The Causal Effect of Increasing Area-Level Income on Birth Outcomes and Pregnancy-Related Health: Estimates From the Marcellus Shale Boom Economy

Abstract

Introduction

Health at birth underlies developmental processes that affect health throughout the life course. Infants born prematurely or with low birth weight have a greater risk of developing cognitive and health challenges (Figlio et al. 2014; Johnson and Schoeni 2011), which are associated with lower educational attainment, lifetime earnings, and adult health (Behrman and Rosenzweig 2004; Currie and Almond 2011; Strauss 2000). Scholars have theorized that higher parental income improves infant health (Rosenzweig and Wolpin 1991)—a theory supported by many observational studies. However, empirical challenges prevent scholars from identifying whether and how exogenous income gains affect birth outcomes. To inform public policies aimed at improving health at birth, we advance scholarship on the causal links between economic circumstances and birth outcomes by testing whether and how plausibly exogenous community-level income increases impact infant health.

Economics and population health research has demonstrated that health improves and mortality declines during economic recessions (Gerdtham and Ruhm 2006; Ruhm 2000, 2015; Tapia Granados and Diez Roux 2009), although such findings have weakened over time and potentially reversed since the Great Recession (Burgard and Kalousova 2015; Catalano et al. 2011; Cutler et al. 2016; Margerison-Zilko et al. 2016). Findings on infant health during recessions, particularly from studies using data from the Global North, are also mixed. Some studies have found that infant health improves during recessions because pregnant people¹ have more time to attend prenatal care visits (Dehejia and Lleras-Muney 2004), whereas others have found that infant health declines because of parental job loss and increases in pregnant people's alcohol consumption and smoking behavior (De Cao et al. 2022). Studies using exogenous increases in means-tested social welfare benefits have largely found that income gains among low-income birthing parents lead to slight increases in birth weight and declines in the incidence of low birth weight (Almond et al. 2011; Currie and Cole 1993; Hoynes et al. 2015; Hoynes et al. 2011; Strully et al. 2010).

Yet, important knowledge gaps remain. First, taking a novel perspective, we examine the causal influence of macroeconomic booms, rather than recessions, on infant health. This shift is important because findings from recessions might not predict the health effects of policies that increase incomes; the effects of economic losses and gains are not likely symmetrical (Kahneman and Tversky 1979). Second, macroeconomic research has rarely directly measured income and has instead used employment rates as a proxy. Yet that approach omits individuals who involuntarily work part-time and those not in the labor force. Third, quasi-experimental studies of increasing social welfare benefits cannot generalize to middle- and high-income individuals because an absolute gain of $100 does not have the same relative value for low-income families as for higher income families (Kolm 1976). Thus, it remains unclear if population-wide income gains affect the health of pregnant individuals and their infants. Finally, although several studies have found that exogenous income gains improve birth weight, the mechanisms remain poorly understood.

We use a natural experiment in which we leverage the economic boom resulting from the extraction of natural gas in the Marcellus Shale geological formation to study how exogenous community-level increases in average family income impacted pregnancy-related and infant health outcomes in Pennsylvania (PA) from 2007 to 2012. We use school districts, a smaller geographic unit than used in similar research, to measure community characteristics because school districts are meaningful administrative units for residential choice, and residential sorting leads to sharp differences in household income and education along school district boundaries (Bayer et al. 2007). Further, we explore the potentially mediating role of health and health behaviors during pregnancy between exogenous income gains and infant health. To address unobserved heterogeneity across birthing parents, we leverage a subsample of nontwin siblings to estimate the effects of income change on pregnancy behaviors and infant health (Rosenzweig and Wolpin 1995) by comparing outcomes between siblings born before and after the Marcellus Shale economic boom. We also test the exogenous effect of income gains across the full population and test for nonlinear income effects by stratifying by the district poverty rate before the economic boom. Finally, we estimate several falsification models to ensure that our results are not explained by selective fertility or migration because the effect of income on infant health partially operates via these pathways (De Cao et al. 2022).

Income, Pregnancy Behaviors, and Infant Health

Income could impact birth outcomes and pregnancy health behaviors via several mechanisms, including financial strain, prenatal stress, food insecurity, and prenatal care access. Financial strain is a chronic stressor linked to coping behaviors, such as smoking cigarettes and overeating (Adam and Epel 2007). Smoking is associated with adverse birth outcomes, including low birth weight (Pereira et al. 2017; Riaz et al. 2018) and preterm birth (Abrevaya 2006; Rosenzweig and Wolpin 1991), whereas prepregnancy obesity (Shin and Song 2015) and excessive gestational weight gain (Bodnar et al. 2015) are associated with infant prematurity. Increased prenatal stress is linked to low birth weight and preterm birth independent of pregnancy behaviors (Almond and Currie 2011; Bussières et al. 2015). Food insecurity is linked to suboptimal gestational weight gain and increased risk of pregnancy complications (Laraia et al. 2010). Finally, income is highly correlated with prenatal care access, which encourages beneficial pregnancy-related behaviors, such as smoking cessation and adequate gestational weight gain, and helps manage pregnancy complications (Reichman et al. 2010; Yan 2017).

Observational studies have found that infants born to low-income individuals are at greater risk of low birth weight and preterm birth (Currie 2009; Finch 2003; Kramer et al. 2000; Martinson and Reichman 2016). Low-income birthing parents are also more likely to smoke before and during pregnancy (Riaz et al. 2018) and are less likely to receive adequate prenatal care (Green 2018), begin pregnancy classified as normal weight, and gain weight within clinically recommended ranges during pregnancy (Paul et al. 2013).

Despite these observational findings, research estimating the causal effects of individual-level income on pregnancy-related health and birth outcomes has proven to be challenging. First, there are measurement limitations. U.S. birth certificates do not record parental income and other characteristics that predict pregnancy behaviors and birth outcomes (e.g., pregnancy intentions and knowledge; Hohmann-Marriott 2009). Second, income is correlated with unobserved parental characteristics that affect pregnancy behaviors and birth outcomes, which might bias the estimated association between observed income and health (Glymour et al. 2014; Mayer 1997). To address these challenges, scholars have leveraged changes in macroeconomic indicators (i.e., area-level unemployment rates) or social welfare policies to estimate the causal effect(s) of economic conditions on infant health. However, such research has yielded mixed findings.

Macroeconomic Conditions and Infant Health

On one hand, some studies have suggested that economic recessions improve infant health. Dehejia and Lleras-Muney (2004) found that periods with relatively high unemployment rates were linked to lower rates of low and very low birth weight (i.e., <2,500 and <1,500 grams, respectively), fewer congenital malformations, and lower postneonatal mortality owing to changes in pregnancy health behaviors and, for Black birthing parents, positive selection into pregnancy (i.e., Black birthing parents tend to be more educated and married during recessions). Similar salutary effects of unemployment on very low birth weight, preterm birth, and neonatal mortality have been documented in Sweden (van den Berg et al. 2020) and Spain (Aparicio et al. 2020). In Sweden, these associations are stronger for economically disadvantaged households (van den Berg et al. 2020). By contrast, the associations in Spain are driven by composition, with fewer first births during recessions, which are positively correlated with adverse outcomes (Aparicio et al. 2020). Declines in air pollution could be a contributing factor; recession-induced declines in air pollution reduce infant mortality in the United States (Chay and Greenstone 2003).

On the other hand, recessions might harm infant health. De Cao et al. (2022) leveraged the Great Recession and a large sample of siblings and found that rising unemployment is associated with decreases in birth weight and increases in small for gestational age and stillbirth, partly owing to smoking and alcohol consumption increases and prenatal care delays. These adverse outcomes are greater for infants born to unpartnered parents (De Cao et al. 2022). With some caveats, similar conclusions about economic recessions harming infant health have been found in studies in Argentina (Bozzoli and Quintana-Domeque 2014), Japan (Kohara et al. 2019), Iceland (Olafsson 2016), and the Northern Netherlands (Alessie et al. 2018). In the Northern Netherlands, the detrimental associations between higher unemployment rates and infant health are stronger for male infants born to older individuals, smokers, or both (Alessie et al. 2018). In Argentina, the associations are stronger for births to low-educated and low-income parents, possibly because of their greater chronic stressors and heightened nutritional deprivation (Bozzoli and Quintana-Domeque 2014).

Individual-level studies leveraging exogenous change in parents’ employment histories also concluded that economic recessions harm infant health and that booms improve infant health. For example, the mere announcement of forthcoming layoffs is associated with declines in birth weight (Carlson 2015), as is fathers’ job loss, especially among economically disadvantaged families (Lindo 2011). Exogenous improvements in low-skilled women's earnings increase prenatal care use and have small, beneficial effects on birth weight and gestational age (Mocan et al. 2015).

To our knowledge, no study has examined the links between macroeconomic booms and infant health, which could have differential effects relative to recessions. For example, economic booms could improve infant health by reducing individuals’ financial strain and food insecurity while improving prenatal care access. Yet, economic booms could harm infant health by boosting air pollution, which increases the likelihood of low birth weight (Bell et al. 2007; Currie et al. 2009; Maisonet et al. 2001; Parker et al. 2005; Savitz et al. 2014), preterm births (Huynh et al. 2006; Ponce et al. 2005; Sagiv et al. 2005), and infant mortality (Currie and Neidell 2005; Knittel et al. 2016). It is unclear a priori whether the detriments of air pollution outweigh the benefits of an economic boom for infant health.

Natural Experiments of Policy Change and Infant Health

Related research has studied the causal effect(s) of income on pregnancy-related health and birth outcomes by leveraging natural experiments resulting from policy-related exogenous income changes. Although most research could not identify individual-level income receipt, three exceptions are studies examining (1) the Gary Income Maintenance Experiment in Gary, Indiana (Kehrer and Wolin 1979), (2) the Alaska permanent fund dividend (Chung et al. 2016), and (3) universal child benefits in England and Wales (Reader 2023). All three studies found that an exogenous increase in income led to small improvements in birth weight and decreases in low birth weight, and two of them (Chung et al. 2016; Reader 2023) also found small improvements in gestational length.

Most natural experiments of exogenous increases in individual or family income have assumed income receipt based on relevant policy guidelines and the birthing parent's demographic characteristics. Studies leveraging geographic variation in the introduction of policies across states or counties have found that increasing income from Aid to Families with Dependent Children (Currie and Cole 1993), food stamps (Almond et al. 2011; Currie and Moretti 2008), and Special Supplemental Nutrition Program for Women, Infants, and Children (Hoynes et al. 2011) led to improvements in birth weight. Almond et al. (2011) found that the positive effect of food stamps on birth weight is greater for births in high-poverty areas. Expansions of the Earned Income Tax Credit (EITC), a refundable payroll tax credit to low-income workers and the most extensive U.S. anti-poverty program, are also associated with modest reductions in low birth weight and preterm birth, especially among those with less than a high school education (Hoynes et al. 2015; Markowitz et al. 2017; Strully et al. 2010; for an exception, see Hamad and Rehkopf 2015). Evidence about whether EITC-associated income gains impact prenatal smoking rates and prenatal care use is mixed: one study found no effects (Markowitz et al. 2017), whereas others found that income gains reduce smoking among lower educated mothers of minor children (Averett and Wang 2013; Cowan and Tefft 2012; Hoynes et al. 2015) and prenatal care use (Hoynes et al. 2015). Studies exploiting state-level minimum wage increases found that income gains lead to significant, albeit small, improvements in birth weight and fetal growth but do not affect prenatal care use or smoking (Komro et al. 2016; Wehby et al. 2020). Finally, Koirala (2021) found that a greater intensity of state-level child support enforcement is associated with a lower incidence of low birth weight among unmarried birthing parents.

The Marcellus Shale Economic Boom

The Marcellus Shale geological formation is one of the largest natural gas fields in the world but remained nonviable for commercial gas development until 2005, when Range Resources–Appalachia, LLC, developed horizontal drilling and hydraulic fracturing (i.e., “fracking”). These technological innovations provide access to 1.9 trillion cubic feet of natural gas trapped in microscopic pockets in the Marcellus Shale geological formation. The pace of drilling accelerated after 2008, when geologists estimated that the Marcellus Shale formation contained 250 times more natural gas than initially estimated (Engelder and Lash 2008).

Mineral rights owners who sign contracts with gas companies receive a one-time leasing payment and monthly royalty payments over the life of the well based on the amount of gas extracted and the contracted royalty rates (i.e., the profit percentage owed to the mineral rights owner upon the sale of the gas from their parcel). The state-mandated minimum royalty rate is 12.5%, and the average reported royalty rate is 13.2% (Brown et al. 2016). In 2010, natural gas companies paid $2.07 billion in Marcellus Shale leasing and royalty payments to PA residents (Considine et al. 2011).

These royalty and leasing payments were eclipsed by the increase in local employment and earnings, although total estimates vary (Brown 2014; Considine et al. 2016; Munasib and Rickman 2015). During the economic boom, employment and wages improved within the oil and gas industry (Cruz et al. 2014) and among locals providing subcontracted work for the gas industry (e.g., commercial truck drivers transporting trillions of gallons of water) or providing housing and food services to out-of-state oil rig workers (Brasier et al. 2011; Kelsey and Hardy 2015; Wrenn et al. 2015). The localized economic boom occurred between 2008 and 2011 because the PA Environmental Protection Agency (EPA) issued voluntary water regulations to the industry in 2011, dramatically cutting the demand for commercial truck drivers (Jacobs 2018).

Marcellus Shale development poses two plausible health risks that adversely affect pregnancy-related health and birth outcomes: (1) exposure to air, water, and noise pollution (Casey et al. 2019; Colborn et al. 2011; Goetz et al. 2015); and (2) community disruptions from increased truck traffic and inflows of transitory workers during the drilling phase (Brasier et al. 2011; Perry 2012; Stedman et al. 2012). Parents’ proximity to natural gas development is correlated with increased risks of infant cardiovascular heart defects, low birth weight, preterm birth, and lower Apgar scores (Casey et al. 2016; Currie et al. 2017; Hill 2018; McKenzie et al. 2014; Stacy et al. 2015). These studies revealed crucial knowledge on the associations between fracking and birth outcomes, but they did not capture the income gains from Marcellus Shale development that could offset its adverse environmental health effects (Currie et al. 2017). Aside from Casey et al. (2019), none of these studies specifically examined the Marcellus Shale impacts on pregnancy-related health or health behaviors.

To instrument for income gains, we do not leverage data about actual industrial activity (i.e., location of active wells, pipelines, gas compressor stations, water reservoirs, or waste landfills) because they would be confounded with the environmental risks of Marcellus Shale development. Instead, we rely on an industry map of anticipated economic returns. Nearly a decade before drilling, gas industry experts identified the Marcellus Shale's more economically valuable areas based solely on the depth, thickness, porosity, thermal maturity, and silica content of the rock (Dell et al. 2008). We refer to the area predicted to be more economically viable—our treatment communities—as the High-Yield Potential Extraction (HYPE) area. We refer to the less economically viable area—our control communities—as the Periphery (see Figure 1). Although the HYPE area strongly predicts where oil and gas companies leased land for drilling, it does not strongly correlate with actual Marcellus Shale development and environmental hazards because those reflect aboveground opportunities, constraints, and human activity (see Table 6). We elaborate on our instrument and test its validity below.

Our study offers three key contributions. First, we test whether community-level income gains from Marcellus Shale development predict changes in pregnancy health behaviors before and during pregnancy. Our examination of potential income effects on prepregnancy health and behavior outcomes is novel and motivated by research that finds prepregnancy health and behaviors are upstream influences on pregnancy experiences and birth outcomes (Lemon et al. 2016; Perry et al. 2019)—the outcomes studied in prior income effects research. We hypothesize that income gains will improve infant health and pregnancy-related health-promoting behaviors and reduce adverse health conditions and suboptimal pregnancy-related behaviors. Second, rather than focusing solely on low-income populations, we study the effects of income gains across the income spectrum while testing for heterogeneous income effects by preboom local poverty levels. We hypothesize that income gains will more strongly affect birth outcomes in districts with higher preboom poverty rates. Finally, in a subset of analyses, we leverage data on sibling pairs—one born before the Marcellus boom and one born after—with birthing parents living in the same district. These analyses control for time-invariant unmeasured characteristics of birthing parents and their environments to refine our estimate of the causal effect of income gains on infant and pregnancy-related health.

Data and Methods

Natural Experiment Approach

As with nearly all natural experiments studying changes in social welfare benefits, we cannot confirm which individuals experienced income gains, but we can measure community-level changes in average income. We calculate the average income gains across school districts and predict the association between district-level income gains and individual-level health outcomes. We match birthing parents to school districts and school district characteristics using their residence's longitudinal and latitudinal coordinates at the time of birth.

School district-level analyses provide several benefits. First, school districts often match municipal government boundaries, with studies of enterprise zones (i.e., impoverished areas with municipal tax breaks to entice business investment) frequently using school districts (Byrnes et al. 1999; Kenyon et al. 2020). Second, such analyses include nonurban populations, which is essential because nearly a third of Pennsylvania's population lives in rural areas and rural residents have, on average, lower per capita income, greater unemployment, and a shortage of health care providers (Carroll et al. 2023; Pennsylvania Office of Rural Health n.d.; Sanders 2023). Third, the number of districts permits sufficient statistical power to estimate stratified models (vs. only 37 PA counties above the Marcellus Shale). Fourth, relative to larger spatial units, school districts are closer to individual-level experiences. Finally, the school district level is the smallest aggregation of PA tax data. Notably, environmental studies of hydraulic fracturing have often used school districts for spatial aggregation and analyses (Bartik et al. 2019; Cai et al. 2021; Harleman et al. 2022; Kovalenko 2023).

All study variables and their sources are listed in Table A1 (online appendix). To create our independent and dependent variables, we use PA vital statistics and tax data, as well as an industry map. For supplemental district-level models, we derive district traits from PA and federal administrative data and U.S. Census Bureau surveys, including the American Community Survey (ACS). Because the area's residential boom occurred between 2008 and 2010, we use 2007 births for the preboom period and 2012 for the postboom period to allow a one-year lag for behavioral change, conception, and completed gestation.

Pregnancy and Birth Outcomes

We predict eight pregnancy-related health and health behavior measures: (1) prepregnancy smoking set to 1 if the birthing parent smoked within three months before pregnancy; (2) any pregnancy smoking set to 1 if the birthing parent smoked during any trimester; (3) prepregnancy underweight (body mass index [BMI] < 18.5) based on self-reported weight and height (Centers for Disease Control and Prevention 2022); (4) prepregnancy overweight/obesity (BMI ≥ 25); (5) first-trimester prenatal care initiation set to 1; (6) adequate prenatal care use set to 1, defined as initiating care in the first trimester and attending at least 80% of recommended visits for gestation length (Kotelchuck 1994); (7) low gestational weight gain set to 1; and (8) high gestational weight gain set to 1. We create gestational weight gain classifications by mapping pregnancy weight gain (i.e., the difference between delivery weight and self-reported prepregnancy weight) and gestational length to well-validated weight-gain-for-gestational-age z scores and defining low gestational weight gain as z score less than 1 standard deviation and high gestational weight gain as z score greater than 1 standard deviation (Hutcheon et al. 2013).

We predict five birth outcomes: (1) birth weight in grams; (2) low birth weight set to 1 if less than 2,500 grams (Almond et al. 2005); (3) very low birth weight set to 1 if less than 1,500 grams (Almond et al. 2005); (4) physician-estimated gestation length in weeks; and (5) preterm birth set to 1 if born before 37 weeks of gestation (Beck et al. 2010).

District Average Income

Using PA Department of Revenue data, our main independent variable is the district's average adjusted gross income, calculated as total income minus specific deductions per household for each school district and adjusted for inflation to 2010 dollars using the CPI-U-RS (Stewart and Reed 2000). For supplemental models, we also measure the following income subcategories: (1) mean royalty income, which consists of income from “royalties, rents, patents, and copyrights”; (2) mean earnings; and (3) mean property sales. To reduce measurement error in the postboom era, we average tax data for 2010 and 2011.

Marcellus Shale Comparison Groups

To classify districts, we calculate the proportion of the district's area above the HYPE and Periphery areas, as well as their area outside the Marcellus Shale formation (see Figure 1). The treatment districts include those with 51+% of their area above the HYPE (n = 188), whereas the comparison districts include those with 51+% of their area above the Periphery (n = 98). To explore the validity of our instrument, we also identify PA districts with 51+% of their area outside the Marcellus Shale.

Control Variables

We include individual-level characteristics that are plausibly associated with birth outcomes, pregnancy-related health, or health behaviors (Cohen et al. 2016; Currie and Moretti 2003; Margerison et al. 2019; Masho et al. 2010). These characteristics are the birthing parent's age, age squared, educational attainment (i.e., high school diploma or less, some college, or college degree or more), marital status, and race and ethnicity. We also include district fixed effects (FE) to control for time-invariant district traits and region FE to control for spatial autocorrelation across districts.

Initial District Poverty

We test for heterogeneous effects of income gains by preboom district poverty rates. We stratify districts by whether they are above or below the median share of PA Marcellus Shale residents with incomes less than 200% of the federal poverty line in 2007 (median share = 31.7%).

Analytic Samples

For our main analytic sample, we restrict the data to singleton births to individuals aged 15 or older because multiple births often necessitate additional prenatal care and increase the risk of adverse birth outcomes (Alexander and Salihu 2018). We also drop all births occurring in two PA school districts that merged in 2009 (n = 348) to ensure consistent district-level data. We exclude births in four district outliers that average approximately five births per year (n = 43) and births above the Marcellus Shale but outside the treatment and comparison districts (n = 10,126). The final primary analytic sample contains 78,249 births in 284 PA school districts, although the exact sample size varies across outcomes because of item nonresponse on birth certificate data (see Tables 3–5 for specific counts).

To create a subsample of nontwin siblings (n = 12,390 in 271 districts), we select birthing parents with live births during 2005–2007 and 2012–2013 who reside in the same school district across births. If the birthing parent has two live births in either the preboom or postboom years, we select the births that occurred closest in time to the Marcellus Shale boom.

Analytic Plan

We estimate the average treatment effect on the treated (ATT) for district income gains (X) on individual pregnancy and birth outcomes (Y) by integrating two quasi-experimental techniques—instrumental variables (IV) and difference-in-differences (DiD)—in an instrumented difference-in-differences design (DDIV) (Duflo 2001; Hudson et al. 2017; Ye et al. 2023). IV methods use a third variable (Z) to estimate the causal associations between X and Y that would otherwise be biased (Angrist et al. 1996), whereas DiD methods compare pre–post outcomes for an exposed treatment group and an unexposed comparison group. We rely on a system of two-stage least squares. The first stage predicts mean district income as a function of the DiD (operationalized as the interaction between the IV [Z = 1 for Marcellus Shale HYPE] and time [t = 1 for postboom]). The second stage predicts pregnancy and birth outcomes as a function of predicted mean district income. The key DDIV coefficient, β, comes from the following two-stage IV system:

where i references the individual, d references school districts, t references period (i.e., 2007 or 2012), r references regions, MeanIncome is the observed average district income, District and Region are vectors of district and region FE, Post is a dichotomous variable equal to 1 for 2012 births to capture period effects, BirthingParentControls_i is a vector of birthing parents’ demographic traits, and HYPE is a dichotomous variable equal to 1 for districts with most of their land above the Marcellus Shale HYPE. The reduced-form regression is

To control for time-invariant birthing parent characteristics that could affect pregnancy or birth outcomes (e.g., genetic or personality traits, their childhood exposures), we estimate an additional model that narrows the sample to siblings born to the same birthing parent living in the same school district at the time of each birth, with one sibling born before and one after the Marcellus boom. Although these sibling models increase the precision of our estimates, they derive from a sample with a slightly different demographic profile (see Table 1). The sibling DDIV models are similar, but the second-stage equation changes as follows:

where all prior terms are defined as before. Pregnancy/Birth Outcome_ijdt refers to birth outcomes and pregnancy-related health for sibling i born to mother j; BirthingParentControls_j refers to time-varying birthing parental characteristics (i.e., age, age squared (in years), marital status, and educational attainment); and Parent is a vector of parent FE.

To estimate these models, we use Stata 18’s ivregress command with cluster-robust standard errors.

Model Assumptions and Falsification Tests

A strong, unbiased DDIV approach requires meeting both IV and DiD conditions. We depict the assumed causal relationships for our DDIV model as a path diagram in Figure 2, where Y is pregnancy and birth outcomes. The instrument in IV methods (Z) must (1) meet the relevance condition by having a causal effect on X; (2) meet the exclusion restriction by affecting Y only through X; and (3) meet the independence condition (or unconfoundedness) by not being confounded with Y. Of the three IV assumptions, the only verifiable condition for IV methods is relevance. As other IV research has done, we test relevance with the first-stage parameter, π (see Table 4). We also use the corresponding F statistic to measure the strength of the instrument (Andrews et al. 2019). The second and third IV conditions—the exclusion restriction and independence—are always unverifiable. However, theoretical justification, falsification tests, and model adjustments provide evidence that these conditions are plausibly met. The most likely violation of the exclusion restriction is that Marcellus HYPE location affects pregnancy behaviors and birth outcomes through drilling-related pollution and environmental hazards (i.e., Figure 2 would display an arrow from Marcellus Shale HYPE status to the oval labeled “Unmeasured Confounders”). Thus, we test whether HYPE location predicts district-level changes in multiple environmental hazards (see Table 6) using district-level FE models with cluster-robust standard errors as follows:

$District Traitdr = κ + υHYPE + oPostt +ϖ(Postt × HYPE) +ξDistrictd +edt,$

(5)

Fig. 2

View largeDownload slide

Path diagram of assumed causal relationships for pregnancy and birth outcomes (Y) in our DDIV model. Boxes indicate observed variables, and the oval indicates latent variables. The subscript t refers to time, whether preboom or postboom.

Fig. 2

View largeDownload slide

Path diagram of assumed causal relationships for pregnancy and birth outcomes (Y) in our DDIV model. Boxes indicate observed variables, and the oval indicates latent variables. The subscript t refers to time, whether preboom or postboom.

Close modal

where all prior terms are defined as before. The key coefficient, ϖ, indicates whether HYPE districts are significantly different postboom. HYPE status is seen as compliant with the independence condition. It is also unconfounded with pregnancy behaviors and birth outcomes because HYPE is based on geological characteristics generated millions of years ago. We also use two placebo tests to support our theoretical justification of independence (Table 4): we examine whether (1) HYPE status leads to significantly higher postboom district income in New York (NY), where hydraulic fracturing is banned (Kaplan 2014); and (2) Periphery status predicts greater postboom district income for PA districts without land above the Marcellus HYPE (n = 96).

DiD methods are valid under three assumptions (Angrist and Imbens 1995; Card and Krueger 1994; Goodman-Bacon 2021): (1) without treatment, the difference between the treatment and comparison groups would be constant over time (i.e., parallel trends); (2) the population composition of the treatment and comparison groups is stable; and (3) there are no dynamic treatment effects. For our design, there are no dynamic treatment effects (the third condition) because we have only two periods (i.e., before and after the boom), given the short window when local residents benefitted from Marcellus Shale development. The first DiD condition is always unverifiable because it pertains to potential outcomes, but we gauge the plausibility of parallel trends by comparing treatment and comparison districts’ preboom and postboom characteristics using t tests of district means (see Table 2). We test the stability of the treatment and comparison group compositions (the second condition) by using descriptive statistics about births in treatment and comparison districts before and after the boom (see Table 1) and district-level FE models with cluster-robust standard errors (Eq. (4)) to look for evidence of differential selective migration or differential selection into fertility across treatment and comparison districts (see Table 6).

All supporting district-level models use inverse probability weights (IPWs) to control for any preboom differences between HYPE and Periphery districts and to limit area-level confounding (Rosenbaum and Rubin 1985). In creating the IPWs, we build a logistic regression model to estimate the probability that a district is above the Marcellus HYPE by considering more than 60 preboom demographic, economic, political, cultural, and environmental measures, including environmental factors linked to infant health and hydraulic fracturing (Black et al. 2021; Fontenot et al. 2013; Hill and Ma 2022; Lauer et al. 2016; Webb et al. 2014; Webb et al. 2018). (See section 10 of Table A1 in the online appendix for all IPW variables.) Using Stata's pscore command, we check for propensity score balance across seven blocks and review the four treatment districts with low propensity scores. We weight all district-level FE models using the product of the estimated IPW and the district's total population in 2000. (Additional details are available upon request. District-level model results without IPW adjustment are presented in Table A2, online appendix.)

Results

Table 1 provides descriptive statistics for the individual birth sample and the sibling subsample separately for each period. Among individual preboom births, 29% and 23% of birthing parents smoked before and during pregnancy, respectively. Before pregnancy, 5% and 45% of birthing parents were classified as underweight and as overweight or obese, respectively; 19% and 14% experienced low and high weight gain during pregnancy, respectively. Nearly 77% received prenatal care in the first trimester, and 71% received adequate prenatal care. Only 1% of births were very low birth weight, 6% were low birth weight, and 8% were preterm. Regarding birthing parents’ demographic characteristics before the boom, mean age was 28 years, 41% had a high school diploma or less, and equal percentages had some college education or a college degree (at roughly 29%). More than 85% of birthing parents were non-Hispanic White, and 60% were married. Approximately 42% of births were first births.

By 2012, there were statistically significant (p < .05) declines in smoking before and during pregnancy, prepregnancy underweight, high gestational weight gain, very low birth weight, and preterm birth, alongside significant increases in prepregnancy overweight/obesity, first-trimester prenatal care, adequate prenatal care usage, mean birth weight, and mean gestation length. In 2012, birthing parents were significantly older and more likely to have a college degree and a first birth, but were slightly less likely to be non-Hispanic White.

The sibling subsample is qualitatively similar, although birthing parents in this sample display higher smoking rates and lower rates of first-trimester prenatal care initiation, adequate prenatal care usage, and lower educational attainment, particularly after the boom.

Treatment and Comparison District Patterns

Table 2 compares district-level descriptive statistics across treatment and comparison districts. Starting with preboom pregnancy and birth outcomes, we find no statistically significant differences in birth weight or gestation length. However, we find a few differences in pregnancy and parental characteristics. Preboom, birthing parents in treatment districts (all p < .05) were less likely to smoke before pregnancy (30% vs. 33%) and more likely to initiate prenatal care in the first trimester (79% vs. 75%). However, they were less likely to receive adequate prenatal care during their pregnancy (71% vs. 75%), slightly older (27.3 years vs. 26.8), less likely to have a high school diploma or less (41% vs. 47%), and more likely to have a college degree (30% vs. 23%). They also had lower parity, with more reporting the focal birth as their first birth (42% vs. 40%). After the boom, only the educational attainment and first-trimester prenatal care initiation gaps between the treatment and comparison districts remained statistically significant. Yet, after the boom, birthing parents in treatment districts were less likely to be classified as overweight/obese relative to birthing parents in the comparison districts (48% vs. 50%; p < .05).

Before the boom, treatment and comparison districts showed no statistically significant economic differences (i.e., average income, homeownership, average rents, poverty rates, and the proportion working in agriculture, forestry, and fishing) and few demographic differences (i.e., no differences in the proportion non-Hispanic White, age distribution, marriage and cohabitation levels, college enrollment). The lone exceptions were that treatment districts had a significantly greater population density (837 vs. 462; p < .05) and a higher proportion of adults with a college degree (21% vs. 17%; p < .01).

Postboom, mean district income was higher in the treatment districts ($52,054 vs. $45,055), largely because income declined in the comparison districts. By 2012, royalty income increased in both the treatment and comparison districts but was significantly greater in the treatment districts ($1,694 vs. $1,044). Earnings were also marginally greater in the treatment districts ($41,938 vs. $37,941; p < .10). Postboom, property sales income was significantly greater in the comparison districts ($11,126 vs. $2,018). All other economic and demographic patterns remained similar.

On average, 97% of the treatment districts’ land is above the HYPE, whereas 91% of the comparison districts’ land is above the Periphery. Treatment and comparison districts are similar in geographic size, but comparison districts are significantly more likely to be rural (78% vs. 57%). Treatment districts are more likely to have coal deposits (64% vs. 44%) and, thus, greater exposure to past coal booms and busts but lower mineral rights ownership (given that banks and other companies bought landowners’ mineral rights during the nineteenth century; Davis 2002).

Environmental differences largely favored treatment districts before the Marcellus Shale boom. In 2007, a greater population proportion in the treatment districts had public water access and fewer roads recently added, as well as lower traces of antimony, arsenic, and selenium pollutants detected in the area's public water. Postboom, however, treatment districts had significantly higher traces of antimony and barium, whereas comparison districts had significantly higher traces of arsenic and selenium. Traffic was also significantly higher in comparison districts after the boom.

These descriptive results demonstrate the economic and demographic similarity of the treatment and comparison districts. Further, the environmental data demonstrate that metal pollutants were present in public water sources (and frequently at higher levels in the comparison districts) before the boom, which aligns with the region's mining and industrial history. These preboom characteristics support our quasi-experimental design by indicating the comparability of the treatment and comparison districts.

Reduced-Form Estimates

We next estimate a reduced-form, generalized DiD model to examine whether location above the HYPE, without instrumenting for income, predicts change in pregnancy and birth outcomes. Each cell in Table 3 presents the coefficient (ν) for being in a HYPE district postboom from separate models predicting each pregnancy and birth outcome using longitudinal data and controlling for period, district, and region FE. Location above the HYPE postboom is associated with a higher likelihood of adequate prenatal care use in the individual birth sample (ν = 0.037; p < .01) and a lower likelihood of low birth weight in the sibling sample (ν = −0.027; p < .01). All other associations are null or marginal.

Validating the Instrument

We next test whether PA districts’ Marcellus HYPE location predicts growth in district-level income between 2007 and 2010/2011. Panel A in Table 4 shows the first-stage π coefficients for location above the HYPE in the postboom period for all PA Shale districts and by initial poverty. PA HYPE districts have, on average, an additional $1,825 (p < .05) in mean district income in 2010/2011 relative to Periphery districts. The first-stage F statistic for all districts (16.56) meets the standard criteria (F > 10) for identifying a strong instrument (Andrews et al. 2019). The average district income increase was larger ($2,141 vs. $1,551) and was significant (p < .05) postboom only in initially high-poverty districts. The F statistics for mean district income for initially low-poverty (9.79) and high-poverty (7.27) areas approach 10 but do not exceed it. Thus, the stratified DDIV results could be subject to bias and less precise.

We next conduct two falsification tests of our identification strategy. First, among NY Shale districts, where hydraulic fracturing is banned, Marcellus HYPE location does not predict significant gains in total mean income. Second, after we exclude all PA HYPE districts, Marcellus Periphery location does not predict significant income gains. These first-stage results provide robust evidence that having 51+% of land above the HYPE in PA, conditional on initial income and district traits, is a valid instrument for income gains.

In supplemental models, we predict whether PA HYPE districts have higher postboom earnings or labor force participation. As shown in Table A3 (online appendix), mean district earnings (recorded in tax data) significantly increased by $1,408 across all districts (p < .01) and by $1,487 in initially low-poverty districts (p < .05). The ACS data reveal that earnings increased by $791 for workers aged 25 or older with a high school diploma (p < .05) in HYPE districts postboom, driven by the gains in initially high-poverty districts (ϖ = $857; p < .05). Regarding labor force participation, we find no statistically significant changes in the unemployment rate or the proportion of adults working full-time, as shown in Table A4 (online appendix). Yet, the proportion of adults not engaged in any hours of paid work declined slightly (ϖ = −0.0088; p < .05) in HYPE districts postboom.

Observational and Causal Income Models

Panel B of Table 4 presents both the simple DiD models examining the observed association between mean district income and our outcome measures and the second-stage DDIV results (β) using the total birth sample. The observational ordinary least-squares (OLS) models show that a $1,000 increase in mean district income is associated with only a 0.02-percentage-point increase in smoking during pregnancy (p < .05). The DDIV results are notably larger in magnitude and substantively different. The DDIV coefficients are generally in the expected direction, but only the effect on adequate prenatal care use is statistically significant: a $1,000 increase in income leads to a 1.8-percentage-point (p < .01) increase in the probability of receiving adequate prenatal care. Although this positive association occurred for both initially low- and high-poverty districts, income gains are statistically significantly associated with the probability of receiving adequate prenatal care only in initially high-poverty districts (β = 0.014; p < .05).

To control for time-invariant, unobserved parental traits that could bias the effect of increasing income, we restrict the sample to nontwin siblings born to the same birthing parent—one born before the boom and one born after—and living in the same district across births. Table 5 provides the second-stage DDIV (β) results for the pregnancy and birth outcomes in the paired sibling sample. Despite the significant reduction in sample size and dependent variable variance, we arrive at a few significant effects of plausibly exogenous income gains, although not all align with population health goals. In the full sibling sample, a $1,000 income increase is associated with a 1.5-percentage-point reduction in low birth weight (p < .05) and a 0.5-percentage-point increase in being underweight before pregnancy (p < .05). No statistically significant associations exist between income gains and pregnancy and birth outcomes in initially low-poverty districts, but a $1,000 gain is associated with a 1.3-percentage-point reduction in low birth weight (p < .05) and a 2.5-percentage-point reduction in first-trimester prenatal care use in initially high-poverty districts.

Robustness Checks

In supplemental models, we test the consistency of our results across different specifications. First, we test the timing of the income effect to have no lag or a two-year lag (i.e., with 2011 and 2013 births, respectively). As shown in Table A5 (online appendix), we find no significant associations with no lag but statistically significant income effects with a two-year lag for adequate prenatal care (β = 0.016; p < .05) and low birth weight (β = −0.005; p < .05). Second, we estimate additional stratified models to examine additional heterogeneous income effects by parity and other district-level characteristics (not shown but available upon request). In the individual birth sample, the income effects on adequate prenatal care are significant among first births (β = 0.021; p < .05) but not later births, and they are concentrated among births in districts with a history of coal mining (β = 0.036; p < .05) and a larger preboom agricultural industry (β = 0.019; p < .01). In the sibling sample, we confirm that the reduction in low birth weight was greater in districts with less affluence initially (β = −0.014; p < .05), as well as in rural districts (β = −0.009; p < .05) and districts with less public water access (β = −0.008; p < .05). Interestingly, the income effect on prepregnancy underweight in the sibling sample was greater in districts that were initially less affluent (β = 0.006; p < .05).

Falsification Tests

Our final models, shown in Table 6, test the DDIV assumptions. First, we use five-year ACS data to test whether our results could be biased as a result of selective migration. As shown in panel A, we estimate whether HYPE districts significantly differed postboom in their log number of women aged 15–44, population percentage aged 15–44, the percentage of women aged 15–44 who are married, the percentage of household heads in cohabiting relationships, the percentage of the adult population enrolled in higher education, and mean property sales income using FE DiD models with district-level data. Across all PA Marcellus districts, postboom Marcellus HYPE location generally does not predict differences in these indicators, although postboom HYPE districts had slightly higher rates of cohabitation (ϖ = 0.0017; p < .05). The rise in cohabitation was statistically significant in low-poverty districts (ϖ = 0.0027; p < .05) but not in high-poverty districts. Thus, selective migration is unlikely to explain our findings.

Second, we explore whether selection into fertility changed in HYPE districts after the boom using birth certificate data. As shown in panel B of Table 6, we predict the log number of total births, the general fertility rate (for the population aged 15–44), the percentage of first births to birthing parents aged 15–44, the percentage of births to married parents, the age distribution of birthing parents, and birthing parents’ educational attainment. Marcellus HYPE location postboom is never statistically significant, making it very unlikely that our estimated effects of income gains on pregnancy and birth outcomes are due to changes in the composition of the birthing population.

Finally, we test the exclusion restriction by examining whether contextual and environmental characteristics were significantly different in HYPE districts postboom. We predict formaldehyde air pollution, multiple metal contaminants detected in public water sources, the number of PA EPA compliance violations, traffic per capita, miles of local roads added in the last year, mean gross rent, and the proportion of owner-occupied housing. As shown in panel C of Table 6, postboom Marcellus HYPE location is not significantly associated with these outcomes in the full district sample or in the initially high-poverty districts, where we find statistically significant income effects. The only significant results are within initially low-poverty Marcellus HYPE districts. Postboom, they have significantly higher antimony levels in public water (ϖ = 0.000028; p < .01) and more local roads (ϖ = 109.7; p < .05). Hence, we find minor evidence that the lack of an income effect in initially low-poverty HYPE districts could be due to water pollution offsetting the effects of income gains.

Discussion

In this study, we leverage the Marcellus Shale boom to provide new evidence about the causal impact of community-level income gains on pregnancy and birth outcomes using birth certificate data for all births, as well as a sibling sample to control for time-invariant characteristics of birthing parents. We also identify heterogeneous income effects based on preboom area poverty rates.

Our key finding is that a plausibly exogenous $1,000 gain in mean area income is associated with a 1.5-percentage-point decline in low birth weight in the sibling sample, with improvements also significant in initially high-poverty districts. This small but significant effect aligns with all quasi-experimental income studies that test whether income gains reduce low birth weight (Chung et al. 2016; Hamad and Rehkopf 2015; Hoynes et al. 2015; Markowitz et al. 2017; Strully et al. 2010) and with several macroeconomic studies that find economic recessions increase low birth weight (Alessie et al. 2018; Bozzoli and Quintana-Domeque 2014; De Cao et al. 2022; Kohara et al. 2019; Olafsson 2016).

Given that prior macroeconomic scholarship used unemployment rates, we also estimated DDIV models using unemployment rates as the causal variable and found that instrumented change in unemployment rates did not lead to significant changes in pregnancy or infant health outcomes (not shown but available upon request). Yet, these DDIV unemployment estimates are likely biased because Marcellus Shale HYPE location is a poor instrument for unemployment change (Table A4, online appendix). Regardless, these DDIV unemployment models suggest that our causal estimates of area-level income gains on low birth weight are not reducible to unemployment declines. Further, the declines in low birth weight are notable given the suggestive evidence (p < .10) that formaldehyde air pollution and arsenic water pollution levels increased in HYPE districts postboom (Table 6) and that birth weight is lower in places with greater air pollution (Bell et al. 2007; Currie et al. 2009; Maisonet et al. 2001; Parker et al. 2005; Savitz et al. 2014) and arsenic exposure (Hopenhayn et al. 2003; Rahman et al. 2017). Finally, documenting even a small causal effect of income on low birth weight is important because although this outcome is rare, children with low birth weight are at greater risk of cognitive delays (Figlio et al. 2014), lower educational attainments (Behrman and Rosenzweig 2004), and poor labor market outcomes (Johnson and Schoeni 2011).

Our finding that income does not have a causal effect on preterm birth aligns with one quasi-experimental study (Hamad and Rehkopf 2015) but contrasts with others (Chung et al. 2016; Hoynes et al. 2015). These varying results suggest that the income effects on gestation length and preterm birth are less consistent than those on low birth weight.

We also explore several prepregnancy and pregnancy health outcomes to clarify the mechanisms by which income gains reduce low birth weight and, in some studies, improve other birth outcomes. In the individual sample, a plausibly exogenous $1,000 gain in mean area income is associated with a 1.8-percentage-point increase in receiving at least adequate prenatal care, with significant improvements also occurring in initially high-poverty districts. Our results provide new evidence that the previously observed deficits in adequate prenatal care receipt among low-income women (Green 2018) are at least partially causal.

The sibling analyses yield somewhat unexpected findings. Specifically, a plausibly exogenous $1,000 gain in mean area income leads to a 0.5-percentage-point increase in prepregnancy underweight in all Marcellus study districts and a 2.5-percentage-point decline in the likelihood of receiving prenatal care during the first trimester in initially high-poverty districts for the later birth (p < .05). The increase in underweight could reflect the previously observed population-wide negative correlation between income and body weight among mothers (Martin and Lippert 2012) or potentially declines in birthing parents’ stress or increases in water pollution not measured well in our falsification tests. Quasi-experimental research on first-trimester prenatal care receipt yielded mixed findings: Chung et al. (2016) and Hoynes et al. (2015) found that income gains via universal transfers and the EITC (respectively) led to earlier prenatal care initiation and increased prenatal care adequacy, whereas Markowitz et al. (2017) found little evidence that state-level EITC generosity impacted first-trimester care initiation. Although none of these studies examined heterogeneous effects by parity, our sibling analyses suggest increasing income lowers first-trimester prenatal care use for later births. Future research should directly study this potential difference by parity.

Interestingly, the Marcellus Shale boom immediately followed the Great Recession. On the one hand, this timing is beneficial for our analyses because the income gains resulting from the Marcellus Shale boom cannot be mistaken for broader macroeconomic improvements. On the other hand, it is challenging because the Marcellus Shale boom largely stemmed the tide against further income declines, as shown in Table 2. Thus, it is difficult to discern whether our causal estimates reflect an absolute effect of income gains or a relative income effect—namely, the importance of income gains (or income stability) when neighboring areas or states experience macroeconomic declines.

Limitations and Strengths

This study has several important limitations. First, the study sample is necessarily restricted to the population living above the Marcellus Shale geological formation, which largely consists of northern and western PA. Thus, our study findings primarily reflect the experiences of non-Hispanic White, U.S.-born birthing populations living outside major metropolitan areas. Nevertheless, this group reflects the demographic profile of more than a quarter of all U.S. births in 2020 (authors’ calculations), suggesting that our findings have relevance for a sizable share of the birthing population. Second, we cannot test the DiD parallel trends assumption (condition 1) because it pertains to unobserved, potential outcomes. We can only observe the relative similarity of treatment and comparison districts in the preboom period (Table 2). If the assumption of parallel trends is violated, then our estimates would be biased. Third, we rely on birth certificate data to capture health outcomes and behaviors (some of which are self-reported), which are subject to measurement error and other forms of bias (Bodnar et al. 2014). Sibling models are more sensitive to measurement error and thus might fail to detect real income effects (Collischon and Eberl 2020). However, we have no reason to believe that the degree or direction of measurement error on birth certificates systematically varies by Marcellus Shale location. Fourth, we can measure income only at the district level, preventing us from precisely estimating the impacts of individual-level income gains on birth outcomes and pregnancy-related health. Moreover, analyses of aggregated data can lead to ecological fallacies and are vulnerable to spatial autocorrelation and modifiable area unit problems (i.e., understandings differ with different aggregations; Buzzelli 2020; Kong and Zhang 2020), although we partially address spatial autocorrelation by including region FEs. Fifth, because we conduct multiple tests for several dependent variables, we risk incorrectly detecting associations with area-level income and health behaviors/outcomes where none exist (i.e., type I error). If we lower our threshold for statistical significance to account for the number of dependent variables (i.e., 13), then only the reduced-form results (Table 3) for adequate prenatal care among all births and low birth weight among siblings remain statistically significant (p < .01). On the other hand, we did not randomly select our outcomes of interest, and our models are theoretically well justified. Further, because FE models are more conservative, we also might have failed to detect existing associations between income and health behavior/outcomes (i.e., type II error). Sixth, HYPE location might have affected pregnancy and birth outcomes via channels other than income gains, although we do not find a significant effect of HYPE location on increasing water or air pollution across all districts and find a significant effect only on antimony levels in initially low-poverty districts. Finally, the DDIV results for the stratified models might be biased, given that the instrument did not pass the stringent standard for a strong estimate in initially high- and low-poverty districts. However, the direction and magnitude of the statistically significant associations found for the initially high-poverty districts are consistent with the estimates for the full district sample, for which our instrument is strong.

Important study strengths balance these limitations. We capture the universe of births in the relevant geographic areas and measure a wider array of reproductive health behaviors and outcomes than other studies. Further, most prior quasi-experimental research on income and health in the United States was limited to studying the effects of income gains for low-income or low–socioeconomic status populations. In contrast, we study the entire income distribution of the birthing population in the study site. Thus, our findings provide important insights into how universal income interventions might impact reproductive health outcomes in similar populations. Given prior theoretical discussions questioning the potential offsetting effects of pollution and income gains during macroeconomic booms, we examine multiple pollution measures across HYPE and Periphery districts and their change after the boom. Interestingly, we do not find a strong association between HYPE location and environmental change, including rising pollution levels, which speaks to the validity of the instrument for identifying income gains. The geographic area of predicted economic benefits overlaps with, but does not perfectly correlate with, Marcellus Shale development and the location of associated environmental hazards. We strongly recommend that future research use alternative geographic indicators to identify the total effect of actual Marcellus Shale development on pregnancy and birth outcomes. Finally, we identify and leverage a robust IV for macroeconomic income gains to provide new causal evidence about an important policy question.

Conclusions

This study makes three important contributions. First, we examine whether community-level income gains affect infant health via an effect on pregnancy health behaviors before and during pregnancy. Second, we study the effects of income gains across the income spectrum while testing for heterogeneous income effects by preboom local poverty levels. Finally, we leverage sibling models to control for time-invariant traits of birthing parents to arrive at our most rigorous estimates of income gains.

In the sibling DDIV models, we find that a $1,000 increase in mean area income results in a 1.5-percentage-point decrease in low birth weight across the full income distribution of mothers. This effect is substantively similar and statistically significant in initially high-poverty districts. The effect of income gains on low birth weight does not appear to operate through birthing parents’ smoking, prepregnancy weight status, or gestational weight gain. Hence this article adds evidence to the growing literature that increasing income leads to a significant improvement in birth weight independent of measured health behaviors (Almond and Currie 2011; Bussières et al. 2015). Data limitations prevent us from exploring whether income gains (1) improved other, unmeasured pregnancy-related health or health behaviors that affect the placenta's nutrient environment (e.g., diet quality); or (2) reduced chronic stress, which can restrict blood flow and, thus, nutrients and oxygen to the fetus (Dunkel Schetter 2011). We encourage future research to pursue these possibilities with appropriate data. Ultimately, our findings suggest that interventions to increase mean area income would be an effective intervention to decrease the proportion of infants with low birth weights—a critical risk factor in the intergenerational transmission of poverty.

Acknowledgments

We are grateful for support from the Russell Sage Foundation (#83-14-16); the Penn State Social Science Research Institute; the Penn State Population Research Institute, which received center grant funding from the National Institutes of Health (P2CHD041025; PI: J. E. Glick); the Penn State Edna Bennett Pierce Prevention Research Center, which received training grant funding from the National Institutes of Health (T32 DA017629; MPIs: J. Maggs and S. Lanza); the Society of Family Planning Research Fund (award SFPRF13–CM4); and the Centennial Scholars/Clinicians Program of the School of Medicine and Public Health at the University of Wisconsin–Madison. The opinions expressed here are those of the authors, not the granting agencies. The birth certificate data were supplied by the Bureau of Health Statistics and Registries, Pennsylvania Department of Health, Harrisburg, Pennsylvania. The Pennsylvania Department of Health specifically disclaims responsibility for any analyses, interpretations, or conclusions. We sincerely thank Arthur Friedson of the New York State Department of Taxation and Finance and Jim Ruberton and David Oliver of the Pennsylvania Department of Health for their assistance in helping us acquire some of the data utilized in this project. We thank Yosef Bodovski and Lisa Ryan for their technical assistance and Elizabeth Ananat for her methodological guidance. We thank the Demography reviewers who strengthened the manuscript. All errors and omissions are our own.

Note

References