Fine Particulate Air Pollution and Birthweight: Differences in Associations along the Birthweight Distribution

Abstract

Background:

Maternal exposure to fine particulate air pollution (PM_2.5) during pregnancy is associated with lower newborn birthweight, which is a risk factor for chronic disease. Existing studies typically report the average association associated with PM_2.5 increase, which does not offer information about potentially varying associations at different points of the birthweight distribution.

Methods:

We retrieved all birth records in Massachusetts between 2001 and 2013 then restricted our analysis to full-term live singletons (n = 775,768). Using the birthdate, gestational age, and residential address reported at time of birth, we estimated the average maternal PM_2.5 exposure during pregnancy of each birth. PM_2.5 predictions came from a model that incorporates satellite, land-use, and meteorologic data. We applied quantile regression to quantify the association between PM_2.5 and birthweight at each decile of birthweight, adjusted for individual and neighborhood covariates. We considered effect modification by indicators of individual and neighborhood socioeconomic status (SES).

Results:

PM_2.5 was negatively associated with birthweight. An interquartile range increase in PM_2.5 was associated with a 16 g (95% CI: 13, 19) lower birthweight on average, 19 g (95% CI: 15, 23) lower birthweight at the lowest decile of birthweight, and 14 g (95% CI: 9, 19) g lower birthweight at the highest decile. In general, the magnitudes of negative associations were larger at lower deciles. We did not find evidence of effect modification by individual or neighborhood SES.

Conclusions:

In full-term live births, PM_2.5 and birthweight were negatively associated with more severe associations at lower quantiles of birthweight.

Background

Exposure to fine particulate air pollution, specifically particulate matter under 2.5 μm in aerodynamic diameter (PM_2.5), is associated with increased mortality and morbidity, and is a leading cause of the global disease burden.¹ During the prenatal period, the fetus is susceptible to maternal environmental exposures such as PM_2.5.² In fact, increased maternal exposure to PM_2.5 during pregnancy has previously been associated with decreased birthweights.^3–7 Lower birthweights appear detrimental to the newborns’ health as they are correlated with subsequent risk for chronic conditions such as including cardiovascular disease, diabetes, obesity, respiratory disease, and premature mortality.^8–13 Prior studies have also shown that associations between particulate air pollution and birthweight are modified by maternal socioeconomic status (SES).^14–16 However, it remains uncertain how SES modifies the associations and if certain population subgroups are more susceptible to PM_2.5 than others.

In prior studies, researchers have almost exclusively used generalized additive models and their simpler linear relatives to estimate the association between an increased exposure to PM_2.5 and mean birthweight.^7,17 While these models inform on the overall associations of PM_2.5 with birthweight in the study population, they neglect possible differences in associations across the birthweight distribution.¹⁸ In this study, we use quantile regression to estimate the associations of PM_2.5 with specific quantiles of birthweight. Compared to more commonly used regression techniques such as simple linear regression, quantile regression is less susceptible to outliers and it allows the parameters of the covariates to differ across the quantiles. While simple linear regression assumes the effect of an exposure to shift the entire distribution of birthweight without changing its shape, quantile regression is more flexible and allows different effects of an exposure on birthweight at different quantiles of birthweight. On the other hand, quantile regression is much more computationally intensive and requires more computing resources to run, especially with large datasets. Quantile regression has previously been applied in studies of birthweights first with maternal behaviors and more recently with prenatal exposures to heavy metals, finding different associations in lower quantiles of birthweight compared to higher quantiles.^19–21

In addition to applying a seldom-used statistical method in air pollution health research, this study uses highly resolved spatial and temporal PM_2.5 predictions to minimize exclusion of participants due to missing exposure data. In some prior studies that relied on data from air pollution monitoring, those living far away from a monitoring station were excluded from analysis.^3,5,22,23 To fill in the missing gaps in air pollution monitoring, land use regression using variables such as traffic density has been used to estimate PM_2.5 concentrations, which can lead to biased estimates of the PM_2.5 effect in areas of low traffic density.^24,25 The generalizability of such models to less densely populated areas is also limited by the dearth of monitoring in such areas for model calibration, and the temporal resolution is often limited, generally to periods longer than a pregnancy. More sophisticated hybrid models, which use satellite remote sensing data in addition to land use and meteorologic variables, have been developed to improve prediction accuracy and expand coverage.²⁶ Using satellite remote sensing data not only increases coverage of the study population, but also the temporal resolution. In this study, PM_2.5 exposure is based on a hybrid model that that predicts at a temporal resolution of 1 day and spatial resolution of 1 km * 1 km.

In summary, we investigate the association between PM_2.5 and birthweight at different points along the birthweight distribution using quantile regression and PM_2.5 exposure estimated at a high spatiotemporal resolution using geocoded maternal residential addresses.

Methods

Study Population

The study population included 978,225 births from the Massachusetts Birth Registry for the period between 1 January 2001 and 31 December 2013. At the time of birth, the residential address of the mother was recorded and later geocoded by the Massachusetts Department of Public Health against TomTom Multinet (American Digital Cartography, Appleton, WI) using AccuMail address and zip code as the input address field and zone. 23,947 births were missing address information and 10,345 births living further than 2 km from a PM_2.5 prediction cell centroid were excluded. In our analysis, we restricted the data set to singleton, full-term live births with a minimum birthweight of 500 grams and a clinical gestational age between 37 and 44 weeks, which excluded 127,076 births. Further restriction to those with complete data on the regression covariates excluded 41,093 births, with the majority missing data on parity, leading to a final sample size of 775,768. This use of birth data was approved by the Massachusetts Department of Public Health as well as the human subjects committee at the Harvard T. H. Chan School of Public Health.

PM_2.5 Exposure

For each birth, we used clinical gestational age and maternal geocoded address to determine the average maternal PM_2.5 exposure during pregnancy. Specifically, each birth was assigned to the closest 1 km * 1 km grid cell within 2 km of a modeled daily PM_2.5 exposure dataset.²⁶ Then, for each day during pregnancy as indicated by the clinical gestational age, the corresponding PM_2.5 value was matched to each birth using statistical program R version 3.4.1 (R Foundation, Vienna, Austria).²⁷ Finally, we took an average of PM_2.5 daily exposures for the entire pregnancy to generate the average PM_2.5 exposure for each birth. PM_2.5 concentrations were from a prediction model uses satellite-derived aerosol optical depth measurements along with land use and meteorologic variables to ascertain ground-level PM_2.5 concentrations. PM_2.5 measurements from two monitoring networks, US EPA Air Quality System and the Interagency Monitoring of Protected Visual Environments, were used to calibrate the model, which, in validation, performed consistently with an out-of-sample R² of 0.8 or higher. Further details on the prediction methodology are found in a previous publication.²⁶

Covariates

In additional to birthweight, the Massachusetts Birth Registry records provide individual variables that are potential confounders in the relationship between PM_2.5 and birthweight. They include maternal age (years), maternal race (white, black, Asian, American Indian, other), maternal marital status (married, not married), maternal smoking (yes, no), maternal education (less than high school, high school, some college, college, advanced degree beyond college), parity (first-born, not first-born), maternal diabetes (yes, no), gestational diabetes (yes, no), maternal chronic high blood pressure (yes, no), maternal high blood pressure during pregnancy (yes, no), Kessner index of adequacy of prenatal care (adequate, intermediate, inadequate, no prenatal care)²⁸, birth mode of delivery (vaginal, forceps, vacuum, first caesarian birth, repeat caesarian birth, vaginal birth after caesarian birth), year of birth (dummy variable for one of 2001 to 2013), clinical gestational age (weeks), newborn sex (male, female), and government support for prenatal care (yes, no). Birthweight was measured and recorded by medical staff at the time of birth; clinical gestational age was determined by a clinician who assessed the fetus’s physical and neurologic maturity during a prenatal care visit. Additionally, we performed analyses with adjustment for Census block group proportion black population and Census block group median household income. These data were based on the trailing 2006–2010 five-year estimate from the American Community Survey.²⁹ No covariates listed here are plausibly causal predictors of PM_2.5 exposure. Rather, variables such as maternal smoking may be surrogates for socioeconomic factors that predict both maternal smoking and neighborhood PM_2.5 exposure. Similar sets of covariates were used in recent studies investigating the PM_2.5 and birthweight.^15,16,30

Statistical Modeling

As a first step, we estimated the association between PM_2.5 and birthweight, adjusted for covariates, using linear regression. This led to an estimate of the difference in birthweight per increase in PM_2.5, on average. Secondly, we tested for effect modification by adding joint product terms between PM_2.5 and each of infant sex, maternal education, government support for prenatal care, median household income, and proportion of black population in Census block group to assess effect modification. We then estimated the association between PM_2.5 and birthweight at specific quantiles of birthweight using quantile regression.

Quantile regression estimates the conditional quantile of the outcome.³¹ In contrast to linear regression, which estimates the expected value of the outcome using the method of least squares, quantile regression relies on least absolute value regression and minimizes the sum of the absolute values of residuals, with weights for positive and negative residuals designed to target specific quantiles. The consequences of this difference in methods are that quantile regression does not assume normality in the distribution of errors or constant variance. Specifically, we assumed that

$$Q_{\tau }\left(B{W}_i\right)={\alpha }_{\tau }+{\beta }_{1,\tau }P{M}_i+{\beta }_{2,\tau }{X}_{2i}+{\beta }_{3,\tau }{X}_{3i}\dots +e_{i,\tau }$$ (1)

where Q_τ is the τ quantile of birthweight (BW), for infant i, given the PM_2.5 exposure (PM_i) and the set of covariates, X_ni. α_τ is the intercept and e_i,τ is the individual error term for the τ quantile. τ has a theoretical range of (0, 1). In this paper, we estimated the association at the 1^st through the 9^th deciles, which correspond to the 10^th, 20^th, 30^th … 90^th percentiles of birthweight. We additionally conducted sensitivity analyses where we did not restrict to full-term births but included all births with a recorded gestational age between 20 and 50 weeks. Furthermore, we considered possible confounding by seasonality and repeated our analyses with year-season included as a covariate. All analyses were performed in R (R Project for Statistical Computing, Vienna, Austria).²⁷

Results

The Table provides descriptive statistics on the study population (n = 775,768). Overall, mean birthweight was 3,441 g and the mean PM_2.5 was 10.1 μg/m³ with an interquartile range width (IQR_w) of 2.3 μg/m³. The IQR_w of PM_2.5 did not vary much across all deciles of birthweight. About two-thirds of the mothers reported as being married and about one-third received government support for prenatal care. The majority of births were delivered vaginally, born to white mothers, received adequate prenatal care, or born to mothers with more than a high school education. Being of higher gestational age, not the first-born, delivered with cesarean section, having adequate prenatal care, born to a mother who reported as married, more educated, was white, having gestational diabetes, other diabetes, or living in a Census block group with higher median household income was associated with higher birthweight while being female, born to a mother who was older, who smoked, received government support for prenatal care, had chronic high blood pressure, high blood pressure during pregnancy, or lived in a Census block group with higher black population proportion was associated with lower birthweight (eTable 1).

Table:

Characteristics of Full-Term Singleton Births in Massachusetts from 2001 to 2013^a

Variable	Overall
Total Births (n)	775,768
Birthweight (g) (mean ± sd)	3441 ± 471
Average PM2.5 over entire Pregnancy (μg/m3) (mean ± sd)	10.1 ± 1.4
Clinical Gestational Age in Weeks (mean ± sd)	39.3 ± 1.2
Mother’s Age in Years (mean ± sd)	30.1 ± 6.0
Census Block Group Median Household Income ($1000s per year ± sd)	67.7 ± 32.8
Census Block Group Black Population Proportion (± sd)	0.08 ± 0.15
Newborn Female Sex (%)	49.0
Mother reported as Married (%)	68.8
Government Support for Prenatal Care (%)	33.1
Smoking During or Prior to Pregnancy (%)	13.7
Gestational Diabetes (%)	4.1
Other Diabetes (%)	0.9
High Blood Pressure during Pregnancy (%)	3.3
Chronic High Blood Pressure (%)	1.2
Parity: First-Born (%)	45.2
Mode of Delivery (%)
vaginal	65.5
forceps	0.6
vacuum	3.6
first cesarean birth	16.7
repeat cesarean birth	12.1
vaginal birth after previous cesarean birth	1.6
Maternal Race (%)
White	71.8
Black	8.5
Asian	7.7
American Indian	0.2
Other	11.8
Kessner Index for Prenatal Care (%)
adequate	78.5
intermediate	17.0
inadequate	3.3
no prenatal care	1.2
Maternal Education (%)
Less than High School	10.8
High School	24.0
Some College	22.3
College	26.2
Advanced Degree	16.7

^a:Inclusion criteria were having complete maternal residence information, living within 2 km from a PM_2.5 prediction cell centroid, being a singleton, full-term (37 to 44 weeks of gestation), live birth weighing at least 500 g.

Higher exposure to PM_2.5 during pregnancy was associated with lower birthweight. On average, an IQR_w increase in PM_2.5 was associated with a 16 g (95% CI: 13, 19) lower birthweight. Other regression coefficients agree with prior observations regarding summary statistics (eTable 2). We assessed strength of confounding from included regression covariates and found that maternal age, parity, smoking, clinical gestational age, and year of birth were strong sources of confounding (eFigure 1). Omitting maternal race or parity from the model would have led to a more negative estimate of the association between PM_2.5 and birthweight while omitting smoking, clinical gestational age, or year of birth would have led to a less negative estimate. We did not find evidence for effect modification of the association by any of infant sex, maternal education, government support for prenatal care, Census block group median household income, and Census block group proportion of black population. Therefore, our final model included main effects of PM_2.5, individual covariates, and Census block group covariates.

Associations between PM_2.5 and birthweight were more negative at lower deciles of birthweight (Figure). The Figure suggests that as birthweight decile increased, the negative association between PM_2.5 exposure and birthweight moved closer to null. At the lowest decile of birthweight, an IQR_w increase of PM_2.5 was associated with a 19 g (95% CI: 15, 23) lower birthweight; at the highest decile, a 14 g (95% CI: 9, 19) lower birthweight. A test of linear trend for this relationship found that for each decile increase, the magnitude in the association between a PM_2.5 IQR_w increase and birthweight was reduced by 0.48 g (95% CI: 0.46, 0.50). Compared to the association between a PM_2.5 IQR_w increase and birthweight on average, the associations at the first through the fifth deciles were more negative while the changes at the sixth, seventh, and ninth deciles were more positive. The most negative association was at the lowest decile of birthweight and the least negative was at the highest decile.

Figure.

Associations between Interquartile Range Increase in PM_2.5 (2.3 μg/m³) and Birthweight in Full-Term Live Singleton Births at Deciles of the Birthweight Distribution. Dotted line shows the average birthweight change, −16.1 g, associated with interquartile range increase in PM_2.5. Model covariates include maternal age, race, marital status, smoking, education, parity, chronic diabetes, gestational diabetes, chronic high blood pressure, high blood pressure during pregnancy, Kessner index of adequacy of prenatal care, mode of delivery, year of birth, clinical gestational age, newborn sex, government support for prenatal care, Census block group median household income, and Census block group black population proportion.

Sensitivity analyses showed similar patterns of associations between PM_2.5 exposure and birthweight. When we did not restrict to full-term births, the difference in magnitude between the associations at the lowest versus the highest deciles was slightly larger (eFigure 2i). When we additionally adjusted for seasonality, the most negative association still occurred at the lowest decile and the least negative association at the highest decile of birthweight (eFigure 2ii and 2iii). The difference between the estimate at the lowest decile compared to the highest decile was smaller compared to the main analysis. Moreover, in contrast to our main analysis, where estimates became progressively less negative as birthweight decile increased, the estimates at the lowest five deciles were within a gram per IQR_w increase in PM_2.5 when seasonality was included. Similar to the main results, the estimates were less negative at the highest four deciles compared to the lowest five when seasonality was included.

Discussion

Our analysis of full-term live singleton births in Massachusetts between 2001 and 2013 found a strong negative association between PM_2.5 exposure during pregnancy and birthweight. Applying quantile regression, we found that negative associations between PM_2.5 and birthweight were more severe for lighter newborns. Each gram reduction in birthweight may be more consequential for newborns in the lowest five deciles of birthweight, making the larger estimated birthweight detriments from PM_2.5 exposure in the lightest newborns more concerning.

Negative associations between PM_2.5 and birthweight have previously been reported. The magnitude of the negative association from our analysis is relatively large compared to existing estimates.¹⁷ In contrast to some recent studies that found that the association between PM_2.5 and birthweight was modified by individual or community level SES,^14–16 we did not find strong evidence of effect modification in our analyses. These differences in findings compared to prior research could be due to differences in study populations. As our study population included only born in Massachusetts, its generalizability to other populations with different characteristics could be limited. The lack of strong evidence of effect modification by SES could be due to differences in which SES variables were included, how SES variables were measured, and that combined SES differences between births vary between studies.^14,16

Importantly, we found that the negative association between PM_2.5 and birthweight was larger in magnitude at the lower end of the birthweight distribution than those at the higher end. In other words, the negative association was estimated to be more severe for lighter newborns than for those who were heavier. Had the severity been equal, the estimated changes in birthweight at different deciles would have been closer together and an increasing linear trend between decile and change in birthweight would not have been observed (Figure). Adding to this, our results imply that as PM_2.5 exposure increased, inequality in birthweight detriments between lighter and heavier newborns also increased. The difference in estimated birthweight detriments from PM_2.5 in a newborn in one of the lowest deciles of birthweight versus one in the highest decile were larger when PM_2.5 exposures was higher. Birthweight detriments from PM_2.5 were similar at the lowest five birthweight deciles as shown in sensitivity analyses that included seasonality; these detriments were nonetheless more severe, though to a lesser degree, at the lowest deciles compared to the highest deciles. Although not investigating the relationship between PM_2.5 and birthweight, a previous study in Texas using quantile regression found more negative associations between ozone and gestational age in the lower deciles of gestational age, which, like birthweight, is an indicator of fetal growth.³² Recently, disparities in environmental health have come to the forefront and there have been studies suggesting factors other than PM_2.5 and commonly used individual and community variables can affect birth outcomes.^33–35 Despite the inclusion of covariates related to SES on the individual and community level, differences in severity in the negative association between PM_2.5 and birthweight remained. This could mean that PM_2.5 does in fact lead to more severe detriments in lighter newborns or that there are other covariates, not included in our analysis, that would explain the disparity of associations between the lower and higher deciles of birthweight.

Our study had several limitations. Although the PM_2.5 data set had high spatial resolution, it was nonetheless gridded, meaning that maternal addresses were assigned to prediction grid cells and there were instances where births were assigned to grid cells in which they were close to the edges of. As a result, there was possibility of non-differential bias due to exposure error. Future work can attempt to predict PM_2.5 exposure around individual locations instead of assignment to existing PM_2.5 data sets to fully take advantage of maternal address data. Another limitation with regards to the maternal address data was that they were reported and potentially inaccurate in representing where the mother was located during her pregnancy. Even though we expect the mothers to have spent more time at home than any place else, duties such as working somewhere away from the residence likely led to exposure misclassification. Since workplaces tend to be in high-density urban areas, which have higher PM_2.5 than lower-density residential areas, PM_2.5 assigned to each birth based on maternal residence may be underestimated compared to the actual full exposure including time at work. Thus, the birthweight change associated with a unit PM_2.5 increase was likely biased away from the null due activities that led to mothers spending time away from the residence. Another source of exposure misclassification was mothers moving residences during pregnancy, which we did not have information on. We expect the number of mothers who moved to be modest, but could not quantify this. This moving should introduce classical exposure error into the PM_2.5 exposure assigned to each birth, biasing coefficients toward the null. Thus, inaccurate reporting of residential address, due to moving residences, was unlikely to have changed our finding of differential associations along the birthweight distribution. With regards to the associations presented at different deciles of birthweight, it was possible that exposure misclassification varied at different deciles. Thus, the confidence intervals of the associations between PM_2.5 and birthweight could potentially be wider to varying degrees at different deciles. As in the case of most studies, our results were limited by residual confounding. While we included many covariates in our regression analyses, we did not have information on some possible confounders such as individual measures of SES. For example, personal or household income could potentially explain our finding of more severe PM_2.5 and birthweight associations at the lower deciles. Other instances of unavailable covariate data prevented us from exploring effect modification by maternal weight. Previous studies have shown that obese or underweight mothers are more likely to give birth to lighter newborns and that maternal obesity modifies the relationship between PM_2.5 and birthweight.^16,36 Finally, because we restricted our main analysis to term births, there was the possibility of an association between PM_2.5 exposure and birthweight induced by collider bias from unmeasured confounders of birthweight and being a term birth. Thus, the estimated PM_2.5 effects on birthweight should be interpreted as direct estimated effects of PM_2.5 exposure not mediated through mechanisms that lead to preterm births.

Our study had several strengths. We had high statistical power due to a large sample size of 775,768 and to exposure assignment using PM_2.5 exposure data with high temporal and spatial resolution. With the high coverage of the exposure data, we minimized the number of people who needed to be excluded based on where they lived. This is in contrast to many prior studies, which relied on data from air monitors and thus had more limited spatial coverage.^3,5,15,37 Avoiding exclusion of births due to limited spatial coverage of air pollution data led to lower bias due to missing exposure data and higher statistical power in our study. This study demonstrated that quantile regression can be applied to a large number of births, motivating future analyses applying this technique with related topics in the current and other existing birth registries. For example, one could investigate critical windows of exposure during pregnancy with quantile regression by assessing PM_2.5 over shorter periods or trimester of exposure. With further advancements in computing, it will become more practical to implement more flexible models with splines or distributed lags for PM_2.5 in a quantile regression model, which is computationally-intensive. Along with this, future applications of quantile regression in birth outcome studies may consider effect modification analyses with product terms between key covariates³⁸ In the present study, we did not have strong prior knowledge that a covariate, when stratified by another covariate, would plausibly affect birthweight in opposite directions; the body of knowledge will likely expand and future work may necessitate including additional product terms. Finally, our finding of more severe associations among lighter newborns could help focus future studies aimed at understanding the mechanism of how PM_2.5 could lead to decreased birthweight, which remains unclear. Two proposed pathways are inflammation and oxidative stress. Pulmonary inflammation stemming from PM_2.5 inhalation lead to poor gas exchange in the mother’s lungs and consequently, low oxygen and nutrition exchange to the fetus, resulting in growth restriction and lower birthweight.^17,39,40 PM_2.5 is also thought to cause oxidative stress, which can lead to DNA damage in both the mother and fetus, hindering fetal growth. These two proposed mechanisms are difficult to test in epidemiologic settings such as ours, which relied on birth certificate data and lacked biomarker data. With the knowledge that lighter newborns may be more severely affected, researchers could consider targeting lighter newborns for resource-intensive biomarker collection.

Conclusions

In full-term singleton live births in Massachusetts between 2001 and 2013, maternal exposure to PM_2.5 during pregnancy and birthweight were negatively associated with higher severity at lower quantiles of birthweight. We revealed the differences in associations by applying quantile regression, a seldom-used technique. Furthermore, we assigned exposure using PM_2.5 data that had high spatial and temporal resolution, which maximized spatial coverage and minimized exposure misclassification. Future work might seek to explain this environmental disparity between PM_2.5 and birthweight, perhaps by incorporating more individual measures of SES and biomarker data during or at birth.