PM_2.5 exposure and birth outcomes: Use of satellite- and monitor-based data

Abstract

Background

Air pollution may be related to adverse birth outcomes. Exposure information from land-based monitoring stations often suffers from limited spatial coverage. Satellite data offer an alternative data source for exposure assessment.

Methods

We used birth certificate data for births in Connecticut and Massachusetts, U.S. (2000-2006). Gestational exposure to PM_2.5 was estimated from US Environmental Protection Agency monitoring data and from satellite data. Satellite data were processed and modeled using 2 methods – denoted satellite (1) and satellite (2) – before exposure assessment. Regression models related PM_2.5 exposure to birth outcomes while controlling for several confounders. Birth outcomes were mean birth weight at term birth, low birth weight at term (LBW <2500g), small for gestational age (SGA, <10th percentile for gestational age and sex), and preterm birth (<37 weeks).

Results

Overall, the exposure assessment method modified the magnitude of the effect estimates of PM_2.5 on birth outcomes. Change in birth weight per inter-quartile range (2.41 μg/m3)-increase in PM_2.5 was -6g (95% confidence interval = -8 to -5), -16g (-21 to -11) and -19g (-23 to -15), using the monitor, satellite (1) and satellite (2) methods, respectively. Adjusted odds ratios, based on the same 3 exposure methods, for term LBW were 1.01 (0.98 to 1.04), 1.06 (0.97 to 1.16), and 1.08 (1.01 to 1.16); for SGA, 1.03 (1.01 to 1.04), 1.06 (1.03 to 1.10) and 1.08 (1.04 to 1.11); and for preterm birth, 1.00 (0.99 to 1.02), 0.98 (0.94 to 1.03) and 0.99 (0.95 to 1.03).

Conclusions

Under exposure assessment methods, we found associations between PM_2.5 exposure and adverse birth outcomes particularly for birth weight among term births and for SGA. These results add to the growing concerns that air pollution adversely affects infant health and suggest that analysis of health consequences based on satellite-based exposure assessment can provide additional useful information.

Air pollution adversely affects human health.^1-3 Specifically, particulate matter is associated with respiratory and cardiovascular disease.^4,5 Maternal exposure to particulate matter, PM_2.5 (particles with aerodynamic diameter ≤2.5μm) is associated with several birth outcomes, although findings are not completely consistent across studies.^6-9 Birth outcomes that have been assessed include birth weight, term low birth weight (LBW; birth weight <2,500g for term births [gestational age ≥37 weeks]), and small for gestational age (SGA; birth weight <10^th percentile for gestational age and sex).

Air pollution and birth outcomes are an important topic of research. The economic burden in the United States associated with preterm birth, which include social and healthcare costs, was $26.2 billion in 2005.¹⁰ Cost of hospitalization for low birth weight/preterm birth in the US was $5.8 billion in 2001.¹¹ Studies have also shown that particulate matter may be associated with inflammation in pregnant women^12,13 and affect fetal growth¹⁴—both of which may be detrimental to a normal course of pregnancy and fetal development. Also, consequences of adverse birth outcomes beyond the perinatal period may include delayed development and decreased academic achievement¹⁵ and short stature¹⁶ in childhood, as well as medical/social disabilities¹⁷ and respiratory disease¹⁸ in adulthood. Given this social, economic and health burden of adverse birth outcomes and the ubiquity of air pollution exposure, there is a need to better understand the health risks posed by airborne particulate matter and other environmental toxins/hazards.^6,7,19,20

International collaborative efforts²¹ and several U.S. studies have found associations between PM _2.5 and birth outcomes (LBW⁹, term LBW,^22,23 birth weight,^9,24-26 and SGA^24,27). However, other studies have found no or null associations between PM_2.5 and birth outcomes (term LBW,^28,29 birth weight,^25,30 SGA,²⁵ and preterm birth^7,31). In almost all of these studies, data for exposure assessment were obtained from central monitoring sites operated and maintained by state and national agencies, such as the US Environmental Protection Agency (EPA), primarily for regulatory purposes.

Use of data from regulatory monitors is a reasonable and cost-effective method to estimate exposure for air pollution studies; however, major challenges of this approach include limited spatial and temporal coverage. In the US, monitors are located primarily in densely populated urban centers. Since monitors record air pollution levels at a specific time and location, exposure estimates for persons located far from monitors may not be possible or, if estimated, may be less reliable. Many studies limit subjects to those within a certain distance from the monitor. The choice of distance depends on the pollutant's spatial heterogeneity, temporal correlation in pollutant levels nearby monitors, and other regional-scale characteristics (e.g., industry type, population density, traffic patterns). Temporal coverage is another limitation of data from existing monitoring systems. For example, in the EPA's Air Quality System (AQS), frequency of data collection can vary by site, pollutant, time of year and start date of measurement. Particles are often measured every three or six days. Therefore, it is not uncommon for data from central monitoring systems to be missing or unreliable. Given these considerations, alternative methods for exposure assessment are needed.

Air quality modeling, land-use regression models, and satellite-based predictions are some of the methodologies being developed to predict air-pollution levels in epidemiology studies.³² For birth outcomes studies, the first two methodologies are more common^22,23,33-36 (although see the paper by Kloog and colleagues²⁶ for a satellite-based approach). Our group^37,38 recently produced estimates of PM_2.5 levels that have higher predictability than current land-use regression models and satellite methods. Specifically, novel methods were developed for calibrating satellite-based measurements of aerosol optical depth (a measure of light extinction due to aerosols in the atmospheric column³⁸), and statistical modeling was used to address missing data (due to cloud cover) in these calibrated data.³⁷ From these studies, we had access to PM_2.5 estimates that were highly predictive of PM measurements (R²=0.88 for cross-validated model, as reported by Lee et al.³⁷). Low predictability between modeled and monitor-based values likely introduces greater uncertainty in health effect estimates.

We investigated PM_2.5 and birth outcomes using a traditional exposure approach (existing monitoring data) and an emerging method (modeled estimates based on satellite data). We consider how these relationships are affected by including observations with satellite-based exposures but no monitor-based exposure estimates to assess the potential added value of satellite-based estimates for exposure.

METHODS

Data and outcome assessment

We obtained detailed birth certificate data for all births in Connecticut and Massachusetts for 2000-2006 (n=834,332). We excluded births that were conceived before the year 2000 because PM_2.5 exposure data from satellite methods was not yet available. The birth certificate data included maternal characteristics (residential address, age, education, parity, tobacco use, marital status, race/ethnicity), birth characteristics (date of last menstrual period, prenatal care and type of birth) and baby's characteristics (date of birth, birth weight, type of birth and gestational age).

We excluded births that were missing residential address (2%), non-singleton deliveries (5%), birth weight <1,000g or >5,500g (1%), and births with implausible gestational age-birth weight combinations (0.02%). These criteria have been applied in similar research.³⁹ Births with gestational age <20 weeks or >46 weeks were excluded (0.3%). Clinical gestational age was used in all analyses; when missing, we used the calculated gestational age when available.

For analysis of mean birth weight and term LBW, only term births (gestational age ≥37 weeks) were included. Preterm births were those with gestation <37 weeks. We classified births as small for gestational age if birth weight was <10^th percentile value for gestational age and sex according to U.S. data-based cut-off values (restricted to gestational ages 22 to 44 weeks).⁴⁰ Therefore, we limited the SGA analysis to births with gestational age in this range. The final number of observations included in each model differed based on the health outcome and exposure assessment method.

We used date of birth and gestational age to establish start and end dates of gestational exposure and to estimate exposure during the entire pregnancy and each trimester. Trimesters were defined as 1-13 weeks, 14-26 weeks, and 27 weeks until birth. Trimester-specific apparent temperature was estimated using data from the National Climatic Data Center.⁴¹

Exposure assessment

Two methods were used for PM_2.5 exposure assessment: monitor data and modeled estimates based on satellite data. Monitor data were obtained from the EPA's Air Quality System system for 2000-2006 (Fig.). We omitted monitor data with qualifier codes indicating uncommon, natural or anthropogenic events, and data with quality assurance issues, based on EPA flag codes. We used the closest monitor to mothers’ residence with a cut-off distance of 50km, based on our previous work (K Ebisu, unpublished data). The average distance between monitor and mothers’ location was 14km (standard deviation = 11km; 25% quartile = 5km; and 75% quartile = 21km). PM_2.5 values were calculated for each of these monitors and for each week of pregnancy. This process avoids biasing exposure estimates because some monitors did not provide concentration data for the entire study period. Analysis excluded births with exposure data for fewer than 75% of weeks in each trimester. The closest monitor to the mother’s residence that met these criteria was used to estimate overall and trimester-specific exposure by averaging weekly values.

Figure.

Map of the study area, which includes counties in Connecticut and Massachusetts (thick black lines). Environmental Protection Agency – Air Quality System (EPA-AQS) monitoring sites located within 50km of the state boundary are shown in grey triangles. Grid cells (10x10km; thin black lines) for satellite methods are overlaid on the study area. Note that not all monitors and grid cells were used in exposure assessment due to data quality and availability issues.

Satellite-based PM_2.5 predictions were modeled under two related yet slightly different methods, which we denote Satellite (1) and Satellite (2) (described below), using measurements from the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument onboard Terra and Aqua satellites. Both modeling approaches produced daily PM_2.5 concentrations for each 10x10km grid cell over our study area (Fig.). These data were available only for 2000 to 2006. Satellite data consisting of aerosol optical depth measurements were obtained from National Aeronautics and Space Administration. Other researchers^42,43 have used AOD measurements directly to estimate PM_2.5, (via a functional relationship between AOD and PM_2.5). We elected not to do so because of lack of high predictability and because of missing data due to cloud cover.

To address these two limitations, we used a calibration and modeling approach, which has been described earlier.^37,38 Briefly, we start by using a mixed-effects model to generate relationships between each day of observed PM_2.5 levels in Northeast US and AOD values corresponding to monitors’ locations. In the mixed effects model, fixed effects explained the average intercept and the slope of the PM_2.5-AOD slope for the entire study period, and random effects accounted for daily variability of PM_2.5-AOD relationships. This daily AOD calibration approach substantially enhanced the PM_2.5 predictive power of AOD, rendering it a robust predictor of PM_2.5. Next, we performed a cluster analysis using the K-means method, which breaks up data into K clusters (K=9 for the Satellite (1) method and K=8 for the Satellite (2) method), such that the data point in each cluster is closest to the mean of the cluster.⁴⁴ This method of classification allowed us to identify the set of days with a similar spatial pattern of PM_2.5. The cluster analysis under the Satellite (1) method was based on PM_2.5 concentration differences between observed PM_2.5 values and regional PM_2.5 values (i.e., daily averages of all available PM_2.5 measurements over the study region on a given day), whereas under the Satellite (2) method we used actual observed PM_2.5 values.

Another difference between the two satellite methods was in how we predicted PM_2.5 values for days with missing AOD data. In the Satellite (1) method, we formulated a general additive model (GAM) for each cluster, in which predicted PM_2.5 concentrations from the mixed effects model were regressed on regional PM_2.5 levels and a spatial smooth function of latitude and longitude. Here, regional PM_2.5 accounted for daily variability in PM_2.5 levels. In this way, we generated a single spatial surface of PM_2.5 concentrations for each cluster and predicted PM_2.5 concentrations for days with missing data. In contrast, in the Satellite (2) method, we assumed that relationships between predicted (from mixed effects model) and regional PM_2.5 concentrations in each grid cell were constant for each cluster. Thus, we derived cluster- and grid-specific relationships using regression models and estimated all the missing PM_2.5 concentrations. Both approaches produced PM_2.5 estimates that were highly predictive, and, therefore better suited to health effects studies. In summary, the main differences between the two satellite methods were in how observed data were clustered to identify spatiotemporal patterns, and how each cluster of data was used to predict PM_2.5. The Satellite (2) approach provided greater spatial heterogeneity in predicted PM_2.5 values.

A recent birth outcomes study in Eastern Massachusetts used a different method to estimate PM_2.5 from satellite data.²⁶.The main differences relate to calibration and modeling of raw satellite data in terms of the modeling approach, and the use of different variables in the calibration and modeling steps.

As with exposures based on monitor data, weekly exposures were used to calculate trimester-specific and overall exposure during gestation for each birth based on mother's residence. Satellite data were unavailable for grid cells containing mostly water (0.5% and 3% of births in Connecticut and Massachusetts, respectively). For these observations, we used satellite data from the closest grid cell (based on grid centroid-to-residence ≤10km). We excluded subjects with a residential address on islands because satellite estimates were unavailable for such geographic areas.

Statistical analysis

We formulated several models based on the combination of data used for exposure assessment (Monitor, Satellite [1], or Satellite [2]; and the included observations (subset of births with exposure estimates under both land-based monitors and satellite methods [Joint], or all subjects with exposure estimates for that exposure method [All]).

For example, a model labeled “Satellite (2), All” includes all subjects with estimated PM_2.5 exposure based on the second method of modeling satellite data. Note that all “Joint” models have the same sample size, whereas “All” models have different sample sizes. “Satellite (1)” and “Satellite (2)” models had identical timeframe/spatial resolution (i.e., same subjects). Each model was applied to the four birth outcomes separately. This approach allows evaluation of the association between exposure method and the health effects for the same study population. Also, we can evaluate how inclusion of subjects with exposure estimates for one exposure method but not another modifies effect estimates.

We used logistic regression for binary outcomes (term LBW, SGA, preterm birth) and linear regression for the continuous outcome birth weight. We controlled for the following variables: mother's age (<20, 20-24, 25-29, 30-34, 35-40, ≥ 40years); marital status (married, not married); mother's education (<12, 12, 13-15, ≥16 years); mother's race/ethnicity (white/non-Hispanic, black/non-Hispanic, Asian/non-Hispanic, Hispanic/other-Hispanic, other/non-Hispanic or unknown ethnicity); prenatal care (Adequacy of Prenatal Care Utilization Index⁶⁴: unknown/missing, inadequate, intermediate, adequate [basic or intensive]); smoking (none, 1-9, 10-20, >20 cigarettes/day); type of birth (vaginal/vaginal after cesarean birth vs. cesarean); parity (0, 1, 2, ≥3 previous live births); season of conception (winter, spring, summer, fall); medical risk factors (0 or ≥1 factors, e.g. anemia); medical risk due to previous preterm birth or SGA (yes, no); baby's sex (boy, girl); and gestational age (continuous). All models controlled for year of conception, trimester-specific apparent temperature, and state of residence (Connecticut or Massachusetts). We also did a sensitivity analysis using mean instead of apparent temperature. For each outcome, we evaluated the effects of overall gestational exposure and of first-, second-, and third-trimester exposure. For trimester models, we included residuals from regressing exposure estimates from the trimester of interest against other trimesters to control for correlation in exposures among trimesters, similar to methods used previously.⁹

RESULTS

Low birth weight was observed in 2% of all term births (n=628,131), with overall mean birth weight of 3,449g (standard deviation 472g). Ten percent (n=662,921) of infants were SGA and 6% (n=656,769) were preterm (Table 1). The sample size for preterm births reflected our exclusion of all births occurring 37 weeks before 31 December 2006 (end of study period). This exclusion rule was necessary to ensure that all births in 2006 had an equal chance of being counted as a preterm birth. For SGA, we included only births with gestation 22-44 weeks, which resulted in different sample sizes for SGA and LBW. Descriptive statistics on other covariates were based on all eligible births (n=662,921) (Table 1). Mothers were mainly white with non-Hispanic ethnicity (68%), educated (41% with >16 years education), and married (70%). For a majority of births, prenatal care was deemed adequate (82%) and the method of delivery was vaginal or vaginal after cesarean (73%).

Table 1.

Descriptive statistics of maternal and child characteristics and birth outcomes from birth certificate data in Connecticut and Massachusetts (2000-2006). N=662,921 for all variables except where limited to births in a specific ranges of gestational length (e.g., term births or small for gestational age). Percentages are rounded to whole numbers.

Variables	No. (%)a
Birth outcomes
Birth weight (in grams, for term births only); mean (SD)	3,449 (472)
Term low birth weight (term births <2500g)	11,641 (2)
Preterm birth (gestational age <37 weeks)	41,868 (6)
Small for gestational age (birth weight <10th percentile for gestational age and sex)	67,842 (10)
Baby's sex
Boy	338,957 (51)
Girl	323,964 (49)
Type of birth
Vaginal/Vaginal birth after cesarean birth	482,358 (73)
Cesarean/Repeated cesarean	180,563 (27)
Adequacy of Prenatal Care Utilization Index64
Unknown/Missing	6,518 (1)
Inadequate	51,826 (8)
Intermediate	57,471 (9)
Adequate (Basic or Intensive)	547,106 (82)
Average number of cigarettes per day
None	613,155 (93)
1-9	28,025 (4)
10-20	21,025 (3)
>20	716 (0.1)
Mother's education (years)
<12	82,526 (12)
12	165,549 (25)
13-15	142,460 (22)
≥ 16	272,386 (41)
Mother's race/ethnicity
White/non-Hispanic	448,330 (68)
Black/non-Hispanic	60,463 (9)
Asian/non-Hispanic	40,698 (6)
Hispanic/other-Hispanic	99,211 (15)
Other/non-Hispanic or unknown ethnicity	14,219 (2)
Mother's marital status
Married	464,155 (70)
Unmarried	198,766 (30)
Mother's age (years)
<20	43,535 (6)
20-24	106,454 (16)
25-29	156,967 (23)
30-34	210,351 (31)
35-39	119,400 (18)
≥ 40	26,214 (4)
Season and weather
Season of conception
Winter	173,499 (26)
Spring	160,963 (24)
Summer	162,735 (25)
Fall	165,724 (25)
Apparent temperature (°C); mean (SD)	48.2 (7.2)
Parity (number of previous births)
0	290,134 (44)
1	228,434 (34)
2	97,127 (15)
≥ 3	47,226 (7)
Previous preterm birth or small-for-gestational-age birth	6,090 (1)
Gestational age (weeks); mean (SD)	39.0 ± 1.6

^aExcept where otherwise specified.

Satellite PM_2.5 exposures were estimated for 367 10x10km grid cells, whereas there were 98 EPA-Air Quality System monitors providing point measurements. Monitoring sites were located within Connecticut or Massachusetts or within 50km of their borders. Mean PM_2.5 exposure during the entire pregnancy based on each method of exposure assessment (Monitor, Satellite [1] and Satellite [2], were similar – 11.9 ,11.2 , and 11.4 μg/m³, respectively, but differed in other statistical properties (Table 2). Satellite-based exposure estimates tended to have smaller standard deviations, narrower ranges and smaller inter-quartile ranges (IQR). These differences were also apparent in trimester-level estimates, where confidence intervals for the third trimester were wider than for other trimesters (Table 2). These wider intervals may be due to variable lengths of exposure in births that reached the third trimester. For all models, we reported results using an increment of 2.41μg/m³ (IQR of exposure during gestation using monitor-based data), in order to make effect estimates comparable across analyses.

Table 2.

Descriptive statistics for exposure data for PM_2.5 (μg/m³), based on 3 exposure methods.

Duration of exposure	Mean	SD	Min	Lower Quartile	Upper Quartile	Max	IQR
Monitor
Entire pregnancy	11.91	1.89	4.02	10.66	13.07	19.97	2.41
1st trimester	12.03	2.66	3.24	10.19	13.66	24.50	3.46
2nd trimester	11.93	2.58	3.24	10.13	13.57	24.50	3.44
3rd trimester	11.81	2.55	0.50	10.02	13.58	43.30	3.56
Satellite (1)a
Entire pregnancy	11.15	0.75	8.77	10.66	11.70	13.75	1.04
1st trimester	11.17	1.48	6.76	10.20	12.03	15.84	1.83
2nd trimester	11.16	1.49	6.75	10.19	12.04	15.75	1.86
3rd trimester	11.11	1.59	4.09	10.08	12.10	24.63	2.02
Satellite (2)a
Entire pregnancy	11.36	0.85	8.82	10.79	11.92	14.81	1.12
1st trimester	11.38	1.59	6.73	10.28	12.34	17.09	2.06
2nd trimester	11.36	1.60	6.71	10.27	12.35	17.09	2.08
3rd trimester	11.32	1.70	4.23	10.18	12.40	31.59	2.22

^asee main text for differences in Satellite(1) and Satellite(2)

IQR indicates interquartile range.

Gestational PM_2.5 exposure was associated somewhat differently with the various birth outcomes (Table 3). Term LBW and SGA were generally associated with PM_2.5 across all exposure methods, although more strongly with Satellite data, and especially Satellite (2) data. PM_2.5 exposures were linked with increased risk of term LBW only in the first trimester, while SGA was linked with exposures in all trimesters (although more weakly) (Table 4). A consistent gradient in risk by exposure method was observed in the models across most trimesters (Table 3). Risk of term LBW per IQR increase in PM_2.5 was 1% (95%CI= -0.02 to 4), 6% (-0.03 to 16) and 8% (1 to 16), using Monitor, Satellite (1), and Satellite (2) methods, respectively. The change in birth weight was negatively associated with PM_2.5 exposure, regardless of window of exposure. The change in birth weight per IQR increase in PM_2.5 was -6g (95%CI= -8 to -5) using the Monitor method and about three times that using either satellite method. The risk of SGA, when using satellite methods was 6% (3 to 10) for Satellite (1) and 8% (4 to 11) for Satellite (2). These risk values are about twice that using the Monitor method (3%, [1 to 4]). For preterm birth, risks were marginally higher risk for some exposure methods, but with no clear excess either overall or by trimester. Our results were not sensitive to using mean instead of apparent temperature (results not shown). Models with satellite-based exposure estimates tended to have much wider confidence intervals (Table 3).

Table 3.

Adjusted^a association between PM_2.5 exposure over the entire pregnancy and birth outcomes, by exposure data source and by type of birth outcomes. Effect estimates (95%CI) are reported as per inter-quartile range (2.41 μg/m³) increase in PM_2.5.

Modelb			No.
Data source	Observations	OR of term LBW
Monitor	All	1.01 (0.98 to 1.04)	619,675
	Joint	1.01 (0.98 to 1.04)	609,813
Satellite (1)	All	1.07 (0.99 to 1.17)	628,131
	Joint	1.06 (0.97 to 1.16)	609,813
Satellite (2)	All	1.09 (1.02 to 1.17)	628,131
	Joint	1.08 (1.01 to 1.16)	609,813
OR of SGA
Monitor	All	1.03 (1.01 to 1.04)	654,193
	Joint	1.03 (1.01 to 1.04)	643,839
Satellite (1)	All	1.07 (1.03 to 1.11)	662,921
	Joint	1.06 (1.03 to 1.10)	643,839
Satellite (2)	All	1.08 (1.05 to 1.11)	662,921
	Joint	1.08 (1.04 to 1.11)	643,839
Change in birth weight (g)
Monitor	All	−6.2 (−7.9 to −4.6)	619,675
	Joint	−6.2 (−7.9 to −4.6)	609,813
Satellite (1)	All	−14.6 (−19.4 to −9.8)	628,131
	Joint	−15.7 (−20.5 to −10.8)	609,813
Satellite (2)	All	−19.1 (−23.1 to −15.1)	628,131
	Joint	−19.0 (−23.0 to −14.9)	609,813
OR of preterm birth
Monitor	All	1.00 (0.99 to 1.02)	647,942
	Joint	1.00 (0.99 to 1.02)	637,586
Satellite (1)	All	0.98 (0.93 to 1.02)	656,769
	Joint	0.98 (0.94 to 1.03)	637,586
Satellite (2)	All	1.00 (0.96 to 1.04)	656,769
	Joint	0.99 (0.95 to 1.03)	637,586

^aAll models controlled for the following confounders mother's age, marital status, education, race/ethnicity, prenatal care, smoking, type of birth, parity, season of conception, medical risk factors, medical risk due to previous preterm birth and/or small for gestational age, baby's sex, and gestational length. Gestational length and baby's sex were not included in the models for SGA. Gestational length was not included in the models for preterm birth.

^bSee main text for explanation of models.

Table 4.

Adjusted^a association between PM_2.5 exposure during each trimester and birth outcomes. Effect estimates with 95%CI are reported as per inter-quartile range (2.41 μg/m³) increase in PM_2.5

Modelb		Trimester of exposure
Data Source	Observations	First	Second	Third
		OR of term LBW
Monitor	All	1.01 (0.99 to 1.03)	1.00 (0.98 to 1.02)	1.01 (0.99 to 1.03)
	Joint	1.01 (0.99 to 1.03)	1.00 (0.98 to 1.02)	1.01 (0.99 to 1.03)
Satellite (1)	All	1.04 (1.01 to 1.08)	0.99 (0.96 to 1.02)	1.02 (0.98 to 1.05)
	Joint	1.04 (1.00 to 1.08)	0.99 (0.95 to 1.02)	1.01 (0.98 to 1.05)
Satellite (2)	All	1.05 (1.01 to 1.08)	1.00 (0.97 to 1.03)	1.02 (0.99 to 1.06)
	Joint	1.04 (1.01 to 1.08)	1.00 (0.97 to 1.03)	1.02 (0.99 to 1.05)
		OR of SGA
Monitor	All	1.03 (1.01 to 1.04)	1.02 (1.01 to 1.02)	1.02 (1.01 to 1.02)
	Joint	1.03 (1.01 to 1.04)	1.02 (1.01 to 1.02)	1.02 (1.01 to 1.02)
Satellite (1)	All	1.02 (1.01 to 1.03)	1.02 (1.01 to 1.04)	1.02 (1.00 to 1.03)
	Joint	1.02 (1.00 to 1.03)	1.02 (1.00 to 1.04)	1.02 (1.01 to 1.04)
Satellite (2)	All	1.01 (1.01 to 1.02)	1.03 (1.01 to 1.04)	1.02 (1.01 to 1.04)
	Joint	1.01 (1.01 to 1.02)	1.02 (1.01 to 1.04)	1.02 (1.01 to 1.04)
		Change in birth weight (g)
Monitor	All	−6.2 (−7.9 to −4.6)	−3.5 (−4.6 to −2.3)	−2.4 (−3.6 to −1.3)
	Joint	−6.2 (−7.9 to −4.6)	−3.5 (−4.7 to −2.4)	−2.5 (−3.7 to −1.4)
Satellite (1)	All	−6.5 (−8.3 to −4.6)	−5.1 (−7.1 to −3.1)	−1.4 (−3.3 to 0.4)
	Joint	−6.3 (−8.1 to −4.4)	−5.5 (−7.5 to −3.4)	−1.9 (−3.8 to 0.0)
Satellite (2)	All	−4.1 (−5.2 to −2.9)	−6.8 (−8.6 to −4.9)	−3.1 (−4.8 to −1.4)
	Joint	−4.0 (−5.2 to −2.8)	−6.8 (−8.6 to −4.9)	−3.2 (−4.9 to −1.5)
		OR of preterm birth
Monitor	All	1.00 (0.99 to 1.02)	1.01 (1.00 to 1.02)	1.00 (0.99 to 1.01)
	Joint	1.00 (0.99 to 1.02)	1.01 (1.00 to 1.02)	1.00 (0.99 to 1.01)
Satellite (1)	All	1.00 (0.98 to 1.01)	0.99 (0.97 to 1.01)	0.99 (0.97 to 1.01)
	Joint	0.99 (0.98 to 1.01)	0.99 (0.97 to 1.01)	0.99 (0.97 to 1.01)
Satellite (2)	All	1.00 (0.99 to 1.01)	1.00 (0.98 to 1.02)	1.00 (0.98 to 1.01)
	Joint	1.00 (0.99 to 1.01)	1.00 (0.98 to 1.02)	0.99 (0.98 to 1.01)

^aAll models controlled for the following confounders mother's age, marital status, education, race/ethnicity, prenatal care, smoking, type of birth, parity, season of conception, medical risk factors, medical risk due to previous preterm birth and/or small for gestational age, baby's sex, and gestational length. Gestational length and baby's sex were not included in the models for SGA. Gestational length was not included in the models for preterm birth.

^bSee main text for explanation of models.

DISCUSSION

Air pollution has previously been associated with birth outcomes using various exposure methods.^6,8 We assessed whether associations between PM_2.5 and birth outcomes were affected by use of monitor or satellite exposure methods. As satellite data become more readily available, their application for exposure assessment will likely become more common, and studies li are needed to evaluate this alternative exposure method.

We are aware of three health studies that have used modeled satellite-based PM_2.5 estimates.^26,45,46 One of these studies, which looked at acute myocardial infarctions,⁴⁵ used the same exposure model was ours.⁴⁷ In the birth outcomes study,²⁵ the authors used land-use and traffic density data and satellite data to model PM_2.5 in Western Massachusetts. They used birth certificate data (2000-2008) and estimated risk of preterm birth and change in birth weight using inclusion criteria similar to ours: their results for change in birth weight are comparable to ours. In addition, we looked at a different and wider geographic region, compared satellite-and monitor-based exposures in our model, and looked at a wider range of birth outcomes (including small for gestational age and low birth weight). Unlike that earlier study, we did not find an association between PM_2.5 and preterm birth. However, our results are not directly comparable to that study due to differences in location, time period, modeling of satellite-based exposure estimates and model covariates. The previous study also did not compare results for risk estimates using monitors; that was done in another study,⁴⁶ although not for birth outcomes. Our estimates of risk (for term LBW and SGA) and change in birth weight are comparable to previous studies using monitor-based exposures.^{9,22,24,27,48}

The magnitude of the association between PM_2.5 and birth outcomes tended to be higher using the Satellite (2) exposure method (Table 3). This could relate to greater variability in PM_2.5 measurements based on monitors rather than satellite methods. Greater variability may attenuate associations towards the null due to exposure misclassification. In other words, areas with very high or low PM_2.5 estimates could influence the fitting of the model to the data and thus affect risk estimates, more than if PM_2.5 estimates were less spatially varied. Therefore, future research is needed on appropriate characterization of spatial heterogeneity for PM_2.5. This is especially true for differentiation of risk estimates based on various exposure methods (e.g., satellite, land-use regression models).

Satellite data can overcome some disadvantages of using monitor data for exposure assessment in health studies. Analysis of monitor data may lead to exposure misclassification and selection bias because sites are typically located for regulatory rather than research purposes. Monitors may not provide full coverage or represent population-based exposure.⁴⁹ Also, U.S. populations at varying distances from monitors differ in other ways.^50,51 For example, populations living in census tracts with a monitor tended to be characterized by having more non-Hispanic blacks, lower education, lower income, greater unemployment and higher poverty.⁵¹ Another limitation of monitor data is that monitors may be discontinued or temporarily out of operation (e.g., under maintenance). For many pollutants (e.g., PM_2.5, ozone), measures are not taken daily, which limits temporal coverage.

In contrast, satellite data provide near-complete spatial coverage of daily pollutant levels. Even so, uncertainty exists in the unprocessed satellite data and its calibration to observed data. Also, uncertainty may be introduced by the statistical procedure used to estimate pollutant levels for days when satellite data are missing due to cloud cover. Satellite data provide near-complete spatial coverage because exposure estimates are not possible to calculate for grid cells containing a substantial fraction of water (e.g., lakes and coastal regions).⁵² Also, coastal populations may have different demographic compositions (e.g., socioeconomic status) than populations living inland. A potential solution would be to use satellite data with finer spatial resolution. Another important issue for studies of air pollution and birth outcomes is the relevant gestational window of exposure. Although time-series analyses controlling for season of conception have been used to identify the relevant exposure window, the daily time scale of the satellite data is especially suited for such analyses.

Satellite methods provide a novel way to estimate health risks associated with air pollution in rural areas with few monitors. Also, it may be possible to investigate rural-urban differences in risk estimates. Rural populations differ from urban populations in terms of health and demographic and socioeconomic characteristics. U.S. studies have suggested that premature mortality, obesity, and cardiovascular disease were higher in rural areas than in urban or semi-urban areas, and that these disparities were related to urban-rural differences in socioeconomic and demographic characteristics.^53,54 The air pollutant mixture in rural areas may differ from urban areas (e.g., due to industry type, traffic patterns). Therefore, it is important to include rural populations in health effects studies. Traditional methods of assessing exposure focus more heavily on urban populations, excluding these rural populations. The use of satellite methods may allow study of broader scientific questions (i.e., whether effect estimates differ by rural vs. urban populations), in addition to increasing sample size through higher temporal and spatial coverage.

Several biologic mechanisms may be responsible for the association between air pollution exposure and adverse birth outcomes. Preterm birth may occur due to environmental disruptors of the endocrine systems that control parturition,⁵⁵ activation of molecular and cellular pathways involved in uterine contraction and quiescence through toxicant-induced inflammation response,¹² and interaction of external compounds with biochemical pathways responsible for the breakdown of the cervical matrix.⁵⁵ Birth weight-related outcomes (birth weight and term LBW and SGA) associated with PM_2.5 exposure may be due to mechanisms similar to those previously found for the effects of maternal smoking on fetal growth and development. Such mechanisms may include oxidative stress, vascular resistance in the placenta, and fetal exposure to toxic chemicals.^56,57 Even though exposure to maternal smoking may be associated with a greater decrease in birth weight than PM_2.5 exposure (150 to 300g⁵⁷ vs. 19g in our study under the Satellite (2) method), the following should be kept in mind. First, the population for maternal smoking exposure is much smaller than for PM_2.5 exposure. Second, it is much more difficult for individuals to avoid PM_2.5 exposure than maternal smoking. Other potential mechanisms involving particulate matter exposure may include, (1) mitochondrial dysfunction (in the placenta) in response to PM₁₀ exposure, which may affect nutrient transfer and growth of the placenta and, in turn, fetal growth and development,⁵⁸ and (2) the production of reactive oxygen species as a detoxification response to maternal smoking or exposure to air pollution further increasing the probability of DNA damage during fetal development and growth. Ongoing animal and human studies, including epigenetic studies looking at gene-environment interactions, continue to improve our limited understanding of these and other biologic mechanisms for adverse birth outcomes.^59-61

There were several limitations of our study. First, smoking habits, alcohol consumption, prenatal care and maternal risk data on birth certificates are less reliable than from other data sources such as questionnaires and cohort data.⁶² Despite these data reliability issues, birth certificate data are frequently used in health effects studies of air pollution and birth outcomes because they provided reliable estimates of birth weight and date of birth, both of which are essential for evaluating several birth outcomes.^8,9,22,26,63 Second, our data were limited in their spatial resolution. We excluded mothers living more than 50km from monitors, which may have introduced exposure misclassification due to spatial heterogeneity of pollutants (i.e., the 10x10km resolution is a necessary limitation of satellite data rather than being selected as an appropriate scale at which to predict PM_2.5 levels). Related to this, even though satellite data were calibrated to monitor data, grid cells where these monitors were located may be better predictors of PM_2.5 levels than grid cells without monitors (e.g., rural areas, suburbs). Lastly, neither method of exposure assessment – monitor- or satellite-based – captures individual-level exposures or identifies sources of air pollution. This drawback is common to most air pollution and health studies; possible solutions include using personal monitors for exposure assessment and simulation models (e.g., regional air quality modeling) or source apportionment.

In summary, our study compares associations between PM_2.5 exposure and birth outcomes, using a traditional data source for exposure assessment (land-based monitoring stations) and a new and emerging exposure method (satellite data that have been calibrated and modeled specifically for use in health effects studies). As satellite data continue to improve in their calibration, modeling and spatial resolution, they will become increasingly useful in health effects studies. Future studies should consider the spatial resolution of satellite data in the context of the specific pollutant under investigation (e.g., satellite data for some pollutants are available but at very large spatial resolution [100km or more]), and should compare associations based on multiple sources of data for exposure assessment so that our results are robust and more useful for policy makers in environmental risk assessments, as each exposure method has its own strengths and challenges.