Associations between ambient air pollutant concentrations and birth weight: a quantile regression analysis

Abstract

INTRODUCTION:

We investigated the extent to which associations of ambient air pollutant concentrations and birth weight varied across birth weight quantiles.

METHODS:

We analyzed singleton births ≥27 weeks gestation from 20-county metropolitan Atlanta with conception dates between 1 January 2002 and 28 February 2006 (N=273,711). Trimester-specific and total pregnancy average concentrations for 10 pollutants, obtained from ground observations that were interpolated using 12-km Community Multiscale Air Quality model outputs, were assigned using maternal residence at delivery. We estimated associations between interquartile range width (IQR_w) increases in pollutant concentrations and changes in birth weight using quantile regression.

RESULTS:

Gestational age-adjusted associations were of greater magnitude at higher percentiles of the birth weight distribution. Pollutants with large vehicle source contributions (carbon monoxide, nitrogen dioxide, PM_2.5 elemental carbon, and total PM_2.5 mass), as well as PM_2.5 sulfate and PM_2.5 ammonium, were associated with birth weight decreases for the higher birth weight percentiles. For example, whereas the decrease in mean birthweight per IQR_w increase in PM_2.5 averaged over pregnancy was −7.8g (95% CI: −13.6g, - 2.0g), the quantile-specific associations were: 10^th percentile −2.4g (−11.5g, 6.7g); 50^th percentile −8.9g (−15.7g, −2.0g); and 90^th percentile −19.3g (−30.6g, −7.9g). Associations for the intermediate and high birth weight quantiles were not sensitive to gestational age adjustment. For some pollutants we saw associations at the lowest quantile (10^th percentile) when not adjusting for gestational age.

CONCLUSIONS:

Associations between air pollution and reduced birth weight were of greater magnitude for newborns at relatively heavy birth weights.

INTRODUCTION

A substantial literature exists on the associations between urban air pollution and newborn birth weight. The authors of recent meta-analyses of observational studies^1–4 have been consistent in their summaries of the evidence base: notwithstanding the substantial between-study heterogeneity in the association estimates, when these estimates are pooled across studies, increases in several ambient air pollutant concentrations (specifically carbon monoxide (CO), nitrogen dioxide (NO₂), particulate matter less than 10 microns in diameter (PM₁₀), and particulate matter less than 2.5 microns in diameter (PM_2.5)) are associated with decreases in birth weight.

Various statistical methods have been used to estimate associations between air pollution and reduced birth weight. Standard approaches involve modeling term low birth weight (<2,500 g and gestational age >36 weeks) using logistic regression or modeling continuous birth weight using linear regression.⁴ These linear models relate fluctuations in air pollutant concentrations to changes in the mean of the birth weight distribution. The pooled effect estimates from such studies, as reported in meta-analyses, show small changes in birthweight associated with increases in pollution, e.g., a 16 g reduction in mean birthweight (95% CI: 5 g, 27 g) per 10 μg/m³ increase in pregnancy-averaged PM_2.5 concentrations.⁴ Associations of this magnitude can be of public health importance; a 16 g reduction in the birth weight distribution will increase the number of children born low birth weight (<2,500 g) and very low birth weight (<1,500 g). However, an outstanding question is whether the entire birth weight distribution shifts downward by a fixed amount (e.g., by 16 g), or if these reductions are localized in certain parts of the distribution. This question is important because a 16 g reduction for newborns weighing 3,500 g may be of less concern than a 16 g reduction for newborns weighing 2,000 g.

Quantile regression is a type of regression analysis well suited to addressing this question.⁵ In quantile regression, investigators can estimate how covariates are associated with changes in the response variable distribution at specific quantile levels. For example, associations with ambient air pollutant concentrations can be estimated for newborns at the low end of the birth weight distribution (e.g., the 10^th percentile of birth weight) as well at the high end.

We know of one previous study where quantile regression was used to estimate associations between ambient air pollution concentrations and birth weight.⁶ In that study, Smith et al. (2015) fit a multi-level quantile regression model to estimate associations between birth weight and ozone concentrations in Texas.⁶ The authors reported that second-trimester ozone exposure was associated with reductions in birth weight (adjusting for gestational age), with the magnitude of the reductions tending to be somewhat larger for the heavier newborns. Our study substantially adds to this evidence base, as here we estimate associations between 10 air pollutants and birth weight in Atlanta, Georgia, USA using quantile regressions to investigate the extent to which the magnitude of the associations vary across birth weight quantiles.

METHODS

Health and Air Quality Data

Daily ambient air pollutant concentrations were estimated by combining gridded (12-km by 12-km) numerical model simulations from the emissions-based Community Multi-scale Air Quality Model (CMAQ) with ground monitoring station observations in Georgia. We used results from a previously developed bias-correction method that optimally combines temporal variation in monitoring measurements at these locations and spatial variation in CMAQ simulations.⁷ Briefly, this approach involved using a modeled observation-based semivariogram and the covariance of observations and simulations to weight the contributions of observations and simulations to create spatially resolved estimates. We examined 10 pollutants: 1-hour maximum carbon monoxide (CO) and nitrogen dioxide (NO₂); 8-hour maximum ozone (O₃); 24-hour average particulate matter ≤10 microns in diameter (PM₁₀) and ≤2.5 microns in diameter (PM_2.5); and the PM_2.5 components sulfate (SO₄), nitrate (NO₃), ammonium (NH₄), elemental carbon (EC), and organic carbon (OC). We calculated daily concentrations for each grid cell. The resultant air pollutant fields agree well with monitor observations, with spatiotemporal Pearson r-square values between 0.80 and 0.98 across pollutants.⁷

We obtained birth records from the Office of Health Indicators for Planning, Georgia Department of Public Health. Emory’s Institutional Review Board approved the use of these data for this study. We included singleton live birth pregnancies with gestational lengths of 27–42 weeks and an estimated date of conception between 1 January 2002 and 28 February 2006. Maternal residence was restricted to 20-county Atlanta (Barrow, Bartow, Carroll, Cherokee, Clayton, Cobb, Coweta, DeKalb, Douglas, Fayette, Forsyth, Fulton, Gwinnett, Henry, Newton, Paulding, Pickens, Rockdale, Spalding, and Walton counties). These counties span 16,079 km². Additional exclusion criteria include: (1) maternal residential address at delivery unsuccessfully geocoded to the 2000 Census block group, (2) birth weight less than 400 grams, (3) mother’s age less than 15 years or greater than 44 years, (4) congenital anomaly identified on birth record, and (5) preterm births (births ≤ 36 weeks) with a procedure code for induction of labor. The final dataset contained 273,711 pregnancies. Each birth record was linked to a community multiscale air quality (CMAQ) grid cell using maternal address at delivery, and average exposures were calculated for first trimester (weeks 1–13), second trimester (weeks 14–26), third trimester (weeks 27-birth), and over the entire pregnancy.

Statistical Analysis

We estimated associations between ambient air pollution exposures during pregnancy and changes in birth weight at the 10^th, 30^th… 90^th percentiles using quantile regression, and we estimated associations for mean birth weight using linear regression. Separate models were created for each pollutant and each exposure window.⁵ Let τ ∈ [0,1], the conditional quantile model is specified as Q_Y(τ) = Xβ, where Q_Y(τ)denotes the τ^th quantile (or τ × 100^th percentile) of birth weight Y, and X denotes the design matrix with corresponding vector of regression coefficients β. Estimates of β are obtained by minimizing the difference between birth weight observations and Xβ subject to a quantile specific loss function. Estimates were adjusted for maternal age modeled as a nonlinear function using a natural cubic spline with 5 degrees of freedom (DF), maternal race (white/black/Asian/Hispanic), maternal education (less than 9^th grade, 9^th to 12^th grade, completed high school, some college and above), reported alcohol use during pregnancy, reported tobacco use during pregnancy, indicators of completed gestational weeks, indicators for county of residence at delivery, marital status (married vs. unmarried), Census block group percent poverty levels modeled using a natural cubic spline with 5 DF, and conception date modeled using a natural cubic spline with 12 DF to capture seasonal and long-term trends.

Regression coefficients from quantile and mean regressions were scaled by interquartile range (IQR) increases in pollutant concentrations averaged over pregnancy. We were consistent in the use of a total pregnancy IQR across analyses (as opposed to using trimester-specific IQRs) to facilitate comparison of results. Birth records were stratified by maternal race (black versus others) and by Census block group poverty (above versus below the median value) to investigate effect modification for the overall pregnancy exposure period by these two covariates. To investigate nonlinear associations, we used a natural cubic spline with three equidistant interior knots to model associations between air pollutant concentrations and birth weight quantiles in trimester-specific models. We also performed analyses to examine the consequences of gestational age adjustment, by 1) restricting analyses to full term births (births ≥37 weeks gestation); and 2) fitting quantile regressions during the first and second trimesters without adjusting for gestational age. We did not fit models for third trimester or total pregnancy exposures that were unadjusted for gestational age. Because the shorter third trimester averaging windows for preterm births yield more extreme air pollution averages (due to seasonality in air pollution), statistical methods that fail to account for gestational length can be biased for seasonal exposures.⁸ All analyses were performed in R version 3.3.1.

RESULTS

Table 1 shows descriptive statistics for the 273,711 births included in the analysis, stratified using the 25^th and the 75^th percentiles. Overall, 8% of births were preterm, 56% were to non-white mothers, and 34% were to mothers living in Census block groups that had >10% of the households below the U.S. federal poverty line. Children born to black mothers and children born to mothers living in Census block groups with >10% of households below the poverty line were more likely to be in the lowest birth weight quantile (< 25^th percentile) and less likely to be in the highest birth weight quantile (> 75^th percentile). Other sociodemographic variables, such as maternal education and maternal smoking, show similar patterns where mothers with no college education and mothers who smoke were overrepresented in the lowest birth weight quantile. Birthweight deciles among all births, term births, and preterm births are shown in eTable 1. Among all births, birth weight was 2,693 g at the 10^th percentile, 3,345 g at the 50^th percentile, and 3,941 g at the 90^th percentile (eTable 1).

Table 1.

Study population characteristics of live born singleton births in Atlanta with estimated dates of conception between 1 January 2002 and 28 February 2006, stratified by selected quantiles of newborn birth weight.

	< 25th birth weight percentile (<3005 g) (N=64,565)		25th - 75th birth weight percentile (3005 – 3657 g) (N=143,802)		> 75th birth weight percentile (>3657 g) (N=65,344)
Characteristic	N	Percent	N	Percent	N	Percent
Maternal race/ethnicity
White	21,597	34%	63,368	44%	36,091	55%
Black	27,908	43%	42,640	30%	13,645	21%
Hispanic	10,919	17%	29,813	21%	13,099	20%
Asian	3,763	6%	7,184	5%	2,167	3%
Other	378	1%	797	1%	342	1%
Maternal education
Less than 9th grade	4,890	8%	12,531	9%	5,235	8%
9th – 11th grade	11,566	18%	20,301	14%	7,369	11%
12th grade	19,913	31%	39,730	28%	16,838	26%
Some college	28,196	44%	71,240	50%	35,902	55%
Maternal age
Less than 25 years	28,892	45%	54,961	38%	19,843	30%
25 – 31 years	22,141	34%	55,116	38%	27,197	42%
More than 31 years	13,532	21%	33,725	24%	18,304	28%
Maternal tobacco use	4,546	7%	6,226	4%	1,872	3%
Maternal alcohol use	437	1%	865	1%	421	1%
Block group percent poverty > 10%	25,924	40%	49,089	34%	19,262	30%
Female infant sex	35,941	56%	71,612	50%	26,190	40%
Preterm birth	16,134	25%	5,506	4%	1,131	2%

Means and IQR_ws for pregnancy-averaged pollutant concentrations are shown in Table 2. Means and standard deviations for trimester-specific pollutant concentrations are provided in eTable 2 and eTable 3 shows pairwise Spearman correlations for total pregnancy average exposures. Traffic-related pollutants (CO, NO₂, PM_2.5, elemental carbon [EC]) are highly correlated (e.g., 0.92 for CO and NO₂; 0.87 for CO and EC). Other strong correlations include O₃ and NO₃ (ρ=−0.74), PM₁₀ and PM_2.5 (ρ=0.79), and PM_2.5 and NH₄ (ρ=0.84). The strong negative correlation between O₃ and NO₃ is related to the opposite dependence on temperature of these two secondary pollutants (O₃ peaks in summer and NO₃ in winter). The strong positive correlations of PM₁₀ and PM_2.5 and of PM_2.5 and NH₄ are related to PM_2.5 being a subset of PM₁₀ (approximately 70% by mass) and NH₄ being chemically associated with both SO₄ and NO₃ in PM_2.5; in total these ionic components account for approximately 50% of PM_2.5 mass.

Table 2.

Mean and interquartile range width (IQRw) of pregnancy-average pollutant exposures.

Pollutant	Mean	IQRw
1-hr carbon monoxide (CO) ppm	0.70	0.34
1-hr nitrogen dioxide (NO2) ppb	22.03	12.96
8-hr ozone (O3) ppb	41.73	6.91
24-hr PM10 μg/m3	22.61	3.03
24-hr PM2.5 μg/m3	15.90	1.86
24-hr PM2.5 sulfate (SO4) μg/m3	4.78	1.10
24-hr PM2.5 nitrate (NO3) μg/m3	0.67	0.22
24-hr PM2.5 ammonium (NH4) μg/m3	1.50	0.26
24-hr PM2.5 elemental carbon (EC) μg/m3	1.12	0.40
24-hr PM2.5 organic carbon (OC) μg/m3	3.02	0.33

Trimester-specific and overall pregnancy associations between interquartile range width increases in pollutant concentrations and changes in birth weight are shown in Figure 1 (numerical results provided in eTable 4). Estimates are presented for the 10^th, 30^th, 50^th, 70^th, 90^th percentiles and for the mean. With respect to mean changes (from linear regression models), decreases in birth weight for the overall pregnancy exposures were present for traffic related pollutants (CO, NO₂, PM_2.5, EC); for the secondary components of PM_2.5 with a strong summertime peak (SO₄ and NH₄); and for PM₁₀. For these pollutants, the quantile regression results showed decreases in birth weight occurring predominantly among children at or above the median birthweight (i.e., the 50^th percentile). For example, whereas the mean decrease in birthweight per IQR_w increase in PM_2.5 was −7.8 g (−13.6 g, −2.0 g), the quantile-specific associations were: 10^th percentile −2.4 g (−11.5 g, 6.7 g); 50^th percentile −8.9 g (−15.7 g, −2.0 g); and 90^th percentile −19.3 g (−30.6 g, −7.9 g). The consistency of this pattern across pollutants is not unexpected, given the high pairwise correlations of several pollutants.

Figure 1.

Trimester-specific (T1, T2, T3) and overall pregnancy associations between IQR increases in pollutant concentrations and changes in birth weight. Estimates are presented for the 10^th, 30^th, 50^th, 70^th, 90^th quantiles and for the mean.

Although trimester-specific differences were present for certain pollutants, the results for the first, second, and third trimester associations were broadly similar (Figure 1). We did not see evidence that one trimester had consistently larger reductions in birth weight. Looking across pollutants and trimesters, the highest magnitude associations were frequently at one of the higher birthweight quantiles (i.e., the 70^th or 90^th percentile). These estimates were typically negative (the exception being an increase in birth weight for third trimester ozone exposure), and they showed a pattern of increasing in magnitude at the higher quantiles. Conversely, the association estimates at the lower quantiles (i.e., the 10^th and 30^th percentiles) almost always had point estimates near 1.0.

With respect to effect modification for the overall pregnancy exposure period, the most divergent estimates were for maternal socioeconomic status (defined as maternal residence in a block group with >7.2% of households below poverty versus residence in a block group with ≤ 7.2% of households below poverty). These results, along with those for effect modification by maternal race (black versus other), are shown in Figure 2 (numerical results provided in eTable 4). Across pollutants, associations tended to be of greater magnitude for children born to mothers residing in high poverty block groups. PM_2.5, NH₄, and the three pollutants considered to be markers of pollution from motor vehicle emissions (CO, NO₂, and EC) all displayed this pattern. Effect modification by maternal race was not evident (Figure 2). For both sets of results, the estimated decreases in birth weight tended to be greatest for the heaviest children (i.e., those in the 90^th percentile of birth weight), similar to the overall results shown in Figure 1.

Figure 2.

Overall pregnancy associations between interquartile range width increases in pollutant concentrations and changes in birth weight, stratified by maternal socioeconomic status (maternal residence in a block group with >7.2% of households below poverty versus residence in a block group with ≤ 7.2% of households below poverty) and by maternal race (black versus other). Estimates are presented for the 10^th, 30^th, 50^th, 70^th, 90^th quantiles and for the mean.

Estimates from trimester-specific quantile regressions that used natural cubic splines with three interior knots within the exposure range to model nonlinear associations are shown in eFigure 1. Qualitatively, the nonlinear estimates showed a similar pattern as the linear estimates, with birth weight reductions tending to be of larger magnitude at the larger (heavier) birth weight quantiles. Confidence intervals from the linear and nonlinear estimates overlapped considerably.

We also performed analyses to examine the impact of gestational age adjustment. When analyses were restricted to full term births (results shown in eFigure 2), the confidence intervals widened, and most associations were either of similar magnitude or increased magnitude when compared with results from our primary analysis. Comparing results in this manner is complicated by the fact that the two birth weight distributions differ, e.g., the median birth weight among full-term births is heavier than the median birth weight among all births. Results from quantile regressions during the first and second trimesters that did not include adjustment for gestational age are shown in Figure 3. Pollutant associations for the intermediate and high birth weight quantiles (30^th, 50^th, 70^th, and 90^th percentiles) were not sensitive to adjustment for gestational age; however, the estimates for the lowest quantile (10^th percentile) were sensitive. In particular, for models that did not include gestational age adjustment, the three pollutants that are markers of motor vehicle emissions (CO, NO₂, and EC) were all associated with decreases in birth weight at the lowest birth weight quantile for both first and second trimester exposures.

Figure 3.

First-trimester (T1) and second-trimester (T2) quantile regression estimates and linear regression estimates for all births with adjustment for gestational age (red) and without adjustment for gestational age (blue).

DISCUSSION

Increases in the concentrations of several ambient air pollutants were associated with reductions in mean birth weight. The direction of these associations is consistent with previously published literature, e.g., we observed a decrease in mean birth weight of −7.8 g (−13.6 g, −2.0 g) per IQR_w increase in pregnancy-averaged PM_2.5 (IQR_w = 1.86 μg/m³), whereas the summary estimate from a recent meta-analysis was −16 g (−27 g, −5 g) per 10 μg/m³ increase in pregnancy-averaged PM_2.5.⁴ Our estimates from quantile regression suggest that much of this reduction occurred in heavier children.

If future studies produce similar findings – that reductions in birth weight occur predominantly at the high end of the birth weight distribution – then the implication is different from findings that suggest the entire distribution shifts downward by a fixed amount. A small change in birth weight caused by air pollution for a relatively heavy infant, although undesirable, is not likely to be clinically significant. Associations between birth weight and several health outcomes resemble a J-shape,⁹ and a wide range of birth weights are considered healthy.¹⁰ There is little evidence that small perturbations of birth weight within a loosely defined “healthy” range (i.e., 2,500 – 4,000 g) could meaningfully affect child and adult health outcomes. Conversely, for low birth weight and very low birth weight infants, small decreases in weight may be clinically meaningful; in this part of the birth weight distribution, infant mortality rates rise sharply in relation to decreases in birth weight.¹⁰ If this relationship is due to a causal effect of birth weight on mortality,¹¹ then small decreases in birth weight would be a cause for concern.

In our primary analyses we chose to adjust for gestational age. When gestational age is not adjusted for, the birth weight quantiles become highly correlated with gestational age – babies born preterm predominate in the lowest quantile, with early term babies (37 – 38 weeks) and full term babies (39+ weeks) distributed throughout the higher quantiles accordingly. In this situation interpretation becomes challenging, i.e., is air pollution associated with a reduction in gestational age, a reduction in birth weight, or both? However, gestational age adjustment raises other concerns. For example, if air pollution causes reductions in gestational age, then gestational age adjustment will bias the effect estimate for the total effect of air pollution on birth weight. This scenario is plausible given that increases in ambient air pollutant concentrations have been associated with small reductions in gestational age in some studies.^1,2 Unmeasured mediator–outcome confounding is also a concern. If there are unmeasured common causes of both gestational age and birth weight, then adjustment for gestational age will induce collider bias (because gestational age would be an effect of both air pollution and the unmeasured common cause). Because the association between air pollution and preterm birth is of small magnitude, this bias is likely to be weak; however, due to these concerns, we present associations between air pollutant concentrations and birth weight quantiles with and without adjustment for gestational age. Pollutant associations for the intermediate and high birth weight quantiles (30^th, 50^th, 70^th, and 90^th percentiles) were not sensitive to adjustment for gestational age; however, the estimates for the lowest quantile (10^th percentile) were sensitive. When we adjusted for gestational age we did not find associations with any of the pollutants for this birth weight quantile. When we did not adjust for gestational age we observed associations with several pollutants (Figure 3).

One limitation of our study is we fit single pollutant models. Associations tended to be similar across highly correlated pollutants (e.g., CO, NO₂, and EC), and we did not attempt to estimate associations for certain pollutants conditional on others, nor did we estimate joint associations for multiple pollutants. We also chose to limit our analyses to trimester-specific and total pregnancy average pollutant exposures because fetal growth occurs throughout pregnancy. Misspecification of the relevant gestational window can lead to biased effect estimates,¹² and data-driven approaches for identifying critical pregnancy windows have been developed.^12,13 These are interesting directions for future research. Exposure measurement error is also a concern, and in this study we used outputs from a chemical transport model that were fused with ground-based observations from monitoring stations.⁷ Although this fusion approach lessens the biases that are present in the CMAQ model outputs, it cannot reduce all errors to zero, and the 12-km resolution outputs will smooth over finer-scale spatial gradients in pollutant concentrations.¹⁴ The estimates of pollutant gas concentrations that originate predominantly from mobile sources, i.e. NO₂ and CO, are most affected by these limitations. The consequence of exposure measurement error (when the errors are classical and additive) for quantile regression is similar to that of linear regression, with slopes expected to be biased towards zero.¹⁵

In quantile regression, it is also important that exposure measurement errors be non-differential with respect to birth weight. For example, if classical measurement errors were larger for low birthweight newborns, then the associations between air pollutant concentrations and birthweight could be more biased towards zero at the lower quantile levels. This could result in a spurious conclusion that air pollutants more strongly impact the heavier infants. Similar to linear regression, exposure measurement errors could also lead to false conclusions about effect modification. For example, in our effect modification analyses, children born to mothers residing in Census block groups where >7.2% of households were below the poverty line tended to have associations of larger magnitude than their counterparts. These block groups with high poverty levels are predominantly (but not exclusively) located in neighborhoods immediately to the south and west of downtown Atlanta. Most block groups in the outlying suburbs, where air quality monitors are sparse, are included in the high socioeconomic status stratum. If there is more exposure measurement error for high socioeconomic status women (perhaps because of longer commute distances or because of larger air quality modeling errors), then it might be an explanation for why weaker associations were observed in the higher socioeconomic status group. Another source of error is that exposures were assigned to mothers based on residence at the time of delivery, as maternal residential history throughout pregnancy is not available on birth records. However, investigators who have quantified the bias caused by maternal mobility (in settings outside of quantile regression) have found it to be negligible.^16–18

Data errors on birth records were also likely present.¹⁹ For example, the prevalence of maternal smoking in our population was 4.6%, which is much lower than the national average of 13%,²⁰ suggesting underreporting of maternal smoking. Maternal smoking is a strong predictor of birth weight,²¹ and if maternal smoking was associated with ambient air pollutant levels in our study then residual confounding would be present. Fortunately the measurements of many variables on birth records, such as birth weight and maternal age, have excellent reliability and validity.¹⁹

Although the research questions were similar, our statistical approach differed from the quantile regression approach used by Smith et al. (2015) to estimate associations between ambient air pollution concentrations and birth weight in Texas.⁶ In that study the authors fit a Bayesian multilevel model with gestational age as the cluster variable, which allowed the association estimates to vary smoothly across gestational weeks. All gestational weeks were modeled simultaneously, and monotonic splines were used to force the effect of each covariate to change monotonically with increasing quantile levels. By borrowing information from nearby quantile levels, the authors were able to reduce the random error in the quantile-specific association estimates.

In our study we also modeled birth weight across all gestational ages simultaneously. Unlike Smith et al. (2015), we did not force a monotonic relationship between the covariate effects and the quantile levels, and instead modeled the quantile functions at specific levels one at a time. In this respect our approach is more similar to standard quantile regression methods that have been used previously.⁵ Although smoothing across quantile levels can reduce random error, for a model with many covariates (such as ours, which had 65 covariates), constraining the splines to be monotone with increasing quantile level across all covariates results in substantial computational demands, and fitting such a model to our data was not feasible. Because we estimated quantile functions one at a time, some of these estimates likely had substantial random error. Even so, we observed similar trends in the associations across multiple pollutants and exposure windows, which somewhat lessens this concern about random error. As statistical research on methods aimed to characterize the entire outcome distribution continues to advance, along with the development of practical software implementation, more nuanced associations of environmental exposures can be explored.

CONCLUSION

We found that increases in the concentrations of several ambient air pollutants were associated with decreases in birth weight in Atlanta. Gestational age-adjusted associations were of highest magnitude for children who were relatively heavy (i.e., the 70^th and 90^th percentile of birth weight). In regression models that did not adjust for gestational age we saw similar patterns, with additional findings for some pollutants at the lowest quantile. Further investigations are needed to evaluate whether similar patterns of associations are present in other populations; fortunately, many investigators have existing datasets that could be reanalyzed using quantile regression to examine this issue.