Ambient air pollution and rate of spontaneous abortion

Abstract

Spontaneous abortion (SAB), defined as a pregnancy loss before 20 weeks of gestation, affects up to 30% of conceptions, yet few modifiable risk factors have been identified. We estimated the effect of ambient air pollution exposure on SAB incidence in Pregnancy Study Online (PRESTO), a preconception cohort study of North American couples who were trying to conceive. Participants completed questionnaires at baseline, every 8 weeks during preconception follow-up, and in early and late pregnancy. We analyzed data on 4,643 United States (U.S.) participants and 851 Canadian participants who enrolled during 2013-2019 and conceived during 12 months of follow-up. We used country-specific national spatiotemporal models to estimate concentrations of particulate matter <2.5 μm (PM_2.5), nitrogen dioxide (NO₂), and ozone (O₃) during the preconception and prenatal periods at each participant’s residential address. On follow-up and pregnancy questionnaires, participants reported information on pregnancy status, including SAB incidence and timing. We fit Cox proportional hazards regression models with gestational weeks as the time scale to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association of time-varying prenatal concentrations of PM_2.5, NO₂, and O₃ with rate of SAB, adjusting for individual- and neighborhood-level factors. Nineteen percent of pregnancies ended in SAB. Greater PM_2.5 concentrations were associated with a higher incidence of SAB in Canada, but not in the U.S. (HRs for a 5 μg/m³ increase=1.29, 95% CI: 0.99, 1.68 and 0.94, 95% CI: 0.83, 1.08, respectively). NO₂ and O₃ concentrations were not appreciably associated with SAB incidence. Results did not vary substantially by gestational weeks or season at risk. In summary, we found little evidence for an effect of residential ambient PM_2.5, NO₂, and O₃ concentrations on SAB incidence in the U.S., but a moderate positive association of PM_2.5 with SAB incidence in Canada.

INTRODUCTION

Spontaneous abortion (SAB), or miscarriage, is defined as the loss of an intrauterine pregnancy before viability (<20 weeks of gestation).¹ Approximately 15% of recognized pregnancies,¹ but up to 30% of pregnancies overall,² end in SAB, and SAB incidence appears to be increasing in several countries.^3–5 SAB has immediate physical and psychological consequences and is associated with increased risk of future obstetrical complications,⁶ cardiovascular disease,⁷ and long-term mental health disorders.⁸ Despite the high population burden of SAB, few modifiable risk factors have been identified, thus limiting opportunities for prevention.

A growing literature indicates that air pollution is a potential risk factor for SAB.⁹ Air pollution could influence SAB risk through multiple hypothesized mechanisms, including greater levels of oxidative stress¹⁰ and inflammation,¹¹ impairing placental function,¹² and altering epigenetic profiles.^13,14 In the epidemiologic literature, the effect of air pollution on SAB has been studied using time-series,^15–19 traditional case-control,^20–23 case-crossover,^24,25 self-matched case-control,^26–28 retrospective cohort,^28–33 and prospective cohort studies.^26,34,35 The specific pollutants most consistently associated with higher SAB risk were particulate matter <2.5 μm (PM_2.5)^{20,21,23–28,35} and nitrogen dioxide (NO₂).^{16,18–22,24,25,29,31,34} Several studies also found associations of particulate matter <10 μm (PM₁₀),^{16,19,24,26,29,32,33} sulfur dioxide (SO₂),^15,20,24 carbon monoxide (CO),²¹ and ozone (O₃)^{16,19,21,31,35} with higher SAB risk.

SAB is a challenging outcome to measure, as it is not directly observable and often occurs early in pregnancy, sometimes before pregnancy recognition.² Studies that enroll couples during the preconception period are more likely to capture early SAB than studies that recruit during pregnancy. Most of the existing preconception cohort studies investigating air pollution and SAB risk have been conducted among couples undergoing in vitro fertilization.^29,31–34 These studies have the unique advantage of being able to identify precise critical windows of exposure, but tend to have small sample sizes and may not be generalizable to populations trying to conceive spontaneously. A 2022 review emphasized the need for large detailed preconception cohort studies of air pollution and SAB.⁹

In the present study, we estimated the effect of residential ambient exposure to PM_2.5, NO₂, and O₃ on SAB incidence in a large prospective preconception cohort study of couples from across the United States (U.S.) and Canada.

METHODS

Study design.

We used data from Pregnancy Study Online (PRESTO), an internet-based prospective preconception cohort study.³⁶ Enrollment began in June 2013 and is ongoing. Eligible participants self-identified as female, were aged 21-45 years, resided in the U.S. or Canada, and were trying to conceive without use of fertility treatments at enrollment. Participants completed a baseline questionnaire on sociodemographic characteristics, lifestyle factors, and reproductive and medical histories. They then completed follow-up questionnaires every 8 weeks for up to 12 months to ascertain information on pregnancy status and to update exposure information. Participants who conceived completed additional questionnaires during early (~8 weeks’ gestation) and late pregnancy (~32 weeks’ gestation) and at 6 months postpartum. On all questionnaires, participants were asked to update their residential address if they had moved. In the current analysis, we used data from 4,643 U.S. and 851 Canadian participants who enrolled in the study between June 2013 and April 2019, who conceived during 12 months of follow-up, and whose residential addresses could be geocoded. The study protocol was approved by the institutional review board at the Boston University Medical Campus and all participants provided informed consent.

Outcome assessment.

We collected information on pregnancy and pregnancy outcomes on follow-up and pregnancy questionnaires. On follow-up questionnaires, participants reported the date of their last menstrual period (LMP), whether they were currently pregnant, and whether they had experienced a miscarriage (or chemical pregnancy), ectopic pregnancy, induced abortion, or blighted ovum since their previous questionnaire. We asked participants when and how often they used home pregnancy tests and the result of each test (positive or negative). Participants who were currently pregnant at the time they completed the follow-up questionnaire were directed to the early pregnancy questionnaire, on which they reported whether they had experienced any of the above pregnancy losses since their previous questionnaire, the date of their first positive pregnancy test, and how their pregnancy was confirmed (e.g., home pregnancy test, blood test in a doctor’s office). Losses that occurred after completion of the early pregnancy questionnaire were identified on the late pregnancy questionnaire.

Participants who reported a loss provided the date the pregnancy ended and how many weeks the pregnancy lasted. For participants with missing data on gestational age at loss, we estimated gestational age using the following formula: (pregnancy end date – (pregnancy due date – 280 days))/7. For participants who reported neither their gestational weeks at loss nor their pregnancy due date, we estimated gestational age as: (pregnancy end date – LMP date)/7.

We attempted to contact non-respondents via phone and/or email to determine their pregnancy status and ascertain any pregnancy losses since their date of last contact. We also linked to birth registries in selected states (CA, FL, MA, MI, NY, OH, PA, TX). Finally, we searched online for baby registries and birth announcements if other methods of outcome ascertainment were unsuccessful. Six percent of pregnancies and 1% of SABs were identified using these alternative methods.

Exposure assessment.

We geocoded all residential addresses reported during the study period using ArcGIS 10.3 (ESRI, Redlands, CA), as previously described.³⁷ We used national-level spatiotemporal models to predict residential ambient concentrations of PM_2.5, NO₂, and O₃. Due to differences in data availability and exposure assessment strategies between the U.S. and Canada, we fit separate models for each country.

In the U.S., we predicted two-week average air pollution concentrations at the residential address of each participant using a hierarchical spatiotemporal model. The model was originally developed for the Multi-Ethnic Study of Atherosclerosis and Air Pollution (MESA Air) to predict two-week average air pollution concentrations in six U.S. cities and has been expanded to a national scale.^38–45 The model operates at a continuous scale and characterizes the air pollution surface at precisely geocoded locations as a linear combination of temporal basis functions with spatiotemporal residuals and spatially-varying coefficients. It incorporates dependence on a large suite of geographic, meteorologic, and census covariates and applies smoothing through a spatial random effect by utilizing universal kriging. It also incorporates extensive regulatory and research ground-level data. The ground-level monitoring dataset comprised approximately 1,500 regulatory and 940 investigator-deployed non-regulatory monitors across the U.S.; these data are supplemented with satellite measures of tropospheric NO₂. We accounted for region-specific features and pollution processes in our nationwide estimates by dividing the country into climatic/topographic regions (nine for PM_2.5 and three for NO₂/O₃) and applied simple smoothing at regional boundaries to avoid artificial discontinuities. The model has been cross-validated using a range of techniques, including both spatial and temporal prediction contrasts.^39,40 The model output consisted of two-week average concentrations of PM_2.5, NO₂, and O₃ at the residential addresses of PRESTO participants during 2012-2019.

In Canada, we predicted monthly PM_2.5, NO₂, and O₃ concentrations at each participant’s residential address from 2012-2019. For PM_2.5 predictions, we used an existing model that combined Aerosol Optical Depth retrievals from satellites (NASA MODIS, MISR, and SeaWIFS) with the GEOS-Chem chemical transport model⁴⁶ and then used geographically-weighted regression to calibrate satellite estimates with measurements of regional ground-based PM_2.5⁴⁶ For NO₂ predictions, we used a nationwide land-use regression (LUR) model developed using National Air Pollution Surveillance (NAPS) monitoring data from 179 sites during 2014-2016 to capture spatial gradients across Canada. We applied geographically-varying monthly scaling factors derived from all NAPS monitoring data from 2012-2019 to the land-use regression model, as has been done in other contexts.⁴⁷ We derived scaling factors by dividing the monthly mean NO₂ concentration at a given monitor by the LUR model prediction at that monitor. We then used Bayesian kriging (applied in ArcGIS software) to create an interpolated surface based on the monitor scaling factors that were used to temporally adjust the NO₂ LUR concentrations. For O₃ predictions, we obtained the monthly O₃ surface for 2015, which was developed using hourly ground-level O₃ concentration estimates from the Global Environmental Multi-scale Modelling Air Quality and Chemistry model (10 km² resolution).⁴⁸ We based monthly O₃ estimates on the monthly average of the highest rolling 8-hour average concentration. We then developed monthly monitor-specific spatiotemporal scaling factors from the NAPS monitoring data, using the same methods described for the NO₂ model⁴⁹ to create monthly interpolated estimates of O₃ surfaces from 2012-2019.

Covariate assessment.

We collected individual-level data on sociodemographics, lifestyle factors, and reproductive and medical histories on the baseline and follow-up questionnaires. We obtained neighborhood-level covariate data from various sources, including the 2010 U.S. and Canadian Census (to estimate the population within 5,000 meters around the residence and census tract characteristics including median household income, % of census tract with ≤high school education, and % of census tract identifying as non-Hispanic white); the Landsat 8 satellite (to estimate residential green space using the Normalized Difference Vegetation Index; NDVI); and the Global Land Data Assimilation System (to obtain data on ambient temperature).

Statistical analysis.

For our primary analysis, we estimated time-varying air pollution concentrations during pregnancy as a predictor of loss during each gestational week. For the U.S. data, we updated the estimates every two weeks; in the Canadian data, we updated the estimates every month. For gestational weeks that fell entirely within a given interval, we assigned exposure as the average value for that interval. For gestational weeks that crossed two intervals, we calculated a weighted average based on the proportion of days in each interval. However, because the etiologically relevant window of exposure for SAB is unknown,⁹ we explored other critical windows in secondary analyses, including 1) time-varying prenatal average concentrations (i.e., average concentrations from pregnancy LMP date through the end of each gestational week), 2) average concentrations during the year before pregnancy, and 3) average concentrations during the two weeks before pregnancy.

Given our use of different exposure assessment methods in the U.S. and Canada, as well as differences in healthcare access and contextual characteristics, we conducted all analyses separately by country. We used life-table methods to calculate the proportion of participants who experienced SAB. We described population characteristics using means and percentages according to air pollution concentrations. We fit Cox proportional hazards models with gestational weeks as the time scale to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association between air pollution concentrations and rate of SAB. Participants entered the risk set at the gestational week when their pregnancy was identified (i.e., the week of their first positive pregnancy test or, if missing, 4 weeks’ gestation) and were censored if/when they experienced an induced abortion, an ectopic pregnancy, were lost to follow-up, or completed 20 weeks of gestation, whichever came first. We fit restricted cubic splines to examine the shape of the exposure-outcome association. We defined exposure cut points a priori as follows: <6, 6-<8, 8-<10, and ≥10 μg/m³ for PM_2.5; <4, 4-<8, 8-<12, and ≥12 ppb for NO₂; and <24, 24-<30, 30-<36, 36-<42, and ≥42 ppb for O₃ (combining the top two categories in the U.S. and the bottom two categories in Canada). We selected these cut points, rather than using country-specific quantiles, to facilitate comparison of results across countries, ensuring that our cut points both captured the variability in exposure and included a sufficient number of participants in each category across countries. We also provided estimates for a 5-unit increase in air pollution concentrations, assuming a linear association between air pollution and rate of SAB.

We used a directed acyclic graph to inform our selection of covariates. In final models, we adjusted for age at pregnancy (<25, 25-29, 30-34, ≥35 years), annual household income (<$50,000, $50,000-$99,999, $100,000-$149,999, ≥$150,000 US dollars), educational attainment (≤high school, some college, college degree, graduate school), self-reported race/ethnicity (conceptualized as a social construct and included as a proxy for exposure to racism at multiple levels; non-Hispanic Asian, non-Hispanic Black, non-Hispanic white, Hispanic, some other race), season of LMP date (March-May, June-August, September-November, December-February; for all analyses except for pollutant concentrations during the year before pregnancy), year of LMP date (2013-2014, 2015-2016, 2017-2019), smoking history (never, former, current occasional, current regular), parous (yes, no), population within 5,000 meters (continuous), census tract median household income (continuous), census tract % population with <high school degree (continuous), census tract % non-Hispanic white population (or census tract % “visible minority” in Canada; continuous), NDVI within 250 meters of residence (continuous), and ambient temperature (continuous). Income, educational attainment, and race/ethnicity were collected on the baseline questionnaire; all other adjustment variables were collected on the preconception questionnaire closest to conception (or, for census tract variables, from the residential address reported on the preconception questionnaire closest to conception). We adjusted for geographic region using the following divisions: in the U.S., Pacific [CA, OR, WA], Mountain [AZ, CO, ID, MT, NM, NV, UT, WY], West North Central [IA, KS, MN, MO, NE, ND, SD], West South Central [AR, LA, OK, TX], East North Central [IL, IN, MI, OH, WI], East South Central [AL, KY, MS, TN], South Atlantic [DE, DC, FL, GA, MD, NC, SC, VA, WV], Mid-Atlantic [NJ, NY, PA], New England [CT, MA, ME, NH, RI, VT]; and in Canada: British Columbia, Alberta, Manitoba/Saskatchewan/Northwest Territories/Yukon, Quebec, Ontario, eastern provinces. Spatial adjustment using more refined (i.e., state or province) or broader (i.e., 4 regions per country) categories gave similar results. We also adjusted models for co-pollutants (e.g., models for PM_2.5 are adjusted for NO₂ and O₃).

Given the strong seasonal patterns in air pollution³⁹ and SAB,⁵⁰ we conducted stratified analyses by season at risk (warm season [June-August] vs. cool season [December-February]). We also stratified by gestational weeks at risk (<8 vs. ≥8) because causes of SAB in early and late losses may differ; to conduct this analysis, we fit one model with follow-up time from weeks of positive pregnancy test through 7 weeks and a second model with follow-up time from 8 through 20 weeks. We stratified by U.S. Census Bureau Division to assess the potential for spatial heterogeneity in the association, which could be in part related to heterogeneity in PM_2.5 composition by region; we did not conduct a similar stratification within Canada due to smaller sample size. In sensitivity analyses, we restricted the analysis to participants residing in urban areas given that air pollution modeling is likely more accurate in urban areas.³⁹ Finally, we restricted to parous participants, who may more consistently recognize early pregnancy.

We used fully conditional specification methods to multiply impute missing covariate and outcome data. We imputed gestational weeks at loss for the 1% of participants who were missing this information. Missingness for covariates was low, ranging from 0% (e.g., age) to 3% (household income). We generated 20 imputed data sets⁵¹ and statistically combined estimates across data sets using Rubin’s rule.⁵²

RESULTS

Table 1 shows descriptive data on characteristics of the study population and air pollution concentrations, stratified by country. Correlations between pollutant concentrations across the four critical windows were generally moderate to strong: in the U.S., the range of correlation coefficients was 0.30-0.68 for PM_2.5, 0.73-0.92 for NO₂, and 0.21-0.75 for O₃ (Table S1). We observed similar correlations in Canada (Table S2). Pollutant concentrations measured during the same critical window were generally weakly correlated (Tables S1 and S2). For example, for time-varying concentrations during pregnancy in the U.S., the Spearman correlation coefficient between PM_2.5 and NO₂ was 0.29, between PM_2.5 and O₃ was −0.09, and between NO₂ and O₃ was −0.33. The distribution of individual- and neighborhood-level characteristics by air pollutant concentrations is shown in Tables S3–S5.

Table 1.

Characteristics of study population, Pregnancy Study Online, 2013-2019.

Characteristic	United States (n=4,643)		Canada (n=851)
	N (%)	Median (IQR)	N (%)	Median (IQR)
Age (years)		30.0 (27.0, 32.0)		29.0 (27.0, 32.0)
Educational attainment (years)
≤12	162 (3.5)		31 (3.6)
13-15	828 (17.8)		200 (23.5)
16	1566 (33.7)		365 (42.9)
≥17	2087 (45.0)		255 (30.0)
Household income (USD/year)
<$50,000	787 (17.0)		105 (12.3)
$50,000-$99,999	1747 (37.6)		369 (43.4)
$100,000-$149,999	1251 (26.9)		263 (30.9)
≥150,000	858 (18.5)		114 (13.4)
Race/ethnicity
Hispanic	314 (6.8)		18 (2.1)
Non-Hispanic Asian	123 (2.7)		5 (0.6)
Non-Hispanic Black	84 (1.8)		24 (2.8)
Non-Hispanic white	3942 (84.9)		773 (90.8)
Non-Hispanic some other race	180 (3.9)		31 (3.6)
Year of conception
2013-2014	1117 (24.1)		96 (11.3)
2015-2016	847 (18.2)		231 (27.1)
2017-2019	2679 (57.7)		524 (61.6)
Body mass index (kg/m2)		25.1 (22.1, 30.3)		25.0 (22.0, 29.1)
Current regular smoker	234 (5.0)		42 (4.9)
Gravid	2412 (52.0)		401 (47.1)
Parous	1611 (34.7)		217 (25.5)
Census tract median household income (dollarsa/year)		59,000 (44,700, 76,500)		76,400 (60,800, 96,300)
Census tract % with <high school education		9.0 (5.0, 14.0)		16.0 (12.0, 21.0)
Census tract % non-Hispanic white		81.0 (64.0, 91.0)		87.7 (71.7, 95.2)
Population within 5,000m		46,800 (13,900, 107,600)		70,600 (17,800, 146,900)
Seasonal maximum NDVI within 500m		0.54 (0.39, 0.68)		0.46 (0.34, 0.58)
Ambient temperature (°C)		13.6 (4.3, 21.2)		6.2 (−2.5, 13.9)
Ambient PM2.5 concentration (μg/m3)		7.1 (5.7, 8.8)		5.8 (4.5, 7.7)
Ambient NO2 concentration (ppb)		6.3 (4.1, 9.5)		6.5 (3.8, 9.8)
Ambient O3 concentration (ppb)		26.1 (21.0, 31.8)		33.0 (27.5, 39.9)

^aU.S. dollars and Canadian dollars.

IQR=interquartile range

The percentage of pregnancies ending in SAB was 19.3% in the U.S. and 18.8% in Canada. In both countries, the median gestational week at SAB was 6 (IQR: 5-9), with 63% of SAB occurring before 8 weeks. The median gestational week at pregnancy recognition was 4.0 (IQR: 3.7-4.6) and did not vary across categories of air pollution concentrations in either country (Table S6).

In the U.S., there was no appreciable association between time-varying concentrations of PM_2.5 during pregnancy and rate of SAB (Table 2). The restricted cubic spline analyses showed little evidence of non-linearity (Figure 1), and the HR for a 5 μg/m³ increase in concentration of PM_2.5 during pregnancy was 0.94 (95% CI: 0.83, 1.08; Figure 2). Associations were relatively consistent across strata of gestational weeks and season, as well as when restricted to participants residing in urban areas and parous participants (Figure 2). Associations did not differ appreciably by geographic region in the U.S., with the exception of the East South Central region, in which PM_2.5 concentrations during pregnancy were associated with lower rate of SAB, although results were imprecise (Figure S1). In Canada, a higher concentration of PM_2.5 during pregnancy was associated with increased rate of SAB, although results were imprecise (Table 2, Figure 1). The HR for a 5 μg/m³ increase in PM_2.5 concentration was 1.29 (95% CI: 0.99, 1.68; Figure 2). This association was stronger for later SAB (≥8 gestational weeks; HR=1.47, 95% CI: 0.97, 2.23) than for early SAB (HR=1.21, 95% CI: 0.85, 1.71), during the summer months (HR=1.79, 95% CI: 1.00, 3.20) relative to the winter months (HR=1.09, 95% CI: 0.43, 2.76), and among participants who resided in urban areas (HR=1.43, 95% CI: 1.10, 1.86; Figure 2). PM_2.5 concentrations during other critical windows of exposure were not appreciably associated with SAB, including average concentrations during pregnancy (Table S7, Figures S2–S3), concentrations during the year before pregnancy (Table S8, Figures S4–S5), or concentrations during the two weeks before pregnancy (Table S9, Figures S6–S7).

Table 2.

Association between prenatal air pollution concentrations and rate of spontaneous abortion, Pregnancy Study Online, 2013-2019.

	United States				Canada
Exposurea	No. of participants	No. of SAB	Unadjusted HR (95% CI)	Adjustedb HR (95% CI)	No. of participants	No. of SAB	Unadjusted HR (95% CI)	Adjustedb HR (95% CI)
PM2.5 (μg/m3)
<6	1416	270	Reference	Reference	380	68	Reference	Reference
6-<8	1589	306	1.03 (0.87, 1.22)	1.04 (0.87, 1.24)	184	27	0.81 (0.51, 1.28)	0.91 (0.55, 1.49)
8-<10	1043	184	0.95 (0.79, 1.16)	0.99 (0.80, 1.22)	92	17	1.05 (0.62, 1.79)	1.19 (0.67, 2.13)
≥10	595	103	0.92 (0.73, 1.15)	0.96 (0.74, 1.25)	65	14	1.35 (0.75, 2.40)	1.44 (0.77, 2.70)
NO2 (ppb)
<4	1093	226	Reference	Reference	235	43	Reference	Reference
4-<8	1926	373	0.91 (0.77, 1.07)	0.98 (0.82, 1.18)	316	56	0.94 (0.63, 1.41)	1.01 (0.59, 1.73)
8-<12	957	159	0.77 (0.63, 0.95)	0.87 (0.67, 1.12)	162	33	1.02 (0.64, 1.63)	1.04 (0.49, 2.21)
≥12	667	105	0.74 (0.58, 0.94)	0.88 (0.63, 1.25)	138	24	0.93 (0.57, 1.54)	1.02 (0.39, 2.69)
O3 (ppb)
<24	1902	357	Reference	Reference	107	17	Reference	Reference
24-<30	1252	226	0.94 (0.79, 1.11)	0.91 (0.75, 1.10)	212	48
30-<36	985	185	0.92 (0.77, 1.11)	0.91 (0.73, 1.14)	196	29	0.68 (0.44, 1.06)	0.53 (0.31, 0.91)
36-<42	414	75	0.98 (0.77, 1.23)	0.94 (0.71, 1.25)	182	32	0.76 (0.49, 1.18)	0.55 (0.30, 1.03)
≥42	90	20			154	30	0.93 (0.60, 1.44)	0.83 (0.42, 1.63)

^aTime-varying week-specific averages during pregnancy.

^bAdjusted for age, education, household income, race/ethnicity, season of conception, year of conception, geographic region, smoking, parity, ambient temperature, population within 5,000 meters, seasonal maximum NDVI within 500 meters, census tract median household income, census tract % with <high school education, census tract % non-Hispanic white, and other pollutant concentrations.

Figure 1.

Association between time-varying prenatal air pollution concentrations and rate of spontaneous abortion, fitted using restricted cubic splines. Figures on the left represent estimates for U.S. participants, and those on the right represent estimates for Canadian participants. The reference value is the lowest observed concentration and there are three knots in each spline (10^th, 50^th, and 90^th percentiles). Splines are adjusted for age, education, household income, race/ethnicity, season of conception, year of conception, geographic region, smoking, parity, ambient temperature, population within 5,000 meters, seasonal maximum NDVI, census tract median household income, % with <high school education, % non-Hispanic white, and other pollutant concentrations.

Figure 2.

Hazards ratios for a 5-unit increase in time-varying prenatal air pollution concentrations and rate of spontaneous abortion, PRESTO, 2013-19. The circles represent effect estimates for PM_2.5, the triangles represent estimates for NO₂, and the diamonds represent estimates for O₃. Estimates are adjusted for age, education, household income, race/ethnicity, season of conception, year of conception, geographic region, smoking, parity, ambient temperature, population within 5,000 meters, seasonal maximum NDVI within 500 meters, census tract median household income, census tract % with <high school education, census tract % non-Hispanic white, and other pollutant concentrations.

NO₂ concentrations during pregnancy were not strongly associated with rate of SAB in the U.S. or Canada (Table 2, Figures 1–2). In Canada, NO₂ concentrations during the winter were associated with higher SAB rate (HR for a 5-ppb increase=1.71, 95% CI: 0.81, 3.62), but this association was not observed in the U.S. (corresponding HR=0.93, 95% CI: 0.73, 1.19). Associations did not vary markedly across strata of gestational weeks or among participants residing in urban areas. NO₂ concentrations assessed during other critical windows were also not appreciably associated with SAB in either country (Tables S7–S9, Figures S2–S7).

Overall, O₃ concentrations during pregnancy were not appreciably related to rate of SAB (Table 2, Figures 1–2). However, in the summer months in Canada, O₃ concentrations during pregnancy were associated with higher rate of SAB (HR for a 5-ppb increase in O₃=1.43, 95% CI: 1.10, 1.86; Figure 2). We did not observe a similar association in the U.S. (corresponding HR=1.06, 95% CI: 0.92, 1.21). O₃ concentrations during other critical windows were not associated with SAB rates (Tables S7–S9, Figures S2–S7). There was some variability in O₃ estimates across strata of gestational weeks and season, but the pattern of variability was not consistent across countries.

DISCUSSION

In this large preconception cohort study, we found little consistent evidence for an effect of ambient concentrations of PM_2.5, NO₂, or O₃ on SAB incidence among participants residing in the U.S. We observed a moderate, but imprecise, association between PM_2.5 concentrations and higher SAB incidence in Canada. When we examined multiple critical windows of exposure including both preconception and prenatal periods, we found little evidence of a consistent association. We also found similar results across strata of season and gestational weeks at risk.

There was some evidence of an association between greater PM_2.5 concentrations during pregnancy and SAB incidence in Canada, but this association was absent among U.S. participants, regardless of geographic region. In addition to chance variation, there are several potential explanations for this inconsistency. Although PM_2.5 levels are similar in the two countries (and thus, differences in the magnitude of exposure are an unlikely explanation for the discrepant results), we used separate modeling approaches based on the best available data in each country. We also assessed pollutant concentrations on different time scales (every two weeks in the U.S. and monthly in Canada), which also could contribute to varying levels of misclassification. If non-differential misclassification of exposure differed across countries, this could lead to non-causal variation in measures of association. Likewise, social and environmental contexts in the U.S. and Canada are different. Therefore, confounding structures and the extent to which we controlled for confounding could vary, although the magnitude of confounding would need to be strong to explain the differences we observed. Finally, there is a potential causal explanation for the discrepant findings. PM_2.5 is a heterogeneous mixture; components of this mixture vary geographically,⁵³ and may have varying influences on health.⁵⁴ While there are many commonalities in particle composition between Canada and the U.S., including substantial contributions from sulfate, nitrate, and elemental and organic carbon, wildfires and residential biofuel combustion contribute proportionately more to PM_2.5 mass across Canada with larger contributions from road transportation and coal in the U.S.,⁵⁵ indicating some potential avenues for future study. In our analysis stratified by geographic region, which may in part reflect spatial variability in PM_2.5 composition, we did not find appreciable spatial heterogeneity. However, a more refined assessment of PM_2.5 components and their association with SAB incidence is warranted. The observed stronger results for the association of PM_2.5 with SAB in Canada during the summer months could also reflect seasonal differences in PM_2.5 composition; it could also reflect higher personal exposures during the summer, with people spending more time outdoors.

Our results generally disagree with the handful of studies that have enrolled participants preconceptionally and followed them throughout early pregnancy for more complete ascertainment of SAB. In a study of 4,581 patients undergoing in vitro fertilization (IVF) in Seoul, South Korea during 2006-2014, district-level concentrations of PM₁₀ and NO₂ during the time-period from embryo transfer to hCG test were associated with increased risk of chemical pregnancy loss (HRs for a one-IQR increase=1.17 [95% CI: 1.03, 1.33] and 1.18 [95% CI: 1.03, 1.34], respectively).²⁹ The corresponding HR for O₃ was 1.07 (95% CI: 0.90, 1.27), whereas PM_2.5 was not assessed in this study. In another study of 275 couples undergoing IVF in Massachusetts, U.S., average NO₂ concentrations were associated with increased risk of SAB 30 days after positive hCG test (HR for a one-IQR increase=1.34, 95% CI: 1.13, 1.58), but not earlier than 30 days (HR=0.83, 95% CI: 0.57, 1.20).³⁴ PM_2.5 and O₃ concentrations were not appreciably associated with pregnancy loss at any time point. In the only existing preconception cohort study of couples trying to conceive spontaneously, concentrations averaged across all gestational weeks of PM_2.5 (HR for a one-IQR increase=1.13, 95% CI: 1.02, 1.24) and O₃ (HR=1.12, 95% CI: 1.07, 1.17), but not NO₂ (HR=1.03, 95% CI: 0.98, 1.08) were weakly associated with increased rate of SAB among 343 participants from Michigan and Texas, U.S.³⁵

In contrast to these three existing preconception studies comprising study populations from relatively small geographic regions (e.g., one city or IVF center), PRESTO participants reside across the U.S. and Canada. While this geographic variability allowed us to enroll more people, it makes adjustment for spatial confounding more challenging. Unlike the other studies, however, we adjusted for both individual- and neighborhood-level confounders. This adjustment is important, given the strong variability in air pollution concentrations by neighborhood sociodemographic characteristics. Another challenge in comparing studies of different geographic areas, as described above, is the potentially heterogeneous composition of PM_2.5 in different places. In the preconception cohort study of couples from Michigan and Texas, U.S., the association between total PM_2.5 and SAB was driven by ammonium ions and sulfate compounds; there was little association of elemental carbon, nitrate compounds, or organic compounds with SAB risk.³⁵ Otherwise, there has been little exploration of the role of various PM_2.5 constituents in the etiology of SAB.

While the possibility of residual confounding due to unmeasured differences between participants exists in traditional cohort and case-control studies, studies that compare risk of SAB among the same individuals at different time points (e.g., case-crossover or self-matched case-control studies) inherently control for confounding by time invariant or slowly varying factors. In fact, studies of air pollution and SAB that used a case-crossover or self-matched case-control design tended to report stronger and more consistent associations than studies using a cohort design. For example, among 1,398 patients admitted for spontaneous pregnancy loss at an emergency department in Utah, U.S., average PM_2.5 concentrations during the prior three days were higher relative to referent periods (3-day windows in which the patient did not experience a loss).²⁵ A similar study conducted among 1,794 hospitalized SAB cases in Chongqing, China also reported higher concentrations of PM_2.5 during the several days before the hospitalization, relative to referent periods.²⁴ Among 19,309 participants in the prospective Nurses’ Health Study II, investigators found little association between PM_2.5 concentrations during the year before pregnancy and risk of SAB.⁵⁶ However, when they conducted a self-matched case-control study among the 3,585 participants who experienced at least one SAB and one live birth during follow-up, PM_2.5 concentrations were higher during the year before the SAB relative to the year before the live birth.⁵⁶ Given that published studies of air pollution and SAB tend to report stronger positive associations when confounding is minimized, residual confounding could potentially explain our relatively null results among U.S. participants.

In addition, many existing studies, particularly case-crossover studies, tended to examine air pollution concentrations during acute windows of exposure (e.g., three days before the loss). The temporal resolution of our air pollution models (every two weeks in the U.S. and monthly in Canada) was not ideal for assessing acute exposures on a daily scale. The data available for our exposure assessment were based on the built-in temporal granularity of the prediction model,^39,45,57 which was designed to leverage reference-grade regulatory monitors (typically available every 3-6 days) and investigator-maintained monitors. Therefore, we were unable to assess more temporally granular windows of association, which could also have contributed to our discrepant findings.

We did not find consistent evidence of an association across the U.S. and Canada between air pollution, at levels experienced by PRESTO participants, on SAB incidence. These results do not preclude an association at higher exposure levels. In a case-crossover study conducted from 2017-2019 in Chongqing, China, the median daily PM_2.5 concentration was 33.6 μg/m³, nearly five times what we estimated in the U.S., with a range of 8.0-177.9 μg/m³;²⁴ the study reported an 18% increase in odds of SAB for every 20 μg/m³ increase in ambient PM_2.5.

Other studies conducted in populations with high exposures have found an association with SAB, although some studies with lower exposures have as well.

A 2022 narrative review emphasized the need for large preconception cohort studies investigating the association between air pollution and SAB.⁹ Our study of nearly 5,500 pregnant individuals enrolled during the preconception period is the largest preconception cohort study on this topic. Because PRESTO participants are actively trying to conceive, most are using home pregnancy tests early (median weeks at first pregnancy test=4). This behavior facilitates early detection of pregnancy and more complete capture of early SAB. This is a key strength of our study design relative to studies that enroll participants later in pregnancy or identify SAB cases in clinical or hospital records. The prospective follow-up of participants throughout pregnancy also allowed us to assess residential mobility, a key advantage over many existing studies. We assessed air pollution using detailed spatiotemporal methods that estimate exposure at the residential addresses during several potential critical windows. Finally, we collected detailed time-varying information on individual- and neighborhood-level confounders.

In addition to the potential for unmeasured confounding described above, there are other limitations of our analysis. The most important limitations are outcome and exposure misclassification. SAB were assessed based on self-report. However, many SAB occur very early in gestation, oftentimes before pregnancy recognition;² therefore, we almost certainly missed some early losses. Unlike some other preconception cohort studies, we did not incorporate a standard pregnancy testing protocol in PRESTO, which could have introduced differential outcome misclassification. However, we found that participants tended to test early for pregnancy (oftentimes before a missed period) and that the timing of pregnancy recognition did not vary by exposure status. We also found little difference when restricting to parous participants, who may recognize subsequent pregnancies earlier than those who have never been pregnant.

Importantly, the unobserved nature of events during early pregnancy makes ascertaining the timing of the outcome, and therefore the relevant window of exposure, challenging. There is a lag between when a pregnancy is no longer viable and when the loss is recognized by the pregnant individual.⁵⁸ In the absence of ultrasound data on early fetal development, we relied on self-reported date of SAB and/or weeks’ gestation at SAB. Therefore, we may have assessed air pollution during the incorrect critical window, which is an important source of exposure misclassification. Finally, although we used validated spatiotemporal models to estimate exposure, these models estimate ambient residential exposure only, without accounting for indoor air pollution or time-activity patterns.

In summary, we found little evidence for an effect of PM_2.5, NO₂, and O₃ on SAB incidence in the U.S., but a moderate positive association between PM_2.5 concentrations and SAB incidence in Canada, although there were few cases with high exposure levels. Nonetheless, given the extensive body of literature documenting an association between prenatal air pollution exposure and risk of other adverse birth outcomes and pregnancy complications,⁵⁹ the protection of individuals from air pollution during preconception and pregnancy remains an important measure to safeguard maternal and fetal health.

HIGHLIGHTS.

• In this preconception cohort study, we found a positive association between ambient concentrations of PM_2.5 and rate of spontaneous abortion in Canada, but not the United States.

• NO₂ and O₃ concentrations were not appreciably associated with rate of SAB.

• Results were relatively consistent across season and weeks of gestation.

• Given the known adverse impacts of air pollution on pregnancy health, the protection of individuals from air pollution during preconception and pregnancy remains an important measure to safeguard maternal and fetal health.