A National Study of the Associations between Hormonal Modulators in Hydraulic Fracturing Fluid Chemicals and Birth Outcomes in the United States of America: A County-Level Analysis

Hydraulic fracturing (HF), more commonly known as fracking, is the injection of water, sand, and chemicals into the ground at high pressures to break up rock and release oil and natural gas. The practice became popular in the United States in the early 2000s, and since then numerous peer-reviewed scientific studies have found that HF is associated with negative health outcomes for those living in the proximity. Associated conditions include preterm birth (PTB), low birth weight (LBW), and congenital heart defects, although these associations have not been seen consistently across observational studies.^1–8 For example, a study conducted in southwest Pennsylvania found an association between HF and LBW but not PTB, whereas a study in northeast Pennsylvania found the opposite: an association between HF and PTB but not LBW.^3,8 A Colorado study found an association between HF and congenital heart defects but no relationship with PTB or LBW.⁴ A North Texas study found an association between HF and PTB but a minimal association with LBW.⁵ Another Pennsylvania study and an Oklahoma study found associations between HF and LBW but did not address PTB.^6,7 Although the exact nature of the relationships between HF and perinatal outcomes are not clear, the research supports the existence of a relationship.

HF can cause water contamination; air, noise, and light pollution; changes in landscape; and seismicity.⁹ HF chemicals have been found in drinking water supplies near HF sites, and some chemicals used in HF are known to affect human development and reproduction processes.^1,4,10,11 More recent iterations of HF, notably enhanced geothermal, or HF to release or create geothermal energy, purport to use more environmentally and human health-friendly chemicals in injection fluid.^12,13 These evolving HF practices demonstrate a growing interest in the ingredients used in HF and their impacts on human health.

Prior studies investigated the associations between HF and poor birth outcomes on a patient-level and calculated HF exposure using the distance between the mother’s residence and HF wells^6,7 or inverse distance squared counts of wells within a given distance (usually $<10\text{\hspace{0.17em}miles}$ or $<16.09\mathrm{kilometers}$) from the mother’s residence,^3–5,8 sometimes including information about the phase of well development and production volume during pregnancy.³ Koehler et al. explored the feasibility and utility of including additional information about the HF process in exposure analysis, including compressor engines, impoundments, and flaring events associated with HF. The authors found that these events are difficult to measure and tend to co-occur in time and space with HF wells, concluding that HF activity can be adequately measured without taking the resources to measure each component.¹⁴

Most HF occurs in rural areas, where land is cheaper and more available.¹⁵ It is especially common in agricultural areas, with many HF operations built on land leased from farmers.¹⁶ Health disparities exist between rural areas and urban areas in the United States. Rural Americans have higher rates of heart disease, cancer, chronic respiratory disease, and strokes than urban Americans do.¹⁷ Proposed explanations for these disparities include less access to health care, exposure to environmental hazards, higher poverty rates, lower rates of health insurance enrollment, and less access to health information.^17–19 Disparities in perinatal outcomes have also been found between rural and urban areas. Rural counties not bordering metropolitan areas and lacking hospital-based obstetrics services see higher rates of PTB and LBW.^20–24 Pesticide exposure from farming operations has been associated with higher rates of PTB and LBW.^25–27 A prior study of HF impacts on PTB and LBW found contradictory results between rural and urban areas, with HF having a positive association with PTB and LBW in rural areas and a negative association in urban areas.²⁸ This paper explores the impact of exposure to HF wells and their chemical ingredients on county level rates of PTB and LBW across the country. We control for rural–urban health disparities by creating and analyzing subsets of rural counties and adjusting all of our models for access to maternal care, rates of insurance enrollment, rates of poverty, the proportion of the county that is agricultural land, and pesticide use within the county.

There are many endocrine disruptors used in HF fluid ingredients and some prior studies have pointed to the potential consequences on human health.²⁹ However, endocrine disruption following HF exposure is complex to study in humans (and hence the motivation for our work). In fish, endocrine disruption has been observed following exposure to HF fluid.³⁰ Furthermore, the prenatal exposure time window has been proposed as a key developmental time point for potential damaging effects of endocrine disruption in humans.^31–33 Specifically focusing on HF, several studies have linked prenatal exposure to HF chemicals with adverse health outcomes in mice,^34,35 but human studies that tie specific chemicals used in HF to health outcomes appear to be lacking. The paucity of human studies in this space was the motivation for our work: to study the effects of prenatal exposure to endocrine-disrupting chemicals (EDCs) used in HF fluid on human birth outcomes, namely PTB and LBW.

This study sought to further explore the associations between HF and rates of PTB and LBW. Prior research has analyzed these relationships within a US state or region of a state, but in this study we examined them on a national scale, responding to calls for future research to characterize specific chemical stressors associated with HF.³⁶ We considered whether the ingredients used in a county’s HF wells were known to impact estrogen, testosterone, or other hormone pathways (where the term other hormone pathways is defined as hormones that are not testosterone or estrogen) to explore whether ingredients that modulate human hormones impact the associations between HF and adverse perinatal outcomes. We hypothesized that wells with ingredients that impact human hormone pathways would increase PTB and LBW rates more than other HF wells.

This study explores the association between four different exposure groupings, namely, a) exposure to HF wells, b) exposure to wells with chemicals known to target estrogen pathways, c) testosterone pathways, or d) other hormone pathways (defined as hormones that are not testosterone or estrogen) on two outcomes: PTB rate and LBW rate. We used only publicly available resources, and our unit of analysis was at the US county level. An overview of our process is shown in Figure S1 and is described in the “Data Sources” section.

We retrieved county-level data on HF well density (wells per square mile area of each county), rates of PTB and LBW, and several demographic covariates from publicly available data (Figure 1). For the data on HF wells, we used data from 2001 to 2018, with a relevant subset of data from 2014 to 2018 used in our sensitivity analyses. Data on our outcomes (PTB and LBW rate) were available for the years 2015 to 2018, and information on maternal care deserts was available from 2021. The US Census Bureau American Community Survey (ACS) data and Centers for Disease Control and Prevention (CDC) WONDER data were available for the 2014–2018 period, the US Department of Agriculture (USDA) data from 2017, and US Census Bureau’s Gazetteer data from 2018. This information is also available in Table S1. We chose to use publicly available data and aggregate it at the county level so that we could expand our study nationwide, a method similar to other nationwide and larger-scale HF studies.^37,38

We used the previously developed and publicly available 2020 WellExplorer database³⁹ (version 1) to obtain data on HF well locations and the chemical ingredients injected into those wells as part of the fracking process. The information in WellExplorer comes from FracFocus, which is self-reported by the HF companies and is regulated by some states. WellExplorer makes use of the FracFocus public database of HF wells, which already provided the ingredients used in the wells. WellExplorer previously linked this information to the hormonal pathways impacted by each well’s ingredients.³⁹ When designing our models, we considered the density of all wells in each county (defined as wells per square mile area of each county), of wells that use chemicals known to target estrogen, testosterone, or other human hormone pathways. From this data, we were able to obtain information on the gene/protein targets of each chemical used in the HF that was mapped within WellExplorer to some specific compound with known chemical activity.³⁹ These data contain information on the link between chemicals used in HF fluid and the proteins/genes involved in one of the three hormonal pathways studied (i.e., estrogen, testosterone, or other hormone) and this information can be found in Excel Table S1. We also provide the information on whether the chemical pairs were linkable to an open-source database, the Toxin and Toxin-Target Database (T3DB),^40,41 using the May 2018 version, that we used to ascertain hormonal activity, and this information can be found in Excel Table S2. The number of unique protein/genes targeted by chemicals used in HF fluid ingredients sorted by the number of chemicals targeting specific proteins are listed in Excel Table S3.

We developed our integrated and harmonized dataset in R studio using R (version 4.0.2; R Core Development Team) using the dplyr and stringr packages.^55,56 We combined the data from our various sources into one table, matching each county/county-equivalent using its corresponding Federal Information Processing Standards (FIPS) code. We cleaned the combined data by excluding FIPS codes outside the 50 US states and the District of Columbia. The Aleutian Islands, Alaska, was one county-equivalent in the USDA dataset but two in the other datasets: Aleutians East Borough and Aleutians West Borough. We therefore split the data on acres of agricultural land and fertilizer and chemical use in half between the two boroughs. All FIPS codes in our final dataset were in line with the 2018 Census Bureau standard.⁵⁷ In addition, a few counties had fracking wells in bodies of water that are considered part of the county (e.g., a well off the coast of Los Angeles County that is listed as located within the county). There were wells located in Alaska’s Harrison Bay, which borders the North Slope Borough of Alaska in a manner like that of the bodies of water associated with other counties, so we counted these wells as located in the borough. We used the US Census Bureau’s 2018 Gazetteer land area data to normalize relevant covariates by county size, which allowed us to compare across counties of various size but caused us to assume an even distribution of HF wells and other normalized variables across the county.⁵⁴ All covariates were quantitative. We used a linear regression model because our outcome was linear, numeric, and continuous. One variable that was binary was maternal care deserts, and this indicated whether some maternal care was available (i.e., not a maternal care desert) or whether no maternal care was available (i.e., a maternal care desert) in that county.⁴⁹ The exposure variables of well density, as well as the hormonal densities, were numerical variables. Many other important covariates, such as race, were also numerical variables given that the variable was the percentage Black population in a particular county.

We calculated the descriptive statistics for each covariate included to provide context for the results of our analysis (Table 1). Excel Table S4 provides the final dataset, which will be helpful for readers who might be interested in a particular covariate.

We constructed four generalized linear models each for PTB rate and for LBW rate as the outcome for each subset. Each of these models is shown in Figure 2, and their model equations are given in the sections “Models for PTB” and “Models for LBW.” Each model looked at the association between county-level density of HF wells (either the total wells, wells with ingredients that impact estrogen pathways, testosterone pathways, or other hormone pathways) and the same set of adverse birth outcomes, namely PTB or LBW (both defined as the percentage of live births within the county from 2015 to 2018) adjusting for all covariates described above in the section “Covariate Data.” We analyzed each generalized linear model in R studio using R (version 4.0.2; R Core Development Team). We considered the coefficient estimate (the beta estimates from our models), 95% confidence interval (CI) of those estimates, the standard error, and level of statistical significance for each association. Statistical significance was determined using standard thresholds where a $p<0.05$ level was determined to be significant.

We adjusted for the same set of covariates in each of our models shown below in Equations 3–6. The only difference between the four models is the type of exposure variable. Equation 3 explores the all_wells_per_sq_mile variable, which is an indicator of all wells regardless of chemicals used. Equation 4 explores the wells that use chemicals that include estrogen targeting chemicals (ETCs) as our exposure variable. Equation 5 explores the wells that use chemicals that include testosterone targeting chemicals (TTCs) as our exposure variable. Finally, Equation 6 explores the wells that use chemicals that include other hormone targeting chemicals (OHTCs) as our exposure variable. Note that ‘targeting’ involves affecting the respective pathway, therefore can be either direct or indirect. Equations 7–10 mirror Equations 3–6 except that the outcome variable is for LBW rather than PTB.

Ultimately, we had seven stratified sets of data from our original, integrated dataset (Figure 2). These stratified sets were based on a) having HF data from mandated reporting states, b) having HF data from a period relevant to our outcome data (2014–2018), and c) exploring the rural vs. urban divide. The full data consists of all data without any stratified exclusions. Importantly, the same set of models as shown in Equations 3–10 were used in each of these stratified subsets.

In the interest of transparency and to help support as much data sharing as possible on the part of fracking companies and research, we started with a primary full analysis as a model that included as much data as possible, without any type of cherry picking or other data exclusionary criteria. Therefore, our first primary model included all data we had access to that was reported, including older fracking data, and we included any fracking data available. In part this is because there might be a temporal lag between the time when fracking is first conducted in a region and the time when birth outcomes will result from the exposure. For example, it is unknown how long between the exposure of a pregnant person, or a female before pregnancy, is needed for there to be impact on later risk to the fetus. Because of this and the variability of time windows reported in other studies, we decided to set our primary (termed the full analysis) as including all data available. We then constructed several stratified analyses that include different slices of our dataset. The subset of data 2014–2018 wells in states with HF exposure fully captured represents the slice of data and models that investigated overlapping time windows and in regions where HF exposure is known to be fully captured owing to explicit legislation in those states dictating reporting. Results in the abstract and conclusion reflect the results from the 2014–2018 stratified subset given that this subset was most relevant.

We visualized some of our results using maps that were made using the usmap package in R Studio (https://cran.r-project.org/web/packages/usmap/index.html). Alaska and Hawaii are excluded from the maps because of missing information on locations of shale plays and basin from Energy Information Administration maps used to create the figure (https://www.eia.gov/analysis/studies/usshalegas/pdf/usshaleplays.pdf, p6). States with HF information not fully captured are represented with stripes (9 states).

The complete dataset and seven stratified subsets that we created can be seen in Table 2 broken down for their corresponding outcomes (PTB and LBW rates) and also shown in Figure 2. In total there are 16 dataset–outcome combos because each outcome was used for the eight subsets (Figure 1). The states with HF exposure fully captured subset still captures most of the wells and counties in FracFocus, 90% and 74%, respectively (Table 2). The farming and mining subset, on the other hand, includes only a third of counties with HF and two thirds of wells. Similarly, the nonmetro subset contains half of counties with HF and two thirds of wells. The farming and mining and nonmetro subsets are not representative of all HF in the United States, although they may still be useful in teasing out how rural–urban disparities may be confounding the associations that are observed.

Figure 3A shows the HF wells active during 2014–2018 and then broken down by ingredients’ chemical pathways. HF wells with ingredients known to target pathways involving estrogen (Figure 3B) that were active during 2014–2018 and HF wells with ingredients known to target pathways involving testosterone (Figure 3C) that were active during 2014–2018 are depicted against a map of the shale plays and basins in the United States to provide context on where most HF in the United States takes place. A shale play is a discovered or possible natural gas accumulation. Shales are located within basins, or large geographical depressions that may contain oil or natural gas.⁶³

We also include in the Supplemental Material files information on each chemical and its corresponding protein or gene target that has been labeled as estrogen, testosterone, or hormonal pathways. Chemicals that target one of the three pathways are given in Excel Table S1. The chemicals in our study used in HF that were linked to gene/protein targets are listed in Excel Table S2. The 106 distinct gene/protein targets for HF ingredients that target one of the three pathways: estrogen, testosterone, or other hormonal pathways are listed in Excel Table S3. The number of counties with hormonal modulating ingredients is provided in Table S2.

Across all models that we constructed, the density of HF wells was positively associated with average county-level PTB and LBW rates. In each model, the coefficient estimates were highest for wells with ingredients that include ETCs, TTCs, and OHTCs. We also explored all well types grouped.

In the following paragraphs, we report the results from our more restrictive stratified analysis using the 2014–2018 subset, because this analysis resulted in the most meaningful time frame for comparison. In our models exploring the association between HF wells that were active between 2014 and 2018 and PTB rates, wells with ETCs had a beta coefficient estimate of 3.789 (95% CI: 1.83, 5.74). This indicates that for every unit increase in well density per square mile of wells that use chemicals that include an ETC, we observed a 3.789-higher PTB rate (95% CI: 1.83, 5.74) compared with counties with no ETC wells. Similarly, we found that for every unit increase in well density per square mile of wells that use TTC, we observed a 3.192-higher PTB rate (95% CI: 1.62, 4.77) compared with counties with no TTC wells. We found that an increase in well density per square mile among wells that use chemicals that include an OHTC resulted in a 2.276-higher PTB rate (95% CI: 1.25, 3.30) compared with counties with no OHTC wells. We also explored the role of HF well exposure in general (regardless of the chemicals used) and found that an increase in total well density (grouped regardless of hormonal targeting status of the chemicals used) resulted in a 1.228-higher PTB rate (95% CI: 0.66, 1.80) compared with counties with no wells.

Similarly, for the equivalent models that explored an association with LBW rates using the 2014–2018 subset. We found that for every 1-unit increase in the well density per square mile of wells that use an ETC, we observed a 1.964-higher LBW rate (95% CI: 0.41, 3.52). For every 1-unit increase in well density per square mile of wells that use TTC, we found a 1.619-higher LBW rate (95% CI: 0.37, 2.87). We also found that an increase in well density per square mile among wells that use chemicals that include an OHTC resulted in a 1.244-higher LBW rate (95% CI: 0.43, 2.06). We also explored the role of HF well density in general (grouped regardless of the chemicals used) and found that a 1-unit increase in density of all wells grouped (regardless of hormonal targeting status of the chemicals used) resulted in 0.602-higher LBW rate (95% CI: 0.15, 1.05) compared with counties with no wells. The association between HF wells and PTB was statistically significant ($p\le 0.05$) for all models from the full dataset, the states with HF exposure fully captured subset, and the nonmetro subset. The association was not statistically significant in the farming and mining subset, although the pattern of coefficient estimates held. Similarly, the association between HF wells and LBW was statistically significant in the full dataset and states with HF exposure fully captured subsets. The associations were not significant in the farming and mining and nonmetro subsets, although the pattern of estimates remained. Exact estimates, standard errors, and $p$-values for all models are given in Table 3. The pattern of coefficient estimate values is visualized in Figure 4 along with the corresponding 95% CIs.

In Figure 4, we plot the results of our models of all US counties, comparing the HF coefficient estimates for the different HF well types to a few covariates that are known to be associated with poor birth outcomes in scientific literature that we had also included in our models to adjust for those known effects.^{43,45,46,64–67} Results for our models are available in Excel Table S5. For example, in our model of all counties and wells active during 2014–2018, for every 1-unit increase in all well density per square mile (regardless of whether chemicals that targeted any hormonal pathways were used in the well), there resulted a 1.228 (95% CI: 0.66, 1.80)-higher PTB rate while adjusting for poverty rate among other relevant covariates (a subset of these are visualized in Figure 4). Similarly, for every 1-unit increase in wells that use chemicals known to target estrogen pathways per square mile, we observed a 3.788 (95% CI: 1.83, 5.74) increase in the PTB rate even after adjusting for poverty rate, maternal care deserts,⁴⁹ and population density among other relevant covariates. In our models, we studied the relationship between HF and the hormonal targets of the HF chemicals on both PTB and LBW while adjusting for known factors that increase the rates of these adverse perinatal outcomes, including race and ethnicity, poverty, population density, education, marriage, drug deaths per 100,000 people, percentage insured, fertilizer, insecticide, herbicide, agricultural land, and maternal care availability. Therefore, the results presented were all after adjustment for the factors mentioned. All model results are provided in detail in Excel Table S5.

We performed several analyses on stratified subsets of our data (Table 3) to explore the relationship between various potential confounders and rates of our perinatal outcomes of interest (PTB and LBW). However, we found that the trend was consistent with higher rates of poor perinatal outcomes being observed when HF chemicals contained estrogen-modulating ingredients, followed by testosterone modulating ingredients, and then other hormonal modulating ingredients. Therefore, ETCs when used in HF appear to have had higher associations with these birth outcomes, whereas TTCs had a statistically significant relationship with these outcomes, but the estimates were less pronounced than the effect of ETCs. This is important for public health researchers because often the exposure to HF is considered as one single constant exposure across various cohorts and this results in misleading and contradictory findings in the literature.^3–8 Currently there is a paucity of researchers studying the role of the chemicals used in HF and their role involving hormonal pathways and human birth outcomes outside of animal models. However, there is a tremendous amount of evidence that endocrine disruptors negatively impact human health and the regulation of these chemicals in other contexts.^68,69

We identified 106 gene/proteins involved in one of the three pathways: estrogen, testosterone, or other hormone pathways that were targeted by HF ingredients (Excel Table S3). The two proteins with the greatest number of distinct ingredients (as measured by distinct Chemical Abstract Service numbers) targeting them were estrogen receptor 1 and 2 (ESR1 and ESR2); these had 804 and 501 distinct ingredients targeting them respectively. In third place, was estrogen sulfotransferase (SULT1E1), an important enzyme involved in the estrogen pathway. Therefore, the estrogen pathway had the largest number of distinct ingredients targeting them.

The general hormone pathway also had many distinct HF ingredients targeting several key gene/proteins involved in key hormonal processes in the body. There are many potential biological mechanisms that link non-estrogen and non-testosterone hormones with biological process that have potential to affect human health and disease.⁷⁰ We found 11 distinct gene/proteins that are involved in thyroid function. Inadequate thyroid function has been linked to LBW⁷¹ and the combination of PTB and LBW.⁷² The thyroid-stimulating hormone receptor gene/protein was targeted by two distinct HF ingredients and mutations in this very gene/protein has been previously linked to both PTB and LBW.⁷²

In the present study, we have focused on chemicals used in HF fluid that target three hormonal pathways: estrogen, testosterone, and general hormone pathways. Prenatal exposure to EDCs is known to cause adverse health outcomes in many species, including fish, mice, and humans.^29,30,34,35 However, prior studies on the human health effects of HF fluid that focused on the chemicals and their endocrine-disrupting effects remain limited due to the complex nature of these mixtures. The purpose of our study was to link all related chemicals together that target the same key pathways to see if we could study the adverse birth outcomes related to these exposures at the county level across the entire nation. Our findings point to the overall potential harm of these EDCs on birth outcomes across the nation. However, we are hypothesizing that water is the major route of this exposure whereby those who are proximal to HF wells are more likely to have water exposure in their drinking water. We did not have access to adequate private water well consumption information; as discussed in the “Limitations” section, data on the use of private well water for drinking is available but has not been updated since the 1990 census.⁷³ We chose to not include this data because we expected that population densities and well water use for human consumption has likely changed since 1990, especially in areas with HF given that local residents may have reduced well water use to reduce exposure. We also did not explore the role of watersheds and complex topology of the landmass throughout the nation because this varies within counties and would require additional methods development to assess adequately.^74–77 We leave this to future researchers interested in using our work as a starting point toward individualized models in specific regions of the United States where topological mapping and assessment of water sheds and private vs. public water consumption would be possible.

We created the farming and mining and nonmetro subsets to explore whether our findings were an artifact of rural–urban health disparities. The association between HF wells and perinatal outcomes was not statistically significant in the farming and mining subset models, although the pattern of coefficient strength remained. The farming and mining subsets include only 457–500 counties each, compared with the much higher number of counties in other subsets (Table 1). The relatively small sample size may explain the loss of statistical significance. Similarly, in the nonmetro counties subset models, the PTB models found statistically significant associations, whereas the LBW models did not. Across all models (Table 3; including the ruralness subsets), PTB models had larger coefficient estimates and lower $p$-values than LBW models, suggesting a stronger relationship between HF and PTB than with LBW.

The FracFocus database used as the primary source for our WellExplorer database³⁹ has multiple limitations. As previously mentioned, only 26 US states are required to report HF operations to FracFocus. Some states that are required to report have only mandated it within the last decade, and some states give well operators the option of reporting to FracFocus or other entities, although FracFocus does not make this clear on their website. It is challenging to tease out who is required to report to FracFocus and when, and there is disagreement by prior studies on when reporting became mandatory in individual states’ legislations.^78,79 In addition, FracFocus’s chemical data is not comprehensive, with some ingredients withheld because they are considered confidential, proprietary, or trade secrets.⁷⁸ Beyond this, some non-withheld ingredient data is only available in the individual PDFs for each well and not available in the downloadable dataset that we used because it was the most easily and publicly accessible. Unfortunately, it is still the most comprehensive record of HF wells and ingredients that is publicly available. Owing to FracFocus’s limitations, our analysis lacks a record of some HF wells and comprehensive well ingredient data to use in classifying wells by impact on human hormone pathway. Some counties’ well densities may be incorrect, and some wells may be misclassified as not impacting hormone pathways because of missing ingredient data.

Although we expanded on prior studies and included information on well ingredients, we were not able to account for many other potentially relevant variables in the HF process that could be relevant to health impacts. We have some data on the start and end dates of HF operations, but we lack records on the phase of oil or gas extraction of each well, which has been found to modulate exposure risk and can vary greatly given that some wells are fractured repeatedly to maintain a steady extraction of gas or oil.³⁶ Deeper wells, wells with horizontal drilling, and wells that are fractured more often would likely pose a higher risk of contaminating nearby water sources, but we lacked information on these practices. In addition, by using well density, we were forced to assume a uniform well distribution over a county’s land area. We did not consider the presence of nearby conventional oil and gas extraction nearby, which has been highlighted as an important consideration in HF research.^80–83

Our study relied on publicly available data at the county level, which is significantly less granular than primary hospital records. Using public sources and county level of analysis allowed us to compare findings across the country but limited us to the variables by which we could adjust for in our models. For example, we did not have information regarding individuals’ exact levels of exposure to HF ingredients, including proxies such as the distance between their houses and a HF well and whether they sourced drinking water from private wells or municipal sources.⁸⁴ Prior research has predominantly found associations for people living within relatively small radii of wells, up to $10\mathrm{km}$, with highest impacts within just a few kilometers.^28,82,85 Data on the use of private well water for drinking is available but has not been updated since the 1990 census.⁷³ County-level exposure data is much less precise and would likely weaken the associations we should see if there is a true relationship. Therefore, we are confident that our findings represent something real and if individual patient-level records were used, we expect that our associations would be even stronger. We were also unable to control many relevant demographic features that impact perinatal outcomes on an individual level, such as whether the pregnancy was a multiple or singleton or if the mother had comorbid conditions. On the positive side, by using counties, we were able to adjust for factors such as exposure to agricultural operations and pesticides, which are difficult to adjust for when using individual-level patient records because that data is often not included or missing in patient records.

We reiterate here that our study does not establish causality between fracking and adverse birth outcomes. Our study rather establishes that there are associations between fracking activities that use chemicals that are known to target estrogen, testosterone, or other hormonal pathways and increases in adverse birth outcomes. We adjusted for many factors that are also known to increase the risk of these adverse birth outcomes, including population per square mile, maternal care deserts,⁴⁹ drug deaths per 100,000 people, proportion of the population reporting belonging to various racial/ethnic minorities (each modeled separately), various education statistics, as well as use of fertilizers and herbicides, and the acres of agricultural land per square mile, along with poverty, insurance status, and marital status. All of these factors have been in the past linked with adverse birth outcomes,^17–27 and we therefore adjusted for these factors using our county-level model. We would like to caution our readers from overinterpreting a county-level model. We constructed this model to determine if the chemicals used in fracking appear to matter across the United States because various state or region-specific studies appeared contradictory. We have established with our work that there does appear to be a signal based on the chemicals used in fracking and the pathways those chemicals target, with ETCs having the highest effect followed by TTCs and, OHTCs. This does not establish causality. The next step would be to construct a spatial model that adjusts for spatial patterns and takes into account how counties that are close to each other might have similar effects while also understanding that state boundaries are a key factor in many policies that affect birth outcomes, insurance access, and even reporting of fracking activities. Therefore, future work involves developing a robust spatial model that explores some of these factors and even models’ political decisions (e.g., requirements for state-level reporting of fracking activity and so forth) in a robust way. Our model is important in establishing that a pattern appears to exist, and it appears to be strong enough that a county-level model is picking up this signal. However, more work is needed. We have constructed spatial models in other studies,⁸⁶ and this remains the subject of our future work. We also encourage other researchers to learn from our findings and expand upon them in their own future studies and research.