Drinking Water Contaminant Concentrations and Birth Outcomes

Abstract

Previous research in the US has found negative health effects of contamination when it triggers regulatory violations. An important question is whether levels of contamination that do not trigger a health-based violation impact health. We study the impact of drinking water contamination in community water systems on birth outcomes using drinking water sampling results data in Pennsylvania. We focus on the effects of water contamination for births not exposed to regulatory violations. Our most rigorous specification employs mother fixed effects and finds changing from the 10th to the 90th percentile of water contamination (among births not exposed to regulatory violations) increases low birth weight by 12% and preterm birth by 17%.

INTRODUCTION

Most residential settings in the U.S., including homes, apartments, and mobile park homes, are served by community water systems (CWS), representing 94% of the nation’s population (United States Environmental Protection Agency [EPA], 2019a). The Safe Drinking Water Act (SDWA) was enacted in 1974 to protect communities from unsafe drinking water in CWS. The act currently regulates over 90 contaminants (e.g., lead, arsenic, disinfection byproducts). Regulatory drinking water violationsⁱ under SDWA occur when a maximum contaminant level (MCL), action level, or maximum residual disinfectant level (MRDL) is exceeded (EPA, 2018).ⁱⁱ Despite successfully reducing drinking water contamination in CWS, up to 45 million people were affected by health-based violations from 1982 to 2015 (Allaire et al., 2018). In 2017, 22 million people, representing 7% of the U.S. population, relied on a CWS with a health-based drinking water violation (EPA, 2019a). Recent lead-related drinking water crises in Newark, New Jersey (Dave & Yang, 2022) and Flint, Michigan (D. S. Grossman & Slusky, 2019) have highlighted the importance of these regulations to protect public health.

The vast majority of the literature studying SDWA has focused on non-compliant drinking water. SDWA violations disproportionately impact minority and low-income communities (Allaire et al., 2018; McDonald & Jones, 2018); egregious cases of water contamination, such as the case of Flint, have major community and health impacts (Christensen et al., 2023; Danagoulian et al., 2022; Danagoulian & Jenkins, 2021; D. Grossman & Slusky, 2019), and health-based violations adversely impact infant health (Currie et al., 2013). Public reporting of these health-based violations is an important mechanism to encourage compliance within the law (Baker et al., 2022; Bennear et al., 2009; Bennear & Olmstead, 2008; Grooms, 2016). Prior research has shown that this public reporting has resulted in fewer violations over time (Baker et al., 2022; Bennear & Olmstead, 2008), however, other work using sampling concentrations has suggested that systems may not return to compliance following health-based violations (Grooms, 2016). Consumers appear to respond to public reporting of violations with avoidance behavior, such as bottled water purchases (Allaire et al., 2019; Marcus, 2022; Zivin et al., 2011). In this paper, we provide new evidence of the health impacts of drinking water quality that is not subject to a health-based violation, which we call compliant drinking water contamination.ⁱⁱⁱ

A large and growing literature focused on the health impacts of air pollution that is compliant with the Clean Air Act has shown that even low levels of air quality can still have meaningful impacts on health, labor market outcomes, and cognition (Currie & Walker, 2019). Much like air quality, where particulate matter is actually a grouping of many air pollutants and not isolating particular pollutant constituents, we propose to measure this compliant contamination using a water quality index. Using an index addresses the multiple comparison issues we would face if we separately estimated the effects of each contaminant, or the multicollinearity issues we would face if we attempted to include the over 90 contaminants that are regulated under SDWA in the same regression.^iv Furthermore, while current toxicological and epidemiological studies of the associations between water quality and health generally reflect a narrow focus on single contaminants (Murphy et al., 2012), researchers have noted that low concentrations of contaminants may not be associated with health outcomes when considered individually but may produce effects when examined as mixtures (Gennings et al., 2018). For these reasons, we create an overall water quality index using contaminant concentrations and regulatory standards, creating a single measure for testing our hypothesis. Following prior literature, we also construct an aggregate “chemical only” measure (Currie et al., 2013). Our aggregation strategy is discussed further in the section “Water Quality Index Construction Details.”

Scientific literature assessing the association between each of the more than 90 contaminants regulated by SDWA and birth outcomes is lacking beyond a subset of contaminants.^v Still, our overall water quality index may not be specific to our outcomes of interest (e.g., low birth weight, preterm birth), and so we create a second index specific to five contaminants with the most associational evidence linking exposure to these contaminants with the outcomes studied in this paper.^vi While a reproductive-specific index can be supportive of the theoretical link between water quality and birth outcomes, lack of empirical evidence for the other contaminants regulated by SDWA is not indicative of no biological effect. Our overall index can be used for hypothesis generation regarding other contaminants impacting birth outcomes. Furthermore, and we consider this most important, the overall measure can be validated on other populations (children, older adults) and health outcomes (cancer), and allow for a useful general index that can be used across systems and susceptible populations, analogous to the EPA’s Air Quality Index (AQI).

We employ the universe of address-specific birth records in Pennsylvania from 2003 to 2014 geocoded to CWS boundaries, and match births to water samples measured at the CWS level that underlie regulatory violations during gestation.^vii Using these novel data on contamination concentrations summarized in an omnibus index, we find evidence of precise harmful effects of water contamination on birth outcomes after removing regulatory violations. Our empirical strategy relies on within-CWS variation (CWS fixed effects) such that our approach ignores correlations between water quality and infant health driven by cross-sectional differences. We also include birth year-month fixed effects to control for seasonality and state-wide time shocks. Using fixed-effects models that isolate only within-mother across-sibling variation, we find that changes in overall contamination from the 10th to the 90th percentile—conditional on no violation—leads to an increase in low-birth weight (LBW) of 12% and preterm birth (PTB) of 17%. Using a reproductive-specific index, we find slightly larger impacts of water quality on our birth outcomes—a 19% increase in LBW and 22% increase in PTB. And in sensitivity analyses, we remove births with samples above the regulatory threshold, which is more stringent than how regulatory violations are determined and continue to find a strong impact on birth outcomes.^viii

Given the recent findings of fertility effects of contamination and the consequent issues of selection (e.g., for lead, see D. S. Grossman & Slusky, 2019, and Clay et al., 2021), we additionally fit CWS panel models of births per person served by the system on various measures of recent water contamination. Our estimates suggest small negative effects of contamination on fertility. We also analyze whether mothers move out of their current CWS in response to contamination. While maternal response would be of concern for our identification strategy, we find no significant effects on mobility, which is consistent with mothers not being aware of variation in contamination that does not trigger regulatory violations. Moreover, our results are robust to excluding mothers who move across pregnancies. Finally, we use the bounding methods of Altonji et al. (2005) and Oster (2016) to assess the importance of selection on unobservables, finding in most specifications that it would have to be greater than selection on observables to conclude that there are no damaging effects of water contamination.

Our work contributes to a well-established fetal origins literature in economics that links shocks, such as pollution, experienced in utero to poorer birth outcomes (Almond et al., 2018). Extensive research in economics establishes that poor birth outcomes lead to worse short- and long-term health and human capital outcomes (see, e.g., Almond et al., 2005; Black et al., 2007; Currie & Moretti, 2007; Oreopoulos et al., 2008; Royer, 2009). Air pollution has attracted particular attention, as findings of harmful effects appear robust, even at levels compliant with air advisory standards (Beatty & Shimshack, 2014; Currie et al., 2015; Currie & Walker, 2011; Knittel et al., 2016; Yang & Chou, 2018). Although federal and state regulation of public drinking water in the U.S. has recently come under scrutiny, in part because of the highly publicized Flint water crisis (Allaire et al., 2018; D. S. Grossman & Slusky, 2019), few papers have studied how the quality of drinking water in the U.S. affects health outcomes using quasi-experimental designs (Keiser & Shapiro, 2019b).^ix The most closely related study to our work is Currie et al. (2013), who found that exposure to a water quality violation during gestation increases the chance of low birth weight and preterm birth, based on 1997 to 2007 data from NJ.^x Importantly, their estimates suggested that non-compliant contamination has negative health consequences. Our analysis addresses health impacts of compliant contamination levels.

Consumers currently learn about the quality of their water either by reviewing the annual Consumer Confidence Reports (CCR), which contain the average concentrations for contaminants in their water system, or through notifications about extreme contamination events that constitute regulatory violations. We argue that, based on the reports as currently designed, consumers are unlikely to respond to concentrations that are compliant with current regulations. Given our findings of adverse birth outcomes from overall contamination concentrations removing regulatory violations, using our index to describe these risks for consumers could greatly improve the available information. Specifically, we propose using our overall water quality index to create a letter grade as a summative measure of water quality that could be included in CCRs to provide more actionable information to consumers. In the section “Discussion,” we evaluate this potential policy, finding that it could lead to economically meaningful improvements in public health as measured by birth outcomes. We find that a policy improving water quality for births served by systems with letter grades B to F to the average quality of systems with letter grade A would reduce the number of LBW infants by 3,960 and reduce the number of PTB infants by 6,930 over 11 years. Such a policy could save around $280 million in medical costs for these avoided adverse birth outcomes. This analysis could serve as a foundation for future refinements to providing a summary measure of water quality for CCRs.

DATA AND SUMMARY STATISTICS

To conduct our analysis, we assembled data on birth outcomes from confidential birth records, temperature, weather, and pollution data derived from weather stations and Toxics Release Inventory (TRI) locations and emissions, and water quality and service areas from community water systems (CWS). We describe each of these in turn. We join these three data sources together using spatial and temporal attributes.

Birth records and control variables

Data on birth certificates spanning the years 2003 to 2014 were provided by the Pennsylvania Department of Health. These data contain detailed information on birth outcomes, clinical characteristics, and demographic characteristics. We describe in detail our control variables obtained from these birth records in the section “Empirical Methods.” Our two main outcomes of interest are binary indicators for low birth weight^xi and preterm birth.^xii In addition, we study two birth outcome measures often used by epidemiologists that better capture growth rates during pregnancy (see e.g., Govarts et al., 2018). These measures are small for gestational age (SGA) and term birth weight (TBW; in grams).^xiii

We obtained daily weather statistics from the website of Schlenker (Schlenker & Roberts, 2009). These data provide estimates of daily precipitation and maximum and minimum temperatures for a grid of points across the United States. We match mothers to their closest grid point using exact location of birth and calculate a set of temperature and weather controls over the mom’s pregnancy. We use these as control variables in our regression specifications described in the section “Empirical Methods.” Note that weather and temperature have been found to be important in previous studies relating health to drinking water contamination (Currie et al., 2013), as well as to air pollution (Beatty & Shimshack, 2014; Knittel et al., 2016). Finally, we also obtained Toxics Release Inventory (TRI) controls (locations and emissions) from the U.S. Environmental Protection Agency (2013). Using maternal address, we calculate the number of TRI facilities operating within 1, 3, and 5 kilometers in the year the infant was born, and the reported total onsite, offsite, and overall releases for TRI facilities weighted by the squared inverse distance between the facility and the mother’s residential location.^xiv

Community water system data

Our data on CWS comes from two sources: the U.S. EPA and the Pennsylvania Department of Environmental Protection (PADEP). We obtained Safe Drinking Water Information System (SDWIS) data from the EPA through a Freedom of Information Act (FOIA) request, which contains information on CWS drinking water violations as well as system characteristics such as recently reported population served. These data include violation reports 2000 to 2014. We observe intervals of time (“compliance periods”) for these violations that are primarily reported quarterly; we define a birth to have been exposed to a violation if its gestation period overlaps with the compliance period of a violation for at least one day.^xv Further discussion of these data is provided in the Online Appendix Section “Data Construction Details.” For our fertility analysis, we make use of the system population served measure from SDWIS, both to normalize the dependent variable across systems (we divide births per month by system population served to calculate a monthly fertility rate), and as weights for the regression model.

Most importantly for our study, we obtained public drinking water sampling information for CWS from PADEP.^xvi These data contain the sample results underlying Pennsylvania’s SDWA violations recorded in the SDWIS data. This drinking water sampling data contains the date of the sample, the contaminant tested, and the result of the test. These data are recorded at a much higher frequency than water quality violations, and they provide information on water contamination at levels that do not trigger a violation.^xvii

These data were augmented with a shapefile obtained from the Pennsylvania Spatial Data Access (PASDA) website containing the service area boundaries of CWS.^xviii Figure 1 is a map of the CWS service area boundaries for the systems in our data. As is apparent, CWS vary substantially in size. There are 1,768 CWS in the boundary file; the average system reports serving about 5,700 people, but the smallest systems (10th percentile) serve only around 60 each, while the top 10 largest systems serve 41% of people served in this data (the Philadelphia Water Department alone serves 16%, reporting about 1.6 million people served).

Figure 1:

Map of Pennsylvania and 1,768 CWS service areas.

Most water quality samples are measured at entry points to the distribution system, or somewhere within the distribution system (e.g., at the tap for lead and copper samples).^xix In the simplest case, a treatment facility pipes water from a single source through its operations, then out into a distribution system. In more complicated cases, a distribution system can be served by several source locations.^xx Analysis requires us to aggregate up to the level of a complete CWS, as that is the level at which we can match the water sample results to birth records using the water system boundary shapefile and geocoded mother addresses.^xxi

Water quality index construction details

In this subsection, we describe how we construct our measure of drinking water quality. For each CWS, contaminant, and day, we use the samples of the contaminant on that day to calculate the average contamination in the CWS’s distributed water. This measure of average daily contamination is in the units in which the contaminant is measured, typically parts-per-billion (ppb) or parts-per-million (ppm). Then, for each birth and contaminant, we average this daily measure of contamination for the CWS where the mother’s residence was at the time of birth. The average is taken across all samples within the birth’s gestation period.^xxii

The raw Pennsylvania sampling data contains the 94 contaminants regulated under the SDWA.^xxiii We also obtained the health based regulatory violations from EPA by FOIA request which includes the maximum contaminant level (MCL), action level or maximum residual disinfectant level (MRDL) for each contaminant and whether the system received a violation during each compliance period. Instead of studying each contaminant separately, in our main analysis we aggregate by normalizing and averaging. Our goal is to construct a simple composite measure of water contamination, in order to avoid the statistical pitfalls of very noisy data combined with multiple hypothesis testing. The measure we choose to employ for this purpose is the average “result relative to MCL” (RRMCL).

$$RRMCL=\frac{SamplingResult}{MCL}$$ #(1)

This approach accommodates the fact that what is atypically large for one contaminant may not be for another contaminant. For example, the MCL for nitrate and nitrite are 10 ppm and 1ppm, respectively. In order to create a single composite measure of contamination, we divide nitrate samples by 10, divide nitrite samples by 1, and divide all other contaminants by their respective MCL.

We then average over all contaminants sampled during pregnancy for each infant to have a measure of “overall” contamination based upon contaminants regulated by the SDWA, weighting contaminants (not samples) equally. Given the large number of chlorine samples (over 4.8 million samples; the next most common in our data is lead at 93,000 samples), we take a 10% random subsample of the chlorine samples before we build the index, to ensure computational feasibility.

Sample nondetects, i.e., samples with results that cannot be distinguish from zero, are recorded as zero in the data and included in our index calculations. This will introduce measurement error into our index measure. In general, when the contamination in the water sample is small, methods that estimate the level of contamination can have substantial relative error. We take the position expressed in the EPA internal document (EPA, 2006) that measured samples below the MCL still contain (albeit noisy) information. We further discuss this issue in Online Appendix Section “Non-Detection and Limits of Measurement.”

We make two notes about our “average RRMCL” construction process. First, it maximizes sample size, but the measure is not composed of the same contaminants across births. For example, one birth may only be matched to total coliform samples, while another birth may include nitrate and arsenic samples. Each birth will have a measure of average RRMCL, but the contaminant used in the first birth is just total coliform, while the RRMCL in the second birth is a simple average of the RRMCL for nitrate and RRMCL for arsenic. To be clear, this is just a stylized example for illustration: 2.47% of births in our data have only one or two unique contaminants measured. Indeed, across births, an average of 39.31 unique contaminants are used to calculate our overall contamination index. Second, and as this example makes clear, the step in this process that likely makes the strongest assumptions is weighting the contaminants available for each birth equally. Moreover, while maximum contaminant level goals (MCLG) may be a more natural denominator given the focus on protecting health, MCLG are often zero, making them not as useful for aggregation. In addition to our main overall measure of contamination, we repeat this exercise, but removing coliform, heterotrophic bacteria, and turbidity. This additional measure we call a measure of “chemical” (i.e., non-bacterial) contamination. Splitting between overall and chemical contamination is consistent with previous literature (Currie et al., 2013). On average, 38.50 unique contaminants are used to calculate this chemical-only index; 3.05% of these births have only one or two contaminants in the index. Finally, we also construct a “reproductive index” measure, using five contaminants that have been found linked to reproductive health outcomes; further discussion of its construction is postponed to Section “Reproductive Health Water Quality Index.”

There is one important contaminant that we treat as an exception to Equation 1: total coliform. Specifically, we define the Result Relative to MCL (RRMCL) for total coliform, for each birth, as

$$RRMCL=\frac{Shareofpositivetests}{\mathit{max}\left\{\frac{2}{\left(\frac{Numberoftests}{\frac{273}{30}}\right)},0.05\right\}}$$ #(2)

In this expression, the ratio (Number of Tests) / ((273/30)) gives the approximate number of tests per month (per 30 days) over the gestation period, assuming full gestation length (273 days = 39 weeks). We use this expression for the RRMCL for this contaminant, since the effective MCL for coliform is 2 tests when the number of tests per month is less than 40, and 5% of tests when the number of tests per month is greater than 40. This implies that the MCL is the maximum in the denominator of the above expression. Indeed, this Equation can be interpreted as the generalization of Equation 1 for total coliform, given that the MCL for total coliform varies based on the number of tests conducted by the system (the number of tests required to be conducted depends on system size; see Bennear et al., 2009 for a detailed analysis of the Total Coliform Rule).

Taken together, our approach to defining a water quality index is the “sampling data analogue” of the drinking water violations measure employed in Currie et al. (2013). To see this, consider that we are combining across contaminants with very different properties and health effects by normalizing by their regulatory thresholds. This is very similar to Currie et al. (2013), who also grouped all contaminants into a single summary measure, weighting equally, but instead defined treatment as exposure to a regulatory violation during gestation. The advantage of our approach is that it allows us to use variation in contamination that does not trigger a health-based violation, which is important for providing policy-relevant evidence.

Throughout our analysis, we refer to and use what we call our “analytic sample,” which restricts births as follows. First, births in our analytic sample are not exposed to any MCL violations; second, they have some measure of contamination within their gestation period; third, they are not plural births; and fourth, their measured contamination does not exceed the 99th percentile across all births in the data for the specific contamination measure under study.^xxiv In addition, when we provide summary statistics, we remove singleton observations based on CWS and birth year-month fixed-effects, or mother and birth year-month fixed-effects, in order to calculate statistics for the exact samples used in our regressions.^xxv

Table 1 describes the distribution of our average RRMCL measures across births; throughout, births exposed to MCL violations are excluded. In preliminary analyses, we noticed that our RRMCL index measures are very right-skewed, even after removing violation-exposed births. Thus, we decided to trim out the top 1% of births based on RRMCL when we formed our analytic sample.^xxvi In panel A, we show the distribution of RRMCL across births, by contaminant group, after our trimming. To illustrate why we trimmed outliers, Panel B shows the distribution before trimming. Finally, since our main specification for identifying causal effects is a within-mother design, Panel C shows the distribution of these index measures after removing mother fixed-effects. We see that the standard deviation in Panel C is about two-thirds the size, as is the spread between the 10th and 90th percentiles.To facilitate comparing the overall, chemical, and reproductive index contamination index measures, in our regressions, we standardize such that a 1-unit change is equivalent to a change from the 10th to the 90th percentile of the average RRMCL, for the RRMCL measure of interest (overall, chemical, or reproductive). These movements are equivalent to about 1.875 standard deviations for the overall measure, about 1.625 standard deviations for the chemical measure, and about 1.889 standard deviations for the reproductive index (five contaminant) measure (Table 1 contains the quantiles of the distribution). We chose the 10th and 90th percentiles to standardize in order to simulate a large, though not impossible, change in contamination exposure.^xxvii For comparison, a movement from the 25th to the 75th percentile is about 33% of the 10th-to-90th percentile movement for the overall contamination measure, and about 31% for the chemical-only measure. The reader can rescale our estimates to represent effects of these movements by multiplying by 0.33 or 0.31, respectively.^xxviii Finally, our most rigorous specifications use within-mother across-sibling variation only; Table 1 includes statistics describing the spread of our contamination indices across births within mother in Panel C, in case the reader wishes to rescale our estimates based on that distribution instead.^xxix

Table 1.

Distribution of drinking water contamination experienced by Pennsylvanian births during gestation.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Contaminant	Mean	SD	10th	25th	50th	75th	90th	Max	Births
Panel A: Trimmed sample (births used in analysis)
All	0.08	0.08	0.02	0.03	0.05	0.08	0.17	0.50	1,216,132
Chemical	0.07	0.08	0.02	0.03	0.04	0.07	0.15	0.87	1,214,709
Repro. index	0.08	0.09	0.00	0.02	0.06	0.11	0.17	2.80	1,147,893
Panel B: Raw sample (including births not used)
All	0.09	2.24	0.02	0.03	0.05	0.09	0.18	1425.00	1,228,531
Chemical	0.09	2.24	0.02	0.03	0.04	0.07	0.17	1425.00	1,227,035
Repro. index	0.08	0.59	0.00	0.02	0.06	0.11	0.18	193.80	1,151,025
Panel C: Trimmed sample, after removing mother fixed effects
All	−0.00	0.05	−0.05	−0.02	−0.00	0.01	0.05	0.37	700,951
Chemical	−0.00	0.05	−0.05	−0.01	−0.00	0.01	0.04	0.54	699,656
Repro. index	−0.00	0.05	−0.05	−0.02	0.00	0.02	0.05	1.26	645,353

Notes: Table displays the average sample result relative to MCL (RRMCL) over gestational periods for births in Pennsylvania, 2003 to 2014. Throughout, the table restricts to non-plural births not exposed to MCL violations. Panel A trims the sample of births at the 99th percentile (separately for each contaminant group). This is the sample we use throughout the paper. Panel B shows the raw statistics for comparison. Panel C shows the distribution after removing mother fixed effects, which is relevant given our causal design. “All” and “Chemical” are our indexes described in the section “Water Quality Index Construction Details”; “Repro. Index” is our five-contaminant index described in the section “Reproductive Health Water Quality Index.”

In the Online Appendix Section “Discussion of Our Main Index Measures,” we further describe the correlation between our index measures and the samples for selected contaminants of interest in Appendix Table A1, as well as provide a full table listing all the contaminants included in each of our index measures (94 for the all index, 91 for the chemical index) in Appendix Table A2.

Reproductive health water quality index

EPA lists 13 contaminants associated with “reproductive difficulties” and another 6 associated with harms to “infants” (EPA, 2018). See Appendix Table A3 for the specific contaminants. We performed a literature review (see Appendix Section “An Alternative Reproductive Health Index of Five Contaminants Based on Data Availability and Prior Research” for discussion of this literature) and the epidemiology literature primarily focuses on arsenic, nitrate, atrazine, and disinfectant byproducts (Haloacetic Acids [HAA5], Total Trihalomethanes [TTHM]) and only nitrate and atrazine overlap with the EPA’s list. We also found some epidemiological literature associating tetrachloroethylene (PERC) and Di(2-ethylhexyl)phthalate (DEHP) and a recent study in Virgina included these plus lead, radon, and fecal bacteria (Young et al., 2022).^xxx We assessed the availability of samples for these contaminants to create an index that would have more uniform coverage in our sample population and limited to five with frequent sampling in Pennsylvania for our reproductive-health specific water quality index: arsenic, atrazine, nitrate, Di(2-ethylhexyl) Phthalate (DEHP), and Tetrachloroethene (PCE or PERC). On average, 3.27 unique contaminants are used in this more specific water quality index. This index could be used to target water alerts communications to pregnant persons and other studies could create water quality indices for specific health outcomes to target water alerts to specifically vulnerable populations.

Summary statistics

Summary statistics for births are provided in Table 2. In column (1) we display statistics for all non-plural births matched to CWS without exposure to MCL violations; in column (2) we restrict to our analytic sample for overall contamination.^xxxi These births include those to mothers who switch CWS between some of their births in the panel. There is very little difference in mother characteristics between columns (1) and (2), suggesting that samples are consistent after addressing data availability and trimming. Columns (3) and (4) examine the births in column (2), split by the tails of average contamination by CWS. These columns suggest that mothers of lower socioeconomic status, and worse birth outcomes on average, tend to give birth in places with higher levels of water contamination (that does not trigger a health-based violation). Column (5) restricts the sample to the mothers with multiple births that are included in our mother fixed-effects regressions. In general, differences between column (5) and column (2) are not large, suggesting that sample composition does not substantially change when we restrict to mothers with multiple births (i.e., siblings).^xxxii

Table 2.

Descriptive statistics of births in Pennsylvania, 2003–2014.

	(1)	(2)	(3)	(4)	(5)
Characteristic	Non-plural	Analytic	Tails of contamination distribution		Mom FE
	No MCL	sample	≤ 10th	≥ 90th	Sample
Sample Sizes
No. of births	1,241,656	1,216,132	122,856	110,523	700,815
No. of moms	832,674	820,173	112,209	103,069	304,934
No. of CWSs	1,537	1,459	1,302	885	1,343
Outcomes
Low birth weight	0.067	0.067	0.057	0.079	0.065
Preterm birth	0.082	0.082	0.074	0.102	0.081
Small for gestational age	0.095	0.096	0.083	0.095	0.091
Term birth weight (g)	3,396 (468.9)	3,396 (468.8)	3,423 (465.0)	3,389 (467.5)	3,402 (467.2)
CWS births per 100k per month	113.4 (107.4)	112.7 (103.3)	141.0 (223.4)	109.6 (108.1)	112.1 (98.59)
Other characteristics
Mom age (years)	27.87 (6.054)	27.87 (6.057)	28.01 (5.854)	27.80 (6.023)	27.69 (5.858)
Mom Black	0.210	0.213	0.074	0.185	0.220
Mom Hispanic	0.068	0.069	0.038	0.060	0.074
Mom white, not Hispanic	0.722	0.718	0.889	0.757	0.713
HS or less	0.401	0.401	0.377	0.402	0.406
Mom smokes	0.224	0.223	0.262	0.238	0.216
Mom married	0.569	0.568	0.642	0.581	0.578
WIC/Medicaid	0.477	0.478	0.422	0.459	0.482
Different CWS next birth	0.369	0.368	0.456	0.356	0.368
Mean RRCML, all contaminants	0.095 (2.237)	0.078 (0.078)	0.014 (0.006)	0.285 (0.082)	0.078 (0.077)
Mean RRCML, chemical only	0.085 (2.238)	0.069 (0.080)	0.013 (0.006)	0.281 (0.112)	0.069 (0.079)

Notes: Sample consists of births matched to a public water system. Means (for indicator variables, proportions) are displayed, with standard deviations in parentheses when appropriate. Column (1) provides statistics for non-plural births without MCLs. Column (2) restricts to our analytic sample, keeping only births with an overall measure of contamination and for which that measure is below the 99th percentile. Columns (3) and (4) describe births in the tails of the overall contamination distribution (among births in column (2)). Column (5) describes births in our mother fixed-effects specifications.

To reiterate, column (1) of Table 2 already restricts to births matched to CWS; in Online Appendix Table A8 we compare this restricted sample to the sample of all births, including those not matched to CWS. The most notable difference is that the sample matched to CWS is more urban, which is expressed most strongly by being of greater share Black (as rural Pennsylvania is predominantly non-Black).

EMPIRICAL METHODS

We employ two regression models to study the effects of drinking water contamination on birth outcomes, one with community water system (CWS) fixed effects (FE) and the other with mother FE (within-mother across-birth variation).

The first model takes the following form:

$${Outcome}_{iwmt}=\beta {RRMCL}_{iwmt}+{\theta }^{\prime }{X}_{iwmt}+{\mathrm{ɳ}}^{\prime }{Temp\&Poll}_{iwmt}+{\alpha }_w+{\gamma }_t+{ϵ}_{iwmt}.$$ #(3)

In this model, $i$ denotes birth to mother $m$ living in CWS $w$ in year-month $t$ at the time of the birth; ${Outcome}_{iwmt}$ is a birth outcome; the birth outcomes we consider are: low birth weight (LBW), preterm birth (PTB), small for gestational age (SGA) and term birth weight (TBW). See the section “Birth Records and Control Variables” for information about how these outcomes are defined.

This regression specification employs a fixed effects estimator, where ${\alpha }_w$ and ${\gamma }_t$ are water system and birth year-month fixed-effects, respectively. The independent variable of interest is ${RRMCL}_{iwmt}$, “mean result relative to MCL,” a measure of overall or chemical drinking water contamination which we construct from the sample concentrations as described in subsection “Water Quality Index Construction Details.” The identification assumption is that, after adjusting for an array of individual mother, weather and pollution characteristics in $X_{iwmt}$ and ${Temp\&Poll}_{iwmt}$, trends in birth outcomes across water systems would be the same on average except for differences due to differences in water contamination trends.

We control rigorously for a wide array of mother, birth, and weather measures. Whenever a variable is missing, we set it to zero and include an indicator for missingness. We include a full set of mother demographic and socioeconomic controls, and risk factors for the pregnancy, in $X_{iwmt}$.^xxxiii These control variables were curated from the birth records.^xxxiv The variables in ${Temp\&Poll}_{iwmt}$ are weather and temperature controls obtained from the daily weather statistics provided by Schlenker (Schlenker & Roberts, 2009), as well as Toxics Release Inventory (TRI) facility statistics from the EPA (EPA, 2013).^xxxv

The CWS fixed-effects model also naturally lends itself to an aggregate water-system level fertility regression,

$$\begin{array}{l}\left({BirthRate}_{wt}=\beta \left(RRMCLlast9{months}_{wt}\right)+{\theta }^{\prime }\left(averageXlast9{months}_{wt}\right)\right.\\ +{\mathrm{ɳ}}^{\prime }\left(averageTemp\&Polllast9{months}_{wt}\right)+{\alpha }_w+{\gamma }_t+{ϵ}_{wt}.\end{array}$$ (4)

This model is motivated by the concern that observed fertility rates may respond to drinking water contamination, either due to behavior such as mobility or selection into pregnancy, or due to health issues, such as infertility, miscarriage, or stillbirth. In this model, ${\left(BirthRate\right)}_{wt}$ is the birth rate for the CWS in the given year-month, defined as the number of births divided by the system’s population served. We also estimate a version using $\mathrm{l}\mathrm{o}\mathrm{g}\left(births\right)$ as the dependent variable to alleviate concerns that we are not accurately measuring CWS-population. Our population served measure is time-invariant, so we cannot ignore the fact that variations in birth rates may be due to mothers leaving or entering CWS over our panel; we address maternal mobility below. The variable of interest ${(RRMCLlast9months}_{wt})$ is a measure of mean contamination in the water system over the last 9 months (by aggregating the RRMCL measures described in the section “Data and Summary Statistics,” weighting months equally); $\left({averageXlast9months}_{wt}\right)$ is a measure of average birth characteristics listed earlier taken over the last 9 months weighting births equally; and $\left({averageTemp\&Polllast9months}_{wt}\right)$ is the same battery of weather and TRI facility-based controls we use in models at the individual birth level, except again averaged over the last 9 months, weighting births equally.^xxxvi In these regressions, we weight systems by their population served.

In light of the possibility that any fertility effects we identify may be due to mothers simply moving out of CWS in response to contamination, it is useful to repeat regression (3), but with a different outcome variable: whether the mom is in a different CWS in her next birth (restricting to mothers with at least two births). These models directly calculate whether mothers can be seen to switch systems more often when there is greater water contamination in their first birth. These estimates indirectly speak to whether any fertility effects we see are due to mother mobility.

The second birth-level regression model we employ is a within-mother analysis, which uses variation only across the births of the same mother. The regression is identical to the CWS fixed- effects model (equation 3) but mother fixed-effects are added, in addition to the current fixed-effects. Controlling for all time-invariant characteristics of mothers is a rigorous way to determine whether water quality variation affects birth outcomes, rather than being confounded with other mother characteristics.

As before, in addition to fitting mother fixed-effects models with birth outcomes as the dependent variable, we can also fit these with mother mobility (switching CWS) as the outcome. We estimate mobility effects within CWS and within mothers. The effects of contamination on mobility are identified off relating within-mother differences in contamination that mothers face across births to within-mother differences in whether the mother has switched CWS between births.^xxxvii

The recent literature on difference-in-differences (DD) and two-way fixed effects (TWFE) estimators have made significant advances in a staggered treatment setting (Callaway et al., 2021; Callaway & SantAnna, 2021; De Chaisemartin & dHaultfoeuille, 2020; Goodman-Bacon, 2021). While we do employ a panel model that leverages within water system or within mother variation over time, we are not explicitly using a difference-in-differences model due to this being a non-standard setting with no clear event. Contamination can steadily increase over time, stay relatively constant, reduce abruptly following a change in technology or following a regulatory violation/remediation, and it is not clear that it follows a standard pattern commonly used in a DD model exploiting pre/post treatment (with continuous treatment). Here, there are also unlikely to be “never treated” units.

While the literature on TWFE DD and event studies with continuous treatment suggests this could still be an issue (Callaway et al., 2021), the literature does not at this time have a solution to employ. The primary issue that the continuous treatment in a TWFE design presents is a selection issue. The decomposition in Callaway et al. (2021) showed that the presence of this selection bias without an easy fix will require credibly arguing that the average treatment effects are equal across all dosage groups at the same dosage level. The identification assumption is strong—that there is no selection into the intensity of treatment. However, we are studying dosages of water contamination that are compliant with current regulations and does not trigger a violation (i.e., information to the consumer). The water systems do not have an incentive to manipulate concentrations except near the regulatory threshold. We show that mothers are not changing water systems in response, nor are there large changes in fertility. Finally, our “dose” is defined by a water quality index that is not observed by the consumer, the operator, or the policymaker, and it is unlikely that any systems or mothers are selecting on the level of treatment.

RESULTS

Tables 3 and 4 display our main estimates of the overall effects of drinking water contamination in community water systems (CWS) on birth outcomes. Tables 5 and 6 present the results for the reproductive-specific water quality index. Table 7 displays our estimates of the effects of water contamination on mother mobility or overall CWS-level fertility. Throughout, we restrict to our analytic sample. In particular, we have removed all births that have a gestation period which overlaps with any MCL violation compliance period. Thus, our estimates are of the effects of water contamination on birth outcomes, mobility, or fertility, for variation in water contamination which does not coincide with MCL violations.

Table 3.

Effects of drinking water contamination on low birth weight and preterm birth.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: low birth weight (LBW)
CWS FE
10th to 90th RRMCL	0.01356*** (0.00302)	0.01456** (0.00133)	0.01249*** (0.00285)	0.01349** (0.00124)	0.01059*** (0.00240)	0.01050** (0.00140)
Observations	1,216,132	1,214,688	1,216,132	1,214,688	1,216,132	1,214,688
Adj R2	0.0050	0.0052	0.0456	0.0458	0.1058	0.1059
AETO’s δ			32.59	34.50	10.78	7.82
Mom & CWS FE
10th to 90th RRMCL	0.00957*** (0.00245)	0.01048** (0.00120)	0.00932*** (0.00232)	0.01003** (0.00115)	0.00819*** (0.00198)	0.00774** (0.00133)
Observations	700,815	699,546	700,815	699,546	700,815	699,546
Adj R2	0.1506	0.1508	0.1689	0.1691	0.2103	0.2104
AETO’s δ			2.29	1.43	1.13	0.55
Panel B: preterm birth (PTB)
CWS FE
10th to 90th RRMCL	0.01997*** (0.00405)	0.02134** (0.00178)	0.01866*** (0.00375)	0.02001** (0.00164)	0.01626*** (0.00306)	0.01623** (0.00175)
Observations	1,216,132	1,214,688	1,216,132	1,214,688	1,216,132	1,214,688
Adj R2	0.0036	0.0040	0.0467	0.0471	0.1316	0.1318
AETO’s δ			40.16	41.35	13.38	9.72
Mom & CWS FE
10th to 90th RRMCL	0.01571*** (0.00363)	0.01716** (0.00166)	0.01533*** (0.00345)	0.01658** (0.00160)	0.01405*** (0.00276)	0.01377** (0.00170)
Observations	700,815	699,546	700,815	699,546	700,815	699,546
Adj R2	0.1486	0.1489	0.1663	0.1665	0.2252	0.2253
AETO’s δ			2.44	1.79	2.00	0.99

Notes: Standard errors in parentheses

^*p < 0.10

^**p < 0.05

^***p < 0.01.

Standard errors are two-way clustered on public water system and mother. Each cell is a separate regression. Observations (number of births) and adjusted R²s (for the full model, i.e., including the fixed-effects) are reported. In each regression, the sample is restricted to births in our analytic sample for whom we observe at least one water quality sample of the given contaminant group during gestation in the public water system of residence at the time of birth. Control variables used vary over supercolumns; independent variable, i.e. the contaminant group studied, vary across columns; the outcomes varies over panels; the specification used (either public water system fixed-effects or mom fixed-effects) varies over sub-panels; temp & TRI ctrls = weather and toxics release inventory (air pollution) controls.

Table 4.

Effects of contamination on small for gestational age and term birth weight.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age (SGA)
CWS FE
10th to 90th RRMCL	0.00341*** (0.00077)	0.00268** (0.00064)	0.00282*** (0.00074)	0.00217** (0.00061)	0.00258*** (0.00079)	0.00183** (0.00062)
Observations	1,214,205	1,212,766	1,214,205	1,212,766	1,214,205	1,212,766
Adj R2	0.0054	0.0054	0.0381	0.0381	0.0390	0.0390
AETO’s δ			13.15	11.59	8.30	5.86
Mom & CWS FE
10th to 90th RRMCL	0.00279*** (0.00100)	0.00179** (0.00083)	0.00276*** (0.00096)	0.00165** (0.00081)	0.00290*** (0.00099)	0.00173** (0.00083)
Observations	698,761	697,487	698,761	697,487	698,761	697,487
Adj R2	0.1386	0.1393	0.1542	0.1548	0.1544	0.1550
AETO’s δ			6.36	0.65	−1.90	1.26
Panel B: term birth weight (TBW)
CWS FE
10th to 90th RRMCL	−10.61*** (1.39)	−10.04*** (1.25)	−9.75*** (1.55)	−9.44*** (1.21)	−9.54*** (1.69)	−8.77*** (1.32)
Observations	1,116,570	1,115,266	1,116,570	1,115,266	1,116,570	1,115,266
Adj R2	0.0169	0.0169	0.0943	0.0944	0.0987	0.0988
AETO’s δ			30.56	42.24	22.73	17.70
Mom & CWS FE
10th to 90th RRMCL	−9.05*** (1.73)	−8.82*** (1.34)	−9.15*** (1.84)	−8.58*** (1.37)	−9.33*** (1.87)	−8.25*** (1.46)
Observations	612,960	611,893	612,960	611,893	612,960	611,893
Adj R2	0.3437	0.3437	0.3827	0.3828	0.3848	0.3849
AETO’s δ			−10.33	3.87	−5.48	1.47

Notes: Standard errors in parentheses

^*p < 0.10

^**p < 0.05

^***p < 0.01.

Standard errors are two-way clustered on public water system and mother. Each cell is a separate regression. Observations (number of births) and adjusted R²s (for the full model, i.e., including the fixed-effects) are reported. In each regression, the sample is restricted to births in our analytic sample for whom we observe at least one water quality sample of the given contaminant group during gestation in the public water system of residence at the time of birth. Control variables used vary over supercolumns; independent variable, i.e. the contaminant group studied, vary across columns; the outcomes varies over panels; the specification used (either public water system fixed-effects or mom fixed-effects) varies over sub-panels; temp & TRI ctrls = weather and toxics release inventory (air pollution) controls.

Table 5.

Effects of drinking water contamination on low birth weight and preterm birth, using five- contaminant reproductive index.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	5 Contam	All	5 Contam	All	5 Contam
Panel A: low birth weight
CWS FEs
10th to 90th RRMCL	0.01356*** (0.00302)	0.01936*** (0.00639)	0.01249*** (0.00285)	0.01833*** (0.00580)	0.01059*** (0.00240)	0.01406** (0.00576)
Observations	1,216,132	1,139,430	1,216,132	1,139,430	1,216,132	1,139,430
Adj R2	0.0050	0.0049	0.0456	0.0452	0.1058	0.1049
Mom & CWS FEs
10th to 90th RRMCL	0.00957*** (0.00245)	0.01709** (0.00405)	0.00932*** (0.00232)	0.01637** (0.00381)	0.00819*** (0.00198)	0.01282** (0.00411)
Observations	700,815	637,615	700,815	637,615	700,815	637,615
Adj R2	0.1506	0.1507	0.1689	0.1686	0.2103	0.2094
Panel B: preterm birth
CWS FEs
10th to 90th RRMCL	0.01997*** (0.00405)	0.02743** (0.00863)	0.01866*** (0.00375)	0.02621** (0.00798)	0.01626*** (0.00306)	0.02096** (0.00788)
Observations	1,216,132	1,139,430	1,216,132	1,139,430	1,216,132	1,139,430
Adj R2	0.0036	0.0034	0.0467	0.0460	0.1316	0.1306
Mom & CWS FEs
10th to 90th RRMCL	0.01571*** (0.00363)	0.02370** (0.00525)	0.01533*** (0.00345)	0.02279** (0.00497)	0.01405*** (0.00276)	0.01834** (0.00543)
Observations	700,815	637,615	700,815	637,615	700,815	637,615
Adj R2	0.1486	0.1487	0.1663	0.1659	0.2252	0.2247

Notes: Standard errors in parentheses

^*p < 0.10

^**p < 0.05

^***p < 0.01.

Standard errors are two-way clustered on public water system and mother. Each cell is a separate regression. Observations (number of births) and adjusted R²s (for the full model, i.e., including the fixed-effects) are reported. In each regression, the sample is restricted to births in our analytic sample for whom we observe at least one water quality sample of the given contaminant group during gestation in the public water system of residence at the time of birth. Control variables used vary over super-columns; independent variable, i.e. the contaminant group studied, vary across columns; the outcomes varies over panels; the specification used (either public water system fixed-effects or mom fixed-effects) varies over sub-panels; temp & TRI ctrls = weather and toxics release inventory (air pollution) controls. Results for the overall index included for comparison, which are the same as the “All” columns in Table 3.

Table 6.

Effects of drinking water contamination on small for gestational age and term birth weight, using five-contaminant reproductive index.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	5 Contam	All	5 Contam	All	5 Contam
Panel A: small for gestational age
CWS FEs
10th to 90th RRMCL	0.00341*** (0.00077)	0.00511*** (0.00143)	0.00282*** (0.00074)	0.00470*** (0.00125)	0.00258*** (0.00079)	0.00390*** (0.00138)
Observations	1,214,205	1,137,738	1,214,205	1,137,738	1,214,205	1,137,738
Adj R2	0.0054	0.0053	0.0381	0.0380	0.0390	0.0390
Mom & CWS FEs
10th to 90th RRMCL	0.00279*** (0.00100)	0.00349** (0.00164)	0.00276*** (0.00096)	0.00350** (0.00167)	0.00290*** (0.00099)	0.00317* (0.00168)
Observations	698,761	635,832	698,761	635,832	698,761	635,832
Adj R2	0.1386	0.1405	0.1542	0.1559	0.1544	0.1561
Panel B: term birth weight
CWS FEs
10th to 90th RRMCL	−10.61*** (1.39)	−13.86*** (3.94)	−9.75*** (1.55)	−13.92*** (3.57)	−9.54*** (1.69)	−12.58*** (3.91)
Observations	1,116,570	1,047,149	1,116,570	1,047,149	1,116,570	1,047,149
Adj R2	0.0169	0.0167	0.0943	0.0945	0.0987	0.0989
Mom & CWS FEs
10th to 90th RRMCL	−9.05*** (1.73)	−13.74*** (2.79)	−9.15*** (1.84)	−13.86*** (2.82)	−9.33*** (1.87)	−12.59*** (2.99)
Observations	612,960	558,355	612,960	558,355	612,960	558,355
Adj R2	0.3437	0.3483	0.3827	0.3867	0.3848	0.3887

Notes: Standard errors in parentheses

^*p < 0.10

^**p < 0.05

^***p < 0.01.

Standard errors are two-way clustered on public water system and mother. Each cell is a separate regression. Observations (number of births) and adjusted R²s (for the full model, i.e., including the fixed-effects) are reported. In each regression, the sample is restricted to births in our analytic sample for whom we observe at least one water quality sample of the given contaminant group during gestation in the public water system of residence at the time of birth. Control variables used vary over super-columns; independent variable, i.e. the contaminant group studied, vary across columns; the outcomes varies over panels; the specification used (either public water system fixed-effects or mom fixed-effects) varies over sub-panels; temp & TRI ctrls = weather and toxics release inventory (air pollution) controls. Results for the overall index included for comparison, which are the same as the “All” columns in Table 4.

Table 7.

Effects of drinking water contamination on mobility and fertility.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother controls		Adding weather controls
	All	Chemical	All	Chemical	All	Chemical
Panel A: Mobility – CWS FE
10th to 90th RRMCL	−0.00108 (0.00189)	−0.00014 (0.00149)	−0.00165 (0.00178)	−0.00068 (0.00151)	−0.00069 (0.00152)	−0.00100 (0.00152)
Observations	413,135	412,785	413,135	412,785	413,135	412,785
Adj R2	0.1271	0.1271	0.1470	0.1470	0.1472	0.1472
B: Mobility – Mom FEs
10th to 90th RRMCL	0.00255 (0.00315)	0.00348 (0.00260)	0.00165 (0.00309)	0.00297 (0.00251)	0.00298 (0.00252)	0.00280 (0.00257)
Observations	172,651	172,498	172,651	172,498	172,651	172,498
Adj R2	0.2494	0.2486	0.2694	0.2689	0.2696	0.2691
Panel C: Fertility (birth rate per 100,000 people served) – CWS FE
10th to 90th RRMCL	−0.07271** (0.03491)	−0.08445** (0.03351)	−0.06937** (0.02934)	−0.07647** (0.03040)	−0.07472** (0.02939)	0.07040** (0.02799)
Observations	59,763	59,406	59,763	59,406	59,763	59,406
Adj R2	0.7397	0.7393	0.7420	0.7417	0.7422	0.7418
Panel D: Fertility (log number of births) – CWS FEs
10th to 90th RRMCL	−0.00114*** (0.00041)	−0.00095*** (0.00033)	−0.00114*** (0.00036)	−0.00089*** (0.00031)	- (0.00037)	- (0.00031)
Observations	33,122	32,912	33,122	32,912	33,122	32,912
Adj R2	0.9924	0.9925	0.9925	0.9926	0.9925	0.9926

Notes: Standard errors in parentheses

^*p < 0.10

^**p < 0.05

^***p < 0.01.

Standard errors are two-way clustered on public water system and mother in panels A and B, and clustered on public water system in panel C. Observations (number of births in panel A and B; public water system months in panel C) and adjusted R²s (inclusive of fixed-effects) are reported. Each cell is a separate regression. In panels A and B, the sample is our analytic sample; in panel A, mothers must have two or more births in the data (to measure mobility), with at least one having a measure of contamination, to be included. Panel B requires mothers to have three or more births in our sample due to the addition of mom fixed-effects. In both panels A and B, the dependent variable (measured for each birth) is whether the mother is in a different public water system at the time of the next birth (being in no public water system in the next birth is considered being in a different water system for this purpose). Observations in panel C are public water system quarters, and the dependent variable is the number of births per 100,000 people served by the public water system. Regressions in panel C are weighted by public water system population served. Regressions in panel D are the same as panel C, except the outcome variable is the natural log of the number of births (community water system quarters with zero births are dropped). Control variables used vary over super-columns; independent variables, i.e., the contaminant group studied, vary across columns.

Birth outcomes

Fitting the model described by Equation 3, we find that the chance of low birth weight (LBW) and preterm birth (PTB) are both significantly positively related to contamination overall and chemical contamination specifically (Table 3). Reading across columns, we find that the addition of our extensive set of control variables does not change inference, despite raising adjusted R² substantially. Note, however, that point estimates are attenuated by 10% to 28% in the fully adjusted specifications (columns 5 and 6) relative to the unadjusted specifications (columns 1 and 2). Generally, effect estimates are larger in the unadjusted models. Reading across rows, we find that our results are more sensitive to whether we use a CWS fixed-effects (CWS FE) specification or a mother fixed-effects (MFE) specification (that also includes CWS FE). In general estimates are larger in the CWS FE specifications. All coefficients are estimated effects of a change from the 10th percentile to the 90th percentile in the Pennsylvanian distribution of exposure to CWS drinking water contamination.^xxxviii Thus they are comparable across specifications. In both cases, effects are economically meaningful: based on columns (5) and (6), in the MFE specification, a 10-to-90th percentile move leads to about a 0.8 percentage point increase in low birth weight (LBW). From the base LBW rate for non-plural births of 6.7% (Table 2), this is about a 12% increase. Analogously for PTB, we find about a 17% increase.

In order to formally calculate the degree to which the addition of control variables affects our inference, we report “AETO’s δ.” By the acronym AETO we refer to the two papers Altonji et al. (2005) and Oster (2016); the latter work substantially builds on the former and provides us the Stata implementation we use.^xxxix The interpretation of δ is the minimum ratio of selection on unobservables to selection on observables that would be needed to make our estimated coefficients zero.^xl A ratio of at least 1 is a plausible rule for saying an effect estimate is “robust to controls.”^xli In our case, AETO’s δ tends to be large in the CWS FE models, but much smaller in the mother (and CWS) fixed-effects models (Table 3). This is partly by construction: the mom fixed-effects absorb many potential unobservables, making it unlikely that the inclusion of control variables will appreciably change the R² in the model. Indeed, consistent with this idea, when we compare for example, columns (1) and (3) in the mom fixed-effects specifications, the control variables do not substantially increase adjusted R². That said, most of our estimates have AETO’s δ’s greater than one. This also lends support for our earlier claim that there is not a clear indication of selection into a particular dose of treatment.

In Table 4, we report the estimated effects on more direct measures of intrauterine growth restriction: small for gestational age (SGA) and term birth weight (TBW). These estimates are consistent with our main findings. In other words, after applying at least some approaches to removing the effects of contamination on gestation length, we continue to find effects of contamination on birth outcomes, suggesting that effects on gestation length are not the only channels by which water contamination affects birth outcome, and intrauterine growth restriction (IUGR) is a possible consequence of contamination.

While our main analysis focuses on indices of water quality that include a large number of contaminants, as discussed in the section “Reproductive Health Water Quality Index” and Appendix section “An Alternative Reproductive Health Index of Five Contaminants Based on Data Availability and Prior Research,” an alternative approach is to restrict to contaminants which are commonly sampled and for which we could find research literature suggesting deleterious health effects for infants. We ultimately chose five contaminants to include in a reproductive health specific water quality index. Estimates using this five-contaminant index to measure water quality are included in Tables 5 and 6 (the tables include estimates for our all index of contamination for comparison). Across the various specifications, we see slightly stronger effects for the reproductive-specific index. In particular, focusing on columns (5) and (6) and mother fixed-effects regressions, we find that a 10th-to-90th percentile movement in RRMCL for our five-contaminant index leads to a 1.3 percentage point increase in low birth weight (about 19% relative to the mean), a 1.8 percentage point increase in preterm birth (about 22% relative to the mean), 0.3 percentage point increase in small for gestational age (significant at only the 10% level), and a 12.59 gram decrease in term birth weight.

For interested readers, in Appendix Section “Effects of MCL Violations on Birth Outcomes” we estimate models studying the effects of MCL violations on health, following Currie et al. (2013). We find that 3% of our sample has a regulatory violation during gestation, compared to 8% in NJ in the prior study. We find similar effects in magnitude and statistical precision for any MCL violation, but not for chemical violations, which is in contrast to Currie et al. (2013).^xlii

Avoidance via migration

Table 7 Panels A and B display the estimated effects of contamination on mother mobility. Mobility is defined for a given birth as an indicator for the mother being in a different CWS at the time of her next birth. Regressions in this table are similar to equation 3, except with mobility as the dependent variable. As before, we restrict to our analytic sample as described in the section “Data and Summary Statistics.” In these models, the independent variable of interest is CWS drinking water contamination, as measured during the gestation period of the current birth.

Mothers would exhibit avoidance behavior via migration if, when they experience higher levels of water contamination during this gestation period, they are more likely to be found in a different CWS at the time of their next birth. We find little evidence for this type of avoidance behavior, which is consistent with mothers not behaviorally responding to contamination levels that do not trigger MCL violations. In Table 2 about 37% of births have their mother in a different CWS at the next birth (this is consistent with statistics from New Jersey in Currie et al., 2013, so the point estimates from Table 7 are also economically small).

While we have evaluated the degree of avoidance behavior through very costly migration decisions, this is not the only possible and probably not the most likely form of avoidance behavior. It is possible that very attentive mothers purchase bottled water instead of using tap water, or they filter their tap water, when contamination rises. This avoidance behavior is less likely in our setting where we study water quality that doesn’t trigger a health-based violation unless their system previously experienced violations. However, mobility looms large as a potential avoidance behavior that could lead to bias in our estimated effects of contamination on birth outcomes, and especially fertility, discussed in the next section, so it is helpful to note little evidence for this particular behavior.

Fertility

Table 7 Panel C displays the estimated effects of contamination on fertility rates from the regression described in equation 4. These regressions are weighted by CWS population served. CWS fertility rates are defined as monthly births per 100,000 people served.^xliii Our findings are consistent with poor water quality reducing fertility. A typical birth is in a CWS with about 113 births per 100,000 served per month (Table 2). We find that changes in overall contamination from the 10th to the 90th percentile reduces this number by about 0.07 births per 100,000 per month, or approximately a 0.06% reduction in fertility. Effects for chemical contamination are similar. Comparing these effects to those found for lead contamination, D. S. Grossman & Slusky (2019) found that the Flint water crisis led to a 12% reduction in fertility, while Clay et al. (2021) found, for topsoil lead, that bringing a typical above-median county to the lead levels of a typical below-median county would lead to approximately an 11% increase in the birth rate. By contrast, we find that going from the 10th to the 90th percentile of water contamination in our sample has a much smaller, though still statistically significant, negative effect on fertility. Lead is known to induce miscarriage, whereas the more general contamination studied in our paper does not have strong evidence for causing pregnancy loss. Given that mobility effects appear to be limited, the effects on fertility are likely due to selection out of pregnancy, either through choice or through contamination leading to reduced fertility.

Our fertility regressions in Panel C of Table 7 use time invariant CWS population served in the denominator of the dependent variable. As a robustness check, in Panel D of the same table we use log births as the outcome variable, thereby avoiding the use of the time-invariant population served variable. This approach is commonly used (Ananat & Hungerman, 2012; González, 2013; Guldi, 2008). Effects in Panel D are larger in magnitude, at about a 0.1% reduction in fertility, but they are of the same sign and statistical significance as our main fertility results in Panel C.

SENSITIVITY ANALYSIS

In this section, we discuss the sensitivity of our findings to different subsamples. For all results presented thus far, we have restricted to non-plural births, and have included all mothers, including those who switch CWS. We show robustness to these sample selection choices in this section. All tables referenced in this section are included in Online Appendix section “Sensitivity Analyses.”

Robustness to sample restrictions based on plurality and mobility

In the Online Appendix, we estimate our models with two alternative samples: first, all births (i.e., including plural births), and second, sub-setting to never-moving mothers. Including plural births does little to change our findings (Appendix Tables A9 and A10). Our findings are also robust to excluding all mothers with multiple births who we ever observe in different CWS across births (Appendix Tables A11 and A12). This is surprising since this sample restriction removes a quarter to a half of the births in our data, depending on specification; moreover, one component of our variation in water quality from the mother fixed-effects analyses was due to mothers switching water systems with different average contamination. The robustness of our results in this very restrictive sample provides additional support for our within-water system estimates.

Table 8:

Largest systems by water quality letter grades based on our index measure.

	(1)	(2)	(3)
Water system name	Population served	Mean RRMCL	Letter grade
Ambler Boro Water Dept	20,000	0.02	A
Lansford Coaldale Jt Water Aut	9,300	0.02	A
Paw Bangor District	9,008	0.02	A
Erie City Water Authority	180,000	0.03	B
York Water Co	159,623	0.04	B
North Penn Water Authority	82,822	0.03	B
Pittsburgh Water & Sewer Auth	250,000	0.06	C
West View Boro Muni Auth	200,000	0.05	C
Westmd Mun Auth-Sweeney Plant	140,000	0.06	C
Philadelphia Water Department	1,600,000	0.07	D
Pa Amer Water Co-Pittsburgh	507,675	0.08	D
City Of Lancaster	120,000	0.09	D
Aqua Pa Main System	784,939	0.06	F
Easton Area Water System	93,400	0.05	F
Capital Region Water	66,540	0.06	F

Notes: Letter grades assigned based on average result relative to MCL for all contaminants available since 2010 (reported in the mean RRMCL column): the cleanest 25% public water systems are assigned grade “A,” the next 25% are assigned grade “B,” and so forth.

Table 9:

Possible effects of an information provision intervention - assigning letter grades to systems on consumer confidence reports.

	(1)	(2)	(3)
	A or B	C, D, or F	Overall
Panel A: Low birth weight
Actual rates	5.73	6.90	6.67
After policy change	5.73	6.60	6.43
Panel B: Pre-term birth
Actual rates	7.28	8.33	8.13
After policy change	7.28	7.81	7.71
Number of births	242,388	1,014,850	1,257,238

Notes: We assume that people in systems assigned to letter grades C, D, or F engage in avoidance behavior and obtain water quality at the average for letter grade A systems. Estimates give an idea of the magnitude of public health gains that might be obtained from adopting a consumer confidence report simplification policy.

Robustness to sample restrictions based on MCL violations

Recall that throughout our analysis, we excluded births exposed to MCL violations. In the Online Appendix, we also re-estimate our models with two additional alternative sample restrictions: we run the analysis keeping births that are exposed to MCL violations, and we run the analysis removing births with any samples exceeding the MCL threshold. The latter analysis is motivated by the complexity of the regulatory environment, in that in many cases violations are triggered only when a particular average of samples exceeds the regulatory threshold (e.g., TTHM and HHA5 use running averages), and a concern that the MCL violation data may be missing some violations. See the Online Appendix sections “Including Births Exposed to MCL Violations” and “Removing Births With Any Samples Exceeding MCL” for further details on these alternative samples. We also include models estimating the effects of MCL violations on birth outcomes in the Appendix section “Effects of MCL Violations on Birth Outcomes.”

Our findings are robust to including births exposed to MCL violations (Online Appendix Tables A13 and A14). This may not be surprising since our sample size increases by only about 3 to 5% (depending on specification) by adding these births. More surprisingly, our findings are robust to removing births exposed to samples exceeding MCL thresholds (Online Appendix Tables A15 through A20). In conducting these further restrictions, we document in the Appendix section “Removing Births With Any Samples Exceeding MCL” that four contaminants very often have single samples that exceed their respective MCL (Coliform, TTHM, HAA5, and lead), and this makes sense because none of these contaminants have violations driven by single-sample exceedances. Our findings are robust to several different ways to treating these four contaminants, which is an important consideration since if we drop births exposed to any samples with exceedances for these contaminants, we lose about 85% of our data. In any case, this robustness check is intended to simulate a stricter definition of MCL violation and reduces sample sizes by 25% to 85% depending on specification. That our results are robust to these stricter definitions of “compliant drinking water” suggests that there are indeed statistically significant effects of water contamination on birth outcomes, even when systems are in compliant with regulations, or all sample concentrations are below the regulatory threshold.

Robustness to restricting to MFE sample using CWS FE

In the main specifications, using mother fixed-effects (MFE) rather than community water system fixed-effects (CWS FE) often appreciably and significantly changed the effect estimates (e.g., the differences between the CWS FE and MFE subpanels in Table 3). In fact, this is the main way in which our results are sensitive, since the addition of mother and birth control variables does little to change our estimates.

In the Online Appendix, we present CWS FE and MFE estimates sub-setting throughout to the sample used in the MFE specifications. Our results are robust in Appendix Tables A21 and A22. Thus, the differences between the CWS FE and MFE specifications are due to differences in the models we fit, not due to differences in the samples we employ.

Study limitations

Our estimates may be affected by a few forms of exposure misclassification, which is common in environmental health research. First, our index is created from the measured contamination during gestation periods. The regulations provide monitoring schedules that vary by size of system,^xliv season, and contaminant (among others). For some contaminants, systems are required to measure monthly, quarterly, annually, every 3 years, and in some cases every 6 years. Unfortunately, this temporal sparsity of the data leads to unobserved contamination levels during gestation for some contaminants that make up our index for any individual pregnancy. We chose not to impute contamination because we do not believe we have adequate information to predict contamination over time across systems. Despite this limitation, we believe using the sampling data provides more temporal frequency than the MCL violations data as they only report violations or not in each monitoring period (approx. quarterly) but when there is no violation, the researcher doesn’t observe whether it is a true zero (measured) or just no violation in the monitoring period (which leads to censoring). Second, our estimates are dependent on the accuracy of the technology measuring the concentrations for CWS. In particular, in the Pennsylvania data, non-detect values are treated as zero, when in fact, the technology may have a minimum detection limit (MDL) such that we cannot be confident that the concentrations are in fact zero. In Appendix section “Non-Detection and Limits of Measurement,” we examine how replacing non-detects with the MDL changes the distribution of our index, and furthermore show that Pennsylvania MDL are typically lower than California’s, suggesting Pennsylvania is a decent setting to conduct this study (Appendix Tables A4 through A6). Third, we rely on water system service areas to define exposure to CWS and assign water quality during gestation at the system-level to each mother residing in that water system. There could be exposure misclassification if the water ingested by an individual mom is more or less contaminated than what is measured prior to the water being released into the distribution system. Ideally, we would have distribution maps and be able to assign samples for sub-areas of the service area. Unfortunately, these distribution maps are not available and this is a data limitation.^xlv Despite this limitation, we believe that this is analogous to what is used in air quality studies extrapolating monitor or satellite data (Currie & Walker, 2019) or water quality studies at the county or city level (Flynn & Marcus, 2023; D. S. Grossman & Slusky, 2019). However, we are using system boundaries providing sub-county variation in water quality, which should improve upon existing approaches that use county-level variation (McDonald et al., 2022). Finally, in our analysis, we used a singular vintage of the water system boundaries. In the Appendix, we checked to see if there was any correlation between systems that changed over time and the water quality index. See Appendix section “Community Water System Boundaries and Linking Births to Water Quality” for details. We found that while there were some changes over time, they are minimally associated with our measure of water quality (Appendix Table A7). We hold the boundaries constant over our study period to ensure that factors that might influence changing systems (e.g., population growth, population decline, water contamination, etc.) are not influencing our assignment of treatment.

DISCUSSION

Policymakers have increasing begun using summary indices to regulate or improve communication of quality in various sectors for consumers. For example, Centers for Medicaid and Medicare Services (CMS) has promoted Hospital Compare, Nursing Home Compare, and Medicare Care Compare as mechanisms for improving health care quality. Most related to our context, the EPA has long provided the U.S. Air Quality Index (AQI; updates by metro area daily) and the AQI is now reported in AirAlerts communicating to the public when air quality could be risky for at-risk individuals. State departments of education provide school report cards and provide summary measures of quality using “grades.” Researchers have studied responses by consumers and suppliers (e.g., providers, schools) to use of these indices for communicating quality. For example, AQI and communicating poor air quality has been shown to change participation in outdoor activities (Wen et al., 2009; Zivin & Neidell, 2009) and following star-ratings for Nursing Home Compare, low scoring nursing homes lost demand and high scoring nursing homes experienced an increase in demand (Werner et al., 2016). With recent scrutiny on water quality, use of a water quality index to communicate more frequent water quality information to society could be beneficial for avoiding poor outcomes for the most sensitive groups.^xlvi

Figure 2 displays an excerpt of Philadelphia’s 2017 Consumer Confidence Report (CCR), which was provided to consumers of Philadelphia water to inform them of water quality in the prior year. Sample result data in the report is displayed by contaminant, for example, the highest level of Barium reported in Philadelphia’s water in 2017 was 0.044ppm. Except for reported regulatory violations, which are rare, consumers are unlikely to extract actionable information from current Consumer Confidence Reports. Motivated by our results, we propose that regulators consider providing a summative, index measure of water quality for CWS, to be displayed on annual CCRs or to be used more frequently as water alerts to community members.

Figure 2:

Philadelphia Water Consumer Confidence Report summary results, 2017.

We run a simple simulation to provide an example policy counterfactual, if systems were to adopt our overall index measure. Our example accountability policy takes our index measure, measured at the birth level, and aggregates to the CWS level by averaging the measure for all births (below the 99th percentile of contamination) since the start of 2010 through the end of 2014.^xlvii This gives us a “mean result relative to MCL” or “mean RRMCL” measure. We then split systems into letter grades, A through F, as follows. We begin by splitting up the CWS-level mean RRMCL measure into four quartiles and assigning letter grades A through D to those quartiles. Then, if a system receives any MCL violation (or for lead and copper, any trigger of the action level) over the 2010 through 2014 period, we assign the system to grade F. Several large water systems with their letter grades assigned according to our method are displayed in Table 8.

In Table 9, we display results from a counterfactual simulation, where we assume that consumers in systems with letter grades C, D, or F engage in avoidance behavior or lobby their water systems for better water treatment, and these responses ultimately move their effective water quality to the mean quality for systems of letter grade A. The table displays back-of-the-envelope estimated effects on birth outcomes.^xlviii To put these numbers in context, there are about 150,000 singleton births in Pennsylvania each year, and about 10,005 are low birth weight (6.67 out of 100); the proposed policy could reduce this to about 9,645 low birth weight infants, thus, about 360 infants per year would avoid being low birth weight (3,960 over our 11-year period). Similarly, for preterm birth, there are about 12,195 such infants born in Pennsylvania each year; the policy could reduce this to 11,565, thus preventing 630 preterm births annually (6,930 over 11 years). This policy could save around $290 million in total costs for PTB and LBW associated hospitalizations and medical services before age 9.^xlix Given the large human capital costs of poor birth outcomes for later life education outcomes, labor market outcomes, and life expectancy (Almond et al., 2018; Black et al., 2007), our estimates clearly have economic and policy significance.

Based on this simulation, we argue that providing a grade for water systems could be an improvement on the current information provided to consumers. We must acknowledge, however, that while this provides more information for the consumer, it puts the burden of preventing exposure on the consumer and may disproportionately affect low-income families. Research on water systems in California, however, suggests that water systems may not be removing contaminants even after regulatory violations (Grooms, 2016). Given this imperfect enforcement, we do believe that adding a summary index (after further validation) to CCR or used in water quality alerts could further protect public health at the very least by informing consumers of water quality in their local systems. For example, the EPA’s Air Alerts indicate levels of concern as well as which populations might be most at risk and should avoid exposure. The actual implementation of this information would need to take into account equity and environmental justice concerns, much like air quality alerts where low-income families may not have air conditioning, access to air filters or be able to avoid exposure due to employment that requires spending time outdoors.

The index that we used in this study is but one possible summary measure, and we encourage researchers and policymakers to consider alternatives. Our index, or an alternative water quality index, could be used to study other health outcomes and other populations that could be at risk of adverse outcomes from exposure to drinking water contamination (e.g., children, older adults). As research causally links various water quality indices that may be relevant for particular populations (such as our reproductive index described in the section “Reproductive Health Water Quality Index”) or how our overall index affects the health of other populations, “water alerts” could be targeted to specific populations at risk with enough research and validation to support such implementation.

We view our letter grading policy proposal as a way to consider the potential benefits of improving water quality, and our study is not a formal cost-benefit analysis of safe drinking water policies. The cost-benefits literature on water contamination is relatively new, and a recent evaluation of the Clean Water Act suggested that its benefits did not exceed its costs (Keiser & Shapiro, 2019a). One approach to assess costs in our context might be to assume that households make bottled water purchases to replace tap water with higher quality water, but it is unclear whether bottled water is a sufficient substitute for quality tap water. In particular, the persistent and large negative house price effects of the Flint water crisis found in Christensen et al. (2023), despite 3 years of free bottled water provision and several programs to distribute free water filters and filter cartridges, suggests that private goods are imperfect substitutes for quality tap water.

Our results using a water quality index have two additional implications for water regulation. The scientific literature has suggested that certain co-occurring contaminants that are not regulated by SDWA and therefore are unmeasured in our data could be explaining some of the effects we observe (Bradley et al., 2018). Using this index could be helpful, given regulators cannot monitor all contaminants in drinking water. Furthermore, as briefly discussed in the introduction, SDWA regulates contaminants on a contaminant-by-contaminant basis. Toxicology literature suggests that mixtures of contamination may be more problematic for human health (Braun et al., 2016; Gennings et al., 2018), and people are exposed to mixtures in real-world settings. Our approach of using a summary index could be a first step towards having drinking water policy take into account mixtures.

In this paper, we define compliance as not having a health-based violation as reported to the EPA. Recent work (Baker et al., 2022) has argued that while there is some evidence that violations go unreported to the national EPA, the majority (79%) are accurately reported.^l Given that we condition on no health-based violations as reported to the EPA, we join other studies (Allaire et al., 2019) that suggest that improving this reporting could benefit communities health through communicating water quality accurately.^li Furthermore, if violations do not result in improved compliance, as Grooms (2016) suggested, our findings could reflect this lack of enforcement resulting in poor birth outcomes.

We address the question of “missed” or unreported violations by crudely removing samples that are above the regulatory threshold for each contaminant (Appendix section “Removing Births With Any Samples Exceeding MCL”). The current regulations do not treat these thresholds as binding for all contaminants, and so our removal of concentrations above any regulatory threshold is much more stringent than the current regulations. We find similar results when we treat thresholds as binding for all contaminants in Appendix Tables A19 and A20. While this suggests that even when concentrations are below the regulatory threshold, water contamination could be impacting infant health, we lose substantial sample and caution that more research is needed to confirm these findings.

For analyses such as those reported in this paper, the lack of regular monitoring for certain contaminants that may be very important for reproductive health outcomes (e.g., lead, uranium) limited our ability to create a reproductive health-specific water quality index containing the full list of contaminants that the EPA has indicated could harm reproductive health or infants.^lii Frequent monitoring has costs and the EPA has already considered these costs when defining monitoring schedules for each contaminant. Still, our analysis has limitations (see the section “Robustness to Restricting to MFE Sample Using CWS FE” for discussion) due to the limited frequency of sampling, and this limitation applies to other analyses that may wish to inform regulators.

CONCLUSIONS

In this paper, we studied the effects of drinking water contamination on birth outcomes, using panel data on births within community water systems (CWS) for the state of Pennsylvania. In order to provide the most policy-relevant estimates possible, we restricted the sample to births that were not exposed to a water quality regulatory violation. We found that when we measure water contamination using either an omnibus measure of water quality or a measure removing microbial contaminants, birth outcomes respond negatively to water contamination, even in this no-violation-exposure sample. We find consistent evidence when using a reproductive-health-specific index and when we remove samples below the regulatory threshold (despite a substantial loss in sample size). In addition, we identify small negative effects on fertility. We find little evidence of mobility in response to contamination. This is consistent with our variation in water contamination not triggering significant avoidance behavior, which is intuitive since we have removed births exposed to regulatory violations that trigger public notification. Finally, we argue that current consumer confidence reports provided to consumers by CWS do not provide much information that can be easily understood by the consumer. Our index could be used to characterize the level of contamination in a simpler way and help consumers avoid exposure.

Although our findings are of short-run health consequences of poor water quality, even at levels that do not trigger regulatory violations, long-run effects are not a distant possibility. Our results call for further research on these important environmental and health policy issues. Future studies could repeat our analysis on different samples and fit models that critique and refine our index measures. In addition, previous research has shown that consumer confidence report policy changes can influence water system behavior (Bennear & Olmstead, 2008), while there is weaker evidence that violations can (Grooms, 2016); our findings motivate developing and adding omnibus water quality measures to these consumer confidence reports, which could be similar to letter grading in K–12 education (e.g., Figlio & Lucas (2004)). Innovative accountability policies such as this may be able to improve water quality even for systems that are not in violation of regulatory standards, an outcome that our results suggest could improve public health.

Appendix for Drinking Water Contaminant Concentrations and Birth Outcomes

DATA CONSTRUCTION DETAILS

While the drinking water regulations are often complicated, in general the maximum contaminant level (MCL) can be seen as the maximum amount of contamination allowed in public drinking water before a regulatory intervention takes place. The intervention may be a violation, which generally includes public reporting requirements (e.g. in the case of coliform or nitrate); or it may be a requirepament for a set of precautionary actions (“treatment techniques”) to be taken to reduce the level of contamination in the water for end-users (e.g. in the case of lead and copper). On the other hand, the maximum contaminant level goals (MCLG) are levels below which there is no known or expected health risks (US Environmental Protection Agency, 2018). As shown in Table A1 for ten selected contaminants, in some instances the MCLG are lower than the MCL.

The main two reasons for MCL to exceed MCLG are that the EPA believes that it is excessively costly for community water systems (CWS) to reduce contamination to the MCLG, or that measurement technology is unable to accurately measure contamination closer to the MCLG (US Environmental Protection Agency (2016b)). Since for some contaminants, the reason MCL are not reduced toward MCLG is due to concerns over imprecise measurements for low levels of contamination, this may raise concerns over the quality of our water sample data. We take the position espoused in the EPA internal document (US Environmental Protection Agency (2006)), describing how measures below MCL still contain (albeit noisy) information: “[i]t is recommended that all values between the PQL and MDL be reported. They are real, the concentration is fuzzy, but their values can give indications or trends and should be reported.” Note that in this quote, both the PQL and MDL are below the MCL: the PQL (practical quantitative limit) is the level at which the EPA believes the samples can be used for administrative purposes while the MDL (method detection level) is a lower level below which the sample reading cannot reliably be distinguished from zero. More detail on how MCL are limited by noise in the water quality measures is available in the EPA six-year review technical report (US Environmental Protection Agency (2016a)).

Discussion of Our Main Index Measures

To illustrate our index measures, in Appendix Table A1, we display ten example contaminants in our data, the number of births with that contaminant measured in our analytic sample (column 1), and the across-birth correlation of that contaminant with our measure of overall contamination (column 2) and chemical-only contamination (column 3). Of course, the correlations reported are conditional on the set of births for which the contaminant is indeed measured, with counts reported in column 1. We find that the disinfectant byproducts trihalomethanes, haloacetic acids, and chlorite, are positively correlated with our measure of overall contamination, as is lead, total coliform (an important microbial measure) and turbidity (a general measure of the cloudiness of water). Note that chlorite is listed by the US EPA as potentially having negative impacts on reproductive health, but it is sampled relatively rarely in our data; yet, for the births for which it is sampled, our index appears to be strongly related to this measure. By contrast, however, nitrate and arsenic, two popular measures in the water policy literature (e.g. Grooms, 2016), are not highly correlated with our measures. Atrazine, a controversial herbicide (Rohr and McCoy, 2010) and as of 2001 one of the most common herbicides detected in streams and ground water (Gilliom and Hamilton, 2006), is only weakly positively correlated with our measures.

Table A1:

Ten selected contaminants, correlation with our two contaminant groups, and regulatory and health standards.

	(1)	(2)	(3)	(4)	(5)
Name	# of births	Cor w/ all	Cor w/ chem	MCL	MCLG
TRIHALOMETHANES	1,144,050	0.179	0.209	0.080	–
HALOACETIC ACIDS (FIVE)	1,130,519	0.134	0.150	0.060	–
NITRATE	1,041,320	0.040	0.094	10.000	10.000
CHLORINE	872,825	0.213	–	4.000	4.000
ARSENIC (IOC)	741,404	−0.035	0.059	0.010	0.000
TOTAL COLIFORM	723,596	0.424	–	0.050	0.000
ATRAZINE (SOC)	590,628	0.120	0.140	0.003	0.003
LEAD	421,453	0.172	0.468	0.015	0.000
TURBIDITY	106,765	0.427	–	2.000	–
CHLORITE	23,247	0.394	−0.037	1.000	0.800

Notes: Table displays the number of births in our analytic sample with a measure on each contaminant (column 1), and the correlation of that measure with the two groups we use in analysis: all in column (2), and chemical-only in column (3). Maximum contaminant limits (MCL) are in column (4) while health goals (MCLG) are in column (5). Dashes (–) indicate that the contaminant is not included in that group, or that there is no MCLG for that contaminant.

For a full list of the contaminants used in our analysis, refer to Appendix Table A2. This table lists all 94 contaminants studied in our analysis, and shows whether it is part of our All or Chem (chemical-only) index measures. It also provides the number of births for which we have non-missing contamination measures (at least one sample during the birth’s gestation period for the CWS at birth) and the total number of drinking water samples for the contaminant in our input data.

Table A2:

Contaminants by group.

Row	ID	Name	All	Chem	Births	Samples
1	999	chlorine	✓	✓	938671	4816645
2	1030	lead	✓	✓	450922	93545
3	1022	copper	✓	✓	465965	93488
4	2950	trihalomethanes	✓	✓	1239417	82338
5	3100	total coliform	✓		785817	77823
6	1040	nitrate	✓	✓	1117361	76439
7	2456	haloacetic acids (five)	✓	✓	1224600	74626
8	1006	chloramine	✓	✓	234710	44909
9	100	turbidity	✓		111782	40059
10	2987	tetrachloroethylene	✓	✓	1039128	38903
11	2984	trichloroethylene	✓	✓	1039602	38891
12	2991	toluene	✓	✓	1039130	38751
13	2980	1,2-dichloroethane	✓	✓	1039073	38747
14	2981	1,1,1-trichloroethane	✓	✓	1039126	38743
15	2990	benzene	✓	✓	1038953	38736
16	2979	trans-1,2-dichloroethylene	✓	✓	1039541	38720
17	2992	ethylbenzene	✓	✓	1037610	38695
18	2378	1,2,4-trichlorobenze	✓	✓	1039042	38691
19	2380	cis-1,2-dichloroethylene	✓	✓	1039068	38674
20	2977	1,1-dichloroethylene	✓	✓	1039169	38673
21	2964	dichloromethane	✓	✓	1038983	38666
22	2982	carbon tetrachloride	✓	✓	1038126	38666
23	2985	1,1,2-trichloroethane	✓	✓	1038971	38662
24	2989	chlorobenzene	✓	✓	1038483	38661
25	2968	o-dichlorobenzene	✓	✓	1038524	38657
26	2969	para-dichlorobenzene	✓	✓	1038518	38650
27	2983	1,2-dichloropropane	✓	✓	1039005	38649
28	2955	xylenes (total)	✓	✓	1037830	38640
29	2996	styrene	✓	✓	1039347	38616
30	1041	nitrite	✓	✓	529308	32221
31	2976	vinyl chloride	✓	✓	643919	27665
32	2941	chloroform (thm)	✓	✓	311636	25217
33	2944	chlorodibromomethane (thm)	✓	✓	309861	25145
34	2943	bromodichloromethane (thm)	✓	✓	311525	25136
35	2942	bromoform (thm)	✓	✓	308017	24730
36	1005	arsenic (ioc)	✓	✓	787789	23956
37	2050	atrazine (soc)	✓	✓	635751	18226
38	2039	di (2-ethyl) phthalate (soc)	✓	✓	698366	17202
39	2037	simazine (soc)	✓	✓	639790	16551
40	2051	alachlor (soc)	✓	✓	592562	16424
41	1025	fluoride (ioc)	✓	✓	450789	16402
42	2042	hexachlorocyclopentadiene(soc)	✓	✓	647018	16276
43	2035	di (2-eth) adipate (soc)	✓	✓	654311	16184
44	2946	ethylene dibromide (edb) (soc)	✓	✓	596121	16124
45	2306	benzo(a)pyrene (soc)	✓	✓	619395	16090
46	2326	pentachlorophenol (soc)	✓	✓	623120	16069
47	2105	2,4 - d (soc)	✓	✓	391691	15890
48	2931	1,2-dribomo-3-chloroprop (soc)	✓	✓	582020	15752
49	2010	lindane (soc)	✓	✓	566031	15515
50	2015	methoxychlor (soc)	✓	✓	542303	15421
51	2959	chlordane (soc)	✓	✓	543058	15175
52	2040	piclorem (soc)	✓	✓	537392	15072
53	2046	carbofuran (soc)	✓	✓	530220	14833
54	2036	oxymal (vydate) (soc)	✓	✓	530677	14805
55	2033	endothall (soc)	✓	✓	527098	14608
56	1010	barium (ioc)	✓	✓	451330	13815
57	1015	cadmium (ioc)	✓	✓	452156	13624
58	1085	thallium (ioc)	✓	✓	451158	13595
59	1075	beryllium (ioc)	✓	✓	450796	13540
60	1074	antimony (ioc)	✓	✓	451315	13536
61	1035	mercury (ioc)	✓	✓	425068	13520
62	1020	chromium (ioc)	✓	✓	452054	13499
63	1045	selenium (ioc)	✓	✓	451787	13484
64	2034	glyphosate (soc)	✓	✓	264887	13452
65	2005	endrin (soc)	✓	✓	314336	13449
66	2274	hexachlorobenzene (soc)	✓	✓	314975	13440
67	2031	dalapon (soc)	✓	✓	338562	13421
68	2065	heptachlor (soc)	✓	✓	310391	13402
69	2020	toxaphene (soc)	✓	✓	297146	13396
70	1024	cyanide (free) (ioc)	✓	✓	438139	13388
71	2110	2,4,5 - tp silvex (soc)	✓	✓	309427	13358
72	2067	heptachlor epoxide (soc)	✓	✓	307172	13347
73	2041	dinoseb (soc)	✓	✓	306771	13267
74	3001	heterotrophic bacteria - enumeration	✓		169026	13123
75	2032	diquat (soc)	✓	✓	277141	13014
76	2383	pcbs (soc)	✓	✓	252770	12321
77	1008	chlorine dioxide	✓	✓	8177	12041
78	1009	chlorite	✓	✓	24122	11888
79	2063	2,3,7,8-tcdd (dioxin) (soc)	✓	✓	185862	10309
80	4030	radium-228	✓	✓	256112	8483
81	4020	radium-226	✓	✓	250946	7979
82	4006	combined uranium	✓	✓	216562	7073
83	4002	alpha/incl. radon & uranium	✓	✓	209737	5128
84	2451	dichloroacetic acid	✓	✓	20452	2783
85	2452	trichloroacetic acid	✓	✓	20452	2781
86	2454	dibromoacetic acid	✓	✓	20452	2779
87	2450	monochloroacetic acid	✓	✓	20410	2775
88	2453	monobromoacetic acid	✓	✓	20408	2774
89	4100	gross beta particle activity	✓	✓	123706	2447
90	1094	asbestos	✓	✓	62226	2157
91	1011	bromate	✓	✓	26251	1839
92	2044	aldicarb sulfone	✓	✓	89377	1568
93	2043	aldicarb sulfoxide	✓	✓	89274	1567
94	2047	aldicarb	✓	✓	81614	1511

Notes: All contaminants used in our analysis are included. The “Births” column reports the number of births for which we observe at least one sample of the given contaminant during its gestation.

Mechanisms for effects of water contamination on health

In stark contrast to air pollution, where research generally focuses on a handful of key contaminants that impact health,¹ there are over 90 drinking water contaminants covered by the National Primary Drinking Water Regulations (NPDWRs), each contaminant may have different potential health effects (US Environmental Protection Agency, 2018), and the research literature appears to provide little guidance on which contaminants are likely to affect birth outcomes.

Bacterial contaminants are especially likely to have short-term effects such as gastrointenstinal illness. While most bacteria that occur in drinking water are harmless, some, such as those from human or animal wastes, may cause short-term, and potentially severe, illnesses. By contrast, US EPA expects most chemical (non-bacterial) contaminants to have health consequences only after long periods of exposure. These health effects may include cancer, liver and kidney damage, reproductive problems, and/or nervous system effects (US Environmental Protection Agency, 2018).

There is a sizeable epidemiological literature on the effects of water pollution on infant health. A detailed summary of the epidemiology literature up to the mid 2000s is provided in Wigle et al. (2008). Wigle et al. (2008) concludes that there is “limited” evidence that prenatal lead exposure leads to greater chance of preterm birth, while there is “inadequate” evidence that arsenic, cadmium, nitrate, and disinfectant byproducts (DBPs) lead to preterm birth. Moreover, they assess that there is “limited” evidence that lead, DBPs, and nitrate exposure will lead to greater chance of low birth weight, while for arsenic and cadmium there is “inadequate” evidence. All contaminants listed here, and many more, are present in our water quality data.

More recently, epidemiologists have found that arsenic exposure in drinking water (Kile et al. (2016)) and, more generally, exposure to toxic metals (Sanders et al. (2014)) is associated with poorer birth outcomes; epidemiologists continue to debate whether nitrate contamination in drinking water affects birth outcomes or infant health (Manassaram et al. (2006)). While useful contributions to our understanding, these studies as a ruledo notuse rigorous econometric methods toisolate exogenous variation in water quality to identify causal effects; one plausible reason for this is small sample sizes. We describe other studies for the five contaminants we include in the reproductive-specific index in Section “An alternative reproductive health index of five contaminants based on data availability and prior research”.

In summary, we do not have a randomized toxicological study for each covered contaminant on human health, so it is unclear whether some of the chemical contaminants may have short-run effects either alone or when combined with other chemical or bacterial contaminants. Indeed, multipollutant exposure analysis is an emerging subfield in epidemiology, especially in the context of air pollution (e.g. see Oakes et al. (2014), Snowden et al. (2015), and Davalos et al. (2017)). These issues motivate exposure analyses that study the effects of pollution sources (e.g. threats to drinking water, such as industrial facilities near source intakes), rather than pollutants per se. In this sense, the approach we take in this paper is traditional: we study the relationship between measured contaminants and health. We leave studies of the effects of drinking water intake exposures to potential point pollution sources as a direction for future research.

In addition, in spite of the NPDWRs regulating a broad range of contaminants, there are still more potentially harmful chemicals or bacteria that could be present in drinking water. Thus, the measures of contamination we observe may be correlated with other, unobserved measures of contamination, which themselves affect fetal health. For example, the EPA currently has over 100 unregulated water contaminants on their “contaminant candidate list,” and in their six-year reviews, US EPA evaluates whether some of these should be added to currently regulated contaminants (US Environmental Protection Agency (2016a)).

For these reasons, in our main analysis we aggregate across contaminants, construct omnibus measures of water quality, and estimate the effects of these overall measures on birth outcomes.

An alternative reproductive health index of five contaminants based on data availability and prior research

The National Primary Drinking Water Regulation provides a brief overview of the main health-based concerns for each contaminant regulated by SDWA. They identify two indicators that could be used to create an index that is specific to birth outcomes: reproductive difficulties, and health effects for infants. There are also studies within epidemiology and economics that have looked at birth outcomes (such as those studied in this paper, i.e., preterm birth, low birth weight) associated with exposure to contamination in drinking water. Appendix Table A3 summarizes the contaminants identified by the EPA and epidemiological/economic literature.

Our goal was to create a summary index of a smaller subset of contaminants that would be more specific to reproductive health and that would be frequently measured in the CWS in Pennsylvania (see Appendix Table A2 for number of births overlapping with contaminant samples). We identified 5 contaminants that had decent overlap between what the NPDWR identifies and the epidemiology literature: arsenic, atrazine, Di(2-ethylhexyl) Phthalate (DEHP), nitrate, and tetrachloroethylene (PCE or PERC). For arsenic, the epidemiological literature is quite large, yet the EPA does not identify it as a contaminant of concern for pregnancy, reproduction, or infants. We choose to include it due to the large associational epi literature (Myers et al., 2010; Saha et al., 2012; Rahman et al., 2017; Bozack et al., 2018).² For atrazine, both the EPA and the epi literature showed concerns for birth outcomes (Rinsky et al., 2012; Migeot et al., 2013; Stayner et al., 2017; Almberg et al., 2018; Porpora et al., 2019).³ DEHP is implicated in reproductive difficulties as well as associated with poor birth outcomes such as preterm birth in drinking water, soil, and food (Ferguson et al., 2014)⁴. Nitrate can be problematic for infants due to blue baby syndrome and has a growing set of studies both in epidemiology and economics suggesting it can be problematic for birth outcomes (Migeot et al., 2013; Stayner et al., 2017; Sherris et al., 2021l; Coffman et al., 2022). ⁵ Finally, we include tetrachloroehtylene due to a growing epi literature suggesting an association with birth outcomes (Sonnenfeld et al., 2001; Aschengrau et al., 2018, 2020).⁶. Much of the literature supporting the NPDWR are over 20 years old, and so we decided to include contaminants that had recent support for a possible effect on birth outcomes (Porpora et al., 2019).

It is important to note that while the EPA and epidemiological literature (including economic literature) suggests strong impacts for lead exposure and infant health (Dave and Yang, 2022; Grossman and Slusky, 2019), lead is measured every 3 years per the regulation and so we decided not to include it in this index due to also wanting an index made up of contaminants that are frequently sampled. We chose not to include disenfectant byproducts in our sub-index due to a growing literature suggesting that early research that found associations between DBP and birth outcomes may not have been replicable with recent data (Savitz et al., 2006; Horton et al., 2011; Mashau et al., 2018). Lead and DBPs are in our omnibus index that is used in the main specification.

Non-detection and Limits of Measurement

Recall that non-detect values are reported in the Pennsylvania sampling data we use as zeroes, and we treat them as zeroes throughout our analysis. One concern is that treating these as zeroes inflates the distance between the 10th and the 90th percentile, which we use to standardize our estimates. It is possible that the 10th percentile is a zero and the 90th percentile is very large, due to the nondetects being zeroes. However, while many states have non-detection levels for reporting that are relatively high, Pennsylvania Department of Environmental Protection’s (PA DEP) policy is to have testing laboratories report their sample results as long as they are above the EPA’s designated Method Detection Limit (MDL), which varies depending on contaminant and method of analysis. As a result, Pennsylvania provides an ideal case to study variation in water quality conditional on regulatory compliance, since less data is suppressed than in many other states. For supporting evidence we compare PA DEP’s reporting thresholds with California’s for a subset of contaminants for which we can do this, and find that PA DEP’s reporting thresholds are much lower. See Tables A4 and A5. PA DEP’s reporting thresholds can be found in documents posted online at http://www.depgreenport.state.pa.us/elibrary/GetFolder?FolderID=690145. California’s detection limits for purposes of reporting can be found at https://www.waterboards.ca.gov/drinking_water/certlic/drinkingwater/documents/mclreview/mcls_dlrs_phgs.pdf.

Table A3:

Drinking water contaminants associated with birth outcomes and reproductive difficulties.

	NPDWR Indicated Health Risks		Epi Lit
Contaminant	Reproductive	Infant	Birth Outcomes
1,2-Dibromo-3- chloropropane (DBCP)	x
Arsenic			x
Atrazine	x		x
Benzo(a)pyrene (PAHs)	x
Carbofuran	x
Chlorine dioxide		x
Chlorite		x
Di(2-ethylhexyl) adipate	x
Di(2-ethylhexyl) phthalate (DEHP)	x		x
Dinoseb	x
Dioxin (2,3,7,8-TCDD)	x
Disenfectant Byproducts (DBP)			x
Ethylene dibromide	x
Fecal coliform and E. coli		x
Glyphosate	x
Hexachlorobenzene	x
Lead		x	x
Methoxychlor	x
Nitrate		x	x
Nitrite		x
Polychlorinated biphenyls (PCBs)	x
Tetrachloroethylene			x
Trichloroethylene			x

Notes: This table lists contaminants identified in the National Primary Drinking Water Regulation Table as having either an impact on reproduction or infants. We also identify some contaminants that have a large literature within epidemiology (or economics, i.e., lead) studying the association between exposure in drinking water and birth outcomes. Some of these contaminants have mixed results in this literature. We bold the 5 contaminants that we include in our 5-contaminant reproductive-specific water quality index.

Table A4:

Comparing non-detect levels in PA and CA.

ID	Name	Method	PA DLR = EPA MDL	CA DLR	% Difference
All VOCs	All VOCs	All Methods	0.0005	0.0005	0
1074	Antimony	ASTM D3697	0.001	0.006	−83.33
1074	Antimony	SM 3113 B	0.003	0.006	−50.00
1074	Antimony	EPA 200.9	8e-04	0.006	−86.67
1074	Antimony	EPA 200.8	4e-04	0.006	−93.33
1005	Arsenic	SM 3114 B; ASTM D1972	0.001	0.002	−50.00
1005	Arsenic	SM 3113 B; ASTM D2972	0.001	0.002	−50.00
1005	Arsenic	EPA 200.9	5e-04	0.002	−75.00
1005	Arsenic	EPA 200.8	0.0014	0.002	−30.00
1010	Barium	SM 3111 D	0.1	0.1	0.00
1010	Barium	SM 3113 B	0.002	0.1	−98.00
1010	Barium	EPA 200.7; SM 3120 B	0.002	0.1	−98.00
1075	Beryllium	SM 3113 B; ASTM D3645	2e-04	0.001	−80.00
1075	Beryllium	EPA 200.9	2e-05	0.001	−98.00
1075	Beryllium	EPA 200.7; SM 3120 B	3e-04	0.001	−70.00
1075	Beryllium	EPA 200.8	3e-04	0.001	−70.00
1015	Cadmium	SM 3113 B	1e-04	0.001	−90.00
1015	Cadmium	EPA 200.7	0.001	0.001	0.00
1020	Chromium	SM 3113 B	0.001	0.01	−90.00
1020	Chromium	EPA 200.7; SM 3120 B	0.007	0.01	−30.00
1024	Cyanide, free	SM 4500-CN- C, E; ASTM D2036A	0.02	0.1	−80.00
1024	Cyanide, free	EPA 335.4	0.005	0.1	−95.00
1024	Cyanide, free	SM 4500-CN- C, F	0.05	0.1	−50.00
1024	Cyanide, free	SM 4500-CN- C, G; ASTM D2036B	0.02	0.1	−80.00
1024	Cyanide, free	Kelada-01	5e-04	0.1	−99.50
1024	Cyanide, free	QuikChem 10-204-00-1-X	6e-04	0.1	−99.40
1024	Cyanide, free	ASTM D6888; OIA-1677, DW (ALPKEM)	5e-04	0.1	−99.50
1035	Mercury	EPA 245.1; SM 3112 B; ASTM D3223	2e-04	0.001	−80.00
1035	Mercury	EPA 245.2	2e-04	0.001	−80.00
1036	Nickel	SM 3113 B	0.001	0.01	−90.00
1036	Nickel	EPA 200.9	6e-04	0.01	−94.00
1036	Nickel	EPA 200.7; SM 3120 B	0.005	0.01	−50.00
1036	Nickel	EPA 200.9	5e-04	0.01	−95.00
1045	Selenium	SM 3114 B; ASTM D3859	0.002	0.005	−60.00
1045	Selenium	SM 3113 B; ASTM D3859	0.002	0.005	−60.00
1085	Thallium	EPA 200.9	7e-04	0.001	−30.00
1085	Thallium	EPA 200.8	3e-04	0.001	−70.00
1094	Asbestos	EPA 100.1,100.2	0.01	0.2	−95.00
1040	Nitrate	SM 4500-NO3- E; ASTM D3867	0.01	0.4	−97.50
1040	Nitrate	EPA 300.0, 300.1; SM 4110 B; ASTM D4327; B-1011 (Waters)	0.01	0.4	−97.50
1040	Nitrate	SM 4500-NO3- D; 601 (ATI Orion)	1	0.4	150.00
1040	Nitrate	EPA 353.2; SM 4500-NO3- F; ASTM D3867	0.05	0.4	−87.50
1040	Nitrate	ASTM D6508	0.076	0.4	−81.00
1041	Nitrite	SM 4500-NO3- E; ASTM D3867	0.01	0.4	−97.50
1041	Nitrite	EPA 300.0, 300.1; SM 4110 B; ASTM D4327; B-1011 (Waters)	0.004	0.4	−99.00
1041	Nitrite	EPA 353.2; SM 4500-NO3- F; ASTM D3867	0.05	0.4	−87.50
1041	Nitrite	SM 4500-NO2- B	0.01	0.4	−97.50
1041	Nitrite	ASTM D6508	0.103	0.4	−74.25

However, notwithstanding PA DEP’s data advantages, it is still theoretically possible that the currently-used 10th to 90th percentile difference is inflated due to the “jump to zero” effect of detection limits. In order to investigate whether this concern is empirically relevant, we provide below our index measure when we use the zeroes provided in the data (which are below EPA MDL) versus when we replace cases below the MDLs with the corresponding EPA MDL (for the contaminant and method combination). We merged the contaminant method detection limits provided by PA with our data, obtaining data on 69 contaminants. To be specific, to construct the following table, for simplicity we (1) took the maximum MDL by contaminant (because contaminants have multiple possible methods of analysis, there are multiple MDLs for each contaminant), then (2) for every birth that is exposed to an average level of contamination (for a particular contaminant) below the MDL, we replace that birth’s average level of contamination with the MDL (since their low level of contamination must be driven by having some nondetects, which are treated as zero). Furthermore, note that after we replace these birth-specific contamination measures in this way, our index measures will increase for every birth, since our index measures are averaged over contaminants including those with replaced-with-MDL values, and those without replaced-with-MDL values. As a result, not only does the 10th percentile rise, but also the 90th percentile will rise, as the entire distribution of contamination shifts to the right (based on how many nondetects each birth experiences during gestation).

Appendix Table A6 shows our findings from this exercise. For these 69 contaminants, the 10th-to-90th percentile difference (for these contaminants) was 0.03 (0.03 – 0) prior to the replacewith-MDL (Panels A1 and B1), but rises to 2.19 (2.25 – 0.06) after replace-with-MDL (because Coliform is not one of the contaminants on the list, there is no difference between the All and Chemical rows in this table). Our interpretation is that, were we to try to adjust our analysis by replacing non-detects with the detection limit, this would in fact increase the 10th-to-90th difference that we use, so our estimates are currently reported in a way that makes them look conservative.

Community Water System Boundaries and Linking Births to Water Quality

Recall that we calculate the average exposure to contamination for each birth during gestation for the system where the mother resides on the birth certificate. Except for differences in gestation length, this means that the contamination measure is the same for two births born on the same date in the same water system.

This raises the concern that there could be within water system heterogeneity in water quality. The water samples we have are matched to water system, but not at any more granular level, and there are a few very large (population) water systems in Pennsylvania. Consider two of the larger systems: the Philadelphia Water Department and the Pittsburgh Water and Sewer Authority.

The Philadelphia Water Department is, roughly-speaking, two water distribution systems, one that intakes from the Delaware River (eastern side of city) and one that intakes from the Schuylkill River (western side of the city). See page 6 of https://water.phila.gov/pool/files/2019-pwd-water-quality-report.pdf for details; we reproduced the map from that page below in Figure A1.

While we analyze this water system as one system, there are indeed two systems involved here, and, actually, three treatment plants (two sourcing from the Schuylkill), although there are places of overlap that would complicate analysis even if we had more granular data. Our sampling data, which we average over the samples reported by the system (at different and unknown to us distribution plants or, in some cases, different points in the distribution) thus represents only a proxy for the water quality experienced by each individual, and not a perfect measure.

Table A5:

Comparing non-detect levels in PA and CA, continued.

ID	Name	Method	PA DLR = EPA MDL	CA DLR	% Difference
4002	Gross Alpha particle	All methods	3	3	0.00
4020	Radium-226	All methods	1	1	0.00
4030	Radium-228	All methods	1	1	0.00
4006	Uranium	All methods	1	1	0.00
4100	Gross beta	All methods	4	4	0.00
4174	Strontium-90	All methods	2	2	0.00
4102	Tritium	All methods	1000	1000	0.00
2051	Alachlor	All methods	0.00044	0.001	−56.00
2050	Atrazine	All methods	0.00022	5e-04	−56.00
2306	Benzo(a)pyrene	All methods	4.4e-05	1e-04	−56.00
2046	Carbofuran	All methods	0.000198	0.005	−96.04
2959	Chlordane	All methods	0.00044	1e-04	340.00
2105	2,4-D	All methods	0.00022	0.01	−97.80
2031	Dalapon	All methods	0.0022	0.01	−78.00
2931	Dibromochloropropane (DBCP)	All methods	4.4e-05	1e-05	340.00
2035	Di(2-ethylhexyl)adipate	All methods	0.00132	0.005	−73.60
2041	Dinoseb	All methods	0.00044	0.002	−78.00
2032	Diquat	All methods	0.00088	0.004	−78.00
2033	Endothall	All methods	0.0198	0.045	−56.00
2005	Endrin	All methods	2.2e-05	1e-04	−78.00
2946	Ethylene dibromide (EDB)	All methods	2.2e-05	2e-05	10.00
2034	Glyphosate	All methods	0.0132	0.025	−47.20
2065	Heptachlor	All methods	8.8e-05	1e-05	780.00
2067	Heptachlor epoxide	All methods	4.4e-05	1e-05	340.00
2274	Hexachlorobenzene	All methods	0.00022	5e-04	−56.00
2042	Hexachlorocyclopentadiene	All methods	0.00022	0.001	−78.00
2010	Lindane	All methods	4.4e-05	2e-04	−78.00
2015	Methoxychlor	All methods	0.00022	0.01	−97.80
2036	Oxymal (Vydate)	All methods	0.0044	0.02	−78.00
2326	Pentachlorophenol	All methods	8.8e-05	2e-04	−56.00
2040	Picloram	All methods	0.00022	0.001	−78.00
2037	Simazine	All methods	0.000154	0.001	−84.60
2063	2,3,7,8-TCDD (Dioxin)	All methods	1.1e-08	5e-09	120.00
2020	Toxaphene	All methods	0.0022	0.001	120.00
2110	2,4,5-TP (Silvex)	All methods	0.00044	0.001	−56.00
2383	PCBs	All methods	0.00022	5e-04	−56.00
1009	Chlorite	All methods	0.02	0.02	0.00

Table A6:

Distribution of drinking water contamination experienced by Pennsylvanian births during gestation – 69 contaminants for which we observe thresholds for reporting by the PA DEP.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Contaminant	Mean	SD	10th	25th	50th	75th	90th	Max	Births
Panel A1: Trimmed sample - nondetects are zero (births used in analysis)
All	0.01	0.03	0.00	0.00	0.01	0.02	0.03	0.28	1,148,370
Chemical	0.01	0.03	0.00	0.00	0.01	0.02	0.03	0.28	1,148,370
Panel B1: Raw sample - nondetects are zero (including births not used)
All	0.02	0.32	0.00	0.00	0.01	0.02	0.03	230.00	1,160,108
Chemical	0.02	0.32	0.00	0.00	0.01	0.02	0.03	230.00	1,160,108
Panel A2: Trimmed sample - nondetects replaced with nondetect limit (births used in analysis)
All	0.44	0.86	0.06	0.08	0.10	0.13	2.25	4.57	1,148,271
Chemical	0.44	0.86	0.06	0.08	0.10	0.13	2.25	4.57	1,148,271
Panel B2: Raw sample - nondetects replaced with nondetect limit
All	0.52	2.14	0.06	0.08	0.10	0.13	2.28	230.00	1,160,108
Chemical	0.52	2.14	0.06	0.08	0.10	0.13	2.28	230.00	1,160,108

Figure A1:

Philadelphia water distribution system map

Turning to a second example, the Pittsburgh Water and Sewer Authority sources its water entirely from the Allegheny River, and has one water treatment plant (https://www.pgh2o.com/your-water/water-quality-treatment). In this case the concern appears less warranted.

We thus recognize inherent measurement limitations in linking mother exposure to water quality using CWS of residence. To address this concern, we include a sensitivity analysis where we run our main analysis, dropping large systems. We include the results in Appendix Tables A23 and A24 (later in this Appendix). We did this by sorting births by system size (number of people served by the CWS) and then dropping the 50% of births in our sample in the largest systems (those in the bottom of the dataset). The largest system in this restricted sample serves 118,000 people (contrast with the largest system, Philadelphia Water Department, which serves 1,600,000 people), and thus this is a meaningful restriction. As the tables show, our main results are robust to this sample restriction.

A second concern about the births to CWS linkage we conduct is that the boundaries of water systems change over time. As mentioned in the main text, the vintage of the community water service map we use as the 2016, month 4, version of the data. We used this vintage since it was the latest available at the start of our research. However, PennState University’s Pennsylvania Spatial Data Access (PASDA), the source of the data, has versions of this data from 2011 to 2022 (https://www.pasda.psu.edu/download/dep/historic/PublicWaterSupply/).

This suggests the possibility of using multiple versions of the boundary data to better measure spatial relationship between mothers and water providers. But since 2011 is near the end of our sample frame (2003–2014), the range of data that is available is of limited use for longitudinal analysis in our study. An additional concern we have about studying multiple years of these service boundaries is it is unclear to us whether the changes from year-to-year are corrections of service boundaries reported, or whether they are true additions or subtractions. With this in mind, the following two maps, in Appendix Figures A2 and A3, compare the 2011/04 vintage with 2016/04 (first map), and 2016/04 with 2022/10 (second map). Blue areas are CWS coverage gains, while orange areas are CWS losses.

Clearly there are some significant changes in these boundaries, but it is less clear whether these are true changes or measurement error (corrections of the maps, or mistakes in reporting).

To shed further light on these issues, we argue that concerns regarding boundary changes could take two forms: (1) if boundaries change randomly (measurement error for example), then that will add noise into our independent variable as births will not be being correctly matched to contamination. This classical measurement error would attenuate our effect estimates. (2) Boundary changes could be an omitted variable (confounder), that is, they could be systematically correlated with contamination and outcomes, and could thus bias our estimates. The two maps we provided above suggest that boundary changes are likely to be correlated with the demographics of the system, as it is the suburban and rural areas that expand or contract, not the major cities. As a result, we admit the possibility that boundary changes may be correlated with outcomes (through socioeconomic status, for example). We instead focus on the question of whether boundary changes are correlated with levels of CWS contamination.

For this auxiliary analysis, we examined CWS level average contamination (taken over the last nine months) for our two omnibus measures, all and chemical-only RRMCL, averaged over the contaminants for which we have samples. We then regressed the change in CWS contamination from 2011/04 to 2015/12 (the first and last month available in our sampling data) on (a) whether the CWS had any meaningful net increase in area (i.e., above the very trivial threshold of 100 square meters) or decrease (i.e. below the threshold of 100 square meters lost) (columns (1) and (3) in the table below), with the excluded group of no meaningful change; and (b) the net change in square kilometers (columns (2) and (3) in the table below). The results of these four regressions we include in the table below. As the table shows, there is no statistically significant relationship between the change in CWS contamination over this period, and CWS boundary changes (though, note that column 4 is statistically significant at the 10% level). This suggests that boundary changes are unlikely to be an omitted variable, since they are not strongly related to CWS contamination levels.

Figure A2:

Changes in PWS boundaries 2011 compared to 2016.

Notes: Blue areas are CWS coverage gains, while orange areas are CWS losses.

SENSITIVITY ANALYSES

This section repeats our main analysis using different sample restrictions and different indices of contamination. Regressions estimating the effects of water contamination on birth outcomes are repeated for each subsample or index definition.

Figure A3:

Changes in PWS boundaries 2016 compared to 2022.

Since much of our sensitivity analysis rests on sample restrictions, we begin by more fully comparing our analytic sample to the full population of births (including those not matched to water systems, plural, and exposed to MCL). To do so, we add an additional column to our summary statistics table (Table 2 in the main text) which includes all births. This is included as Appendix Table A8. (We dropped some columns from this version of the summary statistics table to conserve space.)

Note that in the main text, we exclude plural births. The first change in sample we explore is including plural births. Results are in Appendix Tables A9 and A10.

Concern over to what extent our results are driven by mother mobility motivates the second change in sample: we re-run our models, excluding all mothers (with multiple births) that ever switch a CWS across births. Results are shown in Appendix Tables A11 and A12.

Interest in how our results would change if we keep births exposed to MCL violations, or remove births with high samples more rigorously, motivates our third and fourth sets of sensitivity analyses. We re-run the models including births exposed to MCL, and re-run the models excluding births linked to any samples exceeding the corresponding MCL. For this last sensitivity analyses, four contaminants present a challenge due to samples very often exceeding their MCLs, so we further estimate models that drop these four contaminants from our index construction, and models that keep these four contaminants (with the concomitant substantial reduction in sample size). Our results are included in Appendix Tables A15 through A20.

Table A7:

Change in contamination correlated with changes in CWS boundaries.

	(1)	(2)	(3)	(4)
	ΔAll	ΔAll	ΔChemical	ΔChemical
Meaningful Increase	0.0353 (0.0247)		0.00839 (0.00755)
Meaningful Decrease	0.00406 (0.0238)		−0.00868 (0.0120)
Net Δ Square Km		0.000225 (0.000610)		0.000436* (0.000251)
Constant	−0.0108 (0.00852)	−0.00688 (0.00763)	0.00314 (0.00311)	0.00349 (0.00278)
N	1546	1546	1546	1546

Notes: Standard errors in parentheses

^*p < 0.10

^**p < 0.05

^***p < 0.01

Concern over our sample changing between the mom FE models and the CWS FE models motivates our next analysis, where we force these samples to be identical. We saw in the main text that estimates were sensitive to which fixed-effects structure we employed, in spite of mother and birth characteristics being similar across the two samples; it is interesting to see whether the effect estimates remain different after forcing the samples to be identical. Results are in Appendix Tables A21 and A22.

Finally, concern over mismeasurement of the link between water contamination and exposure, that may be especially acute in large CWS (with complicated distribution systems), motivates a sensitivity analysis estimating our models for smaller CWS only. Results are in Appendix Tables A23 and A24.

We present each of these sensitivity analyses in the following appendix subsections.

Table A8:

Sample means and, when appropriate, standard deviations, for selected statistics and subsamples of births.

	(1)	(2)	(3)	(4)
Characteristic	All Births	NP No MCL	Analytic Sample	Mom FE Sample
Sample Sizes
No. of births	1,608,619	1,241,656	1,216,132	700,815
No. of moms	992,451	832,674	820,173	304,934
No. of CWSs	1,542	1,537	1,459	1,343
Outcomes
Low birth weight	0.083	0.067	0.067	0.065
Preterm birth	0.099	0.082	0.082	0.081
Small for gestational age	0.099	0.095	0.096	0.091
Term birth weight (g)	3,398 (475.2)	3,396 (468.9)	3,396 (468.8)	3,402 (467.2)
CWS births per 100k per month	117.7 (110.8)	113.4 (107.4)	112.7 (103.3)	112.1 (98.59)
Other characteristics
Mom age (years)	28.09 (6.032)	27.87 (6.054)	27.87 (6.057)	27.69 (5.858)
Mom Black	0.177	0.210	0.213	0.220
Mom Hispanic	0.058	0.068	0.069	0.074
Mom white, not Hispanic	0.765	0.722	0.718	0.713
HS or less	0.403	0.401	0.401	0.406
Mom smokes	0.219	0.224	0.223	0.216
Mom married	0.606	0.569	0.568	0.578
WIC/Medicaid	0.442	0.477	0.478	0.482
Different CWS next birth	0.406	0.369	0.368	0.368
Mean RRCML, all contaminants	0.104 (2.978)	0.095 (2.237)	0.078 (0.078)	0.078 (0.077)
Mean RRCML, chemical only	0.095 (3.201)	0.085 (2.238)	0.069 (0.080)	0.069 (0.079)

Notes: Sample consists of all births. Means (for indicator variables, proportions) are displayed, with standard deviations in parentheses when appropriate; NP = non-plural birth. Column (1) provides statistics for for all births, including those not matched to water systems. Column (2) restricts to non-plural births without MCLs that are matched to water systems. Column (3) restricts to our analytic sample, keeping only births with an overall measure of contamination and for which that measure is below the 99th percentile. Column (4) describes births in our mother fixed-effects specifications.

All births (including plural)

Table A9:

Effects of contamination on low birth weight and preterm birth, all births (including plural).

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: low birth weight
CWS FE
10th to 90th RRMCL	0.01817*** (0.00393)	0.01732*** (0.00165)	0.01363*** (0.00309)	0.01308*** (0.00122)	0.01137*** (0.00259)	0.01009*** (0.00137)
Observations	1,262,235	1,260,752	1,262,235	1,260,752	1,262,235	1,260,752
Adj R2	0.0043	0.0047	0.1549	0.1551	0.2079	0.2080
Mom & CWS FE
10th to 90th RRMCL	0.01242*** (0.00308)	0.01214*** (0.00137)	0.01010*** (0.00266)	0.00983*** (0.00120)	0.00884*** (0.00230)	0.00764*** (0.00136)
Observations	754,689	753,372	754,689	753,372	754,689	753,372
Adj R2	0.2738	0.2741	0.3455	0.3456	0.3752	0.3753
Panel B: preterm birth
CWS FE
10th to 90th RRMCL	0.02489*** (0.00506)	0.02355*** (0.00208)	0.02004*** (0.00402)	0.01905*** (0.00160)	0.01721*** (0.00329)	0.01531*** (0.00170)
Observations	1,262,235	1,260,752	1,262,235	1,260,752	1,262,235	1,260,752
Adj R2	0.0034	0.0039	0.1453	0.1457	0.2209	0.2212
Mom & CWS FE
10th to 90th RRMCL	0.01808*** (0.00415)	0.01774*** (0.00178)	0.01553*** (0.00365)	0.01524*** (0.00156)	0.01407*** (0.00297)	0.01255*** (0.00165)
Observations	754,689	753,372	754,689	753,372	754,689	753,372
Adj R2	0.3016	0.3021	0.3681	0.3684	0.4125	0.4127

Notes: see notes to Table 3 in the main text.

Table A10:

Effects of contamination on small for gestational age and term birth weight, all births (including plural).

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age
CWS FE
10th to 90th RRMCL	0.00444*** (0.00080)	0.00380*** (0.00069)	0.00236*** (0.00075)	0.00170*** (0.00060)	0.00209*** (0.00078)	0.00137** (0.00060)
Observations	1,259,696	1,258,220	1,259,696	1,258,220	1,259,696	1,258,220
Adj R2	0.0050	0.0050	0.0569	0.0568	0.0575	0.0575
Mom & CWS FE
10th to 90th RRMCL	0.00388*** (0.00101)	0.00294*** (0.00087)	0.00306*** (0.00095)	0.00195** (0.00082)	0.00319*** (0.00097)	0.00205** (0.00084)
Observations	751,908	750,588	751,908	750,588	751,908	750,588
Adj R2	0.1627	0.1633	0.1886	0.1892	0.1887	0.1893
Panel B: term birth weight
CWS FE
10th to 90th RRMCL	−11.08*** (1.46)	−10.66605*** (1.30313)	−9.66*** (1.51)	−9.36292*** (1.20495)	−9.46*** (1.65)	−8.69350*** (1.31206)
Observations	1,135,217	1,133,896	1,135,217	1,133,896	1,135,217	1,133,896
Adj R2	0.0157	0.0157	0.1199	0.1200	0.1241	0.1242
Mom & CWS FE
10th to 90th RRMCL	−9.93*** (1.68)	−9.43433*** (1.32991)	−9.15*** (1.80)	−8.56855*** (1.34776)	−9.31*** (1.84)	−8.23830*** (1.42809)
Observations	634,768	633,676	634,768	633,676	634,768	633,676
Adj R2	0.3579	0.3580	0.4151	0.4152	0.4171	0.4171

Notes: see notes to Table 3 in the main text.

Only non-plural births to non-moving moms

In this subsection, we repeat our main specifications providing evidence of the effects of contamination on birth outcomes, except we remove moms who switch CWS at any point across births. Of course, this sample restriction only applies to mothers with multiple births in our data. Our intent is to test whether our results are driven by movers.

Table A11:

Effects of contamination on low birth weight and preterm birth, no movers.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: low birth weight
CWS FE
10th to 90th RRMCL	0.01370*** (0.00320)	0.01502*** (0.00137)	0.01260*** (0.00303)	0.01395*** (0.00126)	0.01130*** (0.00246)	0.01131*** (0.00145)
Observations	885,494	884,502	885,494	884,502	885,494	884,502
Adj R2	0.0053	0.0056	0.0464	0.0466	0.1066	0.1067
Mom & CWS FE
10th to 90th RRMCL	0.00882*** (0.00307)	0.01034*** (0.00136)	0.00841*** (0.00288)	0.00963*** (0.00128)	0.00835*** (0.00228)	0.00808*** (0.00147)
Observations	374,367	373,867	374,367	373,867	374,367	373,867
Adj R2	0.2021	0.2023	0.2240	0.2242	0.2577	0.2578
Panel B: preterm birth
CWS FE
10th to 90th RRMCL	0.01948*** (0.00443)	0.02155*** (0.00192)	0.01819*** (0.00409)	0.02023*** (0.00176)	0.01653*** (0.00325)	0.01690*** (0.00193)
Observations	885,494	884,502	885,494	884,502	885,494	884,502
Adj R2	0.0037	0.0042	0.0467	0.0472	0.1314	0.1317
Mom & CWS FE
10th to 90th RRMCL	0.01321*** (0.00463)	0.01616*** (0.00204)	0.01254*** (0.00444)	0.01518*** (0.00194)	0.01260*** (0.00341)	0.01324*** (0.00202)
Observations	374,367	373,867	374,367	373,867	374,367	373,867
Adj R2	0.1917	0.1920	0.2178	0.2180	0.2658	0.2658

Notes: see notes to Table 3 in the main text.

Table A12:

Effects of contamination on small for gestational age and term birth weight, no movers.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age
CWS FE
10th to 90th RRMCL	0.00432*** (0.00090)	0.00330*** (0.00076)	0.00364*** (0.00085)	0.00278*** (0.00071)	0.00344*** (0.00091)	0.00247*** (0.00075)
Observations	884,128	883,138	884,128	883,138	884,128	883,138
Adj R2	0.0059	0.0059	0.0386	0.0386	0.0396	0.0397
Mom & CWS FE
10th to 90th RRMCL	0.00323*** (0.00122)	0.00160 (0.00103)	0.00325*** (0.00123)	0.00149 (0.00104)	0.00365*** (0.00126)	0.00194* (0.00105)
Observations	373,142	372,638	373,142	372,638	373,142	372,638
Adj R2	0.1887	0.1890	0.1984	0.1988	0.1985	0.1989
Panel B: term birth weight
CWS FE
10th to 90th RRMCL	−11.41*** (1.56)	−10.70*** (1.37)	−10.34*** (1.66)	−10.09*** (1.30)	−10.38*** (1.82)	−9.56821*** (1.46)
Observations	812,779	811,873	812,779	811,873	812,779	811,873
Adj R2	0.0182	0.0182	0.0942	0.0944	0.0986	0.0988
Mom & CWS FE
10th to 90th RRMCL	−8.53*** (2.30)	−8.78*** (1.69)	−8.67*** (2.34)	−8.72*** (1.73)	−9.27*** (2.41)	−8.86*** (1.83)
Observations	327,108	326,672	327,108	326,672	327,108	326,672
Adj R2	0.4672	0.4674	0.4941	0.4942	0.4955	0.4957

Notes: see notes to Table 3 in the main text.

Including births exposed to MCL violations

In this subsection, we repeat our main specifications except we do not exclude births exposed to MCL violations.

Table A13:

Effects of contamination on low birth weight and pre-term birth, including births exposed to MCL.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: Low birth weight
CWS FE
10th to 90th RRMCL	0.01306*** (0.00277)	0.01449*** (0.00129)	0.01200*** (0.00263)	0.01342*** (0.00120)	0.01016*** (0.00217)	0.01064*** (0.00130)
Observations	1,257,166	1,255,522	1,257,166	1,255,522	1,257,166	1,255,522
Adj R2	0.0050	0.0053	0.0452	0.0455	0.1050	0.1053
Mom & CWS FE
10th to 90th RRMCL	0.00918*** (0.00225)	0.01059*** (0.00114)	0.00896*** (0.00212)	0.01013*** (0.00111)	0.00780*** (0.00180)	0.00795*** (0.00126)
Observations	738,790	737,372	738,790	737,372	738,790	737,372
Adj R2	0.1482	0.1487	0.1664	0.1668	0.2077	0.2080
Panel B: Preterm birth
CWS FE
10th to 90th RRMCL	0.01948*** (0.00373)	0.02118*** (0.00170)	0.01818*** (0.00346)	0.01987*** (0.00158)	0.01584*** (0.00277)	0.01635*** (0.00162)
Observations	1,257,166	1,255,522	1,257,166	1,255,522	1,257,166	1,255,522
Adj R2	0.0037	0.0042	0.0464	0.0469	0.1308	0.1311
Mom & CWS FE
10th to 90th RRMCL	0.01515*** (0.00333)	0.01688*** (0.00158)	0.01482*** (0.00316)	0.01631*** (0.00152)	0.01347*** (0.00252)	0.01362*** (0.00160)
Observations	738,790	737,372	738,790	737,372	738,790	737,372
Adj R2	0.1465	0.1469	0.1641	0.1644	0.2227	0.2229

Notes: see notes to Table 3 in the main text.

Table A14:

Effects of contamination on small for gestational age and term birth weight, including births exposed to MCL.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age
CWS FE
10th to 90th RRMCL	0.00326*** (0.00073)	0.00291*** (0.00058)	0.00266*** (0.00071)	0.00237*** (0.00054)	0.00242*** (0.00075)	0.00206*** (0.00055)
Observations	1,255,208	1,253,568	1,255,208	1,253,568	1,255,208	1,253,568
Adj R2	0.0054	0.0054	0.0378	0.0378	0.0388	0.0388
Mom & CWS FE
10th to 90th RRMCL	0.00222** (0.00099)	0.00164** (0.00081)	0.00220** (0.00096)	0.00149* (0.00079)	0.00231** (0.00099)	0.00155* (0.00081)
Observations	736,689	735,266	736,689	735,266	736,689	735,266
Adj R2	0.1376	0.1382	0.1532	0.1538	0.1535	0.1541
Panel B: term birth weight
CWS FE
10th to 90th RRMCL	−10.19*** (1.35)	−10.05*** (1.18)	−9.30*** (1.47)	−9.38*** (1.14)	−9.04*** (1.58)	−8.78*** (1.24)
Observations	1,154,988	1,153,455	1,154,988	1,153,455	1,154,988	1,153,455
Adj R2	0.0168	0.0168	0.0941	0.0941	0.0985	0.0985
Mom & CWS FE
10th to 90th RRMCL	−9.15*** (1.69)	−9.19*** (1.29)	−9.28*** (1.79)	−8.93*** (1.35)	−9.40*** (1.83)	−8.58*** (1.43)
Observations	647,223	645,961	647,223	645,961	647,223	645,961
Adj R2	0.3400	0.3400	0.3795	0.3795	0.3816	0.3816

Notes: see notes to Table 3 in the main text.

Removing births with any samples exceeding MCL

In this subsection, we repeat our main specifications except not only do we exclude births exposed to reported MCL violations, we also exclude births exposed to any samples that were in excess of MCL. Independent audits of EPA’s reported MCL violations around the year 2000 found that a large share of drinking water violations were not reported accurately (see the references and discussion in Bennear & Olmstead, 2008, who note that for some contaminants around 85% of MCL violations were not accurately reported), so there is concern that the MCL violations data provided by the US EPA is imperfect. However, by 2011, the middle of our study period, close to 75% of health-based violations were accurately reported across states to EPA (Baker et al., 2023). To provide additional statistics for this question, we requested the MCL violations from Pennsylvania and compared them to the data on health-based violations we FOIA’d the EPA for. We combined the EPA and PA violations from 2000–2015 and found that 80% of them were in both files, 6% were in PA only and 14% were in EPA only. Our main specifications used the EPA violations as the gold standard, making up 94% of the total reported violations.

While our analysis and more recent literature suggests the EPA violations data is of reasonable quality, due to lingering concerns that violations might not be accurately reported or measured, we use the sampling data to identify individual sampling concentrations (results) that are above the MCL. The alternative regressions displayed in Tables A15 through A20 test whether our findings are sensitive to a stronger definition of “compliant drinking water” where we exclude births exposed to any samples in excess of the MCL.⁷

There are four contaminants where births very often have samples that exceed the MCL: Total Coliform, TTHM, HAA5, or lead. Note that none of these contaminants have MCL violations generated using single samples: the Total Coliform rule uses monthly aggregates, TTM and HAA5 use running annual averages, and lead uses 90th percentiles. It is therefore perfectly consistent that samples often exceed their corresponding MCL, but do not trigger regulatory violations, in these cases. Due to these contaminants, we estimate three versions of this sensitivity analysis. In our first version, included in Appendix Tables A15 and A16, we construct our water quality index measure the same way as we do in the text, including these four contaminants, but we do not use these four contaminants to drop births exposed to samples that exceed their MCL. This first version maximizes our sample size for this robustness check, while keeping our index consistent with what we calculate in the main text. In our second version, included in Appendix Tables A17 and A18, we simply do not include these contaminants in the analysis at all: not in the index, nor in the sample restriction. In our third version, included in Appendix Tables A19 and A20, we include these in the index and drop births exposed to exceedences for any and all contaminants (including hese four contaminants), which leads to a substantial reduction in sample size.

The results from these sensitivity analyses are displayed in the following tables. Our findings are, for the most part, robust to these changes.

Table A15:

Effects of contamination on low birth weight and preterm birth, removing births with any samples exceeding MCL, except for four contaminants with widespread exceedances (total coliform, TTHM, HAA5, and lead).

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: low birth weight
CWS FE
10th to 90th RRMCL	0.01461*** (0.00280)	0.01456*** (0.00155)	0.01345*** (0.00262)	0.01345*** (0.00143)	0.01652*** (0.00175)	0.01035*** (0.00174)
Observations	1,004,287	1,002,929	1,004,287	1,002,929	1,004,287	1,002,929
Adj R2	0.0052	0.0054	0.0460	0.0462	0.1065	0.1066
Mom & CWS FE
10th to 90th RRMCL	0.01180*** (0.00194)	0.01116*** (0.00133)	0.01130*** (0.00182)	0.01060*** (0.00125)	0.01302*** (0.00154)	0.00818*** (0.00152)
Observations	511,624	510,569	511,624	510,569	511,624	510,569
Adj R2	0.1533	0.1532	0.1715	0.1714	0.2122	0.2119
Panel B: preterm birth
CWS FE
10th to 90th RRMCL	0.02186*** (0.00382)	0.02154*** (0.00210)	0.02045*** (0.00351)	0.02016*** (0.00192)	0.02476*** (0.00235)	0.01617*** (0.00225)
Observations	1,004,287	1,002,929	1,004,287	1,002,929	1,004,287	1,002,929
Adj R2	0.0040	0.0043	0.0472	0.0475	0.1314	0.1315
Mom & CWS FE
10th to 90th RRMCL	0.01865*** (0.00304)	0.01816*** (0.00191)	0.01806*** (0.00287)	0.01751*** (0.00180)	0.02150*** (0.00221)	0.01451*** (0.00207)
Observations	511,624	510,569	511,624	510,569	511,624	510,569
Adj R2	0.1567	0.1568	0.1740	0.1740	0.2299	0.2296

Notes: see notes to Table 3 in the main text. In this analysis, we do not include births exposed to water quality samples exceeding their respective MCL for any contaminants except four with widespread exceedences (Total Coliform, TTHM, HAA5, and Lead).

Table A16:

Effects of contamination on small for gestational age and term birth weight, removing births with any samples exceeding MCL, except for four contaminants with widespread exceedances (total coliform, TTHM, HAA5, and lead).

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age
CWS FE
10th to 90th RRMCL	0.00309*** (0.00078)	0.00247*** (0.00063)	0.00249*** (0.00075)	0.00197*** (0.00059)	0.00242*** (0.00073)	0.00160*** (0.00058)
Observations	1,002,592	1,001,241	1,002,592	1,001,241	1,002,592	1,001,241
Adj R2	0.0052	0.0052	0.0378	0.0379	0.0389	0.0389
Mom & CWS FE
10th to 90th RRMCL	0.00219* (0.00117)	0.00101 (0.00107)	0.00194* (0.00112)	0.00075 (0.00103)	0.00092* (0.00127)	0.00077 (0.00103)
Observations	510,015	508,959	510,015	508,959	510,015	508,959
Adj R2	0.1378	0.1382	0.1535	0.1538	0.1537	0.1541
Panel B: term birth weight
CWS FE
10th to 90th RRMCL	−10.49*** (1.34)	−9.18398*** (1.19)	−9.73*** (1.39)	−8.73*** (1.16)	−10.72*** (1.43)	−8.16*** (1.37)
Observations	921,294	920,064	921,294	920,064	921,294	920,064
Adj R2	0.0162	0.0162	0.0935	0.0937	0.0979	0.0981
Mom & CWS FE
10th to 90th RRMCL	−9.67*** (1.83)	−8.01*** (1.55)	−9.33*** (1.83)	−7.57*** (1.54)	−9.30*** (1.89)	−7.27*** (1.65)
Observations	445,586	444,713	445,586	444,713	445,586	444,713
Adj R2	0.3422	0.3423	0.3812	0.3813	0.3833	0.3834

Notes: see notes to Table 3 in the main text, and Table A15 in the appendix.

Table A17:

Effects of contamination on low birth weight and preterm birth, removing births with any samples exceeding MCL, excluding four contaminants (total coliform, TTHM, HAA5, and lead) from the analysis entirely.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: low birth weight
CWS FEs
10th to 90th RRMCL	0.00983*** (0.00108)	0.00904*** (0.00100)	0.00903*** (0.00100)	0.00832*** (0.00092)	0.00723*** (0.00080)	0.00663*** (0.00093)
Observations	980,718	979,158	980,718	979,158	980,718	979,158
Adj R2	0.0052	0.0053	0.0461	0.0461	0.1067	0.1066
Mom & CWS FEs
10th to 90th RRMCL	0.00793*** (0.00089)	0.00735*** (0.00083)	0.00738*** (0.00084)	0.00681*** (0.00078)	0.00591*** (0.00068)	0.00545*** (0.00080)
Observations	491,834	490,505	491,834	490,505	491,834	490,505
Adj R2	0.1523	0.1522	0.1702	0.1702	0.2112	0.2110
Panel B: preterm birth
CWS FEs
10th to 90th RRMCL	0.01442*** (0.00139)	0.01322*** (0.00128)	0.01341*** (0.00129)	0.01230*** (0.00118)	0.01068*** (0.00102)	0.01019*** (0.00113)
Observations	980,718	979,158	980,718	979,158	980,718	979,158
Adj R2	0.0041	0.0041	0.0472	0.0473	0.1316	0.1316
Mom & CWS FEs
10th to 90th RRMCL	0.01244*** (0.00117)	0.01156*** (0.00109)	0.01177*** (0.00111)	0.01091*** (0.00103)	0.00947*** (0.00089)	0.00924*** (0.00099)
Observations	491,834	490,505	491,834	490,505	491,834	490,505
Adj R2	0.1556	0.1556	0.1725	0.1725	0.2288	0.2287

Notes: see notes to Table 3 in the main text. In these specifications we include all births except those exposed to any samples that exceed their respective MCL. We discard the four contaminants with widespread exceedences (total coliform, TTHM, HAA5, and lead) from the analysis entirely. Since the indices used in this table do not have the same contaminants as the all or chemical indices used elsewhere (due to the exclusion of the four aforementioned contaminants), the distribution for the indices in this table is different. We do not provide detailed statistics for these distributions, but to aid comparison: the distance between the 10th to 90th RRMCL (a one-unit change in the independent variable) is 0.0611 for the All contaminant index used in this table, and it is 0.0551 for the Chemical index used in this table.

Table A18:

Effects of contamination on small for gestational age and term birth weight, removing births with any samples exceeding MCL, excluding four contaminants (total coliform, TTHM, HAA5, and lead) from the analysis entirely.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age
CWS FEs
10th to 90th RRMCL	0.00139*** (0.00050)	0.00120*** (0.00044)	0.00108** (0.00049)	0.00092** (0.00043)	0.00080* (0.00037)	0.00070 (0.00043)
Observations	979,100	977,545	979,100	977,545	979,100	977,545
Adj R2	0.0052	0.0052	0.0378	0.0378	0.0388	0.0388
Mom & CWS FEs
10th to 90th RRMCL	0.00045 (0.00072)	0.00041 (0.00065)	0.00033 (0.00071)	0.00026 (0.00064)	0.00023 (0.00056)	0.00025 (0.00064)
Observations	490,311	488,990	490,311	488,990	490,311	488,990
Adj R2	0.1368	0.1366	0.1525	0.1524	0.1528	0.1527
Panel B: term birth weight
CWS FEs
10th to 90th RRMCL	−5.35*** (0.79)	−4.99*** (0.70)	−5.08*** (0.79)	−4.79*** (0.71)	−4.15*** (0.61)	−4.59*** (0.82)
Observations	899,901	898,475	899,901	898,475	899,901	898,475
Adj R2	0.0161	0.0161	0.0936	0.0936	0.0980	0.0980
Mom & CWS FEs
10th to 90th RRMCL	−5.28*** (1.06)	−4.90*** (0.97)	−5.21*** (1.05)	−4.71*** (0.95)	−4.09*** (0.82)	−4.65*** (0.97)
Observations	428,339	427,191	428,339	427,191	428,339	427,191
Adj R2	0.3396	0.3396	0.3788	0.3788	0.3810	0.3809

Notes: see notes to Table 3 in the main text, and Table A17 in the appendix.

Table A19:

Effects of contamination on low birth weight and preterm birth, removing births with any samples exceeding MCL, including the four contaminants with widespread exceedances (total coliform, TTHM, HAA5, and lead).

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: low birth weight
CWS FEs
10th to 90th RRMCL	0.01971*** (0.00236)	0.01698*** (0.00201)	0.01840*** (0.00223)	0.01586*** (0.00190)	0.01989*** (0.00239)	0.01232*** (0.00191)
Observations	344,101	343,470	344,101	343,470	344,101	343,470
Adj R2	0.0077	0.0077	0.0477	0.0476	0.1176	0.1174
Mom & CWS FEs
10th to 90th RRMCL	0.01311*** (0.00270)	0.01190*** (0.00227)	0.01277*** (0.00256)	0.01148*** (0.00216)	0.01440*** (0.00271)	0.00933*** (0.00211)
Observations	106,463	106,100	106,463	106,100	106,463	106,100
Adj R2	0.1798	0.1790	0.2007	0.2000	0.2419	0.2411
Panel B: preterm birth
CWS FEs
10th to 90th RRMCL	0.02857*** (0.00294)	0.02466*** (0.00248)	0.02677*** (0.00279)	0.02315*** (0.00235)	0.02904*** (0.00295)	0.01855*** (0.00236)
Observations	344,101	343,470	344,101	343,470	344,101	343,470
Adj R2	0.0081	0.0081	0.0512	0.0513	0.1419	0.1419
Mom & CWS FEs
10th to 90th RRMCL	0.02241*** (0.00334)	0.02082*** (0.00281)	0.02161*** (0.00311)	0.02000*** (0.00263)	0.02509*** (0.00329)	0.01720*** (0.00231)
Observations	106,463	106,100	106,463	106,100	106,463	106,100
Adj R2	0.1807	0.1802	0.2019	0.2013	0.2557	0.2548

Notes: see notes to Table 3 in the main text. In this analysis, we do not include births exposed to any water quality sample with a result that exceeds the respective MCL. We include the four contaminants with widespread exceedences in this analysis. This leads to a substantial reduction in sample sizes, as observed in the table.

Table A20:

Effects of contamination on small for gestational age and term birth weight, removing births with any samples exceeding MCL, including the four contaminants with widespread exceedances (total coliform, TTHM, HAA5, and lead).

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age
CWS FEs
10th to 90th RRMCL	0.00292** (0.00115)	0.00241** (0.00100)	0.00262** (0.00113)	0.00215** (0.00097)	0.00270* (0.00122)	0.00183* (0.00098)
Observations	343,395	342,771	343,395	342,771	343,395	342,771
Adj R2	0.0040	0.0040	0.0364	0.0364	0.0377	0.0377
Mom & CWS FEs
10th to 90th RRMCL	−0.00092 (0.00242)	−0.00115 (0.00206)	−0.00050 (0.00242)	−0.00080 (0.00206)	−0.00100 (0.00258)	−0.00083 (0.00206)
Observations	106,009	105,647	106,009	105,647	106,009	105,647
Adj R2	0.1702	0.1698	0.1824	0.1821	0.1826	0.1823
Panel B: term birth weight
CWS FEs
10th to 90th RRMCL	−10.16*** (2.04)	−8.66*** (1.70)	−10.47*** (1.97)	−8.97*** (1.66)	−11.26*** (2.09)	−8.76*** (1.79)
Observations	314,836	314,283	314,836	314,283	314,836	314,283
Adj R2	0.0130	0.0130	0.0883	0.0884	0.0929	0.0929
Mom & CWS FEs
10th to 90th RRMCL	−5.71 (3.97)	−5.38 (3.44)	−7.59** (3.79)	−7.09** (3.27)	−8.89* (4.10)	−7.03** (3.38)
Observations	92,157	91,851	92,157	91,851	92,157	91,851
Adj R2	0.4015	0.4010	0.4357	0.4352	0.4380	0.4375

Notes: see notes to Table 3 in the main text, and Table A19 in the appendix.

Using the mom FE sample in the CWS FE regressions also

In the main specifications, using mother fixed-effects rather than CWS fixed-effects often appreciably and significantly changes the effect estimates. In fact, this is the main way in which our results are sensitive. In this subsection, we present results subsetting throughout to the sample used in the mother fixed-effects specifications. Note that sample sizes are slightly larger in this subsection than in the main text because, in order to maintain exact balance across CWS fixed-effects and mother fixed-effects regressions, we do not iteratively remove singletons (Correia (2015)); this also means that standard errors in this subsection are slightly understated.

Table A21:

Effects of contamination on low birth weight and preterm birth, using the mother fixed-effects sample throughout.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: low birth weight
CWS FE
10th to 90th RRMCL	0.01291*** (0.00289)	0.01414*** (0.00140)	0.01180*** (0.00274)	0.01298*** (0.00131)	0.00997*** (0.00232)	0.01002*** (0.00146)
Observations	701,006	699,750	701,006	699,750	701,006	699,750
Adj R2	0.0056	0.0058	0.0481	0.0483	0.1092	0.1094
Mom & CWS FE
10th to 90th RRMCL	0.00957*** (0.00245)	0.01049*** (0.00120)	0.00932*** (0.00232)	0.01003*** (0.00115)	0.00819*** (0.00198)	0.00775*** (0.00133)
Observations	701,006	699,750	701,006	699,750	701,006	699,750
Adj R2	0.1505	0.1507	0.1689	0.1690	0.2102	0.2103
Panel B: preterm birth
CWS FE
10th to 90th RRMCL	0.01990*** (0.00412)	0.02139*** (0.00184)	0.01854*** (0.00384)	0.01998*** (0.00170)	0.01632*** (0.00307)	0.01630*** (0.00176)
Observations	701,006	699,750	701,006	699,750	701,006	699,750
Adj R2	0.0042	0.0046	0.0502	0.0506	0.1355	0.1357
Mom & CWS FE
10th to 90th RRMCL	0.01571*** (0.00363)	0.01715*** (0.00166)	0.01533*** (0.00345)	0.01658*** (0.00160)	0.01405*** (0.00276)	0.01377*** (0.00170)
Observations	701,006	699,750	701,006	699,750	701,006	699,750
Adj R2	0.1486	0.1489	0.1663	0.1665	0.2252	0.2252

Notes: standard errors in parentheses

^*p < 0.10

^**p < 0.05

^***p < 0.01.

Standard errors are twoway clustered on public water system and mother. Each cell is a separate regression. Observations (number of births) and adjusted R²s (for the full model, i.e. including the fixed-effects) are reported. In each regression, the sample is restricted to births in our analytic sample for whom we observe at least one water quality sample of the given contaminant group during gestation in the public water system of residence at the time of birth. Control variables used vary over super-columns; independent variable, i.e. the contaminant group studied, vary across columns; the outcomes varies over panels; the specification used (either public water system fixed-effects or mom fixed-effects) varies over sub-panels; temp & TRI ctrls = weather and toxics release inventory (air pollution) controls.

Table A22:

Effects of contamination on small for gestational age and term birth weight, using the mother fixed-effects sample throughout.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age
CWS FE
10th to 90th RRMCL	0.00294** (0.00116)	0.00225*** (0.00087)	0.00228** (0.00109)	0.00164** (0.00083)	0.00212* (0.00114)	0.00141* (0.00085)
Observations	699,702	698,443	699,702	698,443	699,702	698,443
Adj R2	0.0062	0.0062	0.0386	0.0387	0.0394	0.0394
Mom & CWS FE
10th to 90th RRMCL	0.00279*** (0.00100)	0.00180** (0.00084)	0.00276*** (0.00096)	0.00166** (0.00081)	0.00290*** (0.00099)	0.00174** (0.00083)
Observations	699,702	698,443	699,702	698,443	699,702	698,443
Adj R2	0.1391	0.1397	0.1546	0.1552	0.1548	0.1554
Panel B: term birth weight
CWS FE
10th to 90th RRMCL	−9.46*** (1.55)	−9.03*** (1.45)	−8.40*** (1.69)	−8.22*** (1.41)	−8.25*** (1.85)	−7.60*** (1.53)
Observations	644,014	642,882	644,014	642,882	644,014	642,882
Adj R2	0.0194	0.0194	0.0997	0.0998	0.1037	0.1039
Mom & CWS FE
10th to 90th RRMCL	−9.05*** (1.73)	−8.81*** (1.33)	−9.15*** (1.84)	−8.57*** (1.37)	−9.33*** (1.87)	−8.24*** (1.45)
Observations	644,014	642,882	644,014	642,882	644,014	642,882
Adj R2	0.3488	0.3488	0.3875	0.3876	0.3896	0.3896

Notes: see notes to Table A21.

Estimating on Smaller CWS Only

Concern over mismeasurement of the link between water contamination and exposure, that may be especially acute in large CWS (with complicated distribution systems), motivates a sensitivity analysis estimating our models for smaller CWS only. To conduct this analysis, we sorted births by CWS size (number of people served by the CWS) and then dropped the 50% of births in our sample in the largest systems (those in the bottom of the dataset). The largest system in this restricted sample serves 118,000 people (contrast with the largest system, Philadelphia Water Department, which serves 1,600,000 people), and thus this is a meaningful restriction. The results of estimating on this subsample are included below.

Table A23:

Effects of drinking water contamination on low birth weight and preterm birth, small systems only.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: low birth weight
CWS FEs
10th to 90th RRMCL	0.01280*** (0.00124)	0.01161*** (0.00108)	0.01185*** (0.00119)	0.01080*** (0.00104)	0.01259*** (0.00121)	0.00789*** (0.00094)
Observations	597,900	596,418	597,900	596,418	597,900	596,418
Adj R2	0.0040	0.0042	0.0376	0.0378	0.0911	0.0914
Mom & CWS FEs
10th to 90th RRMCL	0.00863*** (0.00129)	0.00801*** (0.00113)	0.00844*** (0.00124)	0.00783*** (0.00110)	0.00913*** (0.00128)	0.00555*** (0.00107)
Observations	282,179	281,162	282,179	281,162	282,179	281,162
Adj R2	0.1584	0.1585	0.1755	0.1756	0.2094	0.2097
Panel B: pre-term birth
CWS FEs
10th to 90th RRMCL	0.01907*** (0.00166)	0.01732*** (0.00143)	0.01782*** (0.00159)	0.01621*** (0.00136)	0.01890*** (0.00159)	0.01240*** (0.00124)
Observations	597,900	596,418	597,900	596,418	597,900	596,418
Adj R2	0.0035	0.0037	0.0400	0.0403	0.1167	0.1171
Mom & CWS FEs
10th to 90th RRMCL	0.01543*** (0.00166)	0.01454*** (0.00141)	0.01519*** (0.00159)	0.01430*** (0.00135)	0.01666*** (0.00157)	0.01138*** (0.00126)
Observations	282,179	281,162	282,179	281,162	282,179	281,162
Adj R2	0.1659	0.1661	0.1842	0.1844	0.2328	0.2330

Notes: see notes to Table 3 in the main text.

Table A24:

Effects of drinking water contamination on small for gestational age and term birth weight, small systems only.

	(1)	(2)	(3)	(4)	(5)	(6)
	No controls		Adding mother ctrls		Adding temp & TRI ctrls
	All	Chemical	All	Chemical	All	Chemical
Panel A: small for gestational age
CWS FEs
10th to 90th RRMCL	0.00133* (0.00069)	0.00139** (0.00057)	0.00088 (0.00066)	0.00105* (0.00055)	0.00122 (0.00064)	0.00069 (0.00055)
Observations	597,188	595,707	597,188	595,707	597,188	595,707
Adj R2	0.0043	0.0043	0.0357	0.0357	0.0366	0.0366
Mom & CWS FEs
10th to 90th RRMCL	0.00026 (0.00128)	0.00052 (0.00107)	0.00016 (0.00126)	0.00046 (0.00105)	0.00054 (0.00123)	0.00029 (0.00104)
Observations	281,525	280,500	281,525	280,500	281,525	280,500
Adj R2	0.1568	0.1576	0.1697	0.1705	0.1700	0.1708
Panel B: term birth weight
CWS FEs
10th to 90th RRMCL	−8.18*** (1.23)	−6.83*** (1.07)	−7.66*** (1.19)	−6.49*** (1.02)	−7.57*** (1.19)	−5.74*** (1.06)
Observations	552,670	551,295	552,670	551,295	552,670	551,295
Adj R2	0.0148	0.0148	0.0894	0.0894	0.0935	0.0936
Mom & CWS FEs
10th to 90th RRMCL	−7.90*** (1.98)	−6.53*** (1.72)	−7.82*** (1.89)	−6.47*** (1.64)	−7.54*** (1.91)	−6.08*** (1.65)
Observations	249,414	248,510	249,414	248,510	249,414	248,510
Adj R2	0.3879	0.3877	0.4231	0.4230	0.4248	0.4247

Notes: see notes to Table 3 in the main text.

EFFECTS OF MCL VIOLATIONS ON BIRTH OUTCOMES

In this section, we closely follow Currie et al. (2013) using our data. We believe this comparison is important since our data represents a more recent time period and a different state. Overall, we find similar results, although ours are less precise. In general, the estimated effect of a reported health-based MCL violation on birth outcomes is less robust than the effects of our continuous measures of water contamination that we use in the main text. On the one hand, this may be surprising since violations indicate contamination in exceedence of regulatory thresholds; on the other hand, the very fact that there was a violation raises the possibility of significant avoidance behavior, such as through bottled water purchases (Zivin et al. (2011)).

As in the main text, we obtained Safe Drinking Water Information System (SDWIS) data on CWS drinking water violations from the US Environmental Protection Agency (US EPA). This SDWIS data records the compliance period, contaminant, and CWS for each drinking water violation received by each CWS in the United States over our sample period. We focus on “MCL violations”, which are cases where the CWS is observed to have exceeded a “maximum contaminant level,” a threshold over which contamination cannot cross without regulatory repercussions. Note that a drinking water compliance period is a period of time that varies based on contaminant and CWS, but is typically measured in quarters (3 months).

We begin by presenting summary statistics in Appendix Table A25, similar to those in Currie et al. (2013). Note that, following Currie et al. (2013), in Table A25 we restrict to mothers with multiple births, and we do not remove plural births. There are more mothers in this sample than in the mom fixed-effects sample in the main text since in the main text we removed plural births, all births exposed to an MCL violation, and births that had no samples during their gestation period (note that even births with no water quality samples during gestation can be flagged as being exposed or not exposed to an MCL violation).

Studying these statistics, in particular, comparing column (1) with column (2) in Appendix Table A25, we find that only about 3% of births in our sample are exposed to any MCL violations; contrast this with Currie et al. (2013)’s finding of about 8% of births exposed in New Jersey from 1997 to 2007. As a result, while our sample of births is about 50% larger than theirs, as would be expected given that we have an additional year and Pennsylvania is a larger state, the number of births exposed to an MCL violation is actually over 40% smaller in our sample. For chemical contamination, these differences are even larger; we have about 75% fewer births exposed to chemical contamination than Currie et al. (2013). We should keep this in mind when interpreting our results, especially for chemical MCL, as we have much less variation in the independent variables as Currie et al. (2013).

Moreover, we find that births that are exposed to an MCL violation are more likely to be from mothers that are advantaged on multiple measures. For example, mothers tend to be older for births exposed to MCL violations, and mothers exposed to MCL violations are less likely to be racial minorities, are more likely to have bachelors degrees or higher education, and are much more likely to be married. These are all opposite of the findings of Currie et al. (2013), who find that mothers exposed to MCL violations tend to be much more disadvantaged on average. For example, focusing on teen birth, they found that, overall, 3.3% of births were to mothers below the age of 19, while among births exposed to any MCL violation, this rate was about 4.6%; by contrast, in our sample these statistics are practically the reverse, at 5% and 2.6%, respectively.

We use the exact model from their paper (Currie et al. (2013)), which is the same regression model as our mom fixed-effects specifications in the main text, except for three differences. First, we use the same controls as used in Currie et al. (2013). Second, the independent variable is an indicator for at least one day of overlap between the gestation period and an MCL violation.⁸ Third, we include plural births (Currie et al. (2013) makes no mention of excluding plural births). For comparison, we also present additional results (a) adding additional controls (for plurality, WIC, and payment type) like in our water sample models in the main text, and (b) removing plural births.

Following Currie et al. (2013), we run the model for all MCL violations and then separately for just chemical violations. Chemical violations include all MCL violations except Total Coliform Rule violations. Note that Total Coliform violations are by far the most common MCL violation in our data (at over 70% of MCL violations).

To account for the fact that in the MCL-violation-based specifications, births with longer gestation periods are more likely to have an intersection with an MCL violation (and these birth tend to have better outcomes on average), in our specifications that have MCL exposure as the independent variable, we instrument for contamination using the same “full term gestation” instrument that was used in Currie et al. (2013). To construct this full term gestation instrument, we fix the date of conception and pretend that birth lasted exactly 39 weeks.⁹ We then construct “full gestation contamination” using this pretend gestation period, and use this constructed variable as an instrument for actual MCL violation exposure. Note that for our water sample regressions in the main text, this instrumental variables approach was unnecessary because there is no mechanical correlation between average water quality and gestational length.

Appendix Tables A26, A27, and A28 report the estimated impacts of MCL violations on low birth weight and prematurity. First, consider Table A26, which reports our attempt at estimating the exact model of Currie et al. (2013) on our sample. We find in this case that the estimated effects are similar to those in Currie et al. (2013), although they are less precisely estimated. In the preferred IV and FE specification (column 4), we find that exposure to an MCL violation for any contaminant increases the probability of low birth weight by about 0.38pp; the analogous result in Currie et al. (2013) was 0.34pp. It is interesting to note that our OLS estimates in column (1) are negative and over four times larger in magnitude than the comparable estimates in Currie et al. (2013), suggesting that the mechanical correlation in our data is more severe than in Currie et al. (2013). While our estimated effects of exposure to a chemical MCL violation on low birth weight has a point estimate of around 0.25pp, compared to Currie et al. (2013) estimate of 0.37pp, it is statistically insignificant. This may have been expected, given the very few births that can be connected to chemical contamination in this data (see Appendix Table A25). Appendix Tables A26 also contains our estimated impacts of MCL violations on pre-term birth, using the same model as Currie et al. (2013). Effects on pre-term birth of “any” MCL violation is about a 0.56pp increase, about twice as large as the estimate in Currie et al. (2013), which was about a 0.25pp increase.

In Appendix Table A27, we add several additional control variables to the Currie et al. (2013) specification, including WIC and Medicaid indicators, Toxics Release Inventory facility measures (as used in the main text), and plurality indicators. From our explorations (not shown), we found that the inclusion of plurality indicators changes our inference the most. While the point estimates for chemical MCL violations are stable (although the standard errors increase), the estimates for “any” violations are cut in half with the inclusion of these controls, and are insignificant in the new specification.

Finally, Appendix Table A28 report results from specifications that include the additional controls and remove plural births from the sample. This is the sample and specification that we use in all tables in this paper, except for tables A26 and A27 which we have just discussed. With this sample selection, the estimated effect of “any” MCL violations on pre-term birth rises back to roughly where it was earlier, and while the effects of “any” MCL violations on low birth weight rise, they remain statistically insignificant.

Table A25:

Sample means for all mothers and those exposed to MCL violations during gestation.

	(1)	(2)	(3)	(4)	(5)
		With MCL		Switchers
Characteristic	All	Any	Chem	Any	Chem
No. of observations	826218	26144	8243	56284	17138
Low birth weight (≤ 2500g)	0.096	0.082	0.080	0.074	0.071
Preterm (≤ 36 weeks)	0.112	0.096	0.097	0.087	0.086
Mom’s age
< 19	0.049	0.024	0.033	0.038	0.045
19 – 24	0.259	0.193	0.246	0.233	0.275
25 – 34	0.554	0.596	0.594	0.582	0.567
35+	0.139	0.186	0.127	0.148	0.113
Mom’s race
African-American	0.212	0.127	0.092	0.155	0.128
Hispanic	0.070	0.034	0.034	0.048	0.042
White, not hispanic	0.723	0.821	0.878	0.792	0.839
Mom’s education
Less than highschool	0.140	0.081	0.102	0.109	0.130
Highschool only	0.255	0.210	0.270	0.227	0.265
Some college	0.178	0.177	0.191	0.176	0.186
College or more	0.420	0.527	0.432	0.482	0.415
Mother smokes	0.215	0.191	0.249	0.203	0.246
Mother is married	0.589	0.700	0.671	0.655	0.633
Mother on WIC/Medicaid	0.471	0.350	0.418	0.401	0.451
Mother moved	0.467	0.570	0.495	0.633	0.584

Notes: Observations are births. Only mothers with multiple births are included. Switcher means births among the set of mothers with at least one birth experiencing and at least one not experiencing an MCL violation. “With MCL” contains births exposed to an MCL violation (i.e., at least one day overlap between the gestation period and an MCL violation in the water system at the time of birth).

Table A26:

Effects of MCL violations during gestation on birth outcomes – Replication of Currie et al. (2013).

	(1)	(2)	(3)	(4)
	OLS	IV	FE	IV and FE
Panel A: Any MCL, Low Birth Weight
Any MCL	−0.0025 (−1.3436)	0.0051*** (2.6142)	−0.0020 (−0.9393)	0.0035 (1.6122)
Panel B: Any MCL, Pre-term Birth
Any MCL	−0.0044** (−2.1996)	0.0058*** (2.6891)	−0.0029 (−1.2721)	0.0047** (1.9725)
Panel C: Chem MCL, Low birth weight
Chem MCL	−0.0061* (−1.9233)	−0.0018 (−0.5492)	−0.0009 (−0.2562)	0.0020 (0.5456)
Panel D: Chem MCL, Pre-term Birth
Chem MCL	−0.0081** (−2.3066)	−0.0014 (−0.3957)	−0.0007 (−0.1734)	0.0042 (1.0172)
No. observations	826218	826218	826218	826218
Full-gestation IV		✓		✓
Mom FE			✓	✓

Notes: t-statistics in parentheses

^*p < 0.10

^**p < 0.05

^***p < 0.01.

Each cell is a separate regression. Sample consists of mothers with multiple births. The independent variables “Any MCL” and “Chem MCL” are indicators for whether an MCL of the specified type occurred during the gestation period. Standard errors are clustered at the mother. “Full-gestation IV” instruments for actual exposure to an MCL violation with predicted exposure given a gestation period of exactly 39 weeks. All regressions include the control variables used in Currie et al. (2013).

Table A27:

Effects of MCL violations during gestation on birth outcomes – Adding additional controls (incl. plurality) to the Currie et al. (2013) model.

	(1)	(2)	(3)	(4)
	OLS	IV	FE	IV and FE
Panel A: Any MCL, Low Birth Weight
Any MCL	−0.0005 (−0.3015)	0.0054*** (3.1734)	−0.0014 (−0.6822)	0.0033 (1.6257)
Panel B: Any MCL, Pre-term Birth
Any MCL	−0.0022 (−1.2310)	0.0063*** (3.2887)	−0.0020 (−0.9140)	0.0048** (2.1166)
Panel C: Chem MCL, Low birth weight
Chem MCL	−0.0023 (−0.8163)	0.0012 (0.4195)	0.0002 (0.0586)	0.0030 (0.8464)
Panel D: Chem MCL, Pre-term Birth
Chem MCL	−0.0042 (−1.3469)	0.0016 (0.4866)	0.0005 (0.1364)	0.0053 (1.3343)
No. observations	826218	826218	826218	826218
Full-gestation IV		✓		✓
Mom FE			✓	✓

Notes: See notes to Table A26.

Table A28:

Effects of MCL violations during gestation on birth outcomes – Removing all plural births.

	(1)	(2)	(3)	(4)
	OLS	IV	FE	IV and FE
Panel A: Any MCL, Low Birth Weight
Any MCL	0.0000 (0.0220)	0.0048*** (3.2193)	−0.0006 (−0.3058)	0.0039** (1.9761)
Panel B: Any MCL, Pre-term Birth
Any MCL	−0.0009 (−0.5614)	0.0063*** (3.7555)	0.0000 (0.0164)	0.0063*** (2.8921)
Panel C: Chem MCL, Low birth weight
Chem MCL	−0.0022 (−0.8821)	0.0009 (0.3669)	0.0003 (0.0835)	0.0033 (0.9750)
Panel D: Chem MCL, Pre-term Birth
Chem MCL	−0.0040 (−1.4103)	0.0015 (0.5061)	0.0007 (0.2001)	0.0055 (1.4425)
No. observations	768228	768228	768228	768228
Full-gestation IV		✓		✓
Mom FE			✓	✓

Notes: See notes to table A26.