Ambient air pollution in critical windows of exposure and spontaneous miscarriage in a preconception cohort.

Of the 446 conceptions, 101 ended in spontaneous miscarriage, and 345 either lasted longer than 20 weeks, live births (n=337) or still births (n=2), or were censored at the time of induced abortion (n=6). The frequency of spontaneous miscarriages was higher in the oldest age group (29%) compared with the younger age group (21%), and this association was similar for partner age ( Table 1 ). Those who reported their race/ethnicity or their partner’s race/ethnicity as a category other than White also had a higher frequency of spontaneous miscarriage. Spontaneous miscarriage was also more frequent among those of higher parity, lower education, higher BMI, and lower income. Characteristics of participants included and excluded from this analysis are shown in Supplemental Table 1 . The median, 25 th , and 75 th percentiles of each pollutant averaged across each exposure window are shown in Supplemental Table 2 . The same statistics were calculated for the peak concentration of each pollutant in each exposure window and are shown in Supplemental Table 3 . The width of the IQR for each pollutant is shown in Supplemental Table 4 . There was little evidence of associations between average ambient air pollutant concentrations and spontaneous miscarriage. ( Figure 2 ) The HRs for ozone in the spermatogenesis (HR[CI]: 1.2 [0.64, 2.3]), early follicle development windows (HR[CI]: 1.2 [0.64, 2.4]), and luteal phase (HR[CI]: 1.2 [0.68, 1.9]) were elevated, but confidence intervals were relatively wide. When these HRs were further adjusted for ozone levels in a previous time window, we found that the HR in the early follicle development window (adjusted for the spermatogenesis window) was 0.94 (CI: 0.19, 4.6). ( Supplemental Table 5 ) The HR for the luteal phase window (adjusted for the early follicle development window) was 1.1 (CI: 0.66, 1.9). There was no conclusive evidence that the association between air pollutants and spontaneous miscarriage differed at early gestational ages compared with late (all interaction p-values >0.05). ( Supplemental Figure 1 ) In the sensitivity analysis accounting for study year, estimates were generally stronger, although for most pollutants the differences were quite small. However, estimates were noticeably stronger for SO 2 , particularly in the luteal, implantation, and luteal-placental shift windows, and for ozone, although confidence intervals were still quite wide. ( Supplemental Figure 2 ) Results were also unchanged when accounting for missing data. ( Supplemental Figure 3 ) When the peak pollutant concentrations for each exposure window were examined, the confidence intervals were generally more precise compared with the average pollutant concentrations. ( Figure 3 ) Again, the estimate for ozone was elevated in the spermatogenesis window (HR[CI]: 1.3 [0.85, 1.9]), but not the early follicle development window. Similarly, the spontaneous miscarriage rate was slightly elevated for increasing NO 2 in the spermatogenesis window (HR[CI]: 1.2 [0.87, 1.5]), the early follicle development window (HR[CI]: 1.2 [0.85, 1.6]), the luteal phase (HR[CI]: 1.1 [0.84, 1.5]), and the implantation window (HR[CI]: 1.2 [0.87, 1.6]). In the luteal-placental shift window, an IQR increase in peak PM 10 concentration was associated with a slight increase in spontaneous miscarriage rate (HR[CI]: 1.1 [0.95, 1.4]), but this was not true for any other exposure window. When we examined the NO 2 results adjusted for NO 2 in a previous window, the HR for the early follicle development window was 1.9 (0.95, 3.9). The HR for the luteal phase window (adjusted for the early follicle development window) was 1.4 (CI: 0.78, 2.4). The HR for the implantation window (adjusted for the early follicle development window) was 1.5 (CI: 0.87, 2.6). When we examined the PM 10 association in the luteal-placental shift window adjusted for the implantation window, the HR was 1.1 (0.94, 1.4). ( Supplemental Table 5 ) There was no conclusive evidence that the associations between air pollutants and spontaneous miscarriage differed for early miscarriages compared with late miscarriages. ( Supplemental Figure 4 ) In the sensitivity analysis accounting for study year, estimates were generally stronger, particularly in the spermatogenesis window. ( Supplemental Figure 5 ) Across time windows, estimates were noticeably stronger for SO 2 but particularly in the luteal, implantation, and luteal-placental shift windows. Results were unchanged when accounting for missing data. ( Supplemental Figure 6 ) When we examined average pollutant concentrations using an exposure mixtures approach, each quartile increase in all average pollutant concentrations was associated with elevated hazard ratios in the spermatogenesis (HR[CI]: 1.2 [0.80, 1.8]), early follicle development (HR[CI]: 1.2 [0.80, 1.8]), and luteal phase windows (HR[CI]: 1.2 [0.80, 1.9]). ( Figure 4 , Supplemental Table 6 ) Increasing ambient air pollutant exposure mixtures were not associated with spontaneous miscarriage rates in the other exposure windows. When accounting for study year, estimates were stronger for the early follicle development and the spermatogenesis windows. ( Supplemental Figure 7 ) When the peak pollutant concentrations were examined as a mixture, again, most estimates were above one but of small magnitude. ( Figure 4 , Supplemental Table 6 ) The spermatogenesis window showed the strongest association with spontaneous miscarriage, where a quartile increase in all single-day peak pollutant concentrations simultaneously was associated with 1.3 times the rate of spontaneous miscarriage (HR[CI]: 1.3 [0.89, 1.9]). When accounting for study year, there was a general increase in the magnitude of the associations for pollutants that were associated with increased miscarriage rate. For example, in the spermatogenesis window, an IQR increase in all air pollutants simultaneously was associated with 1.6 times the rate of spontaneous miscarriage (HR[CI]: 1.6 [1.1, 2.5]) ( Supplemental Figure 7 ) compared with HR: 1.3 in Figure 4 . In general, associations of ambient air pollutants with spontaneous miscarriage were stronger among participants with low vitamin D compared to those with high vitamin D ( Figure 5 ). In the spermatogenesis window, among those with low vitamin D, spontaneous miscarriage was associated with increasing NO (HR[CI]: 1.4 [1.0, 1.9]) and CO (HR[CI]: 1.6 [0.99, 2.7]). However, these associations were not present, or weakly opposite, among those with high vitamin D (NO, HR[CI]: 0.83 [0.59, 1.2] p interaction = 0.030; CO, HR[CI]: 0.94 [0.62, 1.4], p interaction = 0.100). In the follicular phase, average ozone was associated with increased spontaneous miscarriage among those with low vitamin D (HR[CI]: 3.1 [1.3, 7.4], but not among those with high vitamin D (HR[CI]: 0.84 [0.46, 1.5], p interaction = 0.002) ( Figure 5 ). The association between increasing ozone and spontaneous miscarriage among those with low vitamin D was also present in the luteal phase (p interaction = <0.001), implantation (p interaction < 0.001), and luteal-placental shift (p interaction = 0.022). For all of the remaining interactions, p interaction was greater than 0.100. None of the interactions between BMI and average pollutant exposure were significant (all p interaction > 0.150). ( Supplemental Table 7 ) Only one of the interactions between average pollutant exposure and exercise was of borderline significance (PM 2.5 during early follicle development, p interaction = 0.062) ( Supplemental Table 8 ), and this was not a pollutant or exposure window where an interaction with vitamin D was present. No differences in multipollutant models were detected when stratified by vitamin D. ( Supplemental Figure 8 ) In the spermatogenesis window, increasing peak PM 2.5 was associated with increased spontaneous miscarriage rates only among those with high vitamin D (HR[CI]: 1.4 [1.0, 1.9] vs. low vitamin D, HR[CI]: 0.69 [0.40, 1.2], p interaction = 0.026) while increasing NO (p interaction = 0.009) and CO (p interaction = 0.032) were only associated with increased spontaneous miscarriage among those with low vitamin D (NO, HR[CI]: 1.9 [1.1, 3.1] vs. 0.80 [0.51, 1.3] for high vitamin D, CO, HR[CI]: 1.7 [1.1, 2.9] vs. 0.87 [0.58, 1.3]) ( Supplemental Figure 9 , Supplemental Table 9 ). As with average pollutant concentrations, increasing peak ozone was associated with increased spontaneous miscarriage rates among those with low vitamin D across the follicular phase (HR[CI]: 2.0 [0.91, 4.3] vs. 0.67 [0.33, 1.3], p interaction = 0.007), luteal phase (HR[CI]: 2.7 [1.4, 5.4] vs. 0.67 [0.33, 1.4], p interaction = <0.001), implantation (HR[CI]: 3.1 [1.5, 6.4] vs. 0.79 [0.40, 1.6], p interaction < 0.001), and luteal-placental shift windows (HR[CI]: 2.5 [1.1, 5.7] vs. 0.92 [0.41, 2.1], p interaction = 0.007). Interactions between BMI and peak pollutants were largely non-significant, with only two potential exceptions: ozone in the spermatogenesis window (p interaction = 0.016) and the early follicle development window (p interaction = 0.078). ( Supplemental Table 7 ) Interactions between exercise and peak pollutants were also non-significant, with only three of borderline significance: PM 2.5 in the early follicle development (p interaction = 0.073) and follicular phase (p interaction = 0.103) windows and SO 2 in the early follicle development window (p interaction = 0.070). ( Supplemental Table 8 ) Again, these interactions were not in the same exposure windows or for the same pollutants where interactions with vitamin D were present. No differences in multipollutant models were detected when stratified by vitamin D. ( Supplemental Figure 10 )

TTC was a prospective, time-to-pregnancy cohort study that collected data between 2008 and 2016 in the Raleigh-Durham-Chapel Hill Area of North Carolina. Eligible participants were between the ages of 30 and 44 and trying to conceive for less than three months at the time of recruitment ( 30 ). Participants were excluded if they had a known history of infertility (sterilization, previous or current use of fertility treatments, known tubal blockage, or endometriosis), had a partner with a history of infertility, had used injectable hormone contraception within the past year, or were currently breastfeeding. For this analysis, we included all participants with a geocoded address within North Carolina who also reported a positive pregnancy test during the study period (n = 662). ( Figure 1 ) This study was approved by the institutional review board (IRB) at the University of North Carolina at Chapel Hill (#08–0423) and the Duke University IRB (Pro00100281), and written informed consent was provided by all participants prior to participation in the study. Participants kept daily diaries for the first four months of the study and then monthly diaries afterward. In the diaries, they tracked menstrual bleeding and any positive pregnancy tests. They were given urinary pregnancy test kits and asked to take them at the time of their missed period and each third day afterward. If they had a positive home pregnancy test, they were asked to call the study office and report it to study staff. Once a positive test was reported, a pregnancy outcome questionnaire became available in their study portal. Participants reported the outcome of their pregnancy, when it ended, and other pregnancy characteristics. Vitamin D level was defined by 25-hydroxyvitamin D (25(OH)D) levels measured by liquid chromatography tandem mass-spectrometry in blood spots collected at study enrollment ( 24 ). Low vitamin D was defined as a 25(OH)D level of <30 ng/ml, while high vitamin D was defined as ≥30 ng/ml. We derived the PM 2.5 and O 3 metrics from the 2016 US Environmental Protection Agency’s (EPA) Fused Air Quality Surface Using Downscaling (FAQSD) model, which estimated ambient air pollution concentrations at 12 km by 12 km grids ( 32 , 33 ). This space-time Bayesian fusion model integrates monitoring data from National Air Monitoring Stations/State and Local Air Monitoring Stations (NAMS/SLAMS) with estimates from a meteorological grid model called the Community Multi-Scale Air Quality Model (CMAQ) ( 34 ) to predict daily ambient O 3 ( ppb ) and PM 2.5 ( ug / m 3 ). As no existing fusion model existed for the other pollutants of interest, concentrations of CO ( ppb ), NO ( ppb ), NO 2 ( ppb ), SO 2 ( ppb ), and PM 10 ( ug / m 3 ) were derived from the CMAQ alone at a 12 km by 12 km gridded resolution. The performance of CMAQ was evaluated at an earlier version (v4.3) and was designed for accuracy and prediction. Each new version is compared with the previous version to identify improvements or changes in the model estimates ( 33 , 35 ). Briefly, when compared with air monitor data, the FAQSD has a mean bias near zero for both PM 2.5 and O 3 and appropriate coverage (95% for both). To estimate exposures, participant residential addresses at enrollment were geocoded using ArcGIS and spatially linked to the CMAQ and FAQSD air pollution estimates for the 12 km x 12 km grid in which they resided. Participants reported cohabitating with their partners. We assessed six critical windows of exposure: 1) spermatogenesis, 2) early follicle development, 3) follicular phase of the conception cycle, 4) luteal phase of the conception cycle, 5) implantation, and 6) the luteal-placental shift. Last menstrual period (LMP) was prospectively recorded in daily diaries for all participants during the study. The spermatogenesis window was defined as the 105 days immediately prior to LMP since spermatogenesis is estimated to take from 74 days ( 36 ) to 120 days ( 37 ). This window approximates the timing of spermatogenesis which is not well-understood and may be imprecise. The early follicle development window captures the 84 days prior to ovulation, during which the pre-antral and antral follicles are most susceptible to atresia ( 9 ), and is defined as the 70 days prior to LMP plus the follicular phase of the cycle (estimated to be 14 days). Ovulation was assigned as day 15 of the cycle, the mode for the study sample when measured by ovulation predictor kit in a subset of participants ( 38 ). The follicular phase was defined as the first 14 days of the conception cycle. The luteal phase of the conception cycle was defined as the window between day 16 to day 28. The implantation window was defined as days 6 to 12 post-ovulation ( 39 ). The luteal-placental shift was defined as starting 30 days after LMP and ending 60 days after LMP. During the luteal-placental shift, the placenta takes over progesterone production from the corpus luteum, which happens around 3 to 7 weeks after fertilization ( 40 ). For each exposure window and for each pollutant, we estimated the average and the single-day peak concentration (“peak”) where the average concentration was calculated as the mean of the daily average concentrations and the single-day peak concentration was calculated as the maximum of the daily average concentrations. We excluded any participants whose addresses could not be linked to the CMAQ or FAQSD and therefore had missing exposure information (N=5). The primary outcome of interest was spontaneous miscarriage, defined as an intrauterine pregnancy loss prior to 20 weeks gestation based on the last menstrual period. Conceptions were identified by a positive home pregnancy test, and if multiple positive tests were reported, only the first positive test was included. Later in the study, participants who reported a positive test were then asked to report the outcome of that pregnancy. This outcome questionnaire was added later in the study and early participants are missing pregnancy outcome for this reason. Those who participated prior to the collection of pregnancy outcome data or those who did not report outcome data were excluded (N=202). Positive tests that were followed by the report of a negative test within four days were not considered conceptions (N=8). One tubal pregnancy was excluded as a non-intrauterine pregnancy. We used the date that the participant reported their miscarriage was “diagnosed” to define the timing of the outcome. Any pregnancy loss that occurred after 20 weeks (N=3) was not considered a spontaneous miscarriage and was combined with the live birth group. Induced abortions were also included as censored pregnancies (N=6). ( Figure 1 ) Potential confounders were selected based on a directed acyclic graph ( 41 , 42 ) and included age, partner age, parity, education, body mass index (BMI), season, race/ethnicity, partner race/ethnicity, median household income, and daily average land surface temperature (° C). All covariates were categorized except temperature, which was continuous ( Table 1 ). BMI was calculated from self-reported weight and height. Season was defined based on the first day of the conception cycle. Participant race/ethnicity was self-reported. Participants were asked to invite their partners to help them answer questions about their partner, including race/ethnicity, age, and BMI. Census tract median household income estimates were obtained from the 2006–2010 American Community Survey from the U.S. Census Bureau and linked to participant residential locations. We linked National Aeronautics and Space Administration spectroradiometry data at 1 km by 1 km gridded resolution (Moderate Resolution Imaging Spectroradiometer onboard Terra provided Land Surface Temperature 3-Band Emissivity 8-Day L3 Global 1km SIN Grid) to geocoded participant residential addresses to estimate average daily temperature values. Used NASA LP DAAC system to obtain grid tile data layers covering the study area and study time period (one file per day per tile). We assessed the performance of several potential space-time integration algorithms (necessary to compensate for cloud cover) to produce consistent and complete temperature estimates for study domains. We found MOD11A2 v006 L3 smoothing algorithm to have the best performance with nearly complete coverage of the study domain and applied this algorithm with daily interpolation to compute daily land surface temperature estimates for each 1km grid cell in the study in degrees Kelvin. We converted from degrees Kelvin to degrees Celsius and re-projected the coordinate system to latitude and longitude to enable linkage to participants by matching the grid of temperature values to participant residence coordinates. Daily land surface temperatures were averaged across the previously described windows of exposure and are reported in Celsius. Descriptive statistics for couples at baseline and over follow-up were summarized as counts and percentages for categorical variables and medians and interquartile ranges (IQRs) for continuous variables. We used multivariable Cox proportional hazards models to estimate the hazard ratios (HRs) and 95% confidence intervals (CIs) for spontaneous miscarriage for an interquartile range increase in each pollutant concentration, for each pollutant separately. The person-time (in days) for spontaneous miscarriages was defined as the date of the positive pregnancy test to the date of the self-reported miscarriage diagnosis. Induced abortions were included in the analysis as censored pregnancies, with the observation time defined as the date of the positive pregnancy test to the reported date of induced abortion. Person-time for live births and stillbirths was defined as the time from the positive pregnancy test up to 20 weeks of gestation (the end of the risk period for spontaneous miscarriage). We used delayed entry to account for the time from LMP to the positive pregnancy test. Cox proportional hazards models were conducted using R version 4.3.1 and the R package survival , version 3.5–5. Interactions between air pollutants and gestational age, categorized as “early,” ≤ 6 weeks, versus “late,” > 6 weeks, were tested, and time-dependent coefficients are presented as supplemental material . Because air pollution can potentially affect spontaneous miscarriage at multiple time points, our analyses that look at impacts of individual windows of exposure are sensitive to unmeasured confounding by exposure at other time periods. For example, because air pollutants are correlated over time, associations in one window could be, in part, due to confounding by exposure in a correlated window, and vice versa. In illustrative sensitivity analyses, we adjusted some of the observed associations for exposure in a previous time window. Because spermatogenesis was the earliest time window available, it could not be adjusted for a previous time window and was not included in this sensitivity analysis. We used quantile-based g-computation ( 31 ) with a Cox proportional hazards framework to estimate the joint association of spontaneous miscarriage with mixtures of pollutants within each exposure window. This approach estimates the association between spontaneous miscarriage and a simultaneous one quartile increase in all pollutants. As Keil et al. (2020) showed, this approach is equivalent to estimating the impact of a weighted index where the weights are proportional to the effect size of an exposure in a given direction. Similar to any parametric model, the results of quantile-based g-computation can be sensitive to modeling assumptions. This approach assumes that, on a quantile basis, the effects of exposures are linear and additive. As with other generalized linear models, it is sensitive to issues like unmodeled interactions and unmodeled non-linear effects of exposures. In particular, we opted for a linear specification which excluded interaction terms from the model, which is the most common approach for this method that has been used extensively in environmental epidemiology. Quantile-based g-computation can be seen as a linear approximation of the true non-linear model ( 43 ), so that, for example, tests of the coefficients do not have inflated Type I error rates ( 44 ). Thus, we do not explore non-linearity or interactions here because they would not be expected to impact qualitative aspects of the results, and non-linearity would require specification of many interaction terms whose estimation may not be supported by the data. As noted in the original paper on quantile-based g-computation, correlation of elements in a mixture is addressed by focusing on the overall effect, though power can be adversely impacted by negatively correlated exposures. The overall effect is less sensitive to correlation and the method avoids using statistical techniques like shrinkage or selection which can result in bias (but possibly reduce variance) for methods that use them like weighted quantile sum regression and Bayesian kernel machine regression ( 45 , 46 ). We do not use those methods here because they do not permit use of a Cox model to fully permit use of time-to-event outcomes like miscarriage. To assign causal interpretations to results from any statistical method, we must also assume that the causal identification conditions hold ( 47 ). All exposure mixtures models controlled for the same covariates as our single-pollutant models. Quantile-based g-computation was carried out using the R package qgcomp version 2.7.0. We examined the associations between ambient air pollutants and spontaneous miscarriage stratified by vitamin D level. It is possible that potential effect measure modification by vitamin D level could be due to causes of vitamin D level that are also effect measure modifiers, thus, we also investigated interactions between ambient air pollutants and BMI and self-reported exercise at conception (hours per week; 0, ≤1, 1–3, 4–7, >7) to ensure that any heterogeneity across vitamin D categories was not due to heterogeneity in these other factors. Likelihood ratio tests were used to estimate p-values for the interactions (p interaction ) between each air pollutant and either vitamin D level, BMI, or exercise. The first two years of the study (2008/2009) had slightly lower miscarriage rates compared with the rest of the study years and some pollutants had higher levels in those early years. To address this, we also performed a sensitivity analysis excluding data from those years.

Overall, ozone concentrations showed the most consistent, small elevations in spontaneous miscarriage rates. Both average and peak ozone in the spermatogenesis window were associated with increased spontaneous miscarriage. When stratified by participant vitamin D levels, ozone in several exposure windows was associated with increased spontaneous miscarriage rates among participants with low vitamin D. Nitrogen oxides also showed small associations with increased spontaneous miscarriage rates, particularly peak NO 2 and, among participants with low vitamin D, average and peak NO. While the mechanisms underlying these associations are unknown, air pollutants generally are associated with oxidative stress, inflammation, and genetic and genomic perturbations ( 4 , 48 – 50 ). We found that the associations of ambient air pollutants with spontaneous miscarriage differed for people with low levels of vitamin D compared to those who had high levels of vitamin D. The interaction between ambient air pollutants and vitamin D status was particularly strong for O 3 , CO, and NO. While our sample size was small, we did not find interactions between these ambient air pollutants and BMI or exercise during these same exposure windows, suggesting these factors may not explain the heterogeneity across vitamin D levels. Previous research suggests that low vitamin D is associated with delayed ovulation and prolonged menstrual cycles both in humans ( 24 ) and in animals. Higher vitamin D may lead to healthier menstrual cycles, which are then better able to support a pregnancy in the presence of ambient air pollutants. Additionally, in humans, low levels of total 25OHD (<20 or <30 ng/ml) have been associated with a lower probability of spontaneous conception ( 22 , 51 ). Pregnancy rates in those using artificial reproductive technologies (ART) are lower in vitamin D-deficient recipients of donor eggs ( 52 ) and vitamin D supplementation improves endometrial thickness ( 53 ), suggesting that uterine receptivity may be influenced by vitamin D, even if inconsistent ( 54 ). In addition, vitamin D may influence other important processes for pregnancy initiation and development, including uterine decidualization, embryonic implantation, and early placentation. Defects in any of these processes can cause a “ripple effect” of poor pregnancy development leading to spontaneous miscarriage ( 55 ). Vitamin D deficiency in animals has been associated with uterine hypoplasia, subfertility, and fetal resorptions, which suggests a link with uterine dysfunction and subsequent spontaneous miscarriage ( 56 – 60 ). An in vitro study of human endometrial cells reported that vitamin D enhanced decidualization ( 61 ) and upregulated mRNA and protein expression critical for decidualization. ( 62 )( 55 ) Extravillous trophoblast invasion, the invasion of embryonic cells into the maternal vasculature, initiates placentation, and is improved with vitamin D ( 63 ). In total, there is a substantial literature linking vitamin D to improved uterine receptivity, embryonic implantation, and early placentation. The beneficial effects of vitamin D may be able to counterbalance the detrimental effects of toxic exposures, such as ozone. As a mechanistic example, ozone has been associated with increased lung inflammation, potentially through a pathway involving advanced glycation end-products (AGEs) ( 64 ). AGEs have been found in numerous tissues, including ovarian follicles, and are considered a marker of oxidative stress produced through multiple human behaviors, such as smoking and high heat cooking ( 65 ). In ovarian follicles (specifically granulosa cells), studied in vitro , vitamin D inhibited the effects of AGEs ( 65 ). Ozone exposure may trigger the AGE pathway which can be ameliorated with higher levels of vitamin D. Given the limitations of our analysis and the strong epidemiologic, clinical, and biological justification, further study of the interaction between vitamin D status and air pollution on spontaneous miscarriage risk would be important for public health. There are no established interventions that protect people from the deleterious effects of ambient air pollutants after exposure. While our study did not find associations of spontaneous miscarriage with preconception and early pregnancy PM, systematic reviews of ambient air pollution exposure generally have found PM exposure to be associated with increased miscarriage risk ( 2 – 4 ), including both long-term PM exposures across the duration of the pregnancy and acute PM exposure in the week prior to the spontaneous miscarriage ( 4 ). A recent meta-analysis of six studies found that for every 10 μg/m 3 increase in chronic PM exposure, both PM 2.5 and PM 10 were associated with a higher relative risk (RR) of spontaneous miscarriage (RR[CI]: 1.20 [1.01, 1.40] and RR[CI]: 1.09 [1.02, 1.15], respectively) ( 66 ). Consistent with our results, these reviews reported associations between miscarriage and NO 2 and O 3 ( 2 – 4 ), and recent original research also finds associations with these pollutants ( 67 – 75 ). Most of these previous studies focus on gestational ambient air pollution exposures, and to our knowledge, only nine studies also included preconception exposure windows ( 18 – 20 , 76 – 81 ), as we have in this study. Of these nine studies, three found associations between preconception ambient air pollution exposure and increased spontaneous miscarriage risk, including exposure to PM 2.5 , PM 10 , CO, O 3 , SO 2 , and NO 2 ( 18 – 20 ). Each of these studies defined “preconception” differently, which might explain the differences in findings. Our mixtures analysis might help to clarify the “preconception” time frame. Of note, the follicular phase was not identified as an important exposure window, which agrees with two previous studies ( 78 , 79 ). This analysis identified the spermatogenesis, early follicle development, and luteal windows as potentially important exposure windows for ambient air pollutants. The spermatogenesis window was the longest time frame prior to conception and may indicate that longer-term preconception exposures may be relevant for spontaneous miscarriage either through spermatogenesis or another pathway. The early follicle development window may be important for subsequent spontaneous miscarriage risk due to the importance of this time window for the follicle and the oocyte within it to become competent to support a pregnancy. As an example, previous research suggests that disturbances in oocyte metabolism within the ovarian follicle during its development are associated with subsequent meiotic defects, organelle dysfunction, and epigenetic alteration, which result in a higher risk of spontaneous miscarriage ( 82 ). The luteal phase is the time in the menstrual cycle when the uterine endometrium matures and prepares for implantation. Air pollution exposure during this time window could influence uterine receptivity, which could then lead to spontaneous miscarriage. Our exposure window definitions could be used to further examine preconception exposures and timing as targets for intervention. Our study is strengthened by its prospective design, detailed preconception and conception information to identify pregnancies, and incorporation of fused air quality models to estimate ambient air pollution exposure. Moreover, our study is the first to examine air pollutants and spontaneous miscarriage in an exposure mixtures framework and the first to examine the interaction between vitamin D levels and ambient air pollution. Vitamin D levels were measured at baseline using the gold standard method, LC-MS/MS, and while many studies show consistency in vitamin D levels over months or years ( 83 – 85 ), we only measured vitamin D at baseline. The critical exposure windows were defined based on prospective daily diary data where last menstrual period date was recorded. This is a strength of our study, however there is likely measurement error in the critical windows, especially spermatogenesis, as they were not measured biologically, but estimated based on the literature and a number of days prior to LMP. Our study is limited by self-reported BMI, the potential for residual confounding, and the comparatively small number of included miscarriages, especially given that the magnitude of potential associations between air pollutants and spontaneous miscarriage may be small. As with all studies of spontaneous miscarriage, detection is dependent on pregnancy recognition and self-report of the outcome, the exact timing of the pregnancy’s end is unknown. Also, while this is a cohort of pregnancy planners testing early, very early pregnancy spontaneous miscarriages were likely missed. Additionally, air pollutant concentrations were lower in our study than in previous studies, which could limit our power to detect associations. While our exposure assessment is considered the most sophisticated method ( 21 ), it may be limited by the 12 km by 12 km gridded resolution; a finer resolution could provide more precise measures. Our exposure assessment is based on home address rather than personal monitoring, and we do not have information regarding time spent at work or outdoors or changes in address. While we examined pollutant concentrations individually and across time windows, pollutants were correlated within and across time windows, which makes it difficult to isolate their independent associations and negative correlations will reduce power; correlation is expected given the chemical relationships among them (e.g. NO and NO2) and their common sources (e.g., traffic) ( Supplemental Figures 11 and 12 ). The multipollutant models were designed to address correlated pollutant exposures within a time window and are presented here to address that limitation. However, the correlations across time windows are more difficult to address. We tried to address this by adjusting some of our reported results for the exposure level in one previous time window. Point estimates were largely unchanged, and variance increased. Although the results are more defensible in terms of adjustment for confounding, they do not change the interpretation of results that are unadjusted for exposure at other time points. Also, the large increase in the confidence interval width suggested that fully adjusting for exposure at other time points would not provide meaningful results and would be prone to issues of overfit, so instead we focused on the results that are not adjusted for exposures at other time windows. We note that, even given this limitation, only a few time windows were actually associated with the occurrence of spontaneous miscarriage providing some direction toward critical exposure windows. Additionally, the generalizability of our findings to other geographic areas or to unplanned pregnancies may be limited by the inclusion of only pregnancy planners over age 30, living in North Carolina. Nevertheless, our findings suggest several avenues for future research, including examining air pollutants in preconception exposure windows, with a large sample size of pregnancies and spontaneous miscarriages, and among those with low vitamin D levels.

Our study suggests that if associations between ambient air pollutants and spontaneous miscarriage exist, they are small and will require a large sample size to detect. Despite the small effect estimates, given the ubiquitous exposure and the burden of the outcome, these associations are important for public health. An exposure mixtures approach may also increase study power. The spontaneous miscarriage rate was slightly elevated with increasing average and peak ozone, particularly among those with low vitamin D. Nitrogen oxides also showed small associations with increased spontaneous miscarriage rates, particularly peak NO 2, and among those with low vitamin D, average and peak NO. Mixtures of ambient air pollutants during the preconception and early conception cycle exposure windows were associated with small increases in spontaneous miscarriage rates. Future studies should investigate both ambient air pollutants during these critical windows of exposure and the potential of vitamin D status to modify the effects of ambient air pollution exposure.

Ambient or outdoor air pollution is a leading contributor to mortality worldwide ( 1 ). Associations between ambient air pollution and adverse birth outcomes such as reductions in birth weight have been extensively examined ( 2 ). A growing body of literature also investigates associations between ambient air pollution and spontaneous miscarriage (pregnancy loss at <20 weeks gestation) ( 2 – 4 ). Approximately 15% of recognized pregnancies and 30% of pregnancies in prospective cohorts of couples attempting to conceive end in spontaneous miscarriage ( 5 , 6 ). While chromosomal abnormalities may be a factor in up to 60% of these miscarriages, environmental exposures are also hypothesized to contribute ( 5 ). Ambient air pollution is a complex mixture of pollutants emitted from many sources, including automobiles, power plants, industry, and fires. Commonly monitored and regulated pollutants include the gaseous pollutants carbon monoxide (CO), sulfur dioxide (SO 2 ), nitrogen dioxide (NO 2 ), and ozone (O 3 ), and particulate matter (PM), a mixture of solid particles and liquid droplets classified by an aerodynamic diameter as PM 10 (≤10 μm) or PM 2.5 (≤2.5 μm). Once inhaled, these pollutants enter the maternal cardiovascular system, where they can cause oxidative stress, generating free radicals and endocrine-disrupting compounds ( 2 , 7 , 8 ). Follicle maturation begins approximately three months prior to ovulation, and air pollution exposures during this time have been observed to adversely impact ovarian reserve and folliculogenesis, often through inflammatory pathways and oxidative stress ( 9 ). As the follicle is responsible for early pregnancy maintenance, it is hypothesized that adverse exposures could contribute to early pregnancy spontaneous miscarriages ( 10 , 11 ). Air pollutants also may contribute to spontaneous miscarriage by negatively impacting the luteal-placental shift and early pregnancy placental function, through inflammation, compromised oxygen and nutrient delivery, endocrine disruption, and altered fetal and placental DNA methylation ( 2 , 4 , 12 – 15 ). Additionally, spermatogenesis may be affected by air pollution exposure, which has been linked to low semen volume and sperm count, and morphological and motility issues ( 16 , 17 ). The mechanisms by which air pollution impacts spermatogenesis and semen quality are not fully understood, but may include endocrine disruption, inflammation, and oxidative stress in the testes or sperm leading to DNA changes ( 17 ). While prior research suggests outdoor air pollution may contribute to increased spontaneous miscarriage risk, there is heterogeneity across existing studies. For example, of nine previous studies three found an association between preconception ambient air pollution exposure and increased spontaneous miscarriage risk, including PM 2.5 , PM 10 , CO, O 3 , SO 2 , and NO 2 ( 18 – 20 ). Each of these studies defined “preconception” differently which might explain the differences in findings. The variation in study results could be due to differences in study design, enrollment during pregnancy, which could miss early pregnancy spontaneous miscarriages, and different exposure assessment methodologies. Although ambient air pollution is a complex mixture, no prior studies have used a mixtures framework to characterize exposures simultaneously, instead using single and multipollutant models ( 21 ). Additionally, many studies use fixed surface air monitoring data to estimate exposures at the regional level or to the nearest monitor, at times integrating participant addresses ( 21 ). This approach is less precise, with lower resolution and spatiotemporal variability than newer methods that incorporate pollution dispersion, chemical transformation of pollutants, meteorologic effects on pollutants, land use, and topography ( 21 ). Higher levels of vitamin D have been associated with improved reproductive function and might represent a protective factor against environmental insults ( 22 – 24 ). The associations of vitamin D with birth outcomes have been inconsistent, although one recent systematic review found that vitamin D deficiency in pregnancy was associated with increased preterm birth, growth restriction, and recurrent miscarriage while supplementation during pregnancy reduced the risk of preeclampsia and miscarriage ( 25 ). Vitamin D has both antioxidant and anti-inflammatory properties ( 26 , 27 ), which might protect against the toxic effects of air pollutants ( 28 , 29 ). Vitamin D serves as a cell membrane antioxidant and stimulates antioxidant gene expression. It also inhibits the production of pro-inflammatory cytokines ( 27 ). With widespread exposure to air pollution, identification of protective factors is urgently needed. This study uses data from Time to Conceive (TTC), a prospective pregnancy planning cohort in North Carolina, USA, to investigate potential associations between ambient air pollution and spontaneous miscarriage ( 30 ). The preconception enrollment allows the observation of early pregnancy spontaneous miscarriages and to collect detailed participant and partner information. To characterize exposure to PM 10 , PM 2.5 , CO, NO, NO 2 , SO 2 , and O 3 , we linked geocoded residential addresses with fusion and chemical transport model estimates at a 12km by 12km grid resolution that incorporate surface air monitoring data, climate and atmospheric data, and pollutant physiochemical properties. We then used single pollutant and exposure mixtures models (quantile-based g-computation) ( 31 ) to estimate associations between ambient air pollution exposure and spontaneous miscarriage in six critical windows linked to spermatogenesis, early pre-cycle follicle development, the follicular and luteal phases of the conception cycle, implantation, and the luteal-placental shift. Finally, we also investigated whether associations between air pollutants and spontaneous miscarriage differed when stratified by vitamin D level.