retirements in California from 2001 to 2011 resulted in a decrease in PTB at distances of 0–5 km (~0–3.1 miles) and 5–10 km (~3.1–6.2 miles) compared with 10–20 km (~6.2–12.4 miles) with significant decreases in PTB when stratifying for moderate-to-late PTB, and stronger decreases in PTB between power plant retirement and PTB among non-Hispanic Black and Asian gestational parents compared with non-Hispanic white and Hispanic gestational parents.17 Daouda et al evaluated the association between county-level CFPP-attributable air emission across 289 US counties and racial disparities in PTB from 2000 to 2018; they found an association that was stronger among non-Hispanic white gestational parents than non-Hispanic Black gestational parents between PM2.5 emissions from CFPPs and PTB rates.18 Ha et al investigated distance to nearest power plant and adverse birth outcomes in Florida from 2004 to 2005 using logistic regression.19 They found that compared with those who lived greater than 20 km from a power plant, those living near ≥2 CFPPs had increased odds of PTB. Yang and Chou found that the closure of a CFPP in Pennsylvania reduced the likelihood of PTB in New Jersey zip codes downwind of the plant using a DiD study design.23 Although each study, including ours, is not without limitations, the collective results appear generally consistent regardless of differences in exposure assessment (i.e., distance, downwind, modeled air emissions), analytical methods (logistic regression, DiD, Poisson generalized additive mixed models), and geography (California, Florida, New Jersey, NC, eastern US).

The DiD study design is susceptible to confounding by time-varying factors, such as environmental, climate, or economic stressors that could impact PTB. Inability to control for these time-varying covariates may lead to residual confounding. Given that most of the well-established risk factors for PTB are time-invariant (thus, controlled for by design) as well as the relatively short timespan of the study, confounding by time-varying factors may be minimal. However, we discuss some potential time-varying factors not accounted for in our analysis.

The DiD study design controls for time-invariant factors; however, if there were non-static co-exposures during pre- and postinterventions this could lead to unmeasured confounding. At the end of 2015, the CFPPs within the CFPP-scrubber group were still operating as CFPPs, whereas four of the seven CFPPs that had retired their coal EGUs transitioned to operating as natural gas-fired power plants (with natural gas MW capacity ranging from 620 to 920 MW). In addition, all 14 of these CFPPs have a legacy of on-site waste (e.g., coal ash); 89% of the coal ash surface impoundments in NC are unlined,57 increasing the likelihood of constituents leaking out of the containment areas into the groundwater system. Recent evidence suggests these surface impoundments are leaking into surrounding groundwater,58 a primary source for domestic water for many NC residents.59 Additionally, a growing body of NC ecological literature shows bioaccumulation of coal ash constituents in fish or death of fish populations inhabiting nearby water systems to CFPPs.3,60–63

At the 14 large NC CFPPs within this study, the interventions to reduce SO2 decreased the CFPP air emissions. Given that SO2 is a criteria air pollutant regulated under the Clean Air Act, we have confidence that the reported SO2 emissions that we accessed to validate our temporal exposure assignments are reasonably accurate. Relatively high ambient SO2 concentrations were typically observed at monitoring locations near anthropogenic sources, such as coal EGUs.4 Although the interventions decreased CFPP air pollution, it is possible that other co-exposures, such as water contamination from coal ash stored on-site, began to increase during this study period. We were unable to control for this type of time-varying point-source co-exposure, which may confound our results—possibly resulting in the observed increase, rather than expected decrease in PTB within the 0- to <4-mile buffer.

Additionally, there may be live birth bias occurring more strongly in the preintervention time period for the 0- to <4-mile buffer than the postintervention time period, which would likely result in underestimation of adverse outcomes.33 Potentially, the highest exposures occurred closest to the CFPPs in the preintervention time period. This might have led to an underestimation of PTB within that potentially highly exposed time period because pregnancy losses that might have been more likely to result in a live preterm birth than live term birth in a less exposed counterfactual scenario were not included in this study selection of live births with reported clinically estimated gestational age of at least 20 weeks. Studies have shown associations between ambient air pollutant exposure, including CFPP-related air emissions, and increased spontaneous abortion and stillbirth.64–68

Alternatively, one could hypothesize that the dispersion of the SO2 plume that exited the coal EGU stacks may have not returned to ground level until after 4 miles given the tall height of the stacks. Thus, those living closest to the CFPPs may not have been as highly exposed to the point-source SO2 air emissions compared with those living farther away. (See supplemental Material, Table S4;, for the distance between gestational parent residential address at time of delivery and nearest CFPP by race/ethnicity for those living within less than 4 miles of the nearest CFPP.) However, we relied on residential distance as the exposure and did not include ambient SO2 concentrations (or ambient concentrations of other air pollutants) in these DiD analyses.

Another source of potential time-varying point-source co-exposure (i.e., interference) may come from other large CFPPs located near the border of NC. To confirm that there were no nearby CFPPs that had a change in electricity generation or reported SO2 emissions across the study period, we identified CFPPs in surrounding states (i.e., Virginia, Tennessee, Georgia, South Carolina) with capacity >135 MW operating during the study period from 2003 to 2015 within 60 miles of the NC border. For any CFPP identified, if electricity generation and reported SO2 emissions were stable across the pre-/postintervention periods within our study then the impact of the outside CFPP is controlled for by design. We found 11 CFPPs outside of NC but within 60 miles of the border. However, nine were greater than 60 miles from any of the 14 NC CFPPs included in the study, which was far enough to not be of great concern for potentially confounding our results. One CFPP was approximately 22 miles from the NC border and 25 miles from the closest of the 14 NC CFPPs included in the study. However, this power plant remained operating as a CFPP with stable reported annual SO2 emissions across the study period (2003–2015). One CFPP (140 MW) was located within approximately 4 miles of the NC border, but approximately 20 miles from the closest NC CFPP included in our study. During the study period, this CFPP remained operating with a slight reduction in reported electricity generation and SO2 annual emission after 2008. However, we expect this CFPP to have little impact on our main results because of the minimal change in electricity generation and emissions as well as the far distance from most of the CFPPs included in the study. Li and Gibson found that CFPP SO2 emissions decreased at a significantly greater rate in NC from 2002 to 2012 than in the four states neighboring NC as well as across all southeastern states.11

We defined the distance from the CFPPs by the point coordinates (X, Y) of the gestational parent residential address at time of delivery to the point coordinates (X, Y) of the CFPP provided by the US EPA AMPD. The spatial boundaries of the CFPPs extend beyond their point coordinates, and thus, into the 0- to <4-mile buffers. For both CFPP-scrubber and CFPP-retired groups, the closest births ranged from approximately 0.4 to 1.3 miles from the X, Y coordinates of the closest power plant. Additionally, six of the 14 CFPPs (two from CFPP-scrubber and four from CFPP-retired) are within 15 miles of the NC border. Because we only include NC resident births within this study, the precision of our estimates are likely impacted due to the smaller eligible spatial area for births surrounding those six CFPPs. We used the residential address at time of delivery to assign gestational exposure. Residential mobility during pregnancy has been documented to range from 9% to 40%,69–73 although, is generally within short distances,70,72 and if mobility occurs, the residential air pollution exposure misclassification has little influence on the effect estimates.71,74

Space and time defined our exposure—the intervention. The timing of the interventions varied among CFPPs, which may have dulled the temporal contrast of the pre and post time periods for the exposure assignments. For the CFPP-scrubber sites, the overall initial-to-final unit intervention across plants spanned 60 months from November 2005 to October 2010 with the unit installments within plants taking 1–21 months (average: 6.6 months, median: 4 months). For CFPP-retired sites, the retirements across plants spanned 31 months from May 2011 to October 2013 with unit retirements within plants taking 1 to 23 months (average: 4.1 months, median: 1 month). This extended time between the pre- and postintervention exposure periods increases the chance of time-varying factors beyond the static conditions assumed by the DiD study design.

For the majority of the CFPP-scrubber sites, we observed a clear distinction of elevated SO2 emissions before the first intervention and a sharp reduction in SO2 emissions after the last intervention. For most of the CFPP-retired sites, SO2 emissions ramped down prior to the official retirement date of the coal EGU units, although the SO2 emissions immediately after the final coal unit retirement went to near zero as expected (Supplemental Material, Figure S3; Given that all the CFPP-scrubber sites and many of the CFPP-retired sites remained operating as power plants (the CFPP-scrubber sites as CFPPs and the CFPP-retired sites as natural gas-fired power plants), it is unlikely that any shift in demographics surrounding the plants were in response to the intervention. We would expect minimal changes in activity (e.g., workforce, traffic) at the power plants that remained operational after the implementation of SO2 emission reduction strategies. And for the three CFPPs that retired all EGUs at the site upon the retirement of the coal EGUs, subsequent demolition, and land restoration processes took place for at least a couple years following the site retirement date, which extends past the postintervention periods for those sites.

We assumed that the comparison group (10–<15 miles) was unaffected (i.e., unexposed) by the CFPP both before and after interventions. This is a strong assumption that may not be true given that air emissions from CFPPs have been documented to travel over 30 miles.75 Although we believe residents at this 10- to <15-mile distance are less impacted by the CFPPs than those living less than 10 miles from the CFPPs, the resulting DiD estimates are likely underestimates of the true effect. Additionally, the two types of intervention are different. Although these interventions may be similar in terms of reducing SO2, other pollutants may also decline with the retirement intervention while may not be affected by the scrubber intervention. Consequently, the DiD estimates may reflect net impacts of retirement or use of scrubbers beyond specifically a reduction in SO2 air emissions. Regional geography may be an important factor when characterizing exposure from CFPPs because the type of coal (lignite, subbituminous, bituminous, anthracite) burned at CFPPs may differ by state/region, which would impact the emissions profile from combustion. However, this was not a concern for comparison among the 14 NC CFPPs included in our study.

For the other dimension of our exposure (i.e., space), we defined proximity using Euclidean distance from a CFPP. This does not take into account wind direction, topography, ground and surface water flow, and the multitude of other factors potentially important for fate and transport of contaminants through air, surface water, and ground water from point sources. Henneman et al developed the HyADS model that estimates ZIP code-level exposure to emissions from CFPPs in the United States that accounts for changes in emissions across the United States as well as metrological factors that impact dispersion of CFPP emissions.75 The HyADS model has recently been used for exposure assignment in studies investigating exposure to CFPP emissions after 2005 and PTB across the United States, asthma in Louisville, Kentucky, and a range of health outcomes from Medicare beneficiaries across the United States (e.g., mortality and hospitalization outcomes, cardiac and respiratory outcomes).12,18,76,77 The use of the HyADS model was outside the scope of this study because (1) exposure assignments for 1-year LMP preintervention exposure periods begin as early as 2003—HyADS begins in 2005 and (2) the spatial assignment at the ZIP code level does not allow for sufficient variation in exposure assignment for the rural CFPPs that are surrounded by few, large ZIP code boundaries. However, the HyADS model is a valuable metric for developing future CFPP-related studies, especially for the evaluation of citywide or statewide health impacts from changes in CFPPs emissions within and outside of NC.

Similar to the investigation of oil and gas drilling infrastructure and adverse birth outcomes,78–80 proximity may serve as a surrogate for an aggregated “myriad of physical, chemical, and social stressors” that are often too complex to fully and easily characterize.78 Because the DiD approach holds the distances constant, these factors only cause concern if time-varying. However, the assigned distance buffers for the CFPP-scrubber and CFPP-retired cohorts are likely differentially influenced by these factors at the plant level, which could mean we did not optimize the distance buffers for plant-specific factors.

The spatial density of these 14 CFPPs in NC also likely reduced the exposure contrast between the exposed and comparison groups (Supplemental Material, Figure S2; Excluding the two CFPP-scrubber plants that we merged together for the analyses due to their very close proximity to one another, five of the CFPPs (three within CFPP-scrubber and two within CFPP-retired) are located within 30 miles of at least one of the other large CFPPs included in this study. Consequently, births could be within 0-15 miles of one CFPP while also being within 0–15 miles of another CFPP meaning that some births assigned to the comparison group (10–<15 miles) of one CFPP may also fall within 15 miles of another CFPP. Although this impacts the exposure contrast, given the DiD design, as long as the emissions at the other nearby CFPP were static during the pre-to-post intervention periods of the CFPP of interest then the spatial density of the power plants would not introduce uncontrolled co-exposure confounding. Our sensitivity analysis of removing observations from the comparison group at each CFPP that were within 15 miles of one of the other large CFPPs evaluated in this study were consistent with our main analyses (Supplemental Material, Table S2; For example, for our main and sensitivity analysis, the absolute prevalence of PTB decreased [–1.5% (95% CI: –2.6, –0.4) and –1.7% (95% CI: –2.9, –0.4), respectively] with the installation of scrubbers for gestational parents within 4-<10 miles from a CFPP compared with those within 10–<15 miles. Future CFPP-related studies in NC—and generally the Southeast—could consider novel methods, such as bipartite causal inference with interference, to help account for potential interconnectedness of units and interference between observations.81

The meta-analysis results of the individual CFPPs were coherent with our main pooled analysis results (Supplemental Material, Table S2; Although the fixed-effect and random-effects meta-analyses involve different underlying assumptions, in this study, the main analysis (pooled) and sensitivity analysis (meta-analyses) estimates were all very close. Our pooled DiD estimates were likely heavily influenced by one plant in CFPP-scrubber cohort and one plant in CFPP-retired cohort that accounted for approximately half of the births within each cohort. Small sample sizes at many of the CFPPs limited our ability to investigate the association between the intervention and PTB at the individual CFPP level (Supplemental Material, Table S3 and Figure S6; However, future studies should consider ways to navigate that limitation to conduct CFPP-specific evaluations since demographics surrounding CFPPs are not homogenous across plants. For example, in our stratified analyses of race/ethnicity, we were unable to present results for non-Hispanic American Indian gestational parents within the CFPP-scrubber, CFPP-retired, and CFPP-all cohorts or non-Hispanic Asian gestational parents within the CFPP-retired and CFPP-all cohorts due to non-convergence of models. However, stratification by race/ethnicity could be explored in more depth at some individual CFPPs that have higher percentages of residents of color compared with the aggregate percentages nearby all CFPPs or the state. For example, a majority of the non-Hispanic American Indian births included within the CFPP-all cohort (79%) and CFPP-retired cohort (88%) are nearby a single CFPP included in the study.

To define PTB, we used the clinically estimated gestational age provided on the birth certificate, rather than estimate gestational age based on date of LMP. Because clinically estimated gestational age is a composite measure, it may be inaccurate in ways difficult to assess. However, we used clinically estimated gestational age because it does not rely on gestational parent recall, is not affected by menstrual variability and irregularities, and incorporates ultrasound dating.82 We were unable to consider differences by the indication of PTB (i.e., spontaneous or medically induced). Although the underlying biological mechanisms for the type of PTB likely differ, spontaneous PTB accounts for approximately 80% of PTB.83 Our understanding of reproductive and developmental biological mechanisms underlying the potential for CFPP emissions to impact PTB, as well as the critical windows of exposure, is limited. However, there are multiple hypotheses that support various critical windows across pregnancy as well as short- versus long-term exposures. For example, circulating sulfite may result in systemic changes in redox signaling and oxidative stress that may adversely affect the embryo during development, which may lead to PTB or reduction of birth weight.4,15 An integrative review of 30 studies of oxidative stress and PTB found that, in general, PTB specimens (e.g., gestational parent venous blood, placental/membrane tissue, umbilical cord) had higher levels of indicators for oxidative stress (e.g., reactive oxygen species, reactive nitrogen species, or biomarkers of oxidative stress) compared with term birth specimens.83 The increase in oxidative stress suggests that changes to the redox environment influence PTB, but the mechanistic pathways remain unclear.83 Ambient air pollutants have been associated with systemic inflammation and shown to cause proinflammatory responses in toxicological studies.14 The exposure assignment in our study captures the entire pregnancy of all included observations from LMP to live birth within either the preintervention or postintervention time periods; hence, we are not making specific assumptions regarding critical windows of gestational exposure to ambient air pollution and PTB.

NC serves as a unique and important study site to investigate the impact of CFPP-related air pollution on birth outcomes. NC has a historical dependence on CFPPs for electricity generation.7 Additionally, because coal mining and processing does not occur in NC, as it does in many other southeastern states that are similarly dependent on coal for electricity generation, we were unconcerned about co-exposure confounding from other coal-related precombustion processes. Furthermore, as of 2015, NC was one of 16 nonproducing states for crude oil and natural gas. An increasing number of studies show positive associations between unconventional oil and gas development and PTB.80 The lack of oil and gas activity within NC alleviates concerns of time-varying co-exposure confounding by this rapidly expanding method for energy production.

CFPPs are complicated systems with many point-source emissions. Since 2005, seven large CFPPs in NC drastically reduced SO2 emissions by installing desulfurization equipment on their operating coal EGUs and another seven retired their on-site coal EGUs. We leveraged these interventions due to an enacted policy that required CFPPs within the state to substantially reduce their SO2 emissions by two benchmarks in 2009 and 2013 to investigate the association between living near a CFPP and PTB. Our results suggest that living near CFPPs that installed scrubbers to reduce SO2 air emissions resulted in a reduction of PTB among residents 4-<10 miles from CFPPs compared with those living 10–<15 miles from CFPPs. Our study adds to the growing body of literature investigating associations between living near power plants and adverse birth outcomes.15,17–24 Even after these efforts to reduce SO2 emissions, CFPPs in NC remain the dominant anthropogenic source of ambient SO2.4,7 Research investigating impacts of air pollution on health, particularly for reproductive, perinatal, and pediatric outcomes, can help inform decision makers.

Conflicts of interest statement

The authors declare that they have no conflicts of interest with regard to the content of this report.


The authors would like the thank Ms. Nina Forestieri for responding to questions about the NC resident live birth records and Dr. Monica Jimenez for geocoding assistance. We acknowledge Drs. Alison Krajewski and Kristen Rappazzo for input on earlier versions of the manuscript as well as thank the journal peer-reviewers for their thoughtful feedback and suggestions.


1. Bi X. “Cleansing the air at the expense of waterways?” Empirical evidence from the toxic releases of coal-fired power plants in the United States. J Regul Econ. 2017;51:18–40.
2. Massetti EB. Environmental Quality and the U.S. Power Sector: Air Quality, Water Quality, Land Use and Environmental Justice. Oak Ridge National Lab. (ORNL), Oak Ridge, TN (United States). Building Technologies Research and Integration Center (BTRIC); 2017.
3. Wang Z, Cowan EA, Seramur KC, et al. Legacy of coal combustion: widespread contamination of lake sediments and implications for chronic risks to aquatic ecosystems. Environ Sci Technol. 2022;56:14723–14733.
4. U.S. EPA. Integrated Science Assessment (ISA) for Sulfur Oxides – Health Criteria (Final Report, Dec 2017). U.S. Environmental Protection Agency, Washington, DC, EPA/600/R-17/451, 2017. Available at: Accessed 26 December 2022.
5. U.S. EPA. Integrated Science Assessment (ISA) for Particulate Matter (Final Report, Dec 2019). U.S. Environmental Protection Agency, Washington, DC, EPA/600/R-19/188, 2019. Available at: Accessed 26 December 2022.
6. Wikle NB, Hanks EM, Henneman LRF, Zigler CM. A mechanistic model of annual sulfate concentrations in the United States. J Am Stat Assoc. 2022;117:1082–1093.
7. Wilkie AA, Richardson DB, Luben TJ, Serre ML, Woods CG, Daniels JL. North Carolina’s changing energy generation profile and reductions in key air pollutants, 2000-2019. N C Med J. 2022;83:304–310.
8. NCDEQ. NC DEQ: Clean Smokestacks Act (CSA). Available at: Accessed 23 November 2020.
9. PowerEngineering. Progress Energy to retire four coal-fired plants by 2017. Power Engineering. Available at: 2009. Accessed 23 November 2020.
10. Kravchenko J, Lyerly HK, Ross W. The health impacts of environmental policy: the North Carolina clean smokestacks act. N C Med J. 2018;79:329–333.
11. Li YR, Gibson JM. Health and air quality benefits of policies to reduce coal-fired power plant emissions: a case study in North Carolina. Environ Sci Technol. 2014;48:10019–10027.
12. Kim C, Henneman LRF, Choirat C, Zigler CM. Health effects of power plant emissions through ambient air quality. J R Stat Soc Ser A Stat Soc. 2020;183:1677–1703.
13. Leem JH, Kaplan BM, Shim YK, et al. Exposures to air pollutants during pregnancy and preterm delivery. Environ Health Perspect. 2006;114:905–910.
14. Vadillo-Ortega F, Osornio-Vargas A, Buxton MA, et al. Air pollution, inflammation and preterm birth: a potential mechanistic link. Med Hypotheses. 2014;82:219–224.
15. Mohorovic L. First two months of pregnancy—critical time for preterm delivery and low birthweight caused by adverse effects of coal combustion toxics. Early Hum Dev. 2004;80:115–123.
16. Jacobs M, Zhang G, Chen S, et al. The association between ambient air pollution and selected adverse pregnancy outcomes in China: a systematic review. Sci Total Environ. 2017;579:1179–1192.
17. Casey JA, Karasek D, Ogburn EL, et al. Retirements of coal and oil power plants in California: association with reduced preterm birth among populations nearby. Am J Epidemiol. 2018;187:1586–1594.
18. Daouda M, Henneman L, Kioumourtzoglou MA, Gemmill A, Zigler C, Casey JA. Association between county-level coal-fired power plant pollution and racial disparities in preterm births from 2000 to 2018. Environ Res Lett. 2021;16:034055.
19. Ha S, Hu H, Roth J, Kan H, Xu X. Associations between residential proximity to power plants and adverse birth outcomes. Am J Epidemiol. 2015;182:215–224.
20. Luechinger S. Air pollution and infant mortality: a natural experiment from power plant desulfurization. J Health Econ. 2014;37:219–231.
21. Severnini E. Impacts of nuclear plant shutdown on coal-fired power generation and infant health in the Tennessee Valley in the 1980s. Nat Energy. 2017;2:1–9. doi:10.1038/nenergy.2017.51.
22. Yang M, Bhatta RA, Chou SY, Hsieh CI. The impact of prenatal exposure to power plant emissions on birth weight: evidence from a Pennsylvania power plant located upwind of New Jersey. J Pol Anal Manag. 2017;36:557–583.
23. Yang MZ, Chou SY. The impact of environmental regulation on fetal health: evidence from the shutdown of a coal-fired power plant located upwind of New Jersey. J Environ Econ Manage. 2018;90:269–293.
24. Keil AP, Buckley JP, Kalkbrenner AE. Bayesian G-computation for estimating impacts of interventions on exposure mixtures: demonstration with metals from coal-fired power plants and birth weight. Am J Epidemiol. 2021;190:2647–2657.
25. USEIA. U.S. Energy Information Administration (EIA). Electric Power Monthly - March 2004. DOE/EIA-0226. (2004/03). Available at: Accessed 20 December 2022.
26. MarchOfDimes. Perinatal Data Snapshots: North Carolina. 2015. Available at: https://Www.Marchofdimes.Org/Peristats/Reports/North-Carolina/Report-Card. Accessed 26 December 2022.
27. CDC. Stats of the States—Preterm Births. Available at: 2022. Accessed December 20, 2022.
28. Evans S. Statistical Brief, No. 41 (August 2014) -- Preterm Birth Rate Trends in North Carolina 1988-2012. 2014. Available at: Accessed 26 December 2022.
29. UNICEF. Low Birthweight: Country, Regional, and Global Estimates. United Nations Children’s Fund and World Health Organization. 2004. Available at: https://Apps.Who.Int/Iris/Handle/10665/43184. Accessed 26 December 2022.
30. Wilcox AJ. Fertility and Pregnancy: An Epidemiologic Perspective. Oxford University Press; 2010.
31. Miranda ML, Maxson P, Edwards S. Environmental contributions to disparities in pregnancy outcomes. Epidemiol Rev. 2009;31:67–83.
32. NCDHHS. Statistics and Reports: Minority and Health. Available at: Accessed 26 December 2022.
33. Neophytou AM, Kioumourtzoglou MA, Goin DE, Darwin KC, Casey JA. Educational note: addressing special cases of bias that frequently occur in perinatal epidemiology. Int J Epidemiol. 2021;50:337–345. doi:10.1093/ije/dyaa252.
34. Strand LB, Barnett AG, Tong S. Methodological challenges when estimating the effects of season and seasonal exposures on birth outcomes. BMC Med Res Methodol. 2011;11:49.
35. Willis MD, Hill EL, Boslett A, Kile ML, Carozza SE, Hystad P. Associations between residential proximity to oil and gas drilling and term birth weight and small-for-gestational-age infants in Texas: a difference-in-differences analysis. Environ Health Perspect. 2021;129:77002.
36. Wing C, Simon K, Bello-Gomez RA. Designing difference in difference studies: best practices for public health policy research. Annu Rev Public Health. 2018;39:453–469.
37. Rioux C, Weedon S, London-Nadeau K, et al. Gender-inclusive writing for epidemiological research on pregnancy. J Epidemiol Community Health. 2022;76:823–827.
38. Martinez RAM, Andrabi N, Goodwin AN, Wilbur RE, Smith NR, Zivich PN. Conceptualization, operationalization, and utilization of race and ethnicity in major epidemiology journals, 1995–2018: A systematic review. Am J Epidemiol. 2022:kwac146. doi:10.1093/aje/kwac146.
39. CDC. U.S. Census Populations With Bridged Race Categories. Available at: 2020. Accessed 17 September 2021.
40. Grabich SC, Robinson WR, Engel SM, Konrad CE, Richardson DB, Horney JA. County-level hurricane exposure and birth rates: application of difference-in-differences analysis for confounding control. Emerg Themes Epidemiol. 2015;12:19.
41. Bullard RD, Mohai P, Saha R, Wright B. Toxic Wastes and Race at Twenty, 1987-2007: A Report Prepared for the United Church of Christ Justice and Witness Ministries. 2007.
42. Mikati I, Benson AF, Luben TJ, Sacks JD, Richmond-Bryant J. Disparities in distribution of particulate matter emission sources by race and poverty status. Am J Public Health. 2018;108:480–485.
43. Mohai P, Lantz PM, Morenoff J, House JS, Mero RP. Racial and socioeconomic disparities in residential proximity to polluting industrial facilities: evidence from the Americans’ Changing Lives Study. Am J Public Health. 2009;99(Suppl 3):S649–S656.
44. Payne-Sturges DC, Gee GC, Cory-Slechta DA. Confronting racism in environmental health sciences: moving the science forward for eliminating racial inequities. Environ Health Perspect. 2021;129:55002.
45. Almeida J, Becares L, Erbetta K, Bettegowda VR, Ahluwalia IB. Racial/ethnic inequities in low birth weight and preterm birth: the role of multiple forms of stress. Matern Child Health J. 2018;22:1154–1163.
46. Manuck TA. Racial and ethnic differences in preterm birth: a complex, multifactorial problem. Semin Perinatol. 2017;41:511–518.
47. Purisch SE, Gyamfi-Bannerman C. Epidemiology of preterm birth. Semin Perinatol. 2017;41:387–391.
48. Ratnasiri AWG, Parry SS, Arief VN, et al. Recent trends, risk factors, and disparities in low birth weight in California, 2005-2014: a retrospective study. Matern Health Neonatol Perinatol. 2018;4:15.
49. Bower KM, Geller RJ, Perrin NA, Alhusen J. Experiences of racism and preterm birth: findings from a pregnancy risk assessment monitoring system, 2004 through 2012. Women’s Health Issues. 2018;28:495–501.
50. Braveman P, Heck K, Egerter S, et al. Worry about racial discrimination: a missing piece of the puzzle of Black-White disparities in preterm birth? PLoS One. 2017;12:e0186151.
51. Chae DH, Clouston S, Martz CD, et al. Area racism and birth outcomes among blacks in the United States. Soc Sci Med (1982). 2018;199:49–55.
52. Chambers BD, Erausquin JT, Tanner AE, Nichols TR, Brown-Jeffy S. Testing the association between traditional and novel indicators of county-level structural racism and birth outcomes among black and white women. J Racial Ethn Health Disparities. 2018;5:966–977.
53. Braveman P, Dominguez TP, Burke W, et al. Explaining the black-white disparity in preterm birth: a consensus statement from a multi-disciplinary scientific work group convened by the march of dimes. Front Reprod Health. 2021;3:684207. Available at: Accessed 17 December 2022.
54. Adkins-Jackson PB, Chantarat T, Bailey ZD, Ponce NA. Measuring Structural Racism: a guide for epidemiologists and other health researchers. Am J Epidemiol. 2021;191:539–547.
55. Miranda ML, Anthopolos R, Edwards SE. Seasonality of poor pregnancy outcomes in North Carolina. N C Med J. 2011;72:447–453.
56. Vinikoor LC, Messer LC, Laraia BA, Kaufman JS. Reliability of variables on the North Carolina birth certificate: a comparison with directly queried values from a cohort study. Paediatr Perinat Epidemiol. 2010;24:102–112.
57. EarthJustice, WaterkeeperAlliance. North Carolina: Coal Ash Disposal, Damage and Regulation. 2012. Available at: https://Earthjustice.Org/Sites/Default/Files/Nc-Coal-Ash-Factsheet-1112.Pdf. Accessed 26 December 2022.
58. Harkness JS, Sulkin B, Vengosh A. Evidence for coal ash ponds leaking in the Southeastern United States. Environ Sci Technol. 2016;50:6583–6592.
59. MacDonald Gibson J, Pieper KJ. Strategies to improve private-well water quality: a North Carolina perspective. Environ Health Perspect. 2017;125:076001.
60. Besser JM, Giesy JP, Brown RW, Buell JM, Dawson GA. Selenium bioaccumulation and hazards in a fish community affected by coal fly ash effluent. Ecotoxicol Environ Saf. 1996;35:7–15.
61. Brandt JE, Simonin M, Di Giulio RT, Bernhardt ES. Beyond selenium: coal combustion residuals lead to multielement enrichment in receiving lake food webs. Environ Sci Technol. 2019;53:4119–4127.
62. Vengosh A, Cowan EA, Coyte RM, et al. Evidence for unmonitored coal ash spills in Sutton Lake, North Carolina: implications for contamination of lake ecosystems. Sci Total Environ. 2019;686:1090–1103.
63. SoutheastCoalAsh. Map View. Available at: Accessed 26 December 2022.
64. Casey JA, Gemmill A, Karasek D, Ogburn EL, Goin DE, Morello-Frosch R. Increase in fertility following coal and oil power plant retirements in California. Environ Health. 2018;17:44.
65. Conforti A, Mascia M, Cioffi G, et al. Air pollution and female fertility: a systematic review of literature. Reprod Biol Endocrinol. 2018;16:117.
66. Nyadanu SD, Dunne J, Tessema GA, et al. Prenatal exposure to ambient air pollution and adverse birth outcomes: an umbrella review of 36 systematic reviews and meta-analyses. Environ Pollut. 2022;306:119465.
67. Grippo A, Zhang J, Chu L, et al. Air pollution exposure during pregnancy and spontaneous abortion and stillbirth. Rev Environ Health. 2018;33:247–264.
68. Mohorovic L, Petrovic O, Haller H, Micovic V. Pregnancy loss and maternal methemoglobin levels: an indirect explanation of the association of environmental toxics and their adverse effects on the mother and the fetus. Int J Environ Res Public Health. 2010;7:4203–4212.
69. Saadeh FB, Clark MA, Rogers ML, et al. Pregnant and moving: understanding residential mobility during pregnancy and in the first year of life using a prospective birth cohort. Matern Child Health J. 2013;17:330–343.
70. Lupo PJ, Symanski E, Chan W, et al. Differences in exposure assignment between conception and delivery: the impact of maternal mobility. Paediatr Perinat Epidemiol. 2010;24:200–208.
71. Miller A, Siffel C, Correa A. Residential mobility during pregnancy: patterns and correlates. Matern Child Health J. 2010;14:625–634.
72. Bell ML, Belanger K. Review of research on residential mobility during pregnancy: consequences for assessment of prenatal environmental exposures. J Expo Sci Environ Epidemiol. 2012;22:429–438.
73. Fell DB, Dodds L, King WD. Residential mobility during pregnancy. Paediatr Perinat Epidemiol. 2004;18:408–414.
74. Chen L, Bell EM, Caton AR, Druschel CM, Lin S. Residential mobility during pregnancy and the potential for ambient air pollution exposure misclassification. Environ Res. 2010;110:162–168.
75. Henneman LRF, Choirat C, Ivey C, Cummiskey K, Zigler CM. Characterizing population exposure to coal emissions sources in the United States using the HyADS model. Atmos Environ. 2019;203:271–280.
76. Casey JA, Su JG, Henneman LRF, et al. Improved asthma outcomes observed in the vicinity of coal power plant retirement, retrofit, and conversion to natural gas. Nat Energy. 2020;5:398–408.
77. Henneman LR, Choirat C, Zigler CM. Accountability assessment of health improvements in the United States associated with reduced coal emissions between 2005 and 2012. Epidemiology. 2019;30:477–485.
78. Deziel NC. Invited perspective: oil and gas development and adverse birth outcomes: what more do we need to know? Environ Health Perspect. 2021;129:071301.
79. Willis MD, Hill EL, Boslett A, Kile ML, Carozza SE, Hystad P. Associations between residential proximity to oil and gas drilling and term birth weight and small-for-gestational-age infants in Texas: a difference-in-differences analysis. Environ Health Perspect. 2021;129:077002.
80. Deziel NC, Brokovich E, Grotto I, et al. Unconventional oil and gas development and health outcomes: a scoping review of the epidemiological research. Environ Res. 2020;182:109124.
81. Zigler CM, Papadogeorgou G. Bipartite causal inference with interference. Stat Sci. 2021;36:109–123.
82. Rappazzo KM, Lobdell DT, Messer LC, Poole C, Daniels JL. Comparison of gestational dating methods and implications for exposure-outcome associations: an example with PM2.5 and preterm birth. Occup Environ Med. 2017;74:138–143.
83. Moore TA, Ahmad IM, Zimmerman MC. Oxidative stress and preterm birth: an integrative review. Biol Res Nurs. 2018;20:497–512.

Supplemental Digital Content

Copyright © 2023 The Authors. Published by Wolters Kluwer Health, Inc. on behalf of The Environmental Epidemiology. All rights reserved.