Content uploaded by Yanxiao Jiang
Author content
All content in this area was uploaded by Yanxiao Jiang on Sep 25, 2019
Content may be subject to copyright.
Articles
https://doi.org/10.1038/s41893-019-0387-y
1State Key Laboratory of Remote Sensing Science, Faculty of Geographical Science, Beijing Normal University, Beijing, China. 2Beijing Obstetrics and
Gynecology Hospital, Capital Medical University, Beijing, China. 3Laboratory for Climate and Ocean-Atmosphere Studies, Department of Atmospheric and
Oceanic Sciences, School of Physics, Peking University, Beijing, China. 4State Key Laboratory of Resources and Environment Information System, Institute
of Geographical Science and Natural Resources, Chinese Academy of Sciences, Beijing, China. 5School of Surveying and Geo-informatics, Tongji University,
Shanghai, China. 6These authors contributed equally: Liqiang Zhang, Weiwei Liu, Kun Hou. *e-mail: zhanglq@bnu.edu.cn; lffhornet01@163.com;
linjt@pku.edu.cn; zhouch@lreis.ac.cn
Poor air quality is a leading cause of global disease burden1,2.
Considerable evidence has consistently indicated that mater-
nal exposure to air pollution contributes to increased risks of
adverse birth outcomes such as low birthweight3, preterm birth3,4,
gestational hypertension and preeclampsia5,6, and may also affect
maternal health during pregnancy and over the course of a woman’s
life. Approximately 28% of pregnant women risk a loss of pregnancy
in developed countries7. Missed abortion in the first trimester
(within 14 weeks of gestation; MAFT), which is characterized by
the arrest of embryonic or fetal development8, is a common com-
plication of pregnancy. MAFT may occur in up to 15% of all clini-
cally recognized pregnancies, especially in developing countries9–11.
Determining whether or not the risk of MAFT responds to air qual-
ity conditions is important, as a pregnancy loss is devastating for the
expectant parents12.
Few quantitative studies exist that explore how maternal air
pollution exposure affects the MAFT risk. Several studies13,14
have been carried out in high-income countries with relatively
good air quality. A study15 conducted in Tianjin, China using the
official air pollution monitoring data from 2001 to 2007 found
that there were possible adverse impacts of air pollution on
pregnancy outcomes; however, data during this period are unre-
liable16. In contrast, data collected since 2013, when the govern-
ment enforced strict regulations to ensure air quality, are more
reliable. Quantifying the relationship between maternal exposure
to air pollutants and MAFT requires detailed, difficult-to-obtain
information on personal exposures and confounders for a wide
range of pollution exposures.
We investigated how the MAFT risk varies with the level of
maternal ambient air pollution exposure using air pollution mea-
surements and clinical data from pregnant women living in Beijing,
China. Air pollutants considered for the study included particu-
late matter (PM) with diameter below 2.5 μm (PM2.5), sulfur diox-
ide (SO2), ozone (O3) and carbon monoxide (CO). Given its size,
Beijing has diverse terrains (Supplementary Fig. 1) and a consid-
erable range of air quality conditions across space and time17. The
spatial distribution of daily mean concentrations of PM2.5 in Beijing,
averaged from 2008 to 2017, indicates that the temporally average
daily (Supplementary Fig. 2a) and maximum daily (Supplementary
Fig. 2b) concentrations exceeded 100.0 μg m−3 at several locations,
although the minimum concentrations (Supplementary Fig. 2c)
were below 6.0 μg m−3 in many places. Daily mean concentrations
of ambient SO2, O3 and CO in Beijing also showed large spatiotem-
poral variabilities (Supplementary Figs. 3–5). In addition, while
Beijing is a well-developed region, it still has large rural areas with
relatively low household income, such as those in Fangshan District
(Supplementary Fig. 1). Considering the above factors, the expo-
sure–response relationship between air pollution and MAFT risk
derived from Beijing may be representative of the general situation
in China.
We collected the clinical records of 255,668 pregnant women in
Beijing from 2009 to 2017. The dataset contained information on
maternal education level, occupation, residence and working places,
and last menstrual date. Following earlier work18, we computed the
air pollutant exposure level of each pregnant woman on the basis
of measurements at the nearest air monitoring stations from her
Air pollution-induced missed abortion risk
for pregnancies
Liqiang Zhang 1,6*, Weiwei Liu 2,6*, Kun Hou1,6, Jintai Lin 3*, Chenghu Zhou 4*, Xiaohua Tong5,
Ziye Wang1, Yuebin Wang 1, Yanxiao Jiang1, Ziwei Wang 3, Yibo Zheng1, Yonglian Lan2, Suhong Liu1,
Ruijing Ni 3, Mengyao Liu3 and Panpan Zhu1
Fetus death risk reduction is included in the United Nations Sustainable Development Goals. However, little is known about
how missed abortion in the first trimester (MAFT) is related to maternal air pollution exposure. We quantify the link between
air pollution exposure and MAFT in Beijing, China, a region with severe MAFT and air quality problems. We analyse the records
of 255,668 pregnant women from 2009 to 2017 and contrast them with maternal exposure to air pollutants (particulate mat-
ter PM2.5, SO2, O3 and CO). We adjust for confounding factors such as sociodemographic characteristics, spatial autocorre-
lation and ambient temperature. We find that, for all four pollutants, an increased risk of MAFT is associated with rises in
pollutant concentrations and the adjusted odds ratios (ORs) of these associations increase with higher concentrations. For
example, the adjusted OR of MAFT risk for a 10.0 μg m−3 increase in SO2 exposure is between 1.29 and 1.41 at concentrations of
7.1–19.5 μg m−3; it drops to 1.17 below this range and rises to 1.52 above it at higher SO2 concentrations. This means that the risk
increase is not linear but becomes more severe the higher the pollutant concentration. The findings provide evidence linking
fetus disease burden and maternal air pollution exposure.
NATURE SUSTAINABILITY | www.nature.com/natsustain
Articles Nature SuStaiNability
residential and working places (Methods). This choice, as opposed
to previous studies that only considered the distance to residence
place19–21, was preferred because most pregnant women in the first
trimester still went to work.
We used a logistic regression model and a restricted cubic spline
model to explore the quantitative relationship between the MAFT
risk and maternal exposure to air pollutants, including PM2.5, SO2,
O3 and CO, among the 255,668 pregnant women. Potential con-
founders, including maternal age at pregnancy, occupation, spatial
dependence and ambient air temperature were controlled in the
models. We performed several robustness analyses: (1) we tested
whether the exposure–response association was characterized by
different lag periods; (2) we assessed the exposure–response rela-
tionship in individual concentration ranges for each air pollutant;
and (3) we conducted a restricted cubic spline analysis to character-
ize the morphology of the nonlinear association between air pollut-
ant exposure and the MAFT risk. Model results are reported as odds
ratios (ORs)22 and their 95% confidence intervals (CIs).
Results
We grouped pregnant women by age at conception (five groups),
occupation (two groups) and air temperature (four groups). Among
the participating pregnant women, 17,497 (6.8%) experienced
MAFT. We took the Bayes factor as a measure of evidence for the
association between MAFT in different subgroups and air pollution.
As a summary measure, the Bayes factor gave an alternative to the
P-value for the ranking of associations or for the flagging of associa-
tions as significant23. The Bayes factor BF10 > 30 represents strong
evidence for the associations between the MAFT occurring in the
subgroups and air pollution24,25 (Supplementary Table 1). Women
older than 39 years at conception or female farmers and blue-collar
workers had higher percentages of MAFT than their counterparts
did. In all groups, maternal exposure to each air pollutant was asso-
ciated with the risk of MAFT.
Associations between MAFT risk and air pollution exposure. We
used a logistic regression model to calculate the ORs and 95% CIs
for the association between MAFT risk and exposure to each pollut-
ant. We adjusted for potential confounders including maternal age,
occupation, spatial dependence and ambient temperature. As we
estimated the association between PM2.5 and MAFT risk, we did not
control for other air pollutants (SO2, O3 and CO) in the model. The
main reason is that the period of the available PM2.5 data (from 2008
to 2017) was different from those of other pollutants (from June
2014 to December 2017). Moreover, the correlation between PM2.5
and other two pollutants (SO2 and O3) is not high (Supplementary
Table 2), thus controlling for these two pollutants had a little effect
(Supplementary Table 3). As we estimated the association between
O3 and the MAFT risk, we controlled for CO and SO2 but not PM2.5,
given the strong correlation between PM2.5 and CO (Supplementary
Tables 2 and 4). Similarly, as we estimated the association between
SO2 (CO) and the MAFT risk, we controlled for CO (SO2) and O3
but not PM2.5.
We investigated the correlation between maternal exposures
in different time periods and the incidence of MAFT, to test the
effect of time lag between pollution exposure and MAFT. Seven
1.7
ab
cd
1.6
1.5
1.4
1.3
1.2
OR
1.1
1.0
0.9
1.7
1.6
1.5
1.4
1.3
1.2
OR
1.1
1.0
0.9
1.7
1.6
1.5
1.4
1.3
1.2
OR
1.1
1.0
0.9
1.7
1.6
1.5
1.4
1.3
1.2
OR
1.1
1.0
0.9
<63.2
<27.3 27.3–46.2 46.3–74.4 >74.4 <0.9 0.9–1.1 1.2–1.9 >1.9
63.2–71.4 71.5–93.3 93.4–130.2 >130.2
PM
2.5
(µg m
–3
)
<7.1 7.1–11.4 11.5–19.5 >19.5
SO
2
(µg m
–3
)
O
3
(µg m
–3
)C
O
(mg m
–3
)
Fig. 1 | The ORs and 95% CIs of MAFT associated with maternal exposure to each pollutant in Phase 4. a–d, The OR of MAFT with respect to PM2.5 (a),
SO2 (b), O3 (c) and CO (d) exposure. Confounders were controlled here.
NATURE SUSTAINABILITY | www.nature.com/natsustain
Articles
Nature SuStaiNability
time windows of exposures (Phases 1–7; Methods) were consid-
ered. The risk of MAFT was associated with PM2.5 exposure in all
phases (Supplementary Fig. 6a). A 10.0 μg m−3 increase in the tem-
porally averaged PM2.5 concentration in Phase 4 (the period from
90 d before the first day of the last menstrual period (LMP) to the
last exposure time (LET) of the pregnant women with MAFT; see
Methods for definition of LET) was associated with the greatest
risk of MAFT, when compared to the ORs from other phases. The
MAFT risk also changed in different time windows for SO2, O3 and
CO exposures and each had a peak risk with respect to Phase 4
(Supplementary Fig. 6). In the following, exposure to each air pol-
lutant averaged in Phase 4 was used for further analysis.
Due to lag in air pollutant exposure, PM2.5 data for the years
2008–2017 (Methods) were combined with the clinical data
from 2009–2017 for analysis. Maternal PM2.5 exposure in Phase
4 was categorized into five concentration ranges separated by the
25th, 50th, 75th and 95th percentiles of the PM2.5 concentrations:
<63.2 μg m−3, 63.2–71.4 μg m−3, 71.5–93.3 μg m−3, 93.4–130.2 μg m−3
and >130.2 μg m−3. An increase in ambient PM2.5 concentrations
was significantly associated with an increased MAFT risk (Fig. 1).
For a 10.0 μg m−3 increase in PM2.5 concentrations, the OR of MAFT
(after adjusting for the confounders) was 1.08 (95% CI, 0.98–1.18)
for PM2.5 concentrations <63.2 μg m−3, 1.13 (95% CI, 1.03–1.23)
for PM2.5 of 63.2–71.4 μg m−3, 1.28 (95% CI, 1.14–1.42) for PM2.5
of 71.5–93.3 μg m−3, 1.39 (95% CI, 1.23–1.55) for PM2.5 of 93.4–
130.2 μg m−3 and 1.51 (95% CI, 1.33–1.69) for PM2.5 >130.2 μg m−3.
The increase of the OR with increasing PM2.5 concentrations was
evident. Results based on PM2.5 data over the period June 2014–
December 2017 were similar (Supplementary Table 5).
SO2, O3 and CO data were available from 2014 to 2017
(Methods), thus we assessed the correlation between MAFT and
each of SO2, O3 and CO over the same period. SO2 concentration
exposures ranged from 2.6 μg m−3 to 44.0 μg m−3. Maternal SO2
exposure in Phase 4 was categorized into four concentration ranges
separated by the 25th, 50th and 75th percentiles of the SO2 con-
centrations. For a 10.0 μg m−3 increase in SO2 exposure, the ORs
for SO2 exposure in Phase 4 were 1.17 (95% CI, 1.10–1.22) for SO2
concentrations <7.1 μg m−3, increasing to 1.29 (1.22–1.36) for SO2
of 7.1–11.4 μg m−3, 1.41 (1.33–1.49) for SO2 of 11.5–19.5 μg m−3 and
1.52 (1.44–1.60) for SO2 concentrations >19.5 μg m−3 (Fig. 1b).
Maternal O3 exposure in Phase 4 was categorized into four con-
centration ranges separated by the 25th, 50th and 75th percentiles of
the O3 concentrations. For a 10.0 μg m−3 increase in O3 exposure, the
ORs for O3 exposure in Phase 4 were 1.07 (95% CI, 1.00–1.14) for O3
concentrations <27.3 μg m−3, increasing to 1.09 (1.03–1.15) for O3
of 27.3–46.2 μg m−3, 1.14 (1.06–1.22) for O3 of 46.3–74.4 μg m−3 and
1.23 (1.15–1.31) for O3 >74.4 μg m−3 (Fig. 1c).
Maternal CO exposure in Phase 4 was categorized into four con-
centration ranges separated by the 25th, 50th and 75th percentiles of
the CO concentrations. For a 1.0 mg m−3 increase in CO exposure,
the ORs for CO exposure in Phase 4 were 1.05 (95% CI, 1.03–1.07)
for CO concentrations <0.9 mg m−3, increasing to 1.08 (1.05–1.11)
for CO of 0.9–1.1 mg m−3, 1.13 (1.09–1.17) for CO of 1.2–1.9 mg m−3
and 1.17 (1.12–1.22) for CO >1.9 mg m−3 (Fig. 1d).
1.4
ab
1.3
1.2
OR
1.1
1.0
0.9
1.4
cd
1.3
1.2
OR
1.1
1.0
0.9
1.4
1.3
1.2
OR
1.1
1.0
0.9
1.4
1.3
1.2
OR
1.1
1.0
0.9
2014 2015 2016 2017
Year
2015 2016 2017
Year
2015 2016 2017
Year
2015 2016 2017
Year
Fig. 2 | The ORs and 95% CIs of the MAFT risks associated with maternal exposure to different average annual air pollutant concentrations. a–d,
Association of the MAFT with a 10.0 μg m−3 increase in PM2.5 exposure from 2014 to 2017 (a), a 10.0 μg m−3 increase in SO2 exposure from 2015 to 2017
(b), a 10.0 μg m−3 increase in O3 exposure from 2015 to 2017 (c) and a 1.0 mg m−3 increase in CO exposure from 2015 to 2017 (d). Note that since the
datasets of SO2, O3 and CO before June 2014 contained many missing values, for these pollutants we only used the pollutant data from June 2014 to
December 2017 and thus the MAFT data from 2015 to 2017.
NATURE SUSTAINABILITY | www.nature.com/natsustain
Articles Nature SuStaiNability
We further combined a restricted cubic spline model and a
logistic regression model to construct the OR curve for MAFT
and maternal pollution exposure for each pollutant, after other
factors were controlled (Supplementary Fig. 7). The relationship
between pollution exposure and the risk of MAFT was enhanced
with increased pollutant concentrations, consistent with the above
(logistic model based) findings.
Since 2013, the Chinese government has issued new rules to
reduce atmospheric pollution. Air pollutant concentrations have
substantially decreased since 2014 (Supplementary Fig. 8)16. The
MAFT risk has also decreased since 2013 (Fig. 2), which further
suggests the strong quantitative link between maternal air pollution
exposure and MAFT risk.
Maternal characteristics, pollution exposure and risk of MAFT.
Pregnant women with different sociodemographic status might be
exposed to different air pollution levels and, therefore, could be
subject to different MAFT risks. Older pregnant women (>39 years
old), female farmers and blue-collar workers and those conceiving
in low temperature (<5 °C) had higher ORs for the risk of MAFT
associated with air pollution exposure when compared to their
counterparts (for example, pregnant women aged 25–39 years)
(Supplementary Fig. 9). In China, farmers and blue-collar work-
ers usually had a low socioeconomic status and engaged in outdoor
work21. Female farmers or female blue-collar workers were more
exposed to ambient air pollution and thus subjected to a higher OR
and a higher MAFT risk than office workers (as discussed above, the
OR increases with increasing pollutant concentrations). This result
was consistent with the previous finding suggesting that wealthier
households were better able to avoid the adverse health impacts of
hazardous environmental exposures26–28.
Discussion
On the basis of a large record of maternal clinical data and a broad
range of air pollution concentrations, this study has demonstrated
a quantitative association between ambient air pollution exposure
and risk of MAFT. Previous studies have also indicated that mater-
nal long-term exposure to air pollution may mean a higher likeli-
hood of abortion/miscarriage, stillbirth and birth defects7,29,30.
We investigated several possible causal mechanisms to explain
this linkage. Maternal long-term exposure to PM2.5 allows the pol-
lutant to cross the maternal–fetal blood barrier and ultimately
perturb fetal growth and development7,31. Pollutants entering the
bloodstream of a fetus might interact with its tissue components to
produce pathological effects17, leading to irreversible damage to the
dividing cells of the fetus and triggering hypoxic harm or immu-
nomediated injury during critical periods of development32,33. Air
pollution-induced placental epigenetic alterations were observed
during all trimesters of pregnancy34. This suggests that mater-
nal exposure to air pollution might damage placental functions.
Previous studies have shown that perturbations in the maternal
environment could be transmitted to the fetus by changes in placen-
tal functions31 and that ambient environmental insults on placenta
had negative effects on the developing fetus35.
In addition, poor air quality was significantly associated with
the amount of polycyclic aromatic hydrocarbons (which can be
absorbed by or adhere to PM2.5) bound to DNA in both maternal
and fetal cord white blood cells36. Mothers exposed to air pollution
were more likely to have chromosomal abnormalities37. Therefore,
maternal long-term exposure to air pollution increased the chances
of abortion/miscarriage, stillbirth and birth defects.
Furthermore, toxicants could pass through the placenta and
attack the developing fetus by potentially inducing alterations in
immune competence38. CO might interfere with metabolic and
transport function of the placenta and, after crossing the placen-
tal barrier, collect at higher concentrations in the fetus than in the
mother39. Moreover, ambient CO was associated with carboxy-
haemoglobin (COHb) and nucleated red blood cells40. Redundant
COHb in mothers might cause fetal hypoxia, which could lead to
fetal death41.
Since there was strong collinearity between PM2.5 and CO
(Supplementary Tables 2 and 4), we could not separate the indi-
vidual effects of these two pollutants. Although we were able to
adjust for many known risk factors for MAFT that would confound
the association, residual confounding cannot be ruled out, as it is
possible that other factors we were unable to control for, such as
traffic-related noise, may be associated with pregnancy outcome.
The impacts of indoor air pollution on MAFT were not studied due
to the lack of indoor pollution data, although indoor and ambient
pollution (type and severity) are highly correlated.
Associating air pollution with the spatial–temporal variability in
MAFT enhances scientific and policy understanding of pregnant
women’s health in developing countries42. Our findings uncovered
potential opportunities to prevent or reduce harmful pregnancy
outcomes by proactive measures before pregnancy. Meanwhile, our
study helped us understand the relationship between air pollution
exposure and a spectrum of reproductive outcomes.
Pregnant women or those who want to become pregnant, must
protect themselves from air pollution exposure not only for their
own health but also for the health of their fetuses. China is an aging
society and our study provides an additional motivation for the
country to reduce ambient air pollution for the sake of enhanc-
ing the birth rate. Although ambient air pollution has reduced in
China in recent years16, pollution levels are still high and must
reduce further for many reasons, including reducing MAFT. Future
work should explore the human health benefits from air pollution
mitigation through modelling a wide range of environmental con-
ditions using more data sources including land-use and land-cover
change data43.
Methods
Maternal clinical dataset. We collected, processed and selected maternal clinical
data as explained below.
Collection of clinical data. We collected clinical data of 260,231 pregnant women in
Beijing, China. Pregnancy outcomes were classified according to the International
Classification of Diseases, 10th revision44.
Validation of the dataset. Here we presented an independent validation of the
dataset. Specifically, a midwife familiar with clinical coding techniques randomly
selected 926 maternity case notes and then compared these case notes with those
recorded in the dataset. The comparison results were re-checked by an obstetrician
who had undergone training in clinical coding.
The 926 maternity case notes were randomly sampled at the maternity hospital
wards (H-1, H-2, H-3 and H-4). Samples of 235, 247, 203 and 241 singleton
deliveries were selected from these wards, respectively.
Two pregnancy outcome metrics were adopted to measure the quality of
the data. The first metric was percentage agreement between the contents of
ten selected fields recorded in the dataset and their counterpart data extracted
from the maternity notes. The second metric uses Cohen’s Kappa to assess the
consistency between the two datasets.
For the 926 cases examined, the contents of the ten fields were consistent
between the data and case notes, with the percentage agreement exceeding 95%
for all fields and hospitals (Supplementary Table 4). The values of Kappa were also
much larger than 0.6, indicating high agreement45. The validation result provided
confidence in the reliability of the dataset.
Data screening for MAFT based on gestational age. We only considered fetal loss
occurring at <98 d (14 weeks) of gestation in this study.
Gestational age was computed as the number of days between the date of the
LMP and the date of fetal mortality. It was difficult to determine gestational age
precisely since fetal mortality might happen weeks before it was found. In this
study, we estimated the gestational age as follows.
We assumed the date of fetal mortality to be 2 weeks before the abortion date,
when the number of weeks from the LMP to the abortion date was >7 weeks.
Otherwise, we assumed the date of fetal mortality to be the abortion date, due
to fetal heart not being monitored in the first 5 weeks of the first trimester. The
abortion procedure was done on the date a fetus was determined dead for almost
NATURE SUSTAINABILITY | www.nature.com/natsustain
Articles
Nature SuStaiNability
all cases, except in rare occasions in which clinical complications and other issues
led to a 1–2 d delay in the abortion procedure. Ultrasound was used to determine
whether a fetus was alive or dead.
Pregnant women in Beijing usually had one prenatal care visit every month
in the first trimester. They were checked by ultrasound when maternal or fetus
anomalies were suspected based on regular, non-ultrasound examinations. In
addition, vaginal bleeding often occurred after fetuses died. When pregnant
women found vaginal bleeding, they usually went to hospital and were checked
by ultrasound to determine whether their fetuses had died. The above two aspects
helped to detect fetal death as early as possible and decreased the bias in the
gestational age estimation. Moreover, since we assessed the relationship between air
pollution and MAFT using clinical data over 9 years (from 2009 to 2017), the effect
of errors in estimated gestational ages could be reduced.
Very early losses like biochemical pregnancy and stillbirth were beyond the
scope of our research, thus pregnant women whose fetal gestational ages were less
than 40 d or were more than 98 d from the LMP were excluded.
Other data screening. Pregnant women were asked to provide the address they
had lived at for the longest in the half-year before and during pregnancy. Women
were excluded from the current study if there were no records of their addresses.
Those with irregular menstruation cycles or with a history of miscarriage were
also excluded; the dataset recorded the miscarriage dates. We did not account for
maternal smoking status, since most Chinese women do not smoke, especially
before and during pregnancy.
Basic statistics of finally selected data. After all the aforementioned exclusions, data
for a total of 255,668 women in Beijing from 2009 to 2017 were valid for analysis
(Supplementary Table 5).
Air pollution data. We used air pollution data from 34 air quality monitoring sites
(see their spatial distribution in Supplementary Fig. 10). Hourly measurements
were established in 2013 and maintained by the Ministry of Ecology and
Environment (MEE, formally the Ministry of Environmental Protection). Air
pollution measurements at the MEE sites followed the official measurement
standards46. For SO2, O3 and CO measurements, there were few valid data before
June 2014, thus for these pollutants we only used data from June 2014 to December
2017 in this study. Daily mean pollutant concentrations were derived from hourly
data of each day and the average of daily mean values during the exposure phase
was used as the exposed concentration. Measured pollutants included PM10, PM2.5,
SO2, CO, NO2 and O3. The NO2 measurements were not used here due to concerns
regarding contamination by other nitrogen pollutants47. We excluded the PM10
data, which contained many missing values.
To extend the PM2.5 data time period previous to June 2014, we made use of
the long-term measurements taken by the US Embassy (http://www.stateair.net/
web/historical/1/1.html; accessed 12 July 2018). PM2.5 measurements at the US
Embassy site used the beta-attenuation instrument. The US Embassy data were
shown to be consistent with the MEE measurements16. For data at each MEE site
from June 2014 to December 2017, we established a linear relationship with the US
Embassy data on an hourly basis. Measurements at each MEE site were consistent
with those at the US Embassy site: R2 ranged from 0.47 to 0.93 with an average
of 0.70 (Supplementary Table 6). This consistency allowed us to apply the linear
relationship to prior periods when there were no MEE measurements, as done
here. Our further test using only data from June 2014 to 2017 suggested a similar
association between PM2.5 exposure and MAFT (Supplementary Table 7), which
supported our use of US Embassy data for earlier times.
Meteorological measurement data. Three-hourly data for air temperature at 2 m
above ground were taken from the meteorological measurement station near the
southwestern Fourth Ring Road of Beijing (Supplementary Fig. 10). Data at this
station were reported to the World Meteorological Organization and maintained
at the US National Oceanic and Atmospheric Administration National Centers for
Environment Information (https://www.ncdc.noaa.gov/isd/data-access; accessed 24
October 2018). Daily mean air temperature was derived from 3-hourly data.
Interpolation of missing air pollutants and air temperature data. There was
large diurnal variation in each air pollutant and meteorological variable. To fill in
the missing meteorological or air pollution data to accurately quantify the exposure
level of each pregnant woman, we interpolated the missing values using the same
interpolation methods as in our previous study48.
Maternal exposure to air pollutants. Maternal residential and working district
addresses before or at the period of conception and air quality monitoring stations
were geocoded to obtain their latitudes and longitudes. Most of the participating
pregnant women did not change their residences and working places before the
first trimester of pregnancy. We estimated maternal exposure to air pollution by
attributing representative concentrations provided by the air quality monitoring
stations closest to the maternal residence and working place after geolocalization.
Approximately 86% of the pregnant women provided the working addresses.
Since most pregnant women in the first trimester still went to work in China,
the women who provided work addresses were assumed to work; therefore we
estimated pollutant concentrations based on both their residence and work
addresses. For each pregnant woman, the exposure concentration of each air
pollutant Cd was computed as Cd = (Cdw/3) + (2Cdr/3), where Cdw and Cdr denote
air pollutant concentrations at the air monitoring stations closest to the maternal
working and residential addresses, respectively. The weights (1/3 and 2/3)
approximately accounted for the times a pregnant woman spent at work and at
home. For the other 14% of pregnant women who did not provide work addresses,
we assumed that they did not go to work and we only used their residential
addresses to estimate the pollution exposure.
To define the period of pollution exposure, we first determined the LET of
pregnant women before MAFT. We took the date of the LMP plus gestational age
as the LET of the pregnant woman. We examined whether and how the exposure–
response association was affected by different time periods of pollution exposure.
We examined seven time windows of maternal exposure to air pollution (Phases
1–7), each from 0, 30, 60, 90, 120, 150 or 180 d before the first day of the LMP to
the LET. We calculated the mean daily concentrations (the average of 24-h average
across multiple days) of the pollutants in different periods (Phase 1, Phase 2 and
so on) during which a pregnant woman was exposed. We showed that exposure
in Phase 4 (from 90 d before the LMP to the LET) had strongest association with
MAFT and these data were used for the analyses.
Spatial generalized additive model. Spatial autocorrelation was considered in this
study. Supplementary Fig. 11 shows that most pregnant women with MAFT were
clustered in densely populated areas where air pollution was high.
The spatial generalized additive model was used to account for variation.
guð Þ¼β
0
þsPM
2:5
;df
1
ð ÞþsSO
2
;df
2
ð ÞþsO
3
;df
3
ðÞ
þsCO;df4
ð Þþslat;longðÞ ð
1
Þ
where u = E(Y) is the mathematical expectation of Y and Y(t = 1, 2, 3, …, n)
denotes the set of the participating pregnant woman; g is a monotonic link function
of u; β0 is the intercept; df represents the degree of freedom and is used to control
the impact of various pollutants; (lat, long) denotes the location of the residence or
working place of a pregnant women; s is a smooth function; and s(lat, long) denotes
the impact of the spatial autocorrelation on the MAFT risk.
Supplementary Fig. 12 illustrates the partial residuals of s(lat, long) through
controlling the spatial distribution of the pregnant women in the generalized
additive model. It is noted that the spatial distribution of the pregnant women
in the lower right corner of the domain had large residual differences and a
large degree of aggregation, which indicated that the data may have the spatial
autocorrelation.
Logistic regression and restricted cubic spline regression. We used a logistic
regression model to evaluate the risk factors that influence MAFT. The results
of the analysis showed that the mother’s age, mother’s occupation, ambient air
temperature and maternal exposure to each of the pollutants PM2.5, SO2, CO and
O3 were correlated with the MAFT risk.
Potential confounding factors were controlled in the final logistic regression
model. In addition, when associating each pollutant with MAFT, other pollutants
were controlled in the logistic regression model30. Taking into consideration
the possibility that the OR might be influenced by spatial dependence among
participating pregnant women, the logistic regression model was formulated as
ln P
1�P
¼β
0
þβ
1
X
1
þβ
2
X
2
þβ
3
X
3
þβ
4
X
4
þβ5X5þβ6X6þβ7X7þγlatð ÞþδlongðÞ ð
2
Þ
where P denotes the probability of the MAFT risk and β0 is a constant term; β1,
β2, …, β7 are the regression coefficients of the independent variables X1 to X7
(X1 denotes the mother’s age at conception, X2 the mother’s occupation, X3 the
ambient air temperature, X4 is the ambient PM2.5 concentration, X5 is the ambient
SO2 concentration, X6 is the ambient O3 concentration and X7 is the ambient CO
concentration). The OR value of each independent variable is ORi = exp (βi); γ
and δ are the coefficients; (lat, long) is the geographical location of the maternal
residence or work place.
We also used the restricted cubic spline regression model49 to help associate air
pollution exposure and the MAFT risk. In the restricted cubic regression spline,
the function was linear in two intervals, [t0, t1] and [tk−1, tk], of a predicting variable,
so the restricted cubic regression spline, RCS(X), can be described as
RCS x;kð Þ¼
X
k
�
1
i¼1
βiSiXðÞ ð3
Þ
with
S
1
xðÞ ¼ x
SixðÞ ¼ x�ti�1
ðÞ
3
þ�x�tk�1
ðÞ
3
þtk�ti�1
ðÞ
tk�tk�1þx�tk
ðÞ
3
þtk�1�ti�1
ðÞ
tk�tk�1if i≥2
x
�
ti�1
ðÞ
3
þ¼
x�ti�1
ðÞ
3if x≥ti�1
0 else
NATURE SUSTAINABILITY | www.nature.com/natsustain
Articles Nature SuStaiNability
where k denotes the number of the nodes and Si denotes the spline function; x is
the value of a continuous exposure X, i is an integer and t denotes the endpoint of
each interval. We selected four nodes representing the 25th, 50th, 75th and 95th
percentiles of PM2.5 concentrations and three nodes representing the 25th, 50th and
75th percentiles of SO2, O3, CO concentrations, respectively.
The logistic regression model could be combined with the restricted cubic
splines to deal with the nonlinear relationship between the response variables
and independent variables. We combined the two models to assess the exposure–
response relationship. We associated MAFT and each air pollutant (PM2.5,
SO2, CO and O3) separately. Using the spline function RCS(x) to replace the
independent variable x in equation (2), we estimated the nonlinear relationship
between the exposure concentration of each air pollutant and the MAFT risk
through equation (4).
ln P
1
�
P
¼β0þβ1X1þ þ
X
k
�
1
i¼1
βiSiXð Þþγlatð ÞþδlongðÞ ð4
Þ
Correlations between air pollutants. We computed the correlations between
individual air pollutants, on the basis of daily mean air pollution concentration
data from June 2014 to December 2017 that were obtained from 34 air pollution
monitoring stations in Beijing (Supplementary Table 2). We found that PM2.5 and
CO had the strongest correlation. We constructed a multivariate linear model to
further analyse the multicollinearity between pollutants. The variance inflation
factor of PM2.5 was larger than 10.0, suggesting that PM2.5 had collinearity with
other pollutants (primarily CO) (Supplementary Table 8). Similarly, the variance
inflation factor of CO was close to 10.0, reflecting its high collinearity with PM2.5.
Data availability
The collected data are available from the corresponding authors on
reasonable request.
Code availability
The source code is available from the corresponding authors on reasonable
request. It is copyrighted by Beijing Normal University and Beijing Obstetrics and
Gynecology Hospital and is to be used only for educational and research purposes.
Any commercial use is prohibited.
Received: 9 January 2019; Accepted: 27 August 2019;
Published: xx xx xxxx
References
1. Cohen, A. J. etal. Estimates and 25-year trends of the global burden of
disease attributable to ambient air pollution: an analysis of data from the
Global Burden of Diseases Study. Lancet 389, 1907–1918 (2015).
2. Jing, H. etal. Health impact of China’s air pollution prevention and control
action plan: an analysis of national air quality monitoring and mortality data.
Lancet Planet. Health 2, e313–e323 (2018).
3. Shah, P. S. etal. Air pollution and birth outcomes: a systematic review.
Environ. Int. 37, 498–516 (2011).
4. Malley, C. S. etal. Preterm birth associated with maternal ne particulate
matter exposure: a global, regional and national assessment. Environ. Int. 101,
173–182 (2017).
5. Brook, R. D. etal. Insights into the mechanisms and mediators of the eects
of air pollution exposure on blood pressure and vascular function in healthy
humans. Hypertension 54, 659–667 (2009).
6. Dadvand, P. etal. Ambient air pollution and preeclampsia: a spatiotemporal
analysis. Environ. Health Perspect. 121, 1365–1371 (2013).
7. Ha, S. etal. Ambient air pollution and the risk of pregnancy loss: a
prospective cohort study. Fertil. Steril. 109, 148–153 (2017).
8. Gemzell-Danielsson, K. etal. Misoprostol to treat missed abortion in the rst
trimester. Int. J. Gynaecol. Obstet. 99, S182–S185 (2007).
9. Steer, C. etal. Spontaneous abortion rates aer natural and assisted
conception. BMJ 299, 1317–1318 (1989).
10. Daya, S. in Textbook of Gynecology (ed. Copeland, L. J.) 227–271 (WB
Saunder, 2000).
11. Wood, S. Medical management of missed abortion: a randomized clinical
trial. Obstet. Gynecol. 99, 563–566 (2002).
12. Farah, S. & Lucy, K. Intrauterine fetal death. Obstet. Gynaecol. Reprod. Med.
19, 1–6 (2009).
13. Green, R. S. etal. Residential exposure to trac and spontaneous abortion.
Environ. Health Perspect. 117, 1939–1944 (2009).
14. Defranco, E. etal. Air pollution and stillbirth risk: exposure to airborne
particulate matter during pregnancy is associated with fetal death. PLoS ONE
10, e0120594 (2015).
15. Hou, H. Y. etal. Does ambient air pollutants increase the risk of fetal loss?
A case-control study. Arch. Gynecol. Obstet. 289, 285–291 (2014).
16. Liang, X. etal. PM2.5 data reliability, consistency, and air quality assessment in
ve Chinese cities. J. Geophys. Res. Atmos. 121, 10220–10236 (2016).
17. Ji, J. Air pollution and China’s ageing society. Lancet Public Health 3,
e457–e458 (2018).
18. Bowe, B. etal. e 2016 global and national burden of diabetes
mellitus attributable to PM2.5, air pollution. Lancet Planet. Health 2,
e301–e312 (2018).
19. Chen, H. Living near major roads and the incidence of dementia, Parkinson’s
disease, and multiple sclerosis: a population-based cohort study. Lancet 389,
718–726 (2017).
20. Smith, R. etal. Impact of London’s road trac air and noise pollution
on birth weight: retrospective population based cohort study. BMJ 359,
j5299 (2017).
21. Wang, Y. etal. Association of long-term exposure to airborne particulate
matter of 1 μm or less with preterm birth in China. JAMA Pediatr. 8,
e174872 (2018).
22. Mahalingaiah, S. etal. Perimenarchal air pollution exposure and menstrual
disorders. Hum. Reprod. 33, 512–519 (2018).
23. Wakeeld, J. Bayes factors for genome-wide association studies: comparison
with P-values. Genet. Epidemiol. 33, 79–86 (2009).
24. Lee, M. D. & Wagenmakers, E.-J. Bayesian Cognitive Modeling: A Practical
Course (Cambridge Univ. Press, 2013).
25. Schönbrodt, F. D. & Wagenmakers, E. J. Bayes factor design analysis: planning
for compelling evidence. Psychon. Bull. Rev. 25, 128–142 (2018).
26. Patt, A. G. etal. Estimating least-developed countries’ vulnerability to
climate-related extreme events over the next 50 years. Proc. Natl Acad. Sci.
USA 107, 1333–1337 (2010).
27. Smith, K. R. etal. in Climate Change 2014: Impacts, Adaptation, and
Vulnerability (eds Field, C. B. etal.) 709–794 (Cambridge Univ. Press, 2014).
28. Weinberg, C. R., Moledor, E. & Baird, D. D. Is there a seasonal pattern in risk
of early-pregnancy loss. Epidemiology 5, 484–489 (1994).
29. Ciaula, A. & Bilancia, M. Relationships between mild PM10 and ozone
urban air levels and spontaneous abortion: clues for primary prevention.
Int. J. Environ. Health Res. 25, 640–655 (2015).
30. Enkhmaa, D. etal. Seasonal ambient air pollution correlates strongly with
spontaneous abortion in Mongolia. BMC Pregnancy Childbirth 14, 146 (2014).
31. Slama, R. etal. Meeting report: atmospheric pollution and human
reproduction. Environ. Health Perspect. 116, 791–798 (2008).
32. Faiz, A. S. etal. Does ambient air pollution trigger stillbirth? Epidemiology
24, 538–544 (2013).
33. Siddika, N. etal. Prenatal ambient air pollution exposure and the risk of
stillbirth: systematic review and meta-analysis of the empirical evidence.
Occup. Environ. Med. 73, 573–581 (2016).
34. Tsamou., M. etal. Air pollution-induced placental epigenetic alterations in
early life: a candidate miRNA approach. Epigenetics 13, 135–146 (2016).
35. Stillerman, K. P. etal. Environmental exposures and adverse pregnancy
outcomes: a review of the science. Reprod. Sci. 15, 631–650 (2008).
36. Perera, F. P. etal. Molecular epidemiologic research on the eects of
environmental pollutants on the fetus. Environ. Health Perspect. 107(suppl 3),
451–460 (1999).
37. Byrne, J. etal. Morphology of early fetal deaths and their chromosomal
characteristics. Teratology 32, 297–315 (1985).
38. Šrám, R. J. etal. Ambient air pollution and pregnancy outcomes: a review of
the literature. Environ. Health Perspect. 113, 375–382 (2005).
39. Hardy, K. R. & om, S. R. Pathophysiology and treatment of carbon
monoxide poisoning. J. Toxicol. Clin. Toxicol. 32, 613–629 (1994).
40. Ziaei, S. etal. Eects of carbon monoxide air pollution in pregnancy
on neonatal nucleated red blood cells. Paediatr. Perinat. Epidemiol. 19,
27–30 (2005).
41. Singh, J. & Scott, L. H. reshold for carbon monoxide induced fetotoxicity.
Teratology 30, 253–257 (1984).
42. Grace, K. Considering climate in studies of fertility and reproductive health
in poor countries. Nat. Clim. Change 7, 479–485 (2017).
43. Li, X. etal. A new global land-use and land-cover change product at a 1-km
resolution for 2010–2100 based on human–environment interactions. Ann.
Am. Assoc. Geogr. 107, 1040–1059 (2017).
44. International Statistical Classication of Diseases and Related Health Problems
10th revision (WHO, 2007).
45. Altman D. G. Practical Statistics for Medical Research (Chapman and
Hall, 1991).
46. Technical Regulation on Ambient Air Quality Index (on Trial). (Ministry of
Environmental Protection of the People’s Republic of China, 2012).
47. Liu, M. Y. etal. Spatiotemporal variability of NO2 and PM2.5 over Eastern
China: observational and model analyses with a novel statistical method.
Atmos. Chem. Phys. Discuss. 18, 12933–12952 (2018).
48. Zhang, L. etal. Air pollution exposure associates with increased risk of
neonatal jaundice. Nat. Commun. 10, 6741 (2019).
49. Desquilbet, L. & Mariotti, F. Dose–response analyses using restricted cubic
spline functions in public health research. Stat. Med. 29, 1037–1057 (2010).
NATURE SUSTAINABILITY | www.nature.com/natsustain
Articles
Nature SuStaiNability
Acknowledgements
This work was carried out with the support of the National Key Research and
Development Program of China (grant no. 2018YFC0213600), the National Natural
Science Foundation of China (grant nos. 41775115 and 41371324) and the Beijing
Natural Science Foundation (grant no. 7173258).
Author contributions
L.Z. and W.L. jointly designed the study, collected data, performed the analysis and
wrote the manuscript. K.H. contributed to the model of the paper. J.L. contributed to
the research framework, provided air pollution and meteorological data, contributed to
results analysis and edited the paper. C.Z. and X.T. improved the research framework and
edited the paper. Z.Y.W., Y.B.W. and Y.L. organized the neonatal and maternal datasets.
Z.W.W., R.N. and M.L. generated the air pollution and meteorological datasets. Y.J., Y.Z.,
S.L. and P.Z. developed the maps and edited figures.
Competing interests
The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/
s41893-019-0387-y.
Correspondence and requests for materials should be addressed to L.Z., W.L., J.L.
or C.Z.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2019
NATURE SUSTAINABILITY | www.nature.com/natsustain
A preview of this full-text is provided by Springer Nature.
Content available from Nature Sustainability
This content is subject to copyright. Terms and conditions apply.