Same environment, stratified impacts? Air pollution, extreme temperatures, and birth weight in south China^††thanks: This paper has been accepted for publication at Social Science Research: https://doi.org/10.1016/j.ssresearch.2021.102691.

Studies have traced associations between prenatal exposure to various kinds of air pollutants and adverse birth weight and other birth outcomes [72, 74, 50, 60]. [24] found that the introduction of electronic toll collection greatly reduced both traffic congestion and vehicle emissions near highway toll plazas and reduced prematurity and low birth weight among mothers within two kilometers of toll plazas by 10.8% and 11.8%, respectively, relative to mothers two to ten kilometers from toll plazas. [39] analyzed birth records between 2002 and 2006 from the Georgia Department of PUblic Health and found that all traffic-related pollutants, including NO₂ and PM_2.5, were associated with preterm birth. [27] analyzed the impact of a policy that mandated the reduction of power-plant emissions in the eastern United States. The policy greatly reduced emissions of SO₂ and NO_x, which are the major precursors for PM_2.5. Their results suggest that the policy reduced the proportion of newborns downwind born prematurely and experiencing low birth weight, and reduced infant mortality for “high-risk” pregnancies associated with observed maternal health conditions such as diabetes, hypertension or eclampsia. Their conceptualization of risk using observed maternal health conditions is related to but distinct from our focus on unobserved innate vulnerabilities of babies themselves. [58] exploited exogenous air quality improvement during the Guangzhou Asian Games period to estimate the effects of PM₁₀, NO₂, and SO₂ on birth outcomes. They find that effects vary by pregnancy trimesters, sex, and maternal age.

A large-scale study in Brisbane, Australia, indicated that pregnancy exposures to PM_2.5, SO₂, NO₂, and ozone (O₃) were associated with increased risks of low birth weight, as well as pre-term birth [14]. [18] find that incense burning is associated with lower birth weight in boys, and their quantile regression also suggests that the negative associations are larger among the lower quantiles of birth outcomes.

Focusing on particulate matter, a recent United States study analyzed birth certificates data for the period 1999 to 2007 and found significant birth weight effects associated with gestational exposures to one form of coarse particulate matter (PM_10-2.5) [30]. Similarly, a study of 7,772 mothers in the Netherlands between 2001 and 2005 found that PM₁₀ exposure during pregnancy was inversely associated with birth weight [43]. The World Health Organization (WHO) Global Survey on Maternal and Perinatal Health study examined whether outdoor PM_2.5 was associated with adverse birth outcomes among 22 countries from 2004 through 2008 [32]. Results showed that across all countries, PM_2.5 was not associated with preterm birth, but was associated with low birth weight. In China, the country with the greatest PM_2.5 range among those studied, preterm birth and low birth weight both were associated with the highest quartile of PM_2.5 [32]. PM₁₀ and PM_2.5 concentrations are highly correlated in China—especially in southern China [91]. Studies in Lanzhou, China and Wuxi, China also indicated that prenatal exposure to PM₁₀ increases the risk of preterm birth [88, 36].

Several recent reviews have synthesized findings. [60] conducted a systematic review of acute air quality change studies and found mixed results: there was some evidence that maternal exposure to acute changes in air quality of short-to medium-term duration increased the risk of fetal growth restriction and preterm birth, but the relationship for other adverse obstetric or neonatal outcomes was less clear. [72] conducted a broader systematic review of studies of air pollution exposure and birth outcomes. This review suggested that maternal exposure to SO₂ was associated with preterm birth, exposure to PM_2.5 was associated with low birth weight, preterm birth, and small-for-gestational-age birth, and exposure to PM₁₀ was associated with small-for-gestational-age birth. Evidence for NO₂ was inconclusive. Another systematic review and meta-analysis of 62 studies indicated that pooled estimates of effects generally suggested associations between carbon monoxide (CO), NO₂, and particulate matter and pregnancy outcomes, but also that there was a high degree of heterogeneity among studies [74].⁴⁴4Early pollution exposures may endure in their impact. [25] summarized studies from developed and a few developing countries on the short- and long- term effects of in-utero and early-life exposure to pollution. The review highlighted evidence that exposure to pollution significantly impacted a variety of birth, childhood, and later-life outcomes.

A recent paper systematically reviewed Chinese and English publications on links between air pollution exposure and adverse pregnancy outcomes in China [47]. The authors reported that SO₂ was consistently associated with lower birth weight and more preterm births and PM₁₀ was consistently associated with congenital anomalies. Results for NO₂ were inconsistent, and the authors concluded that further studies were required on the effects of PM_2.5, ozone (O₃), and carbon monoxide (CO). While a number of the studies adjusted for education or other measures of socioeconomic status, the review does not report tests of effect modification by these stratifiers. A more recent study in Guangdong Province [59], which covered the period of 2014 to 2015, reported, among other significant findings, associations with low birth weight for PM_2.5, PM₁₀, SO₂, NO₂, and O₃ in the first month and with PM_2.5, PM₁₀, NO₂, and O₃ in the last month of pregnancy.

Extreme temperature events are increasing in frequency, duration, and magnitude [82]. For example, between 2000 and 2016, the number of people exposed to heat waves increased by around 125 million [82]. Temperature extremes have been linked to significant reductions in human health, including height and birth weight [29, 28, 64]. However, to date, evidence about extreme temperatures and birth outcomes remains limited [51, 87, 6], and that which is available is somewhat mixed. For example, a recent analysis of data from Korea found no significant temperature effect on birth outcomes [19]. A 2017 systematic review concluded that the evidence linking preterm birth and low birth weight to ambient temperature was still very limited and not conclusive, though there were more examples of adverse estimated effects for high temperatures than for low temperatures [87]. A 2020 systematic review covered papers about the associations of air pollution and heat exposure with preterm birth, low birth weight, and stillbirth in the United States [9]. Only three papers were identified that considered heat and birth weight, but all found significant risk. One, a paper that analyzed 220,572 singleton births for the years 2002 to 2008 from 12 United States sites, found that whole-pregnancy cold or hot temperature increased term low birth weight risk [35]. A second paper, conducted among 43,629 full-term but low birth weight babies and 2,032,601 normal-weight babies in California for the period 1999 to 2013, found that higher long-term apparent temperature exposure was associated with term low birth weight [6]. The third reviewed paper, conducted among births in Massachusetts from 2000 through December 2008 [51], showed decreased birth weight with increased air temperature.

In China, a study in the city of Guangzhou found that exposure to either low or high temperatures during pregnancy was associated with an increased risk of preterm birth [40]. A near-national scale study by [44] linked extreme heat exposure during pregnancy to significant reductions in multiple dimensions of adult welfare. The authors report that adults who experienced an additional day of high temperature during their prenatal periods, on average, attained 0.02 fewer years of schooling, had a higher risk of illiteracy by 0.18%, showed a 0.48% decrease in standardized word-test scores, and were shorter by 0.02 centimeters.

There is reason to believe that the effects of ambient air pollution and temperature exposures might vary by maternal education. First, because more-educated mothers are likely to enjoy better material resources or better knowledge compared to less-educated mothers, more-educated mothers may well adopt mitigation strategies and thus experience lower realized exposures in the same ambient air pollution and extreme temperature environments. Second, because better-educated mothers, as a group, are likely to have better nutrition and health care access compared to less-educated mothers, they may have more physical capacity to weather the same realized exposures to pollution and extreme temperatures.

There is at least suggestive evidence that the same ambient environment might be experienced differently for more and less economically vulnerable groups, such as groups defined by parental education.⁵⁵5Positive associations between maternal education and birth outcomes, including birth weight but also preterm birth and birth length, have been described in a number of studies [1, 5]. Studies in the United States have often found significant differences in birth weight for infants borne in more disadvantaged and segregated socio-economic communities with distressed physical environments [34, 54]. In a study in Georgia, [39] found that associations of traffic-related pollutants, including NO₂ and PM_2.5, with preterm birth tended to be stronger for mothers with low educational attainment. A study in Korea of multiple air pollutants showed mixed results, but some evidence of socioeconomic disparities in the effects of full-pregnancy exposure at the lowest conditional quantiles of birth weight [55]. [42] conducted a systematic review of papers assessing the effects of particulate matter on birth outcomes from 2000 to 2019 and found suggestive evidence that the risk posed was greater for babies of mothers with lower educational attainment for preterm birth and low birth weight.

However, in a recent review of epidemiological literature about ambient air pollution and birth weight, Westergaard and her colleagues [*]westergaard_ambient_2017 identified only a small number of studies that addressed effect modification by maternal education status. With regard to temperature effects, [87] call for more attention to the possibility of individual effect modifiers; maternal education is one important potential stratifier (see [6, 73] for inconsistent findings on maternal education as an effect modifier). Generally, the resources associated with maternal education might be more likely to be brought to bear under extreme conditions of pollution and temperature, which would exacerbate the maternal protective effect under these conditions. A similar logic suggests that the protective effects of maternal education with respect to pollution and extreme temperature might be most pronounced for infants with greater unobserved innate vulnerabilities.

A recent study from China also examines the heterogeneous association between air pollution and birth weight—but does not consider extreme temperatures—using quantile regression and finds that more pronounced effects were observed in lower and intermediate quantiles [83]. This paper observed stronger associations among well-educated mothers. Although the paper uses data from 16 counties across China during the period of 2014 to 2018, it does not control for county-specific time trends that may capture local policy variations and trends in the economic environment that can also affect birth outcomes. Additionally, [83] consider only full-term births. Except for [83], we have not found other studies of pollution and birth weight that investigate variability in the protective effect of maternal education using a conditional quantile approach.

A handful of studies in the United States have used conditional quantile approaches to consider heterogeneous vulnerability to ambient pollution exposure. A study of PM_2.5 in California [71] shows a tendency of somewhat larger negative associations with outcomes at the lowest conditional quantiles, though primarily among the non-Hispanic Black population. A study of PM_2.5 in Massachusetts [33] found that the negative association between PM_2.5 and birth weight was larger in magnitude at the lower conditional quantiles of birth weight than in the higher quantiles. In contrast, a study in Atlanta [76] showed larger effects for several pollutants at higher conditional quantiles. While results are mixed, all of these studies suggest the potential value in testing for heterogeneity in associations across the distribution of children’s unobserved innate vulnerabilities at birth, along the conditional distribution of birth weight.

Using a case study of three years of singleton live births in one district in Guangzhou, China, we study the protective effect of mothers’ education with respect to pollution and extreme temperature exposure. Specifically, we consider the association between expectant mothers’ college educational attainment and their children’s birth weight and allow the college educational gradient to be heterogeneous at different ambient pollution and extreme temperature levels. By including pollution and extreme temperature in the same specifications, we are able to isolate the separate relationships between these correlated but distinct ambient environmental factors and birth outcomes. Conditional on environmental exposures, we also allow the educational gradient to be heterogeneous for children with different unobserved innate vulnerabilities.

Our analysis relaxes the restrictions imposed by common regression frameworks on how socioeconomic factors might impact child-birth outcomes in a setting where mothers are exposed to high levels of negative environmental factors. A common regression framework includes educational status as an additive variable (for example, see [53] (birth weight) and [37] (BMI) in the context of quantile regressions; see [10] in the context of mean regressions). In the context of (conditional) mean regressions, this framework means that the educational gradient in birth weight is constant across levels of environmental exposure levels. By interacting mothers’ education with multiple potential negative environmental factors, we allow for the educational gradient in child birth outcomes to be magnified at different rates depending on how negative environmental factors jointly worsen. Furthermore, by estimating education-environment interactions separately at different conditional quantiles of birth weight, we allow the educational gradient to have different slopes by quantiles of unobserved innate vulnerabilities of babies. The rich heterogeneous relationships allow for a nuanced look at the potential protective effects of maternal college education in ameliorating negative environmental effects.

The study site for this project is Guangzhou, which is the capital city of Guangdong Province in south China, in the Pearl River Delta. Guangzhou, along with several other cities in Guangdong as well as Hong Kong and Macau, is part of the rapidly-developing “Greater Bay Area” megalopolis. The region is wealthy, with Shenzhen and Guangzhou, respectively, ranked third and fourth highest in city GDP in China, after Beijing and Shanghai [13]. Guangzhou has a typical subtropical climate, with very mild winters and hot, humid summers. The annual mean temperature in Guangzhou is around 22 degrees Celsius [20]. In recent decades, Guangzhou has experienced increases in mean annual temperatures and in the frequency, duration, and intensity of heat waves [86]. Extreme heat events are a particularly significant threat to human health due to a combination of the subtropical climate and urban heat-island effects exacerbated by rapid population growth in recent years [86]. Recent studies in Guangzhou have linked maternal heat exposures to preterm births [40, 79]. At the same time, and perhaps somewhat counter-intuitively in a subtropical climate, cold spells in Guangzhou carry health risks. For example, cold spells have been linked to increased mortality risk for most categories of deaths and to heightened risk of preterm birth [15, 40], possibly due to the lack of widespread access to central heating. As a major city in south China, Guangzhou is a data point that is potentially relevant to other urban settings in south China, and possibly parts of Southeast Asia, characterized by broadly similar climates, high levels of pollution, and economic conditions.

South China has had less air pollution than north China, but as an economic development center, Guangzhou has been one of the most polluted cities in the region. Recent studies in Guangzhou have linked air pollution to respiratory distress syndrome and to student respiratory illness and absenteeism, especially for younger students [16, 57]. Importantly, Guangzhou has a well-established and extensive air-quality monitoring system with both air pollution indices and individual pollutant concentrations available at the city level.

Figure 1 shows levels of air pollution in Guangzhou from 2008 to 2011 for three monitored pollutants: PM₁₀, SO₂, and NO₂. The levels of pollution depicted in Figure 1 reflect substantial improvements in Guangzhou’s air quality compared to prior years. In the early 2000s, the local government implemented progressive air-quality control efforts that included closing down low-efficiency coal-power plants and enforced installation of desulfurization facilities [90]. Even with improvements, annual mean air pollution in Guangzhou during this period exceeded the WHO standard for PM₁₀: a 20 $\mu g/m^{3}$ annual mean. Further, 66.4% of days surpassed the 24-hour mean standard for PM₁₀ of 50 $\mu g/m^{3}$ and 76% of days surpassed the SO₂ standard of 20 $\mu g/m^{3}$ [81]. There is strong seasonality in all three types of air pollution, with more severe air pollution in the winter than in the summer.

We link three forms of data for this analysis: birth certificate data, air pollution data, and meteorological data.

We begin with birth-certificate data for all live births (67,108 or 99.1% of all births) in the district for the period 01 January 2009 to 31 December 2011. To avoid fixed-cohort bias,⁹⁹9Fixed-cohort bias emerges when a sample consists of births during a fixed period—this approach will include only the longer pregnancies at the start of the study and only the shorter pregnancies at the end of the study. This has the potential to bias studies of environmental exposures [75]. we delete births within this period with conception dates earlier than 14 July 2008 and those with conception dates later than 15 February 2011, as gestational age varies between 171 and 319 days, which leaves 58,827 observations. We exclude observations with birth weight less than 500 grams (17 observations dropped), or birth length less than 28 centimeters or longer than 60 centimeters (51 observations dropped). We further restrict our sample to include only those with observed gestational age (855 observations without gestational age dropped) and only live singleton births.¹⁰¹⁰10Because there is no information on the number of multiple births, we count duplicate observations in terms of all observed maternal and paternal characteristics (including mother’s birthdate, parity, number of previous pregnancies, maternal occupation, maternal education, paternal education, gestational age, and an indicator of high risk levels) as well as birthdate of the baby, and treat the observations with one duplicated case as twins (3.90% of 57,108 observations) and those with two duplicated cases as triplets (0.08% of 57,108 observations) and those with three duplicated cases as quadruplets (0.01% of 57,108 observations). We further removed those registered with rural residence (949 observations, or 1.73% of the sample). In the end, our analytic sample consists of 53,879 live singleton births (15,068 born in 2009, 20,368 born in 2010, and 18,443 born in 2011).¹¹¹¹11In Appendix Section A.3, we estimate the model after restricting the analytic sample to only those with more than 36 weeks of gestational age.

Table 1 contains summary statistics for all variables employed in the analysis. Below, we describe our variables.

We model birth weight as related to maternal education, air pollution and extreme temperature exposure in both mean and quantile regression frameworks. We first estimate a baseline main effects (conditional) mean regression model:

where $\text{BirthWeight}_{\text{ymi}}$ is birth weight of individual $i$ who is conceived in year $y$ and month $m$. $\text{Edu}_{i}$ is a binary variable indicating if a mother has at least a college degree. $P_{i}$ is a continuous variable measuring each mother’s prenatal ambient exposure to PM₁₀, or the composite index of PM₁₀, NO₂ and SO₂. $\text{Cold}_{i}$, and $\text{Heat}_{i}$ are continuous variables measuring the percentages of prenatal days exposed to extreme cold or heat. $g\left(\text{Rainfall}_{i}\right)$ is a cubic function of daily mean precipitation that a mother is exposed to during her pregnancy. We control for an individual-specific vector of attributes $\boldsymbol{X}_{i}$, which includes linear and quadratic terms for maternal age, child’s sex, parity, number of multiple births, and birth day of the week. To control for shifts in seasonal patterns of birth as well as possible changes in economic conditions over time, we include conception year and month interaction fixed effects $\zeta_{ym}$.²⁰²⁰20Due to significant day-to-day variations in pollution exposures as shown in Figure 1 and Figure 3, despite the inclusion of year and month interaction fixed effects, there remains sufficient variations in pollution exposures that allow for identification. These variations, in effect, arise due to day-by-day differences in exposures in the first and last months of pregnancy. In our regression framework, we assume homogeneous effects of pollution exposures during the course of pregnancy. Given these, Equation (1) compares the birth weights of infants conceived in the same calendar month and examines the associations of cumulative within-month variation in air pollution and temperature with variation of birth weights.

Equation (1) restricts the birth weight gradient with college education to be constant, which is captured by $\sigma_{0}$ and is invariant of ambient environmental exposures. Equation (2) relaxes this restriction, and allows the birth-weight gradient with college education to potentially vary depending on $P_{i}$, $\text{Cold}_{i}$, and $\text{Heat}_{i}$:

where positive values for $\sigma_{1},\ \sigma_{2}\text{\ \ }$and $\sigma_{3}$ combined with negative values for $\gamma_{1},\ \gamma_{2}\text{\ \ }$and $\gamma_{3}$ would imply protective effects of maternal college education that ameliorate the negative impacts of ambient air pollution, extreme cold, and extreme heat on birth weight.

We estimate Equations (1) and (2) both via conditional mean (OLS) and conditional quantile estimation.²¹²¹21Under OLS, the same estimated coefficients provide both conditional and unconditional mean predictions; under quantile estimations, conditional quantile estimates in general can differ from unconditional quantile results [31]. In this paper, we use conditional quantile estimates to study heterogeneous maternal education and birth weight gradients on levels (grams) of birth weight. While [31] show that in the context of birth weight, conditional and unconditional quantile estimates can be similar, we do not consider the implications of our estimates for the unconditional (marginal) distribution of birth weight. We estimate quantile regressions with the quantreg package [52]. Under OLS, Equation (2) imposes the restriction that the birth weight-college education gradient—conditional on observed variables including the vector of environmental exposure measures $P_{i}$, $\text{Cold}_{i}$, and $\text{Heat}_{i}$—is constant across individuals. The implicit assumption is that the birth weight-college education gradient is invariant across the conditional distribution of birth weight, where the conditional distribution captures unobserved factors—including nutrition, avoidance behavior, genetic factors, or other risk factors—that can also impact birth weight. However, given the same ambient environment, the tendency for college-educated mothers to have babies with higher birth weight compared to non-college-educated mothers may be most pronounced at the lower conditional quantiles of unobserved innate vulnerabilities. Additionally, babies with greater unobserved innate vulnerabilities might be more likely to suffer from air pollution and extreme temperatures.

Our conditional quantile analysis allows birth weight-college education gradients to vary both along observed dimensions of sources of vulnerabilities, i.e., potential negative environmental factors (as seen in the interaction effects in Equation (2)), and along unobserved dimensions of sources of baby vulnerabilities, as captured by different conditional quantiles. The results provide a finer disentangling of heterogeneities of the birth weight-maternal education gradients. For example, the conditional mean estimates could understate the importance of PM₁₀ if the negative relationship between PM₁₀ and birth weight is magnified at lower conditional quantiles where mothers/children have greater unobserved innate vulnerabilities.

Results from mean regressions with PM₁₀ as the only air pollutant are presented in Table 3. Columns 1 and 3 show the Equation (1) estimates. Columns 2 and 4 show the Equation (2) estimates, which include interaction terms between maternal education and environmental exposure variables. In columns 1 and 2, extreme temperature is defined by 1% extreme tails of past temperatures, while columns 3 and 4 use 2.5% as the tail cutoffs.

After controlling for conception year-by-month fixed effects, PM₁₀, extreme cold, and extreme heat are all negatively and substantially associated with birth weight. Column 1 in Table 3 shows that a 1 $\mu g/m^{3}$ increase in average daily PM₁₀ potential exposure during pregnancy is associated with an 17.8 gram reduction of birth weight (s.e. 2.3). An additional percentage point increase in potential exposure to extreme heat or cold is associated with a 22.4 gram (s.e. 9.4) or 30.8 gram (s.e. 9.8) birth weight reduction, respectively. Stated differently, a one standard deviation change in potential exposure to PM₁₀ (i.e., 6.6 $\mu g/m^{3}$), extreme heat (i.e., 1.61% of pregnant days), or extreme cold (i.e., 1.00% of pregnant days) corresponds to 0.25, 0.08, and 0.07 standard deviation changes in birth weights, respectively.²²²²22While the magnitudes of the associations between PM₁₀ or other pollutants and birth weights in relatively low pollution settings are generally smaller than what we find, our results are similar to the findings obtained by [10] from Chile—a setting with similar mean PM₁₀ exposures as our setting here—where results show that a one standard deviation (17 $\mu g/m^{3}$) increase in PM₁₀ exposures is associated with 0.23 standard deviations (125 grams) of birth-weight reduction. The results from Columns 1 and 3 are similar. The exception is that in Column 3, with the less-extreme temperature thresholds, heat exposure is no longer significantly negatively associated with birth weight.

Columns 2 and 4 show the maternal college education interaction coefficient estimates.²³²³23The coefficient estimates for college education appear negative in the interaction specifications (columns 2 and 4). In these specifications, college education coefficients represent intercepts (adjusting for other variables in the model) for the hypothetical situation of no PM₁₀, cold, or heat exposure. The predicted protective effects of maternal education given the observed range of ambient environmental measures are presented in our discussion of college premiums for birth weights. In both columns, mothers’ college education status is associated with reduced vulnerability to ambient air pollution and extreme heat exposures. However, this protective effect of mothers’ college is not significant for extreme cold. Focusing on Column 2, there is a positive college-education interaction coefficient of 2.0 (s.e. 0.7) for PM₁₀ and 11.8 (s.e. 2.9) for extreme heat. Comparing college-educated to non-college-educated mothers, the interactions correspond to a 10% dampening of the negative association of PM₁₀ with birth weight and a 44% dampening of the negative association of extreme heat with birth weight.

We evaluate the college premium in birth weight, which we define to be the predicted gap in birth weight between college-educated and non-college-educated mothers, by considering the negative intercept for college education, -115.7 (s.e. 53.2), and the positive education slopes jointly. Holding extreme heat and cold exposures at their respective means, the college premium amounts to 0.05, 0.08, and 0.16 standard deviations of birth weight at the 1^st (63 $\mu g/m^{3}$), 50^th (70 $\mu g/m^{3}$), and 99^th (89 $\mu g/m^{3}$) percentiles of the marginal distribution of PM₁₀. Similarly, holding PM₁₀ and extreme cold at their respective means, the college premium amounts to 0.05, 0.07 and 0.17 standard deviations of birth weight at the 1^st (0%), 50^th (0.76%) and 99^th (4.71%) percentiles of the marginal distribution of the percentage of pregnancy days exposed to extreme heat.

Conditional quantile results, without and with maternal education interactions with the three environmental measures, are shown in Tables 4 and 5, respectively. These estimates use the more stringent cold and heat exposure measures defined by the 1% cutoffs. Because these are conditional quantile estimates, the quantiles are defined with respect to what we call the children’s unobserved innate vulnerabilities (i.e., the residuals in the birth weight regressions, not the birth weights themselves).

We repeat the analyses parallel to that for PM₁₀ but with the composite index pollution measure. Detailed results are presented in Appendix Tables A.1, A.2, and A.3. Patterns in the temperature results are robust to the choice of pollution measure, and are not discussed here. The results from the composite index analysis closely mirror the patterns revealed in the PM₁₀ analysis. There is a protective effect of maternal education with regard to pollution, and the finding persists across conditional quantiles of birth weight. Further, quantile regressions show that the negative associations between pollution and birth weight, and the protective effects of maternal education with respect to pollution, are magnified at lower conditional quantiles.

A key difference in the composite index analysis results is that for estimates based on Equation (1) without maternal education interactions, the birth weight change associated with one standard deviation change in the composite index is 1.8 times of the corresponding change in the PM₁₀ analysis. In the regressions with interactions, the protective effect of college education is weaker in the composite index analysis than in the PM₁₀ analysis. Together, these factors lead to a pattern in which the college premium in grams of birth weight along the marginal distribution of the composite index is approximately the same as the college premium in birth weight along the marginal distribution of PM₁₀.

Conditional on ambient temperature exposures, we find that one standard deviation increase in ambient PM₁₀ is associated with a 0.25 standard deviation reduction in birth weight. At average ambient PM₁₀ and extreme temperature exposure levels, the protective effect of maternal college education is up to 0.07 standard deviations of birth weight. We also find that the negative association between ambient exposures and birth weight is two times as large at lower conditional quantiles of birth weight than at the median. At lower conditional quantiles of birth weight, a one standard deviation increase in ambient PM₁₀ exposure is associated with a reduction in birth weight of up to 0.48 standard deviations.

Additionally, the protection associated with university-educated mothers with respect to pollution and extreme heat is five times as large at the lowest decile of conditional quantiles of birth weight than at the highest decile. In particular, the protective effect of college education can be up to 0.34 standard deviations of birth weight when children with higher unobserved innate vulnerabilities—those at lower conditional quantiles—face high levels of ambient PM₁₀ and extreme temperature exposures. Though we focus on a continuous measure of birth weight for all terms in the text, we find similar results with binary measures used in some other studies for low birth weight, preterm births, and small for gestational age (Appendix Section A.4). Our results show that there are substantively significant heterogeneous moderating effects of maternal education on the negative associations between ambient environmental exposures and birth weight. We are the first to document these relationships using mean and conditional quantile regressions with an approach in which we consider three key ambient environmental exposures jointly.